bf4a9329-ba54-4241-a6a3-4ec51aba7767

Lecture Notes in Financial Economics
c by Antonio Mele

London School of Economics & Political Science
April 2010
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Foundations
1 The classic capital asset pricing model

1.1 Portfolio selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.1 The wealth constraint . . . . . . . . . . . . . . . . . . . . . . .
1.1.2 Portfolio choice . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.3 Without the safe asset . . . . . . . . . . . . . . . . . . . . . . .
1.1.4 The market portfolio . . . . . . . . . . . . . . . . . . . . . . . .
1.2 The CAPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 The APT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.1 A rst derivation . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.2 The APT with idiosyncratic risk and a large number of assets .
1.3.3 Empirical evidence . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Appendix 1: Some analytical details for portfolio choice . . . . . . . . .
1.4.1 The primal program . . . . . . . . . . . . . . . . . . . . . . . .
1.4.2 The dual program . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Appendix 2: The market portfolio . . . . . . . . . . . . . . . . . . . . .
1.5.1 The tangent portfolio is the market portfolio . . . . . . . . . . .
1.5.2 Tangency condition . . . . . . . . . . . . . . . . . . . . . . . . .
1.6 Appendix 3: An alternative derivation of the SML . . . . . . . . . . . .
1.7 Appendix 4: Broader denitions of risk - Rothschild and Stiglitz theory
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
13
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
14
14
14
15
16
18
20
22
22
23
25
26
26
27
29
29
29
31
32
34
2 The CAPM in general equilibrium

35
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
c
by
A. Mele
Contents
2.2
The static general equilibrium in a nutshell . . .

2.2.1 Walras Law . . . . . . . . . . . . . . . .
2.2.2 Competitive equilibrium . . . . . . . . .
2.2.3 Optimality . . . . . . . . . . . . . . . . .
2.3 Time and uncertainty . . . . . . . . . . . . . . .
2.4 Financial assets . . . . . . . . . . . . . . . . . .
2.5 Absence of arbitrage . . . . . . . . . . . . . . .
2.5.1 How to price a nancial asset? . . . . . .
2.5.2 The Land of Cockaigne . . . . . . . . . .
2.6 Equivalent martingales and equilibrium . . . . .
2.6.1 The rational expectations assumption . .
2.6.2 Stochastic discount factors . . . . . . . .
2.6.3 Optimality and equilibrium . . . . . . .
2.7 Consumption-CAPM . . . . . . . . . . . . . . .
2.7.1 The risk premium . . . . . . . . . . . . .
2.7.2 The beta relation . . . . . . . . . . . . .
2.7.3 CCAPM & CAPM . . . . . . . . . . . .
2.8 Innite horizon . . . . . . . . . . . . . . . . . .
2.9 Further topics on incomplete markets . . . . . .
2.9.1 Nominal assets and real indeterminacy of
2.9.2 Nonneutrality of money . . . . . . . . .
2.10 Appendix 1 . . . . . . . . . . . . . . . . . . . .
2.11 Appendix 2: Proofs of selected results . . . . . .
2.12 Appendix 3: The multicommodity case . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
the equilibrium
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3 Infinite horizon economies

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Consumption-based asset evaluation . . . . . . . . . . . . . . . . . .
3.2.1 Recursive plans: introduction . . . . . . . . . . . . . . . . .
3.2.2 The marginalist argument . . . . . . . . . . . . . . . . . . .
3.2.3 Lucas model . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.4 Arrow-Debreu state prices, the CCAPM and the CAPM . .
3.3 Production: foundational issues . . . . . . . . . . . . . . . . . . . .
3.3.1 Decentralized economy . . . . . . . . . . . . . . . . . . . . .
3.3.2 Centralized economy . . . . . . . . . . . . . . . . . . . . . .
3.3.3 Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.4 Stochastic economies . . . . . . . . . . . . . . . . . . . . . .
3.4 Production-based asset pricing . . . . . . . . . . . . . . . . . . . . .
3.4.1 Firms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.2 Consumers . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.3 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Money, production, asset prices, and overlapping generations models
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
35
36
36
37
41
41
42
42
44
48
49
49
51
54
55
55
55
56
56
56
57
58
59
62
64
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
65
65
65
65
66
67
69
70
70
71
72
74
78
78
82
83
83
c
by
A. Mele
Contents
3.5.1 Introduction: endowment economies . . . . . . . . . . . . . .
3.5.2 Diamonds model . . . . . . . . . . . . . . . . . . . . . . . .
3.5.3 Money . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.4 Money in a model with real shocks . . . . . . . . . . . . . .
3.6 Optimality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6.1 Models with productive capital . . . . . . . . . . . . . . . .
3.6.2 Models with money . . . . . . . . . . . . . . . . . . . . . . .
3.7 Appendix 1: Finite dierence equations, with economic applications
3.8 Appendix 2: Neoclassic growth model - continuous time . . . . . . .
3.8.1 Convergence results . . . . . . . . . . . . . . . . . . . . . . .
3.8.2 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4 Continuous time models

4.1 Lambdas and betas in continuous time . . . . . . . . . . . . . . . . . .
4.1.1 The pricing equation . . . . . . . . . . . . . . . . . . . . . . . .
4.1.2 Expected returns . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.3 Expected returns and risk-adjusted discount rates . . . . . . . .
4.2 An introduction to arbitrage and equilibrium in continuous time models
4.2.1 A reduced-form economy . . . . . . . . . . . . . . . . . . . .
4.2.2 Preferences and equilibrium . . . . . . . . . . . . . . . . . . . .
4.2.3 Bubbles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.4 Reecting barriers and absence of arbitrage . . . . . . . . . . .
4.3 Martingales and arbitrage in a diusion model . . . . . . . . . . . . . .
4.3.1 The information framework . . . . . . . . . . . . . . . . . . . .
4.3.2 Viability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.3 Market completeness . . . . . . . . . . . . . . . . . . . . . . . .
4.4 Equilibrium with a representative agent . . . . . . . . . . . . . . . . . .
4.4.1 Consumption and portfolio choices: martingale approaches . . .
4.4.2 The older, Mertons approach: dynamic programming . . . . . .
4.4.3 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.4 Continuous-time Consumption-CAPM . . . . . . . . . . . . . .
4.5 Market imperfections and portfolio choice . . . . . . . . . . . . . . . .
4.6 Jumps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.1 Poisson jumps . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.2 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.3 Properties and related distributions . . . . . . . . . . . . . . . .
4.6.4 Some asset pricing implications . . . . . . . . . . . . . . . . . .
4.6.5 An option pricing formula . . . . . . . . . . . . . . . . . . . . .
4.7 Continuous-time Markov chains . . . . . . . . . . . . . . . . . . . . . .
4.8 Appendix 1: Convergence issues . . . . . . . . . . . . . . . . . . . . . .
4.9 Appendix 2: Proofs of selected results . . . . . . . . . . . . . . . . . . .
4.9.1 Proof of Theorem 4.2 . . . . . . . . . . . . . . . . . . . . . . . .
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 83
. 86
. 86
. 90
. 91
. 91
. 93
. 95
. 99
. 99
. 100
. 103
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
104
104
104
105
105
106
106
108
110
111
112
112
113
115
117
117
119
120
121
122
123
123
124
125
126
127
127
128
129
129
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
c
by
A. Mele
Contents
4.9.2 Proof of Eq. (4.40). . . . . . . . . . . . . .
4.9.3 Walrass consistency tests . . . . . . . . .
4.10 Appendix 3: The Greens function . . . . . . . . .
4.10.1 Setup . . . . . . . . . . . . . . . . . . . .
4.10.2 The PDE connection . . . . . . . . . . . .
4.11 Appendix 4: Portfolio constraints . . . . . . . . .
4.12 Appendix 5: Models with nal consumption only .
4.13 Appendix 6: Further topics on jumps . . . . . . .
4.13.1 The Radon-Nikodym derivative . . . . . .
4.13.2 Arbitrage restrictions . . . . . . . . . . . .
4.13.3 State price density: introduction . . . . . .
4.13.4 State price density: general case . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5 Taking models to data

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Data generating processes . . . . . . . . . . . . . . . . . .
5.2.1 Basics . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.2 Restrictions on the DGP . . . . . . . . . . . . . . .
5.2.3 Parameter estimators . . . . . . . . . . . . . . . . .
5.2.4 Basic properties of density functions . . . . . . . .
5.2.5 The Cramer-Rao lower bound . . . . . . . . . . . .
5.3 Maximum likelihood estimation . . . . . . . . . . . . . . .
5.3.1 Basics . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.2 Factorizations . . . . . . . . . . . . . . . . . . . . .
5.3.3 Asymptotic properties . . . . . . . . . . . . . . . .
5.4 M-estimators . . . . . . . . . . . . . . . . . . . . . . . . .
5.5 Pseudo (or quasi) maximum likelihood . . . . . . . . . . .
5.6 GMM . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.7 Simulation-based estimators . . . . . . . . . . . . . . . . .
5.7.1 Three simulation-based estimators . . . . . . . . . .
5.7.2 Asymptotic normality . . . . . . . . . . . . . . . .
5.7.3 Advances . . . . . . . . . . . . . . . . . . . . . . .
5.8 Spanning scores . . . . . . . . . . . . . . . . . . . . . . . .
5.9 Asset pricing, prediction functions, and statistical inference
5.10 Appendix 1: Proof of selected results . . . . . . . . . . . .
5.11 Appendix 2: Collected notions and results . . . . . . . . .
5.12 Appendix 3: Theory for maximum likelihood estimation . .
5.13 Appendix 4: Dependent processes . . . . . . . . . . . . . .
5.13.1 Weak dependence . . . . . . . . . . . . . . . . . . .
5.13.2 The central limit theorem for martingale dierences
5.13.3 Applications to maximum likelihood . . . . . . . .
5.14 Appendix 5: Proof of Theorem 5.4 . . . . . . . . . . . . . .
4
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
129
130
131
131
132
133
135
137
137
138
138
139
141
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
142
142
142
142
143
144
144
145
145
145
146
146
148
149
150
153
154
156
158
159
159
164
165
168
169
169
169
169
171
c
by
A. Mele
Contents
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
II
Asset pricing and reality
175
6 On kernels and puzzles

6.1 A single factor model . . . . . . . . . . . . . . . . . . .
6.1.1 The model . . . . . . . . . . . . . . . . . . . . .
6.1.2 Extensions . . . . . . . . . . . . . . . . . . . . .
6.2 The equity premium puzzle . . . . . . . . . . . . . . .
6.3 The Hansen-Jagannathan cup . . . . . . . . . . . . . .
6.4 Multifactor extensions . . . . . . . . . . . . . . . . . .
6.4.1 Exponential ane pricing kernels . . . . . . . .
6.4.2 Lognormal returns . . . . . . . . . . . . . . . .
6.5 Pricing kernels, Sharpe ratios and the market portfolio
6.5.1 What does a market portfolio do? . . . . . . . .
6.5.2 Final thoughts on the pricing kernel bounds . .
6.5.3 The Rolls critique . . . . . . . . . . . . . . . .
6.6 Appendix . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7 Aggregate stock market fluctuations
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . .
7.2 The empirical evidence . . . . . . . . . . . . . . . . . .
7.3 Understanding the empirical evidence . . . . . . . . . .
7.4 The asset pricing model . . . . . . . . . . . . . . . . .
7.4.1 A multidimensional model . . . . . . . . . . . .
7.4.2 A simplied version of the model . . . . . . . .
7.4.3 Issues . . . . . . . . . . . . . . . . . . . . . . .
7.5 Analyzing qualitative properties of models . . . . . . .
7.6 Time-varying discount rates and equilibrium volatility .
7.7 Large price swings as a learning induced phenomenon .
7.8 Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . .
7.8.1 Markov pricing kernels . . . . . . . . . . . . . .
7.8.2 The maximum principle . . . . . . . . . . . . .
7.8.3 Dynamic Stochastic Dominance . . . . . . . . .
7.8.4 Proofs . . . . . . . . . . . . . . . . . . . . . . .
7.8.5 On bond prices convexity . . . . . . . . . . . .
7.9 Appendix 2 . . . . . . . . . . . . . . . . . . . . . . . .
7.10 Appendix 3: Simulation of discrete-time pricing models
References . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8 Tackling the puzzles
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
176
176
176
179
179
180
182
182
184
185
185
187
189
191
194
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
195
195
195
201
201
201
202
204
205
208
213
217
217
218
219
220
221
222
222
224
226
c
by
A. Mele
Contents
8.1
Non-expected utility . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.1 The recursive formulation . . . . . . . . . . . . . . . . . . . . . .
8.1.2 Testable restrictions . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.3 Equilibrium risk-premia and interest rates . . . . . . . . . . . . .
8.1.4 Campbell-Shiller approximation . . . . . . . . . . . . . . . . . . .
8.1.5 Risks for the long-run . . . . . . . . . . . . . . . . . . . . . . . .
8.2 Catching up with the Joneses in a heterogeneous agents economy . . .
8.3 Incomplete markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.4 Limited stock market participation . . . . . . . . . . . . . . . . . . . . .
8.5 Appendix on non-expected utility . . . . . . . . . . . . . . . . . . . . . .
8.5.1 Detailed derivation of optimality conditions and selected relations
8.5.2 Details for the risks for the lung-run . . . . . . . . . . . . . . . .
8.5.3 Continuous time . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.6 Appendix on economies with heterogenous agents . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9 Information and other market frictions

9.1 Introduction . . . . . . . . . . . . . . . . . . . . .
9.2 Prelude: imperfect information in macroeconomics
9.3 Grossman-Stiglitz paradox . . . . . . . . . . . . .
9.4 Noisy rational expectations equilibrium . . . . . .
9.4.1 Dierential information . . . . . . . . . . .
9.4.2 Asymmetric information . . . . . . . . . .
9.4.3 Information acquisition . . . . . . . . . . .
9.5 Strategic trading . . . . . . . . . . . . . . . . . .
9.6 Dealers markets . . . . . . . . . . . . . . . . . . .
9.7 Noise traders . . . . . . . . . . . . . . . . . . . .
9.8 Demand-based derivative prices . . . . . . . . . .
9.8.1 Options . . . . . . . . . . . . . . . . . . .
9.8.2 Preferred habitat and the yield curve . . .
9.9 Over-the-counter markets . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . .
III
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Applied asset pricing theory
10 Options and volatility

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.2 General properties of options . . . . . . . . . . . . . . . . . . .
10.3 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.3.1 On spanning and cloning . . . . . . . . . . . . . . . . .
10.3.2 Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.3.3 Surprising cancellations and preference-free formulae
6
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
226
226
227
228
229
229
230
231
231
233
233
235
236
237
240
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
241
241
241
244
244
244
244
244
244
244
244
244
244
244
244
245
246
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
247
247
247
253
254
254
256
c
by
A. Mele
Contents
10.4 Properties of models . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.4.1 Rational price reaction to random changes in the state variables
10.4.2 Recoverability of the risk-neutral density from option prices . .
10.4.3 Endogeneous volatility . . . . . . . . . . . . . . . . . . . . . . .
10.5 Stochastic volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.5.1 Statistical models of changing volatility . . . . . . . . . . . . . .
10.5.2 Implied volatility and smiles . . . . . . . . . . . . . . . . . . . .
10.5.3 Stochastic volatility and market incompleteness . . . . . . . . .
10.5.4 Trading volatility . . . . . . . . . . . . . . . . . . . . . . . . . .
10.5.5 Pricing formulae . . . . . . . . . . . . . . . . . . . . . . . . . .
10.6 Local volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.6.1 Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.6.2 How does it work? . . . . . . . . . . . . . . . . . . . . . . . . .
10.6.3 Variance swaps . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.7 American options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.8 Exotic options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.9 Market imperfections . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.10Appendix 1: Additional details on Black & Scholes . . . . . . . . . . .
10.10.1 The original arguments . . . . . . . . . . . . . . . . . . . . . . .
10.10.2 Delta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.11Appendix 2: Stochastic volatility . . . . . . . . . . . . . . . . . . . . .
10.11.1 Proof of the Hull and White (1987) equation . . . . . . . . . . .
10.11.2 Simple smile analytics . . . . . . . . . . . . . . . . . . . . . . .
10.12Appendix 3: Technical details for local volatility models . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11 Interest rates
11.1 Prices and interest rates . . . . . . . . . . . . . . . . .
11.1.1 Introduction . . . . . . . . . . . . . . . . . . . .
11.1.2 Markets and interest rate conventions . . . . . .
11.1.3 The yield curve and forward rates . . . . . . . .
11.1.4 Forward martingale probabilities . . . . . . . .
11.2 Common factors aecting the yield curve . . . . . . . .
11.2.1 Methodological details . . . . . . . . . . . . . .
11.2.2 The empirical facts . . . . . . . . . . . . . . . .
11.3 Models of the short-term rate . . . . . . . . . . . . . .
11.3.1 Introduction . . . . . . . . . . . . . . . . . . . .
11.3.2 The basic bond pricing equation . . . . . . . . .
11.3.3 Some famous univariate short-term rate models
11.3.4 Multifactor models . . . . . . . . . . . . . . . .
11.3.5 Ane and quadratic term-structure models . .
11.3.6 Short-term rates as jump-diusion processes . .
11.3.7 Estimation strategies . . . . . . . . . . . . . . .
7
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
256
256
257
257
258
258
258
261
263
265
265
265
266
267
269
269
269
270
270
270
271
271
271
272
275
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
277
277
277
278
280
285
287
287
289
290
290
291
294
298
302
302
305
c
by
A. Mele
Contents
11.4 No-arbitrage models . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.4.1 Fitting the yield-curve, perfectly . . . . . . . . . . . . . . . . .
11.4.2 Ho and Lee . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.4.3 Hull and White . . . . . . . . . . . . . . . . . . . . . . . . . .
11.4.4 Critiques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.5 The Heath-Jarrow-Morton model . . . . . . . . . . . . . . . . . . . .
11.5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.5.2 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.5.3 The dynamics of the short-term rate . . . . . . . . . . . . . .
11.5.4 Embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.6 Stochastic string shocks models . . . . . . . . . . . . . . . . . . . . .
11.6.1 Addressing stochastic singularity . . . . . . . . . . . . . . . .
11.6.2 No-arbitrage restrictions . . . . . . . . . . . . . . . . . . . . .
11.7 Interest rate derivatives . . . . . . . . . . . . . . . . . . . . . . . . . .
11.7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.7.2 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.7.3 The put-call parity in xed income markets . . . . . . . . . .
11.7.4 European options on bonds . . . . . . . . . . . . . . . . . . .
11.7.5 Related pricing problems . . . . . . . . . . . . . . . . . . . . .
11.7.6 Market models . . . . . . . . . . . . . . . . . . . . . . . . . .
11.8 Appendix 1: The FTAP for bond prices . . . . . . . . . . . . . . . . .
11.9 Appendix 2: Certainty equivalent interpretation of forward prices . .
11.10Appendix 3: Additional results on T -forward martingale probabilities
11.11Appendix 4: Principal components analysis . . . . . . . . . . . . . . .
11.12Appendix 5: A few analytics for the Hull and White model . . . . . .
11.13Appendix 6: Expectation theory and embedding in selected models .
11.14Appendix 7: Additional results on string models . . . . . . . . . . . .
11.15Appendix 8: Changes of numeraire . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
12 Risky debt and credit derivatives

12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.2 Conceptual approaches to valuation of defaultable securities . . . . . . . . . .
12.2.1 Firms value, or structural, approaches . . . . . . . . . . . . . . . . . .
12.2.2 Reduced form approaches: rare events, or intensity, models . . . . . . .
12.2.3 Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.3 Derivatives on corporate assets . . . . . . . . . . . . . . . . . . . . . . . . . .
12.3.1 Callable and puttable bonds . . . . . . . . . . . . . . . . . . . . . . . .
12.3.2 Convertibles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.4 (Credit-) risk shifting derivatives and structured products . . . . . . . . . . . .
12.4.1 Securitization, and a brief history of credit risk and nancial innovation
12.4.2 Total Return Swaps (TRS) . . . . . . . . . . . . . . . . . . . . . . . . .
12.4.3 Spread Options (SOs) . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
307
307
309
310
310
311
311
311
312
313
314
314
315
316
316
317
317
318
321
326
333
335
336
337
338
339
341
342
344
.
.
.
.
.
.
.
.
.
.
.
.
348
348
348
348
357
360
364
364
365
368
368
371
372
c
by
A. Mele
Contents
12.4.4 Credit spread options (CSOs) . . . . . . . . . . . . . . . . . . . .

12.4.5 Credit Default Swaps (CDS) . . . . . . . . . . . . . . . . . . . . .
12.4.6 Collateralized Debt Obligations (CDOs) . . . . . . . . . . . . . .
12.4.7 One stylized numerical example of a structured product . . . . . .
12.5 A few hints on the risk-management practice . . . . . . . . . . . . . . . .
12.5.1 Value at Risk (VaR) . . . . . . . . . . . . . . . . . . . . . . . . .
12.5.2 Backtesting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.5.3 Stress testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.5.4 Credit risk and VaR . . . . . . . . . . . . . . . . . . . . . . . . .
12.6 Appendix 1: Proof of selected results . . . . . . . . . . . . . . . . . . . .
12.7 Appendix 2: Details on transition probability matrixes and pricing . . . .
12.8 Appendix 3: Derivation of bond spreads with stochastic default intensity
12.9 Appendix 4: Conditional probabilities of survival . . . . . . . . . . . . . .
12.10Appendix 5: Modeling correlation with copulae functions . . . . . . . . .
12.11Appendix 6: Details on CDO pricing with imperfect correlation . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13 Financial engineering and fixed income securities
13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
13.1.1 Relative pricing in xed income markets . . . . . .
13.1.2 Complexity of xed income securities . . . . . . . .
13.1.3 Many evaluation paradigms . . . . . . . . . . . . .
13.2 Bootstrapping and curve tting . . . . . . . . . . . . . . .
13.2.1 Extracting zeros from bond prices . . . . . . . . . .
13.2.2 Bootstrapping . . . . . . . . . . . . . . . . . . . . .
13.2.3 Curve tting . . . . . . . . . . . . . . . . . . . . . .
13.3 Duration, convexity and asset liability management . . . .
13.3.1 Duration . . . . . . . . . . . . . . . . . . . . . . . .
13.3.2 Convexity . . . . . . . . . . . . . . . . . . . . . . .
13.3.3 Duration and asset-liability management . . . . . .
13.4 Foundational issues on interest rate modeling . . . . . . .
13.4.1 Tree representation of the short-term rate . . . . .
13.4.2 Tree pricing . . . . . . . . . . . . . . . . . . . . . .
13.5 The Ho and Lee model . . . . . . . . . . . . . . . . . . . .
13.5.1 The tree . . . . . . . . . . . . . . . . . . . . . . . .
13.5.2 The price movements and the martingale restriction
13.5.3 The recombining condition . . . . . . . . . . . . . .
13.5.4 Calibration of the model . . . . . . . . . . . . . . .
13.5.5 An example . . . . . . . . . . . . . . . . . . . . . .
13.6 Beyond Ho and Lee: Calibration . . . . . . . . . . . . . . .
13.6.1 Arrow-Debreu securities . . . . . . . . . . . . . . .
13.6.2 The algorithm in two examples . . . . . . . . . . .
13.7 Copying with credit risk . . . . . . . . . . . . . . . . . . .
9
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
372
372
383
392
399
399
403
403
404
407
408
410
411
412
414
415
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
417
417
417
417
418
418
418
419
420
421
421
423
423
430
431
433
446
446
447
448
450
450
454
454
456
461
c
by
A. Mele
Contents
13.7.1 Callable bonds . . . . . .
13.7.2 Convertible bonds . . . . .
13.8 Appendix 1: Proof of Eq. (13.9) .
13.9 Appendix 2: Proof of Eq. (13.24)
References . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
10
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
462
462
464
466
468
Many of the models in the literature are not general equilibrium models in my sense. Of
those that are, most are intermediate in scope: broader than examples, but much narrower
than the full general equilibrium model. They are narrower, not for carefully-spelled-out
economic reasons, but for reasons of convenience. I dont know what to do with models
like that, especially when the designer says he imposed restrictions to simplify the model
or to make it more likely that conventional data will lead to reject it. The full general
equilibrium model is about as simple as a model can be: we need only a few equations to
describe it, and each is easy to understand. The restrictions usually strike me as extreme.
When we reject a restricted version of the general equilibrium model, we are not rejecting
the general equilibrium model itself. So why bother testing the restricted version?
Fischer Black, 1995, p. 4, Exploring General Equilibrium, The MIT Press.
Preface
The present Lecture Notes in Financial Economics are based on my teaching notes for advanced
undergraduate and graduate courses in nancial economics, macroeconomic dynamics, nancial
econometrics and nancial engineering. Part I, Foundations, develops the fundamentals tools
of analysis used in Part II and Part III. These tools span such disparate topics as classical
portfolio selection, dynamic consumption- and production- based asset pricing, in both discrete
and continuous-time, the intricacies underlying incomplete markets and some other market
imperfections and, nally, econometric tools comprising maximum likelihood, methods of moments, and the relatively more modern simulation-based inference methods. Part II, Asset
pricing and reality, is about identifying the main empirical facts in nance and the challenges
they pose to nancial economists: from excess price volatility and countercyclical stock market
volatility, to cross-sectional puzzles such as the value premium. This second part reviews the
main models aiming to take these puzzles on board. Part III, Applied asset pricing theory,
aims just to this: to use the main tools in Part I and cope with the main challenges occurring
in actual capital markets, arising from option pricing and trading, interest rate modeling and
credit risk and their associated derivatives. In a sense, Part II is about the big puzzles we face
in fundamental research, while Part III is about how to live within our current and certainly
unsatisfactory paradigms, so as to cope with demand for intellectual expertise.
These notes are still underground. The economic motivation and intuition are not always developed as deeply as they deserve, some derivations are inelegant, and sometimes, the English
is a bit informal. Moreover, I still have to include material on asset pricing with asymmetric
information, monetary models of asset prices, bubbles, asset prices implications of overlapping
generations models, or nancial frictions. Finally, I need to include more extensive surveys for
each topic I cover, especially in Part II. I plan to revise these notes to ll these gaps. Meanwhile,
any comments on this version are more than welcome.
Antonio Mele
April 2010
Part I
Foundations
13
1
The classic capital asset pricing model
1.1 Portfolio selection

An investor is concerned with choosing a number of assets to include in his portfolio. Which
weigths each asset must bear for the investor to maximize some utility criterion? This section
deals with this problem when our investor maximizes a mean-variance criterion, as in the seminal
approach of Markovitz (1952). First, we derive the wealth constraint. Second, we illustrate the
main results of the model, with and without a safe asset. Third, we introduce the notion of the
market portfolio.
1.1.1
The wealth constraint
The space choice comprises m risky assets, and some safe asset. Let S = [S1 , , Sm ] be the
risky assets price vector, and let S0 be the price of the riskless asset. We wish to evaluate the
value of a portfolio that contains all these assets. Let = [1 , , m ], where i is the number
of the i-th risky asset, and let 0 be the number of the riskless assets, in this portfolio. The
initial wealth is, w = S0 0 + S . Terminal wealth is w+ = x0 0 + x , where x0 is the payo
promised by the riskless asset, and x = [x1 , , xm ] is the vector of the payos pertaining to
the risky assets, i.e. xi is the payo of the i-th asset.
The following pieces of notation considerably simplify the presentation. Let R Sx00 , and
i xi . In words, R is the gross interest rate obtained by investing in a safe asset, and R
i is
R
Si
the gross return obtained by investing in the i-th risky asset. Accordingly, we dene r R 1
i 1 is the rate of return on the i-th
as the safe interest rate; b = [b1 , , bm ], where bi R
asset; and b E(b), the vector of the expected returns on the risky assets. Finally, we let
= [ 1 , , m ], where i i Si is the wealth invested in the i-th asset. We have,
+
w = x0 0 +
m

i=1
xi i R 0 +
m

i=1
ii
R
and w = 0 +
m

i=1
i .
(1.1)
c
by
A. Mele
1.1. Portfolio selection
Combining the two expressions for w+ and w, we obtain, after a few simple computations,
1m R) + Rw = (b 1m r) + Rw + (b b).
w + = (R
We use the decomposition, b b = a u, where a is a m d volatility matrix, with m d,
and u is a random vector with expectation zero and variance-covariance matrix equal to the
identity matrix. With this decomposition, we can rewrite the budget constraint in Eq. (1.1) as
follows:
w+ = (b 1m r) + Rw + a
u.
(1.2)
We now use Eq. (1.2) to compute the expected return and the variance of the portfolio value.
We have,

E w+ () = (b 1m r) + Rw and var w + () =
(1.3)
where aa . Let 2i ii . We assume that has full-rank, and that,
2i > 2j bi > bj all i, j,
which implies that r < minj (bj ).
1.1.2
Portfolio choice
We assume that the investor maximizes the expected return on his portfolio, given a certain
level of the variance of the portfolios value, which we set equal to w2 vp . So we use Eq. (1.3)
to set up the following program
+
+
(vp ) = arg max
E
w
()
s.t.
var
w () = w 2 vp2 .
[1.P1]
m
R
The rst order conditions for [1.P1] are,
(vp ) = (2)1 1 (b 1m r)
and

= w2 vp2 ,
where is a Lagrange multiplier for the variance constraint. By plugging the rst condition
p , where
into the second, we obtain, (2)1 = wv
Sh
Sh (b 1m r) 1 (b 1m r) ,
(1.4)
is the Sharpe market performance. To ensure eciency, we take the positive solution. Substituting the positive solution for (2)1 into the rst order condition, we obtain that the portfolio
that solves [1.P1] is
(vp )
1 (b 1m r)
vp .
(1.5)
w
Sh
We are now ready to calculate the value of [1.P1], E [w+ (
(vp ))] and, hence, the expected
portfolio return, dened as,
p (vp )
E [w+ (
(vp ))] w
= r + Sh vp ,
w
(1.6)
where the last equality follows by simple computations. Eq. (1.6) describes what is known as
the Capital Market Line (CML).
15
c
by
A. Mele

1.1.3
Without the safe asset
Next, let us suppose the investors space choice does not include the riskless asset. In this case,

m
+
his current wealth is w = m
,
and
his
terminal
wealth
is
w
=
i
i=1
i=1 Ri i . By the denition
i 1, and by a few simple computations,
of bi R
+
w =
m

i=1
bi i +
m
i = b + w + a
u,
(1.7)
i=1
where a and u are as dened as in Eq. (1.2). We can use Eq. (1.7) to compute the expected
return and the variance of the portfolio value, which are:

E w+ () = b + w, where w = 1m and var w + () = .
(1.8)
The program our investor solves, now, is:

(vp ) = arg max E w+ ()

s.t. var w+ () = w2 vp2 and w = 1m .
R
[1.P2]
In the appendix, we show that provided 2 > 0 (a second order condition), the solution
to [P2] is,
p (vp ) 1
p (vp ) 1
(vp )
=
1m ,
(1.9)
2 b+
w

2
1
1
where b 1 b, 1
m b and 1m 1m , and p (vp ) is the expected portfolio return,
dened as in Eq. (1.6). In the appendix, we also show that,

2

1
1
2
vp =
1+
p (vp )
.
(1.10)
Therefore, the global minimum variance portfolio achieves a variance equal to vp2 = 1 and an
expected return equal to p = / .
Note that for each vp , there are two values of p (vp ) that solve Eq. (1.10). The optimal choice
for our investor is that with the highest p . We dene the efficient portfolio frontier as the set
of values (vp , p ) that solve Eq. (1.10) with the highest p . It has the following expression,

1 2
p (vp ) = +
vp 1 2 .
(1.11)

Clearly, the ecient portfolio frontier is an increasing and concave function of vp . It can be
interpreted as a sort of production function, one that produces expected returns through
inputs of levels of risk (see, e.g., Figure 1.1). The choice of which portfolio has eectively to
be selected depends on the investors preference toward risk.
Example 1.1. Let the number of risky assets m = 2. In this case, we do not need to
optimize anything, as the budget constraint, w1 + w2 = 1, pins down an unique relation between
the portfolio expected return and the variance of the portfolios value. So we simply have,
E [w+ ()]w
p =
= w1 b1 + w2 b2 , or,
w
p = b1 + (b2 b1 ) 2
2 w

2
v2 = 1 2 2 + 2 1 2 2 12 + 2 2
p
1
2
w
w w
w
16
c
by
A. Mele
0.15
0.14
Expected return, mu
= 1
= 0.5
0.13
=0
= 0.5
0.12
=1
0.11
0.1
0.09
0.05
0.1
0.15
0.2
0.25
Volatility, v
FIGURE 1.1. From top to bottom: portfolio frontiers corresponding to = 1, 0.5, 0, 0.5, 1. Parameters are set to b1 = 0.10, b2 = 0.15, 1 = 0.20, 2 = 0.25. For each portfolio frontier, the ecient
portfolio frontier includes those portfolios which yield the lowest volatility for a given expected return.
whence:
1
vp =
b2 b1
b2 p
2

2
21 + 2 b2 p p b1 1 1 + p b1 22
When = 1,
p = b1 +
(b1 b2 ) ( 1 vp )
.
2 1
In the general case, diversication pays when the asset returns are not perfectly positively
correlated (see Figure 1.1). As Figure 1.1 reveals, it is even possible to obtain a portfolio that
is less risky than than the less risky asset. Moreover, risk can be zeroed when = 1, which
2
1
2
1
corresponds to w1 = 2
and w2 = 2
or, alternatively, to w1 = 2
and w2 = 2
.
1
1
1
1
Let us return to the general case. The portfolio in Eq. (1.9) can be decomposed into two
components, as follows:
d
g
(vp )
= (vp )
+ [1 (vp )] ,
w
w
w

p (vp )
(vp )
,
2
where
1 b
d
,
w
g
1 1m
.
w
17
c
by
A. Mele
g
is the global minimum variance portfolio, for we know from Eq. (1.10)
w

1
that the minimum variance occurs at (vp , p ) =
,
, in which case (vp ) = 0.1 More

generally, we can span any portfolio on the frontier by just choosing a convex combination of
g
d
and
, with weight equal to (vp ). Its a mutual fund separation theorem.
w
w
Hence, we see that
1.1.4
The market portfolio
The market portfolio is the portfolio at which the CML in Eq. (1.6) and the ecient portfolio
frontier in Eq. (1.11) intersect. In fact, the market portfolio is the point at which the CML is
tangent at the ecient portfolio frontier. For this reason, the market portfolio is also referred
to as the tangent portfolio. In Figure 1.2, the market portfolio corresponds to the point M
(the portfolio with volatility equal to vM and expected return equal to M ), which is the point
at which the CML is tangent to the ecient portfolio frontier, AM C.2
As Figure 1.2 illustrates, the CML dominates the ecient portfolio frontier AM C. This is
because the CML is the value of the investors problem, [1.P1], obtained using all the risky
assets and the riskless asset, and the ecient portfolio frontier is the value of the investors
problem, [P2], obtained using only all the risky assets.3 For the same very reason, the CML
and the ecient portfolio frontier can only be tangent with each other. For suppose not. Then,
there would exist a point on the ecient portfolio frontier that dominates some portfolio on the
CML, a contradiction. Likewise, the CML must have a portfolio in common with the ecient
portfolio frontier - the portfolio that does not include the safe asset. Below, we shall use this
insight to characterize, analytically, the market portfolio.
Why is the market portfolio called in this way? Figure 1.2 reveals that any portfolio on the
CML can be obtained as a combination of the safe asset and the market portfolio M (a portfolio
containing only the risky assets). An investor with high risk-aversion would like to choose a
point such as Q, say. An investor with low risk-aversion would like to choose a point such as P ,
say. But no matter how risk averse an individual is, the optimal solution for him is to choose
a combination of the safe asset and the market portfolio M . Thus, the market portfolio plays
an instrumental role. It obviously does not depend on the risk attitudes of any investor - it is a
mere convex combination of all the existing assets in the economy. Instead, the optimal course
of action for any investor is to use those proportions of this portfolio that make his overall
exposure to risk consistent with his risk appetite. Its a two fund separation theorem.
The equilibrium implications, then, are as follows. As we have explained, any portfolio can be
attained by lending or borrowing funds in zero net supply, and in the portfolio M . In equilibrium,
then, every investor must hold some proportions of M. But since in aggregate, there is no net
borrowing or lending, one has that in aggregate, all investors must have portfolio holdings that
sum up to the market portfolio, which is therefore the value-weighted portfolio of all the existing
assets in the economy. This argument is formally developed in the appendix.
1 It
is easy to show that the covariance of the global minimum variance portfolio with any other portfolio equals 1 .
existence of the market portfolio requires a restriction on r, derived in Eq. (1.12) below.
3 Figure 1.2 also depicts the dotted line MZ, which is the value of the investors problem when he invests a proportion higher
than 100% in the market portfolio, leveraged at an interest rate for borrowing higher than the interest rate for lending. In this case,
the CML coincides with rM, up to the point M. From M onwards, the CML coincides with the highest between MZ and MA.
2 The
18
c
by
A. Mele
CML
A
Q
C
r
vM
FIGURE 1.2.
We turn to characterize the market portfolio. We need to assume that the interest rate is
suciently low to allow the CML to be tangent at the ecient portfolio frontier. The technical
condition that ensures this is that the return on the safe asset be less than the expected return
on the global minimum variance portfolio, viz
r<
(1.12)
Let M be the market portfolio. To identify M , we note that it belongs to AM C if

M 1m = w,
where M also belongs to the CML and, therefore, by Eq. (1.5), is such that:
M
1 (b 1m r)
=
vM .
w
Sh
(1.13)
Therefore, we must be looking for the value vM that solves
w = 1
m M = w 1m
i.e.
vM
1 (b 1m r)
vM ,
Sh
Sh
=
.
r
(1.14)
Then, we plug this value of vM into the expression for M in Eq. (1.13) and obtain,4
M
1
=
1 (b 1m r) .
w
r
4 While
the market portfolio depends on r, this portfolio does not obviously include any share in the safe asset.
19
(1.15)
c
by
A. Mele
1.2. The CAPM
Naturally, the market portfolio belongs to the ecient portfolio frontier. Indeed, on the
one hand, the market portfolio can not be above the ecient portfolio frontier, as this would
contradict the eciency of the AM C curve, which is obtained by investing in the risky assets
only; on the other hand, the market portfolio can not be below the ecient portfolio frontier, for
by construction, it belongs to the CML which, as shown before, dominates the ecient portfolio
frontier. In the appendix, we conrm, analytically, that the market portfolio does indeed enjoy
the tangency condition.
1.2 The CAPM

The Capital Asset Pricing Model (CAPM) provides an asset evaluation formula. In this section,
we derive the CAPM through arguments that have the same avor as the original derivation of
Sharpe (1964). The rst step is the creation of a portfolio including a proportion of wealth
invested in any asset i and the remaining proportion 1 invested in the market portfolio.
Mathematically, we are considering an -parametrized portfolio, with expected return and
volatility given by:

p
bi + (1 )M
(1.16)
vp (1 )2 2M + 2(1 ) iM + 2 2i
where we have dened M vM . Clearly, the market portfolio, M , belongs to the -parametrized
portfolio. By the Example 1.1, the curve in (1.16) has the same shape as the curve A M i in
Figure 2.3. The curve A M i lies below the ecient portfolio frontier AM C. This is because
the ecient portfolio frontier is obtained by optimizing a mean-variance criterion over all the
existing assets and, hence, dominates any portfolio that only comprises the two assets i and
M . Suppose, for example, that the A M i curve intersects the AM C curve; then, a feasible
combination of assets (including a proportion of the i-th asset and a proportion 1 of the
market portfolio) would dominate AMC, a contradiction, given that AM C is the most ecient
feasible combination of all the assets. Therefore, the curve A M i is tangent to the ecient
portfolio frontier AMC at M , which in turn, as we know, is tangent to the CML at M .
Let us equate, then, the two slopes of the A M i curve and the ecient portfolio frontier
AM C at M . We shall show that this condition provides a restriction on the expected return bi
on any asset i. Because (1.16) is, mathematically, an -parametrized curve, we may compute

its slope at M through the computation of d
p d and d
vp / d, at = 0. We have,

d
p
(1 ) 2M + (1 2) iM + 2i |=0
d
vp
1
2
= bi M ,
=
.
iM
M
d
d =0
vp |=0
M
Therefore,

d
p ()

=
d
vp () =0
1
M
bi M
.
( iM 2M )
(1.17)
On the other hand, the slope of the CML is (M r)/ M which, equated to the slope in Eq.
(1.17), yields,
iM
bi r = i (M r) , i 2 , i = 1, , m.
(1.18)
vM
20
c
by
A. Mele
1.2. The CAPM
CML
A
M
i
C
r
vM
FIGURE 1.3.
Eq. (1.18) is the celebrated Security Market Line (SML). The appendix provides an alternative
derivation of the SML. Assets with i > 1 are called aggressive assets. Assets with i < 1
are called conservative assets.
Note, the SML can be interpreted as a projection of the excess return on asset i (i.e. bi r)
on the excess returns on the market portfolio (i.e. bM r). In other words,
bi r = i (bM r) + i ,
i = 1, , m.
(1.19)
The previous relation leads to the following decomposition of the volatility (or risk) related to
the i-th asset return:
2
2i = 2i vM
+ var (i ) ,
i = 1, , m.
2
The quantity 2i vM
is usually referred to as systematic risk. The quantity var (i ) 0, instead,
is what we term idiosyncratic risk. In the next section, we shall show that idiosyncratic risk
can be eliminated through a well-diversied portfolio - roughly, a portfolio that contains a
large number of assets. Naturally, economic theory does not tell us anything substantial about
how important idiosyncratic risk is for any particular asset.
The CAPM can be usefully interpreted within a classical hedging framework. Suppose we
hold an asset that delivers a return equal to z - perhaps, a nontradable asset. We wish to
hedge against movements of this asset by purchasing a portfolio containing a percentage of
in the market portfolio, and a percentage of 1 units in a safe asset. The hedging criterion
we wish to use is the variance of the overall exposure of the position, which we minimize by
z ((1 ) r + bM )]. It is straight forward to show that the solution to this basic
min var[
2
problem is,
z cov(
z , bM )/vm
. That is, the proportion to hold is simply the beta of the
asset to hedge with the market portfolio.
21
c
by
A. Mele
1.3. The APT
The CAPM is a model for the required return for any asset and so, it is a very rst tool we
can use to evaluate risky projects. Let
V = value of a project =
E (C + )
,
1 + rC
where C + is future cash ow and rC is the risk-adjusted discount rate for this project. We have:
E (C + )
= 1 + rC
V
= 1 + r + C (M r)

+
C
cov V 1, xM
= 1+r+
(M r)
2
vM
1 cov (C + , xM )
= 1+r+
(M r)
2
V
vM

1
= 1 + r + cov C + , xM
,
V
vM
r
where M
, the unit market risk-premium.
vM
Rearranging terms in the previous equation leaves:
V =
E (C + )
cov (C + , xM )
vM
1+r
(1.20)
The certainty equivalent C is dened as:

E (C + )
C
C : V =
=
,
1 + rC
1+r
or,
C = (1 + r) V,
and using Eq. (1.20),

C = E C +
cov C + , xM .
vM
1.3 The APT

1.3.1
A first derivation
Suppose that the m asset returns we observe are generated by the following linear factor model,
b
m1
a + B f
m1
mk
k1
a + cov(b, f )[var(f)]1 f
(1.21)
where a and B are a vector and a matrix of constants, and f is a k-dimensional vector of factors
supposed to aect the asset returns, with k m. Let us normalize [var(f)]1 = Ikk , so that
B = cov(b, f ). With this normalization, we have,
k
b1 , fj )fj
cov(
cov(b1 , f )
j=1
..
..
b = a +
f
=
a
+
.
.
.

k
cov(bm , f )
cov(bm , fj )fj
j=1
22
c
by
A. Mele
1.3. The APT
Next, let us consider a portfolio including the m risky assets. The return of this portfolio
is,
b = a + Bf,
where as usual, 1m = 1. An arbitrage opportunity arises if there exists some portfolio
such that the return on the portfolio is certain, and dierent from the safe interest rate r, i.e. if
: B = 0 and a = r. Mathematically, this is ruled out whenever Rk : a = B+1m r.
Substituting this relation into Eq. (1.21) leaves,
b = 1m r + B + Bf = 1m r + cov(b, f ) + cov(b, f )f.
Taking the expectation,
bi = r + (B)i = r +
k
j=1
cov(b , f ) ,
i j j
i = 1, , m.
(1.22)
i,j
The APT collapses to the CAPM, once we assume that the only factor aecting the returns
is the market portfolio. To show this, we must normalize the market portfolio return so that its
variance equals one, consistently with Eq. (1.22). So let rM be the normalized market return,
1
dened as rM vM
bM , so that var(
rM ) = 1. We have,
bi = a + i rM ,
i = 1, , m,
1
where i = cov(bi , rM ) = vM
cov(bi , bM ). Then, we have,
bi = r + i ,
i = 1, , m.
(1.23)
1
var(bM ) = vM , and so, by Eq. (1.23),
In particular, M = cov(bM , rM ) = vM
bM r
,
vM
which is known as the Sharpe ratio for the market portfolio, or the market price of risk.
1
By replacing i = vM
cov(bi , bM ) and the expression for above into Eq. (1.23), we obtain,
bi = r +
cov(bi , bM )
(bM r) ,
2
vM
i = 1, , m.
This is simply the SML in Eq. (1.18).

1.3.2
The APT with idiosyncratic risk and a large number of assets
[Ross (1976), and Connor (1984), Huberman (1983).]

How can idiosyncratic risk be eliminated? Consider, for example, Eq. (1.19). Intuitively, we
may form portfolios with a large number of assets, so as to make idiosyncratic risk negligible, by
the law of large numbers. But would the beta-relation still hold, in this case? More in general,
would the APT relation in Eq. (1.22) be still valid? The answer is in the armative, although
it deserves some qualications.
23
c
by
A. Mele
1.3. The APT
Consider the APT equation (1.21), and add a vector of idiosyncratic returns, , which are
independent of f , and have mean zero and variance 2 :
b = a + B f + .
We wish to show that in the absence of arbitrage, to be dened below, it must be that the
number of assets such that Eq. (1.22) does not hold, N (m) say, is bounded as m gets large,
i.e.:
|ai ((B)i + r)| > 0, i = 1, , N (m) ,
(1.24)
where
lim N (m) < .
(1.25)
In other words, we wish to show that in a large market, Eq. (1.22) does indeed hold for most
of the assets, an approach close to that in Huang and Litzenberger (1988, p. 106-108).
By the same arguments leading to Eq. (1.1), the wealth generated by a portfolio of the assets
+
satisfying (1.24), wN(m)
say, is,

+
wN(m)
=
a
1
r
+
Rw
+
B
f
+
,
N(m)
N(m)
N(m)
N(m)
N(m)
N(m)
N(m)
where aN , BN and N are (i) the vector of the expected returns, (ii) the return volatility (or
factor exposures) matrix and (iii) the vector of idiosyncratic return components aecting these
assets, and, nally, N and wN are the portfolio and the initial wealth invested in these assets.
In this context, we may dene an arbitrage as the portfolio N(m) that in the limit, as the
number of all the existing assets m gets large, is riskless and yet delivers an expected return
strictly larger than the safe interest rate, viz
lim
+
E[wN(m)
]
wN(m)
> R,
and
+
lim var[wN(m)
] 0.
(1.26)
We want to show that this situation does not arises, under the condition in (1.25), thereby
establishing that the linear APT relation in Eq. (1.22) is valid for most of the assets, in a large
market.
So suppose the linear relation, aN 1N r = BN , doesnt hold. Then, there exists a portfolio
such that,
BN = 0 and (aN 1N r) = 0.
(1.27)
Consider the portfolio:

1
sign (aN 1N r) ,
N
+
where is as in (1.27). With this portfolio we have, clearly, that E[wN
]=
N (aN 1N r) +
+
RwN > RwN , for each N, and even for N large. That is, limm E[wN (m) ]/wN(m) > R, which
is the rst condition in (1.26). As regards the second condition in (1.26), we have that

+
2
var[wN
]=
N = 2
N ,
N BN BN + INN
N
N =
+
where the second equality follows by the rst relation in (1.27). Clearly, limm var[wN(m)
]0
as N (m) . Hence, in the absence of arbitrage, the condition in (1.25) must hold.
24
c
by
A. Mele
1.3. The APT

1.3.3
Empirical evidence
Fama-MacBeth. Economic forces driving asset returns.
25
c
by
A. Mele
1.4. Appendix 1: Some analytical details for portfolio choice
1.4 Appendix 1: Some analytical details for portfolio choice

We derive Eq. (1.9), which provides the solution for the portfolio choice when the space choice does not
include a safe asset. We derive the solution by proceeding with two programs: (i) the primal program
[P2] in the main text, which consists in maximizing the portfolio expected return, given a certain level
of the variance of the portfolios value; and (ii) a dual program, to be introduced below, by which one
minimizes the variance of the portfolios value, given a certain level of the portfolio expected return.
1.4.1
The primal program
Given Eq. (1.8), the Lagrangian function associated to [P2] is,

L = b + w 1 ( w2 vp2 ) 2 ( 1m w),
where 1 and 2 are two Lagrange multipliers. The rst order conditions are,
1 1
(b 2 1m ) ,
2 1

= w2 vp2 ,
1m = w.
(1A.1)
Using the rst and the third conditions, we obtain,

w = 1
=
m
1
1
1
1
(1
( 2 ).
m b 2 1m 1m )

2 1
2 1
We can solve for 2 , obtaining,
2 =
2w 1
.
By replacing the solution for 2 into the rst condition in (1A.1) leaves,

w 1
1 1
= 1m +
b 1m .
2 1
(1A.2)
Next, we derive the value of the program [P2]. We have,

1
w
1 1
w
E w+ (
) w =
b = 1
1 b +
(b b 1 1 b) = +
m 2 1 m
2 1

2
(1A.3)
It is easy to check that

var w+ (
) = w2 vp2

w 1
1

w
1
1
=
1 +
b 1m
1m +
b 1m
m
2 1
2 1

2
2
2
w
1
=
+
2 1
Let us gather Eqs. (1A.3) and (1A.4),

2
+ (
E
[w
)]
= +
p (vp )
w
2
w
1

2
2
1
1
vp2 = +
2 1 w
26
(1A.4)
(1A.5)
c
by
A. Mele
where we have emphasized the dependence of p on vp , which arises through the presence of the
Lagrange multiplier 1 .
Let us rewrite the rst equation in (1A.5) as follows,

1

1
= 2
p (vp ) .
2 1 w
(1A.6)
We can use this expression for 1 to express

in Eq. (1A.2) in terms of the portfolio expected return,
p (vp ). We have,

1 1m
1
2 1
1
p (vp ) b
=
+
1m .
w
By rearranging terms in the previous equation, we obtain Eq. (1.9) in the main text.
Finally, we substitute Eq. (1A.6) into the second equation in (1A.5), and obtain:
vp2 =

1
2 !
1
1 + 2
p (vp )
,
which is Eq. (1.10) in the main text. Note, also, that the second condition in (1A.5) reveals that,

1
2 1 w
2
vp2 1
Given that 2 > 0, the previous equation conrms the properties of the global minimum variance
portfolio stated in the main text.
1.4.2
The dual program
We now solve the dual program, dened as follows,

+

w ()
= arg minm var

s.t. E w+ () = Ep and w = 1m ,
R
w
[1A.P2-dual]
for some constant Ep . The rst order conditions are
1 w 1
2 w 1
=
b+
1m
w
2
2
b = Ep w
w=
1m ;
(1A.7)
where 1 and 2 are two Lagrange multipliers. By replacing the rst condition in (A14) into the second
one,

1
2 1
2
1
2 1
Ep w =
b = w2 ( b
b
+
1
b
)
+
.
(1A.8)

2
2 m
2
2
By replacing the rst condition in (A14) into the third one,

w=
1m = w2 (

1 1
2
2
1
2 1
b 1m + 1
1
)
+
.
m
2
2 m
2
2
Next, let p
Ep w
w .
By Eqs. (1A.8) and (1A.9), the solutions for 1 and 2 are,

p
1w
=
2
2
27
p
2w
=
2
2
(1A.9)
c
by
A. Mele
Therefore, the solution for the portfolio in Eq. (A14) is,

p 1
p 1
b+
1m .
=
2
w

2
Finally, the value of the program is,
+

2p 2p +
(p )2
w (
1 p
1 p
1
)
1
var
= 2

=
b
+
1
=
=
+ ,
m
2
2
2
2
w
w
w
w

( )
which is exactly Eq. (1.10) in the main text.
28
c
by
A. Mele
1.5. Appendix 2: The market portfolio
1.5 Appendix 2: The market portfolio

1.5.1
The tangent portfolio is the market portfolio
Let us dene the market capitalization for any asset i as the value of all the assets i that are outstanding
in the market, viz
Capi i Si , i = 1, , m,
where i is the number of assets i outstanding in the market. The market capitalization of all the
assets is simply
m

CapM
Capi .
i=1
The market portfolio, then, is the portfolio with relative weights given by,
M,i
Capi
,
CapM
i = 1, , m.
Next, suppose there are N investors and that each investor j has wealth wj , which he invests in two
funds, a safe asset and the tangent portfolio. Let wjf be the wealth investor j invests in the safe asset
and wj wjf the remaining wealth the investor invests in the tangent portfolio. The tangent portfolio

is dened as
T wTj , for some T solution to [P2], and is obviously independent of wj (see Eq.
(1.15) in the main text). The equilibrium in the stock market requires that
CapM
M =
N

j=1
wj wjf
T =
N

j=1
wj
T = CapM
T .

f
where the second equality follows because the safe asset is in zero net supply and, hence, N
j=1 wj = 0;
and the third equality holds because all the wealth in the economy is invested in stocks, in equilibrium.
1.5.2
Tangency condition
We check that the CML and the ecient portfolio frontier have the same slope in correspondence of
the market portfolio. Let us impose the following tangency condition of the CML with the ecient
portfolio frontier in Figure 1.2, AMC, at the point M:
2
Sh =
vM .
M
(1A.10)
The left hand side of this equation is the slope of the CML, obtained through Eq. (1.6). The right hand
side is the slope of the ecient portfolio frontier, obtained by dierentiating p (v) in the expression
for the portfolio frontier in Eq. (1.11), and setting v = vM in

dp (v) 2
2
= (v 1)1 2 v =
v,
dv
p (v)
and where the second equality follows, again, by Eq. (1.11). By Eqs. (1A.10) and (1.14), we need to
show that,
M
1
2 = r .

By plugging M = r + Sh vM into the previous equality and rearranging terms,
Sh
,
vM =
r
29
c
by
A. Mele
1.5. Appendix 2: The market portfolio
where we have made use of the equality Sh = 2r + r2 , obtained by elaborating on the denition
of the Sharpe market performance Sh given in Eq. (1.4). This is indeed the variance of the market
portfolio given in Eq. (1.14).
30
1.6. Appendix 3: An alternative derivation of the SML
c
by
A. Mele
1.6 Appendix 3: An alternative derivation of the SML

The vector of covariances of the m asset returns with the market portfolio are:

M
M
1
cov (
x, x
M ) = cov x
, x
=
=
(b 1m r) ,
w
w
r
(1A.11)
where we have used the expression for the market portfolio given in Eq. (1.15). Next, premultiply the
previous equation by
M
w
to obtain:
2
vM
=
1
1
M M
= M
(b 1m r) =
Sh,
w
w
w r
( r)2
(1A.12)
Sh
or vM = r
, which conrms Eq. (1.14).
Let us rewrite Eq. (1A.11) component by component. That is, for i = 1, , m,
2
1
vM
vM
(bi r) = (bi r) =
(bi r) ,
r
M r
Sh
r
where the last two equalities follow by Eq. (1A.12) and by the relation, Sh = M
vM . By rearranging
terms, we obtain Eq. (1.18).
iM cov (
xi , x
M ) =
31
1.7. Appendix 4: Broader denitions of risk - Rothschild and Stiglitz theory
c
by
A. Mele
1.7 Appendix 4: Broader denitions of risk - Rothschild and Stiglitz theory

The papers are Rothschild and Stiglitz (1970, 1971). Notation, any variable with a tilde is a random
variable. Let us consider the following denition of stochastic dominance:
Definition A.1 (Second-order stochastic dominance). x
2 dominates x
1 if, for each utility function
u satisfying u 0, we have also that E [u (
x2 )] E [u (
x1 )].
We have:
Theorem A.2. The following statements are equivalent:
a) x
2 dominates x
1 , or E [u (
x2 )] E [u (
x1 )];
b) random variable > 0 : x
2 = x
1 + ;
c) x > 0, F1 (x) F2 (x).
Proof. We provide the proof when the support is compact, say [a, b]. First, we show that b) c).
x1 t0 ) = Pr (
x2 t0 + ) Pr (
x2 t0 ) F2 (t0 ). Next, we show
We have: t0 [a, b], F1 (t0 ) Pr (
that c) a). By integrating by parts,
" b
" b
E [u (x)] =
u(x)dF (x) = u(b)
u (x)F (x)dx,
a
where we have used the fact that: F (a) = 0 and F (b) = 1. Therefore,
" b
E [u (
x2 )] E [u (
x1 )] =
u (x) [F1 (x) F2 (x)] dx.
a
Finally, it is easy to show that a) b).
Next, we turn to the denition of increasing risk:

Definition A.3. x
1 is more risky than x
2 if, for each function u satisfying u < 0, we have also
that E [u (
x1 )] E [u (
x2 )] for x
1 and x
2 having the same mean.
This denition of increasing risk does not rely on the sign of u . Furthermore, if var (
x1 ) >
var (
x2 ), x
1 is not necessarily more risky than x
2 , according to the previous denition. The standard
counterexample is the following one. Let x
2 = 1 w.p. 0.8, and 100 w.p. 0.2. Let x
1 = 10 w.p. 0.99, and
1090 w.p. 0.01. We have, E (
x1 ) = E (
x2 ) = 20.8, but var (
x1 ) = 11762.204 and var (
x2 ) = 1647.368.
However, consider u(x) = log x. Then, E (log (
x1 )) = 2.35 > E (log (
x2 )) = 0.92. It is easily seen that
in this particular example, the distribution function F1 of x
1 intersects F2 , which is in contradiction
with the following theorem.
Theorem A.4. The following statements are equivalent:
a) x
2 ;
b) x
1 has more weight in the tails than x
2 , i.e. t,
#t
[F1 (x) F2 (x)] dx
0;
c) x
1 is a mean preserving spread of x
2 , i.e. there exists a random variable : x
1 has the same
distribution as x
2 + , and E ( | x
2 = x2 ) = 0.
32
c
by
A. Mele
Proof. Let us begin with c) a). We have,

E [u (
x1 )] =
=
E [u (
x2 + )]
E [E ( u (
x2 + )| x
2 = x2 )]
E [u (E ( x
2 + | x
2 = x2 ))]
E [u (E ( x
2 | x
2 = x2 ))]
= E [u (
x2 )] .
As regards a) b), we have that:

E [u (
x1 )] E [u (
x2 )] =
"
"
u(x) [f1 (x) f2 (x)] dx

" b
b
= u(x) [F1 (x) F2 (x)]|a
u (x) [F1 (x) F2 (x)] dx
a
" b
u (x) [F1 (x) F2 (x)] dx
=
a

" b

b

= u (x) F1 (x) F2 (x) a
u (x) F1 (x) F2 (x) dx
a
#x

u (x) F1 (x) F2 (x) dx u (b) F1 (b) F2 (b) ,
where Fi (x) = a Fi (u)du. Now, x

2 means that E [u (
x1 )] < E [u (
x2 )] for u < 0.
By the previous relation, then, F1 (x) > F2 (x). Finally, see Rothschild and Stiglitz (1970) p. 238 for
the proof of b) c).
33
c
by
A. Mele
References
Connor, G. (1984): A Unied Beta Pricing Theory. Journal of Economic Theory 34, 13-31.
Huang, C-f. and R.H. Litzenberger (1988): Foundations for Financial Economics. New York:
North-Holland.
Huberman, G. (1983): A Simplied Approach to Arbitrage Pricing Theory. Journal of Economic Theory 28, 1983-1991.
Markovitz, H. (1952): Portfolio Selection. Journal of Finance 7, 77-91.
Ross, S. (1976): Arbitrage Theory of Capital Asset Pricing. Journal of Economic Theory
13, 341-360.
Rothschild, M. and J. Stiglitz (1970): Increasing Risk: I. A Denition. Journal of Economic
Theory 2, 225-243.
Rothschild, M. and J. Stiglitz (1971): Increasing Risk: II. Its Economic Consequences. Journal of Economic Theory 5, 66-84.
Sharpe, W. F. (1964): Capital Asset Prices: A Theory of Market Equilibrium under Conditions of Risk. Journal of Finance 19, 425-442.
34
2
The CAPM in general equilibrium
2.1 Introduction
This chapter develops the general equilibrium foundations to the CAPM, within a framework
that abstracts from the production sphere of the economy. For this reason, we usually refer
the resulting model to as the Consumption-CAPM. First, we review the static model of
general equilibrium, without uncertainty. Then, we illustrate the economic rationale behind the
existence of nancial assets in an uncertain world. Finally, we derive the Consumption-CAPM.
2.2 The static general equilibrium in a nutshell

We consider an economy with n agents and m commodities. Let wij denote the amount of the
i-th commodity the j-th agent is endowed with, and let wj = [w1j , , wmj
]. Let the price vector
be p = [p1 , , pm ], where pi is the price of the i-th commodity. Let wi = nj=1 wij be the total
endowment of the i-th commodity in the economy, and W = [w1 , , wm ] the corresponding
endowments bundle in the economy.
The j-th agent has utility function uj (c1j , , cmj ), where (cij )m
i=1 denotes his consumption
bundle. We assume the following standard conditions for the utility functions uj :
Assumption 2.1 (Preferences). The utility functions uj satisfy the following properties:
(i) Monotonicity; (ii) Continuity; and (iii) Quasi-concavity: uj (x) uj (y), and (0, 1),
2
uj
uj (x + (1 )y) > uj (y) or, c
(c1j , , cmj ) 0 and cu2 j (c1j , , cmj ) 0.
ij
ij
m
m
Let Bj (p1 , , pm ) = {(c1j , , cmj ) : i=1 pi cij i=1 pi wij Rj }, a bounded, closed and
convex set, hence a convex set. Each agent maximizes his utility function subject to the budget
constraint:
max uj (c1j , , cmj ) subject to (c1j , , cmj ) Bj (p1 , , pm ) .
{cij }
[P1]
c
by
A. Mele
2.2. The static general equilibrium in a nutshell
This problem has certainly a solution, for Bj is compact set and by Assumption 2.1, uj is
continuous, and a continuous function attains its maximum on a compact set. Moreover, the
Appendix shows that this maximum is unique.
The rst order conditions to [P1] are, for each agent j,
uj
c1j
uj
c2j
uj
cmj
=
= =
p1
p2
pm
m
m

pi cij =
pi wij
i=1
(2.1)
i=1
These conditions form a system of m equations with m unknowns. Let us denote the solution
to this system with [
c1j (p, wj ), , cmj (p, wj )]. The total demand for the i-th commodity is,
ci (p, w) =
n
cij (p, wj ),
j=1
i = 1, , m.
We emphasize the economy we consider in this chapter is one that completely abstracts from
production. Here, prices are the key determinants of how resources are allocated in the end. The
perspective is, of course, radically dierent from that taken by the Classical school (Ricardo,
Marx and Sraa), for which prices and resources allocation cannot be disentangled from the
production side of the economy. In the next chapter and more advanced parts of the lectures,
we consider the asset pricing implications of production, following the Neoclassical perspective.
2.2.1
Walras Law
Let us plug the demand functions of the j-th agent into the constraint of [P1], to obtain,
p,
0=
m

i=1

pi cij (p, w j ) wij .
(2.2)
Next, dene the total excess demand for the i-th commodity as ei (p, w) ci (p, w) wi . By
aggregating the budget constraint across all the agents,
p, 0 =
n
m

j=1 i=1
pi cij (p, w ) wij =
m
pi ei (p, w).
i=1
The previous equality is the celebrated Walras law.

Next, multiply p by R++ . Since the constraint to [P1] does not change, the excess demand
functions are the same, for each value of . In other words, the excess demand functions are
homogeneous of degree zero in the prices, or ei (p, w) = ei (p, w), i = 1, , m. This property of
the excess demand functions is also referred to as absence of monetary illusion.
2.2.2
Competitive equilibrium
A competitive equilibrium is a vector p in Rm

p, w) 0 for all i = 1, , m, with at
+ such that ei (
least one component of p being strictly positive. Furthermore, if there exists a j : ej (
p, w) < 0,
then pj = 0.
36
c
by
A. Mele
2.2.2.1 Back to Walras law
Walras law holds by the mere aggregation of the agents constraints. But the agents constraints
are accounting identities. In particular, Walras law holds for any price vector and, a fortiori,
it holds for the equilibrium price vector,
0=
m

i=1
pi ei (
p, w) =
m1
pi ei (
p, w) + pm em (
p, w).
(2.3)
i=1
Now suppose that the rst m1 markets are in equilibrium, or ei (

p, w) 0, for i = 1, , m1.
By the denition of an equilibrium, we have that sign (ei (
p, w)) pi = 0. Therefore, by Eq. (2.3),
we conclude that if m 1 markets are in equilibrium, then, the remaining market is also in
equilibrium.
2.2.2.2 The notion of numeraire
The excess demand functions are homogeneous of degree zero. Walras law implies that if m 1
markets are in equilibrium, then, the m-th remaining market is also in equilibrium. We wish
to link these two results. A rst remark is that by Walras law, the equations that dene a
competitive equilibrium are not independent. Once m 1 of these equations are satised, the
m-th remaining equation is also satised. In other words, there are m 1 independent relations
and m unknowns in the equations that dene a competitive equilibrium. So, there exists an
innity of solutions.
Suppose, then, that we choose the m-th price to be a sort of exogeneous datum. The result
is that we obtain a system of m 1 equations with m 1 unknowns. Provided it exists, such
pm ), i = 1, , m 1. Then, we may
a solution is a function f of the m-th price, pi = fi (
refer to the m-th commodity as the numeraire. In other words, general equilibrium can only
determine a structure of relative prices. The scale of these relative prices depends on the price
level of the numeraire. It is easily checked that if the functions fi are homogeneous of degree
one, multiplying pm by a strictly positive number does not change the relative price structure.
Indeed, by the equilibrium condition, for all i = 1, , m,
p1 , p2 , ,
pm , w) = ei (f1 (
pm ), f2 (
pm ), ,
pm , w)
0 ei (
= ei (
p1 ,
p2 ,
pm , w) = ei (
p1 , p2 , pm , w) ,
where the second equality is due to the homogeneity property of the functions fi , and the
last equality holds because the excess demand functions ei are homogeneous of degree zero. In
particular, by dening relative prices as pj = pj / pm , one has that pj = pj pm is a function
that is homogeneous of degree one. In other words, if p1
m , then,

p1
0 ei (
p1 , , pm , w) = ei (
p1 , ,
pm , w) ei
, , 1, w .
pm
2.2.3
Optimality
Let cj = (c1j , , cmj ) be the allocation to agent j, j = 1, , n. The following denition is the
well-known concept of a desiderable resource allocation in a society, according to Pareto.
37
c
by
A. Mele
2.2 (Pareto optimum). An allocation c = (

c1 , , cn ) is a Pareto optimum if
Definition
n
cj wj ) 0 and there is no c = (c1 , , cn ) such that uj (cj ) uj (
cj ), j = 1, , n, with
j=1 (
one strictly inequality for at least one agent.
We have the following fundamental result:
Theorem 2.3 (First welfare theorem). Every competitive equilibrium is a Pareto optimum.
Proof. Let us suppose on the contrary that c is an equilibrium but not a Pareto optimum.
Then, there exists a c : uj (cj ) > uj (

cj ), for some
j . Because
cj is optimal for agent j ,

cj
/ Bj (
p), or pcj > pwj and, by aggregating: p nj=1 cj > p nj=1 wj , which is unfeasible. It
follows that c can not be an equilibrium.
Next, we show that any Pareto optimal allocation can be decentralized. That is, corresponding to a given Pareto optimum c, there exist ways of redistributing endowments around,
and a price vector p : pc = pw, which is an equilibrium for the initial set of resources.
Theorem 2.4 (Second welfare theorem). Every Pareto optimum can be decentralized.
Proof. In the appendix.
The previous theorem can be interpreted as one that supports an equilibrium with transfer
payments. For any given Pareto optimum cj , a social planner can always give pw j to each
agent (with pcj = pwj , where wj is chosen by the planner), and agents choose cj . Figure 2.1
illustratres such a decentralization procedure within the Edgeworths box. Suppose that the
objective is to achieve c. Given an initial allocation w chosen by the planner, each agent is
given pw j . Under laissez faire, c will obtain. In other words, agents are given a constraint of
the form pcj = pw j . If wj and p are chosen so as to induce each agent to choose cj , then p is a
supporting equilibrium price. In this case, the marginal rates of substitutions are identical, as
established by the following celebrated result:
Theorem 2.5 (Characterization of Pareto optima: I). A feasible allocation c = (
c1 , , cn )
m1
R++
is a Pareto optimum if and only if there exists a
such that
$ uj
uj %
c
c
2j
j = ,
j = 1, , n, where u
j
u
, , mj .
(2.4)
uj
c1j
uj
c1j
Proof. A Pareto optimum satises:

1
c arg max
u
c
1
mn
cR+
unj (c ) uj , j = 2, , n

subject to
(cj wj ) 0
j=1
(j , j = 2, , n)
(i , i = 1, , m)
The Lagrangian function associated with this program is

1
L = u1 (c ) +
n

j=2
m
n

j
j uj (c ) uj
i
(cij wij ) ,
i=1
38
j=1
c
by
A. Mele
FIGURE 2.1. Decentralizing a Pareto optimum
and the rst order conditions are
u1
c11
and, for j = 2, , n,
u
1
= m
cm1
uj
c1j
uj
j
cmj
In both systems
of equations,
divide each equation by the the rst, obtaining exactly Eq. (2.4),

= 2 , , m . The converse is straight forward.

with
There is a simple and appealing interpretation of the Kuhn-Tucker multipliers on the

constraints of Theorem 2.5. Note that by Eq. (2.1), in the competitive equilibrium,

p2
pm
uj = p
, ,
.
p1
p1
But because a competitive equilibrium is also a Pareto optimum, then, by Theorem 2.5,

2
m
j =
u
, ,
.
1
1
represents the vector of relative, shadow prices arising within the centralized allocation
Hence,
process.
We provide a further characterization of Pareto optimal allocations.
39
c
by
A. Mele
Theorem 2.6 (Characterization of Pareto optima: II). A feasible allocation c = (

c1 , , cn )
is a Pareto optimum if and only if there exists > 0 such that c is solution to the following
program:
n
n

j
u (w, ) = 1maxn
j uj c
subject to
cj w ( j , j = 1, , m)
[P2]
c , ,c
j=1
j=1
Proof. The if part is simple and at the same time instructive. Let us solve the program in
[P2]. The Lagrangian is,
n
m
n

j
L=
j uj c
i
(cij wij ) ,
j=1
i=1
j=1
and the rst order conditions are, for j = 1, , n,
j uj = ( 1 , , m ) ,
uj
uj
uj
, ,
c1j
cmj
(2.5)
j equals the same vector of constants for all the agents, just as in Theorem 2.5. The
That is, u
converse to this theorem follows by an application of the usual separating theorem, as in Due
(2001, Chapter 1).
Note, if 1 = 1 and j = j for j = 2, , n, then, i = i (i = 1, , m) and so the rst
order conditions in Theorem 2.5 and 2.6 would lead to the same allocation. More generally, we
have:
Theorem 2.7 (Centralization of competitive equilbrium through Pareto weightings). The
outcome of any competitive equilibrium can be obtained, through a central planner who maximizes the program in [P2], with system of social weights equal to j = 1/j , where j is the
marginal utility of income for agent j.
So agents with high marginal utility of income for a given price vector, will receive little
social weight in the centralized planner allocation procedure. This result is particularly useful
when it comes to study nancial markets in economies with heterogeneous agents. Theorem 2.7
is also a point of reference, where to move from, when it comes to study asset prices in a world
of incomplete markets. Chapter 8 contains several examples of these applications.
Proof of Theorem 2.7. In the competitive equilibrium,
uj = j p,
p (p1 , , pm ) ,
(2.6)
where j are the Lagrange multipliers for the agents budget constraint, so that j is the agent
j marginal utility of income:
m
j =
uj (
c1j (p, w1j , , wmj ) , , cmj (p, w1j , , wmj )) , mj
pi wij .
mj
i=1
By comparing the competitive equilibrium solution in Eq. (2.6) with the Pareto optimality
property of the equilibrium in Eq. (2.5), we deduce that, a competitive equilibrium (
c, p) can
be implemented, by a social planner acting as in Theorem 2.6, when
j = 1/j , in which case
it also follows that, necessarily, = p, by the resources constraint, nj=1 cj w, that has to
hold both in the competitive economy and the centralized one.
40
2.3. Time and uncertainty
c
by
A. Mele
2.3 Time and uncertainty

A commodity is characterized by its physical properties, the date and the place at which
it will be available.
Gerard Debreu (1959, Chapter 2)

General equilibrium theory can be used to study a variety of elds, by making an appropriate
use of the previous denition - from the theory of international commerce to nance. To deal
with uncertainty, Debreu (1959, Chapter 7) extended the previous denition, by emphasizing
that a commodity should be described through a list of physical properties, with the structure
of dates and places replaced by some event structure. The following example illustrates the
dierence between two contracts underlying delivery of corn arising under conditions of certainty
(case A) and uncertainty (case B):
A The first agent will deliver 5000 tons of corn of a specified type to the second agent, who
will accept the delivery at date t and in place .
B The first agent will deliver 5000 tons of corn of a specified type to the second agent, who
will accept the delivery in place and in the event st at time t. If st does not occur at
time t, no delivery will take place.
In both cases, the contract is paid at the time it is actually agreed.
The model of the previous section can be used to deal with contracts containg statements such
as that in case B above. For example, consider a two-period economy. Suppose that in the second
period, sn mutually exhaustive and exclusive states of nature may occur. Then, we may recover
the model of the previous section, once we replace m (the number of commodities described
by physical properties, dates and places) with m , where m = sn m. With m replacing m,
the competitive equilibrium in this economy is dened as the competitive equilibrium in the
economy of the previous section.
The important assumption underlying the previous simplifying trick is that markets exists,
where commodities for all states of nature are traded. Such contingent markets are complete
in that a market is open for every commodity in all states of nature. Therefore, the agents may
implement any feasible action plan and, therefore, the resource allocation is Pareto-optimal.
The presumed existence of sn m contingent markets is, however, very strong. We now show
how the presence of nancial assets helps us mitigate this assumption.
2.4 Financial assets

What role might be played by nancial assets in an uncertainty world? Arrow (1953) developed
the following interpretation. Rather than signing commodity-based contracts that are contingent on the realization of events, the agents might wish to sign contracts generating payos
that are contingent on the realization of events. The payos delivered by the assets in the various states of the world could then be collected and used to satisfy the needs related to the
consumption plans.
The simplest nancial asset is the so-called Arrow-Debreu asset, i.e. an asset that payos
some amount of numeraire in the state of nature s if the state s will prevail in the future, and
41
c
by
A. Mele
2.5. Absence of arbitrage
nil otherwise. More generally, a nancial asset is a function x : S R, where S is the set
of all future events. Then, let m be the number of nancial assets. To link nancial assets to
commodities, we note that if the of nature s will occurs, then, any agent could use the payo
xi (s) promised by the i-th assets Ai to nance net transactions on the commodity markets, viz
m

p (s) e (s) =
i xi (s), s S,
(2.7)
i=1
where p(s) and e(s) denote some vectors of prices and excess demands related to the commodities, contingent on the realization of state s, and i is the number of assets i held by the agent.
In other words, the role of nancial assets, here, is to transfer value from a state of nature to
another to nance state-contingent consumption.
Unfortunately, Eq. (2.7) does not hold, in general. A condition is that the number of assets,
m, be suciently high to let each agent cope with the number of future events in S, sn . Market
completeness merely reduces to a size problem - the assets have to be suciently diverse to
span all possible events in the future. Indeed, we shall show that if there are not payos that
are perfectly correlated, then, markets are complete if and only if m = sn . Note, also, that
this reduces the dimension of our original problem, for we are then considering a competitive
equilibrium in sn + m markets, instead of a competitive equilibrium in sn m markets.
2.5 Absence of arbitrage

2.5.1
How to price a financial asset?
Consider an economy in which uncertainty is resolved through the realization of the event:
Tomorrow it will rain. A decision maker, an hypothetical Mr Law, must implement the
following contingent plan: if tomorrow will be sunny, he will need cs > 0 units of money, to
buy sun-glasses; if tomorrow it will rain, Mr Law will need cr > 0 units of money, to buy an
umbrella. Mr Law has access to a nancial market on which m assets are traded. He builds up
a portfolio aimed to reproduce the structure of payments that he will need tomorrow:

m
i Si (1 + xi (r)) = cr
i=1
(2.8)
m
i Si (1 + xi (s)) = cs
i=1
where Si is the price of the i-th asset, i is the number of assets to put in the portfolio, and
xi (r) and xi (s) are the net returns of asset i in the two states of nature, which of course are
known by Mr Law. For now, we do need to assume anything as regards the resources needed
to buy the assets, but we shall come back to this issue below (see Remark 2.6). Finally, and
remarkably, we are not making any assumption regarding Mr Laws preferences.
Eqs. (2.8) form a system of two equations with m unknowns (1 , m ). If m < 2, no perfect
hedging strategy is possible - that is, the system (2.8) can not be solved to obtain the desired
pair (ci )i=r,s . In this case, markets are incomplete. More generally, we may consider an economy
with sn states of nature, in which markets are complete if and only if Mr Law has access to sn
assets. More precisely, let us dene the following payo matrix, dened as
S1 (1 + x1 (s1 ))
Sm (1 + xm (s1 ))
...
X =
,
S1 (1 + x1 (sn ))
Sm (1 + xm (sn ))
42
c
by
A. Mele
where xi (sj ) is the payo promised by the i-th asset in the state sj . Then, to implement any
state contingent consumption plan c Rsn , Mr Law has to be able to solve the following system,
c = X ,
where and Rm , the portfolio. A unique solution to the previous system exists if rank(X) =
sn = m, and is given by = X 1 c. Consider, for example, the previous case, in which sn = 2.
Let us assume that m = 2, for any additional assets would be redundant here. Then, we have,
(1 + x2 (r))cs (1 + x2 (s))cr
1 =
S1 [(1 + x1 (s))(1 + x2 (r)) (1 + x1 (r))(1 + x2 (s))]
(1 + x1 (s))cr (1 + x1 (r))cs
2 =
S2 [(1 + x1 (s))(1 + x2 (r)) (1 + x1 (r))(1 + x2 (s))]
Finally, assume that the second asset is safe, or that it yields the same return in the two states
of nature: x2 (r) = x2 (s) r. Let xs = x1 (s) and xr = x1 (r). Then, the pair (1 , 2 ) can be
rewritten as,
1 = cs cr , 2 = (1 + xs ) cr (1 + xr ) cs .
S1 (xs xr )
S2 (1 + r) (xs xr )
As is clear, the issues we are dealing with relate to the replication of random variables. Here,
the random variable is a state contingent consumption plan (ci )i=r,s , where cr and cp are known,
which we want to replicate for hedging purposes. (Mr Law will need to buy either a pair of
sun-glasses or an umbrella, tomorrow.)
In the previous two-state example, two assets with independent payos are able to generate
any two-state variable. The next step, now, is to understand what happens when we assume
that there exists a third asset, A say, that delivers the same random variable (ci )i=r,s we can
obtain by using the previous pair (1 , 2 ).
We claim that if the current price of the third asset A is H, then, it must be that,
H = V 1 S1 + 2 S2 ,
(2.9)
for the nancial market to be free of arbitrage opportunities, to be dened informally below.
Indeed, if V < H, we can buy and sell at the same time the third asset A. The result is a sure
prot, or an arbitrage opportunity, equal to H V , for generates cr if tomorrow it will rain
and cr if tomorrow it will not rain. In both cases, the portfolio generates the payments that
are necessary to honour the contract committments related to the selling of A. By a symmetric
argument, the inequality V > H would also generate an arbitrage opportunity. Hence, Eq. (2.9)
must hold true.
It remains to compute the right hand side of Eq. (2.9), which in turn leads to an evaluation
formula for the asset A. We have:
H=
1
[P cs + (1 P )cr ] ,
1+r
P =
xr r
.
xs xr
(2.10)
Importantly, then, H can be understood as the discounted (by 1 + r) expectation of payos

promised by A, taken under some articial probability P .
Remark 2.8. In this introductory example, the asset A can be priced without making
reference to any agents preferences. The key observation to obtain this result is that the
43
c
by
A. Mele
payos promised by A can be obtained through the portfolio . This fact does not obviously
mean that any agent should use this portfolio. For example, it may be the case that Mr Law
is so poor that his budget constraint would not even allow him to implement the portfolio .
The point underlying the previous example is that the portfolio could be used to construct an
arbitrage opportunity, arising when Eq. (2.9) does not hold. In this case, any penniless agent
could implement the arbitrage described above.
The next step is to extend the results in Eq. (2.10) to a dynamic setting. Suppose that
an additional day is available for trading, with the same uncertainty structure: the day after
tomorrow, the asset A will pay o css if it will be sunny (provided the previous day was sunny),
and crs if it will be sunny (provided the previous day was raining). By using the same arguments
leading to Eq. (2.10), we obtain that:
H=
2

1
2
P
c
+
P
(1
P
)c
+
(1
P
)P
c
+
P
c
.
ss
sr
rs
rr
(1 + r)2
Finally, by extending the same reasoning to T trading days,

H=
1
E (cT ) ,
(1 + r)T
(2.11)
where E denotes the expectation taken under the probability P .

The key assumption we used to derive Eq. (2.11) is that markets are complete at each trading
day. True, at the beginning of the trading period Mr Law faced 2T mutually exclusive possible
states of nature that would occur at the T -th date, which would seem to imply that we would
need 2T assets to replicate the asset A. However, we have just seen that to price A, we only
need 2 assets and T trading days. To emphasize this fact, we say that the structure of assets
and transaction dates makes the markets dynamically complete in the previous example. The
presence of dynamically complete markets allows one to implement dynamic trading strategies
aimed at replicating the value of the asset A, period by period. Naturally, the asset A could
be priced without any assumption about the preferences of any agent, due to the assumption
of dynamically complete markets.
2.5.2
The Land of Cockaigne
We provide a precise denition of the notion of absence of arbitrage opportunities, as well

as a connection between this notion and the notion and properties of the competitive equilibrium described in Section 2.2. For simplicity, we consider a multistate economy with only
one commodity. The extension to the multicommodities case is dealt with very briey in the
appendix.
Let vi ( s ) be the payo of asset i in the state s , i = 1, , m and s = 1, , d. Consider the
payo matrix:
v1 ( 1 )
vm ( 1 )
...
V
.
v1 ( d )
vm ( d )
Let vsi vi ( s ), vs, [vs1 , , vsm ], v,i [v1i , , , vdi ] . We assume that rank(V ) = m d.
44
c
by
A. Mele

The budget constraint of each agent has the form:
m
Si i
c0 w0 = S =
m
i=1
w
=
v
=
vsi i , s = 1, , d
s
s
s
i=1
Let x1 = [x1 , , xd ] . The second constraint can be written as:

c1 w 1 = V .
We dene an arbitrage opportunity as a portfolio that has a negative value at the rst period,
and a positive value in at least one state of world in the second period, or a positive value in
all states of the world in the second period and a nonpositive value in the rst period.
Notation: x Rm , x > 0 means that at least one component of x is strictly positive while
the other components of x are nonnegative. x 0 means that all components of x are strictly
positive. [Insert here further notes]
Definition 2.9. An arbitrage opportunity is a strategy that yields1 either V 0 with
an initial investment S < 0, or a strategy that produces2 V > 0 with an initial investment
S 0.
As we shall show below (Theorem 2.11), an arbitrage opportunity can not exist in a competitive equilibrium, for the agents program would not be well dened in this case. Introduce,
then, the (d + 1) m matrix,

S
W =
,
V
the vector subspace of Rd+1 ,
'
&
W = z Rd+1 : z = W , Rm ,
and, nally, the null space of W ,
&
'
W = x Rd+1 : xW = 0m .
The economic interpretation of the vector subspace W is that of the excess demand space for
all the states of nature, generated by the wealth transfers generated
by the investments in the
&
'
d+1
assets. Naturally, W and W are orthogonal, as W = x R
: xz = 0m , z W .
Mathematically, the assumption that there are no arbitrage opportunities is equivalent to the
following condition,
(
W Rd+1
= {0} .
(2.12)
+
The interpertation of (2.12) is in fact very simple. In the absence of arbitrage opportunities,
there should be no portfolios generating wealth transfers that are nonnegative and strictly
positive in at least one state, i.e. : W > 0. Hence, W and the positive orthant Rd+1
can
+
not intersect.
1V
2V
0 means that [V ]j 0, j = 1, , d, i.e. it allows for [V ]j = 0, j = 1, , d.

> 0 means [V ]j 0, j = 1, , d, with at least one j for which [V ]j > 0.
45
c
by
A. Mele
The following result provides a general characterization of how the no-arbitrage condition in
(2.12) restricts the price of all the assets in the economy.
Theorem 2.10. There are no arbitrage opportunities
if and only if there
exists a Rd++ :

S = V . If m = d, is unique, and if m < d, dim Rd++ : S = V = d m.
Proof. In the appendix.
The previous theorem provides the foundations for many developments in nancial economics.
To provide its intuition, let us pre-multiply the second constraint by , obtaining,
(c1 w 1 ) = V = S = (c0 w0 ) ,
where the second equality follows by Theorem 2.10, and the third equality is due to the rst
period budget constraint. Critically, then, Theorem 2.10 shows that in the absence of arbitrage
opportunities, each agent has access to the following budget constraint,
d

1

1
0 = c0 w0 + c w = c0 w0 +
s (cs ws ) , with c1 w1 V .
(2.13)
s=1
The budget constraints in (2.13) reveal that can be interpreted as the vector of prices to
the commodity in the future d states of nature, and that the numeraire in this economy is
the rst-period consumption. We usually refer to as the state price vector, or Arrow-Debreu
state price vector. However, it would be misleading to say that the budget constraint in (2.13)
is that we are used to see in the static Arrow-Debreu type model of Section 2.2. In fact, the
Arrow-Debreu economy of Section 2.2 obtains when m = d, in which case V = Rd in (2.13).
This case, which according to Theorem 2.10 arises when markets are complete, also implies the
remarkable property that there exists a unique that is compatible with the asset prices we
observe.
The situation is radically dierent if m < d. In other terms, V is the subspace of excess
demands agents have access to in the second period and can be smaller than Rd if markets
are incomplete. Indeed, V is the subspace generated by the payos obtained by the portfolio
choices made in the rst period,
&
'
V = e Rd : e = V , Rm .
2
Consider, for example, the
v1 case d = 2 and m = 1. In this case, V = {e R : e = V , R},
with V = V1 , where V1 = v2 say, and dim V = 1, as illustrated by Figure 2.2.
Next, suppose we open a new market for)a second nancial asset with payos
given by: V2 =
*
1 v1 +2 v3
v3
v1 v3
2
2
. Then, m = 2, V = ( v2 v4 ), and V = e R : e = 1 v2 +2 v4 , R , i.e. V = R2 . As
v4
a result, we can now generate any excess demand in R2 , just as in the Arrow-Debreu economy
of Section 2.2. To generate any excess demand, we multiply the payo vector V1 by 1 and the
payo vector V2 by 2 . For example, suppose we wish to generate the payo the payo vector
V4 in Figure 2.3. Then, we choose some 1 > 1 and 2 < 1. (The exact values of 1 and 2
are obtained by solving a linear system.) In Figure 2.3, the payo vector V3 is obtained with
1 = 2 = 1.
To summarize, if markets are complete, then, V = Rd . If markets are incomplete, V
is only a subspace of Rd , which makes the agents choice space smaller than in the complete
markets case.
46
c
by
A. Mele
v2
v1
<V>
FIGURE 2.2. Incomplete markets, d = 2, m = 1.
V3
v4
v2
V2
V4
V1
v3
v1
FIGURE 2.3. Complete markets, V = R2 .
47
c
by
A. Mele
2.6. Equivalent martingales and equilibrium
We now present a fundamental result, about the viability of the model. Dene the second
period consumption c1j [c1j , , cdj ] , where csj is the second-period consumption in state s,
and let,
+

c0j w0j = Sj
1
1
c0j , cj arg max1 uj (c0j ) + j E( j (cj )) , subject to
[P3]
1
1
c
= V j
c0j ,cj
j wj
where uj and j are utility functions, both satisfying Assumption 2.1. Naturally, we could use
more general formulations of utilities than that in [P3], and in fact we shall in more advanced
parts of this book. For the sake of this introductory chapter, we only consider additive utility.
We have:
Theorem 2.11. The program [P3] has a solution if and only if there are no arbitrage opportunities.
Proof. Let us suppose on the contrary that the program [P3] has a solution c0j , c1j , j , but
that there exists a : W > 0. The program constraint is, with straight forward notation,
cj = wj + W j . Then, we may dene a portfolio j = j + , such that cj = wj + W (j + ) =
cj + W > cj , which contradicts the optimality of cj . For the converse, note that the absence of
arbitrage opportunities implies that Rd++ : S = V , which leads to the budget constraint
in (2.13), for a given . This budget constraint is clearly a closed subset of the compact budget
constraint Bj in [P1] (in fact, it is Bj restricted to V ). Therefore, it is a compact set and,
hence, the program [P3] has a solution, as a continuous function attains its maximum on a
compact set.
2.6 Equivalent martingales and equilibrium

We provide the denition of an equilibrium with nancial markets, when the nancial assets
are in zero net supply.
Definition 2.12. An equilibrium is given by allocations and prices {(
c0j )nj=1 , ((
csj )nj=1 )ds=1 ,
n
nd
d
(Si )m
i=1 R+ R+ R+ }, where the allocations are solutions of the program [P3] and satisfy:
0=
n

j=1
(
c0j w0j ) ,
0=
n

j=1
(
csj wsj ) (s = 1, , d) ,
0=
n

j=1
ij
(i = 1, , d) .
We now express demand functions in terms of the stochastic discount factor, and then look
for an equilibrium by looking for the stochastic discount factor that clears the commodity
markets. By Walras law, this also implies the equilibrium on the nancial market. Indeed, by
aggregating the agents constraints in the second period,
n

j=1
n

c1j wj1 = V
ij (m).
j=1
For simplicity, we also assume that uj (x) > 0, uj (x) < 0 x > 0 and limx0 uj (x) = ,
limx uj (x) = 0 and that j satises the same properties.
48
c
by
A. Mele

2.6.1
The rational expectations assumption
Lucas, Radner, Green. Every agent correctly anticipates the equilibrium price in each state of
nature.
[Consider for example the models with asymmetric information that we will see later in these
lectures. At some point we will have to compute, E ( v| p (
y ) = p). That is, the equilibrium is a
pricing function which takes some values p (
y ) depending on the state of nature. In this kind of
models, I (p (
y ) , y) + (1 ) U (p (
y ) , y) + y = 0, and we look for a solution p (
y ) satisfying
this equation.]
2.6.2
Stochastic discount factors
Theorem 2.10 states that in the absence of arbitrage opportunities,

Si = v,i =
d
i = 1, , m.
s vs,i ,
s=1
(2.14)
Let us assume that the rst asset is a safe asset, i.e. vs,1 = 1 s. Then, we have
d

1
S1
=
s .
1+r
s=1
(2.15)
Eq. (2.15) conrms the economic interpretation of the state prices in (2.13). Recall, the states
of nature are exhaustive and mutually exclusive. Therefore, s can be interpreted as the price
to be paid today for obtaining, for sure, one unit of numeraire, tomorrow, in state s. This is
indeed the economic interpretation of the budget constaint in (2.13). Eq. (2.15) conrms this
as it says that the prices of all these rights sum up to the price of a pure discount bond, i.e. an
asset that yields one unit of numeraire, tomorrow, for sure.
Eq. (2.15) can be elaborated to provide us with a second interpretation of the state prices in
Theorem 2.10. Dene,
Ps (1 + r)s ,
which satises, by construction,
d
Ps = 1.
s=1
Therefore, we can interpret P (Ps )ds=1 as a probability distribution. Moreover, by replacing

P in Eq. (13.9) leaves,
d
1
1
Si =
Ps vs,i =
E P (v,i ) ,
1 + r s=1
1+r
i = 1, , m.
(2.16)
Eq. (2.16) conrms Eq. (2.10), obtained in the introductory example of Section 2.5. It says
that the price of any asset is the expectation of its future payos, taken under the probability P , discounted at the risk-free interest rate r. For this reason, we usually refer to the
probability P as the risk-neutral probability. Eq. (2.16) can be extended to a dynamic context, as we shall see in later chapters. Intuitively, consider an asset that distributes dividends
in every period, let S (t) be its price at time t, and D (t) the dividend paid o at time t.
49
c
by
A. Mele
Then, the payo it promises for the next period is S (t + 1) + D (t + 1). By Eq. (2.16),
S (t) = (1 + r)1 E P (S (t + 1) + D (t + 1)) or, by rearranging terms,

S (t + 1) + D (t + 1) S (t)
P
E
= r.
(2.17)
S (t)
That is, the expected return on the asset under P equals the safe interest rate, r. In a dynamic
context, the risk-neutral probability P is also referred to as the risk-neutral martingale measure,
or equivalent martingale measure, for the following reason. Dene a money market account as
an asset with value evolving over time as M (t) (1 + r)t . Then, Eq. (2.17) can be rewritten
as S (t) /M (t) = E P [(S (t + 1) + D (t + 1)) /M (t + 1)]. This shows that if D (t + 1) = 0 for

some t, then, the discounted process S (t) /M (t) is a martingale under P .
Next, let us replace P into the budget constraint in (2.13), to obtain, for (c1 w 1 ) V ,
d

1
1

0 = c0 w0 +
s (cs ws ) = c0 w0 +
Ps (cs ws ) = c0 w0 +
E P c1 w1 .
1 + r s=1
1+r
s=1
(2.18)
For reasons developed below, it is also useful to derive an alternative representation of the
budget constraint, in terms of the objective probability P (say). Let us introduce, rst, the
ratio , dened as,
Ps = s Ps , s = 1, , d.
The ratio s indicates how far P and P are. We assume s is strictly positive, which means
that P and P are equivalent measures, i.e. they assign the same weight to the null sets. Finally,
let us introduce the stochastic discount factor, m = (ms )ds=1 , dened as,
ms (1 + r)1 s .
We have,
d
1
1

1

E P c1 w1 =
Ps (cs ws ) =
s (cs ws ) Ps = E m c1 w1 .
1+r
1+r
1+r
s=1
s=1
=ms
Hence, we can rewrite Eq. (2.18) as,

0 = c0 w0 + E m c1 w1 ,

c1 w1 V .
Similarly, by replacing the stochastic discount factor m into Eq. (2.16) we obtain,
Si =
E P (v,i ) = E (m v,i ) ,
1+r
i = 1, , m.
(2.19)
Naturally, despite all such dierent ways to express budget constraints and asset prices, the
key of the model is still ,
ms = (1 + r)1 s = (1 + r)1
Ps
= s,
Ps
Ps
which can be recovered, once we solve for the equilibrium stochastic discount factor m, as we
shall illustrate in the next section.
50
c
by
A. Mele

2.6.3
Optimality and equilibrium
We have argued that in the absence of arbitrage opportunities, the program of any agent j is

1

1
1
1
1
max
subject
to
0
=
c
c
w
u
V .
j (c0j ) + j E( j (cj ))
0j w0j +E m (cj wj ) ,
1
(c0 ,c )
[P4]
2.6.3.1 Complete markets and risk sharing
In the complete markets case, V = Rd , so that the rst order conditions to the program [P4]
are,
uj (
c0j ) = j , j j (
csj ) = j ms , s = 1, , d,
where j is a Lagrange multiplier. So, really, the properties of this model are the same as those
of the static model in Section 2.2. Formally, the complete markets economy in this section is the
same as the static economy in Section 2.2, once we set m = d, where m is the dimension of the
commodity space, in Section 2.2, and ps = s , where ps is the price of the s-th commodity in
Section 2.2, with p1 = 1 (the numeraire), and s is the Arrow-Debreu state price in the unied
budget constraint of Eq. (2.18).
These simple observations have profound implications: an economy subject to uncertainty can
be understood through a static model, in the presence of complete markets! Under the conditions
stated in Section 2.2, even complicated models with heterogeneous agents, with potentially
interesting asset pricing implications, and still, apparently, so hopelessly dicult to analyze,
can actually be centralized, through a dedicated design of Paretos weights, as formalized
in Theorem 2.7. We can actually do much more. First, this centralization property is easily
extended to a dynamic context, as we shall see in more advanced parts of these lectures (see
Chapter 8), provided markets satisfy the property of being dynamically complete, a property
explained in the next two chapters. Second, the assumption agents can exchange Arrow-Debreu
securities for all future states of the world, is clearly unrealistic: markets are pretty likely to
be incomplete, one possible reason why nancial innovation is so pervasive, in practice. Yet
the theory about centralization can be extended to an incomplete markets setting, through a
system of stochastic Pareto weights, as we discuss in detail in Chapter 8. For now, let us
proceed with the next simple and fundamental steps.
To illustrate the equilibrium implications of the rst order conditions in a simple case, consider
an economy with a single agent. In this economy, the rst order conditions immediately lead to
the following stochastic discount factor,
ms =
(ws )
.
u (w0 )
The economic interpretation of this stochastic discount factor is the following. In the autarchic
state,

dc0
(ws )
Ps = ms Ps = s
dcs c0 =w0 ,cs =ws
u (w0 )
is the present consumption the agent is willing to give up to at t = 0, in order to obtain
additional consumption at time t = 1, in state s. In other words, s is the price, in terms of the
present consumption numeraire, of one additional unit of consumption at time t = 1 and state
s. So it is a state price, such that, the agent is happy to consume his own endowment, without
51
c
by
A. Mele

any incentives to trade in the nancial markets. The risk-neutral probability is,
Ps
(ws )
= s Ps = (1 + r) ms Ps = (1 + r)
Ps .
u (w0 )
By the rst
dorderconditions, and the pure discount bond evaluation formula, it is easily checked
that 1 = s=1 Ps . Moreover,
1

Ps
(ws )
(ws )
1 u (w0 )
= ms (1 + r) = ms E
=
m
=
,
s
Ps
u (w0 )
E [ (ws )]
E [ (ws )]
1
where the second equality follows by the pure discount bond evaluation formula: 1+r
= E(m).
In the multi-agent case, the situation is similar as soon as markets are complete. Indeed,
consider the rst order conditions of each agent,
j (
csj )
= ms ,
uj (
c0j )
s = 1, , d,
j = 1, , n.
The previous relation reveals that as soon as markets are complete, agents must have the same
marginal rate of substitution, in equilibrium. This is because by Theorem 2.10, the state price
vector is unique if and only if markets are complete, which then implies uniqueness of ms = Pss
(
csj )
and, hence, the fact that each marginal rate of substitution j uj (c0j ) is independent of j. In this
j
case, the equilibrium allocation is clearly a Pareto optimum, by the discussion at the beginning
of this section, and Theorem 2.5.
The result that agents have the same marginal rate of substitution for each state of the world
is known as risk sharing. It means that, given an initial endowment distribution among the
agents, the market mechanism, through to a system of complete securities markets, is such that
consumption risk is shifted around the economy, so that it is borne by the agents most willing to
take it. For example, suppose that two agents 1 and 2 have the same discount rate, and utility
functions uj = j , with CRRA given by 1 and 2 , where 1 < 2 . Then, Grs1 = (Grs2 )2 /1 ,
where Grsi is consumption growth for the i-th agent in state s. In good times, when Grs2 > 1,
the more risk-averse agent experiences, ex-post, a lower consumption growth rate, Grs2 < Grs1 .
In bad times, however, when Grs2 < 1, the more risk-averse agent experiences, ex-post, a higher
consumption growth rate, Grs2 > Grs1 . In other words, capital markets, when complete, operate
in such a way to have the more risk-averse agent face a less volatile consumption growth.
2.6.3.2 Incomplete markets
If markets are incomplete, marginal rates of substitution cannot be equal, among agents, except
perhaps on a set of endowments distribution with measure zero. The best outcome in this case,
is a set of equilibria called constrained Pareto optima, i.e. constrained by ... the states of nature.
As it turns out, there might not even exist constrained Pareto optima in multiperiod economies
with incomplete marketsexcept perhaps those arising on a set of endowments distributions
with zero measure.
When market are incomplete, the state price vector is not unique. That is, suppose that
is an equilibrium state price. Then, all the elements of

= { Rd++ : ( ) V = 0}
52
(2.20)
c
by
A. Mele
are also equilibrium state prices - there exists an innity of equilibrium state prices that are
consistent with absence of arbitrage opportunities. In other words, there exists an innity of
equilibrium state prices guaranteeing the same observable assets price vector S, for V =
V = S.
How do we proceed in this case? Introduce the following budget constraint:
&
'

C = c Rd++ : 0 = c0 w0 + c1 w 1 , c1 w1 V , Rd++ : S = V .
(2.21)
This budget constraint, and the previous reasoning about the set in (2.20) shows that in the
context of incomplete markets, there exists many constraints to take care of, and the previous
martingale methods, do not apply.
Yet let Val (PI ) be the value of the following program in the incomplete markets at hand:

max uj (c0j ) + j E( j (c1j )) .
[PI ]
cC
Consider, next, the following constraint:

+
,
c Rd++ : 0 = c0 w0 + (c1 w1 ) , (c1 w 1 ) Rd ,
C =
,
for some given Rd++ : S = V
and let Val (P ) be the value of the program in some abstract complete markets case:

max uj (c0j ) + j E( j (c1j )) .
cC
[P ]
Clearly, we have, Val (PI ) Val (P ) for all , for the constraint in the incomplete markets
case, C, is more stringent than that in any complete market setting, C : the solution to the
program in the incomplete markets case [PI ], must satisfy the budget constraints in C, formed
using all of the possible Arrow-Debreu state prices (including the Arrow-Debreu state price
given in C ), as the constraint of Eq. (2.21) shows. Moreover, (c1 w1 ) V . These remarks
suggest to dene the following min-max Arrow-Debreu state price:
= arg min Val (P ) .
The natural question is to know whether

Val (PI ) = Val (P ) .
(2.22)
This is indeed the case, given some regularity conditions. For the characterization of , suppose
: Val (PI ) = Val(P ). Then,
= . Indeed, suppose the contrary, i.e. there exists
there exists
: Val(P ) < Val(P ). Then, we would have,

Val (PI ) Val(P ) < Val(P ) = Val (PI ) ,
a contradiction. Note, again, this is a characterization result about , not an existence proof.
But as mentioned earlier, Eq. (2.22) holds true, as shown in a dynamic setting by He and
Pearson (1991). Chapter 4 provides general guidance about an even more general approach to
solving problems of this kind, arising in a broader context of market imperfections, including
incomplete markets as a special case.
53
c
by
A. Mele
2.7. Consumption-CAPM
2.6.3.3 Computation of the equilibrium
The rst order conditions satised by any agents program are:

c0j = Ij (j ) , csj = Hj 1
j j ms ,
(2.23)
where Ij and Hj denote the inverse functions of uj and j . By the assumptions we made on uj
and j , Ij and Hj inherit the same properties of uj and j . By replacing these functions into
the constraint,
d

0 = c0j w0j + E m (
c1j wj1 ) = Ij (j ) w0j +
Ps ms Hj ( 1
.
j j ms ) wsj
s=1
Dene the function,

1
zj (j ) Ij (j ) + E mHj ( 1
j j m) = w0j + E m wj .
We see that limx0 z(x) = , limx z(x) = 0 and z (x) < 0. Therefore, there exists a unique
solution for j :

j j w0j + E(m wj1 ) ,
where () denotes the inverse function of z. By replacing back into Eqs. (2.23), we obtain:

1
c0j = Ij j w0j + E(m wj1 ) , csj = Hj 1
.
j ms j w0j + E(m wj )
It remains to compute the general equilibrium. The kernel m must be determined. This means
that we have d unknowns (ms , s = 1, , d). We have d + 1 equilibrium conditions (holding in
the d + 1 markets). By Walras law, only d of these are independent. Consider the equilibrium
conditions in the d markets at the second period:
n
n

1

1
Hj j ms j w0j + E(m wj ) =
wsj ws , s = 1, , d.
gs ms ; (ms )s =s
j=1
j=1
These conditions determine the kernel (ms )ds=1 which leads to compute prices and equilibrium
allocations. Finally, once the optimal cs are computed, for s = 0, 1, , d, the portoio
generated them can be inferred through = V 1 (
c1 w1 ).
2.7 Consumption-CAPM

Consider the pricing equation (2.19). It states that for every asset with gross return R
1
S payo,
1 = E(m R),
(2.24)
where m is some pricing kernel.
In the previous section, we learnt that in a complete markets economy, equilibrium leads to
the following identication of the pricing kernel,
ms =
(ws )
.
u (w0 )
For a riskless asset, 1 = E(m R). By combining this equality with Eq. (2.24), leaves E[m
R)] = 0. By rearranging terms,
(R
= R cov( (w ), R) .
E(R)
(2.25)
E [ (w + )]
54
c
by
A. Mele
2.7. Consumption-CAPM
2.7.1
The risk premium
Eq. (2.25) can be rewritten as,

R=
E(R)
cov(m, R)
= R cov(m, R).
E(m)
(2.26)
The risk-premium to invest in the asset is high for securities which pay high returns when
consumption is high (i.e. when we dont need high returns) and low returns when consumption
is low (i.e. when we need high returns).
All in all, if the price p = E (m payo) = E (m) E (payo)+cov (m, payo) = R1 E (payo)
- Premium, where Premium = cov (m, payo), a discounting eect.
2.7.2
The beta relation

m such that
Suppose there is a R
m = 1 (ws ) ,
R
all s.
In this case,
=R+
E(R)
m , R)
cov(R
E [ (w + )]
m) = R +
and E(R
m)
var(R
.
E [ (w+ )]
These relations can be combined to yield,

R = [E(R
m ) R],
E(R)
m , R)
cov(R
.
m)
var(R
2.7.3 CCAPM & CAPM

p be the portfolio return which is the most highly correlated with the pricing kernel m.
Let R
We have,
p ) R = R cov(m, R
p ).
E(R
(2.27)
Using Eqs. (2.26) and (2.27),
and by rearranging terms,
E(R)
cov(m, R)
=
,
p) R
p)
E(R
cov(m, R
R =
E(R)
R,m
p ) R]
[E(R
R p ,m
[CCAPM].
p is perfectly correlated with m, i.e. if there exists : R

p = m, then
If R
R,m
=
and then
p , R)
cov(R
p)
var(R
and R p ,m =
R = p [E(R
p ) R]
E(R)
R,R
[CAPM].
This is not the only way the CAPM obtains. As we shall explain in Chapter 6, the CAPM also
obtains through the so-called maximum correlation portfolio, which is the portfolio that is
the most highly correlated with the pricing kernel m.
55
c
by
A. Mele
2.8. Innite horizon
2.8 Innite horizon

We consider d states of the nature and m = d Arrow securities. We write a unied budget
constraint, as in the valuation equilibria approach of Debreu (1954).
We have,
m
p (c w ) = S (0) (0) =
S (0) (0)
0
or,
i=1
p1s (c1s ws1 ) = (0)

s , s = 1, , d
p0 (c0 w0 ) +
m
(0)
Si
i=1

p1i c1i wi1 = 0.
The previous relation holds in a two-period economy. In a multiperiod economy, in the second
period (as in the following periods) agents save indenitively for the future. In the appendix,
we show that,
-
.

0=E
m0,t pt ct wt ,
(2.28)
t=0
where m0,t are the state prices. From the perspective of time 0, at time t there exist dt states
of nature and, thus, dt possible prices.
2.9 Further topics on incomplete markets

2.9.1
Nominal assets and real indeterminacy of the equilibrium

Rm(d+1)
The equilibrium is a set of prices (
p, S)
Ra++ such that:
++
0=
n
e0j (
p, S),
0=
j=1
n
e1j (
p, S),
0=
j=1
n
j (
p, S),
j=1
where the previous functions are the results of optimal plans of the agents. This system has
m (d + 1) + a equations and m (d + 1) + a unknowns, where a d. Let us aggregate the
constraints of the agents,
p0
n

j=1
e0j = S
n
j ,
j=1
p1
n
e1j = B
j=1
n
j .
j=1
Suppose the nancial markets clearing condition is satised, i.e.
n
j=1
j = 0. Then,
n
m

() ()
0
=
p
e
p
e
=
p0 e0
0
0j
0
0
j=1
=1
m

n
m

()
()
()
()
e1j p1 e1 =
p1 ( 1 )e1 ( 1 ), ,
p1 ( d )e1 ( d )
0d = p1
j=1
=1
=1
Therefore, there is one redundant equation for each state of nature, or d + 1 redundant
equations, in total. As a result, the equilibrium has less independent equations (m (d + 1) 1)
than unknowns (m(d+1)+d), i.e., an indeterminacy degree equal to d+1. This result does not
56
c
by
A. Mele
2.9. Further topics on incomplete markets
rely on whether markets are complete or not. In a sense, it is even not an indeterminacy result
when markets are complete, as we may always assume agents would organize the exchanges
at the beginning. In this case, onle the suitably normalized Arrow-Debreu state prices would
matter for agents.
The previous indeterminacy can be reduced to d1, as we may use two additional homogeneity relations. To pin down these relations, let us consider the budget constaint of each agent
j,
p0 e0j = S j , p1 e1j = Bj .
The rst-period constraint is still the same if we multiply the spot price vector p0 and the
is an equinancial price vector S by a positive constant, (say). In other words, if (
p0 , p1 , S)
is also an equilibrium, which delivers a rst homogeneity relation.
librium, then, (
p0 , p1 , S)
To derive the second homogeneity relation, we multiply the spot prices of the second period by
a positive constant, and increase at the same time the rst period agents purchasing power,
by dividing each asset price by the same constant, as follows:
S
p0 e0j = j ,
p1 e1j = Bj .
is an equilibrium, then, p0 ,
p1 , S is also an equilibrium.
Therefore, if (
p0 , p1 , S)
2.9.2
Nonneutrality of money
The previous indeterminacy arises because nancial contracts are nominal, i.e. the asset payos
are expressed in terms of some unite de compte that, among other things, we did not make
precise. Such an indeterminacy vanishes if we were to consider real contracts, i.e. contracts
with payos expressed in terms of the goods. To show this, note that in the presence of real
contracts, the agents constraints are
+
p0 e0j = Sj
p1 ( s )e1j ( s ) = p1 ( s )As j , s = 1, , d
where As = [A1s , , Aas ] is the m a matrix of the real payos. The previous constraint
now reveals how to recover d + 1 homogeneity relations. For each strictly positive vector
= [0 , 1 , d ], we have that if [
p0 , S, p1 ( 1 ), , p1 ( s ), , p1 ( d )] is an equilibrium, then,
[0 p0 , 0 S, p1 ( 1 ), , p1 ( s ), , p1 ( d )]
is
also
an
equilibrium,
and
so
is
[
p0 , S, p1 ( 1 ), , s p1 ( s ), , p1 ( d )], for s , s = 1, , d.
As is clear, the distinction between nominal and real assets has a precise meaning, when
one considers a multi-commodity economy. Even in this case, however, such a distinctions is
not very interesting without a suitable introduction of a unite de compte. These considerations
led Magill and Quinzii (1992) to solve the indetermincay while still remaining in a framework
with nominal assets. They simply propose to introduce money as a mean of exchange. The
indeterminacy can then be resolved by xing the prices via the d + 1 equations dening the
money market equilibrium in all states of nature:
Ms = ps
n

j=1
wsj , s = 0, 1, , d.
Magill and Quinzii showed that the monetary policy (Ms )ds=0 is generically nonneutral.
57
c
by
A. Mele
2.10. Appendix 1
2.10 Appendix 1
In this appendix we prove that the program [P1] has a unique maximum. Indeed, suppose on the
contrary that we have two maxima:

c = (
c1j , , cmj ) and c = c1j , , cmj .

These two maxima would satisfy uj (
c) = uj (
c),
with m
ij = m
ij = Rj . To check that this
i=1 pi c
i=1 pi c
m
claim is correct, suppose on the contrary that i=1 pi cij < Rj . Then, the consumption bundle,

c = c1j + , , cmj , > 0,
would be preferred to c, by Assumption 2.1, and, at the same time, it would hold that, for suciently
small ,
m
m

pi cij = p1 +
pi cij < Rj .
m
i=1
i=1
[Indeed, we have, A i=1 pi cij . A < Rj > 0 : A + p1 < Rj . E.g., p1 = Rj A , > 0. The
condition is then: > 0 : Rj A > .] Hence, c would be a solution to [P1], thereby contradicting
the optimality of c. Therefore, the existence of two optima would imply a full use of resources. Next,
c + (1 )c, (0, 1). By Assumption 2.1,
consider a point y lying between c and c, viz y =

c + (1 )c > uj (
uj (y) = uj
c) = uj (c).
Moreover,
m

i=1
pi yi =
m

i=1
m

m
pi
cij + (1 )cij =
pi cij + m
i=1 cij
i=1 cij = Rj + Rj Rj = Rj.
i=1
Hence, y Bj (p) and is also strictly preferred to c and c, which means that c and c are not optima,
as initially conjectured. This establishes uniqueness of the solution to [P1].
58
c
by
A. Mele
2.11. Appendix 2: Proofs of selected results
2.11 Appendix 2: Proofs of selected results

We rst provide a useful result, a well-known theorem on separation of two convex sets. We use this
theorem to deal with the proof of the second welfare theorem (Theorem 2.4) and the existence of state
prices tying up all asset prices together (Theorem 2.10). A nal proof we provide in this appendix is
that of Eq. (2.28).
Minkowskis separation theorem.
Let A and B be two non-empty convex subsets of Rd . If A
(
is closed, B is compact and A B = , then there exists a Rd and two real numbers d1 , d2 such
that:
a d1 < d2 b , a A, b B.
We are now ready to prove Theorems 2.4 and 2.10.
&
'
j = cj : uj (cj ) > uj (
Proof of Theorem 2.4. Let c be a Pareto
optimum and B
cj ) . Let us
)
*
n
j n
j
j
= /n B
consider the two sets B

j=1 j and A = (c )j=1 : c 0 j,
j=1 c = w . A is the set of all
possible combinations of feasible allocations. By (
the denition of a Pareto optimum, there are no
or A B
= . In particular, this is true for all compact
elements in A that are(simultaneously in B,
subsets B of B, or A B = . Because A is closed, then, by the Minkowskis separating theorem,

there exists a p Rm and two distinct numbers d1 , d2 such that
p a d1 < d2 p b, a A, b B.
n
This means that for all allocations cj j=1 preferred to c, we have:
p
n
wj < p
j=1
or, by replacing
n
j=1 w
with
n
n
cj ,
j=1
j ,
j=1 c
n
c < p
j=1
n
cj .
(2A.1)
j=1

Next we show that p > 0. Let ci = nj=1 cij , i = 1, , m, and partition c = (
c1 , , cm ). Let us apply
the inequality in (2A.1) to c A and, for > 0, to c = (
c1 + , , cm ) B. We have p1 > 0, or
p1 > 0. By reiterating the argument, pi > 0 for all i. Finally, we choose cj = cj + 1m n , j = 2, , n,
> 0 in (2A.1), p c1 < p c1 + p 1m or,
p c1 < p c1 ,
for suciently small. This means that u1 (c1 ) > u1 (
c1 ) p c1 > p c1 . This means that c1 =
1
1
j
arg maxc1 u1 (c ) s.t. p c = p c . By symmetry, c = arg maxcj uj (cj ) s.t. p cj = p cj for all j.
Proof of Theorem 2.10. The condition in (2.12) holds for any compact subset of Rd+1
+ , and
d+1
therefore it holds when it is restricted to the unit simplex in R+ ,
W
S d = {0} .
Rd+1 : w
d1 < d2 ,
w W , S d .
By the Minkowskis separation theorem,
s , s = 1, , d. On the other hand,
By walking along the simplex boundaries, one nds that d1 <
59
c
by
A. Mele
Rd+1 . Next we show that w

= 0. Assume the contrary,
0 W , which reveals that d1 0, and
++
i.e. w W that satises at the same time w = 0. In this case, there would be a real number
such that w W and w
> d2 , a contradiction. Therefore, we have
with sign() = sign(w )
m
W = (
(S V ) ) = (
0S +
V ), R , where
(d) contains the last d components
0=
(d)
Whence S = V , where = 1 , , d .
of .
The proof of the converse is immediate (hint: multiply by ): shown in further notes.
The proof of the second part is the following one. We have that each point of Rd+1 is equal to
each point of W plus each point of W , or dim W + dim W = d + 1. Since dim W =
rank(W ), dim W = d + 1 dim W , and since S = V in the absence of arbitrage opportunities,
dim W = dim V = m, whence:
dim W = d m + 1.
:
W = 0, or
W . Whence dim W 1 in
In other terms, before we showed that
the absence of arbitrage opportunities. The previous relation provides more information. Specically,
Rd+1 :
W = 0} = 1, which means that
dim W = 1 if and only if d = m. In this case, dim{
+
0S +
, for every positive scalar , but there are no

the relation
d V = 0 also holds true for =
other possible candidates. Therefore, = 1 , , d is such that = (), and then it is unique.
0
0
&
'
d+1
By a similar reasoning, dim{ R+ : W = 0} = dm+1 dim Rd++ : S = V = dm.

(2)()
Proof of Eq. (2.28). Let Ss ,s
be the price at t = 2 in state s if the state in t = 1 was s, for the

(2)
(2)(1)
(2)(m)
Arrow security promising 1 unit of numeraire in state at t = 3. Let Ss ,s = [Ss ,s , , Ss ,s
]. Let
(1)(s)
i
be the quantity purchased at t = 1 in state i of Arrow securities promising 1 unit of numeraire

if s at t = 2. Let p2s,i be the price of the good at t = 2 in state s if the previous state at t = 1 was i.
(1)(i)
(2)()
(1)
(2)
Let S (0)(i) and Ss

correspond to Ss ,s ; S (0) and Ss correspond to Ss ,s .
The budget constraint is
m
(0)
(0)
S (0)(i) (0)(i)
p0 (c0 w0 ) = S =
i=1
m

1

(1) (1)
(1)(i) (1)(i)
(0)(s)
(0)(s)
1
1
Ss s =
Ss
s
, s = 1, , d.
ps cs ws =
i=1
(1)(i)
where Ss
is the price to be paid at time 1 and in state s, for an Arrow security giving 1 unit of
numeraire if the state at time 2 is i.
By replacing the second equation of (3.9) in the rst one:
p0 (c0 w0 ) =
0 = p0 (c0 w0 ) +
= p0 (c0 w0 ) +
= p0 (c0 w0 ) +
m

i=1
m

i=1
m

i=1
m

i=1
!

(1) (1)
S (0)(i) p1i c1i wi1 + Si i
S (0)(i) p1i
m
1

(1) (1)
1
ci wi +
S (0)(i) Si i
i=1
m
m

(1)(j) (1)(j)
S (0)(i) p1i c1i wi1 +
S (0)(i)
Si
i
i=1
i=1
m
m

(1)(j) (1)(j)
S (0)(i) p1i c1i wi1 +
S (0)(i) Si
i
i=1 j=1
60
c
by
A. Mele

At time 2,
m

(1)(s)
(2) (2)
(1)(s)
(2)() (2)()
2
p2s,i c2s,i ws,i
= i
Ss,i s,i = i
Ss,i s,i , s = 1, , d.
=1
(2)
Here Ss,i is the price vector, to be paid at time 2 in state s if the previous state was i, for the Arrow
securities expiring at time 3. The other symbols have a similar interpretation.
By plugging (???) into (???),
0 = p0 (c0 w0 ) +
= p0 (c0 w0 ) +
+
m
m
m
i=1 j=1 =1
m
i=1
m

i=1
!
m
m

(1)(j)
(2) (2)
2
S (0)(i) p1i c1i wi1 +
S (0)(i) Si
p2j,i c2j,i wj,i
+ Sj,i j,i

S (0)(i) p1i c1i wi1 +
(1)(j) (2)() (2)()

Sj,i j,i .
S (0)(i) Si
i=1 j=1
m
m

i=1 j=1
(1)(j) 2
pj,i (c2j,i
S (0)(i) Si
2
wj,i
)
In the absence of arbitrage opportunities, t+1,s Rd++ - the state prices vector for t + 1 if the
state in t is s - such that:
(t)()
Ss ,s = t+1,s e , = 1, , m,
where e Rd+ and has all zeros except in the -th component which is 1. Next, we restate the
()
previous relation in terms of the kernel mt+1,s = (mt+1,s )d=1 and the probability distribution Pt+1,s =
()
(Pt+1,s )d=1 of the events in t + 1 when the state in t is s :

(t)()
()
()
Ss ,s = mt+1,s Pt+1,s ,
= 1, , m.
By replacing in (???), and imposing the transversality condition:

m
m
m
m
1 =1 2 =1 3 =1 4 =1
m
t =1
(1)(2 ) (2)(3 ) (3)(4 )

S2 ,1 S3 ,2
S (0)(1 ) S1
we get eq. (2.28).
61
(t1)( )
t
St1 ,t2
0,
c
by
A. Mele
2.12. Appendix 3: The multicommodity case
2.12 Appendix 3: The multicommodity case

The multicommodity case is interesting, but at the same time is extremely delicate to deal with when
markets are incomplete. While standard regularity conditions ensure the existence of an equilibrium
in the static and complete markets case, only generic existence results are available for the incoplete
markets cases. Hart (1974) built up well-chosen examples in which there exist sets of endowments
distributions for which no equilibrium can exist. However, Due and Shafer (1985) showed that such
sets have zero measure, which justies the terminology of generic existence.
Here we only provide a derivation of the contraints. mt commodities are traded in period t (t = 0, 1).
The states of nature in the second period are d, and the number of traded assets is a. The rst period
budget constraint is:
p0 e0j = Sj , e0j c0j w0j
(1)
(m )
(1)
(m )
where p0 = (p0 , , p0 1 ) is the rst period price vector, e0j = (e0j , , e0j 1 ) is the rst period
excess demands vector, S = (S1 , , Sa ) is the nancial asset price vector, and j = (1j , , aj ) is
the vector of assets quantities that agent j buys at the rst period.
The second period budget constraint is,
E1 p1 = B j
ddm2
where
E1
ddm2
e1 ( 1 )
1m2
1m2
1m2
1m2
e1 ( 2 )
1m2
1m2
1m2
e1 ( d )
1m2
1m2
is the matrix of excess demands, p1 = (p1 ( 1 ), , p1 ( d )) is the matrix of spot prices, and
m2 1
B =
da
m2 1
v1 ( 1 )
va (1 )
..
v1 ( d )
va (d )
is the payos matrix. We can rewrite the second period constraint as p1 e1j = B j , where e1j is
dened similarly as e0j , and p1 e1j (p1 ( 1 )e1j ( 1 ), , p1 ( d )e1j ( d )) . The budget constraints are
then,
p0 e0j = Sj , p1 e1j = Bj .
Now suppose that markets are complete, i.e., a = d and B can be inverted. The second constraint
is then: j = B 1 p1 e1j . Consider without loss of generality Arrow securities, or B = I. We have
j = p1 e1j , and by replacing into the rst constraint,
0 = p0 e0j + Sj
= p0 e0j + Sp1 e1j
= p0 e0j + S (p1 ( 1 )e1j ( 1 ), , p1 ( d )e1j (d ))
d

= p0 e0j +
Si p1 ( i )e1j ( i )
i=1
m
1
(h) (h)
p0 e0j +
d
Si
m2
()
p1 ( i )e1j ( i )
i=1
=1
d m
2

(h) (h)
()
p0 e0j +
p1 ( i )e1j ( i )
i=1 =1
h=1
h=1
m
1
62
c
by
A. Mele

()
()
where p1 ( i ) Si p1 ( i ). The price to be paid today for the obtention of a good in state i is equal
()
to the price of an Arrow asset written for state i multiplied by the spot price p1 ( i ) of this good in this
()
state; here the Arrow-Debreu state price is p1 ( i ). The general equilibrium can be analyzed by making
reference to such state prices. From now on, we simplify and set m1 = m2 m. Then we are left with
(1)
(m)
(1)
(m)
determining m(d + 1) equilibrium prices, i.e. p0 = (p0 , , p0 ), p1 ( 1 ) = (
p1 ( 1 ), , p1 ( 1 )),
(1)
(m)
, p1 ( d ) = (
p1 ( d ), , p1 (d )). By exactly the same arguments of the previous chapter, there
exists one degree of indeterminacy. Therefore, there are only m(d + 1) 1 relations that can determine
the m(d + 1) prices. (Price normalization can be done by letting one of the rst period commodities
be the numeraire.) On the other hand, in the initial economy we have to determine m(d + 1) + d prices
Rm(d+1) Rd++ which are the solution to the system:
(
p, S)
++
n
j=1
= 0,
e0j (
p, S)
n
= 0,
e1j (
p, S)
j=1
n
= 0,
j (
p, S)
j=1
where the previous functions are obtained as solutions to the agents programs. When we solve for
Arrow-Debreu prices, in a second step we have to determine m(d + 1) + d prices starting from the
knowledge of m(d + 1) 1 relations dening the Arrow-Debreu prices, which implies a price indeterminacy of the initial economy equal to d + 1. In fact, it is possible to show that the degree of
indeterminacy is only d 1.
63
c
by
A. Mele
References
Arrow, K. J. (1953): Le role des valeurs boursières pour la repartitition la meilleure des
risques. Econometrie 41-48. CNRS, Paris. Translated and reprinted in 1964: The Role
of Securities in the Optimal Allocation of Risk-Bearing. Review of Economic Studies 31,
91-96.
Debreu, G. (1954): Valuation Equilibrium and Pareto Optimum. Proceedings of the National
Academy of Sciences 40, 588-592.
Debreu, G. (1959): Theory of Value: An Axiomatic Analysis of Economic Equilibrium. New
Haven: Yale University Press.
Due, D. (2001): Dynamic Asset Pricing Theory. Princeton: Princeton University Press.
Due, D. and W. Shafer (1985): Equilibrium in Incomplete Markets: I. A Basic Model of
Generic Existence. Journal of Mathematical Economics 13 285-300.
Hart, O. (1974): On the Existence of Equilibrium in a Securities Model. Journal of Economic
Theory 9, 293-311.
He, H. and N. Pearson (1991): Consumption and Portfolio Policies with Incomplete Markets
and Short-Sales Constraints: The Innite Dimensional Case. Journal of Economic Theory
54, 259-304.
64
3
Infinite horizon economies
3.1 Introduction
We study asset prices in multiperiod economies, where agents either live forever, and have access
to a set of complete markets, or belong to overlapping generations. We consider models without
and with production, without and with money, and develop the fundamental tools we need in
subsequent chapters, to analyze nancial frictions, bubbles and sunspots in capital markets.
3.2 Consumption-based asset evaluation

3.2.1
Recursive plans: introduction
We consider a simple, benchmark case, arising in the absence of any risks for a decision maker.
Consider an agent endowed with initial wealth equal to w0 , who solves the following problem:
V (w0 ) max
(ct )t=0
t u (ct )
t=0
s.t. wt+1 = (wt ct )Rt+1 ,
[3.P1]
(Rt )
t=0
given
The previous problem can be reformulated in a recursive format:

V (wt ) = max [u (ct ) + V (wt+1 )]
ct
s.t. wt+1 = (wt ct ) Rt+1 .
(3.1)
By replacing the wealth constraint into the maximand, it is easily checked that the rst-order
condition for c leads to, u (ct ) = V (wt+1 )Rt+1 . Therefore, the consumption policy is a function
of both wealth and the interest rate, which for sake of simplicity we denote as c (wt ). The value
function and the rst-order condition, then, can be written as:
V (wt ) = u (c(wt )) + V ((wt c (wt )) Rt+1 ) ,
u (c (wt )) = V ((wt c (wt )) Rt+1 ) Rt+1 .
By dierentiating the value function, and using the rst-order condition,

V (wt ) = u (c (wt )) c (wt ) + V ((wt c (wt )) Rt+1 ) (1 c (wt )) Rt+1 = u (c (wt )) .
c
by
A. Mele
3.2. Consumption-based asset evaluation
Therefore, V (wt+1 ) = u (c (wt+1 )) too, and by substituting back into the rst-order condition,
1
u (c (wt+1 ))
=
.
u (c (wt ))
Rt+1
(3.2)
The economic intuition underlying Eq. (8.7) is the same as that we saw in the two-period
economy analyzed in Chapter 2. Eq. (8.7) says that the present consumption I give up, at t,
to obtain addition consumption at t + 1, has to equal a pure discount bond issued at t and
expiring the next period, along an optimal consumption path.
We can arrive at the very same conclusions, following an alternative approach, based on
Lagrange multipliers. This approach is useful when dealing with more intricate issues relating
to production economies or economies with nancial frictions, as we shall see in this and further
chapters. So consider the constraint in program [3.P1]. Savings at time t are savt wt ct .
Using this denition, the constraint in [3.P1] is: ct+1 + savt+1 = Rt+1 savt , with sav1 = w0 ,
given. Let t be a sequence of Lagrange multipliers associated to these constraints. Consider
the program,
L (sav1 )
max
(ct ,savt )
t=0

t

u (ct ) t (ct + savt Rt savt1 ) ,
t=0
where t is a sequence of Lagrange multipliers. The rst-order condition for consumption ct is,
t u (ct ) = t , and the rst-order condition for savings savt leads to: t = t+1 Rt+1 . Putting all
together yields precisely Eq. (8.7). Note that the same program can be cast, and solved, in a
recursive format,
L (savt1 ) = max [u (ct ) t (ct + savt Rt savt1 ) + L (savt )] .
ct ,savt ,t
The rst-order condition for consumption and savings are u (ct ) = t and t = L (savt ),
respectively. By replacing the rst-order condition for t , i.e. the budget constraint, and dierentiating L (savt1 ), leaves L (savt1 ) = L (savt ) Rt . These conditions lead to Eq. (8.7).
As a simple example, consider the case of a logarithmic utility function, u (c) = ln c. Let us
guess that the value function is V (wt ) V (wt ; Rt ) = at + b ln wt . The rst-order condition
then yields c (w) = b1 w. By Eq. (8.7), then, wt+1 = wt Rt+1 . Comparing the right hand
side of this equation with the right hand side of the constraint in the program [3.P1], leaves
c (wt ) = (1 ) wt ; in other terms, b = (1 )1 .1
Next, we introduce uncertainty.
3.2.2
The marginalist argument
Consider the following thought experiment. At time t, I give up to a small quantity of consumption equal to ct . The reduction in the (current) utility is, then, equal to t u (ct )ct .
But by investing ct in a safe asset, I can have access to ct+1 = Rt+1 ct additional units
of consumption at time t + 1. These additional consumption units lead to an expected utility
gain equal to t+1 E (u (ct+1 ) ct+1 ). If ct and ct+1 are part of an optimal consumption plan,
1 To pin down the coecient series a , use the denition of the value function, V (w ; R ) u (c (w )) + V (w
t
t
t
t
t+1 ; Rt+1 ). By
plugging V (w, Rt ) = at + b log w and c (w) = (1 )1 w into this denition leaves, at = ln (1 ) + at+1 + 1
ln (Rt+1 ). If
R is constant, at is also constant, and equal to (ln (1 ) +
ln (R))/ (1 ).
66
c
by
A. Mele
I should be left with no incentives to implement these intertemporal consumption transfers.

Therefore, along an optimal consumption plan, any reductions and gains in the welfare of the
type considered above need to be identical:
u (ct ) = E (u (ct+1 )Rt+1 ) .
This relation generalizes Eq. (8.7). Next, suppose that at time t, ct can be invested in a risky
asset whose price is St . I can buy ct / St units of this asset. Come time t + 1, I could sell the
asset for St+1 , pocket its divend Dt+1 , if any, and nance additional units of consumption equal
to ct+1 = ( ct / St ) (St+1 + Dt+1 ). The reduction in the current utility is t u (ct )ct , and
the boost in the expected utility at time t + 1 is t+1 E (u (ct+1 )ct+1 ). Again, if I am following
an optimal consumption policy, the incentives for these kind of intertemporal transfers should
not exist. Therefore, the celebrated Lucas asset pricing equation holds:

St+1 + Dt+1
u (ct ) = E u (ct+1 )
.
(3.3)
St
We now derive Eq. (3.3) through dynamic programming methods, which are essential, once we
wish to work through more complex models such as those including nancial frictions.
3.2.3
Lucas model
3.2.3.1 The optimality condition
We consider markets for m trees, and assume that the only source of risk stems from the
dividends related to these trees: D = (D1 , , Dm ). We assume D is a Markov process and
denote its conditional distribution function with P ( Dt+1 | Dt ). A representative agent solves
the following program:
.
-

i
V (t ) =
max E
u(ct+i ) Ft
[3.P2]

(ct+i ,t+i )i=0
i=0
s.t. ct + St t+1 = (St + Dt ) t
where t+1 Rm is Ft -measurable, that is, t+1 needs to be chosen at time t. We can solve the
program [3.P2], using the same recursive approach in Section 3.2.1, once due account is made
of uncertainty. The Bellmans equation is:
V (t , Dt ) = max E [ u(ct ) + V (t+1 , Dt+1 )| Ft ] s.t. ct + St t+1 = (St + Dt ) t .
ct ,t+1
Similarly as we did for Eq. (3.1), let us replace the budget constraint into the maximand. The
following rst-order condition holds for i :
0 = E [u ((St + Dt ) t St t+1 ) Si,t + V1i ( t+1 , Dt+1 )] ,
(3.4)
where the subscript in the value function on the right hand side denotes a partial derivative:
V1i (, D) = (, D) /i . The optimal policy, t+1 is a function of the current state, (t , Dt ),
say t+1 = T (t , Dt ). By dierentiating the value function with respect to i , and using the
previous rst-order condition, leaves:
$
%
.
m
m

V1i (t , Dt ) = E u (ct ) Si,t + Di,t
Sj,t T1ji (t , Dt ) +
V1i ( t+1 , Dt+1 ) T1ji (t , Dt )
j=1
j=1
= u (ct ) (Si,t + Di,t ) ,
67
c
by
A. Mele
where we have dened T1ji ( t , Dt ) = Ti (, D) /j and Ti is the i-th component of the vector
T . Substituting this result into Eq. (3.4) yields precisely the Lucas equation (3.3), holding for
each asset i:

Si,t+1 + Di,t+1
u (ct ) = E u (ct+1 )
.
(3.5)
Si,t
3.2.3.2 Rational expectations equilibrium
(0)
The asset market clears when for each t, t = 1m and t = 0, where (0) denotes the amount
of
mthe riskless asset. By the budget constraint, then, the market for goods alsoclears, ct =
i=1 Dit Dt . A rational expectation equilibrium is a sequence of asset prices (St )t=0 such that
t , and each asset price is
the optimality condition in Eq. (3.5) holds, the markets clear, ct = D
a function of the state, Si,t = Si (Dt ) say. All in all,
"

t+1 ) Si (D
t+1 ) + Di,t+1 dP (Dt+1 | Dt ) .
u (Dt )Si (Dt ) = u (D
(3.6)
This is a functional equation in Si (). Let us focus, rst, on the IID case: P (Dt+1 | Dt ) =
P (Dt+1 ).
IID shocks
Eq. (3.6) simplies to:

t )Si (Dt ) =
u (D
"

t+1 ) Si (D
t+1 ) + Di,t+1 dP (Dt+1 ) .
u (D
t )Si (Dt )
Note that the right hand side of this equation is independent of D. Therefore, u (D
equals some constant (say), which we can easily nd by substituting it back into the previous
equation, leaving:
"
t+1 )Di,t+1 dP (Dt+1 ) .

i =
u (D
1
The solution for Si (D) is then:
i
Si (Dt ) = .
u (Dt )
D)
Note, the elasticity of the price to dividend equals uu ((D)
Di , which collapses to relative riskaversion, once we assume only one tree exists, as it is customary. For example, if relative
risk-aversion is constant and equal to ,
"
S(Dt ) = Dt ,
D1 dP (D) .
1
Figure 3.1 depicts the behavior of the asset price function S (D), under the assumption that
is not increasing in .
Only when the representative agents are risk-neutral, = 0, does the asset price collapse to
the constant (1 )1 E(D).
Dependent shocks
i (D) and hi (D)

Dene gi (D) u (D)S
functions, Eq. (3.6) is:
gi (D) = hi (D) +
"
t+1 )Di,t+1 dP ( Dt+1 | D). In terms of these new

u (D
gi (Dt+1 ) dP ( Dt+1 | D) .
68
c
by
A. Mele

S(Dt)
0<<1
(1)1
=1
>1
Dt
FIGURE 3.1. The asset pricing function S (Dt ) in the IID case and constant relative risk-aversion,
equal to .
It is a functional equation in gi , which we can show it admits a unique solution, under the
conditions contained in the celebrated Blackwells theorem below:
Theorem 3.1. Let B(X) the Banach space of continuous bounded real functions on X Rn
endowed with the norm f = supX |f |, f B(X). Introduce an operator T : B(X) B(X)
with the following properties:
(i ) T is monotone: x X and f1 , f2 B(X), f1 (x) f2 (x) T [f1 ] (x) T [f2 ] (x);
(ii ) x X and c 0, (0, 1) : T [f + c] (x) T [f ] (x) + c.
Then, T is a -contraction and, f0 B(X), it has a unique fixed point lim T [f0 ] = f =
T [f ].
So let us introduce the following operator:
T [gi ] (D) = hi (D) +
"
gi (D ) dP ( D | D) .
The existence of gi and, hence, Si , relies on the existence of a xed point of T : gi = T [gi ].
It is easily checked that conditions (i) and (ii) in Theorem 3.1 hold here. To establish that
T : B(D) B(D) as well, it is sucient to show that hi B(D). A sucient condition given
by Lucas (1978) is that u is bounded, and bounded away by a constant u.2
3.2.4
Arrow-Debreu state prices, the CCAPM and the CAPM
Let us consider the case of a single tree. We have the following consumption-based asset pricing
equation:
u (Dt+1 )
St = Et [mt+1 (St+1 + Dt+1 )] , mt+1
.
u (Dt )
2 In this case, concavity of u implies that for each D, 0 = u (0) u (D) + u (D) (D) u
Du (D), which implies that
for each D, Du (D) u
and, hence, hi (D) u
. Then, it is possible to show that the solution is in B(D), which implies that
T : B(D)
B(D).
69
c
by
A. Mele
3.3. Production: foundational issues
By using the same arguments as those in Section 2.6 of the previous chapter, we can show that
the Radon-Nikodym derivative of the risk-neutral probability, P , with respect to P , is:
dP
u (Dt+1 )
( Dt+1 | Dt ) =
.
dP
E [u ( Dt+1 | Dt )]
In the Lucas model, then, the Arrow-Debreu state-price density is:
dP ( Dt+1 | Dt ) = dP ( Dt+1 | Dt ) R1
t .
It is the price to pay, in state Dt , to obtain one unit of the good the next period in state Dt+1 .
as, R
t+1 St+1 +Dt+1 . Then, all the considerations made in
Finally, dene the gross return R
St
Section 2.7 of the previous chapter, are also valid here.
3.3 Production: foundational issues

In the economy of the previous section, the asset reward, is an exogenous datum. In this
chapter, we lay down the foundations for the analysis of production-based economies, where
rms maximize their value and set dividends endogenously. In these economies, production and
capital accumulation are endogenous. In this section, we review the foundational issues that
arise in economies with productive capital. In the next section, we develop the asset pricing
implications of these economies, in absence of frictions. In Part II, we extend the framework
in this and the next section, and examine the asset price implications deriving from nancial
frictions.
3.3.1
Decentralized economy
A continuum of identical rms in (0, 1) have access to capital and labor markets, and the following technology: (K, N) Y (K, N),where Yi (K, N) > 0, yii (K, N) < 0, limK0+ Y1 (K, N ) =
limN 0+ Y2 (K, N ) = , limK Y1 (K, N ) = limN Y2 (K, N ) = 0, and subscripts denote
partial derivatives. We assume Y is homogeneous of degree one, i.e. Y (K, N ) = Y (K, N)
for all > 0. Per capita production is y(k) Y ( K/ N, 1), where k K/ N is per-capita
capital, Population growth can be non-zero, i.e. N satises Nt /Nt1 = (1 + n). Firms purchase
capital and labor at prices R = Y1 (K, N ) and w = Y2 (K, N) = w. We have,
R = y (k) ,
w = y (k) ky (k) .
The Nt consumers live forever. We assume each consumer oers inelastically one unit of labor,
and that, for now, that N0 = 1 and n = 0. The resource constraint for the consumer is:
ct + st = Rt st1 + wt Nt ,
Nt 1,
t = 1, 2, .
(3.7)
At each time t 1, the consumer saves st1 units of capital, which he lends to the rm. At time
t, the consumer receives the gross return on savings from the rm, Rt st1 , where Rt = y (kt ),
plus the wage receipts wt Nt . Then, he uses these resources to consume ct and lend st to the
rm. At time zero,
c0 + s0 = V0 Y1 (K0 , N0 )K0 + w0 N0 ,
70
N0 1.
c
by
A. Mele
Following the approach developed in Chapter 2, we can write down a single budget constraint,
obtained iterating Eq. (3.7):
T

ct wt Nt
sT
0 = c0 +
+ 0T
V0 ,
0t
i=1 Ri
i=1 Ri
t=1
and imposing the transversality condition:
lim sT
so as to have:
max
(ct )
t=0

t=1
T
1
Ri1 = 0,
(3.8)
i=1

ct wt Nt
u(ct ), s.t. V0 = c0 +
.
0t
i=1 Ri
t=1
t
[3.P3]
The economic interpretation of the transversality condition (4.17) is the following. The rstorder conditions of the program [3.P3] are:
t u (ct ) = l 0t
i=1
Ri
(3.9)
where l is a Lagrange multiplier. In equilibrium, current savings equal next period capital, or
kt+1 = st . Therefore, Eq. (4.17) is:
lim T u (cT ) kT +1 = 0.
(3.10)
That is, the economic value of capital is capital weighted by discounted marginal utility, which
needs to be zero, eventually.
The rst-order condition (3.9) leads to the usual optimality condition in Eq. (8.7), where this
t ) satisfying
time, Rt+1 = y (kt+1 ). In this economy, an equilibrium is a sequence ((
c, k)
t=0
kt+1 = y (kt ) ct
u (ct+1 )
1
(3.11)
=

u (ct )
y (kt+1 )
and the transversality condition in Eq. (3.10). The rst equation in this system is simply this:
capital available for producing the next period, kt+1 , is equal to savings, st y (kt ) ct .
3.3.2
Centralized economy
The market solution in (3.11) can be implemented by a social planner, who solves the following
program:

V (k0 ) max
i u (ct )
[3.P4]
(ct ,kt )t=0
i=0
s.t. kt+1 = y (kt ) ct , k0 given
under the further transversality condition in Eq. (3.10).
71
c
by
A. Mele
The program in [3.P4] is easily solved. By replacing the constraint into the utility function, and taking derivatives with respect to kt , leads directly to the second equation in (3.11).
Alternatively, let us introduce the Lagrangian,
L (k0 ) =
max
(ct ,kt+1 )
t=0

t

u(ct ) t (kt+1 y(kt ) + ct ) .
t=0
The rst-order condition with respect to consumption is t = t u (ct ), and the condition for
capital is t1 = t y (kt ). Putting these conditions together, leads to the second equation in
(3.11). The same argument can be made, following a recursive approach. We have:
L (kt ) = max [u (ct ) t (kt+1 y (kt ) + ct ) + L (kt+1 )] .
ct ,kt+1 ,t
The rst-order condition for consumption is t = u (ct ), and that for capital is t = L (kt+1 ).
By replacing the rst-order condition for t (i.e., the constraint in program [3.P4]), and differentiating with respect to kt , yields L (kt ) = L (kt+1 ) y (kt ). These three conditions lead,
again, to the second equation in (3.11).
Finally, consider the Bellmans equation:
V (kt ) = max [u (ct ) + V (kt+1 )] , s.t. kt+1 = y (kt ) ct .
ct
The rst-order condition leads to, u (ct ) = V (y (kt ) ct ). Let us denote the policy with
ct = c (kt ). In terms of the policy c function, the value function and the rst-order conditions
are:
V (kt ) = u (c (kt )) + V (y (kt ) c (kt )) , u (c (kt )) = V (y (kt ) c (kt )) .
By dierentiating the value function:
V (kt ) = u (c (kt )) c (kt ) + V (y (kt ) c (kt )) (y (kt ) c (kt )) = u (c (kt )) y (kt ) .
By replacing back into the rst-order condition, we obtain the second equation in (3.11).
3.3.3
Dynamics
We study the dynamics of the system in (3.11) in a small neighborhood of the stationary state,
dened as the pair (c, k), solution to:
c = y (k) k,
1
.
(k)
A rst-order expansion of each equation in (3.11) around its stationary state, yields the
following linear system:
$
%

y
(k)
1
kt+1
kt
=A
, A
.
(3.12)
uu(c)
y (k) 1 + uu(c)
y (k)
ct+1
ct
(c)
(c)
The solution to this system is obtained with the tools reviewed in Appendix 1 of this chapter.
It is:
kt = v11 1 t1 + v12 2 t2 , ct = v21 1 t1 + v22 2 t2 ,
(3.13)
72
c
by
A. Mele

ct
c0 = c + (v21/v11) (k0 k)
c
c = y(k) k
c0
k0
kt
k k*
FIGURE 3.2.

where: i are constants that depend on the initial state, i are the eigenvalues of A, and vv11
,
21
v12
are the eigenvectors associated with i . In Appendix 1, we show that 1 (0, 1) and
v22
2 > 1. The proof we provide in the appendix is important, as it illustrates precisely how the
neoclassical model reviewed in this section, needs to be modied to induce indeterminacy in
the dynamics of capital and consumption. A critical step in that proof relies on the assumption
of diminishing returns, i.e. y (k) > 0.
Let us return to the equations in (3.13). First, we need to rule out an explosive behavior
of kt and ct , for otherwise we would contradict (i) that (c, k) is a stationary point, and (ii)
the optimality of the trajectories. Since 2 > 1, the only possibility is to lock the initial
state (k0 , c0 ) in such a way that 2 = 0, which yields the following set of initial conditions:
k0 = v11 1 and c0 = v21 1 , or kc0 = vv21
.3 Therefore, the set of initial points that ensure a
11
0
non-explosive path must lie on the line c0 = c + vv21
(k0 k). Since k is a predetermined variable,
11
there exists one, and only one, value of c0 , which ensures a non-explosive path of the system
around its steady state, as Figure 3.2 illustrates. In this gure, k is dened as the solution of
1 = y (k ) k = (y )1 [1], and k = (y )1 [ 1 ].
The usual word of caution is in order. A linear approximation might turn out to be misleading.
We develop one example where the dynamics of the system could be quite dierent from those
analyze here, when we start away from the stationary state. Let y(k) = k , u(c) = ln c. It is
easy to show that the exact solution is:
ct = (1 ) kt , kt+1 = kt .
Figure 3.3 depicts the nonlinear manifold associated with this system, and its linear approximation. For example, let = 0.99 and = 0.3. Then, the (linear) saddlepath is, approximately,
ct = c + 0.7101 (kt k) , where: c = (1 ) k , k = 1/(1) ,
where kt = 1 kt1 , and 1 = 0.3.
3 In
fact, Appendix 1 shows that the converse is also true, i.e.
c
0
0
k
73
v21
v11
2 = 0.
c
by
A. Mele

ct
linear approximation
steady state
nonlinear stable manifold
kt
FIGURE 3.3.
3.3.4
Stochastic economies
Real business cycle theory is the application of general equilibrium theory to the quantitative analysis of business cycle uctuations. Edward Prescott (1991, p. 3)
The Kydland and Prescott model is a complete markets set-up, in which equilibrium and
optimal allocations are equivalent. When it was introduced, it seemed to manymyself
includedto be much too narrow a framework to be useful in thinking about cyclical
issues. Robert Lucas (1994, p. 184)
In its simplest version, real business cycle theory is an extension of the neoclassical model
of Section 3.3.3, in which random productivity shocks are added. The engine of uctuations,
then, comes from the real sphere of the economy. This approach is in contrast with the Lucas
approach of the 1970s, based on information and money, where uctuations arise due to information delays with which agents discover the nature of a shock (real or monetary). As further
reviewed in Chapter 9, the Lucas information-theoretic approach has been, instead, more successful in inspiring work on the formation of asset prices, leading to the development of market
microstructure theory and, more generally, to information driven explanations of asset prices.
Despite the remarkable switch in the economic motivation, the paradigm underlying real
business cycle theory is the same as the information-based approach of Lucas, as it relies on
rational expectations: macroeconomic uctations and, then, as we shall explain, asset prices
uctuations, stem from the optimal response of the agents vis-à-vis exogeneous shocks: agents
implement action plans that are state-contingent, i.e. they decide to consume, to work and to
invest according to the history of shocks as well as the present shocks they observe.
3.3.4.1 Basic model
We consider an economy with complete markets and no frictions, such that its equilibrium
allocations are Pareto-optimal. To characterize these allocations, we implement them through
the following program of a social planner:
-
.

E
t u(ct ) ,
(3.14)
V (k0 , s0 ) = max
(ct )t=0
74
t=0
c
by
A. Mele
subject to a capital accumulation constraint, with capital depreciation. Let It denote new
investment. It is:
It = Kt+1 (1 ) Kt .
(3.15)
At time t 1, the available productive capital is Kt . At time t, a portion Kt of this capital is

lost, due to depreciation. Therefore, at time t, the productive system is left with (1 ) Kt units
of capital. The capital available at time t, Kt+1 , equals the capital already in place, (1 ) Kt ,
plus new investments, which is exactly Eq. (3.15).
Next, normalize population normalized to one, such that Kt = kt . The goods market clearing
condition is:
y (kt , t ) = ct + It ,
where y(kt , st ) is the production function, which is Ft -measurable, and s is the source of
randomnessthe engine for random uctuations of the endogeneous variables. By replacing
Eq. (3.15) into the equilibrium condition,
kt+1 = y (kt , t ) ct + (1 ) kt .
(3.16)
So the planner maximizes the utility in Eq. (3.14), under the capital accumulation constraint
in Eq. (3.16).
We assume that y (kt , st ) st y (kt ), where y is as in Section 3.2, and (st )
t=0 is solution to:
st+1 = st t+1 ,
(3.17)
where (0, 1), and (t )

t=0 is a IID sequence with support s.t. st 0. In this economy, every
asset is priced as in the Lucas model of the previous section. Therefore, the gross return on
savings s y (k ) satises:
u (ct ) = Et (u (ct+1 ) (st+1 y (kt+1 ) + 1 )) .
(3.18)
A rational expectation equilibrium is a stochastic process (ct , kt )

t=0 , satisfying Eq. (3.16), the
Euler equation in (3.18), for given k0 and s0 .
We show the existence of a saddlepoint path for the linearized version of Eqs. (3.16)-(3.17)(3.18), which implies determinacy of the stochastic (linearized) equilibrium.4 We study the
behavior of (c, k, s)t in a neighborhood of E(t ). Let (c, k, s) be consumption, capital and
productivity shock, corresponding to , obtained replacing into Eqs. (3.16)-(3.17)-(3.18), and
assuming no uncertainty takes place:
1
c = sy (k) k, s = 1 , =
sy (k)
1
.
+1
A rst-order approximation to Eqs. (3.16)-(3.17)-(3.18) around (k, c, s), leaves:

zt+1 =
zt + Rut+1 ,
(3.19)
where we have dened xt xtxx , and zt = (kt , ct , st ) , ut = (uc,t , us,t ) , uc,t = ct Et1 (
ct ),
us,t = st Et1 (
st ) = t , and, nally,
1
kc
s y(k)
0 0
k
(c)
u (c)
= cuu (c)
R = 1 0 .
sky (k) 1 + u
sy (k) cu
(c) s(sy(k)y (k) + y (k)) ,
(c)
u (c)
0 1
0
0
4A
stochastic equilibrium is the situation where there is a stationary measure (denition: p(+) =
the transition measure) generating (ct , kt )
t=1 .
75
(+/)dp(), where is
c
by
A. Mele
Let us consider the characteristic equation:

u (c)
2
1
1
0 = det ( I) = ( ) + 1 + sy (k) +
.
u (c)
A solution is 1 = . By the same arguments produced for the deterministic case of Section
3.3.3 (see Appendix 1), one nds that 2 (0, 1) and 3 > 1.5 As for the deterministic case in
Section 3.3.3, we can diagonalize the system by rewriting = P P 1 , where is a diagonal
matrix that has the eigenvalues of on the diagonal, and P is a matrix of the eigenvectors
associated to the roots of . The system in (3.19) is, then:
yt+1 =
yt + wt+1 ,
(3.20)
where yt P 1 zt and wt P 1 Rut . The third equation of this system is:

y3,t+1 = 3 y3t + w3,t+1 ,
(3.21)
and y3 explodes unless y3t = 0 for all t, which is only possible when w3t = 0 for all t.6
The condition that y3t 0 carries an interesting economic interpretation: it tells us that the
only sources of uncertainty in this system can stem from shocks to the fundamentals, or that
there can not be extraneous sources of noise, or sunspots. The reasons for this are easy to
zt . We have:
explain. Let yt = P 1 zt
0 = y3t = 31 kt + 32 ct + 33 st .
(3.22)
Eq. (3.22) shows that the three state variables, kt , ct and st , are are mutually linked through a
two-dimensional plane. This plane is the saddlepoint of the economy, where the state variables
do exhibit a stable behavior, and is formally dened as:

'
&
S = x R3 3. x = 0 , 3. = ( 31 , 32 , 33 ).
Furthermore, Eq. (3.22) implies that a linear relation exists between the two expectational
errors:
33
For all t, uct =
ust (no-sunspots).
(3.23)
32
Eq. (3.23) is a no-sunspots condition, as it says that the expectational error to consumption
can not be independent of the expectational shock on the fundamentals of the economy, which in
this simple economy relates to technological shock. In other words, the source of uncertainty we
have assumed in this economy, relates to the technological shock. The remaining expectational
errors can only be perfectly correlated to the expectational shock in technology or, there are
no sunspots.
The manifold S brings, mathematically, the same meaning as the stable relation depicted in
Figure 3.2, for the deterministic case. In this section, S is convergent subspace, with dim(S) = 2,
5 The
linearized model in this section has state variables expressed in growth rates here. However, we can always reformulate this
model in terms of rst dierences, by pre- and post- multiplying by appropriate normalizing matrices. As an example, if G i the
3 3 matrix that has k1 , 1c and 1s on its diagonal, (3.19) can be written as: E(zt+1 z) = G1 G (zt z), where zt = (kt , ct , st ),
and we would arrive at the same conclusions. It is tedious but easy to check that the model in this section collapses to that in
Section 3.3.3, once we set t = 1, for each t, and s0 = 1.
(T t)
6 In other words, Eq. (3.21) implies that y
3t = 3
Et (
y3,t+T ), and for all T . Because 3 > 1, this relation holds only when
y3t = 0 for all t.
76
c
by
A. Mele
which is the number of roots with modulus less than one. In other words, in this economy with
two predetermined variables, k0 and s0 , there exists one, and only one, value of of c0 in S, which
0
33 s
ensures stability, and is given by c0 = 31 k0+
. This reasoning generalizes that we made
32
for the deterministic case in Section 3.3.3, and is generalized further in Appendix 1.
The solution to the linearized model can be computed by generalizing the reasoning for the
deterministic case. First, by Eq. (3.20) y is:
yit =
ti yi0
+ it ,
it
t1
ji wi,tj ,
j=0
which implies the solution for z is:

zt = P yt = (v1 v2 v3 )
yt =
3

i=1
vi yit =
3

i=1
vi yi0 ti +
3
vi it .
i=1
To pin down the components of y0 , note that z0 = P y0 y0 = P 1 z0

z0 . The stability
(3)
condition then requires that the state variables be in S, or y0 = 0, which we now use to
implement the solution. We have:
zt = v1 t1 y10 + v2 t2 y20 + v3 t3 y30 + v1 1t + v2 2t + v3 3t .
t
Moreover,
t1thej term v3 3 y30 + v3 3t needs to be zero, because y30 = 0. Finally, we have that
3t = j=0 3 w3,tj , and since w3,t = 0, then, then 3t = 0 as well. Therefore, the solution for
zt is:
zt = v1 t1 y10 + v2 t2 y20 + v1 1t + v2 2t .
3.3.4.2 Frictions, indeterminacy and sunspots
In the neoclassical model that we are analyzing, the equilibrium is determinate. As explained,
this property arises because the number of predetermined variables equals the dimension of the
convergent subspace of the economy. If we managed to increase the dimension of the converging
subspace, the equilibrium would be indeterminate, as further formalized in Appendix 1. As it
turns out, indeterminacy goes hand in hand with sunspots, the expectational shocks extraneous
to those in the economic fundamentals, as we discussed earlier, just after Eq. (3.23).
Introducing sunspots in macroeconomics has been an approach pursued in detail by Farmer
in a series of articles (see Farmer, 1998, for an introductory account of this approach). The
idea is quite interesting, as we know that the basic real business cycle model of this section
needs many extensions in order not to be rejected, empirically, as originally shown by Watson
(1993). In other words, the basic model in this section oers little room for a rich propagation
mechanism, as it entirely relies on impulses, the productivity shocks, which we hardly read
about in the Wall Street Journal, as provocatively put by King and Rebelo (1999). Sunspots
oer an interesting route to enrich the propagation mechanism, although their asset pricing
implications in terms of the model analyzed in this section, have not been explored yet.
In a series of articles, David Cass showed that a Pareto-optimal economy can not harbour
sunspots equilibria. On the other hand, any market imperfection has the potential to be a
source of sunspots. The typical example is the presence of incomplete markets. The neoclassical
model analyzed in this section can not generate sunspots, as it relies on a system of perfectly
competitive markets and absence of any sort of frictions. To introduce sunspots in the economy
77
3.4. Production-based asset pricing
c
by
A. Mele
of this section, we need to think about some deviation from optimality. Two possibilities analyzed in the literature are the presence of imperfect competition and/or externality eects. We
provide an example of these eects, by working out the deterministic economy in Section 3.3.3.
(Generalizations to the stochastic economy in this section are easy, although more cumbersome.)
How is it that a deterministic economy might generate stochastic outcomes, that is, outcomes driven by shocks entirely unrelated to the fundamentals of the economy? Let us imagine
this can be possible. Then, both optimal consumption and capital accumulation in Section
3.3.3 are necessarily random processes. The system in (3.12), then, must be rewritten in an
expectation format,

kt+1
kt
Et
=A
.
ct+1
ct
Next, let us introduce the expectational error process uc,t ct Et1 (
ct ), which we plug back
into the previous system, to obtain:

0
kt+1
kt
=A
+
.
uc,t+1
ct+1
ct
Naturally, we still have 1 (0, 1) and 2 > 1, as in Section 3.3.3. Therefore, we decompose A
as P P 1 , and have:
yt+1 =
yt + P 1 (0 uc,t+1 ) .
y2,t+T ) to hold for all T , we need to have y2t = 0, for all t. Therefore,
Moreover, for y2t = T
2 Et (
the second element of the vector P 1 (0 uc,t+1 ) must be zero, or, for all t,
0 = 22 uc,t 0 = uc,t .
There is no room for expectational errors and, hence, sunspots, in this model. The fact that
2 > 1 implies the dimension of the saddlepoint is less than the number of predetermined
variables. So a viable route to pursue here, is to look for economies such that the saddlepoint
has a dimension larger than one, i.e. such that 2 < 1. In these economies, indeterminancy
and sunspots will be two facets of the same coin. As shown in the appendix, the reasons for
which 2 > 1 relate to the classical assumptions about the shape of the utility function u and
the production function y. We now modify the production function, to see the eect on the
eigenvalues of A.
[Economy with increasing returns]
[Asset pricing implications in further chapters]
3.4 Production-based asset pricing

3.4.1
Firms
For each rm, capital accumulation does satisfy the identity in Eq. (3.15), reproduced here for
convenience:
Kt+1 = (1 ) Kt + It .
(3.24)
The additional assumption we make, is that capital
adjustment is costly: investing It per unit
of capital already in place, Kt , entails a cost KItt , expressed in terms of the price of the nal
good, which we take to be the numeraire, thereby allowing the investment goods to dier from
78
c
by
A. Mele

the nal good the rm produces. An investment of It , then, leads to a cost
that the prot the rm makes at time t is,

It
D (Kt , It ) y (Kt , N (Kt )) wt N (Kt ) pt It
Kt ,
Kt
It
Kt
Kt , such
(3.25)
where y (Kt , Nt ) is the rms production at time t, obtained with capital Kt and labor Nt , and
subject to the same random productivity shocks as those in Section 3.3.4, wt is the real wage,
N (K) is the labor demand schedule, solution to the optimality condition, yN (Kt , N (Kt )) = wt
for all t, and pt is the real price of the investment goods, or uninstalled capital. Finally, the
adjustment-cost function satises 0, 0, 0. In words, capital adjustment is costly
when the adjustment is made fastly. Naturally, is zero in the absence of adjustment costs.
What is the value of the prot, from the perspective of time zero? This question can be
answered, by utilizing the Arrow-Debreu state prices introduced in Chapter 2. At time t, and
in state s, the prot Dt (s) (say) is worth,
0,t (s) D (Kt (s) , It (s)) = m0,t (s) Dt (Kt (s) , It (s)) P0,t (s) ,
with the same notation as in Chapter 2.
3.4.1.1 The value of the rm
We assume that in each period, the rm distributes all the prots it makes, and that for a given
capital K0 , it maximizes its cum-dividend value,
$
%.

m0,t D (Kt , It )
,
Vc (K0 ) = max D (K0 , I0 ) + E
(Kt ,It1 )t=1
t=1
subject to the capital accumulation in Eq. (3.24).

The value of the rm at time t, V c (Kt ), can be found recursively, through the Bellmans
equation,
Vc (Kt ) = max [D (Kt , It ) + E (mt+1 Vc (Kt+1 ))] ,
It
where the expectation is taken with respect to the information set as of time t. The rst-order
conditions for It lead to,
DI (Kt , It ) = E [mt+1 Vc (Kt+1 )] .
(3.26)
That is, along the optimal capital accumulation path, the marginal cost of new installed capital
at time t, DI , must equal the expected marginal return on the investment, i.e. the expected
value of the marginal contribution of capital to the value of the rm at time t + 1, Vc (Kt+1 ).
By Eq. (3.26), optimal investment is a function I (Kt ), and the value of the rm satises,
Vc (Kt ) = D (Kt , I (Kt )) + E [mt+1 Vc (Kt+1 )] .
Dierentiating the value function in the previous equation, with respect to Kt , and using Eq.
(3.26), yields the following envelope condition:
Vc (Kt ) = DK (Kt , I (Kt )) + DI (Kt , I (Kt )) I (Kt ) + E [mt+1 Vc (Kt+1 ) ((1 ) + I (Kt ))]
= DK (Kt , I (Kt )) (1 ) DI (Kt , I (Kt )) .
79
c
by
A. Mele

By replacing this expression for the value function back into Eq. (3.26), leaves:
DI (Kt , I (Kt )) = E [mt+1 (DK (Kt+1 , I (Kt+1 )) (1 ) DI (Kt+1 , I (Kt+1 )))] .
(3.27)
Along the optimal capital accumulation path, the marginal cost of new installed capital
at time t, which by Eq. (3.26) is the expected marginal return on the investment, equals the
expected value of (i) the very same marginal cost at time t+1, corrected for capital depreciation,
(1 ), and (ii) capital productivity, net of adjustment costs. Analytically,

It
Kt ,
DK (Kt , I (Kt )) yK (Kt , N (Kt )) wt N (Kt )
Kt
Kt

It
DI (Kt , I (Kt )) pt +
.
Kt
We now proceed to introduce a fundamental concept in investment theory.
3.4.1.2 q theory
The Tobins marginal q is dened as the ratio of the expected marginal value of an additional
unit of capital over its replacement cost:
TQt Tobins marginal q
E [mt+1 Vc (Kt+1 )]
.
pt
It is easy to see that the numerator, E [mt+1 Vc (Kt+1 )], is simply the shadow price of installed
capital. Consider the Lagrangian at time t,
L (Kt ) = max [D (Kt , It ) qt (Kt+1 (1 ) Kt It ) + E (mt+1 L (Kt+1 ))] ,
It ,Kt+1 ,qt
(3.28)
which, integrated, gives rise to the value of the rm:

$
%.

L (K0 ) =
max V c (K0 ) E
m0,t qt (Kt+1 (1 ) Kt It )
.
(It ,Kt+1 qt )t=0
t=0
The rst-order condition for investment, It , is, qt = DI (Kt , It ), and that for capital, Kt+1 ,
is qt = E (mt+1 L (Kt+1 )). By Eq. (3.26), then, L (Kt ) = Vc (Kt+1 ) and, therefore, qt is the
expected marginal return on the investment, that is, the shadow price of installed capital.
Therefore, Tobins marginal q is the ratio of the shadow price of installed capital to its replacement cost:
qt
TQt = .
pt
Next, replace the rst-order condition for qt , i.e. Eq. (3.24), into Eq. (3.28), dierentiate L (Kt )
with respect to Kt , and use the rst-order condition for Kt+1 , obtaining, L (Kt ) = DK (Kt , It )+
qt (1 ). These conditions imply that qt satises the valuation equation (3.27):
qt = E [mt+1 (DK (Kt+1 , It+1 ) + (1 ) qt+1 )] ,
and therefore that:
qt = pt +
80
It
Kt
(3.29)
(3.30)
c
by
A. Mele
The shadow price of installed capital, qt , has to equal the marginal cost of new installed capital,
and is larger than the price of uninstalled capital, pt . It is natural: to install new capital requires
some (marginal) adjustment costs, which add to the row price of uninstalled capital, pt .
Therefore, in the presence of adjustment costs, Tobins marginal q is larger than one.
Eq. (3.29) can be solved forward, leaving:
-
.

qt = E
(1 )s1 m0,t+s DK (Kt+s , It+s ) .
s=1
The shadow price of installed capital is worth the sum of all its future marginal net productivity,
discounted at the depreciation rate. Moreover, Eq. (3.30) can be inverted for It /Kt , to deliver:
It
= 1 (qt pt ) ,
Kt
(3.31)
where 1 denotes the inverse of , and is increasing, since is increasing. Given Kt , and the
fact that Kt+1 is predetermined, the rm evaluates qt through Eq. (3.29), and then determines
the level of new investments through Eq. (3.31). These investments are increasing in the difference between the shadow price of installed capital, qt , and that of uninstalled capital, pt , as
originally assumed by Tobin (1969).
In the absence of adjustement costs, when qt = pt , Eq. (3.29) delivers the usual condition,
1 = E [mt (
yK (Kt , N (Kt )) + (1 ))| Ft1 ] .
Empirically, however, the marginal productivity of capital, yK (Kt , N (Kt )), is not volatile
enough, to rationalize asset returns. Moreover, as we argue in a moment, Tobins marginal
q can be approximated by market-to-book ratios, which are typically time-varying. Therefore,
adjustment costs are important for asset pricing.
A diculty with Tobins marginal q is that it is quite dicult to estimate. Yet in the special
case we are analyzing in this section, where rms act competitively and have access to an
homogeneous production function and adjustment costs, Tobins marginal q can be proxied by
the market-to-book ratio of a given rm. Let V (Kt ) denote the ex-dividend value of the rm,
which is its stock market value, since it nets out the dividend it pays to its holder in the current
period. It is:
V (Kt ) Vc (Kt ) D (Kt , I (Kt )) = E [mt+1 Vc (Kt+1 )] .
The Tobins average q is dened as the ratio of the stock market value of the rm over the
replacement cost of the capital:
Tobins average q
Stock Mkt Value of the Firm

V (Kt )
=
.
Replacement Cost of Capital
pt Kt+1
The next result was originally obtained by Hayashi (1982) in a continuous-time setting.
Theorem 3.2. Tobins marginal q and average q coincide. That is, we have,
V (Kt ) = qt Kt+1 .
Proof. By the homogeneity properties of the production function and the adjustment costs,
D (Kt , It ) = DK (Kt , It ) Kt + DI (Kt , It ) It .
81
c
by
A. Mele

Therefore, the ex-dividend value of the rm is:
-
.

V (K0 , I0 ) = E
m0,t D (Kt , It )
=E
- t=1

t=1
m0,t (DK (Kt , It ) (1 ) DI (Kt , It )) Kt + E
-
m0,t DI (Kt , It ) Kt+1 ,
t=1
where the second line follows by Eq. (3.24). By Eq. (3.27), and the law of iterated expectations,
-
.
-
.

E
m0,t (DK (Kt , It ) (1 ) DI (Kt , It )) Kt = DI (K0 , I0 ) K1 E
m0,t Kt+1 DI (Kt , It ) .
t=1
t=1
Hence, V (K0 , I0 ) = DI (K0 , I0 ) K1 = q0 K1 .

This result, in conjunction with that in Eq. (3.30), provides a simple rule of thumb for
investement decisions. Consider, for example, the case of quadratic adjustment costs, where
(x) = 12 1 x2 , for some > 0. Then, Eq. (3.31) is:

Stock Mkt Value of the Firm
1 pt Kt ,
It = (qt pt ) Kt =
Replacement Cost of Capital
where the second equality follows by Theorem 3.2. Thus, according to q theory, we expect rms
with a market value larger than the cost of reproducing their capital to grow, and rms which
are not worth the cost of reproducing their capital to shrink. This basic observation constitutes
a rst assessment that we can use to assess developments of rms future.
3.4.2
Consumers
We now generalize the budget constraint obtained in the program [3.P3], to the uncertainty
case. We claim that in this case, the relevant budget constraint is,
-
.

V0 = c0 + E
m0,t (ct wt Nt ) .
(3.32)
t=1
We have: ct + St t+1 = (St + Dt ) t + wt Nt and, then:

-
.
-
.
-
.

E
m0,t (ct wt Nt ) = E
m0,t (St + Dt ) t E
m0,t St t+1
t=1
- t=1
t=1
-
.
.

m0,t
=E
E
mt1,t (St + Dt ) t
E
m0,t1 St1 t
t1 mt1,t
- t=1
.
-
. t=2

m0,t1 St1 t E
m0,t1 St1 t
=E

t=1
t=2
= S0 1 = V0 c0 .
where the third line follows by the properties of the discount factor,
mt1,t .
82
m0,t
mt1,t
= m0,t1 and mt
3.5. Money, production, asset prices, and overlapping generations models
c
by
A. Mele
Therefore, the program consumers solve is:

-
.

max
E
t u (ct ) , s.t. Eq. (3.32).
(ct )t=0
t=1
We now have two optimality conditions, one intertemporal and another, intratemporal:
mt+1 =
3.4.3
u1 (ct+1 , Nt+1 )
u2 (ct , Nt )
(intertemporal); wt =
(intratemporal).
u1 (ct , Nt )
u1 (ct , Nt )
Equilibrium
For all t,
y (Kt , Nt ) = ct + pt It +
It
Kt
Kt .
(3.33)
It is easily seen that the condition t = 1 in the nancial market, implies that ct = Dt + wt Nt ,
which, upon substitution of the prots in Eq. (3.25), delivers the equilibrium condition in Eq.
(3.33). Implicit in this reasoning, is the idea the adjustment costs are not paid to anyone. They
represent, so to speak, capital losses incurred along the way of growth.
3.5 Money, production, asset prices, and overlapping generations models

3.5.1
Introduction: endowment economies
3.5.1.1 A deterministic model
We initially assume the population is constant, and made up of one young and one old. The
young agent maximizes his intertemporal utility subject to his budget constraint:
+
savt + c1t = w1t
[3.P5]
max [u (c1t ) + u (c2,t+1 )] subject to
c2,t+1 = savt Rt+1 + w2,t+1
(c1t ,c2,t+1 )
where w1t and w2,t+1 are the endowments the agent receives at his young and old age.
The agent born at time t 1, then, faces the constraints: savt1 + c1,t1 = w1,t1 and c2t =
savt1 Rt + w2t . By combining his second period constraint with the rst period constraint of
the agent born at time t,
savt1 Rt + wt = savt + c1t + c2t ,
wt w1t + w2t .
(3.34)
The equilibrium in the intergenerational lending market is, naturally:

savt = 0,
(3.35)

and implies that the goods market is also in equilibrium, in that wt = 2i=1 ci,t , and for all t.
Therefore, we can analyze the model, by just analyzing the autarkic equilibrium.
As Figure 3.4 illustrates, the rst-order condition for the program [3.P5] requires that the
slope of the indierence curve be equal to the slope of the lifetime budget constraint, c2,t+1 =
Rt+1 c1,t + Rt+1 w1t + w2,t+1 , and leads to:
1
u (c2,t+1 )
=
.
u (c1,t )
Rt+1
83
(3.36)
c
by
A. Mele
c2,t+1
w2,t+1
c2,t+1 = Rt+1 c1,t + Rt+1 w1t + w2,t+1
c1,t
w1,t
FIGURE 3.4.
The equilibrium, then, is a sequence of gross returns Rt satisfying Eqs. (3.34), (3.35) and (3.36),
or:
1
u (w2,t+1 )
.
(3.37)
bt
=
Rt+1
u (w1t )
In this relation, bt is the shadow price of a bond issued at t, and promising one unit of numeraire
at t + 1: the sequence of prices, bt , satisfying Eq. (3.37), is such that agents are happy with not
being able to lend and borrow, intergenerationally.
The previous model is easy to extend to the case where agents are heterogeneous. The program
each agent j solves is, now:
+

savj,t + c1j,t = w1j,t
max
uj (c1j,t ) + j uj (c2j,t+1 )
subject to
c2j,t+1 = savj,t Rt+1 + w2j,t+1
(c1j,t ,c2j,t+1 )
with obvious notation. The rst-order condition is, for all time t and agent j,
j
uj (c2j,t+1 )
1
=
bt ,
uj (c1j,t )
Rt+1
and the equilibrium is a sequence of bond prices bt satisfying the previous relation and the
equilibrium in the intrageneration lending market:
J
savj,t = 0,
(3.38)
j=1
where J denotes the constant number of agents in each generation.

To illustrate, suppose agents have all the same utility, uj (c) = ln c and discount rate, j = .
Then,
c1j,t
1
=
1+

w2j,t+1
1
w2j,t+1
w1j,t +
, c2j,t+1 =
(Rt+1 w1j,t + w2j,t+1 ) , savj,t =
w1j,t
,
Rt+1
1+
1+
Rt+1
and using the equilibrium condition in Eq. (3.38),

Jj=1 w1j,t
1
bt =
= J
.
Rt+1
j=1 w2j,t+1
84
(3.39)
c
by
A. Mele
3.5.1.2 A tree in a stochastic economy
Suppose, next, that we introduce a tree, which yields a stochastic dividend Dt in each period.
Each agent solves the following program:
+
St t + c1t = w1t
max [u (c1t ) + E ( u(c2,t+1 )| Ft )] subject to
[3.P6]
c2,t+1 = (St+1 + Dt+1 )t + w2,t+1
(c1t ,c2,t+1 )
where St denotes the asset price and the units of the asset the agent chooses in his young age.
The agent born at time t 1 faces the constraints St1 t1 + c1,t1 = w1,t1 and w2t + (St +
Dt )t1 = c2,t . By combining the second period constraint of the agent born at time t 1 with
the rst period constraint of the agent born at time t,
(St + Dt ) t1 St t + wt = c1,t + c2,t .
The clearing condition in the asset market, t = 1, implies that the market for goods also clears,
for all t: Dt + w1t + w2t = c1,t + c2,t . A characterization of the solution to the program [3.P6]
can be obtained by eliminating c from the constraint,
max [u (w1t St ) + E ( u ((St+1 + Dt+1 ) )| Ft )] .
The equilibrium is one where t = 1, implying that (i) c1t = w1t St and (ii) c2,t+1 = St+1 +
Dt+1 + w2,t+1 . Using (i) and (ii), the rst-order condition for the program [3.P6] leads to:
u (w1t St ) St = E [u (St+1 + Dt+1 + w2,t+1 ) (St+1 + Dt+1 )| Ft ] .
t+1 = (St+1 + Dt+1 ) /St . We
Consider, for example, the case where u (c) = ln c, and set R
have:
.

1
1

= E
Rt+1 Ft , where savt St t , t = 1.
(3.40)

w1t savt
savt Rt+1 + w2,t+1
In a deterministic setting,
1
1
=
Rt+1 , where savt = 0,
w1t savt
savt Rt+1 + w2,t+1
(3.41)
which leads to the equilibrium bond price in Eq. (3.39). Eqs. (3.40) and (3.41) are formally
equivalent. Their fundamental dierence is that in the tree economy, savings have to stay
positive, as the tree must be held by the young agent, in equilibrium: savt St 0. In an
economy without a tree, instead, the interest rate, Rt , has to be such that savings are zero for
all t, savt = 0.
Eq. (3.40) can be solved explicitly for the price of the tree, St , once we assume w2t = 0 for
all t. In the absence of a tree, we cannot assume endowments are zero in the old age, since
the autarkic economy in this case would be such that the old generation would not consume
anything. In the presence of a tree, instead, this assumption is innocuous, conceptually, as the
autarkic equilibrium in this case is such that the old generation could consume the fruits of the
tree, as well as the proceedings arising from selling the tree to the young generation. Solving
Eq. (3.40) for St when w2t = 0, then, leads to a price for the tree, equal to:
St =
w1t .
1+
85

3.5.2
Diamonds model
3.5.3
Money
c
by
A. Mele
We consider a version of the previous model with endowment (not with capital), and assume
that agents can now transfer value through a piece of paper, interpreted as money. The young
agent, then, maximizes his intertemporal utility, subject to a new budget constraint:
mt
+ c1t = w1t
pt
max [u (c1t ) + u (c2,t+1 )] subject to
[3.P7]
mt
(c1t ,c2,t+1 )
+ w2,t+1
c2,t+1 =
pt+1
where mt is the amount of money he holds at time t, and pt is the price of the consumption
good as of time t.
Let
mt
pt
savt
, Rt+1
.
(3.42)
pt
pt+1
Then, the budget constraint for program [3.P7] is formally identical to that for program [3.P5].
The dierence is that in the monetary economy of this section, the young agent may wish to
transfer value over time, by saving money, earning a gross interest rate equal to the rate
of deation: the lower the price level the next period, the higher the purchasing power of the
money he transfers from the young to the old age. Naturally, then, by aggregating the budget
constraints of the young and the old generation, we obtain, formally, Eq. (3.34), where now,
savt and Rt+1 are as in (3.42). However, in the setting of this section, savt is not necessarily
zero, as money can be transferred from a generation to another one. In equilibrium, savt = mptt ,
where m
t denotes money supply. Therefore, the real value of money is strictly positive, if the
equilibrium price pt stays bounded over time, which might actually occur, as we shall study
below. As we see, the role of money as a medium for transferring value, is, in this context,
similar to that of a tree in the stochastic overlapping generations economy of Section 3.5.1.2.
pt
Substituting the equilibrium savings savt = mptt and Rt+1 = pt+1
into Eq. (3.34), we obtain,
m
t1 = m
t + pt (c1t + c2t wt ), which used again in Eq. (3.34), delivers,
savt1 Rt = savt
m
t
.
pt
We need a law of movement for money creation. We assume that:7

sequence t . Replacing this into Eq. (3.43), leaves:
(1 + t ) savt1 Rt = savt .
(3.43)
m
t
m
t1
= t , for some bounded

(3.44)
t1 pt1
The last relation can be obtained even more simply, noting that by denition, (1 + t ) mpt1
=
pt
m
t
. The previous relation can be generalized when population grows. Suppose that at time t,
pt
t
Nt individuals are born, and that NNt1
= (1 + n), for some constant n. Let money supply be
Mt
given by Mt Nt m
t , and assume that for all t, M
= t . Then, by a reasoning similar to that
t1
leading to Eq. (3.44),
1 + t
sav (Rt ) Rt = sav (Rt+1 ) ,
(3.45)
1+n
7 In this section, we assume that money transfers are made to the young generation: the money the young generation has to
absorb is that from the old generation, m
t1 , and that created by the central bank, t m
t1 . One might consider an alternative
model in which transfers are made to old.
86
c
by
A. Mele
where now, we have set the real savings equal to a function of the interest rate, savt1 sav (Rt ),
as it should be, by the solution to the program [3.P7].
Next, suppose that t is independent of R, and that limt t = , say, a constant. Eq.
(3.45) leads to two stationary equilibria:
(a) R = 1+n
. This stationary equilibrium relates to the golden rule, once we set = 0, as
1+
we shall
= 0, the price is, in this stationary equilibrium,
1+say
t in the next section. For
m
t
t
pt = 1+n p0 . Then, we have: (i) pt = NMt ptt = NM0 p00 , and (ii) pm
= NM0 p00 1+n
. All in all, the
1+
t+1
agents budget constraints are bounded and the real value of money is strictly positive.
In this stationary equilibrium, agents trust money.
(b) Ra : sav (Ra ) = 0. This stationary equilibrium relates to an autarkic state. Generally, we
have that Ra < R: prices increase more rapidly than per-capita money stocks. Analytit+1 /Mt
t+1
cally, Ra < R pt+1
> 1+
= M
= mmt+1
mpt+1
< mptt , whence limt mptt 0.
pt
1+n
Nt+1 /Nt
t
t
t
As for pm
, we have that pm
= mptt Ra < mptt R = mptt 1+n
, and since limt mptt 0, then
1+
t+1
t+1
t
0. In this stationary equilibrium, agents do not trust money.
limt pm
t+1
If sav() is dierentiable and sav () = 0, the dynamics of (Rt )
t=0 can be studied through the
slope,
dRt+1
sav (Rt )Rt + sav(Rt ) 1 + t
.
(3.46)
=
dRt
sav (Rt+1 )
1+n
There are three cases:
(i) sav (R) > 0. Gross substituability: the income eect is dominated by the substitution
eect.
(ii) sav (R) = 0. Income and substitution eects compensate each other.
(iii) sav (R) < 0. Complementarity: the income eect dominates the substitution eect.
An example of gross substituability was provided during the presentation of the introductory
examples of the present section (log utility functions). The second case can be obtained with
the same examples after imposing that agents have no endowments in the second period. The
equilibrium is seriously compromised in this case, however. Another example is obtained with
2
Cobb-Douglas utility functions: u(c1t , c2,t+1 ) = cl1t1 cl2,t+1
, which generates a real savings function
sav(Rt+1 ) =
w
l2
w R2,t+1
l1 1t
t+1
l
1+ l2
1
which also implies
m
t
pt
the derivative of which is nil when one assumes that w2,t = 0 for all t,
= savt = 1 w1t ,
l1 +l2
l2
and, by reorganizing,
m
t = pt w1t ,
an equation supporting the view of the Quantitative Theory of money. In this case, the sequence
pt
t w1,t+1
of gross returns is Rt+1 = pt+1
= mmt+1
, or
w1,t
Rt+1 =
(1 + n) (1 + gt+1 )
,
1 + t+1
where gt+1 denotes the growth rate of endowments of young between time t and time t + 1. The
ination factor R1
is equal to the monetary creation factor corrected for the the growth rate
t
of the economy.
87
c
by
A. Mele
y
1.5
1.0
0.5
0.0
0.0
0.2
0.4
0.6
0.8
1.0
1.2
FIGURE 3.5. = 2
Another example is u(c1t , c2,t+1 )
(1)/
lc1t
1l
, Cobb-Douglas. We have
lim1 u(c1t , c2,t+1 ) = cl1t c2,t+1
=
c1t
c2,t+1 =
sav
=
t
(1)/
+ (1 l)c2,t+1
Rt+1 w1t +w2,t+1

, K
Rt+1 +K Rt+1
Rt+1 w1t +w2,t+1
1+K R1
t+1
K Rt+1
w1t w2,t+1
Rt+1 +K Rt+1
/(1)
. Note that
1l
l
To simplify, suppose that K = 1, and 0 = w2t = t = n, and w1t = w1,t+1 t. It is easily

checked that
sign (sav (R)) = sign ( 1) .
The interest factor dynamics is:
1
Rt+1 = (R
1)1/(1) .
t +R
(3.47)
The stationary equilibria are the solutions of

R1 = R + R1 1,
and one immediately veries the existence of the monetary state R = 1.
When > 1, the dynamics is quite simple studying: in general, one has Ra = 0 and R = 1,
and Ra is stable and R is unstable. Figure 3.5 illustrates this situation in the case = 2.
When < 1, the situation is more delicate. In this case, Ra is generically ill-dened, and
R = 1 is not necessarily stable. One can observe a dynamics converging towards R, or even the
emergence of more or less regular cycles. In this respect, it is important to examine the slope
of the slope of the map in (3.42) in correspondence with R = 1:

dRt+1
+1
.
=

dRt Rt+1 =Rt =1 1
Here are some hints concerning the general case. Figure 3.6 depicts the shape of the map
Rt Rt+1 in the case of gross substituability (in fact, the following arguments can also be
(x)x
adapted verbatim to the complementarity case whenever x, sav
< 1: indeed, in this case
sav(x)
dRt+1
dRt
> 0 since the numerator is negative and the denominator is also negative by assumption.
88
c
by
A. Mele
Rt+1
A
Ra
Rt
FIGURE 3.6. Gross substituability
Such a case does not have any signicative economic content, however). This is an increasing
sav (Rt )Rt +sav(Rt ) 1+t
t+1
function since the slope dR
=
> 0. In addition, the slope (3.41) computed
dRt
sav (Rt+1 )
1+n
1+n
in correspondence with the monetary state R = 1+ is:

sav(R)
dRt+1
=
1
+
.
dRt Rt+1 =Rt =R
Rsav (R)
This is always greater than 1 if sav (.) > 0, and in this case the monetary state M is unstable
and the autarchic state A is stable. In particular, all paths starting from the right of M are
unstables. They imply an increasing sequence of R, i.e. a decreasing sequence of p. This can
not be an equilibrium because it contradicts the budget constraints (in fact, there would not
be a solution to the agents programs). It is necessary that the economy starts from A and M ,
but we have not endowed with additional pieces of information: there is a continuum of points
R1 [Ra , R) which are candidates for the beginning of the equilibria sequence. Contrary to
the models of the previous sections, here we have indeterminacy of the equilibrium, which is
parametrized by p0 .
Is the autarchic state the only possible stable conguration?
The answer is no. It is sucient

dRt+1
that the map Rt Rt+1 bends backwards and that dRt
< 1 to make M a stable
state. A condition

dRt+1
<
dRt
Rt+1 =Rt =R
Rt+1 =Rt =R
(R)R
for the curve to bend backward is that sav
> 1, and
sav(R)
(R)R
sav (R)R
sav
1 to hold is that sav(R) > 12 . If sav(R) > 12 , M is
the condition for
attained starting
from a suciently small neighborhood of M . Figure 3.7 shows the emergence of a cycle of
R2 9
order two,8 in which R = R
. Notice that in this case, the dynamics of the system has been
analyzed in a backward-looking manner, not in a forward-looking manner. The reason is that

there is an indeterminacy of the forward-looking dynamics, and it is thus necessary to analyze
8 There are more complicated situations in which cycles of order 3 may exist, whence the emergence of what is known as chaotic
trajectories.
9 Here is the proof. Starting from relation (3.40), we have that for a 2-cycle,
1+
R s(R ) = s(R )
1+n
1+
R s(R ) = s(R )
1+n
By multiplying the two sides of these equations, one recovers the desired result.
89
c
by
A. Mele
Rt+1
A
R*
R**
Rt
FIGURE 3.7.
the system dynamics in a backward-looking manner ... In any case, the condition sav (R) < 0
is not appealing on an economic standpoint.
3.5.4
Money in a model with real shocks
The Lucas (1972) model10 is the rst to address issues concerning the neutrality of money in a
context of overlapping generations with stochastic features. Here we present a very simplied
version of the model.11
The agent works during the rst period and consumes in the second period: this is, of course,
nothing but a pleasant simplication. Disutility of work is v. We suppose that v (x) > 0,
v (x) > 0, x R+ . The second period consumption utility is u, and has the usual properties.
The program is:
max+
{n,c} {v(nt ) + E [u(ct+1 )| Ft ]}
m = pt yt , yt = t nt
s.t.
pt+1 ct+1 = m
where m denotes money, y denotes the production obtained by means of labour n, and {t }t=0,1,
is a sequence of strictly positive shocks aecting labour productivity. The price level is p.
pt
Let us rewrite the program by using the single constraint ct+1 = Rt+1 t nt , Rt+1 pt+1
,
max {v(nt ) + E [u(Rt+1 t nt )| Ft )]} .
n
The rst-order condition is v (nt ) = E(u (Rt+1 t nt )Rt+1 t / Ft )

1
1
v (nt )
= E u (Rt+1 t nt )
Ft
pt t
pt+1
At the equilibrium, yt = ct ct = t nt , and

ct+1 =
m
= t+1 nt+1 .
pt+1
10 Lucas,
R.E., Jr. (1972): Expectations and the Neutrality of Money, J. Econ. Theory, 4, 103-124.
of this simplied version of the model are taken from Stokey et Lucas (1989, p. 504): Stokey, N.L. and R.E. Lucas (with
E.C. Prescott) (1989): Recursive Methods in Economic Dynamics, Harvard University Press.
11 Parts
90
c
by
A. Mele
3.6. Optimality
By replacing the previous relation into the rst-order condition, and simplifying,
v (nt )nt = E [u (t+1 nt+1 )(t+1 nt+1 )| Ft ] .
(3.48)
As in the model of section 3.6, the rational expectation assumption consists in regarding all the
models variables as functions of the state varibales. Here, the states of natures are generated
by , and we have:
n = n().
By plugging n() into (3.43) we get:
v (n())n() =
"
supp()

u + n(+ ) + n(+ )dP ( + ).
(3.49)
A rational expectations equilibrium is a statistical sequence n() satisfying the functional

equation (3.44).
Eq. (3.44) considerably simplies when shocks are i.i.d.: P (+ / ) = P (+ ),
"

v (n())n() =
u + n(+ ) + n(+ )dP (+ ).
(3.50)
supp()
In this case, the r.h.s. of the previous relationship does not depend on , which implies that
the l.h.s. does not depend on neither. Therefore, the only candidate for the solution for n is
a constant n
:12
n() = n
, .
Provided such a n
exists, this is a result on money neutrality. More precisely, relation (3.45)
can be written as:
"
n) =
u (+ n
)+ dP (+ ),
v (
supp()
and it is always possible to impose reasonable conditions on v and u that ensure existence and
unicity of a strictly positive solution for n
, as in the following example.
Example. v(x) = 12 x2 and u(x) = ln x. The solution is n
=
, y() = and p() =
Exercise. Extend the previous model when the money supply follows the stochastic process:
= t , where {t }t=0,1, is a i.i.d. sequence of shocks.
mt
mt1
3.6 Optimality
3.6.1
Models with productive capital
The starting point is the relation

Kt+1 = St = Y (Kt , Nt ) Ct , K0 given.
12 A
rigorous proof that n() = n

, is as follows. Lets suppose the contrary, i.e. there exists a point 0 and a neighborhood
of 0 such that either n (0 + A) > n (0 )or n (0 + A) < n (0 ), where the constant A > 0. Lets consider the rst case (the proof
of the second case being entirely analogous). Since the r.h.s. of (4.43) is constant for all , we have, v (n (0 + A)) n (0 + A) =
v (n (0 )) n (0 ) v (n (0 + A)) n (0 ), where the inequality is due to the assumption that v > 0 always holds. We have thus
shown that, v (n (0 + A)) [n (0 + A) n (0 )] 0. Now v > 0 always holds, so that n (0 + A) < n (0 ), a contradiction with
the assumption n (0 + A) > n (0 ).
91
c
by
A. Mele
3.6. Optimality
Dividing both sides by Nt one gets:
kt+1 =
1
(y(kt ) ct ) , k0 given and: inf kt 0, sup kt < .
1+n
(3.51)
The stationary state of this economy is:

c = y(k) (1 + n)k,
and we see that the per-capita consumption attains its maximum at:
= 1 + n.
k : y (k)
This is the golden rule.
< 1 + n, it is always possible to increase per-capita consumption at the stationary
If y (k)
state; indeed, since y(k) is given, one can decrease k and obtain dc = (1 + n)dk > 0, and in
the next periods, one has dc = (y (k) (1 + n)) dk > 0. We wish to provide a formal proof of
these facts along the entire capital accumulation path of the economy. Notice that the foregoing
results are valid whatever the structure of the economy is (e.g., a nite number of agents living
forever or overlapping generations). As an example, in the overlapping generations case one can
interpret relation (3.46) as the one describing the capital accumulation path in the Diamonds
c2,t+1
Ct
= c1t + 1+n
.
model once ct is interpreted as: ct N
t
The notion of eciency that we use is the following one:13 a path {(k, c)t }
t=0 is consumption
c)t }
inecient if there exists another path {(k,
t=0 satisfying (3.46) and such that, for all t,
ct ct , with at least one strict inequality.
We now present a weaker version of thm. 1 p. 161 of Tirole (1988), but easier to show.14
Theorem 3.2 ((weak version of the) Cass-Malinvaud theory). (a) A path {(k, c)t }
t=0 is
y (kt )
consumption efficient if 1+n 1 t. (b) A path {(k, c)t }t=0 is consumption inefficient if
y (kt )
< 1 t.
1+n
Proof. (a) Let kt = kt + t , t = 0, 1, , an alternative consumption ecient path. Since k0
is given, t = 0. Furthermore, by relation (3.46) one has that
(1 + n) (kt+1 kt+1 ) = y(kt ) y(kt ) (
ct ct ),
and because k has been supposed to be ecient, ct ct with at least one strictly equality over
the ts. Therefore, by concavity of y,
0 y(kt ) y(kt ) (1 + n)(kt+1 kt+1 )
= y(kt ) y(kt ) (1 + n)t+1
< y(kt ) + y (kt )t y(kt ) (1 + n)t+1 ,
or
t+1 <
y (kt )
t .
1+n
13 Tirole,
J. (1988): Ecacit
e intertemporelle, transferts interg
en
erationnels et formation du prix des actifs: une introduction,
in: Melanges
economiques. Essais en lhonneur de Edmond Malinvaud. Paris: Editions Economica & Editions EHESS, p. 157-185.
14 The proof we present here appears in Touz
e, V. (1999): Financement de la s
ecurit
e sociale et
equilibre entre les g
en
erations,
unpublished PhD dissertation Univ. Paris X Nanterre.
92
c
by
A. Mele
3.6. Optimality
y'(kt)
1
1
kt
k*
FIGURE 3.8. Non-necessity of the conditions of thm. 3.2 in the model with a representative agent.
(k0 )
Evaluating the previous inequality at t = 0 yields 1 < y1+n
0 , and since 0 = 0, one has that
y (kt )
1 < 0. Since 1+n 1 t, t , which contradicts (3.46).
t
(b) The proof is nearly identical to the one of part (a) with the obvious exception that
lim inf t >> here. Furthermore, note that there are innitely many such sequences that
allow for eciency improvements.
Are actual economies dynamically ecient? To address this issue, Abel et al. (1989)15 modied somehow the previous setup to include uncertainty, and conclude that the US economy
does satisfy their dynamic eciency requirements.
The conditions of the previous theorem are somehow restrictive. As an example, let us take the
model of section 3.2 and x, as in section 3.2, n = 0 to simplify. As far as k0 < k = (y )1 [ 1 ],
per-capita capital is such that y (kt ) > 1 t since the dynamics here is of the saddlepoint type
and then monotone (see gure 3.8). Therefore, the conditions of the theorem are fullled. Such
conditions also hold when k0 [k, k ], again by the monotone dynamics of kt . Nevertheless, the
conditions of the theorem do not hold anymore when k0 > k and yet, the capital accumulation
path is still ecient! While it is possible to show this with the tools of the evaluation equilibria
of Debreu (1954), here we provide the proof with the same tools used to show thm. 3.2. Indeed,
let
= inf {t : kt k } = inf {t : y (kt ) 1} .
+
y (kt ) < 1, t = 0, 1, , 1
We see that < , and since the dynamics is monotone,
. By
y (kt ) 1, t = , + 1,
using again the same arguments used to show thm. 3.2, we see that since is nite, <
+1 < 0. From onwards, an explosive sequence starts unfolding, and t .
t
3.6.2
Models with money
The decentralized economy is characterized by the presence of money. Here we are interested
in rst best optima i.e., optima that a social planner may choose by acting directly on agents
15 Abel, A.B., N.G. Mankiw, L.H. Summers and R.J. Zeckhauser (1989): Assessing Dynamic Eciency: Theory and Evidence,
Review Econ. Studies, 56, 1-20.
93
c
by
A. Mele
3.6. Optimality
consumptions.16 Let us rst analyze the stationary state R = 1+n

and show that it corresponds
1+
to the stationary state in which consumptions and endowments are constants, and agents utility
is maximized when = 0. Indeed, here the social planner allocates resources without caring
about monetary phenomena, and the only constraint is the following natural constraint:
wn w1 +
w2
c2
= c1 +
,
1+n
1+n
in which case the utility of the stationary agent is:

u(c1 , c2 ) = u(wn
c2
, c2 ).
1+n
1
The rst-order condition is uuc2
= 1+n
. In the market equilibrium, the rst-order condition is
c1
uc2
1
= R , which means that in the market equilibrium, the golden rule is attained at R if and
uc1
only if = 0, as claimed in section 3.?
The convergence of the optimal policy of the social planner towards the gloden rule can be
veried as it follows. The social planner solves the program:
max t u(t) (c , c
)
1t
w w +
nt
1t
t=0
w2t
1+n
2,t+1
= c1t +
c2t
1+n

t
c2t
(t)
or max
wnt 1+n
, c2,t+1 . Here the notation u(t) (.) has been used to stress the fact
t=0 u
that endowments may change from one generation to another generation, and is the weighting
coecient of generations that is used by the planner. The rst-order conditions,
(t1)
uc2
(t)
uc1
,
1+n
lead towards the modied Golden Rule at the stationary state (modied by ).
16 In
our terminology, a second best optimum is the one in which the social planner makes the thought experiment to let the
market play rst (with money) and then parametrizes such virtual equilibria by t . The resulting indirect utility functions are
expressed in terms of such t s and after creating an aggregator of such indirect utility functions, the social planner maximises such
an aggregator with respect to t .
94
3.7. Appendix 1: Finite dierence equations, with economic applications
c
by
A. Mele
3.7 Appendix 1: Finite dierence equations, with economic applications

Let z0 Rd , and consider the following linear system of nite dierence equations:
zt+1 = A zt ,
t = 0, 1, ,
(3A1.1)
where A is conformable matrix. Solution is,

zt = v1 1 t1 + + vd d td ,
where i and vi are the eigenvalues and the corresponding eigenvectors of A, and i are constants
which will be determined below.
The classical method of proof is based on the so-called diagonalization of system (3A1.1). Let us
consider the system of characteristic equations for A, (A i I) vi = 0d1 , i scalar and vi a column
vector d 1, i = 1, , n, or in matrix form, AP = P , where P = (v1 , , vd ) and is a diagonal
matrix with i on the diagonal. By post-multiplying by P 1 one gets the decomposition17
A = P P 1 .
(3A1.2)
By replacing (3A1.2) into (3A1.1), P 1 zt+1 = P 1 zt , or

yt P 1 zt .
yt+1 = yt ,
The solution for y is:
yit = i ti ,

and the solution for z is: zt = P yt = (v1 , , vd )yt = di=1 vi yit = di=1 vi i ti .
To determine the vector of the constants = (1 , , d ) , we rst evaluate the solution at t = 0,
z0 = (v1 , , vd ) = P ,
whence
(P ) = P 1 z0 ,
where the columns of P are vectors the space of the eigenvectors. Naturally, there is an innity of
such P s, but the previous formula shows how (P ) must adjust to guarantee the stability of the
solution with respect to changes of P .
3A.1 Example. d = 2. Let us suppose that 1 (0, 1), 2 > 1. The system is unstable in correspondence with any initial condition but a set of zero measure. This set gives rise to the so-called
saddlepoint path. Let us compute its coordinates. The strategy consists in nding the set of initial
conditions for which 2 = 0. Let us evaluate the solution at t = 0,

1
v11 1 + v12 2
x0
= z0 = P = (v1 , v2 )
=
,
y0
2
v21 1 + v22 2
where we have set z = (x, y) . By replacing the second equation into the rst one and solving for 2 ,
2 =
v11 y0 v21 x0
.
v11 v22 v12 v21
This is zero when

y0 =
v21
x0 .
v11
95
c
by
A. Mele

y
y0
x
x0
y = (v21/v11) x
FIGURE 3.9.
Here the saddlepoint is a line with slope equal to the ratio of the components of the eigenvector
associated with the root with modulus less than one. The situation is represented in gure 3.9, where
x0 , and corresponds to the case 1 = 0.
the divergent line has as equation y0 = vv22
12
The economic content of the saddlepoint is the following one: if x is a predetermined variable, y
must jump to y0 = vv21
x0 to make the system display a non-explosive behavior. Notice that there is
11
a major conceptual diculty when the system includes two predetermined variables, since in this case
there are generically no stable solutions. Such a possibility is unusual in economics, however.
4A.2 Example. The previous example can be generated by the neoclassic growth model. In section
3.2.3, we showed that in a small neighborhood of the stationary values k, c, the dynamics of (kt , ct )t
(deviations of capital and consumption from their respective stationary values k, c) is:

kt+1
kt
=A
ct+1
ct
where
A
1
y (k)
u (c)
u (c)
u (c) y (k) 1 + u (c) y (k)
(0, 1).
By using the relationship y (k) = 1, and the conditions imposed on u and y, we have
- det(A) = y (k) = 1 > 1;
- tr(A) = 1 + 1 + uu(c)
(c) y (k) > 1 + det(A).
The two eigenvalues are solutions of a quadratic equation, and are: 1/2 =
tr(A)
tr(A)2 4 det(A)
.
2
Now,
a tr(A)2 4 det(A)
!2
(k)
1 + 1 + uu(c)
y
4 1
(c)
1
2
> + 1 4 1
2

= 1 1
> 0.
17 The previous decomposition is known as the spectral decomposition if P = P 1 . When it is not possible to diagonalize A,
one may make reference to the canonical transformation of Jordan.
96
c
by
A. Mele
1+det(A)
a
It follows that 2 = tr(A)
+ 2a > 1 + 2a > 1.
2 + 2 >
2
To show that 1 (0, 1), notice rst that since det(A) > 0, 21 = tr(A)
2
that 1 < 1. But 1 < 1 tr(A)
tr(A) 4 det(A) > 0. It remains 2 to be shown
tr(A)2 4 det(A) < 2, or (tr(A) 2) < tr(A)2 4 det(A), which can be conrmed to be always
true by very simple calculations.
Next, we wish to generalize the previous examples to the case d > 2. The counterpart of the
saddlepoint seen before is called the convergent, or stable subspace: it is the locus of points for which
the solution does not explode. (In the case of nonlinear systems, such a convergent subspace is termed
convergent, or stable manifold. In this appendix we only study linear systems.)
Let P 1 , and rewrite the system determining the solution for :
= z0 .
We suppose that the elements of z and matrix A have been reordered in such a way that s : |i | < 1,
for i = 1, , s and |i | > 1 for i = s + 1, , d. Then we partition in such a way that:
= sd z0 .
u
(ds)d
As in example (3A.1), the objective is to make the system stay prisoner of the convergent space,
which requires that
s+1 = =
d = 0,
or, by exploiting the previous system,
s+1
..
. = u z0 = 0(ds)1 .
(ds)d
Let d k + k (k free and k predetermined), and partition u and z0 in such a way to distinguish
the predetermined from the free variables:

z0free
(1)
(2)
(1)
u
u
k1
0(ds)1 = u z0 =
z0free + (2)
z0pre ,
pre = u
u
(ds)k
(ds)k
z
(ds)d
0
(ds)k k1
(ds)k k 1
k 1
or,
(1)
z0free = (2)
z0pre .
u
u
(ds)k k1
(ds)k k 1
The previous system has d s equations and k unknowns (the components of z0free ): this is so be(1)
(2)
cause z0pre is known (it is the k -dimensional vector of the predetermined variables) and u , u are
(1)
primitive data of the economy (they depend on A). We assume that u has full rank.
(ds)k
Therefore, there are three cases: 1) s = k ; 2) s < k ; and 3) s > k . Before analyzing these case,
let us mention a word on terminology. We shall refer to s as the dimension of the convergent subspace
(S). The reason is the following one. Consider the solution:
zt = v1
1 t1 + + vs
s ts + vs+1
s+1 ts+1 + + vd
d td .
In order to be in S, it must be the case that
s+1 = =
d = 0,
97
c
by
A. Mele
in which case the solution reduces to:

zt = v1
1 t1 + + vs
s ts = (v1
1 , , vs
s ) t1 , , ts ,
d1
i.e.,
d1
d1
zt
d1
t ,
= V
ds s1

t t , , t . Now for each t, introduce the vector subspace:
where V (v1
1 , , vs
s ) and
1
s
t ,
t Rs }.
V t {zt Rd : zt = V
ds s1
d1
Clearly, for each t, dimV t = rank(V ) = s, and we are done.

We analyze these three cases below.
(i) d s = k, or s = k . The dimension of the divergent subspace is equal to the number of the free
variables or, the dimension of the convergent subspace is equal to the number of predetermined
variables. In this case, the system is determined. The previous conditions are easy to interpret.
The predetermined variables identify one and only one point in the convergent space, which
allows us to compute the only possible jump in correspondence of which the free variables can
(1)1 (2) pre
jump to make the system remain in the convergent space: z0free = u
u z0 . This is exactly
the case of the previous examples, in which d = 2, k = 1, and the predetermined variable was
x: there x0 identied one and only one point in the saddlepoint path, and starting from such a
point, there was one and only one y0 guaranteeing that the system does not explode.
(ii) d s > k, or s < k . There are generically no solutions in the convergent space. This case was
already reminded at the end of example 4A.1.
(iii) d s < k, or s > k . There exists an innite number of solutions in the convergent space, and
such a phenomenon is typically referred to as indeterminacy. In the previous example, s = 1,
and this case may emerge only in the absence of predetermined variables. This is also the case
in which sunspots may arise.
98
c
by
A. Mele
3.8. Appendix 2: Neoclassic growth model - continuous time
3.8 Appendix 2: Neoclassic growth model - continuous time

3.8.1
Convergence results
Consider chopping time in the population growth law as,

Nhk Nh(k1) = n
h Nh(k1) ,
k = 1, , ,
where n
is an instantaneous rate, and = ht is the number of subperiods in which we have chopped a
given time period t. The solution is Nh = (1 + n
h) N0 , or
Nt = (1 + n
h) t/h N0 .
By taking limits:
N(t) = lim (1 + n
h) t/h N(0) = en t N(0).
h0
On the other hand, an exact discretization yields:

N(t ) = en (t) N(0).
N(t)
= en 1 + n .
N(t )
n
=
1
ln (1 + n ) .
E.g., = 1, n = n1 n : n
= ln (1 + n).
Now lets try to do the same thing for the capital accumulation law:

Kh(k+1) = 1 h Khk + Ih(k+1) h, k = 0, , 1,
where is an instantaneous rate.

By iterating on as in the population growth case we get:
Kh

j
= 1 h K0 +
1 h
Ihj h,
j=1
or
t/h

t/h
t/hj
Kt = 1 h
K0 +
1 h
Ihj h,
j=1
As h 0 we get:
K(t) = et K0 + et
"
eu I(u)du,
or in dierential form:
K(t)
= K(t) + I(t),
and starting from the IS equation:

Y (t) = C(t) + I(t),
we obtain the capital accumulation law:
K(t)
= Y (t) C(t) K(t).
99
t
,
h
c
by
A. Mele

Discretization issues
An exact discretization gives:
K(t + 1) = e
(t+1)
K(t) + e
"
t+1
eu I(u)du.
By identifying with the standard capital accumulation law in the discrete time setting:
Kt+1 = (1 ) Kt + It ,
we get:
= ln
It follows that
1
.
(1 )
(0, 1) > 0 and = 0 = 0.

Hence, while can take on only values on [0, 1), can take on values on the entire real line.
An important restriction arises in the continuous time model when we note that:
lim ln
1
= ,
(1 )
It is impossible to think about a maximal rate of capital depreciation in a continuous time model
because this would imply an innite depreciation rate!
Finally, substitute into the exact discretization (?.?):
" t+1
eu I(u)du
K(t + 1) = (1 ) K(t) + e(t+1)
t
so that we have to interpret investments in t + 1 as e
(t+1) t+1
t
eu I(u)du.
Per capita dynamics
Consider dividing the capital accumulation equation by N(t):
K(t)
F (K(t), N(t))
=
c(t) k(t) = y(k(t)) c(t) k(t).
N(t)
N(t)
8
By plugging the relationship k = d ( K(t)/ N(t)) = K(t)

N(t) n
k(t) into the previous equation we
get:

k = y(k(t)) c(t) + n
k(t)
This is the constraint we use in the problem of the following subsection.
3.8.2
The model
We consider directly the social planner problem. The program is:
"
max
et u(c(t))dt
c
0

s.t. k(t)
= y(k(t)) c(t) + n
k(t)
(3A2.1)
where all variables are expressed in per-capita terms. We suppose that there is no capital depreciation
(in the discrete time model, we supposed a total capital depreciation). More general results can be
obtained with just a change in notation.
100
c
by
A. Mele
The Hamiltonian is,

H(t) = u(c(t)) + (t) y(k(t)) c(t) + n
k(t) ,
where is a co-state variable.

As explained in Appendix 4 of the present chapter, the rst-order conditions for this problem are:
(t)
(t) = u (c(t))
0
=
(3A2.2)
(t) = k(t)

H (t) = (t)
+ (t)
(t)
= + + n
y (k(t)) (t)
k
By dierentiating the rst equation in (4A2.2) we get:

u (c(t))
(t)
=
c(t)
(t).
u (c(t))
By identifying with the help of the last equation in (4A2.2),
c(t)
=

u (c(t))
+ + n
y (k(t)) .
u (c(t))
(3A2.3)
The equilibrium is the solution of the system consisting of the constraint of (4A2.1), and (4A2.3).
As in section 3.2.3, here we analyze the equilibrium dynamics of the system in a small neighborhood of
the stationary state.18 Denote the stationary state as the solution (c, k) of the constraint of program
(4A2.1), and (4A2.3) when c(t)

= k(t)
= 0,

+
c
= y(k) + n
k
+ + n
= y (k)
Warning! these are instantaneous figures, so that dont worry if they are not such that
y (k) 1 + n!. A rst-order approximation of both sides of the constraint of program (4A2.1) and
(4A2.3) near (c, k) yields:
u (c)
c(t)
= y (k) (k(t) k)
u (c)

k(t) = (k(t) k) (c(t) c)
where we used the equality + + n

= y (k). By setting x(t) c(t) c and y(t) k(t) k the previous
system can be rewritten as:
z(t)
= A z(t),
(3A2.4)
where z (x, y) , and
u (c)
0
y
(k)
.
A
u (c)
1
Warning! There must be some mistake somewhere. Let us diagonalize system (4A2.4) by
setting A = P P 1 , where P and have the same meaning as in the previous appendix. We have:
(t)
= (t),
18 In addition to the theoretical results that are available in the literature, the general case can also be treated numerically with
the tools surveyed by Judd (1998).
101
c
by
A. Mele
where P 1 z.
The eigenvalues are solutions of the following quadratic equation:
0 = 2
We see that 1 < 0 < 2 , and 1
1
2
u (c)
y (k).
u (c)
(3A2.5)
2 + 4 uu(c)
(c) y (k). The solution for (t) is:
i (t) = i ei t , i = 1, 2,
whence
z(t) = P (t) = v1 1 e1 t + v2 2 e2 t ,
where the vi s are 2 1 vectors. We have,
+
x(t) = v11 1 e1 t + v12 2 e2 t
y(t) = v21 1 e1 t + v22 2 e2 t
Let us evaluate this solution in t = 0,

1
x(0)
= P = v1 v2
.
y(0)
2
By repeating verbatim the reasoning of the previous appendix,
2 = 0
y(0)
v21
=
.
x(0)
v11
As in the discrete time model, the saddlepoint path is located along a line that has as a slope
the ratio of the components of the eigenvector associated with the negative root. We can explicitely
compute such ratio. By denition, A v1 = 1 v1
u (c)
y (k) = 1 v11
u (c)
v11 + v21 = 1 v21

i.e.,
v21
v11
1
u (c)
y (k)
u (c)
and simultaneously,
v21
v11
1
1 ,
102
which can be veried with the help of (3A2.5).
c
by
A. Mele
References
Farmer, R. (1998): The Macroeconomics of Self-Fulfilling Prophecies. Boston: MIT Press.
Hayashi, F. (1982): Tobins Marginal q and Average q: A Neoclassical Interpretation. Econometrica 50, 213-224.
Kamihigashi, T. (1996): Real Business Cycles and Sunspot Fluctuations are Observationally
Equivalent. Journal of Monetary Economics 37, 105-117.
King, R. G. and S. T. Rebelo (1999): Resuscitating Real Business Cycles. In: J. B. Taylor
and M. Woodford (Editors): Handbook of Macroeconomics, Elsevier.
Lucas, R. E. Jr. (1978): Asset Prices in an Exchange Economy. Econometrica 46, 1429-1445.
Lucas, R. E. Jr. (1994): Money and Macroeconomics. In: General Equilibrium 40th Anniversary Conference, CORE DP no. 9482, 184-187.
Prescott, E. (1991): Real Business Cycle Theory: What Have We Learned? Revista de Analisis Economico 6, 3-19.
Tobin, J. (1969): A General Equilibrium Approach to Monetary Policy. Journal of Money,
Credit and Banking 1, 15-29.
Watson, M. (1993): Measures of Fit for Calibrated Models. Journal of Political Economy
101, 1011-1041.
103
4
Continuous time models
4.1 Lambdas and betas in continuous time

4.1.1
The pricing equation
Let St be the price of a long lived asset as of time t, and let Dt the dividend paid by the asset
at time t. In the previous chapters, we learned that in the absence of arbitrage opportunities,
St = Et [mt+1 (St+1 + Dt+1 )] ,
(4.1)
where Et is the conditional expectation given the information set at time t, and mt+1 is the
usual stochastic discount factor.
Next, let us introduce the pricing kernel, or state-price process,
t+1 = mt+1 t ,
0 = 1.
In terms of the pricing kernel , Eq. (4.1) is,

0 = Et t+1 St+1 St t + Et t+1 Dt+1 .
For small trading periods h, this is,

0 = Et t+h St+h St t + Et t+h Dt+h h .
As h 0,
0 = Et [d ( (t) S (t))] + (t) D (t) dt.
(4.2)
Eq. (4.2) can now be integrated to yield,

" T

(t) S (t) = Et
(u) D (u) du + Et [ (T ) S (T )] .
t
Finally, let us assume that limT Et [ (T ) S (T )] = 0. Then, provided it exists, the price St
of an innitely lived asset price satises,
"

(u) D (u) du .
(4.3)
(t) S (t) = Et
t
c
by
A. Mele
4.1. Lambdas and betas in continuous time

4.1.2
Expected returns
Let us elaborate on Eq. (4.2). We have,

d (S) = Sd + dS + ddS = S
d dS d dS
+
+
S
S
By replacing this expansion into Eq. (4.2) we obtain,

dS
D
d
d dS
Et
+ dt = Et
Et
.
S
S
(4.4)
This evaluation equation holds for any asset and, hence,

for the assets that do not distribute
dS0
dividends and are locally riskless, i.e. D = 0 and Et d
= 0, where S0 (t) is the price of
S0
0 (t)
these locally riskless assets, supposed to satisfy dS
= r (t) dt, for some short term rate process
S0 (t)
rt . By Eq. (4.4), then,

d
Et
= r (t) dt.
By
this into eq. (4.4) leaves the following representation for the expected returns
dSreplacing
D
Et S + S dt,

dS
D
d dS
Et
+ dt = rdt Et
.
(4.5)
S
S
S
In a diusion setting, Eq. (4.5) gives rise to a partial dierential equation. Moreover, in a
diusion setting,
d
= rdt dW,
where W is a vector Brownian motion, and is the vector of unit risk-premia. Naturally, the
price of the asset, S, is driven by the same Brownian motions driving r and . We have,

d dS
dS
Et
= Vol
dt,
S
S
which leaves,
Et
dS
S

D
dS
+ dt = rdt + Vol

dt.
S
S
lambdas
betas
4.1.3
Expected returns and risk-adjusted discount rates
The dierence between expected returns and risk-adjusted discount rates is subtle. If dividends
and asset prices are driven by only one factor, expected returns and risk-adjusted discount rates
are the same. Otherwise, we have to make a distinction. To illustrate the issue, let us make a
simplication. Suppose that the price of the asset S takes the following form,
S (y, D) = p (y) D,
(4.6)
where y is a vector of state variables that are suggested by economic theory. In other words, we
assume that the price-dividend ratio p is independent of the dividends D. Indeed, this scaleinvariant property of asset prices arises in many model economies, as we shall discuss in detail
105
4.2. An introduction to arbitrage and equilibrium in continuous time models
c
by
A. Mele
in the second part of these lectures. By Eq. (4.6),

dS
dp dD
=
+
.
S
p
D
By replacing the previous expansion into Eq. (4.5) leaves,

dS
D
dp d
Et
+ dt = Disc dt Et
.
S
S
p
(4.7)
where we dene the risk adjusted discount rate, Disc, to be:

9
dD d
Disc = r Et
dt.
D
If the price-dividend ratio, p, is constant, the risk adjusted discount rate Disc has
the usual
dD d
interpretation. It equals the safe interest rate r, plus the premium Et D that arises
to compensate the agents for the uctuations of the uncertain ow of future dividends. This
premium equals,

dD d
dD
Et
= Vol

dt.
D
D
lambdas
cash-ow betas
In this case, expected returns and risk-adjusted discount rates are the same thing, as in the
simple one-factor Lucas economy of Section 2.
If, instead, the price-dividend ratio is not constant, the last term in Eq. (4.7) introduces a
wedge between expected returns and risk-adjusted discount rates. As we shall see, the riskadjusted discount rates play an important role in explaining returns volatility, i.e. the beta
related to the uctuations of the price-dividend ratio. Intuitively, this is because risk-adjusted
discount rates aect prices through rational evaluation and, hence, price-dividend ratios and
price-dividend ratios volatility. To illustrate these properties, note that Eq. (4.3) can be rewritten as,

"

D ( ) Disc(y(u))du
p (y (t)) = Et
e t
(4.8)
y (t) ,
D (t)
t
where the expectation is taken under the risk-neutral probability, but the expected dividend
( )
( )
growth DD(t)
is not risk-adjusted (that is E( DD(t)
) = eg0 ( t) ). Eq. (4.8) reveals that risk-adjusted
discount rates play an important role in shaping the price function p and, hence, the volatility
of the price-dividend ratio p. These points are developed in detail in Chapter 7.
4.2 An introduction to arbitrage and equilibrium in continuous time models

4.2.1
A reduced-form economy
Let us consider the Lucas (1978) model with one tree and a perishable good taken as the
numeraire. The Appendix shows that in continuous time, the wealth process of a representative
agent at time , V ( ), is solution to,

dS( ) D( )
+
d rd ( ) + rV d c( )d ,
(4.9)
dV ( ) =
S( )
S( )
106
c
by
A. Mele
where D ( ) is the dividend process, S(1) , and (1) is the number of trees in the portfolio
of the representative agent.
We assume that the dividend process, D ( ), is solution to the following stochastic dierential
equation,
dD
= D d + D dW,
D
for two positive constants D and D . Under rational expectation, the price function S is such
that S = S(D). By Itos lemma,
dS
= S d + S dW,
S
where
DS (D) + 12 2D D2 S (D)
D DS (D)
S = D
; S =
.
S(D)
S(D)
Then, by Eq. (4.9), the value of wealth satises,

D
dV = S + r + rV c d + S dW.
S
Below, we shall show that in the absence of arbitrage, there must be some process , the unit
risk-premium, such that,
D
S + r = S .
(4.10)
S
Let us assume that the short-term rate, r, and the risk-premium, , are both constant. Below,
we shall show that such an assumption is compatible with a general equilibrium economy. By
the denition of S and S , Eq. (4.10) can be written as,
1
0 = 2D D2 S (D) + (D D ) DS (D) rS (D) + D.
2
(4.11)
Eq. (4.11) is a second order dierential equation. Its solution, provided it exists, is the the
rational price of the asset. To solve Eq. (4.11), we initially assume that the solution, SF say,
tales the following simple form,
(4.12)
SF (D) = K D,
where K is a constant to be determined. Next, we verify that this is indeed one solution to
Eq. (4.11). Indeed, if Eq. (4.12) holds, then, by plugging this guess and its derivatives into Eq.
(4.11) leaves, K = (r D + D )1 and, hence,
SF (D) =
1
D.
r + D D
(4.13)
This is a Gordon-type formula. It merely states that prices are risk-adjusted expectations of
future expected dividends, where the risk-adjusted discount rate is given by r + D . Hence,
in a comparative statics sense, stock prices are inversely related to the risk-premium, a quite
intuitive conclusion.
Eq. (4.13) can be thought to be the Feynman-Kac representation to Eq. (4.11), viz
"

r( t)
SF (D (t)) = Et
e
D ( ) d ,
(4.14)
t
107
c
by
A. Mele
where Et [] is the conditional expectation taken under the risk neutral probability Q (say), the
dividend process follows,
dD
,
= (D D ) d + D dW
D
( ) = W ( )+ ( t) is a another standard Brownian motion dened under Q. Formally,
and W
the true probability, P , and the risk-neutral probability, Q, are tied up by the Radon-Nikodym
derivative,
1 2
dQ
(4.15)
=
= e(W ( )W (t)) 2 ( t) .
dP
4.2.2
Preferences and equilibrium
The previous results do not let us see how precisely preferences aect asset prices. In Eq. (4.13),
the asset price is related to the interest rate, r, and the risk-premium, . In equilibrium, the
agents preferences aect the interest rate and the risk-premium. However, such an impact can
have a non-linear pattern. For example, when the risk-aversion is low, a small change of riskaversion can make the interest rate and the risk-premium change in the same direction. If the
risk-aversion is high, the eects may be dierent, as the interest rate reects a variety of factors,
including precautionary motives.
To illustrate these points, let us rewrite, rst, Eq. (4.9) under the risk-neutral probability Q.
We have,
.
dV = (rV c) d + S dW
(4.16)
We assume that the following transversality condition holds,

lim Et er( t) V ( ) = 0.
(4.17)
By integrating Eq. (4.16), and using the previous transversality condition,

"

r( t)
V (t) = Et
e
c ( ) d .
(4.18)
By comparing Eq. (4.14) with Eq. (4.18) reveals that the equilibrium in the real markets, D = c,
also implies that S = V . Next, rewrite (4.18) as,
"
"

r( t)
V (t) = Et
e
c(t)d = Et
mt ( )c(t)d ,
t
where
1 2
( )
= e(r+ 2 )( t)(W ( )W (t)) .
(t)
We assume that a representative agent solves the following intertemporal optimization problem,
"

"

mt ( )
max Et
c
e( t) u (c( )) d
s.t. V (t) = Et
mt ( )c( )d
for some instantaneous utility function u (c) and some subjective discount rate .
To solve the program [P1], we form the Lagrangean
"

"

( t)
L = Et
e
u(c( ))d + V (t) Et
mt ( )c( )d
,
t
108
[P1]
c
by
A. Mele
where is a Lagrange multiplier. The rst order conditions are,

e( t) u (c( )) = mt ( ).
Moreover, by the equilibrium condition, c = D, and the denition of mt ( ),
u (D ( )) = e(r+ 2
1
)( t)(W ( )W (t))
(4.19)
That is, by Itos lemma,

u (D)D
1 2 2 u (D)
u (D)D
du (D)
+
D
=
d
+
D dW.
u (D)
u (D) D 2 D
u (D)
u (D)
Next, let us dene the right hand side of Eq. (A14) as U ( ) e(r+ 2
By Itos lemma, again,
dU
= ( r) d dW.
U
1
(4.20)
)( t)(W ( )W (t)) .
(4.21)
By Eq. (A14), drift and volatility components of Eq. (4.20) and Eq. (4.21) have to be the same.
This is possible if
r =
u (D) D
1 2 2 u (D)
D
;
u (D) D 2 D
u (D)
and =
u (D) D
D .
u (D)
Let us assume that is constant. After integrating the second of these relations two times, we
obtain that besides some irrelevant integration constant,
u (D) =
D1 1
,
1
,
D
where is the CRRA. Hence, under CRRA preferences we have that,

r = + D
( + 1) 2
D ,
2
= D .
Finally, by replacing these expressions for the short-term rate and the risk-premium into Eq.
(4.13) leaves,
1

D,
S(D) =
(1 ) D 12 2D
provided the following conditions holds true:

1 2
> (1 ) D D .
2
109
(4.22)
c
by
A. Mele
We are only left to check that the transversality condition (4.17) holds at the equilibrium
S = V . We have that under the previous inequality,

lim Et er( t) V ( ) = lim Et er( t) S( )

lim Et [mt ( )S( )]
!
1 2
lim Et e(r+ 2 )( t)(W ( )W (t)) S( )
= S (t) lim Et e(D 2 D r 2

)( t)+( D )(W ( )W (t))
= S (t) lim e(rD +D )( t)
= S (t) lim e[(1)(D 2 D )]( t)

1
= 0.
4.2.3
(4.23)
Bubbles
The transversality condition in Eq. (4.17) is often referred to as a no-bubble condition. To

illustrate the reasons underlying this denition, note that Eq. (4.11) admits an innite number
of solutions. Each of these solutions takes the following form,
S(D) = KD + AD , K, A, constants.
(4.24)
Indeed, by plugging Eq. (4.24) into Eq. (4.11) reveals that Eq. (4.24) holds if and only if the
following conditions holds true:
0 = K (r + D D ) 1,
1
and 0 = (D D ) + ( 1) 2D r.
2
(4.25)
The rst condition implies that K equals the price-dividend ratio in Eq. (4.13), i.e. K =
SF (D)/ D. The second condition leads to a quadratic equation in , with the two solutions,
1 < 0
and
2 > 0.
Therefore, the asset price function takes the following form:

S(D) = SF (D) + A1 D1 + A2 D2 .
It satises:
lim S (D) = , if A1 0,
D0
lim S (D) = 0 if A1 = 0.
D0
To rule out an explosive behavior of the price as the dividend level, D, gets small, we must set
A1 = 0, which leaves,
S (D) = SF (D) + B (D) ,
110
B (D) A2 D2 .
(4.26)
c
by
A. Mele
The component, SF (D), is the fundamental value of the asset, as by Eq. (4.14), it is the
risk-adjusted present value of the expected dividends. The second component, B (D), is simply
the dierence between the market value of the asset, S (D), and the fundamental value, SF (D).
Hence, it is a bubble.
We seek conditions under which Eq. (4.26) satises the transversality condition in Eq. (4.17).
We have,

lim Et er( t) S( ) = lim Et er( t) SF (D ( )) + lim Et er( t) B (D ( )) .

By Eq. (4.23), the fundamental value of the asset satises the transversality condition, under
the condition given in Eq. (4.22). As regards the bubble, we have,
!

lim Et er( t) B (D ( )) = A2 lim Et er( t) D ( )2

= A2 D (t)2 lim Et e(2 (D D )+ 2 2 (2 1)D r)( t)

= A2 D (t)2 ,
(4.27)
where the last line holds as 2 satises the second condition in Eq. (4.25). Therefore, the bubble
can not satisfy the transversality condition, except in the trivial case in which A2 = 0. In other
words, in this economy, the transversality condition in Eq. (4.17) holds if and only if there are
no bubbles.
4.2.4
Reflecting barriers and absence of arbitrage
Next, suppose that insofar as the dividend D ( ) uctuates above a certain level D > 0, everything goes as in the previous section but that, as soon as the dividends level hits a barrier
D, it is reected back with probability one. In this case, we say that the dividend follows a
process with reflecting barriers. How does the price behave in the presence of such a barrier?
First, if the dividend is above the barrier, D > D, the price is still as in Eq. (4.24),
S(D) =
1
D + A1 D1 + A2 D2 .
r D + D
First, and as in the previous section, we need to set A2 = 0 to satisfy the transversality condition
in Eq. (4.17) (see Eq. (4.27)). However, in the new context of this section, we do not need to set
A1 = 0. Rather, this constant is needed to pin down the behavior of the price function S (D)
in the neighborhood of the barrier D.
We claim that the following smooth pasting condition must hold in the neighborhood of D,
S (D) = 0.
(4.28)
This condition is in fact a no-arbitrage condition. Indeed, after hitting the barrier D, the dividend is reected back for the part exceeding D. Since the reection takes place with probability
one, the asset is locally riskless at the barrier D. However, the dynamics of the asset price is,
dS
D DS
= S d +
dW.
S
S

S
111
c
by
A. Mele
4.3. Martingales and arbitrage in a diusion model
Therefore, the local risklessness of the asset at D is ensured if S (D) = 0. [Warning: We need
to add some local time component here.] Furthermore, rewrite Eq. (4.10) as,
S +
D
D DS (D)
r = S =
.
S
S (D)
If D = D then, by Eq. (4.28), S (D) = 0. Therefore,

S +
D
= r.
S
This relations tells us that holding the asset during the reection guarantees a total return
equal to the short-term rate. This is because during the reection, the asset is locally riskless
and, hence, arbitrage opportunities are ruled out when holding the asset will make us earn no
more than the safe interest rate, r. Indeed, by previous relation into the wealth equation (4.9),
and using the condition that S = 0, we obtain that

D
dV = S + r + rV c d + S dW = (rV c) d .
S
This example illustrates how the relation in Eq. (4.10) works to preclude arbitrage opportunities.
Finally, we solve the model. We have, K SF (D)/ D, and
0 = S (D) = K + 1 A1 D1 1 ;
Q S (D) = KD + A1 D1 ,
where the second condition is the value matching condition, which needs to be imposed to
ensure continuity of the pricing function with respect to D and, hence absence of arbitrage.
The previous system can be solved to yield1
Q=
1 1
KD
1
and A1 =
K 11
D
.
1
Note, the price is an increasing and convex function of the fundamentals, D.
4.3 Martingales and arbitrage in a diusion model

4.3.1
The information framework
We still consider a Lucas type economy, but consider a nite horizon T < . The primitives
include a probability space (, F , P ). Let W be a standard Brownian motion in Rd . Dene
F = {F (t)}t[0,T ] as the P -augmentation of the natural ltration F W ( ) = (W (s), s )
generated by W , with F = F (T ).
We consider m trees and an accumulation factor. These assets, and other inside money
assets (i.e., assets in zero net supply) to be introduced later, are exchanged without frictions.
The trees entitle to receive the usual fruits, or dividends, Di = {Di ( )} [t,T ] , i = 1, , m, which
are positive F( )-adapted bounded processes. Fruits are the numeraire. Let S+ = {S+ ( ) =
1 In this model, we take the barrier D as given. In other context, we might be interested in controlling the dividend D in such
a way that as soon as the price, q, hits a level Q, the dividend level D is activate to induce the price q to increase. The solution for
1
Q reveals that this situation is possible when D =
K 1 Q, where Q is an exogeneously given constant.
1 1
112
c
by
A. Mele
(S0 ( ), , Sm ( )) } [t,T ] be the positive F ( )-adapted asset price process. The accumulation
factor does not distribute dividends. Its price satises:
"

S0 ( ) = exp
r(u)du ,
t
#T
where r( ) is F ( )-adapted process satisfying E( t r( )du) < . We assume the dynamics of
the last components of S+ , i.e. S (S1 , , Sm ) , satisfy:
dSi ( ) = Si ( )ai ( )dt + Si ( ) i ( )dW ( ), i = 1, , m,
(4.29)
where a
i ( ) and i ( ) are processes satisfying the same properties as r, with i ( ) Rd . We
assume that rank(( ; )) = m d a.s., where ( ) ( 1 ( ), , m ( )) .
We assume that Di is solution to
dDi ( ) = Di ( )aDi ( )d + Di ( ) Di ( )dW ( ),
where aDi ( ) and Di ( ) are F( )-adapted, with Di Rd .
A strategy is a predictable process in Rm+1 , denoted as: = {( ) = ( 0 ( ), , m ( )) } [t,T ] ,
#T
and satisfying E( t ( )2 d ) < . The value of a strategy, net of dividends, is: V S+ ,
where S+ is a row vector. By generalizing Section 4.3.1, we say a strategy is self-nancing if its
value V , is the solution to:

dV = (a 1m r) + V r c dt + dW,
(4.30)
where 1m is a m-dimensional vector of ones, ( 1 , , m ) , i i Si , i = 1, , m,

1
m
a (a1 + D
, , a
m + D
) . The solution to the previous equation is, for each [t, T ],
S1
Sm
V x,,c ( )
= x
S0 ( )
"
c (u)
du+
S0 (u)
"
(u) (a (u) 1m r (u))

du+
S0 (u)
"
(u) (u)
dW (u), (4.31)
S0 (u)
where x denotes the initial wealth. We require V to be strictly positive.

4.3.2
Viability
#
Let gi = SS0i + zi , i = 1, , m, where d
zi = S10 dzi and zi ( ) = t Di (u)du. Let us generalize the
denition of the risk-neutral probability in Eq. (4.15), and introduce the set Q of risk-neutral,
or equivalent martingale, probabilities, dened as:
Q {Q P : gi is a Q-martingale} .
The aim of this section is to show the equivalent of Theorem 2.8 in Chapter 2: Q is not empty
if and only if there are not arbitrage opportunities.
Associated to every F(t)-adapted process {(t)}t[0,T ] satisfying some basic regularity conditions (essentially, the Novikovs condition),
"
W0 (t) = W (t) +
(u)du, [t, T ],
(4.32)
t
113
c
by
A. Mele
is a standard Brownian motion under a probability Q which is equivalent to P , with RadonNikodym derivative equal to,

"
" T
dQ
1 T
2
(T )
= exp
( ) d
( )dW ( ) .
(4.33)
dP
2 t
t
The process ( ) [t,T ] is a martingale under P . This result is the celebrated Girsanovs theorem.
Now let us rewrite Eq. (4.29) under such a new probability by plugging W0 in it. Under Q,
dSi ( ) = Si ( ) (ai ( ) i ( ) ( )) dt + Si ( ) i ( )dW0 ( ), i = 1, , m.
We also have

Si
d
gi ( ) = d
S0
( ) + d
zi ( ) =
Si ( )
[(ai ( ) r( )) d + i ( )dW ( )] .
S0 ( )
If gi is a Q-martingale, i.e.

" T

S0 ( )
Q S0 ( )
Si (T ) +
Di (s)ds F ( ) , i = 1, , m,
Si ( ) = E
S0 (T )
S0 (s)
(4.34)
it is necessary and sucient that ai i = r, i = 1, , m, or

a( ) 1m r( ) = ( ) ( ) .
(4.35)
Therefore, by Eqs. (4.30), (4.32) and (4.35), we have that, for [t, T ],
V x,,c ( )
= x
S0 ( )
"
c (u)
du +
S0 (u)
"
(u) (u)
dW0 (u).
S0 (u)
(4.36)
Consider the following denition:

Definition 4.1 (Arbitrage
opportunity). A portfolio
is an arbitrage opportunity if V x,,0 (t)

S01 (T ) V x,,0 (T ) and Pr S01 (T ) V x,,0 (T ) x > 0 > 0.
We have:
Theorem 4.2. There are no arbitrage opportunities if and only if Q is not empty.
A proof of this theorem is in the Appendix. The if part follows easily, by Eq. (4.36). The
only if part is more elaborated, but its basic structure can be understood as follows. By the
Girsanovs theorem, the statement absence of arbitrage opportunities Q Q is equivalent
to absence of arbitrage opportunities satisfying Eq. (4.35). If Eq. (4.35) didnt hold, one
could implement an arbitrage, and nd a nonzero : = 0 and (a1m r) = 0. Once could
then use when a 1m r > 0 and when a 1m r < 0, and obtain an appreciation rate of V
greater than r in spite of having zeroed uncertainty through = 0. If Eq. (4.35) holds, such
an arbitrage opportunity would never occur, as in this case for each , (a 1m r) = .
Let
&
'
x L2t,T,m : x = 0d
114
c
by
A. Mele

and
&
z L2t,T,m : z = u,
'
for u L2t,T,d .
Then, we may formalize the previous reasoning as follows. The excess return vector, a 1m r,
must be orthogonal to all vectors in , and since and are orthogonal, a 1m r
, or L2t,T,d : a 1m r = .2
4.3.3
Market completeness
Let Y L2 (, F, P ). Consider the following denition:

Definition 4.3 (Market completeness). Markets are dynamically complete if for each random variable Y L2 (, F , P ), we can find a portfolio process : V x,,0 (T ) = Y a.s.
The previous denition is the natural continuous-time counterpart to that we gave in the
discrete-time case (see Chapter 2). In analogy with the conclusions in Chapter 2, we shall prove
that in continuous-time, markets are dynamically complete if and only if (i) m = d and (ii) the
price volatility matrix of the available assets (primitives and derivatives) is nonsingular. We shall
provide a sketch of the proof for the suciency part of this statement (see, e.g., Karatzas (1997
pp. 8-9) for the converse), which relates to the existence of fully spanning dynamic strategies.
So given a Y L2 (, F , P ), let m = d and suppose the volatility matrix is nonsingular. Let
us consider the Q-martingale:

M ( ) E Q S0 (T )1 Y F( ) .
(4.37)
By the representation theorem of continuous local martingales as stochastic integrals with

respect to Brownian motions (e.g., Karatzas and Shreve (1991) (thm. 4.2 p. 170)), there exists
L20,T,d (, F , Q) such that M can be written as:
"
M ( ) = M (t) +
(u)dW0 (u).
t
We wish to nd out a portfolio process such that the discounted wealth process, net of
consumption, S01 ( ) V x,,0 ( ) equals M ( ) under P (or, equivalently, under Q) a.s. By Eq.
(4.36),
"
V x,,0 ( )
(u) (u)
= x+
dW0 (u),
S0 ( )
S0 (u)
t
and so, by identifying, the portfolio we are looking for is

= S0 1 . Set, then, x = M (t).
Then, M ( ) = S01 ( ) V M(t),,0 ( ), and in particular, M (T ) = S01 (T ) V M(t),,0 (T ) a.s. By
comparing with Eq. (4.37), V M(t),,0 (T ) = Y .
Armed with this result, we can now easily state:
2 To
see that and are orthogonal spaces, note that:

x L2t,T ,m : x z = 0, z
=
=

x L2t,T,m : x u = 0, u L2t,T,d

x L2t,T,m : x = 0d

x L2t,T,m : x = 0d
.
115
c
by
A. Mele

Theorem 4.4. Q is a singleton if and only if markets are complete.
Proof. There exists a unique : a 1m r = m = d. The result follows by the

Girsanovs theorem.
When markets are incomplete, there is an innity of risk-neutral probabilities belonging to
Q. Absence of arbitrage does not allow us to recover a unique risk-neutral probability, just as
in the discrete time model of Chapter 2. One could make use of general equilibrium arguments,
but in this case we go beyond the edge of knowledge, although we shall see something in Part
II of these lectures on Asset pricing and reality.
The next results, provide a further representation of the set of risk-neutral probabilities Q,
in the incomplete markets case. Let L20,T,d (, F, P ) be the space of all F (t)-adapted processes
#T
x in Rd satisfying: 0 < 0 x(u)2 du < , and dene,
&
x L20,T,d (, F , P ) : (t)x(t) = 0m
where 0m is a vector of zeros in Rm . Let
'
a.s. ,

= 1 (a 1m r) .
can be interpreted as the process of unit risk-premia.

Under the usual regularity conditions,
In fact, all processes belonging to the set:
)
*
Z = : (t) = (t) + (t),
are bounded and, hence, can be interpreted as unit risk-premia processes. More precisely, dene
the Radon-Nikodym derivative of Q with respect to P on F(T ):

"
" T
:
dQ
1 T:
: :2
(t)dW (t) ,
(T )
= exp
:(t): dt
dP
2 0
0
and the density process of all Q P on (, F ),
" t

" t
1
2
(t) = (t) exp

(u) du
(u)dW (u) , t [0, T ]),
2 0
0
a strictly positive P -martingale. We have the following results, which follows for example by
He and Pearson (1991, Proposition 1 p. 271) or Shreve (1991, Lemma 3.4 p. 429):
Proposition 4.5. Q Q if and only if it is of the form: Q(A) = E(1A (T )), A F (T ).
To summarize, we have that dim( ) = d m. The previous result shows quite nitidly that
markets incompleteness implies the existence of an innity of risk-neutral probabilities. Such a
result was shown in great generality by Harrison and Pliska (1983).3
3 The so-called F
ollmer and Schweizer (1991) measure, or minimal equivalent martingale measure, is dened as: P (A)
E(1A
(T )), for each A F(T ).
116
c
by
A. Mele
4.4. Equilibrium with a representative agent
4.4 Equilibrium with a representative agent

4.4.1
Consumption and portfolio choices: martingale approaches
For now, we assume that markets are complete, m = d, and that there are no portfolio constraints or any other frictions. We consider the problem of an agent, who maximizes the expected
utility from his consumption ows, u (), plus the expected utility from terminal wealth, U (),
under the constraint in Eq. (4.31):4

" T
x,,c
J(0, V0 ) = max E U(V
(T )) +
u(c( ))d , s.t. Eq. (4.31) holds.
(,c,v)
The rst approach to solve this problem was introduced by Merton, which we shall see later.
We wish to present another approach, which makes use of Arrow-Debreu state prices, similarly
as in Chapter 2. Our rst task is to derive a budget constraint paralleling the budget constraint
in Chapter 2:

(4.38)
0 = c0 w0 + E m c1 w 1 ,
where c and w are consumption and endowments, and m is the discount factor m. In Chapter
2, such a budget constraint arises after having multiplied the initial budget constraint by the
Arrow-Debreu state prices,
s = ms Ps ,
ms (1 + r)1 s ,
s =
Qs
,
Ps
and after having taken the sum over all the states of nature. We wish to apply the same logic
here. First, we dene Arrow-Debreu state price densities:
t,T mt,T dP,
mt,T = S0 (T )1 (T ),
(T ) =
dQ
.
dP
(4.39)
As in the nite state space of Chapter 2, we multiply the budget constraint in Eq. (4.31) by
these Arrow-Debreu densities, and then, we take the integral over all states of nature. The
original problem, one with an innity of trajectory constraints, will then be reduced to one with
only one constraint, just as for the budget constraint in Eq. (4.38). Accordingly, multiply both
sides in Eq. (4.31) by 0,T = S0 (T )1 dQ, and rearrange terms, to obtain:
-"
.

x,,c

" T
T
(a 1m r) (u)du + ( )(u)dW (u)
(T )
c(u)
V
0=
+
du x dQ
dQ.
S0 (T )
S0 (u)
t S0 (u)
t
Next, take the integral over all states of nature. By the Girsanovs theorem,
x,,c

" T
V
(T )
c(u)
0=E
+
du x .
S0 (T )
t S0 (u)
We can retrieve back the budget constraint under the probability P . We have, by a change of
measure and computations in the Appendix, that:
x,,c

" T
" T
V
(T )
c(u)
x,,c
x=E
+
du = E mt,T V
(T ) +
mt,u c(u)du .
(4.40)
S0 (T )
t S0 (u)
t
4 Moreover, we assume that the agent only considers the choice space in which the control functions satisfy the elementary
Markov property and belong to L20,T ,m (, F, P ) and L20,T ,1 (, F, P ).
117
c
by
A. Mele

So the program is,

" T
(T t)
x,,c
J(t, x) = max E e
U (V
(T )) +
u( , c( ))d ,
(c,v)
t

" T
x,,c
s.t. x = E mt,T V
(T ) +
mt, c( )d .
t
Because of its emphasis on the equivalent martingale measure, this approach to solve the original
problem is known as relying on martingale methods. Critically, market completeness is needed
to use these methods, as in this case, there is one and only one Arrow-Debreu density process.
However, the same martingale methods can be applied in the presence of portfolio constraints
(which include incomplete markets as a special case) too, although in a slightly modied manner,
as we shall see in Section 4.5.
To solve the problem, consider the Lagrangean,
" T

max E
[u ( , c( )) mt, c( )] d + U (v) mt,T v + x ,
(c,v)
where is the constraints multiplier, and by Eqs. (4.33) and (4.39),

"

"
1
2
mt, = exp
r (u) + (u) du
(u) dW (u) .
2
t
t
(4.41)
The rst order conditions are:

uc ( , c( )) = mt, , for [t, T ), and U (V x,,c (T )) = mt,T .
(4.42)
To compute the portfolio-consumption policy, note that for c ( ) 0, the proof is just that
leading to Theorem 4.4. In the general case, dene,

" T

Q
1
1

M ( ) E S0 (T ) v +
S0 (u) c(u)du F ( ) .
t
Notice that:
M( ) = E
S01 (T )
v +
"

"

S0 (u) c(u)du F ( ) = E mt,T v +
By the predictable representation theorem, such that:

"
M ( ) = M (t) +
(u)dW (u).

mt,u c(u)du F ( ) .
Consider the process {m0,t V x,,c ( )} [t,T ] . By Itos lemma,

m0,t V
x,,c
( ) +
"
mt,u c(u)du = x +
By identifying,
( ) = V
x,,c
"

mt,u V x,,c (u)dW (u).

( ) 1
( ) ( ) +
( ) ,
mt,
118
(4.43)
c
by
A. Mele

where V x,,c ( ) can be computed from the constraint:

" T

x,,c
V
( ) = E m ,T v +
m ,u c(u)du F( ) ,
once that the optimal trajectory of c has been computed.

As an example, let U(v) = ln v and u(x) = ln x. By the rst order conditions (4.42), c(1 ) =
mt, , v1 = m ,T . By plugging these conditions into the constraint, one obtains the solution
for the Lagrange multiplier: = T +1
. By replacing this back into the previous rst order
x
x
x
1
1
conditions, one eventually obtains: c(t) = T +1
, and v = T +1
. As regards the portfolio
mt,
mt,T
process, one has that:

" T

M ( ) = E m ,T v +
mt,u c(u)du F( ) = x,
t
which shows that = 0 in the representation of Eq. (4.43). So by replacing = 0 into (4.43),
( ) = V x,,c ( ) ( ) 1 ( ) .
We can compute V x,,c in (4.14) by using c:

" T
x
m ,T
m ,u
x T + 1 ( t)
x,,
c
V
( ) =
E
+
du F ( ) =
,
T +1
mt,T
mt,
T +1
mt,u
where we used the property that m satises: mt,a mt,b = mt,b , t a b. The solution is:
( ) =
x T + 1 ( t)
( ) 1 ( )
mt,
T +1
whence, by taking into account the relation: a 1m r = ,

( ) =
4.4.2

x T + 1 ( t)
( )1 (a 1m r) ( ).
mt,
T +1
The older, Mertons approach: dynamic programming
The Mertons approach derives optimal consumption and portfolio through Bellmans dynamic
programming. Let us see how it works in the innite horizon case. The problem the agent faces
is:
"

( t)
J (V (t)) = max E
e
u (c( )) d
c
t

s.t. dV = (a 1m ) + rV c d + dW
Under regularity conditions,

1
0 = max E u(c) + J (V ) (a 1m r) + rV c + J (V ) J(V ) .

c
2
The rst order conditions lead to:
u (c) = J (v)
J (V )
and =
J (V )
119
1
(a 1m r) .
(4.44)
(4.45)
c
by
A. Mele
By plugging these expressions back to the Bellmans Equation (4.44) leaves:

J (V )
1
J (V ) 2
Sh J(V ),
0 = u(c) + J (V )
Sh + rV c + J (V )
J (V )
2
J (V )
(4.46)
where:
Sh (a 1m r) ( )1 (a 1m r),
with limT e(T t) E [J (V (T ))] = 0.

As an example, consider the CRRA utility u (c) = (c1 1) / (1 ). Conjecture that:
J(x) = A
x1 B
,
1
where A, B are constants to be determined. Using the rst condition in (4.45), leaves c =
A1/ V . By plugging this expression into Eq. (4.46), and using the conjectured analytical form
of J, we obtain:

1 Sh
1
1/
A
+r
(1 AB) .
0 = AV
+
1
2
1
1
This equation must hold for every V . Therefore

r(1 ) (1 )Sh
1 r(1 ) (1 )Sh
A=
,
B=
2 2
2 2
Clearly, lim1 J(V ) = 1 ln V .
4.4.3
Equilibrium
In a complete markets setting, an equilibrium is (i) a consumption plan satisfying the rst order
conditions (4.42); (ii) a portfolio process having the form in Eq. (4.43), and (iii) the following
market clearing conditions:
c ( ) = D( )
m

i=1
Di ( ), for [t, T ), q(T )
0 ( ) = 0, ( ) = S( ), for [t, T ] .
m
Si (T )
(4.47)
i=1
(4.48)
We now derive equilibrium allocations and Arrow-Debreu state price densities. First, note
that the dividend process, D, satises:
dD( ) = aD ( )D( )d + D ( )D( )dW ( ),

m
where aD D m
i=1 aDi Di and D D
i=1 Di Di .
We have:
d ln uc ( , D( )) = d ln uc ( , c( ))
= d ln mt,

1
2
= r( ) + ( ) dt ( )dW ( ),
2
120
(4.49)
c
by
A. Mele
where the rst equality holds in an equilibrium, the second equality follows by the rst order
conditions in (4.42), and the third equality is true by the denition of mt, in Eq. (4.41).
Finally, by Itos lemma, ln uc ( , D( )) is solution to:
-
u c
ucc 1 2 2
d ln uc =
+ aD D
+ D D
uc
uc
2
uccc
uc
ucc
uc
2 %.
dt +
ucc
D D dW.
uc
(4.50)
By identifying drifts and diusion terms in Eqs. (4.49)-(4.50), we obtain, after a few simplications, the expression for the equilibrium short term rate and the prices of risk:

u c ( , D( ))
ucc ( , D( )) 1
2
2 uccc ( , D( ))
r( ) =
+ aD ( )D( )
+ D ( ) D( )
uc ( , D( ))
uc ( , D( ))
2
uc ( , D( ))
ucc ( , D( ))
( ) =
D ( ) D ( ) .
uc ( , D( ))
For example, consider the CRRA utility function, if u ( , c) = e( t) (c1 1) / (1 ), and
m = 1. Then,
1
r( ) = + aD ( ) ( + 1) D ( )2 ,
2
( ) = D ( ) .
Appendix 2 performs Walrass consistency tests: Eq. (4.47) Eq. (4.48).

4.4.4
Continuous-time Consumption-CAPM
By Eq. (4.34),

" T

S0 ( )
S0 ( )
Si ( ) = E
Si (T ) +
Di (s)ds F ( )
S0 (T )
S0 (s)

" T

mt,T
mt,s

= E
Si (T ) +
Di (s)ds F( ) ,
mt,
mt,
Q
where the second line follows by the same arguments leading to Eq. (4.40). Replacing the
rst order condition in (4.42), and the equilibrium conditions in Eq. (4.47), we obtain the
consumption CAPM evaluation of each asset:

-
.

" T

u q(T )
u (D(s))

Si ( ) = E
Si (T ) +
Di (s)ds F ( ) , i = 0, 1, , m.

u (D( ))
u (D( ))
As an example, consider a pure discount bond, with price b. We have that its dividend is zero
and that b(T ) = 1. Therefore,
-
.

u q(T )
mt,T
b( ) = E
F ( ) ,
F( ) = E
u (D( ))
mt,
where mt, is as in Eq. (4.41).
121
c
by
A. Mele
4.5. Market imperfections and portfolio choice
4.5 Market imperfections and portfolio choice

The setup is as in Section 4.3, where we x m = d. To allow for frictions such as market
incompleteness or short sale constraints, we assume that the vector of normalized portfolio
shares in the risky assets, p (t) (t) /V x,,c (t), is constrained to lie in a closed convex set
K Rd .
We follow the approach put forward by Cvitanic and Karatzas (1992), which consists in
embedding the constrained portfolio choice of the investor in a set of unconstrained portfolio
optimization problems. Under regularity conditions that we shall not deal with in these lectures,
it is shown that in this set of unconstrained problems, there exists one, which happens to be the
solution to the original constrained portfolio problem. So the constrained portfolio problem is
solved, once we solve for the unconstrained, which we can do through the martingale methods in
Section 4.3. This approach is closely related to the discrete time minimax probability mentioned
in Chapter 2. It is a systematic approach to consumption and portfolio policies in a context of
constrained portfolio choices, and generalizes results from He and Pearson (1991).
The starting point is the denition of the support function,
() = sup(p ),
pK
and its effective domain,
Rd ,
(4.51)
= { Rd : () < }.
K
The role of the support function is to tilt the dynamics of the price system in Section 4.3,
as follows:
dS0 (t)
= r (t) dt,
Si (t)
dSi (t)
= ai (t) dt + i (t) dW (t)
Si (t)
(i = 1, , d)
(4.52)
where:
r r + () ,
a
i ai + + () ,
and a
i is as in Section 4.3.
The main result is as follows. Denote with Val (x; K) the value of the problem faced by an
investor facing a portfolio constraint K Rd , when his initial wealth is x. Let Val (x) be the
corresponding value of the problem faced by an unconstrained investor in the market (4.52).
Clearly, this value is just Val0 (x) for the market considered in Sections 4.3-4.4. Moreover, for
each Rd , the unconstrained program the investor faces in the market (4.52), can be solved
through martingale methods, using the unique risk-neutral probability Q , equivalent to P ,
with Radon-Nikodym derivative equal to,
" T

"
:2
1

dQ
1 T:
0
1
: (t) (t): dt . (4.53)
(T )
= (T ) exp
(t) (t) dW (t)
dP
2 0
0
Then, under regularity conditions, we have that:
Val (x; K) = inf (Val (x)) ,
(4.54)
and optimal consumption and portfolio choices for this unconstrained problem are exactly those
chosen by the investor constrained to have p K. Appendix 4 provides an informal sketch of
the arguments leading to Eq. (4.54).
122
c
by
A. Mele
4.6. Jumps
Examples of the support function in Eq. (4.51) are the unconstrained case: K = Rd , in
= {0} and = 0 on K;
prohibition of short-selling: K = [0, )d , in which case
which case K
= K and = 0 on K,
or: incomplete markets: K = {p Rd : pM+1 = = pD = 0} (i.e.
K
= { Rd : 1 = = M = 0} and
the rst M assets can only be traded), in which case K
= 0 on K.
In the context of log-utility functions, we have that,

:2
:
= arg min 2 () + : + 1 : ,
where = 1 (a 1d r). Applications of this will be worked out in Part II on Asset pricing
and reality.
4.6 Jumps
Brownian motions are well suited to model the price behavior of liquid assets or assets issued by
names or Governments not subject to default risk. There is, however, a fair amount of interest in
modeling discontinuous changes in asset prices. Fixed income instruments may undergo liquidity
dry-ups, or even default, causing price discontinuities that we wish to model. This section is
an introduction to Poisson models, a class of processes that is particularly useful in addressing
these issues.
4.6.1
Poisson jumps
Let (t, T ) be a given interval, and consider events in that interval which display the following
properties:
(i) The random number of events arrivals on any disjoint time intervals of (t, T ) are independent.
(ii) Given two arbitrary disjoint but equal time intervals in (t, T ), the probability of a given
random number of events arrivals is the same in each interval.
(iii) The probability that at least two events occur simultaneously in any time interval is zero.
Next, let Pk ( t) be the probability that k events arrive during the time interval t. We
make use of the previous three properties to determine the functional form of Pk ( t). First,
Pk ( t) must satisfy:
P0 ( + d t) = P0 ( t) P0 (d ) ,
(4.55)
and we impose
P0 (0) = 1,
Pk (0) = 0 for k 1.
(4.56)
Eq. (4.55) and the rst condition in (4.56) are satised by P0 ( ) = ev , for some constant v,
which we take to be positive, so as to ensure that P0 [0, 1]. Furthermore, we have that:
P1 ( + d t) = P0 ( t) P1 (d ) + P1 ( t) P0 (d )
..
.
(4.57)
Pk ( + d t) = Pk1 ( t) P1 (d ) + Pk ( t) P0 (d )
..
.
123
c
by
A. Mele
4.6. Jumps
The rst equation in (4.57) can be rearranged as follows:
P1 ( + d t) P1 ( t)
1 P0 (d )
P1 (d )
=
P1 ( t) +
P0 ( t) .
d
d
d
For small d , P1 (d ) 1 P0 (d ) and P0 (d ) = 1 vd + O (d 2 ) 1 vd . Therefore,
P1 ( t) = vP1 ( t) + vP0 ( t). By a similar reasoning,
Pk ( t) = vPk ( t) + vPk1 ( t) .
The solution to this equation is:
v k ( t)k v( t)
e
.
Pk ( t) =
k!
4.6.2
Interpretation
A Poisson model is one of rare events. Moreover, by:

E (event arrival in d ) = P1 (d ) = vd .
For this reason, we usually refer to the parameter v as the intensity of event arrivals.
To provide additional intuition about the mathematics of rare events, consider the expression
for the probability of k arrivals in n trials, predicted by a binomial distribution:

n k nk
n!
pk q nk ,
p q
=
p, q > 0, p + q = 1,
Pn,k =
k! (n k)!
k
where p is the probability of arrival for each trial. We want to model the probability p as a
function of n, with the feature that limn p(n) = 0, so as to make each arrival rare. One
possible choice is p (n) = na , for some constant a > 0. Under this assumption, we have:
Pn,k =
=
=
=
=
n!
p(n)k (1 p(n))nk
k! (n k)!
a k
n!
a nk
1
k! (n k)! n
n
a k
n!
a n
a k
1
1
k! (n k)! n
n
n

k
n
n!
a
a
a k
1
nk (n k)! k!
n
n
k
n n1
nk+1a
a n
a k
1
1
,
n
n k!
n
n
n
k times
leaving,
ak a
e .
n
k!
Next, we split the interval ( t) into n subintervals of length nt , and then make the probability of one arrival in each sub-interval proportional to each sub-interval length, as illustrated
in Figure 4.1,
a
t
,
a v( t).
p(n) = v
n
n
124
lim Pn,k Pk =
c
by
A. Mele
4.6. Jumps
n 1 ( t)
t
n subintervals
FIGURE 4.1. Heuristic construction of a Poisson process from a binomial distribution.
The Poisson model in the previous section is thus as that we consider here, with n ,
which is continuous-time, as each sub-interval in Figure 4.1 shrinks to d . The probability there
is one arrival in d is vd , which is also the expected number of events in d as shown below:
E (# arrivals in d )
= Pr (one arrival in d ) one arrival + Pr (zero arrivals in d ) zero arrivals
= Pr (one arrival in d ) 1 + Pr (zero arrivals in d ) 0
= vd .
The heuristic construction in this section opens the way to how we can simulate Poisson
processes. We can just simulate a Uniform random variable U (0, 1), with the continuous-time
process being approximated by Y , where:
+
0 if 0 U < 1 vh
Y =
1 if 1 vh U < 1
where h is a discretization interval.
4.6.3
Properties and related distributions
We ckeck that Pk is a probability. We have:
Pk = e
k=0
since
k=0

ak
k=0
k!
= 1,

ak k! is the McLaurin expansion of ea . Second, we compute the mean,
Mean =

k=0
k Pk = e

k=0
ak
= a.
k!
A related distribution is the exponential (or Erlang) distribution. Remember, the probability
of zero arrivals in t predicted by the Poisson model is P0 ( t) = ev( t) , from which it
follows that:
G ( t) 1 P0 ( t) = 1 ev( t)
is the probability of at least one arrival in t. The function G can be also interpreted as the
probability the rst arrival occurred before , starting from t. The density function of G is:
G ( t) = vev( t) .
The rst two moments of the exponential distribution are:

"
"

2
vx
1
Mean =
xve dx = v , Variance =
x v 1 vevx dx = v2 .
g ( t) =
125
c
by
A. Mele
4.6. Jumps
The expected time of the rst arrival occurred before starting from t equals v 1 . More generally, v 1 can be interpreted as the average time from an arrival to another.5
A more general distribution than the exponential is the Gamma distribution with density:
t)]1
.
( 1)!
v( t) [v (
g ( t) = ve
The exponential distribution obtains when = 1.

4.6.4
Some asset pricing implications
This section is a short introduction to modeling asset prices as being driven by Brownian
motions and jumps processes. We model jumps by interpreting the arrivals in the previous
sections as those events upon which a certain random variable experiences a jump of size S,
where S is another random variable with a xed probability p. A simple model is:
dS( ) = b(S( ))d + (S( ))dW ( ) + (S( )) S dZ( ),
(4.58)
where b, , are given functions (with > 0), W is a standard Brownian motion, and Z is a
Poisson process with intensity equal to v, i.e.
(i) Pr (Z(t)) = 0.
(ii) t 0 < 1 < < N < , Z( 0 ) and Z( k ) Z( k1 ) are independent for each
k = 1, , N.
(iii) > t, Z( ) Z(t) is a random variable with Poisson distribution and expected value
v( t), i.e.:
v k ( t)k v( t)
Pr (Z( ) Z(t) = k) =
.
e
k!
In this framework, k is the number of jumps over the time interval t.6 From this, we have
that Pr (Z( ) Z(t) = 1) = v ( t) ev( t) and for t small,

Pr (dZ( ) = 1) Pr ( Z( ) Z(t)| t = 1) = v ( t) ev(t) t vd .
More generally, the process {Z( ) v ( t)} t is a martingale.

Armed with these preliminary facts, we can provide a heuristic derivation of Itos lemma for
jump-diusion processes. Consider any function f with enough regularity conditions, a rational
function of time and S in Eq. (4.58), i.e. f ( ) f (S( ), ). Consider the following expansion
of f :

df ( ) =
+ L f (S( ), ) d + fS (S( ), ) (S(t))dW ( )
+ [f (S( ) + (S( )) S, ) f (S( ), )] dZ( ).
5 Suppose arrivals are generated by Poisson processes, and consider the random variable time interval elapsing from one arrival
to next one. Let be the instant at which the last arrival occurred. Then, the probability the time which will elapse from
the last arrival to the next is less than is the same as the probability that during the time interval , there is at least one
arrival.
6 For simplicity, we take v to be constant. If v is a deterministic function of time, we have that

k

v(u)du
Pr (Z( ) Z(t) = k) = t
exp
v(u)du , k = 0, 1,
k!
t
and there is also the possibility to model v as a function of the state: v = v(q), for example. Cox processes.
126
c
by
A. Mele
4.7. Continuous-time Markov chains
The rst two terms in are the usual Itos lemma terms, with
+L denoting the innitesimal
generator for diusions. The third term accounts for jumps. If there are no jumps from time
to time (where d = ), then dZ( ) = 0. If there is a jump then dZ( ) = 1, and in this
case f , as a rational function, needs also instantaneously jump to f (S( ) + (S( )) S, ).
The jump will be exactly f (S( ) + (S( )) S, ) f (S( ), ), where S is another random
variable with a xed probability measure. Clearly, if f(S, ) = S, we are back to the initial
jump-diusion model in Eq. (4.58).
To derive the innitesimal generator for jumps-diusion, LJ f say, note that:

E (df) =
+ L fd + E [(f (S + S, ) f (S, )) dZ( )]

=
+ L fd + E [(f (S + S, ) f (S, )) v d ] ,
or
J
L f = Lf + v
"
supp(S)
[f (S + S, ) f (S, )] p (dS) ,
where supp (S) denotes

theJ support of S. Therefore, the innitesimal generator for jumpsdiusion is simply,
+ L f.
4.6.5
An option pricing formula
Merton (1976, JFE), Bates (1988, working paper), Naik and Lee (1990, RFS) are the seminal
papers.
4.7 Continuous-time Markov chains

Needed to model credit risk.
127
c
by
A. Mele
4.8. Appendix 1: Convergence issues
4.8 Appendix 1: Convergence issues

We have,
(1)
(2)
(1)
(2)
ct + St t+1 + bt t+1 = (St + Dt ) t + bt t

(1)
(2)
where Vt St t + bt t
(1)
Vt + Dt t ,
is wealth net of dividends. We have,

(1)
(2)
Vt Vt1 = St t + bt t Vt1

(1)
(2)
(1)
(2)
(1)
= St t + bt t ct1 + St1 t + bt1 t Dt1 t1
(1)
(2)
(1)
= (St St1 ) t + (bt bt1 ) t ct1 + Dt1 t1 ,
and more generally,

(1)
(2)
(1)
Vt Vt = (St St) t + (bt bt ) t (ct ) + (Dt ) t .

Now let 0 and assume that (1) and (2) are approximately constant between t and t . We
have:
dV ( ) = (dS( ) + D( )d ) (1) ( ) + db( ) (2) ( ) c( )d .
Assume that
db( )
= rd .
b( )
The budget constraint can then be written as:

dV ( ) = (dS( ) + D( )d ) (1) ( ) + rb( )(2) ( )d c( )d

= (dS( ) + D( )d ) (1) ( ) + r V S( ) (1) ( ) d c( )d
= (dS( ) + D( )d rS( )d ) (1) ( ) + rV d c( )d

dS( ) D( )
=
+
d rd (1) ( )S( ) + rV d c( )d
S( )
S( )

dS( ) D( )
=
+
d rd ( ) + rV d c( )d .
S( )
S( )
128
c
by
A. Mele
4.9 Appendix 2: Proofs of selected results

4.9.1
Proof of Theorem 4.2
As mentioned in the main text, we have that by by the Girsanovs theorem, Q is non-empty if and
only if Eq. (4.35) holds true. Therefore, the proof will rely on Eq. (4.35). If part. With c 0, Eq.
(4.36) is:
"
(u) (u)
V x,,0 ( )
=x+
dW0 (u), [t, T ],
S0 ( )
S0 (u)
t
which implies, x = EQ [S0 (T )1 V x,,0 (T )]. An arbitrage opportunity is V x,,0 (t) S0 (T )1 V x,,0 (T )
a.s., which combined with the previous equality leaves: V x,,0 (t) = S0 (T )1 V x,,0 (T ) Q-a.s. (if a r.v.
y 0 and
y ) = 0, this means thaty = 0 a.s.) and, hence, P -a.s. The last equality is in contradiction
Et (
with Pr S0 (T )1 V x,,0 (T ) x > 0 > 0, as required by Denition 4.3.
Only if part. We combine portions of proofs in Karatzas (1997, thm. 0.2.4 pp. 6-7) and ksendal
(1998, thm. 12.1.8b, pp. 256-257). We let:
Z ( ) = { : Eq. (4.35) has no solutions}
= { : a( ; ) 1m r( ; )
/ }
)
*
= : ( ; ) : ( ; ) ( ; ) = 0 and ( ; ) (a( ; ) 1m r( ; )) = 0 ,
and consider the following portfolio,

+
k sign ( ; ) (a( ; ) 1m r( ; )) ( ; )
( ; w) =
0
for Z( )
for
/ Z( )
Clearly
is ( ; )-measurable, and generates, by Eq. (4.31),
"
"
V x,,0 ( )
(u) (a (u) 1m r (u))
(u) (u)
=x+
IZ(u) du +
IZ(u) dW (u)
S0 ( )
S
(u)
S0 (u)
0
t
t
%
$
"
(u) (a (u) 1m r (u))

IZ(u) du
=x+
S0 (u)
t
x.
So the market has no arbitrage only if IZ(u) = 0, i.e. only if Eq. (4.35) has at least one solution.
4.9.2
Proof of Eq. (4.40).
We have:

" T
V x,,c (T )
c(u)
x=E
+
du
S0 (T )
t S0 (u)

" T
(T ) V x,,c (T )
(T )c(u)
+
du
= (t)1 E
S0 (T )
S0 (u)
t

" T
x,,c
(T ) V
(T )
(T )c(u)
1
= (t) E
+
E
F(u) du
S0 (T )
S0 (u)
t

" T
(T ) V x,,c (T )
E ( (T )| F(u)) c (u)
1
= (t) E
+
du
S0 (T )
S0 (u)
t

" T
(T ) V x,,c (T )
(u) c (u)
1
= (t) E
+
du
S0 (T )
S0 (u)
t

" T
x,,c
= E mt,T V
(T ) +
mt,u c(u)du ,
t
129
c
by
A. Mele
where we used the fact that c is adapted, the law of iterated expectations, the martingale property of
, and the denition of m0,t .
4.9.3
Walrass consistency tests
First, we show that Eq. (4.47) Eq. (4.48). To grasp intuition about the ongoing proof, consider the
two-period economy of Chapter 2. In that economy, absence of arbitrage opportunities implies that
Rd : (c1 w1 ) = S = (c0 w0 ), whence cs = ws , s = 0, , d = 0m . In the model of
the present chapter, absence of arbitrage opportunities implies that ! Q Q : such that:
"
"
0 ( )S0 ( ) + ( )1m
(u)
c (u)
V x,,c ( )
= 1
S(t)
+
dW
(u)
du.
0
m
S0 ( )
S0 ( )
S
(u)
S
0
0 (u)
t
t
That is,

0 ( )S0 ( ) + ( ) S ( ) 1m S ( )1m
+
S0 ( )
S0 ( )

"
"
"
(u) S (u) (u)

c (u)
S (u)(u)
= 1m S(t) +
dW0 (u)
du +
dW0 (u).
S0 (u)
S0 (u)
t
t S0 (u)
t

#

#

Plugging the solution SS0i ( ) = Si (t)+ t S01 Si (u)i (u)dW0 (u) t S01 Di (u)du in the previous
relation,

" T
" T
0 (T )S0 (T ) + (T ) S (T ) 1m
(u) S (u)
D(u) c(u)
=
(u)dW0 (u) +
du. (4A.1)
S0 (T )
S0 (u)
S0 (u)
t
t
When Eq. (4.47) holds, we have that V x,,c (T ) = 0 (T )S0 (T ) + (T )1m = q(T ) = S (T )1m , and
D = c, and Eq. (4A.1) becomes:
0 = x(T )
"
(u) S (u)
(u)dW0 (u),
S0 (u)
a martingale starting at zero, satisfying:

dx( ) =
( ) S ( )
( )dW0 ( ) = 0.
S0 ( )
Since ker() = {} then, we have that ( ) = S( ) a.s. for [t, T ] and, hence, ( ) = S( ) a.s. for
[t, T ]. It is easily checked that this implies 0 (T ) = 0 P -a.s. and that in fact, 0 ( ) = 0 a.s.
Next, we show that Eq. (4.48) Eq. (4.47). When Eq. (4.48) holds, Eq. (4A.1) becomes:
0 = y(T )
"
D(u) c(u)
du,
S0 (u)
a martingale starting at zero. We conclude by the same arguments used in the proof of the previous
part.
130
c
by
A. Mele
4.10. Appendix 3: The Greens function
4.10 Appendix 3: The Greens function

4.10.1
Setup
In Section 4.5, it is shown that in frictionless markets, the value of a security as of time is:

" T
mt,T
mt,s
V (x( ), ) = E
V (x(T ), T ) +
h (x(s), s) ds ,
(4A.2)
mt,
mt,
where mt, is the stochastic discount factor,
mt,

(t, )
1
dQ
=
=
.
S0 ( )
S0 ( ) dP F( )
Arrow-Debreu state price densities are:
t,T = mt,T dP =
S0 (t)
dQ,
S0 (T )
which now we want to characterize in terms of PDEs.

By the same reasoning produced in Section 4.5, Eq. (4A.2) can be rewritten as:

" T
S0 ( )
S0 ( )
V (x(T ), T ) +
h (x(s), s) ds .
V (x( ), ) = E
S0 (T )
S0 (s)
Let
(4A.3)

S0 (t )
a t , t =
.
S0 (t )
In terms of a, Eq. (4A.3) is:

"
V (x( ), ) = E a ( , T ) V (x(T ), T ) +

a ( , s) h (x(s), s) ds .
Next, consider the augmented state vector:

y(u) (a ( , u) , x(u)) , u T,
and let P ( y(t )| y( )) be the density function of the augmented state vector under the risk-neutral
probability. We have,

" T
a ( , s) h (x(s), s) ds
V (x( ), ) = E a ( , T ) V (x(T ), T ) +
=
"
"
a ( , T ) V (x(T ), T ) P ( y(T )| y( )) dy(T ) +
"
"
a ( , s) h (x(s), s) P ( y(s)| y( )) dy(s)ds.
If V (x(T ), T ) and a ( , T ) are independent,

" "
a ( , T ) V (x(T ), T ) P ( y(T )| y( )) dy(T ) =
a ( , T ) P ( y(T )| y( )) dy(T ) V (x(T ), T ) dx(T )
X
A
"
G( , T )V (x(T ), T ) dx(T )
X
where:
G( , T )
"
a ( , T ) P ( y(T )| y( )) dy(T ).
131
c
by
A. Mele
4.10. Appendix 3: The Greens function

If we assume the same thing for h, we eventually get:
"
"
V (x( ), ) =
G( , T )V (x(T ), T ) dx(T ) +
X
"
G( , s)h (x(s), s) dx(s)ds.
The function G is known as the Greens function:

"
G (t, ) G (x, t; , ) =
a (t, ) P ( y()| y(t)) da.
A
It is the value in state x Rd as of time t of a unit of numeraire at > t if future states lie in a
neighborhood (in Rd ) of . It is thus the Arrow-Debreu state-price density.
For example, a pure discount bond has V (x, T ) = 1 x, and h(x, s) = 1 x, s, and
"
G (x( ), ; , T ) d,
V (x( ), ) =
X
with
lim G (x( ), ; , T ) = (x( ) ) ,

T
where is the Dirac delta.
4.10.2
The PDE connection
We have,
V (x (t) , t) =
"
G (x(t), t; (T ) , T ) V ( (T ) , T ) d (T ) +
"
"
G(x(t), t; (s), s)h ((s), s) d(s)ds.
(4A.4)
Consider the scalar case. Eq. (4A.3) and the connection between PDEs and Feynman-Kac tell us that
under standard regularity conditions, V is solution to:
Now take the

Vt =
Vx =
Vxx =
1
0 = Vt + Vx + 2 Vxx rV + h.
(4A.5)
2
appropriate partial derivatives in Eq. (4A.4),
"
"
" T"
"
" T"
Gt V d
(x )hd +
Gt hdds =
Gt V d h +
Gt hdds
X
X
X
X
X
t
"
" T"
Gx V d +
Gx hdds
X
X
t
"
" T"
Gxx V d +
Gxx hdds
X
and replace them into Eq. (4A.5) to obtain:

"
1 2
0 =
Gt + Gx + Gxx rG V ( (T ) , T ) d(T )
2
X

" T"
1 2
+
Gt + Gx + Gxx rG h ( (s) , s) d (s) ds.
2
t
X
This shows that G is solution to
1
0 = Gt + Gx + 2 Gxx rG,
2
and
lim G (x, t; , T ) = (x ) .
tT
132
c
by
A. Mele
4.11. Appendix 4: Portfolio constraints
4.11 Appendix 4: Portfolio constraints

We are looking for a portfolio-consumption policy (p , c ) such that
" T

x,p ,c
Val (x; K) = E
u (t, c (t)) dt + U (V
(T )) Val (x) ,
(4A.6)
and p (t) K for all t [0, T ].

Note that because K contains the origin, then, the support function in Eq. (4.51) satises () 0
Moreover, an intuitive and important property of is that,
for each K.
p K () + p 0,
K.
(4A.7)
Next, dene the standard Brownian motion under the probability Q , dened through the RadonNikodym in Eq. (4.53):
W (t) = W (t) +
"
t
(u) +
(u) (u) du W0 (t) +
"
t

1 (u) (u) du,
where = 1 (a 1d r), and W0 is the usual Brownian under the risk-neutral probability in a market
without any frictions. If the price system is as in Eqs. (4.52), then, for any unconstrained portfolioconsumption (p, c), the dynamics of wealth, Vx,p,c say, are easily seen to be:

dVx,p,c = p + () Vx,p,c + rVx,p,c c dt + p dW0 .
So we have that under Q0 ,
" T
Vx,p,c (T )
c (t)
+
dt
S0 (T )
0 S0 (t)
" T x,p,c
" T x,p,c
!
V
(t)
V
(t)
=x+
p (t) (t) + ( (t)) dt +
p (t) (t) dW0 (t) .
S0 (t)
S0 (t)
0
0
Therefore, for any normalized portfolio-consumption (p, c), we have that the wealth dierence, (t)
Vx,,c (T )V x,,c (T )
, satises:
S0 (T )
d (t) =
!
Vx,,c (t)
p (t) (t) + ( (t)) dt + (t) p (t) (t) dW (t) ,
S (t)
0

(0) = 0.
m(t)
Next, consider, the simpler equation,
(t) =
(t) p (t) (t) dW (t) ,
d
(0) = 0.
(4A.8)
Because m (t) 0 by Eq. (4A.7), then, by a comparison theorem (e.g., Karatzas and Shreve (1991,
(t) = 0, where the last equality follows because the solution to Eq. (4A.8) is
p. 291-295)), (t)
(t) = (0) L (t), for some positive process L (t). Therefore, we have,
Vx,p,c (t) V x,p,c (t) ,
with an equality if ( (t)) + p (t) = 0 for all t.
(4A.9)
Finally, suppose there is a constrained portfolio-consumption pair (p , c ), such that

(
(t)) + p (t) (t) = 0.
133
(4A.10)
c
by
A. Mele
4.11. Appendix 4: Portfolio constraints

Naturally, we have that Val (x; K) Val (x) for all and, hence,
Val (x; K) inf (Val (x)) .
(4A.11)
Moreover, we have,
Val (x; K) = E
"
E
= E
"
"

u (t, c (t)) dt + U (V
(T )) , p (t) K

x,p ,c
u (t, c (t)) dt + U (V
(T ))

x,p ,c

u (t, c (t)) dt + U V
(T )
x,p,c
= Val (x) ,
(4A.12)
where the second line follows, because the value of the unconstrained problem is, of course, the largest
we may have, once we consider any arbitrary constrained portfolio-consumption (p , c ). The third
line follows by Eq. (4A.10) and (4A.9). The fourth line is the denition of Val (x). Combining (4A.11)
with (4A.12) leaves,
Val (x; K) = Val (x) .
that minimizes Val (x), then, the corresponding
The converse, namely if there exists a K
portfolio-consumption process (p , c ) is optimal for the constrained problem, is also true, but its
arguments (even informal) are omitted here.
134
4.12. Appendix 5: Models with nal consumption only
c
by
A. Mele
4.12 Appendix 5: Models with nal consumption only

Sometimes, we may beinterested in models with consumption taking place in at the end of the period
only. Let S = S (0) , S and = ((0) , ), where and S are both m-dimensional. Dene as usual
wealth as of time t as Vt St t . There are no dividends. A self-financing strategy satises,
St+ t+1 = St t Vt ,
t = 1, , T.
Therefore,
Vt = St t + St1 t1 St1 t1
= St t + St1 t1 St1 t (because is self-nancing)
= Vt1 + St t , St St St1 , t = 1, , T,
or,
Vt = V1 +
t
Sn n .
n=1
Next, suppose that

(0)
St
(0)
t = 1, , T,
= rt St1 ,
with {rt }Tt=1 given and to be dened more precisely below. The term St +
t can then be rewritten as:
(0) (0)
St t = St t + St t
(0)
(0)
(0)
(0)
= rt St1 t + St t
=
=
=
=
rt St1 t + rt St1 t rt St1 t + St t

rt St1 t rt St1 t + qt t
rt St1 t1 rt St1 t + St t (because is self-nancing)
rt Vt1 rt St1 t + St t ,
and we obtain
or,
Vt = (1 + rt ) Vt1 rt St1 t + St t ,
Vt = V1 +
t
n=1
(rn Vn1 rn Sn1 n + Sn n ) .
Next, considering small time intervals. In the limit we obtain:

dV (t) = r(t)V (t)dt r(t)S(t)(t)dt + dS(t)(t).
Such an equation can also be arrived at by noticing that current wealth is nothing but initial wealth
plus gains from trade accumulated up to now:
" t
(u).
V (t) = V (0) +
dS(u)
0
(t)+
dV (t) = dS(t)
= dS0 (t)0 (t) + dS(t)(t)
= r(t)S0 (t)0 (t)dt + dS(t)(t)
= r(t) (V (t) S(t)(t)) dt + dS(t)(t)
= r(t)V (t)dt r(t)S(t)(t)dt + dS(t)(t).
135
c
by
A. Mele
4.12. Appendix 5: Models with nal consumption only

Now consider the sequence of problems of terminal wealth maximization:
+
maxt E [ u(V (T ))| Ft1 ] ,
For t = 1, , T, Pt :
s.t. Vt = (1 + rt ) Vt1 rt St1 t + St t
Even if markets are incomplete, agents can solve the sequence of problems {Pt }Tt=1 as time unfolds.
Each problem can be written as:
- $
%
.
T

max E u V1 +
(rt Vt1 rt St1 t + St t ) Ft1 .

t
t=1
The FOC for t = 1 is:
whence

E u (V (T )) (S1 (1 + r0 ) S0 ) F0 ,
S0 = (1 + r0 )1
In general
St = (1 + rt )1
E [ u (V (T )) S1 | F0 ]
.
E [ u (V (T ))| F0 ]
E [ u (V (T )) St+1 | Ft ]
,
E [ u (V (T ))| Ft ]
t = 0, , T 1.
The previous relations suggest that we can dene a martingale measure Q for the discounted price
process by dening

u (V (T ))
dQ
=
.
dP
E [ u (V (T ))| Ft ]
Ft
Connections with the CAPM. Its easy to show that:

u (V (T ))
E (
rt+1 ) rt = cov
, rt+1 ,
E [ u (V (T ))| Ft ]
where rt+1 (St+1 St )/ St .
136
c
by
A. Mele
4.13. Appendix 6: Further topics on jumps
4.13 Appendix 6: Further topics on jumps

4.13.1
The Radon-Nikodym derivative
To derive heuristically the Radon-Nikodym derivative, consider the jump times 0 < 1 < 2 < <
n = T;. The probability of a jump in a neighborhood of i is v( i )d , and to nd what happens under
the risk-neutral probability or, in general, under any equivalent measure, we just write vQ ( i )d under
measure Q, and set vQ = vJ . Clearly, the probability of no-jumps between any two adjacent random
points i1 and i and a jump at time i1 is, for i 2, proportional to
v( i1 )e
i
i1
v(u)du
under the probability P,
and to
vQ ( i1 )e
i
i1
vQ (u)du
= v( i1 )J ( i1 )e
i
i1
v(u)J (u)du
under the probability Q.
As explained in section 6.9.3 (see also formula # 6.76)), these are in fact densities of time intervals
elapsing from one arrival to the next one.
Next let A be the event of marks at time 1 , 2 , , n . The Radon-Nikodym derivative is the
likelihood ratio of the two probabilities Q and P of A:
e
Q(A)
=
P (A)
1
t
v(u)J (u)du
v( 1 )J ( 1 )e
1
t
v(u)du
v( 1 )e
2
1
v(u)J (u)du
2
1
v(u)du
v( 2 )J ( 2 )e
v( 2 )e
3
2
v(u)du
3
2
v(u)J (u)du
where we have used

fact that given that at
the
1 0 = Jt, there are no-jumps, the probability of no-jumps
1 v(u)du
t
from t to 1 is e
under P and e t v(u) (u)du under Q, respectively. Simple algebra then
yields,
Q(A)
P (A)
= ( 1 ) ( 2 ) e
=
n
1
i=1
J ( i ) e
n
t
1
t
v(u)(J (u)1)du
2
1
v(u)(J (u)1)du
3
2
v(u)(J (u)1)du
v(u)(J (u)1)du
- $n
%.
n
1
J
= exp ln
J ( i ) e t v(u)( (u)1)du
= exp
= exp
- n
i=1
T
-"
i=1
ln ( i )
"
.
J

v(u) (u) 1 du
ln J (u)dZ(u)
"
.
J

v(u) (u) 1 du ,
where the last equality follows from the denition of the Stieltjes integral.
The previous results can be used to say something substantive on an economic standpoint. But
before, we need to simplify both presentation and notation. We have:
Definition (Doleans-Dade exponential semimartingale): Let M be a martingale. The unique solution to the equation:
"
L( ) = 1 +
L(u)dM(u),
t
is called the Doleans-Dade exponential semimartingale and is denoted as E(M).
137
c
by
A. Mele

4.13.2
Arbitrage restrictions
As in the main text, let now S be the price of a primitive asset, solution to:
dS
S
= bd + dW + SdZ
= bd + dW + S (dZ vd ) + Svd
= (b + Sv) d + dW + S (dZ vd ) .
Next, dene

dZ = dZ vQ d vQ = vJ ;
= dW + d .
dW
are Q-martingales. We have:

Both Z and W

dS
+ SdZ.
= b + Sv Q d + dW
S
The characterization of the equivalent martingale measure for the discounted price is given by the
following Radon-Nikodym density of Q with respect to P :
" T

" T

J
dQ
=E
( )dW ( ) +
( ) 1 (dZ( ) v( )) d ,
dP
t
t
where E () is the Doleans-Dade exponential semimartingale, and so:
b = r + v Q ES (S) = r + vJ ES (S).
Clearly, markets are incomplete here. It is possible to show that if S is deterministic, a representative
1
agent with utility function u(x) = x 11 makes J (S) = (1 + S) .
4.13.3
State price density: introduction
We have:
"
L(T ) = exp

v( ) J ( ) 1 d +
"

ln J ( )dZ( ) .
The objective here is to use Itos lemma for jump processes to express L in dierential form. Dene
the jump process y as:
"
"

v(u) J (u) 1 du +
ln J (u)dZ(u).
y( )
t
In terms of y, L is L( ) = l(y( )) with l(y) = ey . We have:
or,

dL( ) = ey( ) v( ) J ( ) 1 d + ey( ) + jump ey( ) dZ( )

J
= ey( ) v( ) J ( ) 1 d + ey( ) eln ( ) 1 dZ( )

dL( )
= v( ) J ( ) 1 d + J ( ) 1 dZ( ) = J ( ) 1 (dZ( ) v( )d ) .
L( )
The general case (with stochastic distribution) is covered in the following subsection.
138
c
by
A. Mele

4.13.4
State price density: general case
Suppose the primitive is:

dx( ) = (x( ))d + (x( ))dW ( ) + dZ( ),
and that u is the price of a derivative. Introduce the P -martingale,
dM( ) = dZ( ) v(x( ))d .
By Itos lemma for jump-diusion processes,
du(x( ), )
u(x( ), )
= u (x( ), )d + u (x( ), )dW ( ) + J u (x, ) dZ( )
= (u (x( ), ) + v(x( ))J u (x, )) d + u (x( ), )dW ( ) + J u (x, ) dM ( ),

where u = t
+ L u u, u = u
u, t
+ L is the generator for pure diusion processes and,
x
nally:
u(x( ), ) u(x( ), )
.
J u (x, )
u(x( ), )
Next generalize the steps made some two subsections ago, and let
= dW + d ;
dW
dZ = dZ vQ d .
and Z are Q-martingales.

The objective is to nd restrictions on both and v Q such that both W
Q
Below, we show that there is a precise connection between v and J , where J is the jump component
in the dierential representation of :
d( )
= (x( ))dW ( ) + J (x, ) dM( ), (t) = 1.
( )
The relationship is
v Q = v (1 + J ) ,
and a proof of these facts will be provided below. What has to be noted here, is that in this case,

d( )
= (x( ))dW ( ) + J 1 dM( ), (t) = 1,
( )
which clearly generalizes what stated in the previous subsection.

Finally, we have:
du
u
= (u + vJ u ) d + u dW + J u (dZ vd )

+ J u dZ
= u + vQ J u u d + u dW
+ J u dZ.
= (u + v (1 + J ) J u u ) d + u dW
Finally, by the Q-martingale property of the discounted u,

u r = u vQ Ex (J u ) = u v Ex {(1 + J ) J u } ,
where Ex is taken with respect to the jump-size distribution, which is the same under Q and P .
139
c
by
A. Mele
Proof that v Q = v (1 + J ). As usual, the state-price density has to be a P -martingale in order

to be able to price bonds (in addition to all other assets). In addition, clearly depends on W and
Z. Therefore, it satises:
d( )
= (x( ))dW ( ) + J (x, ) dM( ), (t) = 1.
( )
We wish to nd vQ in dZ = dZ vQ d such that Z is a Q-martingale, viz
) = E[Z(T
)],
Z(
i.e.,
E(Z(t))
=

)
E (T ) Z(T
(t)
i.e.,
= E[(T )Z(T
)],
= Z(t)
(t)Z(t)
is a P -martingale.
(t)Z(t)
By Itos lemma,
= d Z + dZ + d dZ
d( Z)

= d Z + dZ v Q d + d dZ

Q
= d Z + [dZ
vd + v v d ] + d dZ

dM

= d Z + dM + v vQ d + d dZ.
Because , M and Z are P -martingales,

" T
"

Q
T , 0 = E
( ) v( ) v ( ) d +
t
d( ) dZ( ) .
But

d dZ = (dW + J dM) dZ vQ d = [dW + J (dZ vd )] dZ v Q d ,
and since (dZ)2 = dZ,
= J v d ,
E(d dZ)
and the previous condition collapses to:

" T

Q
T , 0 = E
( ) v( ) v ( ) + J (x)v( ) d ,
t
which implies
vQ ( ) = v( ) (1 + J (x)) , a.s.

140
c
by
A. Mele
References
Cvitanic, J. and I. Karatzas (1992): Convex Duality in Constrained Portfolio Optimization.
Annals of Applied Probability 2, 767-818.
Harrison, J.M. and S. Pliska (1983): A Stochastic Calculus Model of Continuous Trading:
Complete Markets. Stochastic Processes and Their Applications 15, 313-316.
He, H. and N. Pearson (1991): Consumption and Portfolio Policies with Incomplete Markets
and Short-Sales Constraints: The Innite Dimensional Case. Journal of Economic Theory
54, 259-304.
Follmer, H. and M. Schweizer (1991): Hedging of Contingent Claims under Incomplete Information. In: Davis, M. and R. Elliott (Editors): Applied Stochastic Analysis. New York:
Gordon & Breach, 389-414.
Karatzas, I. and S.E. Shreve (1991): Brownian Motion and Stochastic Calculus. Springer Verlag, Berlin.
Shreve, S. (1991): A Control Theorists View of Asset Pricing. In: Davis, M. and R. Elliot
(Editors): Applied Stochastic Analysis. New York: Gordon & Breach, 415-445.
141
5
Taking models to data
5.1 Introduction
This chapter surveys methods to estimate and test dynamic models of asset prices. It begins
with foundational issues on identication, specication and testing. Then, it surveys classical
estimation and testing methodologies such as the Method of Moments, in which the number of
moment conditions equals the dimension of the parameter vector (Pearson (1894)); Maximum
Likelihood (ML) (Gauss (1816), Fisher (1912)); the Generalized Method of Moments (GMM),
in which the number of moment conditions exceeds the dimension of the parameter vector,
and thus leads to minimum chi-squared (Neyman and Pearson (1928), Hansen (1982)); and
nally the recent developments based on simulations, which aim at implementing ML and
GMM estimation for models that are analytically quite complex - but that can be simulated.
The chapter concludes with an illustration of how asset pricing can help achieve asymptotic
eciency in the estimation of dynamic models. The chapter emphasizes the asymptotic theory,
that is, what happens when the sample size is large).
5.2 Data generating processes

5.2.1
Basics
Given is a multidimensional stochastic process yt , a data generating process (DGP). While we

do not know the probability distribution that has generated yt , we use the available data to
get insights into its nature. Let us begin with a few denitions. A DGP is characterized by a
conditional law, say the law of yt given the set of past values yt1 = {yt1 , yt2 , }, and some
exogenous variable z, with zt = {zt , zt1 , zt2 , },
DGP : 0 ( yt | xt ),
where xt = (yt1 , zt ), and 0 denotes the conditional density of the data, the true law. Then,
we have three basic denitions. First, we dene a parametric model as a set of conditional laws
for yt , indexed by a parameter vector Rp ,
(M ) = { ( yt | xt ; ) , Rp } .
5.2. Data generating processes
c
by
A. Mele
Second, we say that the model (M ) is well-specified if,

0 : ( yt | xt ; 0 ) = 0 ( yt | xt ) .
Third, we say that the model (M) is identifiable if 0 is unique. The main concern in this
chapter is to draw inference about the true parameter 0 , given the observations.
5.2.2
Restrictions on the DGP
The previous denition of the DGP is very rich. In practice, it is extremely dicult to cope
with DGP dened at such a level of generality. In this chapter, we only describe estimation
methods applying to DGPs satisfying some restrictions. There are two fundamental restrictions
usually imposed on the DGP.
Restrictions related to the heterogeneity of the stochastic process - which pave the way
to the concept of stationarity.
Restrictions related to the memory of the stochastic process - which pave the way to the
concept of ergodicity.
5.2.2.1 Stationarity
Stationary processes describe phenomena that approach a sort of long run equilibrium in a statistical sense: as time unfolds, the probability generating the observations settles down to some
long-run probability density, a time invariant probability. In the early 1980s, economy theory even dened a long-run equilibrium as a well-dened stationary, or invariant probability
distribution generating economic outcomes. We have two notions of stationarity.
Strong, or strict, stationarity. Denition: Homogeneity in law.
Weak stationarity, or stationarity of order p. Denition: Homogeneity in moments.
Even with stationary DGP, we might encounter situations where the number of parameters
to be estimated increases with the sample size. For example, consider two stochastic processes:
(i) one, for which cov(yt , yt+ ) = 2 , and (ii) and another, for which cov(yt , yt+ ) = exp ( | |).
In both cases, the DGP is stationary. Yet for the rst process, the dependence increases with
, and for the second, the dependence decreases with . This issue relates to the memory
of the process: as this simple example reveals, a stationary stochastic process may have long
memory. Ergodicity further restricts DGP, so as to make this memory play a more limited
role.
5.2.2.2 Ergodicity
We shall consider situations in which the dependence between yt1 and yt2 decreases with |t2 t1 |.
Lets introduce some concepts and notation. Two events A and B are independent when P (A
B) = P (A)P (B). A stochastic process is asymptotically independent if, for some function ,
|F (yt1 , , ytn , yt1 + , , ytn + ) F (yt1 , , ytn ) F (yt1 + , , ytn + )| ,
we also have that lim 0. A stochastic process is p-dependent if = 0 p < . A
stochastic process is asymptotically uncorrelated if there exists such that for all t,
143
c
by
A. Mele
5.2. Data generating processes

cov(yt , yt+ )/ var(yt ) var(yt+ ), and that 0 1 with
=0 < . For example,
(1+)
=
, > 0, in which case 0 as .
Let Bt1 denote the -algebra generated by {y1 , , yt } and A Bt, B B
t+ , and dene:
= sup |P (A B) P (A)P (B)| ,
( ) = sup |P (B | A) P (B)| , P (A) > 0.
We say that (i) y is strongly mixing, or -mixing if lim 0; (ii) y is uniformly mixing
if lim 0. Clearly, a uniformly mixing
process is also strongly mixing. A second order
T
stationary process is ergodic if limT =1 cov (yt , yt+ ) < . If a second order stationary
process is strongly mixing, it is also ergodic.
5.2.3
Parameter estimators
Consider an estimator of the parameter vector of the model,

(M ) = {( yt | xt ; ), Rp } .
Naturally, any estimator does necessarily depend on the sample size, which we write as T
tT (y). Of a given estimator T , we say that it is:
Correct (or unbiased ), if E( T ) = 0 . The dierence E(T ) 0 is the distortion, or bias.
a.s.
Weakly consistent if plimT = 0 . And strongly consistent if T 0 .
(1)
(2)
Finally, an estimator T is more efficient than another estimator T if, for any vector of
(1)
(2)
constants c, we have that c var( ) c < c var( ) c.
T
5.2.4
Basic properties of density functions
We have T observations yT1 = {y1 , , yT }. Suppose these observations are the realization
of a T -dimensional random variable with joint density, f (
y1 , , yT ; ) = f y1T ; . We have
1
momentarily put the tilde on yi , to emphasize that we view them as random variables.
How#
we write yi instead of yi , from now on. By construction, f ( y| ) dy
notation,
#ever,#to ease

f y1T dy1T = 1 or,
"
,
f (y; ) dy = 1.
Now suppose that the support of y doesnt depend on . Under regularity conditions,
"
"
f (y; ) dy = f (y; ) dy = 0p ,
where 0p is a column vector of zeros in Rp . Moreover, for all ,
"
0p = f (y; ) dy = E [ ln f (y; )] .
(5.1)
1 Therefore, we follow a classical perspective. A Bayesian statistician would view the sample as given. We do not review Bayesian
methods in this chapter.
144
c
by
A. Mele
5.3. Maximum likelihood estimation

Finally we have,
"
0pp = [ ln f (y; )] f (y; ) dy

"
"
= [ ln f (y; )] f (y; ) dy + | ln f (y; )|2 f (y; ) dy,
where |x|2 denotes the outer product, i.e. |x|2 = x x . Hence, by Eq. (5.1),
E [ ln f (y; )] = E | ln f (y; )|2 = var [ ln f (y; )] J (), .

The matrix J is known as the Fishers information matrix.
5.2.5
The Cramer-Rao lower bound
Let t(y) some unbiased estimator of . Let p = 1. We have,

"
E [t(y)] = t(y)f (y; ) dy.
Under regularity conditions,

"
E [t(y)] = t(y) [ ln f (y; )] f (y; ) dy = cov (t(y), ln f (y; )) .
By the basic inequality, cov (x, y)2 var (x) var (y),
[cov (t(y), ln f (y; ))]2 var [t(y)] var [ ln f (y; )] .
Therefore,
[ E (t(y))]2 var [t(y)] var [ ln f (y; )] = var [t(y)] E [ ln f (y; )] .
But if t(y) is unbiased, or E [t(y)] = ,
var [t(y)] [E ( ln f (y; ))]1 J ()1 .
This is the celebrated Cramer-Rao bound. The same results holds with a change in notation
in the multidimensional case (see, e.g., Amemiya (1985, p. 14-17)).
5.3 Maximum likelihood estimation

5.3.1
Basics

The density of the data, f( y1T ), maps every possible sample and parameter values of on
to positive numbers, the likelihood of occurence of a any given sample, given the parameter
: RnT R+ . We trace the joint density of the entire sample through a thought experiment,
in which we change the sample y1T . We ask: Which value of makes the sample we observed
the most likely to have occurred? We introduce the likelihood function, L( | y1T ) f (y1T ; ).
It is the function f (y; ) for y1T given and equal to y, say:
L( | y) f (
y ; ).
Then, we maximize L( | y1T ) with respect to : that is, we look for the value of , which
maximizes the probability to observe the sample we have eectively observed. The resulting
estimator is called maximum likelihood estimator (MLE). If the model is not misspecied, the
MLE attains the Cramer-Rao lower bound, as we shall see.
145
c
by
A. Mele

5.3.2
Factorizations
Consider a series of events {Ai }. In the Appendix, we show that,

$
%
n
1
n
i1
(
(

Pr
Ai =
Pr Ai Aj .
j=1
i=1
i=1
(5.2)
Next, consider the previous denition of the MLE. Given Eq. (5.2), we have that the MLE
satises:

1
T = arg max LT () = arg max
ln LT () ,
T
where
ln LT () ln
T
T
T
T
1

f yt y1t1 ; =
ln f yt y1t1 ;
ln f (yt ; )
t (),
t=1
t=1
t=1
t=1
and t () is the log-likelihood of a single observation.

5.3.3
Asymptotic properties
We consider the i.i.d. case only. Moreover, we provide heuristic arguments, leaving more rigorous
proofs and general results in the Appendix.
5.3.3.1 The limiting problem
The MLE satises the following rst order conditions,

0p = ln LT ()|=T ln LT (T ).
Consider a Taylor expansion of the rst order conditions around 0 ,
d
0p = ln LT ()|=T = ln LT (0 ) + ln LT (0 )(T 0 ),
(5.3)
where the notation xT = yT means that the dierence xT yT = op (1). Here 0 is solution to
the limiting problem,

1
0 = arg max lim
ln LT ()
= arg max [E (())] ,
T T
where L satises all of the regularity conditions needed to ensure that,

0 : E [ (0 )] = 0p .
To show that this is indeed the solution, suppose 0 is identied; that is, = 0 and , 0
f ( y| ) = f( y| 0 ). Suppose, further, that for each , E [ln f ( y| )] < . Then, we
have that 0 = arg max E [ln f ( y| )], and such a value of is unique. The proof is, indeed,
very simple. We have,

f ( y| )
f ( y| )
E0 ln
> ln E0
f ( y| 0 )
f( y| 0 )
"
f (y| )
= ln
f( y| 0 )dy
f ( y| 0 )
"
= ln f ( y| )dy = 0.
146
c
by
A. Mele

5.3.3.2 Consistency and asymptotic normality
p
a.s.
Under regularity conditions, T 0 and even T 0 if the model is well-specied. One
example of conditions required to obtain weak consistency is that the following UWLLN (Uniform weak law of large numbers) holds,
lim Pr sup |T () E ( ())| 0.
Next, consider again the asymptotic expansion in Eq. (5.3), which can be elaborated, so as to
have,

1
1
1
d
ln LT ( 0 )
T (T 0 ) =
ln LT (0 )
T
T
.
1
T
T
1
1
=
t (0 )
t (0 ).
T t=1
T t=1
By the law of large numbers reviewed in the Appendix (weak law no. 1),
T
1
p
t (0 ) E0 [ t (0 )] = J (0 ) .
T t=1
Therefore, asymptotically,
T
d
1 1
T (T 0 ) = J (0 )
t (0 ).
T t=1
We also have,
T
1
d
t (0 ) N (0, J ( 0 )) .
T t=1
Indeed, by the central limit theorem reviewed in the Appendix, let (0 )T =

and note that E ( t (0 )) = 0. Then,

T
T
(
)
E
(
(
))
0 T
t 0
( )
1
t=1 t 0 =

,
T var [ t (0 )]
var [ t ( 0 )]
where, for each t, var [ t (0 )] = J (0 ).

Finally, by the Slutzkys theorem reviewed in the Appendix,

d
T (T 0 ) N 0, J (0 )1 .
Therefore, the ML estimator attains the Cramer-Rao bound.

147
1
T
T
t=1
t ( 0 ),
c
by
A. Mele
5.4. M-estimators
5.4 M-estimators
Consider a function g of the unknown parameters . Given a function , a M-estimator of the
function g() is the solution to,
max
gG
T
(xt , yt ; g) ,
t=1
where y and x are as in Section 5.2.1. We assume that a solution to this problem exists, that it
is interior and that it is unique. Let us denote the M-estimator with gT (xT1 , y1T ). Naturally, the
M-estimator satises the following rst order conditions,
T

1
g yt , xt ; gT (xT1 , y1T ) .
0=
T t=1
To simplify the presentation, we assume that (x, y) are independent in time, and that they have
the same law. By the law of large numbers,
""
""
T
1
p
(yt , xt ; g)
(y, x; g) dF (x, y) =
(y, x; g) dF (y| x) dZ (x) Ex E0 [ (y, x; g)] ,
T t=1
where E0 is the expectation operator taken with respect to the true conditional law of y given
x and Ex is the expectation operator taken with respect to the true marginal law of x. The
limit problem is,
g = g (0 ) = arg max Ex E0 [ (y, x; g)] .
gG
Under standard regularity conditions,2 there exists a sequence of M-estimators gT (x, y) converging a.s. to g = g (0 ). Under some additional regularity conditions, the M-estimator is
asymptotic normal:

Theorem 5.1: Let I Ex E0 g (y, x; g ( 0 )) [g (y, x; g (0 ))]

and assume that the
matrix J Ex E0 [gg (y, x; g)] exists and has an inverse. We have,

d
T (
gT g (0 )) N 0, J 1 IJ 1 .
Sketch of the proof. The M-estimator satises the following rst order conditions,
T
1
0 =
g (yt , xt ; gT )
T t=1
.
T
T

1
1
d
=
g (yt , xt ; g ) + T
gg (yt , xt ; g ) (
gT g ) .
T t=1
T t=1
2G
is compact; is continuous with respect to g and integrable with respect to the true law, for each g;
Ex E0 [ (y, x; g)] uniformly on G; the limit problem has a unique solution g = g (0 ).
148
1
T
t=1
a.s.
(yt , xt ; g)
c
by
A. Mele
5.5. Pseudo (or quasi) maximum likelihood

By rearranging terms,
.1 .
T
T

1
1
d
T (
gT g ) =
gg (yt , xt ; g )
g (yt , xt ; g)
T t=1
T t=1
T
1
d
= [Ex E0 (gg (y, x; g))]1
g (yt , xt ; g)
T t=1
d
=J
T
1
g (yt , xt ; g ) .
T t=1

By the limiting problem, Ex E0 [g (y, x; g )] = 0. Since var (g ) = E g [g ] = I,
then,
T
1
d
g (yt , xt ; g ) N (0, I) .
T t=1
The result follows by the Slutzkys theorem and the symmetry of J .
One simple example of M-estimator is the Nonlinear Least Square estimator,

T = arg min
T

t=1
[yt m (xt ; )]2 ,
for some function m. In this case, (x, y; ) = [y m (x; )]2 .
5.5 Pseudo (or quasi) maximum likelihood

The maximum likelihood estimator is a M-estimator: set = ln L, the log-likelihood function.
Assume that the model is well-specied, in which case J = I, which conrms we are back to
the MLE.
Next suppose that we implement the MLE to estimate a model, when in fact the model
is misspecified : the true DGP 0 ( yt | xt ) doesnt belong to the family of laws spanned by our
model,
0 ( yt | xt )
/ (M) = {f ( yt | xt ; ) , } .

So suppose we insist in maximizing = ln L, where L = t f ( yt | xt ; ). In this case,

d
T (T 0 ) N 0, J 1 IJ 1 ,
where 0 is the pseudo-true value,3 and

!
!
J = Ex E0 ln f ( yt | yt1 ; 0 ) , I = ExE0 ln f( yt | yt1 ; 0 ) ln f (yt | yt1 ; 0 )

.
In the presence of specication errors, J = I. By comparing the two estimated matrices
leads to detect specication errors. Finally, note that in this general case, the variance-covariance
3 That is, is, clearly, the solution to some misspecied limiting problem. It is possible to show that has an appealing
0
0
interpretation in terms of some entropy distance minimizer.
149
c
by
A. Mele
5.6. GMM
matrix J 1 IJ 1 depends on the unknown law of (yt , xt ). To assess the precision of the estimates
of gT , one needs to estimate such a variance-covariance matrix. A common practice is to use
the following a.s. consistent estimators,
1
J =
T
T

t=1
gg (yt , xt ; gT ),
1
and I =
T
T

g (yt , xt ; gT ) g (yt , xt ; gT ) .
t=1
5.6 GMM
Economic theory often places restrictions on models that have the following format,
E [h (yt ; 0 )] = 0q ,
(5.4)
where h : Rn Rq , 0 is the true parameter vector, yt is the n-dimensional vector of

the observable variables and Rp . The MLE is often unfeasible here. Moreover, it requires
specifying a density function. Hansen (1982) proposed the following Generalized Method of
Moments (GMM) estimation procedure. Consider the sample counterpart to the population in
Eq. (5.4),
T

h y T ; = 1
h (yt ; ) ,
1
T t=1
where we have rewritten h asa function
of the parameter vector . The basic idea of GMM

is to nd a which makes h y1 ; as close as possible to zero. Precisely, we have,

Definition (GMM estimator): The GMM estimator is the sequence T satisfying,

T = arg min h
y T ; WT h
yT ; ,
1
1
p
R
1q
qq
q1
where {WT } is a sequence of weighting matrices, with elements that may depend on the observations.
The simplest situation arises in the so-called just-identified case p = q, in which case, the
GMM estimator is simply:
T : h(y
T ; T ) = 0q .
1
In general, p q. If p < q, we say that the GMM estimator imposes overidentifying restrictions.
We analyze the i.i.d. case only. Under regularity conditions, there exists a matrix WT that
minimizes the asymptotic variance of the GMM estimator, which satises asymptotically,
W =

!1
T ; T ) h(y
T ; T )
lim T E h(y
1
1
1
0 .
An estimator of 0 can be:

T
!
1
h(yt ; T ) h(yt ; T ) .
T =
T t=1
150
c
by
A. Mele
5.6. GMM
Note that T depends on the weighting matrix T and viceversa. Therefore, we need to implement an iterative procedure. The more one iterates, the less the nal results will depend on the
(0)
(0)
initial weighting matrix T . For example, one can start with T = Iq .
We have:
p
Theorem 5.2: Suppose to be given a sequence of GMM estimators T such that: T 0 . We

have,

!1
T ( T 0 ) N 0p , E (h ) 0 E (h )
, where h h(y; 0 ).
p
Sketch of the proof: The assumption that T 0 is easy to check under mild regularity
conditions. Moreover, the GMM satises,
T ; T )1 h(y
T ; T ).
0p = h(y
1
1
T
pq
qq
(5.5)
q1
T ; T ) = 0. This is so because if
Eq. (5.5) conrms that if p = q the GMM satises T : h(y
1
1
= 0. In the general
p = q, then hT is full-rank, and Eq. (5.5) can only be satised with h
case, q > p, we have,

T ; T ) = T h
y T ; 0 + h
y T ; 0 T (T 0 ) + op (1).
T h(y
1
1
1
q1
q1
qp
T ; T )1 ,
By premultiplying both sides of the previous equality by h(y
1
T
1T ; T )1 h(y
1T ; T )
T h(y
T

1T ; T )1 h
1T ; T )1 h
y1T ; 0 T ( T 0 ) + op (1).
y1T ; 0 + h(y
=
T h(y
T
T
The l.h.s. of this equality is zero by the rst order conditions in Eq. (5.5). By rearranging
terms,

T

1

d
y1T ; 0 1 h
y1T ; 0
1T ; T )1 T h
y1 ; 0
T (T 0 ) = h
h(y
T
T
1
T
T
T
T

1
1
1 1
y ; 0
=
h(yt ; T )T
[ h(yt ; T )]
h(yt ; T )1
Th
1
T
T t=1
T t=1
T t=1
T

1
1
d
= E (h ) 1
E
(h
)
E
(h
)
h (yt ; 0 ) .
0
0
T t=1
Next, consider the term
1
T
T
h (yt ; 0 ), which satises a central limit theorem:
t=1
T
1
d
h (yt ; 0 ) N (E(h), var(h)) ,

T t=1

where, by Eq. (5.4), E(h) = 0, and var(h) = E h h = 0 . Then, we have:
T
1
d
h (yt ; 0 ) N (0, 0 ) .
T t=1
151
c
by
A. Mele
5.6. GMM
Therefore,
T ( T 0 ) is asymptotic normal with expectation 0p , and variance,

1

1
1
1
E (h ) 1
E (h ) 1
E (h ) 1
= E (h ) 1
.
0 E (h )
0 0 0 E (h )
0 E (h )
0 E (h )
A global specication test is that of the celebrated overidentifying restrictions. First, consider the behavior of the statistic,
T 1 T d 2
y ; 0
y ; 0 (q).
Th
Th
1
0
1
One might be led to think that the same result applies when 0 is replaced with T (which is a
consistent estimator of 0 ). Wrong. Consider,
T ; T ) 1 T h(y
T ; T ).
CT = T h(y
1
1
T
We have,
T

d
T ; T ) =
y ; 0 + h
y T ; 0 T (T 0 )
T h(y
T
h
1
1
1
!1
T
T

d T
1
y1 ; 0
y1 ; 0
= T h y1 ; 0 h
E (h ) 1
E
(h
)
E
(h
)
Th
0
0
!1
T

d T
1
y ; 0
= T h y1 ; 0 E (h ) E (h ) 0 E (h )
E (h ) 1
Th
0
1
T
y ; 0 ,
= (Iq P) T h
1
qq
q1
and
Pq E (h ) E (h ) 1
0 E (h )
!1
E (h ) 1
0
is the orthogonal projector in the space generated by the columns of E (h ) by the inner product
1
0 . We have thus shown that,
T

d T
y ; 0 .
CT = T h
y1 ; 0 (Iq Pq ) 1
(I
P
)
Th
q
1
T
But,
and by a classical result,
T d
y ; 0 N (0, 0 ) ,
Th
1
d
CT 2 (q p) .
Hansen and Singleton (1982, 1983) started the literature on the estimation and testing of dynamic asset pricing models within a fully articulated rational expectations framework. Consider
the classical system of Euler equations arising in the Lucas tree,

u (ct+1 )
E
(1 + ri,t+1 ) 1 Ft = 0, i = 1, , m,
u (ct )
where u is the utility function, ri is the return on asset i, is the time-discount factor, Ft is
the information set as of time t, and m is the number of assets. Consider the CRRA utility
152
c
by
A. Mele
5.7. Simulation-based estimators
function, u(x) = x1 / (1 ). If the model is well-specied, there exist some 0 and 0 such
that:
.
-
0

ct+1

(1 + ri,t+1 ) 1 Ft = 0, i = 1, , m.
E 0

ct
To sumup, the dimension of the parameter vector is p = 2. To estimate the true parameter
vector 0 ( 0 , 0 ), we may build up a system of orthogonality conditions. This system can
be based on projecting observable variables predicted by the model onto other variables, which
we call instruments, and that are included in the information set Ft ,
E [h (yt ; 0 )] = 0,
where, for some vector of n instruments, say, Instrt = [i1,t , , in,t ] ,

ct+1
(1 + r1,t+1 ) 1 Instrt
ct
..
h (yt ; ) =
.

mn

ct+1
(1 + rm,t+1 ) 1 Instrt
ct
The instruments used to produce the orthogonality restrictions, may include constants, past
values of consumption growth, ct+1
, past returns, ri .
ct
5.7 Simulation-based estimators

Ideally, MLE should be the preferred estimation method of parametric Markov models, as it
leads to rst-order eciency. Yet economic theory places restrictions that make these models
impossible to estimate through maximum ML. Instead, GMM is a natural estimation method.
But GMM can be unfeasible in situations of interest. For example, assume that the data generating process is not i.i.d. Instead, data are generated by the transition function,
yt+1 = H (yt , t+1 , 0 ) ,
(5.6)
where H : Rn Rd Rn , and t is a vector of i.i.d. disturbances in Rd . Assume the

econometrician knows the function H. Let zt = (yt , yt1 , , ytl+1 ), l < . In many cases of
can be written as,
interest, the function h
T

h y T ; = 1
[f E (f (zt , ))] ,
1
T t=1 t
(5.7)
where,
ft = f (zt , 0 ) ,
is a vector-valued moment function, or observation function, a function that summarize satisfactorily the data, so to speak. The GMM estimator is unfeasible, if we are not able to compute
the expectation E (f (zt , )) in closed form, for each . Simulation-based methods can make the
method of moments feasible in such cases.
153
c
by
A. Mele

5.7.1
Three simulation-based estimators
The basic idea underlying simulation-based methods is quite simple. While the moment conditions are too complex to be evaluated analytically, the model in Eq. (5.6) can be simulated.
Accordingly, draw t from its distribution, and save the simulated values t . Compute recursively,

yt+1
= H yt , t+1 , ,
and create simulated moment functions as follows,

ft f zt , .
Consider the following parameter estimator,
T = arg min GT () WT GT () ,
(5.8)
in Eq. (5.7),
where GT () is the simulated counterpart to h
$
%
T
S(T
)

1
1
GT () =
ft
fs ,
T t=1
S (T ) s=1
and S (T ) is the simulated sample size, which is written as a function of the sample size T , for
the purpose of the asymptotic theory.
The estimator T , also known as the Simulated Method of Moments (SMM) estimator, aims to
match the sample properties of the actual and simulated processes ft and ft . It was introduced
in a series of works, by McFadden (1989), Pakes and Pollard (1989), Lee and Ingram (1991)
and Due and Singleton (1993). The simulated pseudo-maximum likelihood method of Laroque
and Salanie (1989, 1993, 1994) can also be interpreted as a SMM estimator.
Another approach to statistical inference based on simulations relies on the indirect inference
principle (IIP), and was initiated by Gourieroux, Monfort and Renault (1993) and Smith (1993).
A IIP-based estimator works as follows. Instead of minimizing the distance of some moment
conditions, the IIP relies on minimizing the parameters of an auxiliary, possibly misspecied
model. For example, consider the following auxiliary parameter estimator,

(5.9)
T = arg max ln L y1T ; ,
where L is the likelihood of some (possibly misspecied) model. Consider simulating S times
the process yt in Eq. (5.6), and computing,
sT () = arg max ln L(ys ()T1 ; ),
s = 1, , S,
where ys ()T1 = (yt,s )Tt=1 are the simulated variables (for s = 1, , S) when the parameter vector
is . The IIP-based estimator is dened similarly as T in Eq. (5.8), but with the function GT
given by,
S
1 s
GT () = T
() .
(5.10)
S s=1 T
The diagram in Figure 5.1 illustrates the main ideas underlying the IIP.
154
c
by
A. Mele
Model
Model-simulated data
y t = f ( y t 1 , t ; )
~
y ( ) = ( ~
y1 ( ), L , ~y T ( ))
Estimation of an
auxiliary model on
model-simulated data
Auxiliary
parameter estimates
~
bT ( )
Sss
Auxiliary
parameter estimates
y = ( y1 , L , y T )
bT
Observed data
Indirect Inference Estimator
Estimation of the
same auxiliary model
on observed data
T arg min bT ( ) bT
FIGURE 5.1. The Indirect Inference principle. Given the true model yt = f (yt1 , t ; ), an estimator of based on the indirect inference principle ( T say) makes the parameters of some auxiliary
model bT (T ) as close as possible to the: parameters:bT of the same auxiliary model estimated on the
:
:
observations. That is, T = arg min :bT () bT : , for some norm .
Finally, Gallant and Tauchen (1996) propose a simulation-based estimation method they
label efficient method of moments (EMM). Their estimator sets,
GT () =
N

1
; T mT (, T ),
ln f yn zn1
N n=1
where
ln f ( y| z; ) is the score of some auxiliary model f , T is the Pseudo ML estimator
of the auxiliary model, and (yn )N
n=1 is a long simulation (i.e. N is very large) of Eq. (5.6), with
parameter vector set equal to . Finally, the weighting matrix WT in Eq. (5.8) is taken to be
any matrix IT1 converging in probability to:

I = E
ln f ( y2 | z1 ; ) .
(5.11)
To motivate this choice of GT (), note that the auxiliary score,

the following rst order conditions:
T
1
ln f ( yt | zt1 ; T ) = 0,
T n=1
which is the sample equivalent of

ln f ( y2 | z1 ; ) = 0,
E
155

ln f ( yt | zt1 ; T ), satises
c
by
A. Mele
for some . Likewise, we must have that with = 0 , mT (0 , T ) = 0, for large N. Moreover,
by the theory of pseudo maximum likelihood reviewed in Section 5.5, we know that under
regularity conditions,

d
T ( T ) N 0, J 1 IJ 1 ,
(5.12)
2
where I is as in Eq. (5.11) and J = E(

ln f ( y2 | y1 ; )). Moreover, for large N ,
d
d
T mT (0 , T ) = J T ( T ) N (0, I) ,
where the last convergence follows by (5.12). By the usual MM theory, then,

1
d
1
T ( T ) N 0, D0 0 D0
,
where D0 = m and 0 =
5.7.2
j= E
ln f ( yt | zt1 ; ) ,
ln f ( ytj | ztj1 ; ) .
Asymptotic normality
We show, heuristically, how asymptotic normality obtains for the three estimators of Section
5.7.1.
5.7.2.1 SMM
Let,
0 =
j=
(ft
(ft ))
ftj
ftj
!
and suppose that

p
WT W0 = 1
0 .
We now demonstrate that under this condition, as T and S (T ) ,

1 1
d
T (T 0 ) N 0p , (1 + ) D0 0 D0
,
(5.13)

T
0
where = limT S(T
, D0 = E ( G (0 )) = E f
, and the notation G means that
)
G. is drawn from its stationary distribution.
Indeed, the rst order conditions satised by the SMM in Eq. (5.8) are,
0p = [ GT ( T )] WT GT (T ) = [ GT (T )] WT [GT (0 ) + GT (0 ) (T 0 )] + op (1) .
That is,

1
d
T (T 0 ) = [ GT (T )] WT GT (0 )
[ GT (T )] WT T GT (0 )

1
d
= D0 W0 D0
D0 W0 T GT (T )

1 1
= D0 1
D0 0 T GT ( 0 ) .
(5.14)
0 D0
156
c
by
A. Mele

We have,
$
%
T
S(T
)

1
1
ft
T GT (0 ) =
T
f 0
T t=1
S (T ) s=1 s
S(T )
T

0
1
1
T
=

(ft E (f ))
fs0 E f
T t=1
S (T )
S (T ) s=1
d
N (0, (1 + ) 0 ) ,

0
. By replacing this result into Eq. (5.14)
where we have used the fact that E (f
) = E f
T
produces the convergence in Eq. (5.13). If = limT S(T
= 0 (i.e. if the number of simula)
tions grows more fastly than the sample size), the SMM estimator is as ecient as the GMM
T
< .
estimator. Moreover, inspection of (5.13) reveals that we must have = limT S(T
)
This condition means that the number of simulations S (T ) can not grow more slowly than the
sample size.
5.7.2.2 Indirect inference
The IIP-based estimator works slightly dierently. For this estimator, even if the number of
simulations S is xed, asymptotic normality obtains without the need to impose that S goes
to innity more fastly than the sample size. Basically, what really matters here is that ST goes
to innity.
By Eq. (5.14), and the discussion in Section 5.7.1, we know that asymptotically, the rst
order conditions satised by the IIP-based estimator are,

1
d
T (T 0 ) = D0 W0 D0
D0 W0 T GT (0 ) ,
where GT is as in Eq. (5.10), D0 = b (), and b () is solution to the limiting problem

corresponding to the estimator in (5.9), viz,

T
1
ln L y1 ; .
b () = arg max lim
T T
We need to nd the distribution of GT in Eq. (5.10). We have,

S
1
T GT (0 ) =
T ( T sT ( 0 ))
S s=1
=
=
S
1
T [( T 0 ) ( sT (0 ) 0 )]
S s=1
1
T ( T 0 )
T ( sT ( 0 ) 0 ) ,
S s=1
where 0 = b (0 ). Hence, given the independence of the sample and the simulations,

1
d
T GT (0 ) N 0, 1 +
Asy.Var
T T
.
S
That is, asymptotically S can be xed with respect to T .

157
c
by
A. Mele

5.7.2.3 Ecient method of moments
We have,
T = arg min mT (, T ) WT mT (, T ) ,
mT (, T ) =
The rst order conditions are:
N

1
ln f yn zn1
; T .
N n=1
0 = mT (T , T ) WT mT (T , T )
d
= mT (0 , T ) WT (mT (0 , T ) + mT (0 , T ) (T 0 )) ,
or

1
d
T (T 0 ) = mT ( 0 , T ) WT mT (0 , T )
mT (0 , T ) WT T mT (0 , T ) .
We have, for some ,
d
d
T mT (0 , T ) = JT T ( T ) N (0, I) .
Hence,
where,
d
T (T 0 ) N (0, V ) ,

1

1
V = m W m
m W IW m m W m
.
With W = I 1 , this variance collapses to,

1
.
V = m I 1 m
5.7.3
(5.15)
Advances
The three estimators that we have examined in Sections 5.7.1-5.7.2, are general-purpose, but
in general, they do not lead to to asymptotic eciency unless the true score belongs to the
span of the moment conditions, as we shall explain in Section 5.8. There exist other simulationbased methods, which aim to approximate the likelihood function through simulations (e.g.,
Lee (1995), Hajivassiliou and McFadden (1998)). While these methods lead to asymptotically
ecient estimators, they address specic estimation problems.
There exist estimators that are both general purpose and that can lead to asymptotic eciency. Fermanian and Salanie (2004) consider an estimator that relies on approximating the
likelihood function through kernel estimates obtained simulating the model of interest. Carrasco,
Chernov, Florens and Ghysels (2007) rely on a continuum of moment conditions matching
model-based (simulated) characteristic functions to data-based characteristic functions. Altissimo and Mele (2009) propose an estimator that minimizes a certain distance between conditional densities on the observations and the conditional densities from data simulated from
the model, both estimated through kernel methods.
158
c
by
A. Mele
5.8. Spanning scores
5.8 Spanning scores

This section provides a heuristic discussion about the conditions under which the EMM achieves
the Cramer-Rao lower bound. Consider the following denition, which is similar to that in
Tauchen (1997). Of a given span of moment conditions sf , say that of the EMM, we say that
it also spans the true score if,
var ( s| sf ) = 0,
(5.16)
where s denotes the true score. From Eq. (5.15), we know that the asymptotic variance of the
EMM, say varEMM , satises:
1
V 1 = m var (sf )1 m.
varEMM
By the linear projection,

s = Bsf + ,
B = cov (s, sf ) var (sf )1 ,
we have,
1
varMLE
= var (s) = Bvar (sf ) B + var ( s| sf ) = cov (s, sf ) var (sf )1 cov (s, sf ) + var ( s| sf ) ,
(5.17)
where varMLE denotes the asymptotic variance of the MLE. We claim that:
cov (s, sf ) = m.
(5.18)
Indeed,

"
m (0 , ) =
ln f (y; ) p (y, ) dy
=0
"
ln f (y; ) p (y, 0 ) dy
=

"

ln f (y; )
ln p (y, 0 ) p (y, 0 ) dy
=
= cov (s, sf ) ,
where p (y, ) is the true density. Next, replace Eq. (5.18) into Eq. (5.17),
1
1
varMLE
= m var (sf )1 m + var ( s| sf ) = varEMM
+ var ( s| sf ) .
Therefore, the EMM estimator achieves the Cramer-Rao lower bound under the spanning condition in Eq. (5.16).
5.9 Asset pricing, prediction functions, and statistical inference

We develop conditions, which ensure the feasibility of the estimation methods when applied
to general equilibrium models, that is as soon as an unobservable multidimensional process is
estimated in conjunction with predictions functions suggested by standard asset pricing theories.
We assume that the data generating process is a multidimensional partially observed diusion
process solution to,
dy ( ) = b (y ( ) ; ) d + (y ( ) ; ) dW ( ) ,
(5.19)
159
c
by
A. Mele
5.9. Asset pricing, prediction functions, and statistical inference
where W is a multidimensional process and (b, ) satisfy some regularity conditions we single
out below. This appendix analyzes situations in which the original partially observed system
(5.19) can be estimated by augmenting it with a number of observable deterministic functions
of the state. In many situations of interest, such deterministic functions are suggested by asset
pricing theories in a natural way. Typical examples include derivative asset price functions or
any deterministic function(als) of asset prices (e.g., asset returns, bond yields, implied volatility, etc.). The idea to use predictions of asset pricing theories to improve the t of models
with unobservable factors is not new (see, e.g., Christensen (1992), Pastorello, Renault and
Touzi (2000), Chernov and Ghysels (2000), Singleton (2001, sections 3.2 and 3.3)), and Pastorello, Patilea and Renault (2003). In this appendix, we provide a theoretical description of
the mechanism leading to eciency within the class of our estimators.
We consider a standard Markov pricing setting. For xed t 0, we let M be the expiration
date of a contingent claim with rational price process c = {c(y( ), M )} [t,M) , and let
{z(y( ))} [t,M] and (y) be the associated intermediate payo process and nal payo function,
respectively. Let / + L be the usual innitesimal generator of (5.19) taken under the riskneutral measure. In a frictionless economy without arbitrage opportunities, c is the solution to
the following partial dierential equation:

0=
+ L R c(y, M ) + z(y), (y, ) Y [t, M)
(5.20)
c(y, 0) = (y), y Y
where R R(y) is the short-term rate. We call prediction function any continuous and twice
dierentiable function c (y; M ) solution to the partial dierential equation (5.20).
We now augment system (5.19) with d q prediction functions. Precisely, we let:
C( ) (c (y( ), M1 ) , , c (y( ), Mdq )) ,
[t, M1 ]
where {Mi }dq

i=1 is an increasing sequence of xed maturity dates. Furthermore, we dene the
measurable vector valued function:
(y( ); , ) (y o ( ), C (y( ))) ,
[t, M1 ],
(, ) ,
where Rp is a compact parameter set containing additional parameters. These new parameters arise from the change of measure leading to the pricing model (5.20), and are now
part of our estimation problem.
We assume that the pricing model (5.20) is correctly specied. That is, all contingent claim
prices in the economy are taken to be generated by the prediction function c(y, M ) for some
(0 , 0 ) . For simplicity, we also consider a stylized situation in which all contingent
claims have the same contractual characteristics specied by C (z, ). More generally, one
may dene a series of classes of contingent claims {Cj }Jj=1 , where class of contingent claims j
has contractual characteristics specied by Cj (zj , j ).4 The number of prediction functions

that we would introduce in this case would be equal to d q = Jj=1 M j , where M j is the
number of prediction functions within class of assets j. To keep the presentation simple, we do
not consider such a more general situation here.
4 As
an example, assets belonging to class C1 can be European options; assets belonging to class C1 can be bonds; and so on.
160
c
by
A. Mele

(y; 0, 0)
Y
1(y; 0, 0)
FIGURE 5.2. Asset pricing, the Markov property, and statistical efficiency. Y is the domain on which
the partially observed primitive state process y (y o y u ) takes values, is the domain on which
the observed system (y o C(y)) takes values in Markovian economies, and C(y) is a contingent
claim price process in Rdq . Let c = (y o , c(y, 1 ), , c(y, dq )), where {c(y, j )}dq
j=1 forms an
intertemporal cohort of contingent claim prices, as in Denition 5.3. If the local restrictions of are
one-to-one and onto, statistical inference about and can be made, using information about the price
of derivative contracts, c . If is also globally invertible, statistical inference can lead to rst-order
asymptotic eciency, once conditioned upon c .
Our objective is to provide estimators of the parameter vector (0 , 0 ) under which observations were generated. In exactly the same spirit as for the estimators considered in the main
text, we want any estimator of ( 0 , 0 ) to make the nite dimensional distributions of implied
by model (5.19) and (5.20) as close as possible to their sample counterparts. Let Rd be the
domain on which takes values. As illustrated in Figure 5.2, our program is to move from the
unfeasible domain Y of the original state variables in y (observables and not) to the domain
on which all observable variables take value. Ideally, we would like to implement such a change
in domain in order to recover as much information as possible on the original unobserved process in (5.19). Clearly, is fully revealing whenever it is globally invertible. However, we will
show that estimation is feasibly even when is only locally one-to-one.
An important feature of the theory in this section is that it does not hinge upon the availability
of contingent prices data covering the same sample period covered by the observables in (5.19).
First, the price of a given contingent claim is typically not available for a long sample period.
As an example, available option data often include option prices with a life span smaller than
the usual sample span of the underlying asset prices; in contrast, it is common to observe long
time series of option prices having the same maturity. Second, the price of a single contingent
claim depends on time-to-maturity of the claim; therefore, it does not satisfy the stationarity
assumptions maintained in this paper. To address these issues, we deal with data on assets
having the same characteristics at each point in time. Precisely, consider the data generated by
the following random processes:
Definition 5.3. (Intertertemporal (, N)-cohort of contingent claim prices) Given a prediction
function c (y; M ) and a N -dimensional vector (1 , , N ) of fixed maturities, an
intertemporal (, N)-cohort of contingent claim prices is any collection of contingent claim
price processes c ( , ) (c(y( ), 1 ), , c(y( ), N )) ( 0) generated by the pricing model
(5.20).
Consider for example a sample realization of three-months at-the-money option prices, or
a sample realization of six-months zero-coupon bond prices. Long sequences such as the ones
in these examples are common to observe. If these sequences were generated by (5.20), as
161
c
by
A. Mele
in Denition 5.3, they would be deterministic functions of y, and hence stationary. We now
develop conditions ensuring both feasibility and rst-order eciency of the class of simulationbased estimators, as applied to this kind of data. Let a
denote the matrix having the rst q
rows of , where a is the diusion matrix in (5.??). Let C denote the Jacobian of C with
respect to y. We have:
Theorem 5.4. (Asset pricing and Cramer-Rao lower bound) Suppose to observe an intertemporal (, d q )-cohort of contingent claim prices c ( , ), and that there exist prediction functions
C in Rdq with the property that for = 0 and = 0 ,

a
( ) ( )1
= 0, P d -a.s. all [t, t + 1],
(5.21)
C( )
where C satisfies the initial condition C(t) = c (t, ) (c(y(t), 1 ), , c(y(t), dq )). Let
ct = (y o (t), c(y(t), 1 ), , c(y(t), dq )). Then, any simulation-based estimator applied to ct
is feasible. Moreover, asssume ct is also Markov. Then, any estimator with a span of moment
conditions for ct that also spans the true score, attains the Cramer-Rao lower bound, with
respect to the fields generated by ct .
According to Theorem 5.4, any estimator is feasible, whenever is locally invertible for a
time span equal to the sampling interval. As Figure 5.2 illustrates, condition (5.21) is satised
whenever is locally one-to-one and onto.5 If is also globally invertible for the same time span,
c is Markov. The last part of this theorem says that in this case, any estimator is obviously
asymptotically ecient. We emphasize that this conclusion is about rst-order eciency in the
joint estimation of and given the observations on c . It is not a claim about some estimator
being rst-order ecient in the estimation of , when y is fully observable.
Naturally, condition (5.21) does not ensure that is globally one-to-one and onto. In other
terms, might have many locally invertible restrictions.6 In practice, might fail to be globally
invertible because monotonicity properties of may break down in multidimensional diusion
models. In models with stochastic volatility, for example, option prices can be decreasing in the
underlying asset price (see Bergman, Grundy and Wiener (1996)); and in the corresponding
stochastic volatility yield curve models, medium-long term bond prices can be increasing in the
short-term rate (see Mele (2003)). Intuitively, these pathologies may arise because there is no
guarantee that the solution to a stochastic dierential system is nondecreasing in the initial
condition of one if its components - as it is instead the case in the scalar case.
When all components of vector y o represent the prices of assets actively traded in frictionless
markets, (5.21) corresponds to a condition ensuring market completeness in the sense of Harrison
and Pliska (1983). As an example, condition (5.21) for Hestons (1993) model is c/ =
0 P d -a.s, where denotes instantaneous volatility of the price process. This condition is
satised by the Hestons model. In fact, Romano and Touzi (1997) showed that within a fairly
general class of stochastic volatility models, option prices are always strictly increasing in
whenever they are convex in Q. Theorem 5.4 can be used to implement ecient estimators in
other complex multidimensional models. Consider for example a three-factor model of the yield
curve. Consider a state-vector (r, , ), where r is the short-term rate and , are additional
5 Local
invertibility of means that for every y Y , there exists an open set Y containing y such that the restriction of to
Y is invertible. And is locally invertible on Y if det J = 0 (where J is the Jacobian of ), which is condition (5.21).
6 As an example, consider the mapping R2 R2 dened as (y , y ) = (ey1 cos y , ey1 sin y ). The Jacobian satises
1 2
2
2
det J(y1 , y2 ) = e2y1 , yet is 2-periodic with respect to y2 . For example, (0, 2) = (0, 0).
162
c
by
A. Mele
factors (such as, say, instantaneous short-term rate volatility and a central tendency factor). Let
u(i) = u (r( ), ( ), ( ); Mi ) be the time rational price of a pure discount bond expiring
at Mi , i = 1, 2, and take M1 < M2 . Let (r, u(1) , u(2) ). Condition (5.21) for this model
is then,
(2)
(1) (2)
u(1)
(5.22)
u u u = 0, P dt-a.s. [t, t + 1],
where subscripts denote partial derivatives. It is easily checked that this same condition must be
satised by models with correlated Brownian motions and by yet more general models. Classes
of models of the short-term rate for which condition (5.22) holds are more intricate to identify
than in the European option pricing case seen above (see Mele (2003)).
163
c
by
A. Mele
5.10. Appendix 1: Proof of selected results
5.10 Appendix 1: Proof of selected results

Proof of Eq. (5.2). We have: P (A1
We still have,
Pr (A3 |A1
That is,
Pr
3
(
i=1
Ai
= Pr (A1
A2 ) = P (A1 ) P (A2 |A1 ). Consider the event E A1
A2 ) = Pr (A3 |E ) =
A2 ) Pr (A3 |A1
Continuing, we obtain Eq. (5.2).
(
( (
Pr (A3 E)
Pr (A3 A1 A2 )
(
=
.
Pr (E)
Pr (A1 A2 )
A2 ) = Pr (A1 ) Pr (A2 |A1 ) Pr (A3 |A1
164
A2 ) .
A2 .
c
by
A. Mele
5.11. Appendix 2: Collected notions and results
5.11 Appendix 2: Collected notions and results

Convergence in probability. A sequence of random vectors {xT } converges in probability to the
random vector x
if for each > 0, > 0 and each i = 1, 2, , N, there exists a T, such that for
every T T, ,
Pr (|xT i x
i | > ) < .
p
This is succinctly written as xT x

, or plim xT = x
, if x
x
, a constant.
Convergence in probability generalizes the standard notion of a limit of a deterministic sequence.
Of a deterministic sequence xT , we say it converges to some limit x
if, for > 0, there exists a T :
for each T T we have that |xT x
| < . Convergence in probability can also be restated as saying
that:
i | > ) = 0.
lim Pr (|xT i x
T
The following is a stronger notion of convergence:

Almost sure convergence. A sequence of random vectors {xT } converges almost surely to the
random vector x
if, for each i = 1, 2, , N, we have:
Pr ( : xT i () x
i ) = 1,
a.s.
where denotes the entire random sequence xT i . This is succinctly written as xT x

.
Almost sure convergence implies convergence in probability. Convergence in probability means
that for each > 0, limT Pr ( : |xT i () x
i | < ) = 1. Almost sure convergence requires that
i ) = 1 or that
Pr (limT xT i x
$
%
$
%
/
lim
Pr sup |xT i x
i | > = lim
Pr
|xT i x
i | > = 0.
T T
T T
Next, assume that the second order moments of all xi are nite. We have:
Convergence in quadratic mean. A sequence of random vectors {xT } converges in quadratic

mean to the random vector x
if for each i = 1, 2, , N, we have:
!
2
E
(x
)
lim
0.
Ti
i
q.m.

.
Remark. By Chebyshevs inequality,
Pr (|xT i x
i | > )
E[(xT i x
i ) 2 ]
,
2
which shows that convergence in quadratic mean implies convergence in probability.

We now turn to a weaker notion of convergence:
Convergence in distribution. Let {fT ()}T be the sequence of probability distributions (that is,
fT (x) = pr (xT x)) of the sequence of the random vectors {xT }. Let x
be a random vector with
probability distribution f(x). A sequence {xT } converges in distribution to x
if, for each i = 1, 2, , N,
we have:
lim fT (x) = f(x).
T
165
c
by
A. Mele

d

.
The following two results are useful to the purpose of this chapter:
p
Slutzkys theorem. If yT y and xT x

, then:
d
yT xT y x
.
Cramer-Wold device. Let be a N-dimensional vector of constants. We have:
d
xT x
xT x
.

d
d
The following example illustrates the Cramer-Wold device. If xT N 0; , then xT
N (0; ).
We now state two laws about convergence in probability.
Weak law (No. 1) (Khinchine). Let {xT } be a i.i.d. sequence satistfying E(xT ) = < T . We
have:
T
1
p
x
T
xt .
T t=1
Weak law (No. 2) (Chebyshev). Let
{xT } be a sequence independent but not
identically distributed,
satisfying E(xT ) = T < and E (xT T )2 = 2T < . If limT T12 Tt=1 2t 0, then:
T
T
1
1
p
x
T
xt
T
.
T t=1
T t=1 t
We now state and provide a proof of the central limit theorem in a simple setting.

Central Limit Theorem. Let {xT } be a i.i.d. sequence, satisfying E(xT ) = < and E (xT )2

= 2 < T . Let x
T T1 Tt=1 xt . We have,
T (
xT ) d
N(0, 1).
The multidimensional version of this theorem requires a mere change in notation. For the proof, the
classic method relies on the characteristic functions. Let:
"

(t) E eitx = eitx f(x)dx, i 1.

r

tr (t) t=0
= ir m(r) , where m(r) is the r-th order moment. By a Taylors expansion,

1 2

t2 + = 1 + im(1) t m(2) 1 t2 + .
(t) = (0) + (t) t +
(t)

2
t
2 t
2
t=0
t=0

Next, let x
T = T1 Tt=1 xt , and consider the random variable,
We have
T
T (
xT )
1 xt
YT
=
.
T t=1
166
c
by
A. Mele

The characteristic function of YT is the product of the characteristic functions of at
all the same: YT (t) = (a (t))T , where a (t) = 1
t2
2T
which are
+ . Therefore,

T
t T
1 t2
YT (t) =
= 1
+ o T 1
.
2T
T
1 2
xt
,
T
Clearly, limT YT (t) = e 2 t , which is the characteristic function of a standard Gaussian variable.
167
c
by
A. Mele
5.12. Appendix 3: Theory for maximum likelihood estimation
5.12 Appendix 3: Theory for maximum likelihood estimation

a.s.
Assume that T 0 , and that H(y, ) ln L( |# y) exists, it is continuous in uniformly in y
and that we can dierentiate twice inside the integral L( | y)dy = 1. We have:
T
1
ln L ( | y) .
sT () =
T t=1
Consider the c-parametrized curves (c) = c(0 T ) + T where, for all c (0, 1)p and , c
denotes a vector in where the ith element is c(i) (i) . By the intermediate value theorem, there exists
then a c in (0, 1)p such that we have almost surely:
sT (T ) = sT (0 ) + HT ( ) (T 0 ),
where (c ) and:
HT () =
T
1
H( | yt ).
T t=1
The rst order conditions tell us that sT ( T ) = 0. Hence,
0 = sT (0 ) + HT (T ) (T 0 ).
We also have that:
|HT (T ) HT (0 )|
T
1
|H (T ) HT (0 )| sup |H (T ) HT (0 )| ,
T t=1
(5A.1)
a.s.
where the supremum is taken over the set of all the observations. Since T 0 , we also have that
a.s.
T 0 . Moreover, by the law of large numbers,
T
1
p
H ( 0 | yt ) E [H ( 0 | yt )] = J (0 ) .
HT (0 ) =
T t=1
(5A.2)
Since H is continuous in uniformly in y, the inequality in (5A.1), and (5A.2) both imply that:
a.s.
HT (T ) J (0 ) .
Therefore, as T ,
T T 0 = HT1 (0 ) sT (0 ) T = J 1 T sT ( 0 ).
By the central limit theorem, and E (sT ) = 0, the score, sT (0 ) =
1
T
T
t=1 s (0 , yt ),
is such that
d
T sT (0 ) N (0, var (s ( 0 , yt ))) ,
where
var (s (0 , yt )) = J .
The result follows by the Slutzkys theorem and the symmetry of J .

Finally, one should show the existence of a sequence T converging a.s. to 0 . Proofs on this type
of convergence can be found in Amemiya (1985), or in Newey and McFadden (1994).
168
c
by
A. Mele
5.13. Appendix 4: Dependent processes
5.13 Appendix 4: Dependent processes

5.13.1
Weak dependence
Let 2T = var(
T
t=1 xt ),

and assume that that 2T = O(T ), and that 2T = O T 1 . If
1
T
T

t=1
(xt E(xt )) N (0, 1) ,
we say that {xt } is weakly dependent. Of a process, we say it is nonergodic, when it exhibits such a
strong dependence that it does not even satisfy the law of large numbers.
Stationarity
Weak dependence
Ergodicity
5.13.2
The central limit theorem for martingale differences

1 T
Let xt be a martingale dierence sequence with E x2t = 2t < for all t, and dene x
T
xt ,
T t=1
1 T
2 . Let,
and
2T
T t=1 t
T
1 2
x I
2 = 0,
T T
2T t=1 t |xt |T T
> 0, lim
Under the previous condition,
5.13.3
T
1 2
p
xt
2T 0.
T t=1
and
T x
T d
N (0, 1) .
Applications to maximum likelihood
We use the central limit theorem for martingale dierences to prove asymptotic normality of the MLE,
in the case of weakly dependent processes. We have,
ln LT () =
T
t () ,
t=1
t () (; yt | xt ) .
The MLE satises the following rst order conditions,

d
0p = ln LT ()|=T =
whence
We have:
which shows that
T

t=1
t ()|=0 +
T

t=1
t ()|=0 (T 0 ),
.1
T
T
1
1
d
T ( T 0 ) =
t (0 )
t (0 ).
T t=1
T t=1
t (0 )
E0 [ t+1 ( 0 )| Ft ] = 0p ,
is a martingale dierence. Naturally, here we also have that:
E0 ( | t+1 ( 0 )|2 | Ft ) = E0 ( t+1 (0 )| Ft ) Jt (0 ) .
169
(5A.3)
c
by
A. Mele
5.13. Appendix 4: Dependent processes

Next, for a given constant c Rp , let:
xt c t (0 ).
Clearly, xt is also a martingale dierence. Furthermore,

E0 x2t+1 Ft = c Jt ( 0 ) c,
and because xt is a martingale dierence, E (xt xti ) = E [E ( xt xti | Fti )] = E [E ( xt | Fti ) xti ] =
0, for all i. That is, xt and xti are mutually uncorrelated. It follows that,
$ T
%
T

var
xt =
E x2t
t=1
t=1
T
c E0 (| t ( 0 )|2 ) c
t=1
T

t=1
c E0 [E0 ( | t ( 0 )|2 | Ft1 )] c
T
c E0 [Jt1 (0 )] c
t=1
= c
- T
E0 (Jt1 (0 )) c.
t=1
Next, dene:
T
1
xt
x
T
T t=1
.
T
T

1
1
and
2T
E x2t = c
E (Jt1 (0 )) c.
T t=1
T t=1 0
Under the conditions underlying the central limit theorem for weakly dependent processes provided
earlier, to be spelled out below,
Tx
T d
N (0, 1) .
T
By the Cramer-Wold device,
-
T
1
E (Jt1 (0 ))
T t=1 0
.1/2
T
1
d
t ( 0 ) N (0, Ip ) .
T t=1
The conditions that need to be satised are,

T
T
1
1
p
t (0 )
E0 [Jt1 ( 0 )] 0,
T t=1
T t=1
and plim
T
1
E [Jt1 ( 0 )] J (0 ) .
T t=1 0
Under the previous conditions, it follows from Eq. (5A.3) that,

d
T (T 0 ) N 0p , J (0 )1 .
170
c
by
A. Mele
5.14. Appendix 5: Proof of Theorem 5.4
5.14 Appendix 5: Proof of Theorem 5.4

Let t t ( (y(t + 1), M (t + 1)1dq )| (y(t), M t1dq )) denote the transition density of
(y(t), M t1dq ) (y(t)) (y o (t), c(y(t), M1 t), , c(y(t), Mdq t)),
where we have emphasized the dependence of on the time-to-expiration vector:
M t1dq (M1 t, , Mdq t).
By ( ) full rank P d -a.s., and Itos lemma, satises, for [t, t + 1],
+ o
dy ( ) = bo ( )d + F ( )( )dW ( )
dc( ) = bc ( )d + c( )( )dW ( )
where bo and bc are, respectively, q -dimensional and (d q )-dimensional measurable functions, and
F ( ) a
( )( )1 P d -a.s. Under condition (5.21), t is not degenerate. Furthermore, C (y(t); )
C(t) is deterministic in (1 , , dq ). That is, for all (
c, c+ ) Rd Rd , there exists a function
c+ )) of (
c+ )
such that for any neighbourhood N(
c+ ) of c+ , there exists another neighborhood N((
such that,

&
'
c+ ) (y(t), M t1dq ) = c
: (y(t + 1), M (t + 1)1dq ) N(
&
= : (y o (t + 1), c(y(t + 1), M1 t)), , c(y(t + 1), Mdq t)) N((
c+ ))
| (y(t), M t1dq ) = c}
&
= : (y o (t + 1), c(y(t + 1), M1 t)), , c(y(t + 1), Mdq t)) N((
c+ ))
|(y o (t), c(y(t), M1 t), , c(y(t), Mdq t)) = c}
where the last equality follows by the denition of . In particular, the transition laws of ct given
ct1 are not degenerate; and ct is stationary. The feasibility of simulation based method of moments
estimation is proved. The eciency claim follows by the Markov property of , and the usual score
martingale dierence argument.
171
c
by
A. Mele
References
Altissimo, F. and A. Mele (2009): Simulated Nonparametric Estimation of Dynamic Models.
Review of Economic Studies 76, 413-450.
Amemiya, T. (1985): Advanced Econometrics. Cambridge, Mass.: Harvard University Press.
Bergman, Y. Z., B. D. Grundy, and Z. Wiener (1996): General Properties of Option Prices.
Journal of Finance 51, 1573-1610.
Carrasco, M., M. Chernov, J.-P. Florens and E. Ghysels (2007): Ecient Estimation of General Dynamic Models with a Continuum of Moment Conditions. Journal of Econometrics
140, 529-573.
Chernov, M. and E. Ghysels (2000): A Study towards a Unied Approach to the Joint Estimation of Objective and Risk-Neutral Measures for the Purpose of Options Valuation.
Journal of Financial Economics 56, 407-458.
Christensen, B. J. (1992): Asset Prices and the Empirical Martingale Model. Working paper,
New York University.
Due, D. and K. J. Singleton (1993): Simulated Moments Estimation of Markov Models of
Asset Prices. Econometrica 61, 929-952.
Fermanian, J.-D. and B. Salanie (2004): A Nonparametric Simulated Maximum Likelihood
Estimation Method. Econometric Theory 20, 701-734.
Fisher, R. A. (1912): On an Absolute Criterion for Fitting Frequency Curves. Messages of
Mathematics 41, 155-157.
Gallant, A. R. and G. Tauchen (1996): Which Moments to Match? Econometric Theory 12,
657-681.
Gauss, C. F. (1816): Bestimmung der Genanigkeit der Beobachtungen. Zeitschrift f
ur Astronomie und Verwandte Wissenschaften 1, 185-196.
Gourieroux, C., A. Monfort and E. Renault (1993): Indirect Inference. Journal of Applied
Econometrics 8, S85-S118.
Hajivassiliou, V. and D. McFadden (1998): The Method of Simulated Scores for the Estimation of Limited-Dependent Variable Models. Econometrica 66, 863-896.
Hansen, L. P. (1982): Large Sample Properties of Generalized Method of Moments Estimators. Econometrica 50, 1029-1054.
Hansen, L. P. and K. J. Singleton (1982): Generalized Instrumental Variables Estimation of
Nonlinear Rational Expectations Models. Econometrica 50, 1269-1286.
Hansen, L. P. and K. J. Singleton (1983): Stochastic Consumption, Risk Aversion, and the
Temporal Behavior of Asset Returns. Journal of Political Economy 91, 249-265.
Harrison, J. M. and S. R. Pliska (1983): A Stochastic Calculus Model of Continuous Trading:
Complete Markets. Stochastic Processes and their Applications 15, 313-316.
172
c
by
A. Mele
Heston, S. (1993): A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options. Review of Financial Studies 6, 327-343.
Laroque, G. and B. Salanie (1989): Estimation of Multimarket Fix-Price Models: An Application of Pseudo-Maximum Likelihood Methods. Econometrica 57, 831-860.
Laroque, G. and B. Salanie (1993): Simulation-Based Estimation of Models with Lagged
Latent Variables. Journal of Applied Econometrics 8, S119-S133.
Laroque, G. and B. Salanie (1994): Estimating the Canonical Disequilibrium Model: Asymptotic Theory and Finite Sample Properties. Journal of Econometrics 62, 165-210.
Lee, B-S. and B. F. Ingram (1991): Simulation Estimation of Time-Series Models. Journal
of Econometrics 47, 197-207.
Lee, L. F. (1995): Asymptotic Bias in Simulated Maximum Likelihood Estimation of Discrete
Choice Models. Econometric Theory 11, 437-483.
McFadden, D. (1989): A Method of Simulated Moments for Estimation of Discrete Response
Models without Numerical Integration. Econometrica 57, 995-1026.
Mele, A. (2003): Fundamental Properties of Bond Prices in Models of the Short-Term Rate.
Review of Financial Studies 16, 679-716.
Newey, W. K. and D. L. McFadden (1994): Large Sample Estimation and Hypothesis Testing. In: Engle, R. F. and D. L. McFadden (Editors): Handbook of Econometrics, Vol. 4,
Chapter 36, 2111-2245. Amsterdam: Elsevier.
Neyman, J. and E. S. Pearson (1928): On the Use and Interpretation of Certain Test Criteria
for Purposes of Statistical Inference. Biometrika 20A, 175-240, 263-294.
Pakes, A. and D. Pollard (1989): Simulation and the Asymptotics of Optimization Estimators. Econometrica 57, 1027-1057.
Pastorello, S., E. Renault and N. Touzi (2000): Statistical Inference for Random-Variance
Option Pricing. Journal of Business and Economic Statistics 18, 358-367.
Pastorello, S., V. Patilea, and E. Renault (2003): Iterative and Recursive Estimation in
Structural Non Adaptive Models. Journal of Business and Economic Statistics 21, 449509.
Pearson, K. (1894): Contributions to the Mathematical Theory of Evolution. Philosophical
Transactions of the Royal Society of London, Series A 185, 71-78.
Romano, M. and N. Touzi (1997): Contingent Claims and Market Completeness in a Stochastic Volatility Model. Mathematical Finance 7, 399-412.
Singleton, K. J. (2001): Estimation of Ane Asset Pricing Models Using the Empirical Characteristic Function. Journal of Econometrics 102, 111-141.
Smith, A. (1993): Estimating Nonlinear Time Series Models Using Simulated Vector Autoregressions. Journal of Applied Econometrics 8, S63-S84.
173
c
by
A. Mele
Tauchen, G. (1997): New Minimum Chi-Square Methods in Empirical Finance. In D. Kreps

and K. Wallis (Editors): Advances in Econometrics, 7th World Congress, Econometrics
Society Monographs, Vol. III. Cambridge UK: Cambridge University Press, 279-317.
174
Part II
Asset pricing and reality
175
6
On kernels and puzzles
This chapter discusses theoretical restrictions that can be used to perform statistical validation
of asset pricing models. We reconsider the Lucas model, and give more structure on the data
generating process. We present a simple setting which allows us to obtain closed-form solutions.
We then discuss how the models predictions can be used to test the validity of the model.
6.1 A single factor model

6.1.1
The model
There is a representative agent with CRRA utility, viz u (x) = x1 / (1 ). Cum-dividends

gross returns (St + Dt )/ St1 are generated by:
log(St + Dt ) = log St1 + S 1 2S + S,t

2
(6.1)
1
log Dt
= log Dt1 + D 2D + D,t
2
where
S,t
D,t
2

S
SD
NID 02 ;
.
SD 2D
Given the stochastic price and dividend process in (13.26), we now derive restrictions between
the various coecients S , D , 2S , 2D and SD that are imposed by economic theory. The cum
dividend process in (13.26) has been assumed so for analytical purposes only.
By standard consumption-based asset pricing theory (see Part I),

u (Dt+1 )
St = E
(St+1 + Dt+1 ) Ft ,
u (Dt )
with the usual notation. By the preferences assumption,

1 = E eZt+1 +Qt+1 Ft ,
(6.2)
c
by
A. Mele
6.1. A single factor model

where
Zt+1
$
%
Dt+1
= log
;
Dt
Qt+1 = log
St+1 + Dt+1
St
In fact, eq. (8.6) holds for any asset. In particular, it

holds
1
for a one-period bond with price
b
b
b
b
St bt , St+1 1 and Dt+1 0. Dene,
Qt+1 log bt log Rt . By replacing this into eq.

1
Zt+1
(8.6), one gets Rt = E e
Ft . We are left with the following system:

1
= E eZt+1 Ft
(6.3)
R
1 t = E eZt+1 +Qt+1 F
t
To obtain closed-form solutions, we will need to use the following result:
Lemma 6.1: Let Z be conditionally normally distributed. Then, for any R,

1 2
E eZt+1 Ft = eE( Zt+1 |Ft )+ 2 var( Zt+1 |Ft )

2
var [eZt+1 | Ft ] = eE( Zt+1 |Ft )+ var( Zt+1 |Ft ) 1 e 2 var( Zt+1 |Ft )
By the denition of Z, eq. (13.26), and Lemma 6.1,

1 2
1 2 2
1
1
= E eZt+1 Ft = eE[ Zt+1 |Ft ]+ 2 var[ Zt+1 |Ft ] = elog (D 2 D )+ 2 D .
Rt
The equilibrium interest rate thus satises,
log Rt = log + D
( + 1) 2
D ,
2
a constant.
(6.4)
The D term reects intertemporal substitution eects; the last term term reects precautionary motives.
The second equation in (6.3) can be written as,

1 2
1 2
1 = E [exp (Zt+1 + Qt+1 )| Ft ] = elog (D 2 D )+S 2 S E en t+1 Ft ,
where n
t+1 S,t+1 D,t N(0, 2S +2 2D 2 SD ). The above expectation can be computed
through Lemma 6.1. The result is,
0 = log D +

( + 1) 2
D + S SD .
2

log Rt
By dening Rt ert , and rearranging terms,
r = SD .
S
risk premium
To sum up,
S = r + SD
rt
= log + D
177
( + 1) 2
D
2
c
by
A. Mele
6.1. A single factor model
Let us compute other interesting objects. The expected gross return on the risky asset is,

St+1 + Dt+1
1 2
E
Ft = eS 2 S E [eS,t+1 | Ft ] = eS = er+SD .

St

Ft , as expected.
Therefore, if SD > 0, then E [ (St+1 + Dt+1 )/ St | Ft ] > E b1
t
Next, we test the internal consistency of the model. The coecients of the model must satisfy
some restrictions. In particular, the asset price volatility must be determined endogeneously.
We rst conjecture that the following no-sunspots condition holds,
S,t = D,t .
(6.5)
We will demonstrate below that this is indeed the case. Under the previous condition,
S = r + D ;
and
Zt+1
D ,

1 2
= r + uD,t+1 ;
2
uD,t+1
D,t+1
.
D
Under condition (6.5), we have a very instructive way to write the pricing kernel. Precisely,
dene recursively,
mt+1 = t+1 exp (Zt+1 ) ; 0 = 1.

t
This is reminiscent of the continuous time representation of Arrow-Debreu state prices (see
Chapter 4).
Next, lets iterate the asset price equation (8.6),
.
.
%
%
-$ n
-$ i
n

1

1

St = E
eZt+i St+n Ft +
E
eZt+j Dt+i Ft

j=1
i=1
j=1

n

t+n
t+i
= E
St+n Ft +
E
Dt+i Ft .
t
t
i=1
By letting n and assuming no-bubbles, we get:

t+i
St =
E
Dt+i Ft .
t
i=1
(6.6)
The expectation is, by Lemma 6.1,

i

t+i
Z
E
Dt+i Ft = E e j=1 t+j Dt+i Ft = Dt e(D rD )i .
t
Suppose that the risk-adjusted discount rate r + D is higher than the growth rate of the
economy, viz.
r + D > D k eD rD < 1.
Under this condition, the summation in eq. (6.6) converges, and we obtain:
St
k
.
=
Dt
1k
178
(6.7)
c
by
A. Mele
6.2. The equity premium puzzle
This is a version of the celebrated Gordons formula. It predicts that price-dividend ratios are
constant, a counterfactual feature addressed in Chapter 8.
To nd the nal restrictions of the model, notice that eq. (6.7) and the second equation in
(13.26) imply that
1
log(St + Dt ) log St1 = log k + D 2D + D,t .
2
By the rst equation in (13.26),

1
1
S 2S = D 2D log k
2
2
S,t
= D,t , t
The second condition conrms condition (6.5). It also reveals that, 2S = SD = 2D . By replacing
this into the rst condition, delivers back S = D log k = r + D .
6.1.2
Extensions
In Chapter 3 we showed that in a i.i.d. environment, prices are convex (resp. concave) in the
dividend rate whenever > 1 (resp. < 1). The pricing formula (6.7) reveals that in a dynamic
environment, such a property is lost. In this formula, prices are always linear in the dividends
rate. It would be possible to show with the techniques developed in the next chapter that in
a dynamic context, convexity properties of the price function would be inherited by properties
of the dividend process in the following sense: if the expected dividend growth under the riskneutral measure is a convex (resp. concave) function of the initial dividend rate, then prices are
convex (resp. concave) in the initial dividend rate. In the model analyzed here, the expected
dividend growth under the risk-neutral measure is linear in the dividends rate, and this explains
the linear formula (6.7).
6.2 The equity premium puzzle

Average excess returns on the US stock market [the equity premium] is too high to be
easily explained by standard asset pricing models. Mehra and Prescott
To be consistent with data, the equity premium,

S r = D ,
= D
must be high enough - as regards US data, approximately an annualized 6%. If the asset we
are trying to price is literally a consumption claim, then D is consumption volatility, which is
very low (approximately 3%). To make S r high, one needs very high values of (lets say
30), But assuming = 30 doesnt seem to be plausible. This is the equity premium puzzle
originally raised by Mehra and Prescott (1985).
Even if we dismiss the idea that = 30 is implausible, there is another puzzle, the interest
rate puzzle. As we showed in eq. (6.4), very high values of can make the interest rate very
high (see Figure 6.1).
In the next section, we show how this failure of the model can be detected with a general
methodology that can be applied to a variety of related models - more general models.
179
c
by
A. Mele
6.3. The Hansen-Jagannathan cup
r
0.1
0.0
10
20
30
40
eta
-0.1
FIGURE 6.1. The risk-free rate puzzle: the two curves depict the graph
r() = log + 0.0183 (0.0328)2 (+1)
, with = 0.95 (top curve) and = 1.05
2
(bottom curve). Even if we accept the idea that risk aversion is as high as = 30, we would obtain a
resulting equilibrium interest rate as high as 10%. The only way to make low r consistent with high
values of is to make > 1.
6.3 The Hansen-Jagannathan cup

Suppose there are n risky assets. The n asset pricing equations for these assets are,
1 = E [mt+1 (1 + Rj,t+1 )| Ft ] ,
j = 1, , n.
By taking the unconditional expectation of the previous equation, and dening Rt = (R1,t ,
, Rn,t ) ,
1n = E [mt (1n + Rt )] .
Let m
E(mt ). We create a family of stochastic discount factors mt parametrized by m
by
projecting m on to the asset returns,
P roj ( m| 1n + Rt ) mt (m)
=m
+ [Rt E(Rt )] m ,
1n
n1
where1
m
n1
= 1 cov (m, 1n + Rt ) = 1 [1n mE

(1n + Rt )] ,
nn
n1
!
and E (Rt E(Rt )) (Rt E(Rt )) . As shown in the appendix, we also have that,
(1n + Rt )] .
1n = E [mt (m)
We have,

var (mt (m))
= m m = (1n mE
(1n + Rt )) 1 (1n mE
{1n + Rt }).
This is the celebrated Hansen-Jagannathan cup (Hansen and Jagannathan (1991)). The
interest of this object lies in the following theorem.
1 We
have, cov (m, 1n + Rt ) = E [m (1n + R)] E (m) E (1n + Rt ) = 1n mE

(1n + Rt ).
180
6.3. The Hansen-Jagannathan cup
c
by
A. Mele
Theorem 6.2: Among all stochastic discount factors with fixed expectation m,
mt (m)
is the
one with the smallest variance.
Proof: Consider another discount factor indexed by m,
i.e. mt (m).
Naturally, mt (m)
satises
1n = E [mt (m)
(1n + Rt )]. And since it also holds that 1n = E [mt (m)
(1n + Rt )], we deduce
that
0n =
=
=
=
E [(mt (m)
mt (m))
(1n + Rt )]
E {[mt (m)
mt (m)]
[(1n + E(Rt )) + (Rt E(Rt ))]}
E {[mt (m)
mt (m)]
[Rt E(Rt )]}
cov [mt (m)

mt (m),
Rt ]
where the third line follows from the fact that E [mt (m)]
= E [mt (m)]
= m,
and the fourth line
follows because E [(mt (m)

mt (m))]
= 0. But mt (m)
is a linear combination of Rt . By the
previous equation, it must then be the case that,
0 = cov [mt (m)
mt (m),
mt (m)]
.
Hence,
=
var [mt (m)]
=
=
var [mt (m)

+ mt (m)
mt (m)]
var [mt (m)]

+ var [mt (m)
mt (m)]
+ 2 cov [mt (m)
mt (m),
mt (m)]
var [mt (m)]

+ var [mt (m)
mt (m)]
var [mt (m)]

.

The previous bound can be improved by using conditioning information as in Gallant, Hansen
and Tauchen (1990) and the relatively more recent work by Ferson and Siegel (2003). Moreover,
these bounds typically diplay a nite sample bias: they typically overstate the true bounds and
thus they reject too often a given model. Finite sample corrections are considered by Ferson
and Siegel (2003).
For example, let us consider an application of the Hansen-Jagannathan testing methodology
to the model in Section 6.1. That model has the following stochastic discount factor,

t+1
1 2
D,t+1
mt+1 =
= exp (Zt+1 ) ;
Zt+1 = r + uD,t+1 ;
uD,t+1
.
t
2
D
First, we have to compute the rst two moments of the stochastic discount factor. By Lemma
6.1 we have,

1 2
2
m
= E(mt ) = er and m = var (mt (m))
= er+ 2 1 e
(6.8)
where
( + 1) 2
D and = D .
2
For given D and 2D , system (6.8) forms a -parametrized curve in the space (m-
m ). The
objective is to see whether there are plausible values of for which such a -parametrized
181
r = log + D
c
by
A. Mele
6.4. Multifactor extensions
curve enters the Hansen-Jagannathan cup. Typically, this is not the case. Rather, one has the
situation depicted in Figure 6.2 below.
The general message is that models can be consistent with data with high volatile pricing
kernels (for a xed m).
Dismiss the idea of a representative agent with CRRA utility function.
Consider instead models with heterogeneous agents (by generalizing some ideas in Constantinides and Due (1996); and/or consider models with more realistic preferences - such as for
example the habit preferences considered in Campbell and Cochrane (1999); and/or combinations of these. These things will be analyzed in depth in the next chapter.
6.4 Multifactor extensions

A natural way to increase the variance of the pricing kernel is to increase the number of factors.
We consider two possibilities: one in which returns are normally distributed, and one in which
returns are lognormally distributed.
6.4.1
Exponential affine pricing kernels
Consider again the simple model in Section 6.1. In this section, we shall make a dierent
assumption regarding the returns distributions. But we shall maintain the hypothesis that the
pricing kernel satises an exponential-Gaussian type structure,

1 2
Zt+1 = r + uD,t+1 ;
uD,t+1 NID (0, 1) ,
mt+1 = exp (Zt+1 ) ;
2
where r and are some constants. We have,
t+1 ) = E (mt+1 ) E(R
t+1 ) + cov(mt+1 , R
t+1 ),
1 = E(mt+1 R
t+1 St+1 + Dt+1 .

R
St
By rearranging terms,2 and using the fact that E (mt+1 ) = R1 ,

t+1 ) R = R cov(mt+1 , R
t+1 ).
E(R
(6.9)
The following result is useful:

Lemma 6.3 (Steins lemma): Suppose that two random variables x and y are jointly normal.
Then,
cov [g (x) , y] = E [g (x)] cov (x, y) ,
for any function g : E (|g (x)|) < .
is normally distributed. This assumption is inconsistent with the

We now suppose that R
is lognormally distributed in equilibrium
model in Section 6.1. In the model of Section 6.1, R
= D 1 2 +S , with S normal. But lets explore the asset pricing implications of
because log R
2 D
2 With
a portfolio return that is perfectly correlated with m, we have:

M )
Et (R
t+1
1
t (mt+1 )
M ).
=
t (R
t+1
Et (mt+1 )
Et (mt+1 )
In more general setups than the ones considered in this introductory example, both
182
t (mt+1 )
Et (mt+1 )
M ) should be time-varying.
and t (R
t+1
c
by
A. Mele
t+1 and Zt+1 are normal, and mt+1 = m (Zt+1 ) = exp (Zt+1 ),
this tilting assumption. Because R
we may apply Lemma 6.3 and obtain,
t+1 ) = E [m (Zt+1 )] cov(Zt+1 , R
t+1 ) = R1 cov(uD,t+1 , R
t+1 ).
cov(mt+1 , R
Replacing this into eq. (6.9),
t+1 ) R = cov(uD,t+1 , R
t+1 ).
E(R
We wish to extend the previous observations to more general situations. Clearly, the pricing
kernel is some function of K factors m (1t , , Kt ). A particularly convenient analytical assumption is to make m exponential-ane and the factors (i,t )K
i=1 normal, as in the following
denition:
Definition 6.4 (EAPK: Exponential Ane Pricing Kernel): Let,
Zt 0 +
K
i i,t .
i=1
A EAPK is a function
mt = m(Zt ) = exp(Zt ).
2
If (i,t )K
i=1 are jointly normal, and each i,t has mean zero and variance i , i = 1, , K, the
EAPK is called a Normal EAPK (NEAPK).
In the previous denition, we assumed that each i,t has mean zero. This entails no loss of
generality insofar as 0 = 0.
is normally distributed. By Lemma 6.3 and the NEAPK structure,
Now suppose that R
t+1 ) = cov[exp (Zt+1 ) , R
t+1 ] = R1 cov(Zt+1 , R
t+1 ) = R1
cov(mt+1 , R
K
t+1 ).
i cov(i,t+1 , R
i=1
By replacing this into eq. (6.9) leaves the linear factor representation,
t+1 ) R =
E(R
K

i=1
We have thus shown the following result:
).
i cov(i,t+1 , R

t+1
(6.10)
betas
is normally distributed. Then, NEAPK linear factor

Proposition 6.5: Suppose that R
representation for asset returns.
The APT representation in eq. (6.10), is close to one result in Cochrane (1996).3 Cochrane
(1996) assumed that m has a linear structure, i.e. m (Zt ) = Zt where Zt is as in Denition 5.1.
3 To recall why eq. (6.10) is indeed a APT equation, suppose that R
is a n-(column) vector of returns and that R
= a + bf , where
f is K-(column) vector with zero mean and unit variance and a, b are some given vector and matrix with appropriate dimension.
f ). A portfolio delivers R
= a + cov(R,
f )f . Arbitrage opportunity is: : cov(R,
f) = 0
Then clearly, b = cov(R,
and a = r. To rule that out, we may show as in Part I of these Lectures that there must exist a K-(column) vector s.t.
f ) + r. This implies R
= a + bf = r + cov(R,
f ) + bf . That is, E(R)
= r + cov(R,
f ).
a = cov(R,
183
c
by
A. Mele

t+1 ) =
This assumption implies that cov(mt+1 , R
eq. (6.9),
t+1 ) R = R
E(R
K
K
i=1
t+1 ). By replacing this into

i cov(i,t+1 , R
t+1 ),
i cov(i,t+1 , R
where R =
i=1
1
1
= .
E (m)
0
The advantage to use the NEAPKs is that the pricing kernel is automatically guaranteed to be
strictly positive - a condition needed to rule out arbitrage opportunities.
6.4.2
Lognormal returns
is lognormally distributed, and that NEAPK holds. We have,
Next, we assume that R

!

t+1 e0 = E e K
t+1 .
i=1 i i,t+1 R
1 = E mt+1 R
(6.11)
t be normally distributed. The previous

Consider rst the case K = 1 and let yt = log R
equation can be written as,

1
2 2
2
e0 = E e1 t+1 +yt+1 = eE(yt+1 )+ 2 (1 +y +21 y ) .
This is,

1 2 2
2
E (yt+1 ) = 0 + (1 + y + 21 y ) .
2
By applying the pricing equation (6.11) to a bond price,

1 2 2
e0 = E e1 t+1 elog Rt+1 = elog Rt+1 + 2 1 ,
and then
log Rt+1
The expected excess return is,

1 2 2
= 0 + 1 .
2
1
E (yt+1 ) log Rt+1 + 2y = 1 y .
2
This equation reveals how to derive the simple theory in Section 6.1 in an alternate way.
Apart from Jensens inequality eects ( 12 2y ), this is indeed the Lucas model of Section 6.1 once
1 = . As is clear, this is a poor model because we are contrived to explain returns with only
one stochastic discount-factor parameter (i.e. with 1 ).
Next consider the general case. Assume as usual that dividends are as in (13.26). To nd the
price function in terms of the state variable , we may proceed as in Section 6.1. In the absence
of bubbles,

1 K
t+i
2
E
Dt+i = Dt
e(D +0 + 2 i=1 i (i i +2i,D ))i , i,D cov (i , D ) .
St =
t
i=1
i=1
Thus, if
K

1 2
k D + 0 +
i i i + 2 i,D < 0,
2 i=1
184
6.5. Pricing kernels, Sharpe ratios and the market portfolio
c
by
A. Mele
then,
St
k
=
.
Dt
1 k
Even in this multi-factor setting, price-dividend ratios are constant - which is counterfactual.
Note that the various parameters can be calibrated so as to make the pricing kernel satisfy
the Hansen-Jagannathan theoretical test conditions in Section 6.3. But the resulting model
always makes the boring prediction that price-dividend ratios are constant. This multifactor
model doesnt work even if the variance of the implied pricing kernel is high - and lies inside
the Hansen-Jagannathan cup. Living inside the cup doesnt necessarily imply that the resulting
model is a good one. We need other theoretical test conditions. The next chapter develops
such theoretical test conditions (When are price-dividend ratios procyclical? When is returns
volatility countercyclical? Etc.).
6.5 Pricing kernels, Sharpe ratios and the market portfolio

6.5.1
What does a market portfolio do?
When is the market portfolio perfectly correlated with the pricing kernel? When this is the case,
we say that the market portfolio is -CAPM generating. The answer to the previous question
is generically negative. This section aims at clarifying the issue.4
e
i,t+1 Rt+1 be the excess return on a risky asset. By standard arguments,
Let ri,t+1
=R

e
e
e
0 = Et mt+1 ri,t+1 = Et (mt+1 ) Et ri,t+1 + i,t V art (mt+1 ) V art ri,t+1
,
e
where i,t = corrt (mt+1 , ri,t+1
). Hence,
e

Et ri,t+1
V art (mt+1 )
Sharpe Ratio
e = i,t E (m ) .
t
t+1
V art ri,t+1
Sharpe ratios computed on Fama-French data take on an average value of 0.45.

We have,

e

Et ri,t+1

V art (mt+1 ) = V art (mt+1 ) Rt+1 .

Et (mt+1 )
e
V art ri,t+1

M
The highest possible Sharpe ratio is bounded. The equality is obtained with returns Rt+1
(say) that are perfectly conditionally negatively correlated with the price kernel, i.e. M,t = 1.
This portfolio is a -CAPM generating portfolio. Should it be ever possible to create such a
portfolio, we could also call it market portfolio. The reason is that a feasible and attainable
portfolio lying on the kernel volatility bounds is clearly mean-variance ecient.
If M,t = 1,
e

Et rM,t+1
V art (mt+1 )

e
= slope of the capital market line = E (m ) .
t
t+1
V ar r
t
4 An
M,t+1
additional article to read is Cecchetti, Lam, and Mark (1994).
185
c
by
A. Mele

e
As explained in Chapter 2, the Sharpe ratio Et (rM,t+1
)
tation of unit market risk-premium. Therefore:
unit market risk premium =
e
V art (rM,t+1
) has also the interpre-
V art (mt+1 )
.
Et (mt+1 )
For example, the Lucas model in Section 5.1 has,

V art (mt+1 ) 2 2
= e D 1 D .
Et (mt+1 )
In Section 5.1, we also obtained that (S r)/ D = D . As the previous relation reveals,
is only approximately equal to D because the asset in Section 6.1 is simply not a -CAPM
generating portfolio. For example, suppose that the economy in Section 6.1 has only a single
risky asset. It would then be very natural to refer this asset to as market portfolio. Yet this
asset wouldnt be -CAPM generating.
= eS , R = e log +(D 12 2D ) 12 2 2D and var(R)
=
In Section 6.1, we found that E(R)
8
2
R)
is:
e2S (eD 1). Therefore, S E(R
var(R)
2
1 eD
S Sharpe Ratio = 2
.
eD 1
Indeed, by simple computations,
1 eD
2
= 2 2
.
e D 1 eD 1
This is not precisely minus one. Yet in practice 1 when D is low. However, consumption
claims are not acting as market portfolios - in the sense of Chapter 2. If that consumption claim
is very highly correlated with the pricing kernel, then it is also a good approximation to the CAPM generating portfolio. But as the previous simple example demonstrates, that is only an
approximation. To summarize, the fact that everyone is using an asset (or in general a portfolio
in a 2-funds separation context) doesnt imply that the resulting return is perfectly correlated with
the pricing kernel. In other terms, a market portfolio is not necessarily -CAPM generating.5
We now describe a further complication: a -CAPM generating portfolio is not necessarily
the tangency portfolio. We show the existence of another portfolio producing the same -pricing
relationship as the tangency portfolio. For reasons developed below, such a portfolio is usually
referred to as the maximum correlation portfolio.
= 1 . By the CCAPM (see Chapter 3),
Let R
E(m)

= Ri ,m E (Rp ) R
,
E Ri R
Rp ,m
where Rp is a portfolio return. Next, let
Rp = R m
m
.
E(m2 )
5 As is well-known, things are the same in economies with one agent with quadratic utility. This fact can be seen at work in the
previous formulae (just take = 1). You should also be able to show this claim with more general quadratic utility functions - as
in chapter 3.
186
c
by
A. Mele
This is clearly perfectly correlated with the kernel, and by the analysis in Chapter 3,

.
= Ri ,Rm E (Rm ) R
E Ri R
This is not yet the -representation of the CAPM, because we have yet to show that there
is a way to construct Rm as a portfolio return. In fact, there is a natural choice: pick m = m ,
where m is the minimum-variance kernel leading to the Hansen-Jagannathan bounds. Since
m is linear in all asset retuns, Rm can be thought of as a return that can be obtained by
investing in all assets. Furthermore, in the appendix we show that Rm satises,

1 = E m Rm .
Where is this portfolio located? As shown in the appendix, there is no portfolio yielding the
same expected return with lower variance (i.e., Rm is mean-variance ecient). In addition, in
the appendix we show that,

r Sh
1+r
E Rm 1 =
=r
Sh < r.
1 + Sh
1 + Sh
Mean-variance eciency of Rm and the previous inequality imply that this portfolio lies in
the lower branch of the mean-variance ecient portfolios. And this is so because this portfolio
is positively correlated with the true pricing kernel. Naturally, the fact that this portfolio is
-CAPM generating doesnt necessarily imply that it is also perfectly correlated with the true
pricing kernel. As shown in the appendix, Rm has only the maximum possible correlation
with all possible m. Perfect correlation occurs exactly in correspondence of the pricing kernel
m = m (i.e. when the economy exhibits a pricing kernel exactly equal to m ).
Proof that Rm is -capm generating. The relations 1 = E(m Ri ) and 1 = E(m Rm )

imply

E(Ri ) R = R cov m , Ri

E(Rm ) R = R cov m , Rm
and,
E(Ri ) R
cov (m , Ri )
=
.
E(Rm ) R
cov (m , Rm )
By construction, Rm is perfectly correlated with m . Precisely, Rm = m / E(m2 ) 1 m ,

E(m2 ). Therefore,

cov Rm , Ri
cov Rm , Ri
cov (m , Ri )
=
=
= Ri ,Rm .
cov (m , Rm )
cov (Rm , Rm )
var (Rm )

6.5.2
Final thoughts on the pricing kernel bounds
Figure 6.2 summarizes the typical situation that arises in practice. Points are the ones generated by the Lucas model in correspondence of dierent . The model has to be such that points
lie above the observed Sharpe ratio ((m)/ E(m) greatest Sharpe ratio ever observed in
187
c
by
A. Mele
the dataSharpe ratio on the market portfolio) and inside the Hansen-Jagannathan bounds.
Typically, very high values of are required to enter the Hansen-Jagannathan bounds.
There is a beautiful connection between these things and the familiar mean-variance portfolio
frontier described of Chapter 1. As shown in Figure 6.3, every asset or portfolio must lie inside
the wedged region bounded by two straight lines with slopes (m)/ E(m). This is so because,
for any asset (or portfolio) that is priced with a kernel m, we have that

E(Ri ) R (m) Ri .
E(m)
As seen in the previous section, the equality is only achieved by asset (or portfolio) returns that
are perfectly correlated with m. The point here is that a tangency portfolio such as T doesnt
necessarily attain the kernel volatility bounds. Also, there is no reason for a market portfolio
to lie on the kernel volatility bound. In the simple Lucas-Breeden economy considered in the
previous section, for example, the (only existing) asset has a Sharpe ratio that doesnt lie on
the kernel volatility bounds. In a sense, the CCAPM doesnt necessarily imply the CAPM, i.e.
there is no necessarily an asset acting at the same time as a market portfolio and -CAPM
generating that is also priced consistently with the true kernel of the economy. These conditions
simultaneously hold if the (candidate) market portfolio is perfectly negatively correlated with
the true kernel of the economy, but this is very particular (it is in this sense that one may
say that the CAPM is a particular case of the CCAPM). A good research question is to nd
conditions on families of kernels consistent with the previous considerations.
(m )
H ansen-Jagannathan bounds
Sharpe ratio
E (m )
On the other hand, we know that there exists another portfolio, the maximum correlation
portfolio, that is also -CAPM generating. In other terms, if R : R = m, for some positive
constant , then the -CAPM representation holds, but this doesnt necessarily mean that R
is also a market portfolio. More generally, if there is a return R that is -CAPM generating,
then
i,m
i,R =
, all i.
(6.12)
R ,m
Therefore, we dont need an asset or portfolio return that is perfectly correlated with m to
make the CCAPM shrink to the CAPM. In other terms, the existence of an asset return that is
188
c
by
A. Mele
perfectly negatively correlated with the price kernel is a sucient condition for the CCAPM to
shrink to the CAPM, not a necessary condition. The proof of eq. (6.12) is easy. By the CCAPM,
E(Ri ) R = i,m
(m)
(Ri );
E(m)
That is,
and E(R ) R = R ,m
(m)
(R ).
E(m)
i,m (Ri )
E(Ri ) R
=
E(R ) R
R ,m (R )
(6.13)
E(Ri ) R
cov(Ri , R )
(Ri )
=
=
i,R
E(R ) R
(R )2
(R )
(6.14)
But if R is -CAPM generating,
Comparing eq. (6.13) with eq. (6.14) produces (6.12).
E(R)
kernel volatility bounds

mean-variance efficient portfolios
efficient portfolios frontier
T
tangency portfolio
1 / E(m)
maximum correlation portfolio
(R)
A nal thought. Many recent applied research papers have important result but also a surprising motivation. They often state that because we observe time-varying Sharpe ratios on8(proxies

of) the market portfolio, one should also model the market risk-premium V art (mt+1 ) Et (mt+1 )
as time-varying. However, this is not rigorous
motivation. The Sharpe
8
8 ratio of the market portfo

lio is generally less than V art (mt+1 ) Et (mt+1 ). V art (mt+1 ) Et (mt+1 ) is only a bound.
8

On a strictly theoretical point of view, V art (mt+1 ) Et (mt+1 ) time-varying is not a necessary nor a sucient condition to observe time-varying Sharpe ratios. Figure 6.3 illustrates this
point.
6.5.3
The Rolls critique
In applications and tests of the CAPM, proxies of the market portfolio such as the S&P 500 are
used. However, the market portfolio is unobservable, and this prompted Roll (1977) to point
out that the CAPM is inherently untestable. The argument is that a tangency portfolio (or in
general, a -CAPM generating portfolio) always exists. So, even if the CAPM is wrong, the
189
c
by
A. Mele
proxy of the market portfolio will incorrectly support the model if such a proxy is more or
less the same as the tangency portfolio. On the other hand, if the proxy is not mean-variance
ecient, the CAPM can be rejected even if the CAPM is wrong. All in all, any test of the CAPM
is a joint test of the model itself and of the closeness of the proxy to the market portfolio.
190
c
by
A. Mele
6.6. Appendix
6.6 Appendix
Proof of the Equation, 1n = E [mt (m)
(1n + Rt )]. We have,
E [mt (m)
(1n + Rt )] = E

!
m
+ (Rt E(Rt )) m
(1
+
R
)
n
t
!
= mE
(1n + Rt ) + E (Rt E(Rt )) m
(1
+
R
)
n
t
!
= mE
(1n + Rt ) + E (1n + Rt ) (Rt E(Rt )) m
!
= mE
(1n + Rt ) + E ((1n + E(Rt )) + (Rt E(Rt ))) (Rt E(Rt )) m
!
= mE
(1n + Rt ) + E (Rt E(Rt )) (Rt E(Rt )) m
= mE
(1n + Rt ) + m
= mE
(1n + Rt ) + 1n mE
(1n + Rt ) ,
where the last line follows by the denition of m

.
Proof that Rm can be generated by a feasible portfolio

Proof of the Equation, 1 = E m Rm . We have,
E(m Rm ) =
1
E (m m ) ,
E [(m )2 ]
where
!
E (m m ) = m
2 + E m (Rt E(Rt )) m
!
!
= m
2 + E m (1 + Rt ) m
E
m
(1
+
E(R
))
m
t
= m
2 + m
E (m) [1 + E(Rt )] m
2
= m
+ 1n m
(1 + E(Rt )) m
!
= m
2 + 1n m
(1 + E(Rt )) 1 [1n m
(1n + E(Rt ))]
= m
2 + var (m ) ,
where the last line is due to the denition of m .
Proof that Rm is mean-variance

efficient. Let p = (p0 , p1 , , pn ) the vector of n + 1

portfolio weights (here pi i w is the portfolio weight of asset i, i = 0, 1, , n. We have,
p 1n+1 = 1.
1

The returns we consider are rt = m
1, r1,t , , rn,t . We denote our benchmark portfolio
return as rbt = rm 1. Next, we build up an arbitrary portfolio yielding the same expected return
E(rbt ) and then we show that this has a variance greater than the variance of rbt . Since this portfolio
191
c
by
A. Mele
6.6. Appendix
is arbitrary, the proof will be complete. Let rpt = p rt such that E(rpt ) = E(rbt ). We have:
cov (rbt , rpt rbt ) = E [rbt (rpt rbt )]
= E [Rbt (Rpt Rbt )]

= E (Rbt Rpt ) E R2bt

!

1
1
=
1
+
p
r
E m2
E
m
t
2
2
2
E (m )
[E (m )]
*
)
1
=
p
E
[m
(1
+
r
)]
1
n+1
t
E (m2 )
= 0.
The rst line follows by construction since E(rpt ) = E(rbt ). The last line follows because
p E [m (1n+1 + rt )] = p 1n+1 = 1.
Given this, the claim follows directly from the fact that
var (Rpt ) = var [Rbt + (Rpt Rbt )] = var (Rbt ) + var (Rpt Rbt ) var (Rbt ) .

Proof of the Equation, E Rm 1 = r
E(Rm ) 1 =
1+r
1+Sh Sh.
We have,
1.
E[(m )2 ]
In terms of the notation introduced in Section 6.8, m is:

1
m = m
+ (a) m
(1n m
{1n + b}) .
, m
=
We have,
E[(m )2 ] =
m
+ (a) m
!2
= m
2 + E (a) m
!2
!2
= m
2 + E (a) m
(a)

!
= m
2 + E
a m
m
a
= m
2 +
m
m

!
2
= m
+ 1n m
1
+
b
1 [1n m
(1n + b)]
n

1
1
1
= m
2 + 1
1
+
1
b
n
n
n
n
n

*
)
1
1
1
1
1
1
1
m
1
n 1n + b 1n m
n 1n + b 1n + 1n b + b b
1
1
Again in terms of the notation of Section 6.8 ( 1
n 1n and 1n b), this is:
!

E (m )2 = 2m
( + ) + m
2 1 + + 2 + b 1 b .
The expected return is thus,

E R

m
+ 2m
( + ) m
2 1 + + 2 + b 1 b
! 1=
1=
.
2m
( + ) + m
2 (1 + + 2 + b 1 b)
E (m )2
E (m )
192
c
by
A. Mele
6.6. Appendix
Now recall two denitions:
m
=
1
1+r
Sh = (b 1m r) 1 (b 1m r) = b 1 b 2r + r2 .
In terms of r and Sh, we have,

E Rm 1 =
E (m )
E (m )2
! 1
(1 + r)2 (1 + r) (1 + 2 + 2) + 1 + + 2 + b 1 b
(1 + r)2 (1 + r) (2 + 2) + 1 + + 2 + b 1 b
r Sh
=
1 + Sh
1+r
= r
Sh
1 + Sh
< r.
This is positive if r Sh > 0, i.e. if b 1 b (2 + 1) r + r2 < 0, which is possible for suciently

low (or suciently high) values of r.
Proof that Rm is the m-maximum correlation portfolio. We have to show that for any
price kernel m, |corr(m, Rbt )| |corr(m, Rpt )|. Dene a -parametrized portfolio such that:
E [(1 )Ro + Rpt ] = E (Rbt ) , Ro m
1 .
We have
corr (m, Rpt ) = corr [m, (1 )Ro + Rpt ]
= corr [m, Rbt + ((1 )Ro + Rpt Rbt )]
cov (m, Rbt ) + cov (m, (1 )Ro + Rpt Rbt )

=
(m) var ((1 )Ro + Rpt )
cov (m, Rbt )

=
(m) var ((1 )Ro + Rpt )
The rst line follows because (1 )Ro + Rpt is a nonstochastic ane translation of Rpt . The last
equality follows because
cov (m, (1 )Ro + Rpt Rbt ) = E [m ((1 )Ro + Rpt Rbt )]
= (1 ) E (mRo ) + E (mRpt ) E (m Rbt )

=1
=1
=1
= 0.
where the rst line follows because E((1 )Ro + Rpt ) = E(Rbt ).
Therefore,
corr (m, Rpt ) =
cov (m, Rbt )

cov (m, Rbt )

= corr (m, Rbt ) ,

(m) var ((1 )Ro + Rpt )
(m) var(Rbt )
where the inequality follows because Rbt is mean-variance ecient (i.e. feasible portfolios with the
same expected return as Rbt and variance less than var(Rbt )), and then var((1 )Ro + Rpt )
var(Rbt ), all Rpt .
193
c
by
A. Mele
6.6. Appendix
References
Campbell, J. Y. and J. Cochrane (1999): By Force of Habit: A Consumption-Based Explanation of Aggregate Stock Market Behavior. Journal of Political Economy 107, 205-251.
Cecchetti, S., Lam, P-S. and N. C. Mark (1994): Testing Volatility Restrictions on Intertemporal Rates of Substitution Implied by Euler Equations and Asset Returns. Journal of
Finance 49, 123-152.
Cochrane, J. (1996): A Cross-Sectional Test of an Investment-Based Asset Pricing Model.
Journal of Political Economy 104, 572-621.
Constantinides, G. M. and D. Due (1996): Asset Pricing with Heterogeneous Consumers.
Journal of Political Economy 104, 219-40.
Ferson, W. E. and A. F. Siegel (2003): Stochastic Discount Factor Bounds with Conditioning
Information. Review of Financial Studies 16, 567-595.
Gallant, R. A., L. P. Hansen and G. Tauchen (1990): Using the Conditional Moments of
Asset Payos to Infer the Volatility of Intertemporal Marginal Rates of Substitution.
Journal of Econometrics 45, 141-179.
Hansen, L. P. and R. Jagannathan (1991): Implications of Security Market Data for Models
of Dynamic Economies. Journal of Political Economy 99, 225-262.
Mehra, R. and E. C. Prescott (1985): The Equity Premium: A Puzzle. Journal of Monetary
Economics 15, 145-161.
Roll, R. (1977): A Critique of the Asset Pricing Theorys Tests Part I: On Past and Potential
Testability of the Theory. Journal of Financial Economics 4, 129-176.
194
7
Aggregate stock market fluctuations
7.1 Introduction
This chapter reviews the progress made to address the empirical puzzles relating to the neoclassical asset pricing model. We rst provide a succinct overview of the main empirical regularities
of aggregate stock market uctuations. For example, we emphasize that price-dividend ratios
and returns are procyclical, and that returns volatility and risk-premia are both time-varying
and countercyclical. Then, we discuss the extent to which these empirical features can be explained by rational models. For example, many models with state dependent preferences predict
that Sharpe ratios are time-varying and that stock market volatility is countercyclical. Are these
appealing properties razor-edge? Or are they general properties of all conceivable models with
state-dependent preferences? Moreoover, would we expect that these properties show up in
other related models in which asset prices are related to the economic conditions? The nal
part of this chapter aims at providing answers to these questions, and develops theoretical test
conditions on the pricing kernel and other primitive state processes that make the resulting
models consistent with sets of qualitative predictions given in advance.
7.2 The empirical evidence

One fascinating aspect of stock market behavior lies in a number of empirical regularities closely
related to the business cycle. These empirical regularities seem to persist after controlling for
the sample period, and are qualitatively the same in all industrialized countries. These pieces of
evidence are extensively surveyed in Campbell (2003). In this section, we summarize the salient
empirical features of aggregate US stock market behavior. We rely on the basic statistics and
statistical models presented in Table 7.1. There exist three fundamental sets of stylized facts
that deserve special attention.
Fact 1. P/D, P/E ratios and returns are strongly procylical, but variations in the general business
cycle conditions do not seem to be the only force driving these variables.
For example, Figure 7.1 reveals that price-dividend ratios decline at all NBER recessions.
But during NBER expansions, price-dividend ratios seem to be driven by additional factors not
c
by
A. Mele
7.2. The empirical evidence
necessarily related to the business cycle conditions. As an example, during the roaring 1960s,
price-dividend ratios experienced two major drops having the same magnitude as the decline
at the very beginning of the chaotic 1970s. Ex-post returns follow approximately the same
pattern, but they are more volatile than price-dividend ratios (see Figure 7.2).1
A second set of stylized facts is related to the rst two moments of the returns distribution:
Fact 2. Returns volatility, the equity premium, risk-adjusted discount rates, and Sharpe ratios
are strongly countercyclical. Again, business cycle conditions are not the only factor explaining both short-run and long-run movements in these variables.
Figures 7.3 through 7.5 are informally very suggestive of the previous statement. For example,
volatility is markedly higher during recessions than during expansions. (It also appears that the
volatility of volatility is countercyclical.) Yet it rocketed to almost 23% during the 1987 crash a crash occurring during one of the most enduring post-war expansions period. As we will see
later in this chapter, countercyclical returns volatility is a property that may emerge when the
volatility of the P/D ratio changes is countercyclical. Table 7.1 reveals indeed that the P/D
ratios variations are more volatile in bad times than in good times. Table 7.1 also reveals that
the P/D ratio (in levels) is more volatile in good times than in bad times. Finally, P/E ratios
behave in a dierent manner.
A third set of very intriguing stylized facts regards the asymmetric behavior of some important
variables over the business cycle:
Fact 3. P/D ratio changes, risk-adjusted discount rates changes, equity premium changes, and
Sharpe ratio changes behave asymmetrically over the business cycle. In particular, the
deepest variations of these variables occur during the negative phase of the business cycle.
As an example, not only are risk-adjusted discount rates counter-cyclical. On average, riskadjusted discount rates increase more during NBER recessions than they decrease during NBER
expansions. Analogously, not only are P/D ratios procyclical. On average, P/D increase less
during NBER expansions than they decrease during NBER recessions. Furthermore, the order
of magnitude of this asymmetric behavior is very high. As an example, the average of P/D
percentage (negative) changes during recessions is almost twice as the average of P/D percentage
(positive) changes during expansions. It is one objective in this chapter to connect this sort of
concavity of P/D ratios (with respect to the business cycle) to convexity of risk-adjusted
discount rates.2
1 We
use smoothed ex-post returns to eliminate the noise inherent to high frequency movements in the stock-market.
of changes in risk-premia related objects appears to be higher during expansions. This is probably a conservative
view because recessions have occurred only 16% of the time. Yet during recessions, these variables have moved on average more
than they have done during good times. In other terms, economic time seems to move more fastly during a recession than during
an expansion. For this reason, the physical calendar time-based standard deviations in Table 6.1 should be rescaled to reect
unfolding of economic calendar time. In this case, a more appropriate concept of volatility would be the standard deviation
average of expansions/recessions time.
2 Volatility
196
c
by
A. Mele
P/D
P/E
P/Dt+1 P/Dt
P/D
log P/Dt+1 100 12
t
P/Et+1 P/Et
P/E
log P/Et+1 100 12
t
log DDt+1
100 12
t
returns (real)
smooth returns (real)
risk-free rate (nominal)
risk-free rate (real)
excess returns
smooth excess returns
equity premium ()
t t1
t
100
log t1
risk-adjusted rate (Disc)
Disct Disct1
Disct
log Disc
100
t1
excess returns volatility
Sharpe ratio (Sh)
Sht Sht1
t
log ShSht1
100
total
average std dev
33.951
15.747
16.699
6.754
0.063
1.357
2.34
42.228
0.033
0.762
2.208
47.004
4.901
5.655
7.810
52.777
8.143
16.303
5.253
2.762
1.346
3.097
6.464
52.387
6.782
15.802
8.773
3.459
0.002
0.544
0.023
8.243
10.296
3.630
0.004
0.574
0.048
6.553
14.938
3.026
0.586
0.192
0.0001 0.053
0.027 10.129
NBER expansions
average std dev
35.085
15.538
17.210
6.429
0.118
1.255
4.032
37.236
0.0003
0.709
10.512
41.688
5.210
5.710
9.414
48.423
11.997
13.342
5.019
2.440
1.416
2.866
7.998
48.125
10.560
12.828
8.138
3.215
0.087
0.506
0.788
8.537
9.635
3.376
0.091
0.530
0.711
6.625
14.458
2.711
0.565
0.194
0.003
0.051
0.548
10.260
NBER recessions
average std dev
28.162
15.451
14.089
7.672
0.219
1.758
6.264 60.876
0.201
0.967
8.964
67.368
3.324
5.049
0.377 70.183
11.529 15.758
6.442
3.796
0.988
4.045
1.366 69.491
12.501 15.395
12.012
2.758
0.459
0.496
4.163
4.644
13.669
2.917
0.493
0.536
3.919
4.438
17.386
3.335
0.696
0.139
0.017
0.061
2.630
8.910
TABLE 7.1. P/D and P/E are the

S&P Comp. price-dividends and price-earnings ratios. Smooth
12
St +Dt
returns as of time t are dened as i=1 (Rti Rti ), where Rt = log( St1 ), and R is the risk-free
rate. Volatility is the excess returns volatility. With the exception of the P/D and P/E ratios, all
gures are annualized percent. Data are sampled monthly and cover the period from January 1954
through December 2002. Time series estimates of equity premium t (say), excess return volatility t
(say) and Sharpe ratios t / t are obtained through Maximum Likelihood estimation (MLE) of the
following model,

t Rt = t + t , t | Ft1 N 0, 2t
R
t = 0.162 + 0.766 t1 0.146 IPt1 ; t = 0.218 + 0.106 |t1 | + 0.868 t1
(0.004)
(0.005)
(0.015)
(0.089)
(0.010)
(0.029)
where (robust) standard errors are in parentheses; IP is the US real, seasonally adjusted industrial
production rate; and IP is generated by IPt = 0.2IPt1 +0.8IPt1 . Analogously, time series estimates
of the risk-adjusted discount rate Disc (say) are obtained by MLE of the following model,

t int = Disct + ut , ut | Ft1 N 0, vt2
R
Disct = 0.191 + 0.767 Disct1 0.152 IPt1 ; vt = 0.214 + 0.105 |ut1 | + 0.869 vt1
(0.036)
(0.042)
(0.081)
(0.012)
where int is the ination rate as of time t.
197
(0.004)
(0.003)
c
by
A. Mele
100
pt
p t
p t
p t
pt
pt p t
pt
pt
75
P/D ratio
50
25
P/E ratio
0
1954 1958 1962 1966 1970 1974 1978 1982 1986 1990 1994 1998 2002
FIGURE 6.1. P/D and P/E ratios
60
p t
pt
p t
p t
pt
pt p t
pt
p t
40
20
-20
-40
-60
1954 1958 1962 1966 1970 1974 1978 1982 1986 1990 1994 1998 2002
FIGURE 6.2. Monthly smoothed excess returns (%)
198
c
by
A. Mele
27.5
p t
p t
p t
p t
pt
pt p t
p t
p t
25.0
22.5
20.0
17.5
15.0
12.5
10.0
7.5
1954 1958 1962 1966 1970 1974 1978 1982 1986 1990 1994 1998 2002
FIGURE 6.3. Excess returns volatility
25
pt
pt
p t
p t
pt
pt p t
pt
p t
20
15
10
equity premium
0
long-averaged industrial production rate
-5
1954 1958 1962 1966 1970 1974 1978 1982 1986 1990 1994 1998 2002
FIGURE 6.4. Equity premium and long-averaged real industrial production rate
199
c
by
A. Mele
1.4
pt
p t
p t
pt
pt p
pt
pt
1.2
1.0
0.8
0.6
0.4
0.2
0.0
1954 1958 1962 1966 1970 1974 1978 1982 1986 1990 1994 1998 2002
FIGURE 6.5. Sharpe ratio

Stylized fact 1 has a simple and very intuitive consequence: price-dividend ratios are somewhat related to, or predict, future medium-term returns. The economic content of this prediction is simple. After all, expansions are followed by recessions. Therefore in good times the
stock market predicts that in the future, returns will be negative. Indeed, dene the excess
e R
t Rt . Consider the following regressions,
return as R
t
e
t+n
R
= an + bn P/Dt + un,t ,
n 1,
2
where u is a residual term. Typically, the estimates of bn are signicantly negative, and the
R on
et+n P/Dt ] =
these regressions increases with n. In turn, the previous regressions imply that E[ R
an + bn P/Dt . They thus suggest that price-dividend ratios are driven by expected excess
returns. In this restrictive sense, countercyclical expected returns (stated in stylized fact 2) and
procyclical price-dividend ratios (stated in stylized fact 1) seem to be the two sides of the same
coin.
There is also one apparently puzzling feature: price-dividend ratios do not predict future
dividend growth. Let gt log(Dt / Dt1 ). In regressions of the following form,
gt+n = an + bn P/Dt + un,t ,
n 1,
the predictive content of price-dividend ratios is very poor, and estimates of bn even come with
a wrong sign.
The previous simple regressions thus suggest that: 1) price-dividend ratios are driven by timevarying expected returns (i.e. by time-varying risk-premia); and 2) the role played by expected
dividend growth seems to be somewhat limited. As we will see later in this chapter, this view
can however be challenged along several dimensions. First, it seems that expected earning
200
c
by
A. Mele
7.3. Understanding the empirical evidence
growth does help predicting price-dividend ratios. Second, the fact that expected dividend
growth doesnt seem to aect price-dividend ratios can in fact be a property to be expected in
equilibrium.
7.3 Understanding the empirical evidence

Consider the following decomposition,
t+1 log St+1 + Dt+1 = gt+1 + log 1 + pt+1 ;
log R
St
pt
where gt log
Dt
St
and pt
.
Dt1
Dt
The previous formula reveals that properties of returns can be understood through the corresponding properties of dividend growth gt and price-dividend ratios pt . The empirical evidence
discussed in the previous section suggests that our models should take into account at least
the following two features. First, we need volatile price-dividend ratios. Second we need that
price-dividend ratios be on average more volatile in bad times than in good times. For example, consider a model in which prices are aected by some key state variables related to the
business cycle conditions (see Section 7.4 for examples of models displaying this property). A
basic property that we should require from this particular model is that the price-dividend
ratio be increasing and concave in the state variables related to the business cycle conditions.
In particular, the concavity property ensures that returns volatility increases on the downside
- which is precisely the very denition of countercyclical returns volatility. One of the ultimate
scopes in this chapter is to search for classes of promising models ensuring this and related
properties.
The Gordons model in Chapter 6 predicts that price dividend ratios are constant - which is
counterfactual. It is thus unsuitable for the scopes we are pursuing here. We need to think of
multidimensional models. However, not all multidimensional models will work. As an example,
in the previous chapter we showed how to arbitrarily increase the variance of the kernel of the
Lucas model by adding more and more factors. We also showed that the resulting model is
one in which price-dividend ratios are constant. We need to impose some discipline on how to
increase the dimension of a model.
7.4 The asset pricing model

7.4.1
A multidimensional model
Consider a reduced-form rational expectation model in which the rational asset price Si is a
twice-dierentiable function of a number of factors,
Si = Si (y),
y Rd ,
i = 1, , m (m d),
where y = [y1 , , yd ] is the vector of factors aecting asset prices, and Si is the rational pricing
function. We assume that asset i pays o an instantaneous dividend rate Di , i = 1, , m, and
that Di = Di (y), i = 1, , m. We also assume that y is a multidimensional diusion process,
viz
dyt = (yt )dt + v(yt )dWt ,
201
c
by
A. Mele
7.4. The asset pricing model
where is d-valued, v is d d valued, and W is a d-dimensional Brownian motion. By Itos

lemma,
1d
dd
Si
dSi
LSi
v dW,
=
dt +
Si
Si
Si
where LSi is the usual innitesimal operator. Let r = {rt }t0 be the instantaneous short-term
rate process. We now invoke the FTAP to claim that under mild regularity conditions, there
exists a measurable d-vector process (unit prices of risk) such that,
LS
S
D1
1
1
r
+
S1
S1
S1
..

,
where = ... v.
(7.1)
=
.
LSm
Dm
S
m
md d1
r + Sm
Sm
Sm
The usual interpretation of is the vector of unit prices of risk associated with the uctuations
of the d factors. To simplify the structure of the model, we suppose that,
r r (yt )
and (yt ).
(7.2)
The previous assumptions impose a series of severe restrictions on the dimension of the model.
We emphasize that these restrictions are arbitrary, and that they are only imposed for simplicity
sake.
Eqs. (7.1) constitute a system of m uncoupled partial dierential equations. The solution to
it is an equilibrium price system. For example, the Gordons model in Chapter 6 is a special case
of this setting.3 We do not discuss transversality conditions and bubbles in this chapter. Nor
we discuss issues related to market completeness.4 Instead, we implement a reverse-engineering
approach and search over families of models guaranteeing that long-lived asset prices exhibit
some properties given in advance. In particular, we wish to impose conditions on the primitives
P (a, b, r, ) such that the aggregate stock market behavior exhibits the same patterns
surveyed in the previous section. For example, model (7.1) predicts that returns volatility is,

1
dSi,t
volatility
V(yt )dt 2 Si (yt )v(yt )2 dt.
Si,t
Si,t
In this model volatility is thus typically time-varying. But we also wish to answer questions such
as, Which restrictions may we impose to P to ensure that volatility V(yt ) is countercyclical?
Naturally, an important and challenging subsequent step is to nd models guaranteeing that
the restrictions on P we are looking for are economically and quantitatively sensible. The most
natural models we will look at are models which innovate over the Gordons model due to
time-variation in the expected returns and/or in the expected dividend growth. These issues
are analyzed in a simplied version of the model.
7.4.2
A simplified version of the model
We consider a pure exchange economy endowed with a ow of a (single) consumption good. To

make the presentation simple, we assume that consumption equals the dividends paid by a single
3 Let m = d = 1, = y, (y) = y and (y) = y, and assume that and r are constant. By replacing these things into eq. (7.1)
0
and assuming no-bubbles yields the (constant) price-dividend ratio predicted by the Gordons model, qt / t = ( r 0 )1 .
4 As we explained in chapter 4, in this setting markets are complete if and only if m = d.
202
c
by
A. Mele
long-lived asset (see below). We also assume that d = 2, and take as given the consumption
endowment process D and a second state variable y. We assume that D, y are solution to,
+
dD( ) = m (y( )) D ( ) d + 0 D( )dW1 ( )
dy( ) = (D( ), y( ))d + v1 (y( )) dW1 ( ) + v2 (y( )) dW2 ( )
where W1 and W2 are independent standard Brownian motions. By the connection between
conditional expectations and solutions to partial dierential equations (the Feynman-Kac representation theorem) (see Chapter 4), we may re-state the FTAP in (7.1) in terms of conditional
expectations in the following terms. By (7.1), we know that S is solution to,
LS + D = rS + (SD 0 D + Sy v1 ) 1 + Sy v2 2 , (D, y) Z Y.
(7.3)
Under regularity conditions, the Feynman-Kac representation of the solution to Eq. (7.3) is:

"

"

S(D, y) =
C(D, y, )d ,
C(D, y, ) E exp
r(D(t), y(t))dt D( ) D, y ,
0
(7.4)
where E is the expectation operator taken under the risk-neutral probability Q (say). Finally,
(Z, Y ) are solution to
5
1 ( )
dD( ) = m(y(
))D( )d + 0 D( )dW
1 ( ) + v2 (y( )) dW
2 ( )
dy( ) =
(D( ), y( ))d + v1 (y( )) dW
(7.5)
1 and W
2 are two independent Q-Brownian motions, and m
where W
and
are risk-adjusted drift
functions dened as m
(D, y) m(y)D 0 D1 (D, y) and
(D, y) (D, y)v1 (y) 1 (D, y)
v2 (y) 2 (D, y).6 Naturally, Eq. (7.4) can also be rewritten under the physical measure. We have,

"

C(D, y, ) = E exp
r(D(t), y(t))dt D( ) D, y = E [ ( ) D( )| D, y] ,
0
where is the stochastic discount factor of the economy:

( ) =
( )
;
(0)
(0) = 1.
Given the previous assumptions on the information structure of the economy, necessarily
satises,
d( )
= [r(D( ), y( ))d + 1 (D( ), y( ))dW1 ( ) + 2 (D( ), y( ))dW2 ( )] .
( )
(7.6)
In the appendix (Markov pricing kernels), we provide an example of pricing kernel generating
interest rates and risk-premia having the same functional form as in (7.2).
5 See,
for example, Huang and Pag`

es (1992) (thm. 3, p. 53) or Wang (1993) (lemma 1, p. 202), for a series of regularity conditions
underlying the Feynman-Kac theorem in innite horizon settings arising in typical nancial applications.
6 See, for example, Huang and Pag`
es (1992) (prop. 1, p. 41) for mild regularity conditions ensuring that Girsanovs theorem
holds in innite horizon settings.
203
c
by
A. Mele

7.4.3
Issues
We analyze general properties of long-lived asset prices that can be streamlined into three categories: monotonicity properties, convexity properties, and dynamic stochastic dominance
properties. We now produce examples illustrating the economic content of such a categorization.
Monotonicity. Consider a model predicting that S(D, y) = Dp(y), for some positive func (y)
tion p C 2 (Y). By Itos lemma, returns volatility is vol(D)+ pp(y)
vol(y), where vol(D) > 0
is consumption growth volatility and vol(y) has a similar interpretation. As explained in
the previous chapter, actual returns volatility is too high to be explained by consumption
volatility. Naturally, additional state variables may increase the overall returns volatility.
In this simple example, state variable y inates returns volatility whenever the pricedividend ratio p is increasing in y. At the same time, such a monotonicity property would
ensure that asset returns volatility be strictly positive. Eventually, strictly positive volatility is one crucial condition guaranteeing that dynamic constraints of optimizing agents
are well-dened.
Concavity. Next, suppose that y is some state variable related to the business cycle conditions. Another robust stylized fact is that stock market volatility is countercyclical. If
S(D, y) = D p(y) and vol(y) is constant, returns volatility is countercyclical whenever p
is a concave function of y. Even in this simple example, second-order properties (or nonlinearities) of the price-dividend ratio are critical to the understanding of time variation
in returns volatility.
Convexity. Alternatively, suppose that expected dividend growth is positively aected by
a state variable g. If p is increasing and convex in y g, price-dividend ratios would
typically display overreaction to small changes in g. The empirical relevance of this
point was rst recognized by Barsky and De Long (1990, 1993). More recently, Veronesi
(1999) addressed similar convexity issues by means of a fully articulated equilibrium model
of learning.
Dynamic stochastic dominance. An old issue in nancial economics is about the relation
between long-lived asset prices and volatility of fundamentals.7 The traditional focus of the
literature has been the link between dividend (or consumption) volatility and stock prices.
Another interesting question is the relationship between the volatility of additional state
variables (such as the dividend growth rate) and stock prices. In some models, volatility
of these additional state variables is endogenously determined. For example, it may be
inversely related to the quality of signals about the state of the economy.8 In many other
circumstances, producing a probabilistic description of y is as arbitrary as specifying the
preferences of a representative agent. In fact, y is in many cases related to the dynamic
specication of agents preferences. The issue is then to uncover stochastic dominance
properties of dynamic pricing models where state variables are possible nontradable.
In the next section, we provide a simple characterization of the previous properties. To achieve
this task, we extend some general ideas in the recent option pricing literature. This literature
7 See,
8 See,
for example, Malkiel (1979), Pindyck (1984), Poterba and Summers (1985), Abel (1988) and Barsky (1989).
for example, David (1997) and Veronesi (1999, 2000)
204
c
by
A. Mele
7.5. Analyzing qualitative properties of models
attempts to explain the qualitative behavior of a contingent claim price function C(D, y, ) with
as few assumptions as possible on D and y. Unfortunately, some of the conceptual foundations
in this literature are not well-suited to pursue the purposes of this chapter. As an example,
many available results are based on the assumption that at least one state variable is tradable.
This is not the case of the European-type option pricing problem (7.4). In the next section,
we introduce an abstract asset pricing problem which is appropriate to our purposes. Many
existing results are specic cases of the general framework developed in the next section (see
Theorems 7.1 and 7.2). In sections 7.6 and 7.7, we apply this framework of analysis to study
basic model examples of long-lived asset prices.
7.5 Analyzing qualitative properties of models

Consider a two-period, risk-neutral environment in which there is a right to receive a cash
premium at the second period. Assume that interest rates are zero, and that the cash premium
is a function of some random variable x, viz = (
x). Finally, let c E[(
x)] be the price
of this right. What is the relationship between the volatility of x and c? By classical secondorder stochastic dominance arguments (see appendix to Chapter 1), c is inversely related to
mean preserving spreads in x if is concave. Intuitively, this is so because a concave function
exaggerates poor realizations of x and dampens the favorable ones.
Do stochastic dominance properties still hold in a dynamic setting? Consider for example
a multiperiod, continuous time extension of the previous risk-neutral environment. Assume
that the cash premium is paid o at some future date T , and that x = x(T ), where X =
{x( )} [0,T ] (x(0) = x) is some underlying state process. If the yield curve is at at zero,
c(x) E[(x(T ))| x] is the price of the right. Clearly, the pricing problem E[(x(T ))| x] is
dierent from the pricing problem E[(
x)]. Analogies exist, however. First, if X is a proportional
process (one for which the risk-neutral distribution of x(T )/x is independent of x),
c(x) = E [(x G(T ))] ,
G(T )
x(T )
,
x
x > 0.
As this simple formula reveals, standard stochastic dominance arguments still apply: c decreases
(increases) after a mean-preserving spread in G whenever is concave (convex) - consistently
for example with the prediction of the Black and Scholes (1973) formula. This point was rst
made by Jagannathan (1984) (p. 429-430). In two independent papers, Bergman, Grundy and
Wiener (1996) (BGW) and El Karoui, Jeanblanc-Picque and Shreve (1998) (EJS) generalized
these results to any diusion process (i.e., not necessarily a proportional process).9,10 But one
crucial assumption of these extensions is that X must be the price of a traded asset that
does not pay dividends. This assumption is crucial because it makes the risk-neutralized drift
function of X proportional to x. As a consequence of this fact, c inherits convexity properties of
, as in the proportional process case. As we demonstrate below, the presence of nontradable
9 The
proofs in these two articles are markedly distinct but are both based on price function convexity. An alternate proof
directly based on payoff function convexity can be obtained through a direct application of Hajeks (1985) theorem. This theorem
states that if is increasing and convex, and X1 and X2 are two diusion processes (both starting o from the same origin) with
integrable drifts b1 and b2 and volatilities a1 and a2 , then E[(x1 (T ))] E[(x2 (T ))] whenever m2 ( ) m1 ( ) and a2 ( ) a1 ( )
for all (0, ). Note that this approach is more general than the approach in BGW and EJS insofar as it allows for shifts in
both m and a. As we argue below, both shifts are important to account for when X is nontradable.
10 Bajeux-Besnainou and Rochet (1996) (section 5) and Romano and Touzi (1997) contain further extensions pertaining to
stochastic volatility models.
205
c
by
A. Mele
state variables makes interesting nonlinearites emerge. As an example, Proposition 7.1 reveals
that in general, convexity of is neither a necessary or a sucient condition for convexity of
c.11 Furthermore, dynamic stochastic dominance properties are more intricate than in the
classical second order stochastic dominance theory (see Proposition 7.1).
To substantiate these claims, we now introduce a simple, abstract pricing problem.
Canonical pricing problem. Let X be the (strong) solution to:
( ),
dx( ) = b (x( )) d + a (x( )) dW
is a multidimensional P -Brownian motion (for some P ), and b, a are some given
where W
functions. Let and be two twice continuously differentiable positive functions, and define

" T

exp
c(x, T ) E
(x(t))dt (x(T )) x
(7.7)
0
to be the price of an asset which promises to pay (x(T )) at time T .
In this pricing problem, X can be the price of a traded asset. In this case b(x) = x(x). If in
addition, = 0, the problem collapses to the classical European option pricing problem with
constant discount rate. If instead, X is not a traded risk, b(x) = b0 (x)a(x)(x), where b0 is the
physical drift function of X and is a risk-premium. The previous framework then encompasses
a number of additional cases. As an example, set (x) = x. Then, one may 1) interpret X as
consumption #process; 2) restrict a long-lived asset price S to be driven by consumption only,
and set S = 0 c(x, )d . As another example, set (x) = 1 and (x) = x. Then, c is a zerocoupon bond price as predicted by a simple univariate short-term rate model. The importance
of these specic cases will be claried in the following sections.
In the appendix (see Proposition 7.A.1), we provide a result linking the volatility of the state
variable x to the price c. Here I characterize slope (cx ) and convexity (cxx ) properties of c. We
have:
Proposition 7.1. The following statements are true:
a) If > 0, then c is increasing in x whenever 0. Furthermore, if = 0, then c is
decreasing (resp. increasing) whenever > 0 (resp. < 0).
b) If 0 (resp. 0) and c is increasing (resp. decreasing) in x, then c is concave
(resp. convex ) in x whenever b < 2 (resp. b > 2 ) and 0 (resp. 0). Finally, if
b = 2 , c is concave (resp. convex ) whenever < 0 (resp. > 0) and 0 (resp. 0).
Proposition 7.1-a) generalizes previous monotonicity results obtained by Bergman, Grundy
and Wiener (1996). By the so-called no-crossing property of a diusion, X is not decreasing
in its initial condition x. Therefore, c inherits the same monotonicity features of if discounting
does not operate adversely. While this observation is relatively simple, it explicitly allows to
address monotonicity properties of long-lived asset prices (see the next section).
Proposition 7.1-b) generalizes a number of existing results on option price convexity. First,
assume that is constant and that X is the price of a traded asset. In this case, = b = 0.
11 Kijima
(2002) recently produced a counterexample in which option price convexity may break down in the presence of convex
payo functions. His counterexample was based on an extension of the Black-Scholes model in which the underlying asset price had
a concave drift function. (The source of this concavity was due to the presence of dividend issues.) Among other things, the proof
of proposition 2 reveals the origins of this counterexample.
206
c
by
A. Mele
The last part of Proposition 7.1-b) then says that convexity of propagates to convexity of
c. This result reproduces the ndings in the literature that surveyed earlier. Proposition 7.1b) characterizes option price convexity within more general contingent claims models. As an
example, suppose that = = 0 and that X is not a traded risk. Then, Proposition 7.1-b)
reveals that c inherits the same convexity properties of the instantaneous drift of X. As a nal
example, Proposition 7.1-b) extends one (scalar) bond pricing result in Mele (2003). Precisely,
let (x) = 1 and (x) = x; accordingly, c is the price of a zero-coupon bond as predicted by
a standard short-term rate model. By Proposition 7.1-b), c is convex in x whenever b (x) < 2.
This corresponds to Eq. (8) (p. 688) in Mele (2003).12 In analyzing properties of long-lived asset
prices, both discounting and drift nonlinearities play a prominent role.
An intuition of the previous result can be obtained through a Taylor-type expansion of c (x, T )
in Eq. (7.7). To simplify, suppose that in Eq. (7.7), 1, and that
(x) = g (x) Disc (x) .
The economic interpretation of the previous decomposition is that g is the growth rate of some
underlying dividend process and Disc is some risk-adjusted discount rate. Consider the
following discrete-time counterpart of Eq. (7.7):
c(x0 , N) E
For small values of the exponent,
,
+ N

[g(x
)Disc(x
)]
i
i
x0 .
e i=0

c(x, N ) 1 + [g (x) Disc (x)] N +
N
i

i=0 j=1
[g (xj ) Disc (xj )| x] ,

E
(7.8)
where g (xj ) g (xj ) g (xj1 ). The second term of the r.h.s of Eq. (7.8) make clear that
convexity of g can potentially translate to convexity of c w.r.t x; and that convexity of Disc
can potentially translate to concavity of c w.r.t x. But Eq. (7.8) reveals that higher order terms
[g (xj ) Disc (xj )| x] plays some role. Intuare important too. Precisely, the expectation E
itively, convexity properties of c w.r.t x also depend on convexity properties of this expectation.
In discrete time, these things are dicult to see. But in continuous time, this simple observation
translates to a joint restriction on the law of movement of x. Precisely, convexity properties
of c w.r.t x will be somehow inherited by convexity properties of the drift function of x. In
continuous time, Eq. (7.8) becomes, for small T ,13
c(x,
+ T ) 1 + [g (x) Disc (x)] T
,
d
1
d2
2
2
+ [g (x) Disc (x)] + b (x)
[g (x) Disc (x)] + a (x) 2 [g (x) Disc (x)] T 2 .
dx
2
dx
(7.9)
Naturally, this formula is only an approximation. Importantly, it doesnt work very well for
large T . Proposition 7.1 gives the exact results.
12 In
the appendix, we have developed further intuition on this bounding number.

(7.9) can derived with operation-theoretic arguments based on the functional iteration of the innitesimal generator for
Markov processes.
13 Eq.
207
7.6. Time-varying discount rates and equilibrium volatility
c
by
A. Mele
7.6 Time-varying discount rates and equilibrium volatility

Campbell and Cochrane (1999) model of external habit formation is certainly one of the most
well-known attempts at explaining some of the empirical features outlined in Section 7.2. Consider an innite horizon, complete markets economy. There is a (representative) agent with
undiscounted instantaneous utility given by
(c x)1 1
u(c, x) =
,
1
(7.10)
where c is consumption and x is a (time-varying) habit, or (exogenous) subsistence level. The

total endowment process Z = {D( )}0 satises,
dD ( )
= g0 d + 0 dW ( ) .
D ( )
(7.11)
In equilibrium, C = Z. Let s (D x)/D, the surplus consumption ratio. By assumption,

S = {s( )} 0 is solution to:

1 2
2
ds( ) = s( ) (1 )(
s log s( )) + 0 l(s( )) d + 0 s( )l(s( ))dW ( ),
(7.12)
2
where l is a positive function given in appendix 6.2. This function l turns out to be decreasing
in s; and convex in s for the empirically relevant range of variation of s. The Sharpe ratio
predicted by the model is:
(7.13)
(s) = 0 [1 + l(s)]
(see appendix 6.2 for details). Finally, Campbell and Cochrane choose function l so as to make
the short-term rate constant.
The economic interpretation of the model is relatively simple. Consider rst the instantaneous
utility in (7.10). It is readily seen that CRRA = s1 . That is, risk aversion is countercyclical in
this model. Intuitively, during economic downturns, the surplus consumption ratio s decreases
and agents become more risk-averse. As a result, prices decrease and expected returns increase.
Very nice, but this is still a model of high risk-aversion. Furthermore, Barberis, Huang and
Santos (2001) have a similar mechanism based on behavioral preferences.
The rationale behind Eq. (7.12) is simply that the log of s is a mean reverting process. By taking logs, we are sure that s remains positive. Finally, log s is also conditionally heteroskedastic
since its instantaneous volatility is 0 l. Because l is decreasing in s and s is clearly pro-cyclical,
the volatility of log s is countercyclical. This feature is responsible of many interesting properties
of the model (such as countercyclical returns volatility).
Finally, the Sharpe ratio in (7.13) is made up of two components. The rst component
is 0 and coincides with the Sharpe ratio predicted by the standard Gordons model. The
additional component 0 l(s) arises as a compensation related to the stochastic uctuations of
x = D(1s). By the functional form of l picked up by the authors, is therefore countercyclical.
Since is high, this generates a slowly-varying, countercyclical risk-premium in stock returns.
Finally, numerical results revealed that one important prediction of the model is that the pricedividend ratio is a concave function of s. In appendix 6.3, we provide an algorithm that one
may use to solve this and related models numerically (we only consider discrete time models in
appendix 6.3). Here we now clarify the theoretical link between convexity of l and concavity of
the price-dividend ratio in this and related models.
208
c
by
A. Mele
We aim at writing the solution in the canonical pricing problem format of Section 7.5, and
then at applying Proposition 7.1. Our starting point is the evaluation formula (7.4). To apply
it here, we might note that interest rate are constant. Yet to gain in generality we continue
to assume that they are state dependent, but that they only depend on s. Therefore Eq. (7.4)
becomes,

"
"

S(D, s)
C(D, s, )
D(
)
0 r(s(u))du

=
d =
E e
D, s d .
(7.14)
D
D
D
0
0
To compute the inner expectation, we have to write the dynamics of Z under the risk-neutral
probability measure. By Girsanov theorem,

1 2
D( )
= e 2 0 +0 W ( ) eg0 0 0 (s(u))du ,
D
is a Brownian motion under the risk-neutral measure. By replacing this into Eq.
where W
(7.14),
"

!

1 2
S(D, s)

=
eg0 E e 2 0 +0 W ( ) e 0 Disc(s(u))du D, s d ,
(7.15)
D
0
where
Disc (s) r (s) + 0 (s)
is the risk-adjusted discount rate. Note also, that under the risk-neutral probability measure,
( ),
ds( ) =
(s( )) d + v(s ( ))dW
where
(s) = (s) v (s) (s), (s) = s[(1 )(
s log s) + 12 20 l(s)2 ] and v (s) = 0 sl(s).
Eq. (7.15) reveals that the price-dividend ratio p (D, s) S (D, s)/ D is independent of D.
1 2
Therefore, p (D, s) = p (s). To obtain a neat formula, we should also get rid of the e 2 0 +0 W ( )
term. Intuitively this term arises because consumption and habit are correlated. A convenient
change of measure will do the job. Precisely, dene a new probability measure P (say) through

1 2
the Radon-Nikodym derivative dP dP = e 2 0 +0 W ( ) . Under this new probability measure,

the price-dividend ratio p (s) satises,
"
!

e 0 Disc(s(u))du s d ,
p (s) =
eg0 E
(7.16)
0
and
( ),
ds( ) =
(s( )) d + v(s ( ))dW
( ) = W
( ) 0 is a P -Brownian motion, and
(s) = (s) v (s) (s) + 0 v (s).
where W
The inner expectation in Eq. (7.16) comes in exactly the same format as in the canonical
pricing problem of Section 7.5. Therefore, we are now ready to apply Proposition 7.1. We have,
1. Suppose that risk-adjusted discount rates are countercyclical, viz
d
price-dividend ratios are procyclical, viz ds
p (s) > 0.
d
Disc(s)
ds
0. Then
2. Suppose that price-dividend ratios are procyclical. Then price-dividend ratios are cond2
cave in s whenever risk-adjusted discount rates are convex in s, viz ds
2 Disc(s) > 0, and
d2
d
(s) 2 ds Disc(s).
ds2
209
c
by
A. Mele
So we have found joint restrictions on the primitives such that the pricing function p is
consistent with certain properties given in advance. What is the economic interpretation related
to the convexity of risk-adjusted discount rates? If price-dividend ratios are concave in some
state variable Y tracking the business cycle condititions, returns volatility increases on the
downside, and it is thus countercyclical (see Figure 7.6.) According to the previous predictions,
price-dividend ratios are concave in Y whenever risk-adjusted discount rates are decreasing
and sufficiently convex in Y . The economic signicance of convexity in this context is that in
good times, risk-adjusted discount rates are substantially stable; consequently, the evaluation
of future dividends does not vary too much, and price-dividend ratios are relatively stable. And
in bad times risk-adjusted discount rates increase sharply, thus making price-dividend ratios
more responsive to changes in the economic conditions.
Heuristically, the mathematics behind the previous results can be explained as follows. For
small , Eq. (7.9) is,
!

e 0 [g0 Disc(s(u))]du s 1 + [g0 Disc (s)] + h.o.t.
p (s, ) E
Hence convexity of Disc(s) translates to concavity of p (s, ). But as pointed out earlier, the
additional higher order terms matter too. The problem with these heuristic arguments is how
well the approximation works for small . Furthermore
p (s, ) is not the price-dividend ratio.
#
The price dividend ratio is instead p (s) = 0 p (s, ) d . Anyway the previous predictions
conrm that the intuition is indeed valid.
210
c
by
A. Mele
Price-dividend
ratio
Risk-adjusted
discount rates
good
times
bad
times
bad
times
good
times
FIGURE 7.6. Countercyclical return volatility
What does empirical evidence suggest? To date no empirical work has been done on this.
Here is a simple exploratory analysis. First it seems that real risk-adjusted discount rates Disct
are convex in some very natural index summarizing the economic conditions (see Figure 7.7).
In Table 7.2 , we also run Least Absolute Deviations (LAD) regressions to explore whether P/D
dividend ratios are concave functions in IP.14 And we run LAD regressions in correspondence of
three sample periods to better understand the role of the exceptional (yet persistent) increase
in the P/D ratio during the late 1990s. Figure 7.8 depicts scatter plots of data (along with
tted regressions) related to these three sampling periods.
14 We run LAD regressions because this methodology is known to be more robust to the presence of outliers than Ordinary Least
Squares.
211
c
by
A. Mele
Expected Returns & Industrial Production
Predictive regression
16
Predicted Expected Returns (annualized, %)
Expected Stock Returns (annualized, %)
25
20
15
10
14
12
10
4
-1.2
-0.6
0.0
0.6
1.2
1.8
2.4
-1.2
Industrial Production Growth Rate (%)
-0.6
0.0
0.6
1.2
1.8
2.4
Industrial Production Growth Rate (%)
FIGURE 7.7.The left-hand side of this picture plots estimates of the expected returns (annualized,
percent) (Et say) against one-year moving averages of the industrial production growth (IPt ). The expected returns are estimated through the predictive regression of S&P returns on to default-premium,

< t 12 |Exct+1i | , where Exct is the return
term-premium and return volatility dened as Vol
i=1
2
12
in excess of the 1-month bill return as of month t. The one-year moving average of the industrial
1 12
production growth is computed as IPt 12
i=1 Indt+1i , where Indt is the real, seasonally adjusted
industrial production growth as of month t. The right-hand side of this picture depicts the prediction
of the static Least Absolute Deviations regression: Et = 8.56 4.05 IPt +1.18 IP2t + wt , where wt is a
(0.15)
(0.30)
(0.31)
residual term, and standard errors are in parenthesis. Data are sampled monthly, and span the period
from January 1948 to December 2002.
TABLE 7.2. Price-dividend ratios and economic conditions. Results of the LAD regression P/D =
a + bIP+cIP2 + w, where P/D is the S&P Comp. price-dividend ratio, IPt = (It + + It11 )/ 12; It
is the real, seasonally adjusted US industrial production growth rate, and wt is a residual term. Data
are sampled monthly, and cover the period from January 1948 through December 2002.
a
b
c
1948:01 - 1991:12
estimate std dev
27.968
0.311
2.187
0.419
2.428
0.429
1948:01 - 1996:12
estimate std dev
29.648
0.329
2.541
0.475
3.279
0.480
212
1948:01 - 2002:12
estimate std dev
30.875
0.709
3.059
1.074
3.615
1.091
7.7. Large price swings as a learning induced phenomenon
c
by
A. Mele
7.7 Large price swings as a learning induced phenomenon

Consider a static scenario in which consumption D is generated by D = + w, where and
w are independently distributed, with p Pr( = A) = 1 Pr( = A), and Pr(w = A) =
Pr(w = A) = 12 . Suppose that the state is unobserved. How would we update our prior
probability p of the good state upon observation of D? A simple application of the Bayes
Theorem gives the posterior probabilities Pr( = A| Di ) displayed in Table 7.3. Considered as
a random variable dened over observable states Di , the posterior probability Pr( = A| Di )
has expectation E [Pr ( = A| D)] = p and variance var [Pr ( = A| D)] = 12 p(1 p). Clearly,
this variance is zero exactly where there is a degenerate prior on the state. More generally, it
is a -shaped function of the a priori probability p of the good state. Since the lter,
g E ( = A| D)
is linear in Pr ( = A| D), the same qualitative conclusions are also valid for g.
Pr(Di )
Pr ( = A| D = Di )
Di (observable state)
D1 = 2A
D2 = 0
D3 = 2A
1
1
1
p
(1 p)
2
2
2
1
p
0
TABLE 7.3. Randomization of the posterior probabilities Pr ( = A| D) .
To understand in detail how we computed the values in Table 7.3, let us recall Bayes Theorem. Let (Ei )i be a partition of the state space . (This partition can be nite or uncountable,
i.e. the set of indexes i can be nite or uncountable - it really doesnt matter.) Then Bayes
Theorem says that,
Pr (Ei | F ) = Pr (Ei )
Pr ( F | Ei )
Pr ( F | Ei )
= Pr (Ei )
.
Pr (F )
j Pr ( F | Ej ) Pr (Ej )
(7.17)
By applying Eq. (7.17) to our example,
Pr ( = A| D = D1 ) = Pr ( = A)
Pr ( D = D1 | = A)
Pr ( D = D1 | = A)
=p
.
Pr (D = D1 )
Pr (D = D1 )
But Pr (D = D1 | = A) = Pr (w = D1 A) = Pr (w = A) = 12 . On the other hand, Pr (D = D1 ) =

1
p. This leaves Pr ( = A| D = D1 ) = 1. Naturally this is trivial, but one proceeds similarly to
2
compute the other probabilities.
The previous example conveys the main ideas underlying nonlinear ltering. However, it leads
to a nonlinear lter g which is somewhat distinct from the ones usually encountered in the
literature.15 In the literature, the instantaneous variance of the posterior probability changes
d (say) is typically proportional to 2 (1 )2 , not to (1 ). As we now heuristically
demonstrate, the distinction is merely technical. Precisely, it is due to the assumption that w is
a discrete random variable. Indeed, assume that w has zero mean and unit variance, and that
15 See,
e.g., Liptser and Shiryaev (2001a) (chapters 8 and 9).
213
c
by
A. Mele
it is absolutely continuous with arbitrary density function . Let (D) Pr ( = A| D dD).

By an application of the Bayesian learning mechanism in Eq. (7.17),
(D) = Pr ( = A)
Pr ( D dD| = A)
.
Pr ( D dD| = A) Pr ( = A) + Pr ( D dD| = A) Pr ( = A)
But Pr (D dD| = A) = Pr (w = D A) = (D A) and similarly, Pr (D dD| = A) =

Pr (w = D + A) = (D + A). Simple computations then leave,
(D) p = p(1 p)
(D A) (D + A)
.
p(D A) + (1 p)(D + A)
(7.18)
That is, the variance of the probability changes (D) p is proportional to p2 (1 p)2 .
To add more structure to the problem, we now assume that w is Brownian motion and set
A Ad . Let D0 D(0) = 0. In appendix, we show that by an application of Itos lemma to
(D),
d( ) = 2A ( )(1 ( ))dW ( ), (D0 ) p,
(7.19)
where dW ( ) dD( ) g( )d and g( ) E ( | D ( )) = [A( ) A(1 ( ))]. Naturally,
this construction is heuristic. Nevertheless, the result is correct.16 Importantly, it is possible to
show that W is a Brownian motion with respect to the agents information set (D(t), t ).17
Therefore, the equilibrium in the original economy with incomplete information is isomorphic
in its pricing implications to the equilibrium in a full information economy in which,
+
( )
dD( ) = [g( ) ( ) 0 ] d + 0 dW
(7.20)
( )
dg( ) = ( )v(g( ))d + v(g( ))dW
is a Q-Brownian motion, is a risk-premium process, v(g) (A g)(g + A)/ 0
where W
and 0 1.18 In fact, if the variance per unit of time of w is 20 , eqs. (7.20) hold for any
0 > 0.19 Furthermore, a similar result would hold had drift and diusion of Z been assumed
to be proportional to D. In this case, (Z, G) would be solution to
dD( ) = [g( ) ] d + dW
( )
0
0
(7.21)
D( )
dg( ) = (g( )) d + v (g( )) dW ( )

where = v and v is as before. In all cases, the instantaneous volatility of G is -shaped.
Under positive risk-aversion, this makes the risk-neutralized drift of Z a convex function of g.
The economic implications of this result are very important, and will be analyzed with the help
of Proposition 7.1.
System (7.20) is related to a model studied by Veronesi (1999). This model regards an innite horizon economy in which a representative agent with CARA = . This agent observes
realizations of Z generated by:
dD( ) = d + 0 dw1 ( ),
16 See,
for example, Liptser and Shiryaev (2001a) (theorem 8.1 p. 318; and example 1 p. 371).
Liptser and Shiryaev (2001a) (theorem 7.12 p. 273).
18 Such an isomorphic property has been pointed out for the rst time by Veronesi (1999) in a related model.

17 See
19 More
precisely, we have dW ( ) = 1
dz ( ) E | z (t)t d
0
214
= 1
0 (dz ( ) g ( ) d ).
(7.22)
c
by
A. Mele
where w1 is a Brownian motion, and is a two-states (, ) Markov chain. is unobserved,

and the agent implements a Bayesian procedure to learn whether she lives in the good state
> . All in all, such a Bayesian procedure is similar to the one in Eq. (7.17). Therefore, it
would be relatively simple to show that all equilibrium prices in this economy are isomorphic
to all equilibrium prices in an economy in which (Z, G) are solution to:
+
( )
dD( ) = (g( ) 20 ) d + 0 dW
( )
dg( ) = (k(
g g( )) 0 v (g( ))) d + v (g( )) dW

is a P 0 -Brownian motion, v(g) = ( g)(g ) 0 , k, g are some positive constants.
where W
Veronesi (1999) also assumed that the riskless asset is innitely elastically supplied, and therefore that the interest rate r is a constant. It is instructive to examine the price implications
of the resulting
# economy. In terms of the representation in Eq. (7.4), this model predicts that
S(D, g) = 0 C(D, g, )d , where
"
r
r
C(D, g, ) = e (D 0 ) + D (g, ) , and D (g, ) e
E [g(u)| g] du, 0.
0
(7.23)
We may now apply Proposition 7.1 to study convexity properties of D. Precisely, function
E [g(u)| g] is a special case of the auxiliary pricing function (7.7) (namely, for 1 and
(g) = g). By Proposition 7.1-b), E [g(u)| g] is convex in g whenever the drift of G in (7.20)
is convex. This condition is automatically guaranteed by > 0. Technically, Proposition 7.1
implies that the conditional expectation of a diusion process inherits the very same second
order properties (concavity, linearity, and convexity) of the drift function.
The economic implications of this result are striking. In this economy prices are convex in
the expected dividend growth. This means that in good times, prices may well rocket to very
high values with relatively small movements in the underlying fundamentals.
The economic interpretation of this convexity property is that risk-aversion correction is nil
during extreme situations (i.e. when the dividend growth rate is at its boundaries), and it is
the highest during relatively more normal situations. More formally, the risk-adjusted drift
of g is
(y) = (g) 0 v (g), and it is convex in g because v is concave in g.
Finally, we examine model (7.21). Also, please notice that this model has been obtained as
a result of a specic learning mechanism. Yet alternative learning mechanisms can lead to a
model having the same structure, but with dierent coecients and v. For example, a model
related to Brennan and Xia (2001) information structure is one in which a single innitely lived
agent observes Z, where Z is solution to:
dD( )
= g( )d + 0 dw1 ( ),
D( )
= {
where G
g ( )} >0 is unobserved, but now it does not evolve on a countable number of
states. Rather, it follows an Ornstein-Uhlenbeck process:
d
g ( ) = k(
g g( ))d + 1 dw1 ( ) + 2 dw2 ( )
where g, 1 and 2 are positive constants. Suppose now that the agent implements a learning
procedure similar as before. If she has a Gaussian prior on g(0) with variance 2 (dened below),
the nonarbitrage price takes the form S(D, g), where (Z, G) are now solution to Eq. (7.6), with
215
c
by
A. Mele
m0 (D, g) = gD, (D) = 0 D, 0 (D, g) = k(

g g), v2 = 0, and v1 v1 ( ) = ( 1 + 10 )2 ,
where is the positive solution to v1 () = 21 + 22 2k.20
Finally, models making expected consumption another observed diusion may have an interest in their own (see for example, Campbell (2003) and Bansal and Yaron (2004)).
Now lets analyze these models. Once again, we may make use of Proposition 7.1. We need
to set the problem in terms of the notation of the canonical pricing problem in Section 7.5. To
simplify the exposition, we suppose that is constant. By the same kind of reasoning leading
to Eq. (7.16), one nds that the price-dividend ratio is independent of D here too, and is given
by function p below,
"
!

e 0 [g(u)r(g(u))]du0 g d ,
(7.24)
p (g) =
E
0
where,
dD( )
( )
= [g( ) 0 ] d + 0 dW
D( )
( )
dg( ) = [ (g( )) + 0 v (g( ))] d + v (g( )) dW
( ) = W
( ) 0 is a P -Brownian motion. Under regularity conditions, monotonicity
and W
and convexity properties are inherited by the inner expectation in Eq. (7.24). Precisely, in the
notation of the canonical pricing problem,
(g) = g + R (g) + 0 and b (g) = 0 (g) + ( 0 ) v (g) ,
where 0 is the physical probability measure. Therefore,
1. The price-dividend ratio is increasing in the dividend growth rate whenever
d
R (g)
dg
< 1.
2. Suppose that the price-dividend ratio is increasing in the dividend growth rate. Then it is
d2
d2
d
convex whenever dg
2 R (g) > 0, and dg 2 [0 (g) + ( 0 ) v (g)] 2 + 2 dg R (g).
For example, if the riskless asset is constant (because for example it is innitely elastically
supplied), then the price-dividend ratio is always increasing and it is convex whenever,
d2
[ (g) + ( 0 ) v (g)] 2.
dg 2 0
The reader can now use these conditions to check predictions made by all models with stochastic
dividend growth presented before.
20 In
their article, Brennan and Xia considered a slightly more general model in which consumption and dividends dier. They
obtain a reduced-form model which is identical to the one in this example. In the calibrated model, Brennan and Xia found that
the variance of the ltered g is higher than the variance of the expected dividend growth in an economy with complete information.
The results on in this example can be obtained through an application of theorem 12.1 in Liptser and Shiryaev (2001) (Vol. II, p.
22). They generalize results in Gennotte (1986) and are a special case of results in Detemple (1986). Both Gennotte and Detemple
did not emphasize the impact of learning on the pricing function.
216
c
by
A. Mele
7.8. Appendix 1
7.8 Appendix 1
7.8.1
Markov pricing kernels
Let
( ) (D( ), y( ), ) = e

0
(D(s),y(s))ds
(D( ), y( )),
(7.25)
for some bounded positive function , and some positive function (D, y) C 2,2 (Z Y). By the
assumed functional form for , and Itos lemma,
R(D, y) = (D, y)
L(D, y)
(D, y)
log (D, y) v1 (D, y) log (D, y)

D
y
2 (D, y) = v2 (D, y) log (D, y)

y
1 (D, y) = 0 D
Example A1 below is an important special case of this setting. Finally, to derive Eq. (7.3) in this
setting, let us dene the (undiscounted) Arrow-Debreu adjusted asset price process as:
w(D, y) (D, y) S(D, y).
By the results in Section 7.4.2, we know that the following price representation holds true:
"

S( )( ) = E
(s)D(s)ds , 0.
Under usual regularity conditions, the previous equation can then be understood as the unique
Feynman-Kac stochastic representation of the solution to the following partial dierential equation
Lw(D, y) + f(D, y) = (D, y)w(D, y), (D, y) Z Y,

d
where f D. Eq. (7.3) then follows by the denition of Lw( ) ds
E [S]s= .
Example A1 (Innite horizon, complete markets economy.) Consider an innite horizon, complete
markets economy in which total consumption Z is solution to Eq. (7.6), with v2 0. Let a (single)
agents program be:
"

"

max E
e u(c( ), x( ))d s.t. V0 = E
( )c( )d , V0 > 0,
0
where > 0, the instantaneous utility u is continuous and thrice continuously dierentiable in its
arguments, and x is solution to
dx( ) = (D( ), g( ), x( ))d + (D( ), g( ), x( ))dW1 ( ).
In equilibrium, C = Z, where C is optimal consumption. In terms of the representation in (7.25), we
have that (D, x) = , and (D( ), x( )) = u1 (D( ), x( ))/ u1 (D(0), x(0)). Consequently, 2 = 0,
u11 (D, x)
u12 (D, x)
m0 (D, g)
(D, g, x)
u1 (D, x)
u1 (D, x)
1
u111 (D, x) 1
u122 (D, x)
u112 (D, x)
(D, g)2
(D, g, x)2
(D, g, x)(D, g)
2
u1 (D, x)
2
u1 (D, x)
u1 (D, x)
R(D, g, x) =
(D, g, x) =
u11 (D, x)
u12 (D, x)
(D, g)
(D, g, x).
u1 (D, x)
u1 (D, x)
217
(7.26)
(7.27)
c
by
A. Mele
7.8. Appendix 1
7.8.2
The maximum principle
Suppose we are given the dierential equation:

dx( )
= ( ), (t, T ),
d
where satises some basic regularity conditions (essentially an integrability condition: see below).
Suppose we know that
x(T ) = 0,
and that
sign (( )) = constant on (t, T ).
We wish to determine the sign of x(t). Under the previous assumptions on x(T ) and the sign of , we
have that:
sign (x(t)) = sign () .
The proof of this basic result can be grasped very simply from Figure 7.11, and it also follows easily
analytically. We obviously have,
" T
" T
0 = x(T ) = x(t) +
( )d x(t) =
( )d .
t
Next, suppose that,

dx( )
= ( ), (t, T ),
d
where
x( ) = f (y( ), ) , (t, T ),
and
dy( )
= D( ), (t, T ).
d
With enough regularity conditions on , f, D, we have that

dx
=
+ L f, (t, T ),
d
where Lf =
f
y
D. Therefore,
+ L f = , (t, T ),
(7.28)
and the previous conclusions hold here as well: if f(y, T ) = 0 y, and sign(( )) = constant on
(t, T ), then,
sign (f (t)) = sign () .
Again, this is so because
f (y(t), t) =
"
( )d .
Such results can be extended in a straightforward manner in the case of stochastic dierential
equations. Consider the more elaborate operator-theoretic format version of (7.28) which typically
emerges in many asset pricing problems with Brownian information:

0=
+ L k u + , (t, T ).
(7.29)
218
c
by
A. Mele
7.8. Appendix 1
x(t)
<0
x(T)
t
>0
x(T)
t
x(t)
FIGURE 6.9. Illustration of the maximum principle for ordinary dierential equations
Let
y( ) e

t
k(u)du
u( ) +
"
u
t
k(s)ds
(u)du.
I claim that if (7.29) holds, then y is a martingale under some regularity conditions. Indeed,

t
k(u)du
du( ) + e t k(u)du ( )d

t k(u)du
t k(u)du
= k( )e
u( ) + e
+ L u( ) d + e t k(u)du ( )d
+ local martingale

t k(u)du
= e
k( )u( ) +
+ L u( ) + ( ) d + local martingale

+ L k u + = 0.
= local martingale - because
dy( ) = k( )e
u( )d + e
k(u)du
Suppose that y is also a martingale. Then
y(t) = u(t) = E [y(T )] = E e
T
t
k(u)du
"
!
u(T ) + E
u
t
k(s)ds

(u)du ,
and starting from this relationship, you can adapt the previous reasoning on deterministic dierential
equations to the stochastic dierential case. The case with jumps is entirely analogous.
7.8.3
Dynamic Stochastic Dominance
We have,
219
c
by
A. Mele
7.8. Appendix 1
Proposition 7.A.1. (Dynamic Stochastic Dominance) Consider two economies A and B with two
fundamental volatilities aA and aB and let i (x) ai (x)i (x) and i (x) (i = A, B) the corresponding
risk-premium and discount rate. If aA > aB , the price cA in economy A is lower than the price price
cB in economy B whenever for all (x, ) R [0, T ],
V (x, ) [A (x) B (x)] cB (x, ) [A (x) B (x)] cB
x (x, ) +

1 2
aA (x) a2B (x) cB
xx (x, ) < 0.
2
(7.30)
If X is the price of a traded asset, A = B . If in addition is constant, c is decreasing (increasing)

in volatility whenever it is concave (convex) in x. This phenomenon is tightly related to the convexity
eect discussed earlier. If X is not a traded risk, two additional eects are activated. The rst one
reects a discounting adjustment, and is apparent through the rst term in the denition of V . The
second eect reects risk-premia adjustments and corresponds to the second term in the denition of
V . Both signs at which these two terms show up in Eq. (7.30) are intuitive.
7.8.4
Proofs

#T

Proof of proposition 7.A.1. Function c(x, T s) E[ exp( s (x(t))dt) (x(T )) x(s) = x] is
solution to the following partial dierential equation:
+
0 = c2 (x, T s) + L c(x, T s) (x)c(x, T s),
(x, s) R [0, T )
(7.31)
c(x, 0) = (x),
x R
where L c(x, u) = 12 a(x)2 cxx (x, u) + b(x)cx (x, u) and subscripts denote partial derivatives. Clearly, cA
and cB are both solutions to the partial dierential equation (7.31), but with dierent coecients. Let
bA (x) b0 (x) A (x). The price dierence c(x, ) cA (x, ) cB (x, ) is solution to the following
partial dierential equation: (x, s) R [0, T ),
1
0 = c2 (x, T s) + B (x)2 cxx (x, T s) + bA (x)cx (x, T s) A (x)c(x, T s) + V (x, T s),
2
with c(x, 0) = 0 for all x R, and V is as in Eq. (7.30) of the proposition. The result follows by the
maximum principle for partial dierential equations.
Proof of proposition 7.1. By dierentiating twice the partial dierential equation (7.31) with
respect to x, I nd that c(1) (x, ) cx (x, ) and c(2) (x, ) cxx (x, ) are solutions to the following
partial dierential equations: (x, s) R++ [0, T ),
1
1
(1)
2 (1)
0 = c2 (x, T s) + a(x)2 c(1)
xx (x, T s) + [b(x) + (a(x) ) ]cx (x, T s)
2
2

(x) b (x) c(1) (x, T s) (x)c(x, T s),
with c(1) (x, 0) = (x) x R; and (x, s) R [0, T ),
1
(2)
2 (2)
0 = c2 (x, T s) + a(x)2 c(2)
xx (x, T s) + [b(x) + (a(x) ) ]cx (x, T s)
2

1
2
(x) 2b (x) (a(x) ) c(2) (x, T s)
2

(1)
2 (x) b (x) c (x, T s) (x)c(x, T s),
with c(2) (x, 0) = (x) x R. By the maximum principle for partial dierential equations, c(1) (x, T
s) > 0 (resp. < 0) (x, s) R[0, T ) whenever (x) > 0 (resp. < 0) and (x) < 0 (resp. > 0) x R.
This completes the proof of part a) of the proposition. The proof of part b) is obtained similarly.
220
c
by
A. Mele
7.8. Appendix 1
Derivation of Eq. (7.19). We have,

dD = gd + dW,
and, by Eq. (7.18),
(D) =
pe
p(D A)
=
p(D A) + (1 p)(D + A)
p + (1 p)e2AD
1 2
where the second equality follows by the Gaussian distribution assumption (x) e 2 x , and straight
forward simplications. By simple computations,
(1 p) e2AD
1 (D)
=
;
(D)
p
(D) = 2A (D)2
(1 p) e2AD
;
p
(D) = 2A (D) [1 2 (D)] .

(7.32)
By construction,
g = (D) A + [1 (D)] (A) = A [2 (D) 1] .
Therefore, by Itos lemma,
1
d = dD + d = dD + A (1 2) d = [g + A (1 2)] d + dW = dW.
2
By using the relations in (7.32) once again,
d = 2A (1 ) dW.
7.8.5
On bond prices convexity
Consider a short-term rate process {r( )} [0,T ] (say), and let u(r0 , T ) be the price of a bond expiring
at time T when the current short-term rate is r0 :

" T

u(r0 , T ) = E exp
r( )d r0 .
0
As pointed out in Section 7.6, a restricted version of Proposition 7.1-b) implies that in all scalar
(diusion) models of the short-term rate, u11 (r0 , T ) < 0 whenever b < 2, where b is the risk-netraulized
drift of r. This specic result was originally obtained in Mele (2003). Both the theory in Mele (2003)
and the proof of Proposition 7.1-b) rely on the Feynman-Kac representation of u11 . Here we provide
a more intuitive derivation under a set of simplifying assumptions.
By Mele (2003) (Eq. (6) p. 685),
-"
.
2 " T 2
" T
=
T
r
r
u11 (r0 , T ) = E
( )d
r( )d
.
2 ( )d exp
0 r0
0 r0
0
Hence u11 (r0 , T ) > 0 whenever

"
2r
( )d <
r02
"
r
( )d
r0
2
To keep the presentation as simple as possible, we assume that r is solution to:

dr( ) = b(r( ))dt + a0 r( )dW ( ),
221
(7.33)
c
by
A. Mele
7.9. Appendix 2
where a0 is a constant. We have,
r
( ) = exp
r0
and
"
1
b (r(u))du a20 + a0 W ( )
2
2r
r
( ) =
( )
2
r0
r0
2 r( )/r02
"

r(u)
b (r(u))
du .
r0
Therefore, if
< 0, then
< 0, and by inequality (11.43), u11 > 0. But this result can
considerably be improved. Precisely, suppose that b < 2 (instead of simply assuming that b < 0).
By the previous equality,
"

2r
r
r(u)
( ) < 2
( )
du ,
r0
r0
r02
0
and consequently,
"
2r
( )d < 2
r02
"
r
( )
r0
"

" T
2
r(u)
r(u)
du d =
du ,
r0
r0
0
which is inequality (11.43).
7.9 Appendix 2
In their original article, Campbell and Cochrane considered a discrete-time model in which consumption is a Gaussian process. The diusion limit of their model is simply Eq. (8.17) given in the main
text. By example A1 (Eq. (7.27)),

1
(D, x) =
0 (D, x) .
(7.34)
s
D
To nd the diusion function of x, notice that x = D(1 s), where s solution to Eq. (7.12). By
Itos lemma, then, = [1 s sl(s)] D0 . Finally, we replace this function into (7.34), and obtain
(s) = 0 [1 + l(s)], as we claimed in the main text. (This result holds approximately in the original
discrete time framework.) Finally, the real interest rate is found by an application of formula (7.26),

1 2
1
R(s) = + g0 0 + (1 )(
s log s) 2 20 [1 + l(s)]2 .
2
2
Campbell
and Cochrane choose l so as
to make the real interest rate constant. They took l(s) =

s), which leaves R = + g0 12 20
S1 1 + 2(s log s) 1, where S = 0 /(1 ) = exp(
1
2 (1 ).
7.10 Appendix 3: Simulation of discrete-time pricing models

The pricing equation is

S = E m S + D ,
uc (D , x )
m=
=
uc (D, x)
Hence, the price-dividend ratio p S/ D satises:

D
1 + p ,
p=E m
D
222

s
D
.
s
D
D
= eg0 +w .
D
c
by
A. Mele
7.10. Appendix 3: Simulation of discrete-time pricing models

This is a functional equation having the form,
&

'
p(s) = E g s , s 1 + p s s ,

g s , s =
1
s
D
.
s
D
A numerical solution can be implemented as follows. Create a grid and dene pj = p (sj ), j = 1, , N,
for some N. We have,

b1
a11 aN1
p1
p1
.. .. ..
.. .. ,
..
. = . + .
.
. .
pN
bN
a1N aNN
pN
bi =
N
j=1
aji ,
aji = gji pji ,
gji = g (sj , si ) ,
pji = Pr ( sj | si ) s,
where s is the integration step; s1 = smin , sN = smax ; smin and smax are the boundaries in the
approximation; and Pr ( sj | si ) is the transition density from state i to state j - in this case, a Gaussian
transition density. Let p = [p1 pN ] , b = [b1 bN ] , and let A be a matrix with elements aji .
The solution is,
p = (I A)1 b.
(7.35)
The model can be simulated in the following manner. Let s and s be the boundaries of the underlying
s s
. Draw states. State s is drawn. Then,
state process. Fix s =
N
1. If min (s s, s s ) = s s, let k be the smallest integer close to
and smax = smin + N s.
2. If min (s s, s s ) = s s , let k be the biggest integer close to
and smin = smax N s.
s s
s .
s s
s .
Let smin = s ks,

Let smax = s + ks,
The previous algorithm avoids interpolations. Importantly, it ensures that during the simulations,
p is computed in correspondence of exactly the state s that is drawn. Precisely, once s is drawn,
1 ) create the corresponding grid s1 = smin , s2 = smin + s, , sN = smax according to the previous
rules; 2 ) compute the solution from Eq. (7.35). In this way, one has p (s ) at hand - the simulated
P/D ratio when state s is drawn.
223
c
by
A. Mele
References
Abel, A. B. (1988): Stock Prices under Time-Varying Dividend Risk: An Exact Solution
in an Innite-Horizon General Equilibrium Model. Journal of Monetary Economics 22,
375-393.
Bajeux-Besnainou, I. and J.-C. Rochet (1996): Dynamic Spanning: Are Options an Appropriate Instrument? Mathematical Finance 6, 1-16.
Bansal, R. and A. Yaron (2004): Risks for the Long Run: A Potential Resolution of Asset
Pricing Puzzles. Journal of Finance 59, 1481-1509.
Barberis, N., M. Huang and T. Santos (2001): Prospect Theory and Asset Prices. Quarterly
Journal of Economics 116, 1-53.
Barsky, R. B. (1989): Why Dont the Prices of Stocks and Bonds Move Together? American
Economic Review 79, 1132-1145.
Barsky, R. B. and J. B. De Long (1990): Bull and Bear Markets in the Twentieth Century.
Journal of Economic History 50, 265-281.
Barsky, R. B. and J. B. De Long (1993): Why Does the Stock Market Fluctuate? Quarterly
Journal of Economics 108, 291-311.
Black, F. and M. Scholes (1973): The Pricing of Options and Corporate Liabilities. Journal
of Political Economy 81, 637-659.
Brennan, M. J. and Y. Xia (2001): Stock Price Volatility and Equity Premium. Journal of
Monetary Economics 47, 249-283.
Campbell, J. Y. (2003): Consumption-Based Asset Pricing. In: Constantinides, G.M., M.
Harris and R. M. Stulz (Editors): Handbook of the Economics of Finance (Volume 1B:
Chapter 13), 803-887.
Campbell, J. Y., and J. H. Cochrane (1999): By Force of Habit: A Consumption-Based
Explanation of Aggregate Stock Market Behavior. Journal of Political Economy 107,
205-251.
David, A. (1997): Fluctuating Condence in Stock Markets: Implications for Returns and
Volatility. Journal of Financial and Quantitative Analysis 32, 427-462.
Detemple, J. B. (1986): Asset Pricing in a Production Economy with Incomplete Information. Journal of Finance 41, 383-391.
El Karoui, N., M. Jeanblanc-Picque and S. E. Shreve (1998): Robustness of the Black and
Scholes Formula. Mathematical Finance 8, 93-126.
Fama, E. F. and K. R. French (1989): Business Conditions and Expected Returns on Stocks
and Bonds. Journal of Financial Economics 25, 23-49.
224
c
by
A. Mele
Gennotte, G. (1986): Optimal Portfolio Choice Under Incomplete Information. Journal of

Finance 41, 733-746.
Huang, C.-F. and Pagès, H. (1992): Optimal Consumption and Portfolio Policies with an
Innite Horizon: Existence and Convergence. Annals of Applied Probability 2, 36-64.
Hajek, B. (1985): Mean Stochastic Comparison of Diusions. Zeitschrift fur Wahrscheinlichkeitstheorie und Verwandte Gebiete 68, 315-329.
Jagannathan, R. (1984): Call Options and the Risk of Underlying Securities. Journal of
Financial Economics 13, 425-434.
Kijima, M. (2002): Monotonicity and Convexity of Option Prices Revisited. Mathematical
Finance 12, 411-426.
Liptser, R. S. and A. N. Shiryaev (2001): Statistics of Random Processes. Berlin, SpringerVerlag. [2001a: Vol. I (General Theory). 2001b: Vol. II (Applications).]
Malkiel, B. (1979): The Capital Formation Problem in the United States. Journal of Finance
34, 291-306.
Mele, A. (2005): Rational Stock Market Fluctuations. WP FMG-LSE.
Pindyck, R. (1984): Risk, Ination and the Stock Market. American Economic Review 74,
335-351.
Poterba, J. and L. Summers (1985): The Persistence of Volatility and Stock Market Fluctuations. American Economic Review 75, 1142-1151.
Rothschild, M. and J. E. Stiglitz (1970): Increasing Risk: I. A Denition. Journal of Economic Theory 2, 225-243.
Veronesi, P. (1999): Stock Market Overreaction to Bad News in Good Times: A Rational
Expectations Equilibrium Model. Review of Financial Studies 12, 975-1007.
Veronesi, P. (2000): How Does Information Quality Aect Stock Returns? Journal of Finance
55, 807-837.
Wang, S. (1993): The Integrability Problem of Asset Prices. Journal of Economic Theory
59, 199-213.
225
8
Tackling the puzzles
8.1 Non-expected utility

The standard intertemporal additively separable utility function confounds intertemporal substitution eects and attitudes towards risk. This fact is problematic. Epstein and Zin (1989,
1991) and Weil (1989) consider a class of recursive, but not necessarily expected utility, preferences. In this section, we present the details of this approach, without insisting on the theoretic
underpinnings which the reader will nd in Epstein and Zin (1989). We provide a basic denition and derivation of this class of preferences, and a heuristic analysis of the resulting asset
pricing properties.
8.1.1
The recursive formulation
Let utility as of time t be vt . We have,

vt = W (ct , vt+1 ) ,
where W is the aggregator function and vt+1 is the certainty-equivalent utility at t+1 dened
as,
h (
vt+1 ) = E [h (vt+1 )] ,
where h is a von Neumann - Morgenstern utility function. That is, the certainty equivalent
depends on some agents risk-attitudes encoded into h. Therefore,

vt = W ct , h1 [E (h (vt+1 ))] .
The analytical example used in the asset pricing literature is,

1/
W (c, v) = c + e v
and h (
v) = v1 ,
for three positive constants , and . In this formulation, risk-attitudes for static wealth
gambles have still the classical CRRA avor. More precisely, we say that is the RRA for
static wealth gambles and (1 )1 is the IES.
c
by
A. Mele
8.1. Non-expected utility

We have,
1
1 1 1
vt+1 = h1 [E (h (vt+1 ))] = h1 E vt+1
= E vt+1
.
The previous parametrization of the aggregator function then implies that,

!1/
1 1
vt = ct + e E(vt+1
)
.
(8.1)
This collapses to the standard intertemporal additively separable case when = 1

RRA = IES1 . Indeed, it is straight forward to show that in this case,
1

n 1 1
vt = E
e ct+n
.
n=0
Let us go back to Eq. (8.1). Function V = v 1 / (1 ) is obviously ordinally equivalent to

v, and satises,
! 1
.
(8.2)
Vt =
ct + e ((1 ) E(Vt+1 )) 1
1
The previous formulation makes even more transparent that these utils collapse to standard
intertemporal additive utils as soon as RRA = IES1 .
8.1.2
Testable restrictions
Let us dene cum-dividend wealth as xt

xt evolves as follows:
m
i=1
(pit + Dit ) it . In the Appendix, we show that
xt+1 = (xt ct )
t (1m + rt+1 ) (xt ct ) (1 + rM,t+1 ) ,
(8.3)
where is the vector of proportions of wealth invested in the m assets, rt+1 is the vector of
it+1 pit
asset returns, with any component i being equal to, rit+1 Pit+1 +D
, and rM,t is the return
Pit
on the market portfolio, dened as,

Pit it+1
rM,t+1 = m
it
,
i=1 rit+1 it ,
Pit it+1
where Pit and Dit are the price and the dividend of asset i at time t.
Let us consider a Markov economy in which the underlying state is some process y. We
consider stationary consumption and investment plans. Accordingly, let the stationary util be
a function V (x, y) when current wealth is x and the state is y. By Eq. (8.2),
)
* 1

1
max c + e [(1 ) E(V (x , y ))]
V (x, y) = max W (c, E (V (x , y )))
.
c,
1 c,
(8.4)
In the Appendix, we show that the rst order conditions for the representative agent lead to
the following Euler equation,
E [m (x, y; x y ) (1 + ri (y ))] = 1,
i = 1, , m,
(8.5)
where the stochastic discount factor m is,

c (x , y )
1

m (x, y; x y ) = e
(1 + rM (y )) ,
c (x, y)
where =
= 1 .
1
1
1
. Note, the stochastic discount factor is more volatile than the usual one, once
227
c
by
A. Mele

8.1.3
Equilibrium risk-premia and interest rates
So the Euler equation is,

-
.

c
1
E e
(1 + rM
) (1 + ri ) = 1.
c
(8.6)
Eq. (8.6) obviously holds for the market portfolio and the risk-free asset. Therefore, by taking
logs in Eq. (8.6) for i = M , and for the risk-free asset, i = 0 (say) yields the two following
conditions:

c
, RM = log (1 + rM
0 = log E exp log
+ RM
),
(8.7)
c
and,

c
Rf = log (1 + r0 ) = log E exp log
.
(8.8)
+ ( 1) RM
c

Next, suppose that consumption growth, log cc , and the market portfolio return, RM , are
jointly normally distributed. In the appendix, we show that the expected excess return on the
market portofolio is given by,
1
E(RM ) Rf + 2RM = RM ,c + (1 ) 2RM

2
(8.9)
where 2RM = var(RM ) and RM ,c = cov(RM , log (c /c)), and the term 12 2RM in the left hand
side is a Jensens inequality term. Note, Eq. (8.9) is a mixture of the Consumption CAPM (for
the part RM ,c ) and the CAPM (for the part (1 ) 2RM ).
The risk-free rate is given by,

1
1 2
c
1
(1 ) 2RM
,
(8.10)
Rf = + E log
c
2
2 2 c
where 2c = var(log (c /c)).
Eqs. (8.9) and (8.10) can be elaborated further. In equilibrium, the asset price and, hence,
the return, is certainly related to consumption volatility. Precisely, let us assume that,
2RM = 2c + 2
RM ,c = 2c ,
(8.11)
where 2 is a positive constant that may arise when the asset return is driven by some additional
state variable. (This is the case, for example, in the Bansal and Yaron (2004) model described
below.) Under the assumption that the asset return volatility is as in Eq. (8.11), the equity
premium in Eq. (8.9) is,
1
E(RM ) Rf + 2RM = 2c + (1 ) 2 = 2c +
2
1
2 .
As we can see, disentangling risk-aversion from intertemporal substitution does not suce per
se to resolve the equity premium puzzle. To increase theequity premium, it is important that
2 > 0, i.e. that some additional state variables drives the variation of the asset return.
228
c
by
A. Mele
On the other hand, in this framework, the volatility of these state variables can only aect the
asset return if risk-aversion is distinct from the inverse of the intertemporal rate substitution.
In particular, suppose that 2 does not depend on and . If > 1, then, the equity premium
increases with 2 whenever > 1 .
Next, let us derive the risk-free rate. Assume that E [log (c /c)] = g0 12 2c , where g0 is the
expected consumption growth, a constant. Furthermore, use the assumptions in Eq. (8.11) to
obtain that the risk-free rate in Eq. (8.10) is,

1
1
1
1
1 2
2
R f = + g0 1 +
c
.
2 1 1
As we can see, we may increase the level of relative risk-aversion, , without substantially
aecting the level of the risk-free rate, Rf . This is because the eects of on Rf are of a
second-order importance (they multiply variances, which are orders of magnitude less than the
expected consumption growth, g0 ).
8.1.4
Campbell-Shiller approximation
Consider the denition of the return on the market portfolio,

zt+1

Pt+1 + Ct+1
+1
e
RM,t+1 = log
= log
+ gt+1 f (zt+1 , zt ) + gt+1 ,
Pt
ezt
where Pt is the value of the market portfolio, gt+1 = log CCt+1
is the growth of the aggregate
t
Pt
dividend, and zt = log Ct is the log of the aggregate price-dividend ratio. A rst-order linear
approximation of f (zt+1 , zt ) around the average level of z leaves,

RM,t+1 0 + 1 zt+1 zt + gt+1 ,
(8.12)
where 0 = log e e+1

+ ezz+1 , 1 = eze+1 and z is the average level of the log price-dividend
z
ratio. Typically, 1 0.997 for US data.

Cambell and Shiller (1988) derive Eq. (8.12) and show how useful it is to address a number of
questions. We now use Eq. (8.12) to illustrate how non-expected utility may be used to address
the equity premium puzzle.
8.1.5
Risks for the long-run
Bansal and Yaron (2004) argue that persistence in the expected consumption growth may
explain the equity premium puzzle. To understand the potential for this, let us assume that
consumption growth is solution to,

Ct+1
1
= g0 2c + xt + t+1 , t+1 N 0, 2c ,
(8.13)
gt+1 = log
Ct
2
where xt is a small persistent component of consumption growth, solution to,

xt+1 = xt + t+1 , t+1 N 0, 2x .
(8.14)
To nd an approximate solution to the log of the price-dividend ratio, replace the CampbellShiller approximation in Eq. (8.12) into the Euler equation (8.7) for the market portfolio,

Ct+1
0 = log E exp log

+ (0 + 1 zt+1 zt + gt+1 ) .
(8.15)
Ct
229
8.2. Catching up with the Joneses in a heterogeneous agents economy
c
by
A. Mele
Let us conjecture that the log of the price-dividend ratio takes the simple form, zt = a0 + a1 xt ,
where a0 and a1 are two coecients to be determined. Substituting this guess into Eq. (8.15),
and identifying terms, one nds that there exixts a constant a0 such that,
zt = a0 +
1 1
xt
(8.16)
The appendix provides details on these calculations.

Next, use RM,t+1 0 + 1 zt+1 zt + gt+1 (or alternatively, the stochastic discount factor)
to compute 2 , volatility, risk-premium, etc. [In progress]
8.2 Catching up with the Joneses in a heterogeneous agents economy

Chan and Kogan (2002) study an economy with heterogeneous agents and catching up with the
Joneses preferences. In this economy, there is a continuum of agents indexed by a parameter
[1, ) of the instantaneous utility function,
c 1
u (c, x) =
where c is consumption, and x is the standard living of others, to be dened below.

The total endowment in the economy, D, follows a geometric Brownian motion,
dD ( )
= g0 d + 0 dW ( ) .
D ( )
(8.17)
By assumption, then, the standard of living of others, x ( ), is a weighted geometric average of

the past realizations of the aggregate consumption D, viz
"
log x ( ) = log x (0) e

+
e( s) log D (s) ds, with > 0.
0
Therefore, x ( ) satises,
dx ( ) = s ( ) x ( ) d ,
where s ( ) log (D ( )/ x ( )) .
(8.18)
By Eqs. (8.17) and (8.18), s ( ) is solution to,

1 2
ds ( ) = g0 0 s ( ) d + 0 dW ( ) .
2
In this economy with complete markets, the equilibrium price process is the same as the price
process in an economy with a representative agent with the following utility function,
"
"
u (D, x) max
u (c , x) f () d s.t.
c d = D,
[P1]
c
where f()1 is the marginal utility of income of the agent . (See Chapter 2 in Part I, for the
theoretical foundations of this program.)
230
c
by
A. Mele
8.3. Incomplete markets
In the appendix, we show that the solution to the static program [P1] leads to the following
expression for the utility function u (D, x),
"
1
1
1
u(D, x) =
f () V (s) d,
1
1
where V is a Lagrange multiplier, which satises,
"
1
1
s
e =
f () V (s) d.
1
The appendix also shows that the unit risk-premium predicted by this model is,
(s) = 0 #
1
exp(s)
1
f
() V (s) d
(8.19)
This economy collapses to an otherwise identical homogeneous economy if the social weighting
function f () = ( 0 ), the Diracs mass at 0 . In this case, (s) = 0 0 , a constant.
A crucial assumption in this model is that the standard of living X is a process with bounded
variation (see Eq. (8.18)). By this assumption, the standard living of others is not a risk which
agents require to be compensated for. The unit risk-premium in Eq. (8.19) is driven by s through
nonlinearities induced by agents heterogeneity. By calibrating their model to US data, Chan
and Kogan nd that the risk-premium, (s), is decreasing and convex in s.1 The mechanism at
the heart of this result is an endogenous wealth redistribution in the economy. Clearly, the less
risk-averse individuals put a higher proportion of their wealth in the risky assets, compared to
the more risk-avers agents. In the poor states of the world, stock prices decrease, the wealth
of the less risk-averse lowers more than that of the more risk-averse agents, which reduces the
fraction of wealth held by the less risk-averse individuals in the whole economy. Thus, in bad
times, the contribution of these less risk-averse individuals to aggregate risk-aversion decreases
and, hence, the aggregate risk-aversion increases in the economy.
8.3 Incomplete markets

8.4 Limited stock market participation
Basak and Cuoco (1998) consider a model with two agents. One of these agents does not
invest in the stock-market, and has logarithmic instantaneous utility, un (c) = log c. From his
perspective, markets are incomplete. The second agent, instead, invests in the stock market,
and has an instantaneous utility equal to up (c) = (c1 1)/ (1 ). Both agents are innitely
lived.
Clearly, in this economy, the competitive equilibrium is Pareto inecient. Yet, Basak and
Cuoco show how aggregation obtains in this economy. Let ci ( ) be the general equilibrium
allocation of the agent i, i = p, n. The rst order conditions of the two agents are,
up (
cp ( )) = wp e ( ) ;
cn ( )1 = wn e
R(s)ds
(8.20)
1 Their numerical results also revealed that in their model, the log of the price-dividend ratio is increasing and concave in s.
Finally, their lemma 5 (p. 1281) establishes that in a homogeneous economy, the price-dividend ratio is increasing and convex in s.
231
c
by
A. Mele
8.4. Limited stock market participation
where wp , wn are two constants, and is the usual pricing kernel process, solution to,
d ( )
= R ( ) dt ( ) dW ( ) .
( )
(8.21)
Let
u (D, x) max [up (cp ) + x un (cn )] ,
cp +cn =D
where
up (
cp )
cp )
cn
x
= up (
un (
cn )
(8.22)
is a stochastic social weight. By the denition of , x ( ) is solution to,

dx ( ) = x ( ) ( ) dW ( ) ,
(8.23)
where is the unit risk-premium, which as shown in the appendix equals,

(s) = 0 s1 ,
where s ( ) cp ( )/ D ( ) .
Then, the equilibrium price system in this

tive agent with utility u (D, x). Intuitively,
construction,
up (cp ( ))
=
un (cn ( ))
economy is supported by a ctitious representathe representative agent allocations satisfy, by

up (
cp ( ))
= x ( ) ,
un (
cn ( ))
where starred allocations are the representative agents allocations. In other words, the trick
underlying this approach is to nd a stochastic social weight process x ( ) such that the rst
order conditions of the representative agent leads to the market allocations. This is shown more
rigorously in the Appendix.
Guvenen (2005) makes an interesting extension of the Basak and Cuoco model. He consider
two agents in which only the rich invests in the stock-market, and is such that ISErich >
IESpoor . He shows that for the rich, a low IES is needed to match the equity premium. However,
US data show that the rich have a high IES, which can not do the equity premium. (Guvenen
considers an extension of the model in which we can disentangle IES and CRRA for the rich.)
232
c
by
A. Mele
8.5. Appendix on non-expected utility
8.5 Appendix on non-expected utility

8.5.1
Detailed derivation of optimality conditions and selected relations
Derivation of Eq. (8.3). We have,

xt+1 = (pit+1 + Dit+1 ) it+1

= (pit+1 + Dit+1 pit ) it+1 + pit it+1

pit+1 + Dit+1 pit pit it+1

= 1+
pit it+1
pit
pit it+1

= (1 + rit+1 it ) (xt ct )

where the last line follows by the standard budget constraint ct + pit it+1 = xt , the denition of
rit+1 and the denition of it given in the main text.
Optimality. Consider Eq. (8.4). The rst order condition for c yields,

= W2 c, E V x , y
E V1 x , y 1 + rM y
,
W1 c, E V x , y
(A1)
where subscripts denote partial derivatives. Thus, optimal consumption is some function c (x, y). Hence,

x = (x c (x, y)) 1 + rM y
We have,

V (x, y) = W c (x, y) , E V x , y
.
By dierentiating the value function with respect to x,

V1 (x, y) = W1 c (x, y) , E V x , y
c1 (x, y)

+W2 c (x, y) , E V x , y
E V1 x , y 1 + rM y
(1 c1 (x, y)) ,
where subscripts denote partial derivatives. By replacing Eq. (A1) into the previous equation we get
the Envelope Equation for this dynamic programming problem,

V1 (x, y) = W1 c (x, y) , E V x , y
.
(A2)
By replacing Eq. (A2) into Eq. (A1), and rearranging terms,

W2 (c (x, y) , (x, y))
E
W1 c x , y , x , y
1 + rM y
= 1,
W1 (c (x, y) , (x, y))

(x, y) E V x , y .
Below, we show that by a similar argument the same Euler equation applies to any asset i,

W2 (c (x, y) , (x, y))
E
W1 c x , y , x , y
1 + ri y
= 1, i = 1, , m.
W1 (c (x, y) , (x, y))
Derivation of Eq. (A3). We have,

V (x, y) = max W c, E V x , y
= max
W x pi i , E V x , y
;
c,
The set of rst order conditions is,

i : 0 = W1 () pi + W2 () E V1 x , y pi + Di ,
233
x =
(A3)

pi + Di i .
i = 1, , m.
c
by
A. Mele
Optimal consumption is c (x, y). Let (x, y) E (V (x , y )), as in the main text. By replacing Eq.
(A2) into the previous equation,

pi + Di
W2 (c (x, y) , (x, y))
E
W1 c x , y , x , y
= 1, i = 1, , m.

W1 (c (x, y) , (x, y))
pi
Derivation of Eq. (8.5). We need to compute explicitly the stochastic discount factor in Eq.
(A3),

W2 (c (x, y) , (x, y))

W1 c x , y , x , y .
m x, y; x y =
W1 (c (x, y) , (x, y))
We have,
W (c, ) =
From this, it follows that,
W1 (c, ) =
W2 (c, ) =
1
c + e ((1 ) ) 1
1
c +e
c +e
((1 ) )
((1 ) )
and,
! 1 1
! 1 1
c1
e ((1 ) ) 1 1 ,
1
1

! 1

1
c1 = W c , 1 (1 ) 1 c1
W1 c , = c + e (1 ) 1
where (x , y ). Therefore,

W2 (c, )

W1 c , = e
m x, y; x y =
W1 (c, )
W (c , )
1
1
(A4)
1
c
.
c
Along any optimal consumption path, V (x, y) = W (c (x, y) , (x, y)). Therefore,

1 1

1

c

E (V (x , y ))
m x, y; x y = e
.
V (x , y )
c
(A5)
(x ,y ))
We are left with evaluating the term E(V
is that v (x, y) = b (y)1/(1) x,
V (x ,y ) . The conjecture to make

for some function b. From this, it follows that V (x, y) = b (y) x1 (1 ). We have,
V1 (x, y)
1
1
1
1

= W1 c (x, y) , E V x , y
= W (c, ) 1 (1 ) 1 c1 = V (x, y) 1 (1 ) 1 c1 .
where the rst equality follows by Eq. (A2), the second equality follows by Eq. (A4), and the last
equality follows by optimality. By making use of the conjecture on V , and rearraning terms,
c (x, y) = a (y) x,

Hence, V (x , y ) = b (y ) x1 (1 ), where
and
a (y) b (y) (1)(1) .

x = (1 a (y)) x 1 + rM y ,
(x , y ))
E (V
V (x , y )
E (y ) (1 + rM (y ))1
(y ) (1 + rM (y ))1
234
(A6)
(A7)
!
(A8)
c
by
A. Mele
Along any optimal path, V (x, y) = W (c (x, y) , E (V (x , y ))). By plugging in W (from Eq. 8.4)) and
the conjecture for V ,
Moreover,

1 ! 1
E y 1 + rM y
= e
a (y)
1 a (y)
(1)(1)

1

! (1)(1)
y 1 + rM y
= a y 1 + rM y 1
.
(A9)
(A10)
By plugging Eqs. (A9)-(A10) into Eq. (A8),

E (V (x , y ))
= e
V (x , y )
1
= e
1
= e
. (1)(1)
a (y)
(1 a (y)) a (y ) (1 + rM (y ))
-
. (1)(1)
c 1
x
c
(1 a (y)) x (1 + rM (y )) 1
-
. (1)(1)
c 1
1
1
c
(1 + r (y )) 1
M
where the rst equality follows by Eq. (A6), and the second equality follows by Eq. (A7). The result
follows by replacing this into Eq. (A5).
Proof of Eqs. (8.9) and (8.10). By using the standard property that log E(ey) = E(
y )+ 12 var (
y),
for y normally distributed, in Eq. (8.7), we obtain,

c
0 = log E exp log
+ RM
c
-
.

c
1
2 2
2
2 2
= E log
+ E(RM ) +
c + RM 2 RM ,c .
(A11)
c
2
We do the same in Eq. (8.8), and obtain,

-
.

c
1
2 2
( 1)
2 2
Rf = + E log
( 1) E(RM )
c + ( 1) RM 2
RM ,c . (A12)
c
2
By replacing Eq. (A12) into Eq. (A11), we obtain Eq. (8.9) in the main text.
To obtain the risk-free rate Rf in Eq. (8.10), we replace the expression for E(RM ) in Eq. (8.9) into
Eq. (A12).
8.5.2
Details for the risks for the lung-run
Proof of Eq. (8.16). By substituting the guess zt = a0 + a1 xt into Eq. (8.15),

Ct+1
0 = (0 (1 1 ) a0 ) + log Et exp log
+ 1 a1 xt+1 a1 xt + gt+1
Ct

1
1 2
1
= (0 (1 1 ) a0 ) + 1
g0 c
+ log Et exp 1
t + 1 a1 t

1
+ (1 1) a1 + 1
xt
const1 + const2 xt ,
235
c
by
A. Mele
where the second equality follows by Eqs. (8.13) and (8.14). Note, then, that this equality can only
hold if the two constants, const1 and const2 are both zero. Imposing const2 = 0 yields,
a1 =
1 1
as in Eq. (8.16) in the main text. Imposing const1 = 0, and using the solution for a1 , yields the solution
for the constant a0 .
8.5.3
Continuous time
Due and Epstein (1992a,b) extend the framework on non-expected utility to continuous time. Heuristically, the continuation utility is the continuous time limit of,

1/
1
1
t
vt = ct t + e
E(vt+t )
.
Continuation utility vt solves the following stochastic dierential equation,
!

dvt = f (ct , vt ) 12 A (vt ) vt 2 dt + vt dBt
vT = 0
Here (f, A) is the aggregator. A is a variance multiplier - it places a penalty proportional to utility
volatility vt 2 . (f, A) somehow corresponds to (W, v) in the discrete time case.
The solution to the previous stochastic dierential utility is,
+" T
,
1
2
vt = E
f (cs , vs ) + A (vs ) vs
.
2
t
The standard additive utility case is obtained by setting,
f (c, v) = u (c) v
236
and A = 0.
c
by
A. Mele
8.6. Appendix on economies with heterogenous agents
8.6 Appendix on economies with heterogenous agents

When every agent faces a system of complete markets, the equilibrium can be computed along the
lines of Huang (1987), who generalizes the classical approach described in Chapter 2 of Part I of these
Lectures. To keep the presentation as close as possible to some of the models of this chapter, we rst
consider the case of a continuum of agents indexed by the instantaneous utility function ua (c, x),
where c is consumption, a is some parameter belonging to some set A, and x is some variable. (For
example, x is the standard of living of others in the Chan and Kogan (2002) model.)
In the context of this chapter, the equilibrium allocation is Pareto ecient if every agent a A
faces a system of complete markets. But by the second welfare theorem, we know that for each Pareto
allocation (ca )aA , there exists a social weighting function f such that the Pareto allocation can be
implemented by means of the following program,
"
"
u (D, x) = max
ua (ca , x) f (a) da,
s.t.
ca da = D,
[Soc-Pl]
ca
aA
aA
where D is the aggregate endowment in the economy. Then, the equilibrium price system can be
computed as the Arrow-Debreu state price density in an economy with a single agent endowed with
the aggregate endowment D, instantaneous utility function u (c, x), and where for a A, the social
weighting function f (a) equals the reciprocal of the marginal utility of income of the agent a.
The practical merit of this approach is that while the marginal utility of income is unobservable, the
thusly constructed Arrow-Debreu state price density depends on the innite dimensional parameter,
f, which can be calibrated to reproduce the main quantitative features of consumption and asset price
data.
We now apply this approach to indicate how to derive the Chan and Kogan (2002) equilibrium
conditions.
catching up with the Joneses (Chan and Kogan (2002)). In this model, markets are
complete, and we have that A = [1, ] and u (c , x) = ( c / x)1 / (1 ). The static optimization
problem for the social planner in [Soc-Pl] can be written as,
u (D, x) = max
c
"
( c / x)1
f () d,
1
s.t.
"
( c / x) d = D/ x.
(A13)
The rst order conditions for this problem lead to,

( c / x) f () = V ( D/ x) ,
(A14)
where V is a Lagrange multiplier, a function of the aggregate endowment D, normalized by x. It is

determined by the equation,
"
V ( D/ x) f () d = D/ x,
which is obtained by replacing Eq. (A14) into the budget constraint of the social planner.
The general equilibrium allocations and prices can be obtained by setting f () equal to the marginal
utility of income for agent . Then, the expression for the unit risk-premium in Eq. (8.19) follows by,
(s) =
2 u (D, x)
D2

u (D, x)
0 D,
D
and lenghty computations, after setting D/ x = es . The short-term rate can be computed by calculating
the expectation of the pricing kernel in this ctitious representative agent economy.
237
c
by
A. Mele
It is instructive to compare the rst order conditions of the social planner in Eq. (A14) with those
in the decentralized economy. Since markets are complete, we have that the rst order conditions in
the decentralized economy satisfy:
et ( c (t)/ x (t)) = () (t) x (t) ,
(A15)
where () is the marginal utility of income for the agent , and (t) is the usual pricing kernel.
By aggregating the market equilibrium allocations in Eq. (A15),
"
"
! 1
1
c (t) d = x (t)
() d.
D (t) =
et (t) x (t)
1
By aggregating the social weighted allocations, with f = 1 ,

"
1
1
V ( D (t)/ x (t)) () d.
D (t) = x (t)
1
Hence, it must be that,
x (t)1 V ( D (t)/ x (t)) = et (t) .
(A16)
1 ,
That is, if f =
then, Eq. (A16) holds. The convers to this result is easy to obtain. Eliminating
( c / x) from Eq. (A14) and Eq. (A15) leaves,

et x (t) (t)
1
=
,
V ( D (t)/ x (t))
f () ()
(t, ) [0, ) [1, ).
Hence if Eq. (A16) holds, then, f = 1 .

To summarize, the equilibrium allocations and prices can be centralized through the social planner
program in (A13), with f = 1 .
Restricted stock market participation (Basak and Cuoco (1998)). We rst show that
Eq. (8.23) holds true. Indeed, by the denition of the stochastic social weight in Eq. (8.22), we have
that

wp
x ( ) = up (
cp ( ))
cn ( ) =
( ) e 0 R(s)ds
wn
where the second line follows by the rst order conditions in Eq. (8.20). Eq. (8.23) follows by the
previous expression for x and the dynamics for the pricing kernel in Eq. (8.21).
By Chapter 7 (Appendix 1), the unit risk premium satises,
(D, x) =
u12 (D, x)x

u11 (D, x)
0 D +
(D, x).
u1 (D, x)
u1 (D, x)
This is:
u1 (D, x)u11 (D, x)
0 D
u1 (D, x) u12 (D, x)x u1 (D, x)

u (
ca )
= a
0 D
u1 (D, x)
u (
ca )
ca
= a
0 s1 .
ua (
ca )
(D, x) =
where the second line follows by Basak and Cuoco (identity (33), p. 331) and the third line follows by
the denition of u(D, x) and s. The Sharpe ratio reported in the main text follows by the denition
of ua . The interest rate is also found through Chapter 7 (Appendix 1). We have,
R(s) = +
1 ( + 1)20
g0
.
( 1)s 2 s( ( 1)s)
238
c
by
A. Mele
Finally, by applying Itos lemma to s = cDa , and using the optimality conditions for agent a, we nd
that drift and diusion functions of s are given by:

(1 )(1 s)
1 ( + 1)20
1 ( + 1)20
(s) = g0
s
+
+ 0 (s 1),
+ (1 )s
2 + (1 ) s 2
s
and (s) = 0 (1 s).
239
c
by
A. Mele
References
Bansal, R. and A. Yaron (2004): Risks for the Long Run: A Potential Resolution of Asset
Pricing Puzzles. Journal of Finance 59, 1481-1509.
Basak, S. and D. Cuoco (1998): An Equilibrium Model with Restricted Stock Market Participation. Review of Financial Studies 11, 309-341.
Campbell, J. and R. Shiller (1988): The Dividend-Price Ratio and Expectations of Future
Dividends and Discount Factors. Review of Financial Studies 1, 195228.
Chan, Y.L. and L. Kogan (2002): Catching Up with the Joneses: Heterogeneous Preferences
and the Dynamics of Asset Prices. Journal of Political Economy 110, 1255-1285.
Due, D. and L.G. Epstein (1992a): Asset Pricing with Stochastic Dierential Utility. Review of Financial Studies 5, 411-436.
Due, D. and L.G. Epstein (with C. Skiadas) (1992b): Stochastic Dierential Utility. Econometrica 60, 353-394.
Epstein, L.G. and S.E. Zin (1989): Substitution, Risk-Aversion and the Temporal Behavior of
Consumption and Asset Returns: A Theoretical Framework. Econometrica 57, 937-969.
Epstein, L.G. and S.E. Zin (1991): Substitution, Risk-Aversion and the Temporal Behavior of
Consumption and Asset Returns: An Empirical Analysis. Journal of Political Economy
99, 263-286.
Guvenen, F. (2005): A Parsimonious Macroeconomic Model for Asset Pricing: Habit formation or Cross-Sectional Heterogeneity. Working paper, University of Rochester.
Huang, C.-f. (1987): An Intertemporal General Equilibrium Asset Pricing Model: the Case
of Diusion Information. Econometrica 55, 117-142.
Weil, Ph. (1989): The Equity Premium Puzzle and the Risk-Free Rate Puzzle. Journal of
Monetary Economics 24, 401-421.
240
9
Information and other market frictions
9.1 Introduction
The assumption agents have imperfect information about the fundamentals of the economy was
rst used by Phelps (1970) and Lucas (1972), to explain the relation between monetary policy
and the business cycle. This information-based approach to the business cycle, summarized
in Lucas (1981), was, in fact, abandoned in favour of the real business cycle theory, reviewed
in Chapter 3, partly because imperfect information can not be considered as the sole engine
of macroeconomic uctuations. Instead, it is widely acknowledged that the merit of Lucas
approach was the introduction of a systematic way of thinking about uctuations, in a context
of rational expectations. Moreover, his information approach has inspired work in nancial
economics, where imperfect information is likely to play a quite fundamental role. In Section
9.2, we provide a succinct account of the Lucas framework, and solve a model relying on a
simplied version of Lucas (1973). We solve this model, following the perspective we think a
nance theorist would typically have. It is quite useful to present this model, as this is very
simple and at the same time, contributes to give us a big picture of where imperfect information
can lead us, in general. Section 9.2 through 9.7 review the many models in nancial economics
that have been used to explain the price formation mechanism in contexts with imperfect
information, be it asymmetric or dierential, as we shall make precise below.
Sections 9.7 and 9.9 conclude this chapter, and present additional market frictions that are
potentially apt to explain certain features in the asset price formation process.
9.2 Prelude: imperfect information in macroeconomics

There are n islands, where n goods are produced. Let yis denote log-production supplied in the
i-th island. (All prices and quantities are in logs, in this section.) It is assumed that this supply
is set so as to equal the expected wedge of the price in the island, pi , over the average price in
the economy, p,
n
1
s
yi = E ( pi p| pi ) , where p =
pj .
n j=1
9.2. Prelude: imperfect information in macroeconomics
c
by
A. Mele
The previous equation can be easily derived, once we assume p is common knowledge, as for
example in the model of monopolistic competion of Blanchard and Kiyotaki (1987). If, instead,
p is not common knowledge, it is more problematic to derive the exact functional form assumed
for yis , although this describes a quite plausible decision mechanism.
Information is disseminated dierentially, not asymmetrically, in that producers in the i-th
island do not know the price in the remaining islands, and guess economic developments in
the other islands with the same precision. We assume and, later, verify, that all variables,
exogeneous and endogeneous, are normally distributed. Under this presumption, we shall show,
the price index p gathers all the available information in the economy eciently, i.e. it is a
sucient statistics for all that information.
We have, by the Projection theorem,
yis = E ( pi p| pi ) (pi E (p)) ,
where we have used the fact that information is symmetrically disseminated and, then, (i)
the expectation E (pi ) = E (pj ) = E (p) for every i and j, and (ii) both the numerator and
i p,pi )
denominator of the ratio, cov(p
, are the same across all islands. This coecient will
var(pi )
be determined below, as a result of the equilibrium.
Aggregating across all islands, yields the celebrated Lucas supply equation:
n
1 s
y = (p E (p)) .
y
n j=1 j
s
(9.1)
Next, assume the demand for the good produced in the i-th island is given by:

yid = m p + ui (pi p) , where ui N 0, 2u
where money is
m = E (m) + ,

where N 0, 2 .
n
(9.2)
Finally, we assume that E (ui ) = 0, and that ui are a sectoral shocks, in that: j=1 uj = 0.
The functional form assumed for the demand function, yid , can be easily derived, assuming the
goods in the islands are imperfect substitutes, as for example in Blanchard and Kiyotaki (1987).
In this context, the equilibrium price in the islands plays two roles. A rst, standard role, is
to clear the market in each island, being such that yis = yid , or:
(pi E (p)) = m p + ui (pi p) , for all i.
(9.3)
Its second role is to convey information about the two shocks, the macroeconomic, monetary
shock, , and the real shocks in all the islands, uj , j = 1, , n. Let us assume, then, that the
only real shock that matters for the price in the i-th island is ui . Below, we shall verify this
conjecture holds, in equilibrium. Then, the price is a function pi = P (, ui ), which we conjecture
to be ane, in and ui , viz
P (, ui ) = a + b + cui ,
(9.4)
where the coecients a, b and c have to be determined, in equilibrium. Under these conditions,
the average price is a function p = P (), equal to:
P () = a + b.
242
(9.5)
9.2. Prelude: imperfect information in macroeconomics
c
by
A. Mele
Let us replace Eqs. (9.4), (9.5) and (9.2) into Eq. (9.3). By rearranging terms, we obtain:
0 = (b + b 1) + (c + c 1) ui + a E (m) .
This equation has to hold for all and ui . Therefore,
a = E (m) ,
and the coecients for and ui must both equal zero, leading to the following expressions for
b and c:
1
1
, c=
.
(9.6)
b=
1+
+
We are left with determining , which given Eqs. (9.4)-(9.5), and Eq. (9.6), is easily shown to
equal:
2u
=
(9.7)

2 .
+
2
2
u + 1+
The positive xed point to this equation, which is easily shown to exist, delivers , which can
then be replaced back into Eqs. (9.6), to yield the solutions for b and c, which are both positive.
We can now gure out the implications of this equilibrium. By replacing Eqs. (9.4)-(9.5) into
the Lucas supply equation (9.1), leaves:
y s = b.
This is Lucas celebrated neutrality result. Anticipated monetary policy, E (m), does not aect
the equilibrium outcome, y s . Instead, it is the monetary shock that aects y s . Agents in any
one island do not observe the price in the remaining islands and, hence, the aggregate price
level, p. Therefore, they are unable to tell whether an increase in the price of the good they
produce, pi , is due to a real shock, ui , or to a monetary shock, . In other words, they can
not disentangle a monetary shock from a real shock. If the agents were informed about the
real shocks in the other islands, they would of course infer , and a monetary shock would not
exert any eect on the equilibrium production. Formally, in equilibrium, the price dierence,
pi p = cui , which does not depend on , a standard dichotomy prediction reminiscent of
classical theory. But pi p is not observed, as p is not observed. Instead, the producers in the i-th
island can only guess pi E ( p| pi ) = b+cui , which co-varies positively with the observed price,
pi , cov (pi p, pi ) = c2 2u . This covariance is zero precisely when we remove the assumption of
imperfect knowledge about the real shocks, so that 2u = 0, in which case = 0. By contrast,
with imperfect knowledge, producers act so as to compensate for their partial lack of knowledge,
and produce to the maximum extent they can justify, on the basis of the positive statistical
co-movements, cov (pi p, pi ) > 0. Note, if E (m) = m1 , i.e. money supply in the previous
period, then from Eq. (9.5), the ination rate, p p1 = b + (1 b) 1 . Therefore, output and
ination are positively correlated, and generate a Phillips curve, which policy makers can not
exploit anyway, as anticipated monetary policy, E (m), is rationally factored out, and does
not aect output. This is the essence of the Lucas critique (Lucas, 1977).
In the next sections, we present a number of models that work due to a similar mechanism.
Why should we ever purchase an asset from any one else, who is insisting in selling it to the
market? Trading seems to be a dicult phenomenon to explain, in a world with imperfect
information. Yet trading does occur, if imperfect information has the same nature as that of
243
c
by
A. Mele
9.3. Grossman-Stiglitz paradox
the Phelps-Lucas model. Agents might well be imperfectly informed about the nature of, say,
unusually high market orders. For example, huge sell orders might arrive to the market, either
because the asset is a lemon or because the agents selling it are hit by a liquidity shock. In
the models of this section, an equilibrium with rational expectation exists, precisely because of
this noiseliquidity, in this example. There is a chance the sell order arrives to the market,
simply because the agents selling it are hit by a liquidy shock. Imperfectly informed agents,
therefore, might be willing to buy, if it is in their interest to do so.
9.3 Grossman-Stiglitz paradox

9.4 Noisy rational expectations equilibrium
9.4.1
Differential information
Hellwig (1980). Diamond and Verrecchia (1981).

9.4.2
Asymmetric information
Grossman and Stiglitz (1980).

9.4.3
Information acquisition
9.5 Strategic trading

Kyle (1985). Foster and Viswanathan (1996).
9.6 Dealers markets

Glosten and Milgrom (1985).
9.7 Noise traders

DeLong, Shleifer, Summers and Waldman (1990).
9.8 Demand-based derivative prices

9.8.1
Options
Garleanu, Pedersen and Poteshman (2007).

9.8.2
Preferred habitat and the yield curve
Vayanos and Vila (2007), Greenwood and Vayanos (2008).
9.9 Over-the-counter markets

Due, Garleanu and Pedersen (2005, 2007).
244
c
by
A. Mele
9.9. Over-the-counter markets
References
Blanchard, O. and N. Kiyotaki (1987): Monopolistic Competition and the Eects of Aggregate
Demand. American Economic Review 77, 647-666.
Lucas, R. E., Jr. (1972): Expectations and the Neutrality of Money. Journal of Economic
Theory 4, 103-124.
Lucas, R. E., Jr. (1973): Some International Evidence on Output-Ination Tradeos. American Economic Review 63, 326-334.
Lucas, R. E., Jr. (1977): Econometric Policy Evaluation: A Critique. Carnegie-Rochester
Conference Series on Public Policy 1: 19-46.
Lucas, R. E., Jr. (1981): Studies in Business-Cycle Theory. Boston, MIT Press.
Phelps, E. S. (1970): Introduction. In: Phelps, E. S. (Editor): Microeconomic Foundations
of Employment and Inflation Theory, New York: W. W. Norton.
245
Part III
Applied asset pricing theory
246
10
Options and volatility
10.1 Introduction
This chapter is under construction. I shall include material on futures, American options, exotic
options, and evaluation of contingent claims through trees. I shall illustrate more systematically the general ideas underlying implied trees, and cover details on how to deal with market
imperfections.
10.2 General properties of options

A European call (put) option is a contract by which the buyer has the right, but not the
obligation, to buy (sell ) a given asset at some price, called the strike, or exercise price, at some
future date. Let C and p be the prices of the call and the put option. Let S be the price of
the asset underlying the contract, and K and T be the exercise price and the expiration date.
Finally, let t be the current evaluation time. The following relations hold true,
+
+
0
if S(T ) K
K S(T ) if S(T ) K
C(T ) =
p(T ) =
S(T ) K if S(T ) > K
0
if S(T ) > K
or more succinctly, C(T ) = (S(T ) K)+ and p(T ) = (K S(T ))+ .
Figure 10.1 depicts the net prots generated by a number of portfolios that include one asset
(a share, say) and/or one option written on the share: buy a share, buy a call, short-sell a
share, buy a put, etc. To simplify the exposition, we assume that the short-term rate r = 0.
The rst two panels in Figure 10.1 illustrate instances in which the exposure to losses is less
important when we purchase an option than the share underlying the option contract. For
example, consider the rst panel, in which we depicts the two net prots related to (i) the
purchase of the share and (ii) the purchase of the call option written on the share. Both cases
generate positive net prots when S(T ) is high. However, the call option provides protection
when S(T ) is low, at least insofar as C(t) < S(t), a no-arbitrage condition we shall demonstrate
later. It is this protecting feature to make the option contract economically valuable.
c
by
A. Mele
10.2. General properties of options
(S)T = S (T) S (t)
(c)T = c(T) c(t)
S(T)
K + c(t)
S(T)
S(t)
c(t)
S(t)
Buy share
S(t)
Buy call
(-S)T = S(t) - S(T)
(p)T = p(T) - p(t)
S(T)
K - p(t)
S(t)
S(T)
-p(t)
Short-sell share
Buy put
(-c)T = c(t) - c(T)
(-p)T = p(t) - p(T)
c(t)
p(t)
K + c(t)
K
S(T)
K
K - p(t)
p(t) - K
Sell call
Sell put
FIGURE 10.1.
248
S(T)
c
by
A. Mele

A
(c)T = c(T) - c(t)
K
-S(t)
K+c(t)
S(T)
-c(t)
-(K+c(t))
-S(t) B
FIGURE 10.2.
Figure 10.2 illustrates this. It depicts two cases. The rst case is such that the share price is
less than the option price. In this case, the net prots generated by the purchase of the share,
the AA line, would strictly dominate the net prots generated by the purchase of the option,
(C)T , an arbitrage opportunity. Therefore, we need that C (t) < S (t) to rule this arbitrage
opportunity. Next, let us consider a second case, which arises when the share and option prices
are such that the net prots generated by the purchase of the share, the BB line, is always
dominated by the net prots generated by the purchase of the option, (C)T , an arbitrage
opportunity. Therefore, the price of the share and the price of the option must be such that
S (t) < K + C (t), or C (t) > S (t) K. Finally, note that the option price is always strictly
positive as the option payo is nonnegative. Therefore, we have that (S (t) K)+ < C (t) <
S (t). Theorem 8.1 below provides a formal derivation of this result in the more general case in
which the short-term rate is positive.
Let us go back to Figure 10.1. This gure suggests that the price of the call and the price
of the put are intimately related. Indeed, by overlapping the rst two panels, and using the
same arguments used in Figure 10.2, we see that for two call and put contracts with the same
exercise price K and expiration date T , we must have that C (t) > p (t) K. The put-call parity
provides the exact relation between the two prices C (t) and p (t). Let P (t, T ) be the price as
of time t of a pure discount bond expiring at time T . We have:
Theorem 10.1 (Put-call parity). Consider a put and a call option with the same exercise
price K and the same expiration date T . Their prices p(t) and C(t) satisfy, p (t) = C (t)
S (t) + KP (t, T ).
Proof. Consider two portfolios: (A) Long one call, short one underlying asset, and invest
KP (t, T ); (B) Long one put. The table below gives the value of the two portfolios at time t
249
c
by
A. Mele

and at time T .
Portfolio A
Portfolio B
Value at t
C (t) S (t) + KP (t, T )
p (t)
Value at T
S(T ) K
S(T ) > K
S(T ) + K S(T ) K S(T ) + K
K S (T )
0
The two portfolios have the same value in each state of nature at time T . Therefore, their values
at time t must be identical to rule out arbitrage.
By the put-cal parity in Theorem 8.1, the properties of European put prices can be mechanically deduced from those of the corresponding call prices. We focus the discussion on European
call options. The following result gathers some basic properties of European call option prices
before the expiration date, and generalizes the reasoning underlying Figure 10.2.
Theorem 10.2. The rational option price C (t) = C (S (t) ; K; T t) satisfies the following
properties: (i) C (S (t) ; K; T t) 0; (ii) C (S (t) ; K; T t) S(t) KP (t, T ); and (iii)
C (S (t) ; K; T t) S (t).
Proof. Part (i) holds because Pr {C (S (T ) ; K; 0) > 0} > 0, which implies that C must be
nonnegative at time t to preclude arbitrage opportunities. As regards Part (ii), consider two
portfolios: Portfolio A, buy one call; and Portfolio B, buy one underlying asset and issue debt
for an amount of KP (t, T ). The table below gives the value of the two portfolios at time t and
at time T .
Value at T
Value at t
S(T ) K S(T ) > K
Portfolio A
C(t)
0
S(T ) K
Portfolio B
S (t) KP (t, T )
S(T ) K S(T ) K
At time T , Portfolio A dominates Portfolio B. Therefore, in the absence of arbitrage, the value
of Portfolio A must dominate the value of Portfolio B at time t. To show Part (iii), suppose the
contrary, i.e. C (t) > S (t), which is an arbitrage opportunity. Indeed, at time t, we could sell
m options (m large) and buy m of the underlying assets, thus making a sure prot equal to
m (C (t) S (t)). At time T , the option will be exercized if S (T ) > K, in which case we shall
sell the underlying assets and obtain m K. If S (T ) < K, the option will not be exercized, and
we will still hold the asset or sell it and make a prot equal to m S (T ).
Theorem 10.2 can be summarized as follows:
max {0, S (t) KP (t, T )} C (S (t) ; K; T t) S (t) .
The next corollary follows by Eq. (10.1).
250
(10.1)
c
by
A. Mele

c(t)
45
B
K b(t,T)
c(t)
S(t)
c(t)
S(t)
S(t)
FIGURE 10.3.
Corollary 10.3. We have, (i) limS0 C (S; K; T t) 0; (ii) limK0 C (S; K; T t)

S; (iii) limT C(S; K; T t) S.
The previous results provide the basic properties, or arbitrage bounds, that option prices
satisfy. First, consider the top panel of Figure 10.3. Eq. (10.1) tells us that C (t) must lie inside
the AA and the BB lines. Second, Corollary 10.3(i) tells us that the rational option price starts
from the origin. Third, Eq. (10.1) also reveals that as S , the option price also goes to
innity; but because C cannot lie outside the the region bounded by the AA line and the BB
lines, C will go to innity by sliding up through the BB line.
How does the option price behave within the bounds AA and BB? That is impossible to
tell. Given the boundary behavior of the call option price, we only know that if the option
price is strictly convex in S, then it is also increasing in S, In this case, the option price could
be as in the left-hand side of the bottom panel of Figure 10.3. This case is the most relevant,
empirically. It is predicted by the celebrated Black and Scholes (1973) formula. However, this
property is not a general property of option prices. Indeed, Bergman, Grundy and Wiener
(1996) show that in one-dimensional diffusion models, the price of a contingent claim written
on a tradable asset is convex in the underlying asset price if the payo of the claim is convex
in the underlying asset price (as in the case of a Europen call option). In our context, the
boundary conditions guarantee that the price of the option is then increasing and convex in the
price of the underlying asset. However, Bergman, Grundy and Wiener provide several counterexamples in which the price of a call option can be decreasing over some range of the price of
the asset underlying the option contract. These counter-examples include models with jumps,
or the models with stochastic volatility that we shall describe later in this chapter. Therefore,
there are no reasons to exclude that the option price behavior could be as that in right-hand
side of the bottom panel of Figure 10.3.
251
c
by
A. Mele

c(t)
T1
T2
T3
45
K b(t,T)
S(t)
FIGURE 10.4.
The economic content of convexity in this context is very simple. When the price S is small,
it is unlikely that the option will be exercized. Therefore, changes in the price S produce little
eect on the price of the option, C. However, when S is large, it is likely that the option
will be exercized. In this case, an increase in S is followed by almost the same increase in C.
Furthermore, the elasticity of the option price with respect to S is larger than one,
dC S
> 1,
dS C
as for a convex function, the rst derivative is always higher than the secant. Overall, the option
price is more volatile than the price of the asset underlying the contract. Finally, call options
are also known as wasting assets, as their value decreases over time. Figure 10.4 illustrate
this property, in correspondence of the maturity dates T1 > T2 > T3 .
These properties illustrate very simply the general principles underlying a portfolio that
mimicks the option price. For example, investment banks sell options that they want to
hedge against, to avoid the exposure to losses illustrated in Figure 10.1. At a very least, the
portfolio that mimics the option price must exhibit the previous general properties. For
example, suppose we wish our portfolio to exhibit the behavior in the left-hand side of the
bottom panel of Figure 10.3, which is the most relevant, empirically. We require the portfolio
to exhibit a number of properties.
(p-i) The portfolio value, V , must be increasing in the underlying asset price, S.
(p-ii) The sensitivity of the portfolio value with respect to the underlying asset price must be
strictly positive and bounded by one, 0 < dV
< 1.
dS
(p-iii) The elasticity of the portfolio value with respect to the underlying asset price must be
strictly greater than one, dV
VS > 1.
dS
The previous properties hold under the following conditions:
(c-i) The portfolio includes the asset underlying the option contract.
(c-ii) The number of assets underlying the option contract is less than one.
252
c
by
A. Mele
10.3. Evaluation
(c-iii) The portfolio includes debt to create a suciently large elasticity. Indeed, let V = S D,
where is the number of assets underlying the option contract, with (0, 1), and D is
debt. Then, dV
> and dV
S = VS > 1 S > V = S D, which holds if and only
dS
dS V
if D > 0.
In fact, the hedging problem is dynamic in nature, and we would expect to be a function
of the underlying asset price, S, and time to expiration. Therefore, we require the portfolio to
display the following additional property:
(p-iv) The number of assets underlying the option contract must increase with S. Moreover,
when S is low, the value of the portfolio must be virtually insensitive to changes in
S. When S is high, the portfolio must include mainly the assets underlying the option
contract, to make the portfolio value slide up through the BB line in Figure 10.3.
The previous property holds under the following condition:
(c-iv) is an increasing function of S, with limS0 (S) 0 and limS (S) 1.
Finally, the purchase of the option does not entail any additional inows or outows until time
to expiration. Therefore, we require that the mimicking portfolio display a similar property:
(p-v) The portfolio must be implemented as follows: (i) any purchase of the asset underlying
the option contract must be nanced by issue of new debt; and (ii) any sells of the asset
underlying the option contract must be used to shrink the existing debt:
The previous property of the portfolio just says that the portfolio has to be self-nancing, in
the sense described in the rst Part of these lectures.
(c-v) The portfolio is implemented through a self-financing strategy.
We now proceed to add more structure to the problem.
10.3 Evaluation
We consider a continuous-time model in which asset prices are driven by a d-dimensional Brownian motion W .1 We consider a multivariate state process

dY (h) (t) = h (y (t)) dt + dj=1 hj (y (t)) dW (j) (t) ,
for some functions h and hj (y), satisfying the usual regularity conditions.
The price of the primitive assets satises the regularity conditions in Chapter 4. The value of
a portfolio strategy, V , is V (t) = (t) S+ (t). We consider a self-nancing portfolio. Therefore,
V is solution to
!
dV (t) = (t) ( (t) 1m r (t)) + r (t) V (t) C (t) dt + (t) (t) dW (t) ,
(10.2)
where ( (1) , ..., (m) ) , (i) (i) S (i) , ((1) , ..., (m) ) , S (i) is the price of the i-th asset,
(i) is its drift and (t) is the volatility matrix of the price process. We impose that V satisfy
the same regularity conditions in Chapter 4.
1 As usual, we let {F (t)}
W (t) = (W (s) , s t) generated by W , with
t[0,T ] be the P -augmentation of the natural ltration F
F = F (T ).
253
c
by
A. Mele
10.3. Evaluation
10.3.1
On spanning and cloning
A set of securities spans a given vector space, say a set of economic payos, if any point
in that space can be generated by a linear combination of the security prices. The set of
economic payos may include those promised by a contingent claim, such as that promised by
a European call, or even a certain targeted consumption, as in Harrison and Kreps (1979) and
Due and Huang (1985). Chapter 4 relies on spanning, to solve for consumption-portfolio
choices through martingale techniques. This section emphasizes how spanning helps dening
replicating strategies that lead to price redundant assets.
Formally, let V x, (t) denote the solution to Eq. (10.2) when the initial wealth is x, the
portfolio policy is , and the intermediate consumption is C 0. We say that the portfolio
almost surely, where X
is any square-integrable F (T )policy spans F (T ) if V x, (T ) = X
measurable random variable.
Chapter 4 provides a characterization of the spanning property relying on the behavior of
asset prices under the risk-adjusted probability measure, Q. In the context of this chapter, it
is convenient to analyze the behavior of the asset prices under the physical probability, P . In
the diusion environment of this chapter, the asset prices are semimartingales under P . More
generally, let consider the following representation of a F(t)-P semimartingale,
dA(t) = dF (t) + (t)dW (t),
where F is a process with nite variation, and L20,T,d (, F , P ). We wish to replicate A
through a portfolio. First, then, we must look for a portfolio satisfying
(t) = (t) (t) .
(10.3)
Second, we equate the drift of V to the drift of F , obtaining,

dF (t)
= (t) ( (t) 1m r (t)) + r (t) V (t) = (t) ( (t) 1m r (t)) + r (t) F (t) .
dt
(10.4)
The second equality holds because if drift and diusion terms of F and V are identical, then
F (t) = V (t).
Clearly, if m < d, there are no solutions for in Eq. (10.3). The economic interpretation is that
in this case, the number of assets is so small that we can not create a portfolio able to replicate
all possible events in the future. Mathematically, if m < d, then V x, (T ) M L2 (, F , P ).
As Chapter 4 emphasizes, there is also a converse to this result, which motivates the denition
of market incompleteness given in Chapter 4 (Denition 4.5).
10.3.2
Pricing
Eq. (10.4) is all we need to price any derivative asset which promises to pay o some F (T )measurable random variable. The rst step is to cast Eq. (10.4) in a less abstract format. Let
us consider a semimartingale in the context of the diusion model of this section. For example,
let us consider the price process H (t) of a European
call option.
This price process is rationally

formed if H (t) = C (t, y (t)), for some C C 1,2 [0, T ) Rk . By Itos lemma,
C
dC =
Cdt +
254

C
J dW,
Y
c
by
A. Mele
10.3. Evaluation
where
C C = C
+
t
d d. Finally,
k
C
l=1 yl l
(t, y) +
1
2
k k
l=1
2C
j=1 yl yj cov (yl , yj );
C/ Y is 1 d; and J is
L2 (, F , P ) .
C (T, y) = X
In this context,
C C and (C/ Y ) J play the same roles as dF / dt and in the previous
C
section. In particular, the volatility identication Y
J = , corresponds to Eq. (10.3).
As an example, let m = d = 1, and suppose that the only state variable of the economy is
the price of a share, (s) = s, cov(s) = 2 s2 , and and 2 are constants, then (C/ Y ) J =
CS S, = CS S, and by Eq. (10.4),
1 2C 2 2
C
C C
S = ( r) + rC =
+
S +
S ( r) + rC.
2
t
S
2 S
S
(10.5)
By the boundary condition, C (T, s) = (s K)+ , one obtains that the solution is the celebrated
Black and Scholes (1973) formula,
C (t, S) = S(d1 ) Ker(T t) (d2 )
(10.6)
where
S
ln( K
) + (r + 12 2 )(T t)
,
d1 =
T t
d2 = d1 T t,
1
(x) =
2
"
e 2 u du.
Note that we can obtain Eq. (10.6) even without assuming that a market exists for the
option,2 and that the pricing function C (t, S) is dierentiable. As it turns out, the option price is
dierentiable, but this can be shown to be a result, not just an assumption. Indeed, let us dene
the function C (t, S) that solves Eq. (10.5), with boundary condition C (T, S) = (S K)+ .
Note, we are not assuming this function is the option price. Rather, we shall show this is the
option price. Consider a self-nanced portfolio of bonds and stocks, with = CS S. Its value
satises,
dV = [CS S( r) + rV ] dt + CS SdW.
Moreover, by Itos lemma, C (t, s) is solution to

1 2 2
dC = Ct + SCS + S CSS dt + CS SdW.
2
By subtracting the previous two equations,
1
dV dC = [Ct rSCS 2 S 2 CSS + rV ]dt = r (V C) dt.

2

=rC
Hence, we have that V ( ) C ( , S ( )) = [V (0) C (0, S (0))] exp(r ), for all [0, T ].
Next, assume that V (0) = C (0, S (0)). Then, V ( ) = C ( , S ( )) and V (T ) = C (T, S (T )) =
(S (T ) K)+ . That is, the portfolio = CS S replicates the payo underlying the option
contract. Therefore, V ( ) equals the market price of the option. But V ( ) = C ( , S ( )), and
we are done.
2 The original derivation of Black and Scholes (1973) and Merton (1973) relies on the assumption that an option market exists
(see the Appendix to this chapter).
255
c
by
A. Mele
10.4. Properties of models

10.3.3
Surprising cancellations and preference-free formulae
Due to what Heston (1993a) (p. 933) terms a surprising cancellation, the constant doesnt
show up in the nal formula. Heston (1993a) shows that this property is not robust to modications in the assumptions for the underlying asset price process. Gamma processes, incomplete
markets.
10.4 Properties of models

10.4.1
Rational price reaction to random changes in the state variables
We now derive some general properties of option prices arising in the context of diusion
processes. The discussion in this section hinges upon the seminal contribution of Bergman,
Grundy and Wiener (1996).3 We take as primitive,
dS (t)
= (S (t)) dt + (S (t)) dW (t) ,
S (t)
(s) =

2v (s)
(10.7)
and develop some properties of a European-style option price at time t, denoted as C (S (t) , t, T ),
where T is time-to-expiration. Let the payo of the option be the function (S), where satises (S) > 0. In the absence of arbitarge, C satises the following partial dierential equation
+
0
= Ct + CS r + CSS v(S) rC
for all ( , S) [t, T ) R++
C(S, T, T ) = (S)
for all S R++
Let us dierentiate the previous partial dierential equation with respect to S. The result is
that H CS satises another partial dierential equation,
+
0
= Ht + (r + v (S)) HS + HSS (S) rH
for all ( , S) [t, T ) R++
H(S, T, T ) = (S) > 0
for all S R++
(10.8)
By the maximum principle reviewed in the Appendix of Chapter 7, we have that,
H (S, , T ) > 0 for all ( , S) [t, T ] R++ .
That is, we have that in the scalar diffusion setting, the option price is always increasing in the
underlying asset price.
Next, let us consider how the price behave when we tilt the volatility of the underlying asset
price. That is, consider two economies A and B with prices (C i , S i )i=A,B . We assume that the
asset price volatility is larger in the rst economy than in the second economy, viz
i
dS i ( )
i
( ) , i = A, B,
=
rd
+
S ( ) dW
S i ( )
is Brownian motion under the risk-neutral probability, i is as in Eq. (10.7), and

where W
A
(s) > B (s), for all s. It is easy to see that the price dierence, C C A C B , satises,

B
+
0 = C + rCS + CSS A (S) rC + A B CSS
, for all ( , S) [t, T ) R++
C = 0, for all S
(10.9)
3 Moreover,
Eqs. (10.8), (10.9) and (10.10) below can be seen as particular cases of the general results given in Chapter 6.
256
c
by
A. Mele
10.4. Properties of models
By the maximum principle, again, C > 0 whenever CSS > 0. Therefore, it follows that if
option prices are convex in the underlying asset price, then they are also always increasing in
the volatility of the underlying asset prices. Economically, this result follows because volatility
changes are mean-preserving spread in this context. We are left to show that CSS > 0. Let
us dierentiate Eq. (10.8) with respect to S. The result is that Z HS = CSS satises the
following partial dierential equation,
+
0 = Z + (r + 2v (S)) ZS + ZSS (S) (r (S))Z
for all ( , S) [t, T ) R++
H(S, T, T ) = (S)
for all S R++
(10.10)
By the maximum principle, we have that
H (S, , T ) > 0 for all ( , S) [t, T ] R++ , whenever (S) > 0 S R++ .
That is, we have that in the scalar diffusion setting, the option price is always convex in the
underlying asset price if the terminal payoff is convex in the underlying asset price. In other
terms, the convexity of the terminal payo propagates to the convexity of the pricing function.
Therefore, if the terminal payoff is convex in the underlying asset price, then the option price
is always increasing in the volatility of the underlying asset price.
10.4.2
Recoverability of the risk-neutral density from option prices
Consider the price of a European call,

"
"
+
C (S(t), t, T ; K) = P (t, T )
[S(T ) K] dQ (S(T )| S(t)) = P (t, T )
0
(x K) q (x| S(t)) dx,
where Q is the risk-neutral measure and q( x+ | x)dx dQ( x+ | x). This is indeed a very general
formula, as it does not rely on any parametric assumptions for the dynamics of the price
underlying the option contract. Let us dierentiate the previous formula with respect to K,
"
r(T t) C (S(t), t, T ; K)
e
=
q (x| S(t)) dx,
K
K
where we assumed that limx x q ( x+ | S(t)) = 0. Let us dierentiate again,
er(T t)
2 C (S(t), t, T ; K)
= q ( K| S(t)) .
K 2
(10.11)
Eq. (10.11) provides a means to recover the risk-neutral density using option prices. The
Arrow-Debreu state density, AD (S + = u| S(t)), is given by,

2

+

+

r(T t)
2r(T t) C (S(t), t, T ; K)

AD S = u S(t) = e
q S S (t) S + =u = e
.

K 2
K=u
These results are quite useful in applied work.
10.4.3
Endogeneous volatility
Hedges and crashes

257
c
by
A. Mele
10.5. Stochastic volatility
10.5 Stochastic volatility

10.5.1
Statistical models of changing volatility
One of the most prominent advances in the history of empirical nance is the discovery that
nancial returns exhibit both temporal dependence in their second order moments and heavypeaked and tailed distributions. Such a phenomenon was known at least since the seminal
work of Mandelbrot (1963) and Fama (1965). However, it was only with the introduction of
the autoregressive conditionally heteroscedastic (ARCH) model of Engle (1982) and Bollerslev
(1986) that econometric models of changing volatility have been intensively tted to data.
An ARCH model works as follows. Let {yt }N
t=1 be a record of observations on some asset
returns. That is, yt = ln St /St1 , where St is the asset price, and where we are ignoring dividend issues. The empirical evidence suggests that the dynamics of yt are well-described by the
following model:
yt = a + t ,
t | Ft1 N(0, 2t ),
2t = w + 2t1 + 2t1 ,
(10.12)
where a, w, and are parameters and Ft denotes the information set as of time t. This model
is known as the GARCH(1,1) model (Generalized ARCH). It was introduced by Bollerslev
(1986), and collapses to the ARCH(1) model introduced by Engle (1982) once we set = 0.
ARCH models have played a prominent role in the analysis of many aspects of nancial
econometrics, such as the term structure of interest rates, the pricing of options, or the presence
of time varying risk premia in the foreign exchange market. A classic survey is that in Bollerslev,
Engle and Nelson (1994).
The quintessence of ARCH models is to make volatility dependent on the variability of past
observations. An alternative formulation, initiated by Taylor (1986), makes volatility driven
by some unobserved components. This formulation gives rise to the stochastic volatility model.
Consider, for example, the following stochastic volatility model,
t | Ft1 N(0, 2t );
yt = a + t ;
ln 2t = w + ln 2t1 + ln 2t1 + t ;
t | Ft1 N (0, 2 )
where a, w, , and 2 are parameters. The main dierence between this model and the
GARCH(1,1) model in Eq. (10.12) is that the volatility as of time t, 2t , is not predetermined
by the past forecast error, t1 . Rather, this volatility depends on the realization of the stochastic
volatility shock t at time t. This makes the stochastic volatility model considerably richer than
a simple ARCH model. As for the ARCH models, SV models have also been intensively used,
especially following the progress accomplished in the corresponding estimation techniques. The
seminal contributions related to the estimation of this kind of models are mentioned in Mele
and Fornari (2000). Early contributions that relate changes in volatility of asset returns to
economic intuition include Clark (1973) and Tauchen and Pitts (1983), who assume that a
stochastic process of information arrival generates a random number of intraday changes of the
asset price.
10.5.2
Implied volatility and smiles
Parallel to the empirical research into asset returns volatility, practitioners and academics realized that the assumption of constant volatility underlying the Black and Scholes (1973) and
258
c
by
A. Mele
Merton (1973) formulae was too restrictive. The Black-Scholes model assumes that the price of
the asset underlying the option contract follows a geometric Brownian motion,
dSt
= d + dWt ,
St
where W is a Brownian motion, and , are constants. As explained earlier, is the only
parameter to enter the Black-Scholes-Merton formulae.
The assumption that is constant is inconsistent with the empirical evidence reviewed in
the previous section. This assumption is also inconsistent with the empirical evidence on the
cross-section of option prices. Let CBS (St , t; K, T, ) be the option price predicted by the BlackScholes formula, when the stock price is St , the option contract has a strike price equal to K,
and the maturity is K, and let the market price be Ct$ (K, T ). Then, empirically, the implied
volatility, i.e. the value of that equates the Black-Scholes formula to the market price of the
option, IV say,
CBS (St , t; K, T, IV) = Ct$ (K, T )
(10.13)
depends on the moneyness of the option, dened as,
mo
St er(T t)
,
K
where r is the short-term rate, K is the strike of the option, and T is the maturity date of
the option contract. By the results in Section 10.4.1, we know the Black-Scholes option price is
strictly increasing in . Therefore, the previous denition makes sense, in that there exists an
unique value IV such that Eq. (10.13) holds true. In fact, the market practice is to quote option
prices in terms of their implied volatilities - rather than in terms of prices. Moreover, this same
implied volatility relates to both the call and the put option prices. Consider the put-call parity
in Theorem 10.1,
Pt (K, T ) = Ct (K, T ) St + Ker(T t) ,
Naturally, for each , this same equation must necessarily hold for the Black-Scholes model, i.e.
PBS (St , t; K, T, ) = CBS (St , t; K, T, ) St + Ker(T t) . Subtracting this equation from the
previous one, we see that, the implied volatilities of a call and a put options are the same.
The crucial empirical point is that the IV exhibits a pattern. Before 1987, it did not display
1
a clear pattern, or a -shaped pattern in mo
at best, a smile. After the 1987 crash, the smile
turned in to a smirk, also referred to as volatility skew.
Why a smile? There are many explanations. The rst, is that options (be they call or puts)
that are deep-in-the-money and options (be they call or puts) that are deep-out-of the money are
relatively less liquid and therefore command a liquidity risk-premium. Since the Black-Scholes
1
option price is increasing in volatility, the implied volatility is then, -shaped in mo
.
A second explanation relates to the Black-Scholes assumption that asset returns are lognormally distributed. This assumption may not be correct, as the market might be pricing using
an alternative distribution. One possibility is that such an alternative distribution puts more
weight on the tails, as a result of the market fears about the occurrence of extreme outcomes.
For example, the market might fear the stock price will decrease under a certain level, say K.
As a result, the market density should then have a left tail ticker than that of the log-normal
density, for values of S < K. This implies that the probability deep-out-of-the-money puts (i.e.,
those with low strike prices) will be exercized is higher under the market density than under the
259
c
by
A. Mele
log-normal density. In other words, the volatility needed to price deep-out-of-the-money puts is
larger than that needed to price at-the-money calls and puts.
then, the
At the other extreme, if the market fears that the stock price will be above some K,
market density should exhinit a right tail ticker than that of the log-normal density, for values
which implies a larger probability (compared to the log-normal) that deep-out-of-theof S > K,
money calls (i.e., those with high strike prices) will be exercized. Then, the implied volatility
needed to price deep-out-of-the-money calls is larger than that needed to price at-the-money
calls and puts. The second eect has disappeared since the 1987 crash, leaving the smirk.
Ball and Roma (1994) and Renault and Touzi (1996) were the rst to note that a smile eect
arises when the asset return exhibits stochastic volatility. In continuous time,4
dS (t)
= dt + (t) dW (t)
S (t)
d 2 (t) = b(S (t) , (t))d + a(S (t) , (t))dW (t)
(10.14)
where W is another Brownian motion, and b and a are some functions satisfying the usual
regularity conditions. In other words, let us suppose that Eqs. (10.14) constitute the data
generating process. Then, the fundamental theorem of asset pricing (FTAP, henceforth) tells
us that there is a probability equivalent to P , Q say (the risk-neutral probability), such that
the rational option price C(S(t), (t)2 , t, T ) is given by,

C(S(t), (t)2 , t, T ) = er(T t) E (S(T ) K)+ S(t), (t)2 ,
where E [] is the expectation taken under the probability Q. Next, if we continue to assume
that option prices are really given by the previous formula, then, by inverting the Black-Scholes
formula produces a constant volatility that is -shaped with respect to K.
The rst option pricing models with stochastic volatility are developed by Hull and White
(1987), Scott (1987) and Wiggins (1987). Explicit solutions have always proved hard to derive.
If we exclude the approximate solution provided by Hull and White (1987) or the analytical
solution provided by Heston (1993b),5 we typically need to derive the option price through
some numerical methods based on Montecarlo simulation or the numerical solution to partial
dierential equations.
In addition to these important computational details, models with stochastic volatility raise
serious economic concerns. Typically, the presence of stochastic volatility generates market
incompleteness. As we pointed out earlier, market incompleteness means that we can not hedge
against future contingencies. In our context, market incompleteness arises because the number
of the assets available for trading (one) is less than the sources of risk (i.e. the two Brownian
motions).6 In our option pricing problem, there are no portfolios including only the underlying
4 In an important paper, Nelson (1990) shows that under regularity conditions, the GARCH(1,1) model converges in distribution
to the solution of the following stochastic dierential equation:
d( )2 = ( ( )2 )dt + ( )2 dW ( ),
where W is a standard Brownian motion, and , , and are parameters. See Mele and Fornari (2000) (Chapter 2) for additional
results on this kind of convergence. Corradi (2000) develops a critique related to the conditions underlying these convergence results.
5 The Hestons solution relies on the assumption that stochastic volatility is a linear mean-reverting square-root process. In
a square root process, the instantaneous variance of the process is proportional to the level reached by that process: in model
(10.14), for instance, a(S, ) = a , where a is a constant. In this case, it is possible to show that the characteristic function is
exponential-ane in the state variables S and . Given a closed-form solution for the characteristic function, the option price is
obtained through standard Fourier methods.
6 Naturally, markets can be completed by the presence of the option. However, in this case the option price is not preference
free.
260
c
by
A. Mele
asset and a money market account that could replicate the value of the option
at the expiration

date. Precisely, let C be the rationally formed price at time t, i.e. C ( ) = C S ( ) , ( )2 , , T ,
where ( )2 is driven by a Brownian motion W , which is dierent from W . The value of the
portfolio that only includes the underlying asset is only driven by the Brownian motion driving
the underlying asset price, i.e. it does not include W . Therefore, the value of the portfolio does
not factor in all the random uctuations that move the return volatility, ( )2 . Instead, the
option price depends on this returnvolatility as we have
assumed that the option price, C ( ),
2
is rationally formed, i.e. C ( ) = C S ( ) , ( ) , , T .
In other words, trading with only the underlying asset can not lead to a perfect replication of
the option price, C. In turn, rembember, a perfect replication of C is the condition we need to
obtain a unique preference-free price for the option. To summarize, the presence of stochastic
volatility introduces two inextricable consequences:7
There is an innity of option prices that are consistent with the requirement that there
are no arbitrage opportunities.
Perfect hedging strategies are impossible. Instead, we might, alternatively, either (i) use
a strategy, which is not self-nanced, but that allows for a perfect replication of the claim
or (ii) a self-nanced strategy for some misspecied model. In case (i), the strategy leads
to a hedging cost process. In case (ii), the strategy leads to a tracking error process, but
there can be situations in which the claim can be super-replicated, as we explain below.
10.5.3
Stochastic volatility and market incompleteness
Let us suppose that the asset price is solution to Eqs. (10.14). To simplify, we assume that W
and W are independent. Since C is rationally formed, C( ) = C(S( ), ( )2 , , T ). By Itos
lemma,

1 2 2
1 2
C
+ SCS + bC2 + S CSS + a C2 2 d + SCS dW + aC2 dW .
dC =
t
2
2
Next, let us consider a self-nanced portfolio that includes (i) one call, (ii) shares, and
(iii) units of the money market account (MMA, henceforth). The value of this portfolio is
V = C S P , and satises
dV = dC dS dP

1 2 2
1 2
C
=
+ S (CS ) + bC2 + S CSS + a C2 2 rP d + S (CS ) dW + aC2 dW .
t
2
2
As is clear, only when a = 0, we could zero the volatility of the portfolio value. In this case,
we could set = CS and P = C S V , leaving

C
1 2 2
dV =
+ bC2 + S CSS rC + rSCS + rV d ,
t
2
7 The
mere presence of stochastic volatility is not necessarily a source of market incompleteness. Mele (1998) (p. 88) considers
a circular market with m asset prices, in which (i) the asset price no. i exhibits stochastic volatility, and (ii) this stochastic
volatility is driven by the Brownian motion driving the (i 1)-th asset price. Therefore, in this market, each asset price is solution
to the Eqs. (10.14) and yet, by the previous circular structure, markets are complete.
261
c
by
A. Mele
where we have used the equality V = C. The previous equation shows that the portfolio is
locally riskless. Therefore, by the FTAP,
0=
C
1
+ bC2 + 2 S 2 CSS rC + rSCS + rV = rV.
t
2
The previous equation generalizes the Black-Scholes equation to the case in which volatility
is time-varying and non-stochastic, as a result of the assumption that a = 0. If a = 0, return
volatility is stochastic and, hence, there are no hedging portfolios to use to derive a unique
option price. However, we still have the possibility to characterize the price of the option.
Indeed, consider a self-nanced portfolio with (i) two calls with dierent strike prices and
maturity dates (with weights 1 and ), (ii) shares, and (iii) units of the MMA. We
denote the price processes of these two calls with C 1 and C 2 . The value of this portfolio is
V = C 1 + C 2 S P , and satises,
dV = dC 1 + dC 2 dS dP

= LC 1 + LC 2 S rP d + S CS1 + CS2 dW + a C12 + C22 dW ,
i
i
where LC i C
+ SCSi + bCi 2 + 12 2 S 2 CSS
+ 12 a2 Ci 2 2 , for i = 1, 2. In this context, risk can
t
be eliminated. Indeed, set
C12
= 2 and = CS1 + CS2 .
C2
The value of this portfolio is solution to,

dV = LC 1 + LC 2 S + rV + rS rC 1 rC 2 d .
Therefore, by the FTAP,
0 = LC 1 + LC 2 S + rS rC 1 rC 2

= LC 1 rC 1 CS1 (S rS) + LC 2 rC 2 CS2 (S rS)
where the second equality follows by the denition of , and by rearranging terms. Finally, by
using the denition of , and by rearranging terms,
LC 2 rC 2 CS2 (S rS)
LC 1 rC 1 CS1 (S rS)
=
.
C12
C22
(10.15)
These ratios agree. So they must be equal to some process a (say) independent of both the
strike prices and the maturity of the options. Therefore, we obtain that,
C
1
1
+ rSCS + [b a ] C2 + 2 S 2 CSS + a2 C2 2 = rC.
t
2
2
(10.16)
The economic interpretation of is that of the unit risk-premium required to face the risk
of stochastic uctuations in the return volatility. The problem, the requirement of absence of
arbitrage opportunities does not suce to recover a unique . In other words, by the FeynmanKac stochastic representation of a solution to a PDE, we have that the solution to Eq. (10.16)
is,

C(S(t), (t)2 , t, T ) = er(T t) EQ (S(T ) K)+ S(t), (t)2 ,
(10.17)
262
c
by
A. Mele
where Q is a risk-neutral probability.

Eqs. (10.15) and (10.16) can be interpreted as APT relations. Indeed, let us dene the unit
risk-premium related to the uctuations of the asset price, = ( r) /. Then, Eq. (10.15)
or Eq. (10.16) imply that,

dC
CS
LC
C2
=E
=r+
S +
a ,
C
C
C
C

S
where S is the beta related to the volatility of the option price induced by uctuations in
the stock price, S, and 2 is the beta related to the volatility of the option price induced by
uctuations in the return volatility.
10.5.4
Trading volatility
Buying volatility is a strategy relying on the expectation future volatility will increase. There
are option-based trading strategies that allow us to have views about volatility, such as straddles
or strangles. A straddle is a portfolio of one call option and one put option that have the same
strike price and the same maturity. A strangle is the same as a straddle, with the dierence
that the strike of the call diers from that of the put. In 1995, the 233-year old Barings Bank
collapsed, because of the famous short-straddle Nick Leeson was implementing on the Nikkei
Index. A short-straddle is, of course, a view volatility will not raise. However, in January 1995,
a quite violent earthquake made the Nikkei index crash by almost 7% in a week. The straddle
was naked, i.e. delta-hedged, at most, which leads to losses Leeson was not only unable to
absorb, but also to amplify, given he was insisting on having views the Index would stabilize.
The Index did not.
To understand straddles, and the reason why they are not necessarily the best way to take
views about volatility, consider the simplest strategy, where one buys an option and hedges it
through the Black and Scholes formula. Suppose to live in a world with stochastic volatility,
and purchase a call at time t, with market price equal to C (St , 2t , t) - we are assuming the
world moves exactly as in Eqs. (10.14). Build up a self-nanced portfolio with value Vt ,
Vt = at St + bt Bt ,
(10.18)
where Bt = ert is the money market account,

V0 = C S0 , 20 , 0 , at = BS (St , t; IV0 ) ,
and IV0 is the Black-Scholes implied volatility as of time t = 0, i.e. the time at which we are
to take a view on future volatility.
Consider, rst, the following heuristic arguments. Assume the short-term rate is zero. The,
say, daily, prot and loss (P&L, henceforth) of call options valued at t , with zero-delta satises,
approximately,
-
.

2
1
1
1
1
S
t
P&Lt = t = t t+ t (St )2 = t St2 IV20 t+ t (St )2 = t St2
IV20 t ,
2
2
2
2
St
2
where =
, = S2 , the Gamma, and the second equality follows by a well-known property
t
of the Black-Scholes pricing equation. Aggregating the daily P&L until the maturity of the
263
c
by
A. Mele

option, we obtain:
T
1
P&LT =
t St2
2 t=1
-
St
St
2
IV20 t .
(10.19)
Hence, a portfolio
2 of options is a quite basic way to have views about the movements of future
t
. It may lead to diculties, however, as described below.
volatility, S
St
Next, consider a more rigorous argument, leading to the previous conclusions, originally put
forth by El Karoui, Jeanblanc-Picque and Shreve (1998). Consider the value of the self-portfolio
in Eq (10.18). Because this portfolio is self-nanced,
dVt = at dSt + rbt Bt dt
= rVt dt + at (dSt rSt dt)
= [rVt + at ( r) St ] dt + at t St dWt .
Moreover, by Itos lemma,
dCBS (St , t; K, T, IV0 ) =
CBS 1 2 2 2 CBS
CBS
+ St
+ t St
t
S
2
S 2
dt + t St
CBS
dWt ,
S
where,
CBS 1 2 2 2 CBS
CBS
+ St
+ t St
t
S
2
S 2
2 2 CBS
CBS
CBS
2 CBS
CBS 1 2
2
+ rSt
+ IV20 St2
+
=
+
(
r)
S
IV
t
0 St
S
S 2
S
2 t
S 2
t
rCBS
= rCBS + ( r) St
2 CBS
CBS 1 2
+
t IV20 St2
.
S
2
S 2
Therefore, the tracking error, or P&Lt , dened as the dierence between the Black-Scholes price
and the portfolio value,
P&Lt CBS (St , t; K, T, IV0 ) Vt ,
satises,
At maturity T :

2 2 CBS
1 2
2
dP&Lt = rP&Lt +
IV0 St
dt.
2 t
S 2
P&LT CBS (ST , T ; K, T, IV0 ) VT
= max {ST K, 0} VT
"
2 CBS
1 rT T rt 2
= e
e
t IV20 St2
dt.
2
S 2
0
(10.20)
This expression is the neat version of Eq. (10.19). It is possible to show that a straddle
strategy leads to twice the expression in Eq. (10.20), with the second partial of the straddle
replacing the Black-Scholes Gamma. Eq. (10.20) has the following implications. We know the
Black-Scholes price is convex. Hence, Eq. (10.20) tells us that even if we do not exactly know
the law of movement for volatility, but still hold the view it will increase in the future, we could:
264
c
by
A. Mele
10.6. Local volatility
(i) buy a call option; (ii) short the Black-Scholes replicating portfolio. The P&L in Eq. (10.20)
shows that this strategy leads to positive prots. Naturally, this is not an arbitrage opportunity.
The critical assumption is that volatility will increase.
Eq. (10.20) shows a disturbing feature. Even if the volatility t is larger than IV0 for most of
the time, the nal P&L may not necessarily lead to a prot. The reason is that each volatility
2C
BS
view, 2t IV20 , is weighted by the Dollar Gamma, St2 S
2 . It may be that bad realization of
2
2
the volatility views, i.e. t < IV0 , occur precisely when the Dollar Gamma is large. This feature
is known as price-dependency. Moreover, the strategy is costly, as it relies on expensive hedging. (Naturally, this issue does not apply to straddles.) Volatility contracts overcome these
diculties, and are described in Section 10.6 below.
10.5.5
Pricing formulae
Hull and White (1987) derive the rst pricing formula of the stochastic volatility literature.
They assume that the return volatility is independent of the asset price, and show that,
!
2
C(S (t) , (t) , t, T ) = EV BS(S(t), t, T ; V ) ,

where BS(S(t), t, T ; V ) is the Black-Scholes formula obtained by replacing the constant 2 with
V , and
" T
1
V =
( )2 d .
T t t
This formula tells us that the option price is simply the Black-Scholes formula averaged over
all the possible values taken by the future average volatility V . A proof of this equation is
given in the appendix.8
The most widely used formula is the Hestons (1993b) formula, which holds when the return
volatility is a square-root process.
10.6 Local volatility

10.6.1
Issues
Stochastic volatility models may EXPLAIN the Smile

Obviously, stochastic volatility models do not allow for a perfect hedge. Their main drawback is that they can not perfectly FIT the Smile.
Towards the end of 1980s and the beginning of the 1990s, interest rates modelers invented
models that allow a perfect t of the initial yield curve.
Important for interest rate derivatives.
In 1993 and 1994, Derman & Kani, Dupire and Rubinstein come up with a technology
that could be applied to options on tradables.
Why is it important to exactly t the structure of already existing plain vanilla options?
8 The result does not hold in the general case in which the asset price and volatility are correlated. However, Romano and Touzi
(1997) prove that a similar result holds in such a more general case.
265
c
by
A. Mele
Plain vanilla versus exotics. Suppose you wish to price exotic, or illiquid, options.
The model you use to price the illiquid option must predict that the plain vanilla option
prices are identical to those your company is selling! How can we trust a model that is
not even able to pin down all outstanding contracts? - Arbitrage opportunities for quants
and traders?
10.6.2
How does it work?
As usual in this context, we model the dynamics of asset prices under the risk-neutral proba be a Brownian motion under the risk-neutral probability, and E the
bility. Accordingly, let W
expectation operator under the risk-neutral probability.
1. Start with a set of actively traded (i.e. liquid) European options. Let K and T be strikes
and time-to-maturity. Let us be given a collection of prices:
C$ (K, T ) C (K, T ) , K, T varying.
2. Is it mathematically possible to conceive a diusion process,
dSt
t,
= rdt + (St , t)dW
St
such that the initial collection of European option prices is predicted without errors by
the resulting model?
3. Yes. We should use the function loc , say, that takes the form,
>
? C(K, T )
C(K, T )
?
+ rK
?
T
K
loc (K, T ) = ?
.
@2
2
C(K,
T
)
K2
K 2
(10.21)
We call this function loc local volatility.
4. Now, we can price the illiquid options through numerical methods. For example, we can use
simulations. In the simulations, we use
dSt
t.
= rdt + loc (St , t) dW
St
It turns out, empirically, that loc (x, t) is typically decreasing in x for xed t, a phenomenon known as the Black-Christie-Nelson leverage eect. This fact leads some practitioners to assume from the outset that (x, t) = x f (t), for some function f and some
constant < 0. This gives rise to the so-called CEV (Constant Elasticity of Variance)
model.
More recently, practitioners use models that combine local vols with stoch vol, such
as
dSt
t
= rdt + (St , t) vt dW
(10.22)
St
v
dvt = (vt )dt + (vt )dWt

266
c
by
A. Mele
v is another Brownian motion, and , are some functions ( includes a riskwhere W

premium). It is possible to show that in this specic case,
loc (K, T )
loc (K, T ) =
E ( vT2 | ST )
(10.23)
would be able to pin down the initial structure of European options prices. (Here loc (K, T )
is as in Eq. (10.21).)
In this case, we simulate
10.6.3
Variance swaps
dSt
t
= rdt +
loc (St , t) vt dW
St
dv = (v )dt + (v )dW
v
t
t
t
t
In fact, it is possible to demonstrate the following general result. Let St satisfy,

dSt
t,
= rdt + t dW
St
where t is Ft -adapted: i.e. Ft can be larger than FtS (S : t). Then,
e
r(T t)

E 2T = 2
"
C(K,T )
T
)
+ rK C(K,T
K
dK.
K2
(10.24)
The previous developments can be used to address very important issues. Dene the total
integrated variance within the time interval [T1 , T2 ], with T1 > t, to be
IVT1 ,T2
"
T2
T1
2u du.
For reasons developed below, let us compute the risk-neutral expectation of such a realized variance. This can easily be done. If r = 0, then by Eq. (10.24),
"
Ct (K, T2 ) Ct (K, T1 )
E (IVT1 ,T2 ) = 2
dK,
(10.25)
K2
0
where Ct (K, T ) is the price as of time t of a call option expiring at T and struck at K.
A proof of Eq. (10.25) is in the appendix.
If r > 0 and T1 = t, T2 T , we have,
-"
.
"
F (t)
P
(K,
T
)
C
(K,
T
)
t
t
E (IVt,T ) = 2er(T t)
dK +
dK ,
2
K
K2
0
F (t)
(10.26)
where F (t) is the forward price: F (t) = er(T t) S (t), and Pt (K, T ) is the price as of time
t of a put option expiring at T and struck at K. A proof of Eq. (10.26) is in the appendix.
267
c
by
A. Mele
In September 2003, the Chicago Board Option Exchange (CBOE) changed its stochastic
volatility index VIX to approximate the variance swap rate of the S&P 500 index return
(for 30 days). In March 2004, the CBOE launched the CBOE Future Exchange for trading
futures on the new VIX. Options on VIX are also forthcoming.
A variance swap is a contract that has zero value at entry (at t). At maturity T , the
buyer of the swap receives,
(IVt,T SWt,T ) notional,
where SWt,T is the swap rate established at t and paid o at time T . Therefore, this
contract is not a swap, really, but a forward. If r is deterministic,
SWt,T = E (IVt,T ) ,
where E (IVt,T ) is given by Eq. (10.26). Therefore, (10.26) is used to evaluate these variance
swaps.
Finally, it is worth mentioning that the previous contracts rely on some notions of realized
volatility as a continuous record of returns is obviously unavailable.
268
c
by
A. Mele
10.7. American options
10.7 American options

10.8 Exotic options
10.9 Market imperfections
269
10.10. Appendix 1: Additional details on Black & Scholes
c
by
A. Mele
10.10 Appendix 1: Additional details on Black & Scholes

10.10.1 The original arguments
The original arguments in Black and Scholes (1973) and Merton (1973) rely on the assumption the
option is already traded. Let dS/ S = d + dW . Create a self-nancing portfolio of n
S units of the
underlying asset and nC units of the European call option, where n
S is an arbitrary number. Such a
portfolio is worth V = n
S S + nC C and since it is self-nancing it satises:
dV = n
S dS + nC dC

1 2 2
=n
S dS + nC CS dS + C + S CSS d
2

1
= (
nS + nC CS ) dS + nC C + 2 S 2 CSS d
2
where the second line follows from Itos lemma. Therefore, the portfolio is locally riskless whenever
nC =
nS
1
,
CS
in which case V must appreciate at the r-rate

C1S C + 12 2 S 2 CSS
nC C + 12 2 S 2 CSS d
dV
d = rd .
=
=
V
n
S S + nC C
S C1S C
The last equality, plus the boundary condition, lead to the Black-Scholes partial dierential equation.
10.10.2 Delta
We have:
BS
(10A.1)
= N(d1 ).
S
Indeed, the Black-Scholes formula is homogeneous of degree one in S and K, that is, BS(S, K) =
BS(S, K). Therefore, bu Eulers theoreom,
BS (S, K) =
BS
BS
S+
K,
S
K
and Eq. (10A.1) then follows by identifying terms in the Black-Scholes formula.
270
c
by
A. Mele
10.11. Appendix 2: Stochastic volatility
10.11 Appendix 2: Stochastic volatility

10.11.1 Proof of the Hull and White (1987) equation
By the law of iterated expectations, (10.17) can be written as:

C(S(t), (t)2 , t, T ) = er(T t) E [S(T ) K]+ S(t), (t)2

!

'
&

= E E er(T t) [S(T ) K]+ S(t), ( )2 [t,T ] S(t), (t)2

!

= E BS S(t), t, T ; V S(t), (t)2

!

= E BS S(t), t, T ; V (t)2
"

= BS S(t), t, T ; V Pr V (t)2 dV

!
EV BS S(t), t, T ; V ,
(10A.2)
where Pr(V | (t)2 ) is the density of V conditional on the current volatility value (t)2 .
In other terms, the price of an option on an asset with stochastic volatility is the expectation of
the Black-Scholes formula over the distribution of the average (random) volatility V . To
& understand
'
better this result, all we have to understand is that conditionally on the volatility path ( )2 [t,T ] ,

)
ln S(T
is normally distributed under the risk-neutral probability measure. To see this, note that
S(t)
under the risk-neutral probability measure,

" T
"
S(T )
1 T
2
ln
( ) d +
( )dW ( ).
= r(T t)
S(t)
2 t
t
Therefore, conditionally upon the volatility path { ( )} [t,T ] ,

S(T )
1
E ln
= r(T t) (T t) V
S(t)
2

" T
S(T )
and var ln
=
( )2 d = (T t) V .
S(t)
t
This shows the claim. It also shows that the Black-Scholes formula can be applied to compute the
inner expectation of the second line of Eq. (10A.2). And this produces the third line of Eq. (10A.2).
The fourth line is trivial to obtain. Given the result of the third line, the only thing that matters in
the remaining conditional distribution is the conditional probability Pr(V | (t)2 ), and we are done.
10.11.2 Simple smile analytics
271
c
by
A. Mele
10.12. Appendix 3: Technical details for local volatility models
10.12 Appendix 3: Technical details for local volatility models

In all the proofs to follow, all expectations are taken to be expectations conditional on Ft . However,
to simplify notation, we simply write E ( | ) E ( | , Ft ).
Proof of Eqs. (10.23) and (10.24). We rst derive Eq. (10.23), a result encompassing Eq.
(10.21). By assumption,
dSt
t,
= rdt + t dW
St
where t is some Ft -adapted process. For example, t (St , t) vt , all t, where vt is solution to the
2nd equation in (10.22). Next, by assumption we are observing a set of option prices C (K, T ) with a
continuum of strikes K and maturities T . We have,
and
C (K, T ) = er(T t) E (ST K)+ ,
(10A.3)
C (K, T ) = er(T t) E (IST K ) .

K
(10A.4)
For xed K,

1
2 2
T,
dT (ST K) = IST K rST + (ST K) T ST dT + IST K T ST dW
2
+
where is the Diracs delta. Hence, by the decomposition (ST K)+ + KIST K = ST IST K ,
1

dE (ST K)+
= r E (ST K)+ + KE (IST K ) + E (ST K) 2T ST2 .
dT
2
By multiplying throughout by er(T t) , and using (10A.3)-(10A.4),

+

C (K, T )
1
r(T t) dE (ST K)
e
= r C (K, T ) K
+ er(T t) E (ST K) 2T ST2 .
dT
K
2
(10A.5)
We have,

E (ST K) 2T ST2 =
=
""
"
(ST K) 2T ST2 T ( T | ST ) T (ST ) dST d T

2T
"
joint density of ( T ,ST )
(ST K) ST2 T
= K 2 T (K)
"
(ST ) T ( T | ST ) dST dT
2T T ( T | ST = K) dT

K 2 T (K) E 2T ST = K .
By replacing this result into Eq. (10A.5), and using the famous relation
2 C (K, T )
= er(T t) T (K)
K 2
(10A.6)
(which easily follows by dierentiating once again Eq. (10A.4)), we obtain

C (K, T )
1
dE (ST K)+
2 C (K, T ) 2
= r C (K, T ) K
+ K2
er(T t)
E T ST = K . (10A.7)
2
dT
K
2
K
272
c
by
A. Mele
We also have,
E (ST K)+
C (K, T ) = rC (K, T ) + er(T t)
.
T
T
Therefore, by replacing the previous equality into Eq. (10A.7), and by rearranging terms,
C (K, T ) 1 2 2 C (K, T ) 2
C (K, T ) = rK
+ K
E T ST = K .
2
T
K
2
K
This is,

E 2T ST = K = 2
C (K, T )
C (K, T )
+ rK
T
K
loc (K, T )2 .
2 C (K, T )
K2
K 2
(10A.8)
As an example, let t (St , t) vt , where vt is solution to the 2nd equation in (10.22). Then,

loc (K, T )2 = E 2T ST = K

= E (ST , T )2 vT2 ST = K

= (K, T )2 E vT2 ST = K

loc (K, T )2 E vT2 ST = K ,
which proves Eq. (10.23).

Next, we prove Eq. (10.24). We have,
"

E 2T =
E 2T ST = K T (K) dK
0
"
C(K,T ) + rK C(K,T )
T
K
2
T (K) dK
2 C(K,T )
2
0
K
K 2
" C(K,T )
)
+ rK C(K,T
r(T t)
T
K
dK,
2e
K2
0
where the 2nd line follows by Eq. (10A.8), and the third line follows by Eq. (10A.6). This proves Eq.
(10.24).
Proof of Eq. (10.25). If r = 0, Eq. (10.24) collapses to,

E 2T = 2
Then, we have,
E (IVT1 ,T2 ) =
"
T2
T1

E 2u du = 2
"
1
K2
"
T2
T1
"
C(K,T )
T
dK.
K2

"
C (K, u)
C (K, T2 ) C (K, T1 )
du dK = 2
dK.
T
K2
0
Proof of Eq. (10.26). By the standard Taylor expansion with remainder, we have that for any
function f smooth enough,
" x
f (x) = f (x0 ) + f (x0 ) (x x0 ) +
(x t) f (t) dt.
273
x0
c
by
A. Mele

By applying this formula to ln FT ,
ln FT
1
= ln Ft +
(FT Ft )
Ft
1
= ln Ft +
(FT Ft )
Ft
"
FT
(FT t)
Ft
Ft
"
1
(K FT )
dK
K2
+
"
Ft
(FT K)+
1
dK
K2
"
1
+ 1
(K ST )
dK
(ST K)+ 2 dK,
2
K
K
0
Ft
#x
#x
#
where the second equality follows because x0 (x t) t12 dt = 0 0 (t x)+ t12 dt + x0 (x t)+ t12 dt, and
the third equality follows because the forward price at T satises FT = ST . Hence, by E (FT ) = Ft ,

" Ft

"
FT
Pt (K, T )
Ct (K, T )
r(T t)
E ln
=e
dK +
dK .
(10A.9)
Ft
K2
K2
0
Ft
1
= ln Ft +
(FT Ft )
Ft
"
1
dt
t2
Ft
On the other hand, by Itos lemma,

E
"
2u du

FT
= 2E ln
.
Ft
(10A.10)
By replacing this formula into Eq. (10A.9) yields Eq. (10.26).

Remark A1. The previous proof holds in the case of a constant instantaneous interest rate, r. If
the instantaneous interest rate is stochastic, the formula would be dierent, for

" T

1
+
rs ds (ST K) = P (t, T ) EQT (ST K)+ ,
Ct (K, T ) = P (t, T ) E P (t, T ) exp
F
where QTF is the T -forward measure to be introduced in the next chapter.

Remark A2. Eqs. (10A.9) and (10A.10) reveal that variance swaps can be hedged!
+
dC(K,T )
= dE(STdTK) , then,
dT
2
)
K 2 C(K,T
. The converse is also true.
K 2
Remark A3. Set for simplicity r = 0. The previous proofs 8

show that if
)
volatility must be restricted in a way to make 2 = 2 C(K,T
T
By Fokker-Planck,
1 2 2 2
x = , t, x forward.
2
2 x
t
If we ignore ill-posedness issues related to Eq. (10A.6), such as those8dealt with in Tikhonov and
2
2
)
)
Arsenin (1977), then, we have that = xC2 . Replacing 2 = 2 C(x,T
x2 C(x,T
into the FokkerT
x2
Planck equation,
$ C(x,T ) %
2
T
= ,
)
x2 2 C(x,T
t
2
x
which works for =
2C
x2
274
c
by
A. Mele
References
Ball, C.A. and A. Roma (1994): Stochastic Volatility Option Pricing. Journal of Financial
and Quantitative Analysis 29, 589-607.
Bollerslev, T. (1986): Generalized Autoregressive Conditional Heteroskedasticity. Journal of
Econometrics 31, 307-327.
Bollerslev, T., Engle, R. and D. Nelson (1994): ARCH Models. In: McFadden, D. and R.
Engle (Editors): Handbook of Econometrics (Volume 4), 2959-3038. Amsterdam, NorthHolland
Clark, P. K. (1973): A Subordinated Stochastic Process Model with Fixed Variance for Speculative Prices. Econometrica 41, 135-156.
Corradi, V. (2000): Reconsidering the Continuous Time Limit of the GARCH(1,1) Process.
El Karoui, N., M. Jeanblanc-Picque and S. Shreve (1998): Robustness of the Black and
Scholes Formula. Mathematical Finance 8, 93-126.
Engle, R.F. (1982): Autoregressive Conditional Heteroskedasticity with Estimates of the Variance of United Kingdom Ination. Econometrica 50, 987-1008.
Fama, E. (1965): The Behaviour of Stock Market Prices. Journal of Business 38, 34-105.
Heston, S.L. (1993a): Invisible Parameters in Option Prices. Journal of Finance 48, 933-947.
Heston, S.L. (1993b): A Closed Form Solution for Options with Stochastic Volatility with
Application to Bond and Currency Options. Review of Financial Studies 6, 327-344.
Hull, J. and A. White (1987): The Pricing of Options with Stochastic Volatilities. Journal
of Finance 42, 281-300.
Mandelbrot, B. (1963): The Variation of Certain Speculative Prices. Journal of Business
36, 394-419.
Mele, A. (1998): Dynamiques non lineaires, volatilite et equilibre. Paris: Editions Economica.
Mele, A. and F. Fornari (2000): Stochastic Volatility in Financial Markets. Crossing the Bridge
to Continuous Time. Boston: Kluwer Academic Publishers.
Merton, R. (1973): Theory of Rational Option Pricing. Bell Journal of Economics and
Management Science 4, 637-654.
Nelson, D.B. (1990): ARCH Models as Diusion Approximations. Journal of Econometrics
45, 7-38.
275
c
by
A. Mele
Renault, E. (1997): Econometric Models of Option Pricing Errors. In: Kreps, D., Wallis, K.
(Editors): Advances in Economics and Econometrics (Volume 3), 223-278. Cambridge:
Cambridge University Press.
Scott, L. (1987): Option Pricing when the Variance Changes Randomly: Theory, Estimation,
and an Application. Journal of Financial and Quantitative Analysis 22, 419-438.
Tauchen, G. and M. Pitts (1983): The Price Variability-Volume Relationship on Speculative
Markets. Econometrica 51, 485-505.
Taylor, S. (1986): Modeling Financial Time Series. Chichester, UK: Wiley.
Tikhonov, A. N. and V. Y. Arsenin (1977): Solutions to Ill-Posed Problems. Wiley, New York.
Wiggins, J. (1987): Option Values and Stochastic Volatility: Theory and Empirical Estimates. Journal of Financial Economics 19, 351-372.
276
11
Interest rates
11.1 Prices and interest rates

11.1.1
Introduction
A pure-discount, or zero-coupon, bond is a contract that guarantees one unit of numeraire at

some maturity date. Apart from isolated exceptions, we only consider pure-discount bonds,
which we will then simply call bonds, to simplify the exposition. Also, with the exception of
Section 11.3.7, we assume no default risk. Default risk is, instead, more systematically dealt
with in the next chapter.
Let [t, T ] be a xed time interval, and for each [t, T ], let P ( , T ) be the price as of
time of a bond maturing at T > t. The information in this chapter is Brownian, except
for the jump-diusion models in Section 11.3.4. The price of a bond in this chapter, then, is
driven by some multidimensional diusion process {y ( )} t , which we emphasize by writing
P (y ( ) , , T ) P ( , T ). As an example, y can be a scalar diusion, and r = y can be the
short-term rate. In this particular example, bond prices are driven by short-term rate movements
through the bond pricing function P (r, , T ). The exact functional form of the pricing function
is determined by (i) the assumptions made as regards the short-term rate dynamics and (ii)
the Fundamental Theorem of Asset Pricing (henceforth, FTAP). The bond pricing function in
the general multidimensional case is obtained following the same route. Models of this kind are
presented in Section 11.3.
A second class of models is that in which bond prices can not be expressed as a function
of any state variable. Rather, current bond prices are taken as primitives, and forward rates
(i.e., interest rates prevailing today for borrowing in the future) are multidimensional diusion
processes. There is a relation linking bond prices to forward rates. The FTAP restricts the
dynamic behavior of future bond prices and forward rates. Models belonging to this second
class are analyzed in Section 11.4.
The aim of this chapter is to develop the simplest foundations of the previously described
two approaches to interest rate modeling. In the next section, we provide denitions of interest
rates and markets. Section 11.2.3 develops the two basic representations of bond prices: one in
terms of the short-term rate; and the other in terms of forward rates. Section 11.2.4 develops
11.1. Prices and interest rates
c
by
A. Mele
the foundations of the so-called forward martingale probability, which is a probability measure
under which forward interest rates are martingales. It is an important tool of analysis. [ ... ]
11.1.2
Markets and interest rate conventions
11.1.2.1 Markets for interest rates
There are three main types of markets for interest rates: (i) LIBOR; (ii) Treasure rate; (iii)
Repo rate (or repurchase agreement rate).
LIBOR (London Interbank Offer Rate) and other interbank rates
Many large nancial institutions trade with each other deposits for maturities ranging from
just overnight to one year at a given currency. The LIBOR is the rate at which nancial
institutions are willing to lend, on average. It is an average indicative quote of the interbank
lending market. It is calculated by Thomson Reuters for ten currencies, and published daily by
the British Bankers Association. Instead, the LIBID (London Interbank Bid Rate) is the rate
that these nancial institutions are prepared to pay to borrow money, on average. Normally,
LIBID < LIBOR. The LIBOR is a fundamental point of reference to nancial institutions,
who look at it as an opportunity cost of capital. Moreover, many xed income instruments
are indexed to the LIBOR: forward rate agreements, interest rate swaps, or variable mortgage
rates.
The LIBOR is distinct from the US Federal Funds rate. Banks have to maintain reserves
with the Federal Reserve to back deposits and to clear nancial transactions. Transactions
involve banks with excess reserves with the Fed, which earn no interest, to banks with reserve
deciencies. The Federal Funds rate is the overnight rate at which banks lend these reserves
to each other. The Federal Funds rate is aected by the FDRBNY, which aims to make it lie
within a range of the target rate decided by the governors at Federal Open Market Committee
meetings. This range is maintained through open market operations.
Treasury rate
This is the rate at which a given Government can borrow at a given currency.
Repo rate (or repurchase agreement rate)
A Repo agreement is a contract by which one counterparty sells some assets to the other one,
with the obligation to buy these assets back at some future date. The assets act as collateral.
The rate at which such a transaction is made is the repo rate. One day repo agreements give
rise to overnight repos. Longer-term agreements give rise to term repos.
Spreads
Interest rate spreads isolate interesting pieces of information, as they remove common components of the interest rates generating the spreads, which we are not interested in. An important
example is the overnight interest swap rate (OIS), which is the swap rate in a swap agreement
of xed against variable interest rate payments, where the variable interest rate payments are
made of an overnight reference, typically an average, unsecured interbank overnight rate, such
as the Federal Funds rate in the US, SONIA in the UK or EONIA in the Euro area. (See Section
11.7.5 for denitions of swaps and swap rates.) An interesting indicator, then, is the 3-month
LIBOR 3-month OIS spread, also known as the LIBOR-OIS spread. Because payments
relating to overnight rates are not subject to default risk, and the overnight rate is anchored
to monetary policy, the LIBOR-OIS spread is capable of isolating credit views about nancial
278
c
by
A. Mele
institutions. It is generally at, although then it reached high record levels during the 2007
subprime crisis (see Figure 11.1). Instead, the so-called TED (Treasury bill rate Eurodollar
LIBOR) spread, also captures ight to quality eects occurring during times of crisis, when
Treasury bonds are considered particularly valuable. For this ight to quality reason, the
TED spread might fail isolate views about developments in the interbank market.
FIGURE 11.1. Antonio Mele does not claim any copyright on this picture, which is taken from Brunnermeier (2009). The picture has been put here for illustrative purposes only, and permission to the
author shall be duly asked before the book will be published.
On a historical note, the Federal Funds rate has been the object of much empirical research.
In an attempt to explain how the credit view contributes to growth more than Friedmans
monetary view, Bernanke and Blinder (1992) show that the Federal Funds rate makes the
predicting power of M1 growth insignicant. This nding initially spread enthusiasm about the
ability of this rate to explain short-run aggregate uctuations. However, as surveyed for example
by Stock and Watson (2003), the explanatory power of the Federal Funds rate evaporizes, once
we condition on the term spread.
11.1.2.2 Mathematical denitions of interest rates
The simplest denition is that of simply compounded interest rates. A simply-compounded

interest rate at time , for the time interval [ , T ], is dened as the solution L to the following
equation:
1
.
P ( , T ) =
1 + (T )L( , T )
279
c
by
A. Mele
This denition is intuitive, and is the most widely used in the market practice. As an example,
LIBOR rates are computed in this way. In this case, P ( , T ) is generally interpreted as the initial
amount of money to invest at time to obtain $ 1 at time T .
Given L( , T ), the short-term rate process r is obtained as:
r ( ) lim L ( , T ) .
T
We now introduce an important piece of notation that will be used in Section 11.7. Given a
non decreasing sequence of dates {Ti }i=0,1, , we dene:
L(Ti ) L(Ti , Ti+1 ).
(11.1)
In other terms, L(Ti ) is solution to:

P (Ti , Ti+1 ) =
11.1.3
1
,
1 + i L(Ti )
i Ti+1 Ti , i = 0, 1, .
(11.2)
The yield curve and forward rates
11.1.3.1 The yield curve
The yield-to-maturity is dened to be the function R (t, T ) such that:

P (t, T ) e(T t)R(t,T ) .
(11.3)
Its a sort of average rate for investing from time t to time T > t. The function, T R (t, T ),
is called the yield curve, or the term structure of interest rates.
A related, and widely used concept, is the the par yield curve. Let B (t, T ) be the current
price of a coupon bearing bond. This bond pays o the principal of $1 at expiry T , as well as a
known sequence of coupons C (t, T ) at t + 1, t + 2, ..., T , such that, in the absence of arbitrage
or any other frictions, its price is:
B (t, T ) = C (t, T )
T t
P (t, t + i) + P (t, T ) .
i=1
Please note, C (t, T ) is xed at time t. A par bond, then, is one such that B (t, T ) = 100%, and
the par yield curve is the resulting sequence of the coupon rates C (t, T ), for T varying, viz
B (t, T ) P (t, T )
C (t, T ) = T t
,
i=1 P (t, t + i)
B (t, T ) = 1.
(11.4)
In other words, the coupon rates C (t, T ) have to adjust to make the market happy to have
the coupon bearing bond quote at par, B (t, T ) = 1.
11.1.3.2 A rst representation of bond prices
Let Q be a risk-neutral probability probability. Let E [] denote the expectation operator taken
under Q. By the FTAP, there are no arbitrage opportunities if and only if P ( , T ) satises:
!
T
P ( , T ) = E e r()d , all [t, T ].
(11.5)
A sketch of the if-part (there is no arbitrage if bond prices are as in Eq. (11.5)) is provided in
Appendix 1. The proof is standard and in fact, similar to those oered in the rst part of the
book, but it is developed again here, as it highlights some issues specic to the term-structure.
280
c
by
A. Mele

11.1.3.3 Forward rates, and a second representation of bond prices
In a forward rate agreement (FRA, henceforth), two counterparties agree that the interest
rate on a given principal in a future time-interval [T, S] will be xed at some level K. Let
the principal be normalized to one. The FRA works as follows: at time T , the rst counterparty receives $1 from the second counterparty; at time S > T , the rst counterparty pays
back $ [1 + 1 (S T ) K] to the second counterparty. The amount K is agreed upon at time
t. Therefore, the FRA makes it possible to lock-in future interest rates. We consider simply
compounded interest rates because this is the standard market practice.
The amount K for which the current value of the FRA is zero is called the simply-compounded
forward rate as of time t for the time-interval [T, S], and is usually denoted as F (t, T, S). A
simple argument can be used to express F (t, T, S) in terms of bond prices. Consider the following
portfolio implemented at time t. Long one bond maturing at T and short P (t, T )/ P (t, S) bonds
maturing at S, for the time period [t, S]. The initial cost of this portfolio is zero because,
P (t, T ) +
P (t, T )
P (t, S) = 0.
P (t, S)
At time T , the portfolio yields $1 (originated from the bond purchased at time t). At time S,
P (t, T )/ P (t, S) bonds maturing at S (that were shorted at t) must be purchased. But at time
S, the cost of purchasing P (t, T )/ P (t, S) bonds maturing at S is obviously $ P (t, T )/ P (t, S).
The portfolio, therefore, is acting as a FRA: it pays $1 at time T , and $ P (t, T )/ P (t, S) at
time S. In addition, the portfolio costs nothing at time t. Therefore, the interest rate implicitly
paid in the time-interval [T, S] must be equal to the forward rate F (t, T, S), and we have:
P (t, T )
= 1 + (S T )F (t, T, S).
P (t, S)
(11.6)
Clearly,
L(T, S) = F (T, T, S).
Next, we derive the value of the FRA in the general case in which K = F (t, T, S). Consider
the following trade. At time t, enter a FRA for the time-interval [T, S] as a future borrower.
Come time T , honour the FRA by borrowing $1 for the time-interval [T, S] at a cost of K.
Then, lend this very same $1 at the random interest rate L(T, S). The time S payo deriving
from this trade is:
(S T ) [L (T, S) K] .
The value of the FRA, which we denote as IRS(t, T, S; K), is the current market value of this
future, random payo. By the FTAP,
!
S
IRS (t, T, S; K) = E e t r( )d (S T ) [L (T, S) K]
!
S
= (S T ) E e t r( )d L (T, S) (S T ) P (t, S) K
- S
.
e t r( )d
=E
[1 + (S T ) K] P (t, S)
P (T, S)
= P (t, T ) [1 + (S T ) K] P (t, S) ,
281
(11.7)
c
by
A. Mele
where the second line holds by the denition of L and the third line follows by the following
relation:1
- S
.
e t r()d
P (t, T ) = E
.
(11.8)
P (T, S)
Finally, by replacing Eq. (11.6) into Eq. (11.7),
IRS (t, T, S; K) = (S T ) [F (t, T, S) K] P (t, S) .
(11.9)
As is clear, IRS can take on any sign, and is exactly zero when K = F (t, T, S), where
F (t, T, S) solves Eq. (11.6).
A useful remark. Comparing the second line in Eq. (11.7) with Eq. (11.9) reveals that:
.
- S
e t r( )d
L (T, S) .
F (t, T, S) = E
P (t, S)
That is, forward rates are not unbiased expectations of future interest rates, not even under
the risk-neutral probability. We shall return to this point in Section 11.1.4.2.
Bond prices can be expressed in terms of these forward interest rates, namely in terms of the
instantaneous forward rates. First, rearrange terms in Eq. (11.6) so as to obtain:
F (t, T, S) =
P (t, S) P (t, T )
.
(S T )P (t, S)
The instantaneous forward rate f (t, T ) is dened as

f(t, T ) lim F (t, T, S) =
ST
ln P (t, T )
.
T
(11.10)
It can be interpreted as the marginal rate of return from committing a bond investment for an
additional instant. To express bond prices in terms of f, integrate Eq. (11.10)
f (t, ) =
ln P (t, )
with respect to maturity date , use the condition that P (t, t) = 1, and obtain:
P (t, T ) = e
T
t
f (t,)d
(11.11)
f(t, )d .
(11.12)
11.1.3.4 More on the marginal revenue nature of forward rates
Comparing Eq. (13.1) with Eq. (11.11) yields:

1
R(t, T ) =
T t
"
T
t
1 To show that Eq. (11.8) holds, suppose that at time t, $P (t, T ) are invested in a bond maturing at time T . At time T , this
investment will obviously pay o $1. And at time T , $1 can be further rolled over another bond maturing at time S, thus yielding
$ 1/ P (T, S) at time S. Therefore, it is always possible to invest $P (t, T ) at time t and obtain a payo of $ 1/ P (T, S) at time
S. By the FTAP, there are no arbitrage opportunities if and only if Eq. (11.8) holds true. Alternatively, use the law of iterated
expectations to obtain

S

T
S
e t r()d
e t r( )d e T r( )d
E
=E E
= P (t, T ).
F(T )

P (T, S)
P (T, S)
282
c
by
A. Mele

By dierentiating Eq. (11.12) with respect to T yields:
R(t, T )
1
=
[f (t, T ) R(t, T )] .
T
T t
This relation underscores very clearly the marginal revenue nature of forward rates. Similarly
as for the cost function in the basic theory of production, we have that: (i) If f (t, T ) < R(t, T ),
the yield-curve R(t, T ) is decreasing at T ; (ii) if f(t, T ) = R(t, T ), the yield-curve R(t, T ) is
stationary at T ; (iii) if f(t, T ) > R(t, T ), the yield-curve R(t, T ) is increasing at T .
11.1.3.5 The expectation theory
The expectation theory holds that forward rates equal expected future short-term rates, or
f (t, T ) = E [r (T )] ,
where E() denotes expectation under the physical probability. So by Eq. (11.12), the expectation theory implies that,
" T
1
R (t, T ) =
E [r ( )] d .
T t t
The question whether f (t, T ) is higher than E [r (T )] is very old. The oldest intuition we
have is that only risk-adverse investors may induce f(t, T ) to be higher than the short-term
rate they expect to prevail at T , viz,
f (t, T ) E(r(T )).
In other terms, (11.13) never holds true if all investors are risk-neutral.
(11.13)
(11.14)
The inequality (11.13) is related to the Hicks-Keynesian normal backwardation hypothesis.2

According to Hicks, rms tend to demand long-term funds while fund suppliers prefer to lend
at shorter maturity dates. The market is cleared by intermediaries who demand a liquidity
premium to be compensated for their risky activity consisting in borrowing at short maturity
dates and lending at long maturity dates. As we will see in a moment, we do not really need a
liquidity risk premium to explain (11.13). Pure risk-aversion can be sucient. In other terms,
a proof of statement (11.14) can be sucient. Here is a proof. By Jensens inequality,
!

T
tT f (t, )d
tT r( )d
e
P (t, T ) = E e
e t E[r( )]d .
By taking logs,
"
E[r( )]d
"
f (t, )d .
This shows statement (11.14).

Another simple prediction on yield-curve shapes can produced as
By the same rea! follows.

(T t)R(t,T )
tT r( )d
tT E[r( )]d
soning used to show (11.14), e
P (t, T ) = E e
e
. Therefore,
1
R(t, T )
T t
"
E[r( )]d .
t
2 According to the normal backwardation (contango) hypothesis, forward prices are lower (higher) than future expected spot
prices. Here the normal backwardation hypothesis is formulated with respect to interest rates.
283
c
by
A. Mele
As an example, suppose that the short-term rate is a martingale under the risk-neutral probability, viz. E[r( )] = r(t). The previous relation then collapses to:
R(t, T ) r(t),
which means that the yield curve is not increasing in T . Increasing yield-curves then arise because the short-term rate is not a martingale under the risk-neutral probability, which happens
because of two fundamental, and not necessarily mutually exclusive reasons: (i) interest rates
are expected to increase, (ii) investors are risk-averse.
Finally, a recurrent denition. The dierence
1
R(t, T )
T t
"
E [r( )] d
is usually referred to as yield term-premium.

What does the empirical evidence suggest about the expectation hypothesis? Denote the
)
T
continuously compounded returns on a zero expiring at some date T as rt+1
= ln PP(t+1,T
, and
(t,T )
T
T
the excess returns as rt+1
= rt+1
R (t, t + 1), where R (t, t + 1) = ln P (t, t + 1). Finally, the
P
(t,T
)
T
forward rate is: ft = ln P (t,T 1) . It is easy to see that the log-excess returns can be expressed
as:
T
rt+1
= [R (t, t + 1) R (t, T )] (T t 1) + [R (t, T ) R (t, t + 1)] ,
such that,
T
1
1
Et rt+1
+
[R (t, T ) R (t, t + 1)] .
T t1
T t1
T
The expectation hypothesis implies that the risk-premium Et rt+1
= 0, and to test for it,
we the following regression can be run:
R (t, t + 1) R (t, T ) =
R (t, t + 1) R (t, T ) = T + T
1
[R (t, T ) R (t, t + 1)] + Residualt ,
T t1
and test for the null of T = 0 and T = 1. A widely known empirical feature of US data is that
the estimates of are typically negative for all maturities T , and somewhat increasing
T with T
in absolute value. In fact, Fama and Bliss (1987) show that the risk-premium Et rt+1 relates
to the forward spreads, dened as ftT R (t, t + 1), in that regressing

T
rt+1
= T + T ftT R (t, t + 1) + Residualt ,
delivers statistically signicant and positive values of T for many maturities T .
Cochrane and Piazzesi (2005) go one step further and consider the following regressions:
T
rt+1
= T + 1T R (t, t + 1) +
5
j,T ftj + Residualt ,
j=2

5
and document a tent shape for the estimates of the coecients j,T j=1 , for bond maturities
T {1, , 5}, and where t is in years so as to make returns calculated on a yearly basis. They
284
c
by
A. Mele
document this tent shape is robust to estimating a factor model in that this shape persists in
the estimates of the coecients (bj,T )5j=1 in:
T
rt+1
= T + 1T Zt + Residualt , Zt = b1T R (t, t + 1) +
5
bj,T ftj ,
j=2
where Zt is the common factor among the bond maturities T {1, , 5}. Moreover, they
argue that using the traditional factors known to explain movements in the yield curve (see
Section 11.2) does not destroy the predicting power of their factors, in sample.
11.1.4
Forward martingale probabilities
11.1.4.1 Denition
Let (t, T ) be the T -forward price of a claim S(T ) at T . That is, (t, T ) is the price agreed
at t, that will be paid at T for delivery of the claim at T . Nothing has to be paid at t. By the
FTAP, there are no arbitrage opportunities if and only if:
!

tT r(u)du
0=E e
(S(T ) (t, T )) .
But since (t, T ) is known at time t,
E e
T
t
r(u)du
!
!
T
S(T ) = (t, T ) E e t r(u)du .
Now use the bond pricing equation (11.5), and rearrange terms in the previous equality, to
obtain
- T
.
e t r(u)du
(t, T ) = E
(11.15)
S(T ) = E [T (T ) S(T )] ,
P (t, T )
where3
T
e t r(u)du
.
T (T )
P (t, T )
Eq. (11.15) suggests that we can dene a new probability QTF , as follows,
T
dQTF
e t r(u)du
!.
T (T ) =
T
dQ
E e t r(u)du
(11.16)
Naturally, E[ T (T )] = 1. Moreover, if the short-term rate process is deterministic, T (T ) equals

one and Q and QTF are the same.
In terms of this new probability QTF , the forward price (t, T ) is:
"
"
(t, T ) = E [ T (T ) S(T )] = [T (T ) S(T )] dQ = S(T )dQTF = EQTF [S(T )] ,
(11.17)
3 As
an example, suppose that S is the price process of a traded asset. By the FTAP, there are no arbitrage opportunities if and

T
only if e t r(u)du S( ) is a Q-martingale. In this case, E[e t r(u)du S(T )] = S(t), and Eq. (11.15) collapses to the well-known
formula: (t, T )P (t, T ) = S(t). As is also well-known, entering the forward contract established at t at a later date > t costs.
Apply the FTAP to prove that the value of a forward contract as of time [t, T ] is given by P ( , T ) [( , T ) (t, T )]. [Hint:
Notice that the nal payo is S(T ) (t, T ) and that the discount has to be made at time .]
285
c
by
A. Mele
where EQTF [] denotes the expectation taken under QTF . For reasons that will be clear in a
moment, QTF is referred to as the T -forward martingale probability. The forward martingale
probability is a practical tool to price interest-rate derivatives, as we shall explain in Section
11.7. It was introduced by Geman (1989) and Jamshidian (1989), and further analyzed by
Geman, El Karoui and Rochet (1995). The appendix provides additional details: Appendix
2 relates forward prices to their certainty equivalent, and Appendix 3 illustrates additional
technicalities about the forward martingale probability.
11.1.4.2 Martingale properties
Forward prices
Clearly, (T, T ) = S(T ). Therefore, (11.17) becomes:

(t, T ) = EQTF [(T, T )] .
Forward rates
Forward rates exhibit an analogous property:

f (t, T ) = EQTF [r(T )] = EQTF [f(T, T )] .
(11.18)
where the last equality holds as r(t) = f (t, t). The proof is also simple. We have,
ln P (t, T )
T
9
P (t, T )
=
P (t, T )
T
- T
.
e t r( )d
= E
r(T )
P (t, T )
f(t, T ) =
= E [ T (T ) r(T )]
= EQTF [r(T )] .
Finally, the same result is also valid for the simply-compounded forward rate:
Fi ( ) = EQTi+1 [L(Ti )] = EQTi+1 [Fi (Ti )] , [t, Ti ]
F
where the second equality follows from Eq. (11.68). To show the previous relation, note that
by denition, the simply-compounded forward rate F (t, T, S) satises:
IRS(t, T, S; F (t, T, S)) = 0,
where IRS(t, T, S; K) is the value as of time t of a FRA struck at K for the time-interval [T, S].
By rearranging terms in the rst line of Eq. (11.7),
!

tS r( )d
F (t, T, S)P (t, S) = E e
L(T, S) .
By the denition of S (S),
F (t, T, S) = EQSF [L(T, S)] .

286
c
by
A. Mele
11.2. Common factors aecting the yield curve
Now use the denitions of L(Ti ) and Fi ( ) in Eq. (11.1) and Eq. (11.66) to conclude.
These relations show that it is only under the forward martingale probability that the expectation theory holds true. Consider, for instance, Eq. (13.19). We have,
f(t, T ) = EQTF [r(T )] = EQ [ T (T ) r(T )]
=
E [ T (T )]E [r(T )] + covQ [T (T ) , r(T )]

=1
= E [r(T )] + cov [Ker (T ) , r(T )] + covQ [ T (T ) , r(T )] ,

where Ker(T ) denotes the pricing kernel in the economy. That is, forward rates in general
deviate from the future expected spot rates because of risk-aversion corrections (the second
term in the last equality) and because interest rates are stochastic (the third term in the last
equality).
11.2 Common factors aecting the yield curve

Which systematic risks aect the entire term-structure of interest rates? How many factors are
needed to explain the variation of the yield curve? The standard duration hedging practice,
reviewed in detail in Chapter 13, relies on the idea that most of the variation of the yield curve
is successfully captured by a single factor that produces parallel shifts in the yield curve. How
reliable is this idea, in practice?
Litterman and Scheinkman (1991) demontrate that most of the variation (more than 95%)
of the term-structure of interest rates can be attributed to the variation of three unobservable
factors, which they label (i) a level factor, (ii) a steepness (or slope) factor, and (iii)
a curvature factor. To disentangle these three factors, the authors make an unconditional
analysis based on a fixed-factor model. Succinctly, this methodology can be described as follows.
Suppose that p returns computed from bond prices at p dierent maturities are generated by
a linear factor structure, with a xed number k of factors,
Rt
p1
+ B Ft + t ,
= R
p1
pk k1
p1
(11.19)
where Rt is the vector of returns, Ft is the zero-mean vector of common factors aecting the
is the vector of unconditional expected returns, t is a vector
returns, assumed to be zero mean, R
of idiosyncratic components of the return generating process, and B is a matrix containing the
factor loadings. Each row of B contains the factor loadings for all the common factors aecting
a given return, i.e. the sensitivities of a given return with respect to a change of the factors.
Each comumn of B contains the term-structure of factor loadings, i.e. how a change of a given
factor aects the term-structure of excess returns.
11.2.1
Methodological details
Estimating the model in Eq. (11.19) leads to econometric challenges, mainly because the vector of factors Ft is unobservable.4 However, there exists a simple method, known as principal

, where = BB +.
that in Eq. (11.19), F N (0, I), and that N (0, ), where is diagonal. Then, R N R,
The assumptions that F N (0, I) and that is diagonal are necessary to identify the model, but not sucient. Indeed, any
orthogonal rotation of the factors yields a new set of factors which also satises Eq. (11.19). Precisely, let T be an orthonormal
4 Suppose
287
c
by
A. Mele
components analysis (PCA, henceforth), which leads to empirical results qualitatively similar
to those that hold for the general model in Eq. (11.19). We discuss these empirical results in
the next subsection. We now describe the main methodological issues arising within PCA.
The main idea underlying PCA is to transform the original p correlated variables R into a set
of new uncorrelated variables, the principal components. These principal components are linear
combinations of the original variables, and are arranged in order of decreased importance: the
rst principal component accounts for as much as possible of the variation in the original data,
etc. Mathematically, we are looking for p linear combinations of the demeaned excess returns,

,
Yi = Ci R R
i = 1, , p,
(11.20)
such that, for p vectors Ci of dimension 1 p, (i) the new variables Yi are uncorrelated, and (ii)
their variances are arranged in decreasing order. The logic behind PCA is to ascertain whether
a few components of Y = [Y1 Yp ] account for the bulk of variability of the original data.
Let C = [C1 Cp ] be a p p matrix such that we can write Eq. (11.20) in matrix format,
or, by inverting,
Yt = C Rt R
= C 1 Yt .
Rt R
(11.21)
Next, suppose that the vector Y (k) = [Y1 Yk ] accounts for most of the variability in the
original data,5 and let C (k) denote a p k matrix extracted from the matrix C 1 through
the rst k rows of C 1 . Since the components of Y (k) are uncorrelated and they are deemed
largely responsible for the variability of the original data, it is natural to disregard the last
p k components of Y in Eq. (11.21),
C (k) Yt(k) .
Rt R
pk
p1
k1
(k)
If the vector Yt really accounts for most of the movements of Rt , the previous approximation
to Eq. (11.21) should be fairly good.
Let us make more precise what the concept of variability is in the context of PCA. Suppose
that the variance-covariance matrix of the returns, , has p distinct eigenvalues, ordered from
the highest to the lowest, as follows: 1 > > p . Then, the vector Ci in Eq. (11.20) is the
eigenvector corresponding to the i-th eigenvalue. Moreover,
var (Yi ) = i ,
Finally, we have that
RPCA
i = 1, , p.
k
k
var (Yi )
i
i=1
= p
= i=1
.
p
i=1 var (Ri )
i=1 i
(11.22)
matrix. Then, (BT ) (BT ) = BT T B = BB . Hence, the factor loadings B and BT have the same ability to generate the matrix
. To obtain a unique solution, one needs to impose extra constraints
on B. For example,
J
oreskog (1967) develop a maximum

likelihood approach in which the log-likelihood function is, 12 N log || + Tr S1 , where S is the sample covariance matrix of
R, and the constraint is that B B be diagonal with elements arranged in descending order. The algorithm is: (i) for a given ,
maximize the log-likelihood with respect to B, under the constraint that B B be diagonal with elements arranged in descending
(ii) given B,
maximize the log-likelihood with respect to , thereby obtaining ,
which is fed back into
order, thereby obtaining B;
step (i), etc. Knez, Litterman and Scheinkman (1994) describe this approach in their paper. Note that the identication device they
describe at p. 1869 (Step 3) roughly corresponds to the requirement that B B be diagonal with elements arranged in descending
order. Such a constraint is clearly related to principal component analysis.
5 There are no rigorous criteria to say what most of the variability means in this context. Instead, a likelihood-ratio test is
most informative in the context of the estimation of Eq. (11.19) by means of the methods explained in the previous footnote.
288
c
by
A. Mele
(Appendix 4 provides technical details and proofs of the previous formulae.) It is in the sense
of Eq. (11.22) that in the context of PCA, we say that the rst k principal components account
for RPCA % of the total variation of the data.
11.2.2
The empirical facts
The striking feature of the empirical results uncovered by Litterman and Scheinkman (1991)
is that they have been conrmed to hold across a number of countries and sample periods.
Moreover, the economic nature of these results is the same, independently of whether the
statistical analysis relies on a rigorous factor analysis of the model in Eq. (11.19), or a more
back-of-envelope computation based on PCA. Finally, the empirical results that hold for bond
returns are qualitatively similar to those that hold for bond yields.
Level
Slope
Curvature
FIGURE 11.2. Changes in the term-structure of interest rates generated by changes in the level,
slope and curvature factors.
Figure 11.2 visualizes the eects that the three factors have on the movements of the termstructure of interest rates.
The rst factor is called a level factor as its changes lead to parallel shifts in the termstructure of interest rates. Thus, this level factor produces essentially the same eects
on the term-structure as those underlying the duration hedging portfolio practice. This
factor explains approximately 80% of the total variation of the yield curve.
The second factor is called a steepness factor as its variations induce changes in the
slope of the term-structure of interest rates. After a shock in this steepness factor, the
short-end and the long-end of the yield curve move in opposite directions. The movements
of this factor explain approximately 15% of the total variation of the yield curve.
The third factor is called a curvature factor as its changes lead to changes in the
curvature of the yield curve. That is, following a shock in the curvature factor, the middle
of the yield curve and both the short-end and the long-end of the yield curve move in
opposite directions. This curvature factor accounts for approximately 5% of the total
variation of the yield curve.
289
c
by
A. Mele
11.3. Models of the short-term rate
Understanding the origins of these three factors is still a challenge to nancial economists and
macroeconomists. For example, macroeconomists explain that central banks aect the shortend of the yield curve, e.g. by inducing variations in Federal Funds rate in the US. However, the
Federal Reserve decisions rest on the current macroeconomic conditions. Therefore, we should
expect that the short-end of the yield-curve is related to the development of macroeconomic
factors. Instead, the development of the long-end of the yield curve should largely depend on the
market average expectation and risk-aversion surrounding future interest rates and economic
conditions. Financial economists, then, should expect to see the long-end of the yield curve as
being driven by expectations of future economic activity, and by risk-aversion. Indeed, Ang and
Piazzesi (2003) demonstrate that macroeconomic factors such as ination and real economic
activity are able to explain movements at the short-end and the middle of the yield curve.
However, they show that the long-end of the yield curve is driven by unobservable factors.
However, it is not clear whether such unobservable factors are driven by time-varying riskaversion or changing expectations.
The compelling lesson for practitioners is that reduced-form models with only one factor are
unlikely to perform well, in practice.
11.3 Models of the short-term rate

The short-term rate represents the velocity at which locally riskless investments appreciate
over the next instant. This velocity, or rate of increase, is of course not a traded asset. What it
is traded is a bond and/or a MMA.
11.3.1
Introduction
The fundamental bond pricing equation in Eq. (11.5),

P (t, T ) = E e
T
t
r(u)du
(11.23)
suggests to model the arbitrage-free bond price P by using as an input an exogenously given
short-term rate process r. In the Brownian information structure considered in this chapter, r
would then be the solution to a stochastic dierential equation. As an example,
dr( ) = b(r( ), )d + a(r( ), )dW ( ), (t, T ],
(11.24)
where b and a are well-behaved functions guaranteeing the existence of a strong-form solution
to the previous equation.
This approach to modeling interest rates was the rst to emerge, after the seminal papers of
Merton (1973) (in a footnote!) and Vasicek (1977). This section illustrates the main modeling
and empirical challenges related to this approach. We examine one-factor models of the shortterm rate, such as that in Eqs. (11.23)-(11.24), and also multifactor models, in which the
short-term rate is a function of a number of factors, r ( ) = R(y ( )), where R is some function
and y is solution to a multivariate diusion process.
Two fundamental issues for the models users are that the models they deal with be (i)
fast to compute, and (ii) accurate. As regards the rst point, the obvious target would be to
look for models with a closed form solution, such as for example, the so-called ane models
(see Section 11.3.6). The second point is more subtle. Indeed, perfect accuracy can never be
290
c
by
A. Mele
achieved with models such as that in Eqs. (11.23)-(11.24) - even when this model is is extended
to a multifactor diusion. After all, the model in Eqs. (11.23)-(11.24) can only be taken as it
really is - a model of determination of the observed yield curve. As such the model in Eqs.
(11.23)-(11.24) can not exactly fit the observed term structure of interest rates.
As we shall explain, the requirement to exactly t the initial term-structure of interest rates
is important when the concern of the models user is the pricing of options or other derivatives
written on the bonds. And the good news is that such a perfect t can be obtained indeed, once
we augment Eq. (11.24) with an innite dimensional parameter calibrated to the observed
term-structure. The bad news, such a calibration leads to some intertemporal inconsistencies,
which we shally duly explain in a moment.
The models leading to perfect accuracy are often referred to as no-arbitrage models. These
models work by making the short-term rate process exactly pin down the term-structure we
observe at a given instant. The intertemporal inconsistencies arise because the parameters of
the short-term rate pinning down the term structure today, say, are likely to dier from from the
very same parameters as of tomorrow. Clearly, this methodology goes to the opposite extreme
of the original approach, where the short-term rate is the input of all subsequent movements
of the term-structure of interest rates. This original approach is consistent with the rational
expectations paradigm that permeates modern economic analysis: economically admissible, i.e.
no-arbitrage, bond prices move as a result of random changes in the state variables. Economists
try to explain broad phenomena with the help of a few inputs, a science reduction principle.
Practitioners, instead, implement models to solve pricing problems where bond prices have to
match market data. In these models used by practitioners, it is derivatives written on these
bond prices to move in reaction to changes in the underlying fundamentals, not bond prices,
which instead are perfectly tted, as we shall say. Both activities are important, and the choice
of the right model to use rests on the role that we are playing within a given institution.
11.3.2
The basic bond pricing equation
11.3.2.1 A rst derivation
Suppose bond prices are solutions to the following stochastic dierential equation:
dPi
= bi d + bi dW,
Pi
(11.25)
where W is a standard Brownian motion in Rd , bi and bi are some progressively measurable

functions ( bi is vector-valued), and Pi P ( , Ti ). The exact functional form of bi and bi
is not given, as in the BS case. Rather, it is endogenous and must be found as a part of the
equilibrium.
As shown in Appendix 1, the price system in (11.25) is arbitrage-free if and only if
bi = r + bi ,
(11.26)
for some Rd -dimensional process satisfying some basic regularity conditions. The meaning of
(13.15) can be understood by replacing it into Eq. (11.25), and obtaining:
dPi
= (r + bi ) d + bi dW.
Pi
The previous equation tells us that the growth rate of Pi is the short-term rate plus a termpremium equal to bi . In the bond market, there are no obvious economic arguments enabling
291
c
by
A. Mele
us to sign term-premia. Empirical evidence suggests that term-premia did take both signs over
the last twenty years. But term-premia would be zero in a risk-neutral world. In other terms,
bond prices are solutions to:
dPi
,
= rd + bi dW
Pi
#
= W + d is a Q-Brownian motion and Q is the risk-neutral probability.
where W
To derive Eq. (13.15) with the help of a specic version of theory developed in Appendix 1,
we now work out the case d = 1. Consider two bonds, and the dynamics of the value V of a
self-nancing portfolio in these two bonds:
dV = [ 1 (b1 r) + 2 (b2 r) + rV ] d + (1 1b + 2 b2 ) dW,
where i is wealth invested in bond maturing at Ti : i = i Pi . We can zero uncertainty by
setting
b2
1 = 2.
b1
By replacing this into the dynamics of V ,

b1 r
dV =
b2 + (b2 r) 2 d + rV d .
b1
Notice that 2 can always be chosen so as to make the value of this portfolio appreciate at a
rate strictly greater than r. It is sucient to set:

b1 r
b2 + (b2 r) .
sign( 2 ) = sign
b1
Therefore, to rule out arbitrage opportunities, it must be the case that:
b1 r
r
= b2
.
b1
b2
The previous relation tells us that the Sharpe ratio for any two bonds has to equal a process
, say, and Eq. (13.15) immediately follows. Clearly, such a does not depend on T1 or T2 .
In models of the short-term rate such as (11.24), functions bi and bi in Eq. (11.25) can be
determined after a simple application of Itos lemma. If P (r, , T ) denotes the rational bond
price function (i.e., the price as of time of a bond maturing at T when the state at is r),
and r is solution to (11.24), Itos lemma then implies that:

P
1 2
dP =
+ bPr + a Prr d + aPr dW,
2
where subscripts denote partial derivatives.
Comparing this equation with Eq. (11.25) then reveals that:
b P =
P
1
+ bPr + a2 Prr , b P = aPr .
Now replace these functions into Eq. (13.15) to obtain:

P
1
+ bPr + a2 Prr = rP + aPr , for all (r, ) R++ [t, T ),
2
292
(11.27)
c
by
A. Mele
with the obvious boundary condition

P (r, T, T ) = 1, all r R++ .
Eq. (11.27) shows that the bond price, P , depends on both the drift of the short-term rate, b,
and the risk-aversion correction, . This circumstance occurs as the initial asset market structure
is incomplete, in the following sense. In the Black-Scholes model, the option is redundant, given
the initial market structure. In the context we analyze here, the short-term rate r is not a
traded asset. In other words, the initial market structure has one untraded risk (r) and zero
assets - the factor generating uncertainty in the economy, r, is not traded. Therefore, the drift
of the short term, b, can not equal r r = r2 under the risk-neutral probability, and the bond
price depends on b, a and .
This dependence is, perhaps, a kind of hindrance to practitioners. Instead, it can be viewed
as a good piece of news to policy-makers. Indeed, starting from observations and (b, a), one
may back out information about , which contains information about agents risk-appetite.
Information about agents risk-appetite, then, can help central bankers to take decisions about
the interest rate to set.
By specifying the drift and diusion functions b and a, and by identifying the risk-premium
, the partial dierential equation (PDE, henceforth) (11.27) can explicitly be solved, either
analytically or numerically. Choices concerning the exact functional form of b, a and are often
made on the basis of either analytical or empirical reasons. In the next section, we will examine
the rst, famous short-term rate models in which b, a and have a particularly simple form.
We will discuss the analytical advantages of these models, but we will also highlight the major
empirical problems associated with these models. In Section 11.3.4 we provide a very succinct
description of models exhibiting jump (and default) phenomena. In Section 11.3.5, we introduce
multifactor models: we will explain why do we need such more complex models, and show that
even in this more complex case, arbitrage-free bond prices are still solutions to PDEs such as
(11.27). In Section 11.3.6, we will present a class of analytically tractable multidimensional
models, known as ane models. We will discuss their historical origins, and highlight their
importance as regards the econometric estimation of bond pricing models. Finally, Section
11.3.7 presents the perfectly tting models, and Appendix 5 provides a few technical details
about the solution of one of these models.
11.3.2.2 Derivation based on duration
The idea, here, is to replicate the price of a bond expiring at some time T1 , say P 1 P (r, , T1 ),
with a self-nanced portfolio comprising a money market account and a second bond expiring
at time T2 > T1 . The value of the self-nanced portfolio is V = P 2 + M , where is the
number of the number of the bonds maturing at T2 to be put in the portfolio, P 2 = P (r, , T2 ),
and M is the amount of resources put in the money market account. Since the portfolio is
self-nanced, we have, by the usual arguments, that,
where LP 2 =
P 2

dV = dP 2 + dM = LP 2 + rM d + aPr2 dW,
(11.28)
2
+ bPr2 + 12 a2 Prr
. And, obviously,
dP 1 = LP 1 d + aPr1 dW.
293
(11.29)
c
by
A. Mele
Let the initial value of the portfolio match the bond price. Then, comparing the diusive terms
in Eq. (11.29) and Eq. (11.28), we nd the delta to be:
= P (r, , T1 )/ r .
P (r, , T2 )/ r
Comparing the drift terms in Eq. (11.29) and Eq. (11.28),

LP 1 = LP 2 + rM = LP 2 + r V P 2 = LP 2 + r P 1 P 2 ,
where the last line follows as were using the values (, M ) such that the portfolio matches the
value of the rst bond. Rearranging terms yields, LP 1 rP 1 = (LP 2 rP 2 ), and evaluating
this for = ,
LP 1 rP 1
LP 2 rP 2
=
a,
Pr2
Pr2
for some and independent of calendar time.
can be interpreted as the ratio of the durations of the two bonds, as explained
The delta ,
in Chapter 13.
11.3.3
Some famous univariate short-term rate models
11.3.3.1 Vasicek and CIR
Vasiceks (1977) model is to be considered the seminal contribution to literature. The model
assumes the short-term rate is solution to:
dr( ) = ( r( ))d + dW ( ), (t, T ],
(11.30)
where , and are positive constants. This model generalizes that of Merton (1973), where
0. The intuition behind Eq. (11.30) is simple. Suppose, rst, that = 0. In this case, the
solution is:

( t)
r( ) = + e
r(t)
.
The previous equation reveals that if the current level of the short-term rate r(t) = /, it
will be locked-in
at

/ forever. If, instead, r(t) < /, then, for all > t, r( ) < /

too, but r( ) / will eventually shrink to zero as . An analogous property holds
when r(t) > /. In all cases, the speed of convergence of r to its long-term value / is
determined by : the higher is , the higher is the speed of convergence to /. In other terms,
/ is the long-term value towards which r tends to converge, and determines the speed of
such a convergence.
Eq. (11.30) generalizes the previous ideas to the stochastic dierential case. It can be shown
that a solution to Eq. (11.30) can be written in the following format:

"

( t)
r( ) = + e
r(t)
+ e
es dW (s),
t
where the integral has the so-called Itos sense meaning. The interpretation of this solution
is similar to the one given above. The short-term rate tends to a sort of central tendency
/. Actually, it will have the tendency to uctuate around it. In other terms, there is always
294
c
by
A. Mele
the tendency for shocks to be absorbed with a speed dictated by the value of . In this case,
the short-term rate process r is said to exhibit a mean-reverting behavior. In fact, it can be
shown that the expected future value of r will be given by the solution given above for the
deterministic case, viz

( t)
E [r( )| r (t)] = + e
r(t)
.
Of course, that is only the expected value, not the actual value that r will take at time . As a
result of the presence of the Brownian motion in Eq. (11.30), r can not be predicted, and it is
possible to show that the variance of the value taken by r at time is:
var [r( )| r (t)] =

2
1 e2( t) .
2
Finally, it can be shown that r is normally distributed (with expectation and variance given by
the two functions given above).
The previous properties of r are certainly instructive. Yet the main objective here is to nd
the price of a bond. As it turns out, the assumption that the risk premium process is a
constant allows one to obtain a closed-form solution. Indeed, replace this constant and the
functions b(r) = r and a(r) = into the PDE (11.27). The result is that the bond price
P is solution to the following partial dierential equation:
0=

1
P
+ ( ) r Pr + 2 Prr rP, for all (r, ) R [t, T ),
(11.31)
with the usual boundary condition. It is now instructive to see how this kind of PDE can be
solved. Guess a solution of the form:
P (r, , T ) = eA( ,T )B( ,T )r ,
(11.32)
where A and B have to be found. The boundary condition is P (r, T, T ) = 1, which implies that
the two functions A and B must satisfy:
A(T, T ) = 0 and B(T, T ) = 0.
(11.33)
Now suppose that the guess is true. By dierentiating Eq. (11.32), P

= (A1 B1 r)P , Pr = P B
2
and Prr = P B , where A1 ( , T ) A( , T )/ and B1 ( , T ) B( , T )/ . By replacing
these partial derivatives into the PDE (11.31) we get:

1 2 2
0 = A1 ( )B + B + (B B1 1)r, for all (r, ) R++ [t, T ).

2
This implies that for all [t, T ),
1
0 = A1 ( )B + 2 B 2 , 0 = B B1 1,
2
subject to the boundary conditions (11.33). The solutions are

1
1
1 e(T ) , A ( , T ) = 2
B( , T ) =
2
295
"
B(s, T ) ds ( )
2
"
B(s, T )ds.
c
by
A. Mele

By the denition of the yield curve given in Section 11.1 (see Eq. (13.1)),
ln P (r, t, T )
A(t, T ) B(t, T )
=
+
r.
T t
T t
T t
It is possible to show the existence of a nite asymptotic spot rate, i.e. limT R(t, T ) =
)
limT A(t,T
< .
T t
The model has a number of features that can describe quite a few aspects of reality. Many
textbooks show the typical shapes of the yield-curve that can be generated with the above
formula (see, for example, Hull (2003). However, this model is known to suer from two main
drawbacks. The rst drawback is that the short-term rate is Gaussian and, hence, can take on
negative values with positive probability. That is a counterfactual feature of the model. However,
it should be stressed that on a practical standpoint, this feature is practically irrelevant. If is
low compared to , this probability is really very small. The second drawback is tightly related
to the rst one. It refers to the fact that the short-term rate diusion is independent of the
level of the short-term rate. That is another conterfactual feature of the model. It is well-known
that short-term rates changes become more and more volatile as the level of the short-term rate
increases. In the empirical literature, this phenomenon is usually referred to as the level-effect.
The model proposed by Cox, Ingersoll and Ross (1985) (CIR, henceforth) addresses these
two drawbacks at once, as it assumes that the short-term rate is solution to,

dr( ) = ( r( ))d + r( )dW ( ), (t, T ].
R ( , T )
The CIR model is also referred to as square-root process to emphasize that the diusion
function is proportional to the square-root of r. This feature makes the model address the leveleect phenomenon. Moreover, this property prevents r from taking negative values. Intuitively,
when r wanders just above zero, it is pulled back to the stricly positive region at a strength
of the order dr = d .6 The transition density of r is noncentral chi-square. The stationary
density of r is a gamma distribution.
The expected value is as in Vasicek.7 However, the variance is dierent, although its exact
expression is really not important here.
CIR formulated a set of assumptions
that
on the primitives of the economy (e.g., preferences)
led to a risk-premium
function = r, where is a constant. By replacing this, b(r) = r
and a(r) = r into the PDE (11.27), one gets (similarly as in the Vasicek model), that the
bond price function takes the form in Eq. (11.32), but with functions A and B satisfying the
following dierential equations:
1
0 = A1 B, 0 = B1 + ( + )B + 2 B 2 1,
2
subject to the boundary conditions (11.33).
In their article, CIR also showed how to compute options on bonds. They even provided
hints on how to invert the term-structure, a popular technique that we describe in detail in
Section 11.3.6. For all these features, the CIR model and paper have been used in the industry
for many years. And many of the more modern models are mere multidimensional extensions
of the basic CIR model. (See Section 11.3.6).
6 This is only intuition. The exact condition under which the zero boundary is unattainable by r is
> 12 2 . See Karlin and
Taylor (1981, vol II chapter 15) for a general analysis of attainability of boundaries for scalar diusion processes.
7 The expected value of linear mean-reverting processes is always as in Vasicek, independently of the functional form of the
diusion coecient. This property follows by a direct application of a general result for diusion processes given in Chapter 6
(Appendix A).
296
c
by
A. Mele

11.3.3.2 Nonlinear drifts
An important issue is the analytical tractability of a given model. As demonstrated earlier,

models such as Vasicek and CIR admit a closed-form solution. Among other things, this is
because these models have a linear drift. Is evidence consistent with this linear assumption?
What does empirical evidence suggest as regards mean reversion of the short-term rate?
Such an empirical issue is subject to controversy. In the mid 1990s, three papers by AtSahalia (1996), Conley et al. (1997) and Stanton (1997) produced evidence that mean-reverting
behavior is nonlinear. As an example, Conley et al. (1997) estimated a drift function of the
following form:
b(r) = 0 + 1 r + 2 r2 + 3 r1 ,
which is reproduced in Figure 11.2 below (Panel A). Similar results were obtained in the other
papers. To grasp the phenomena underlying this nonlinear drift, Figure 11.2 (Panel B) also
contrasts the nonlinear shape in Panel A with a linear drift shape that can be obtained by
tting the CIR model to the same data set (US data: daily data from 1981 to 1996).
drift
drift
0.3
0.05
0.2
0.1
0.00
0.04
0.0
0.02
0.04
0.06
0.08
-0.1
0.10
0.12
0.14
0.06
0.16
short-term rate r
0.08
0.10
0.12
short-term rate r
-0.05
-0.2
-0.10
-0.3
Panel A
Panel B
FIGURE 11.3. Nonlinear mean reversion?

The importance of the nonlinear eects in Figure 11.3 is related to the convexity eects in
Mele (2003). Mele (2003) showed that bond prices may be concave in the short-term rate if the
risk-neutralized drift function is suciently convex. While the results in this Figure relate to
the physical drift functions, the point is nevertheless important as risk-premium terms should
look like very strange to completely destroy the nonlinearities of the short-term rate under the
physical probability.
The main lesson is that under the nonlinear drift dynamics, the short-term rate behaves in
a way that can at least be roughly comparable with that it would behave under the linear drift
dynamics. However, the behavior at the extremes is dramatically dierent. As the short-term
rate moves to the extremes, it is pulled back to the center in a very abrupt way. At the
moment, it is not clear whether these preliminary empirical results are reliable or not. New
econometric techniques are currently being developed to address this and related issues.
One possibility is that such single factor models of the short-term rate are simply misspecied.
For example, there is strong empirical evidence that the volatility of the short-term rate is timevarying, as we shall discuss in the next section. Moreover, the term-structure implications of
297
c
by
A. Mele
a single factor model are counterfactual, since we know that a single factor can not explain
the entire variation of the yield curve, as we explained in Section 11.2. We now describe more
realistic models driven by more than one factor.
11.3.4
Multifactor models
The empirical evidence reviewed in Section 11.2 suggests that one-factor models can not explain
the entire variation of the term-structure of interest rates. Factor analysis suggests we need at
least three factors. In this section, we succinctly review the advances made in the literature to
address this important empirical issue.
11.3.4.1 Stochastic volatility
In the CIR model, the instantaneous short-term rate volatility is stochastic, as it depends
on the level of the short-term rate, which is obviously stochastic. However, there is empirical
evidence, surveyed by Mele and Fornari (2000), which suggests that the short-term rate volatility
depends on some additional factors. A natural extension of the CIR model is one in which the
instantaneous volatility of the short-term rate depends on (i) the level of the short-term rate,
similarly as in the CIR model, and (ii) some additional random component. Such an additional
random component is what we shall refer to as the stochastic volatility of the short-term
rate. It is the term-structure counterpart to the stochastic volatility extension of the Black and
Scholes (1973) model (see Chapter 10).
Fong and Vasicek (1991) write the rst paper in which the volatility of the short-term rate
is stochastic. They consider the following model:

dr ( ) = r (
r r ( )) d + v( )r (t) dW1 ( )
dv ( ) = v ( v ( )) d + v v ( )dW2 ( )
(11.34)
in which r , r, v , and v are constants, and [W1 W2 ] is a vector Brownian motion. To obtain
a closed-form solution, Fong and Vasicek set = 0. The authors also make assumptions about
risk aversion corrections. Namely, they assume that the unit-risk-premia for the stochastic uctuations
of the short-term rate, r , and the short-term rate volatility, v , are both proportional

to v ( ), and then they nd a closed-form solution for the bond price as of time t and maturing
at time T , P (r (t) , v (t) , T t).
Longsta and Schwartz (1992) propose another model of the short-term rate in which the
volatility of the short-term rate is stochastic. The remarkable feature of their model is that
it is a general equilibrium model. Naturally, the Longsta-Schwartz model predicts, as the
Fong-Vasicek model, that the bond price is a function of both the short-term rate and its
instantaneous volatility.
Note, then, the important feature of these models. The pricing function, P (r (t) , v (t) , T t)
and, hence, the yield curve R (r (t) , v (t) , T t) (T t)1 ln P (r (t) , v (t) , T t), depends
on the level of the short-term rate, r (t), and one additional factor, the instantaneous variance
of the short-term rate, v (t). Hence, these models predict that we now have two factors that
help explain the term-structure of interest rates, R (r (t) , v (t) , T t).
What is the relation between the volatility of the short-term rate and the term-structure
of interest rates? Does this volatility help track one of the factors driving the variations of
the yield curve? Consider, rst, the basic Vasicek (1997) model. While this model assumes the
short-term rate volatility is constant, it can still be used to develop intuition about models with
298
c
by
A. Mele
stochastic volatility models, such those the Fong and Vasicek (1991) of Eqs. (11.34). For the
Vasicek model, then,
" T

" T
R (r (t) , T t)
1
2
B (T s) ds +
B (T s) ds .
=
T t
t
t
(11.35)

where B (T s) = 1 1 e(T s) . Eq. (11.35) shows that if 0, the term-structure is
decreasing short-term rate volatility. That is, bond prices increase in , a conclusion paralleling
that for options, where option prices are increasing in the volatility of the asset price. As
explained in Chapter 10, this property arises through the optionality of the contractsay the
convexity of a European call price with respect to the asset price.
Interesting properties arise when < 0, which is an empirically relevant case.8 In this case,
)
the sign of R(t,T
is determined by both convexity and slope eects. Convexity eects,
2 (r,T t)
those relating to the second partial P r
= P (r, T t) B (T t)2 , arise through the term
2
#T
t)
t B(T s)2 ds. Slope eects, those relating to P (r,T
= P (r, T t) B (T t), arise,
r
#T
instead, through the term t B (T s) ds. If is negative, and large in absolute value, slope
eects can dominate convexity eects, and the term-structure can actually increase in . For
intermediate values of , the term-structure can be both increasing and decreasing in . At short
maturities, the convexity eects in Eq. (11.35) are typically dominated by slope eects, and
the short-end of the term-structure can be increasing in . At longer maturity dates, however,
convexity eects are more important and, sometimes, dominate slope eects.
To develop further intuition, consider the following binomial example. In the next period, the
short-term rate is either r+ = r + d or r+ = r d with equal probability, where r is the current
interest rate level and d > 0. The price of a two-period bond is P (r, d) = m(r, d)/ (1 + r),
where m(r, d) = E [1/ (1 + r+ )] is the expected discount factor of the next period. By Jensens
inequality, m(r, d) > 1/ (1 + E [r+ ]) = 1/ (1 + r) = m(r, 0). Therefore, two-period bond prices
increase upon activation of randomness. More generally, two-period bond prices are always
increasing in the volatility parameter d in this example (see Figure 11.4). Again, this property
relates to an insight of Jagannathan (1984, p. 429-430) that in a two-period economy with
identical initial underlying asset prices, a terminal underlying asset price y is a mean preserving
spread of another terminal underlying asset price x (in the Rothschild and Stiglitz (1970) sense)
if and only if the price of a call option on y is higher than the price of a call option on x. This
is because if y is a mean preserving spread of x, then E [f(
y )] > E [f (
x)] for f increasing and
9
convex.
These properties arise as we assumed the expected short-term rate is independent of d. In an
alternative setting, say a multiplicative setting where either r+ = r (1 + d) or r+ = r/ (1 + d)
with equal probability, bond prices are decreasing in volatility at short maturities and increasing
in volatility at longer maturities, as originally pointed out by Litterman, Scheinkman and Weiss
(1991). Its because expected future interest rates increase over time at a strength positively
related to d. That is, the expected variation of the short-term rate is increasing in the volatility
of the short-term rate, d, a property that can be re-interpreted as one arising in an economy
8 In
this simple model, the assumption that < 0 is reasonable, as we observe positive risk-premia more often than negative
risk-premia. But in this very same model, ur < 0, which together with < 0, ensures that term-premia are positive.
9 In our case, let m
d (i+ ) = 1/ (1 + i+ ) denote the random discount factor when i+ = i d. We have that x m
d (x)
is increasing and concave and, hence, E [m
d (x)] < E [m
d (x)] d < d , which is what demonstrated in Figure 11.4. In
Jagannathan (1984), f is increasing and convex, and so we must have: E [f (
y)] > E [f (
x)] y is riskier than, or a mean preserving
spread of, x
.
299
c
by
A. Mele
with risk-averse agents. At short maturity dates, such an eect dominates the convexity eect
illustrated in Figure 11.4. At longer maturity dates, the convexity eect dominates.
More generally, then, and as regards the term-structure, volatility changes do not represent
a mean-preserving spread for the risk-neutral distribution, as Eq. (11.35) illustrates for the Vasicek model. In a world with complete markets, as in the Black-Scholes one, the asset underlying
the contract is traded. As regards interest rates, the situation diers, for the very simple reason
the short-term rate is not a traded asset. Therefore, the risk-neutral drift of the short-term rate
does in general depend on the short-term volatility through some risk-adjustementin Vasicek,
for example, this dependence is channeled through the risk-premium parameter . And while
the previous conclusions rely on comparative statics for a constant volatility model, they illustrate the more general situation of stochastic volatility. Mele (2003) shows that in more complex
stochastic volatility cases, provided the risk-premium required to bear the interest rate risk is
negative, and suciently large in absolute value, slope eects dominate convexity eects at any
nite maturity date, thus making bond prices decrease with volatility at any arbitrary maturity
date.
1
a
m(r,d) = (a + A)/2
m(r,d) = (b + B)/2
B
A
r d
r d
r +d
r + d
FIGURE 11.4. If the risk-neutralized interest rate of the next period is either r+ = r + d or
r+ = r d with equal probability, the random discount factor 1/ (1 + r+ ) is either B or b with
equal probability. Hence m(r, d) = E [1/ (1 + r+ )] is the midpoint of bB. Similarly, if volatility
is d > d, m(r, d ) is the midpoint of aA. Since ab > BA, it follows that m(r, d ) > m(r, d).
Therefore, the two-period bond price P (r, d) = m(r, d)/ (1 + r) satises: P (r, d ) > P (r, d) for
d > d.
What are the implications of these conclusions in terms of the classical factor analysis of
the term-structure reviewed in Section 11.2? Clearly, the very short-end of yield curve is not
aected by movements of the volatility, as limT t R (r (t) , v (t) , T t) = r (t), for all possible
where
values of v (t). Also, in these models, we have that limT R (r (t) , v (t) , T t) = R,
is a constant and, hence, independent of of v (t). Therefore, movements in the short-term
R
volatility can only produce their eects on the middle of the yield curve. For example, if the
risk-premium required to bear the interest rate risk is negative and suciently large, an upward
movement in v (t) can produce an eect on the yield curve qualitatively similar to that depicted
300
c
by
A. Mele
in Figure 11.2 (Curvature panel), and would thus roughly mimic the curvature factor that
we reviewed in Section 11.2.
11.3.4.2 Three-factor models
We need at least three factors to explain the entire variation in the yield-curve. A model in
which the interest rate volatility is stochastic may be far from being exhaustive in this respect. A
natural extension is a model in which the drift of the short-term rate contains some predictable
component, r ( ), which acts as a third factor, as in the following model:

dr ( ) = r (
r ( ) r ( )) d + v ( )r (t) dW1 ( )
dv ( ) = v ( v ( )) d + v v ( )dW2 ( )
d
r ( ) = r ( r ( )) d + r r ( )dW3 ( )
(11.36)
where r , , v , , v , r, and r are constants, and [W1 W2 W3 ] is vector Brownian motion.

Balduzzi et al. (1996) develop the rst model in which the drift of the short-term rate changes
stochastically, as in Eqs. (11.36). Dai and Singleton (2000) consider a number of models that generalize that in Eqs. (11.36). The term-structure implications of these models can be understood
very simply. First, the bond price has now the form, P (r (t) , r (t) , v (t) , T t) and, hence, the
yield curve is, under reasonable assumptions on the risk-premia, R (r (t) , r (t) , v (t) , T t)
(T t)1 ln P (r (t) , r (t) , v (t) , T t). Second, and intuitively, changes in the new factor r (t)
should primarily aect the long-end of the yield curve. This is because empirically, the usual
nding is that the short-term rate reverts relatively quickly to the long-term factor r ( ) (i.e. r
is relatively large), where r ( ) mean-reverts slowly (i.e. r is relatively low). This mechanism
makes the short-term rate quite persistent anyway. Ultimately, then, the slow mean-reversion of
r ( ) means that changes in r ( ) last for the relevant part of the term-structure we are usually
interested in (i.e. up to 30 years), despite the fact that limT R (r (t) , r (t) , v (t) , T t) is
independent of the movements of the three factors r (t), r (t) and v (t).
However, it is dicult to see how to reconcile such a behavior of the long-end of the yield
curve with the existence of any of the factors discussed in Section 11.2. First, the short-term rate
can not be taken as a level factor, since we know its eects die o relatively quickly. Instead, a
joint change in both the short-term rate, r (t), and the long-term rate, r (t), should be really
needed to mimic the Level panel of Figure 11.2 in Section 11.2. However, this interpretation
is at odds with the assumption that the factors discussed in Section 11.2 are uncorrelated!
Moreover, and crucially, the empirical results in Dai and Singleton reveal that if any, r (t) and
r (t) are negatively correlated.
Finally, to emphasize how exacerbated these puzzles are, consider the eects of changes in
the short-term rate r (t). We know that the long-end of the term-structure is not aected by
movements of the short-term rate. Hence, the short-term rate acts as a steepness factor, as
in Figure 11.2 (Slope panel). However, this interpretation is restrictive, as factor analysis
reveals that the short-end and the long-end of the yield curve move in opposite directions after
a change in the steepness factor. Here, instead, a change in the short-term rate only modies
the short-end (and, perhaps, the middle) of the yield curve and, hence, does not produce any
variation in the long-end curve.
301
c
by
A. Mele

11.3.5
Affine and quadratic term-structure models
11.3.5.1 Ane
The Vasicek and CIR models predict that the bond price is exponential-ane in the short-term
rate r. This property is the expression of a general phenomenon. Indeed, it is possible to show
that bond prices are exponential-ane in r if, and only if, the functions b and a2 are ane in
r. Models that satisfy these conditions are known as affine models. More generally, these basic
results extend to multifactor models, in which bond prices are exponential-ane in the state
variables.10 In these models, the short-term rate is a function r (y) such that
r (y) = r0 + r1 y,
where r0 is a constant, r1 is a vector, and y is a multidimensional diusion, in Rn , and is solution
to.
dy ( ) = ( y (t)) dt + V (y ( )) dW ( ) ,
where W is a d-dimensional Brownian motion, is a full rank n d matrix, and V is a full
rank d d diagonal matrix with elements,

i = 1, , d,
(11.37)
V (y)(ii) = i +
i y,
for some scalars i and vectors i . Langetieg (1980) develops the rst multifactor model of this
kind, in which i = 0. Under the assumption that the risk-premia are
(y) = V (y) 1 ,
for some d-dimensional vector 1 , Due and Kan (1996) show that the bond price is exponentialane in the state variables y. That is, the price of the zero has the following functional form,
P (y, T t) = exp [A (T t) + B (T t) y] ,
for some functions A and B of time to maturity, T t (B is vector-valued), such that A (0) = 0
and B (0)(i) = 0.
The clear advantage of ane models is that they considerably simplify the econometric
estimation, as explained below.
11.3.5.2 Quadratic
Ane models are known to impose tight conditions on the structure of the volatility of the
state variables. These restrictions arise to keep the square root in Eq. (11.37) real valued. But
these constraints may hinder the actual performance of the models. There exists another class
of models, known as quadratic models, that partially overcome these diculties.
11.3.6
Short-term rates as jump-diffusion processes
Seminal contribution (extension of the CIR general equilibrium model to jumps-diusion): Ahn
and Gao (1988, JF). Suppose that the short-term rate is a jump-diusion process:
dr( ) = bJ (r( ))d + a(r( ))dW ( ) + (r( )) S dZ( ),
10 More generally, we say that ane models are those that make the characteristic function exponential-ane in the state variables.
In the case of the multifactor interest rate models of the previous section, this condition is equivalent to the condition that bond
prices are exponential ane in the state variables.
302
c
by
A. Mele
where the previous equation is written under the risk-neutral probability, and bJ is thus a jumpadjusted risk-neutral drift. For all (r, ) R++ [t, T ), the bond price P (r, , T ) is then the
solution to,
0=

"
Q
+ L r P (r, , T ) + v
[P (r + S, , T ) P (r, , T )] p (dS) ,
supp(S)
(11.38)
and P (r, T, T ) = 1 r R++ .

#
This is because, as usual, {exp( t r(u)du)P (r, , T )} [t,T ] must be a martingale under the
risk-neutral probability in order to prevent arbitrage opportunities.11 Also, we can model the
presence of dierent quality (or types) of jumps, and the previous formula becomes:
0=

"
N
Q
+ L r P (r, , T ) +
vj
[P (r + S, , T ) P (r, , T )] pj (dS) ,
supp(S)
j=1
where N is the number of jump types, but here for simplicity we just set N = 1.
As regards the risk-neutral distribution, the important thing as usual is to identify the riskpremia. Here we simply have:
v Q = v J ,
where v is the intensity of the short-term rate jump under the physical distribution, and J is
the risk-premium demanded by agents to be compensated for the presence of jumps.12
Bonds subject to default-risk can be modeled through partial dierential equations. This is
particularly the case when default is considered as an exogeously given rare event modeled as
a Poisson process. This is the so-called reduced-form approach. Precisely, assume that the
event of default at each instant of time is a Poisson process Z with intensity v, and assume
that in the event of default at point , the holder of the bond receives a recovery payment
P ( ) which can be a deterministic function of time (e.g., a constant) or more generally, a
(r(s) : t s )-adapted process satisfying some basic regularity conditions.
Next, let be the random default time, and lets create an auxiliary state variable g with
the following features:
+
0 if t <
g=
1 otherwise
The relevant information for an investor is thus given by the following risk-neutral dynamics:
+
dr( ) = b(r( ))d + a(r( ))dW ( )

dg( ) = S dN( ), where S 1, with probability one
(11.39)
Denote the rational bond price function as P (r, g, , T ), [t, T ]. It is assumed that
[t, T ] and v (0, ), P (r, 1, , T ) = P ( ) < P (r, 0, , T ) a.s. As shown below, such an
assumption, plus the assumption that P ( ; v ) P ( ; v) v v, is sucient to guarantee
that default-free bond prices are higher than defaultable bond prices.
11 Just use y( ) b( )1 u(r( ), , T ), where b solves db() = r( )b( )d (in dierential form), for the connection between Eq.
(11.38) and martingales.
12 Further details on changes of measures for jump-type processes can be found in Br
emaud (1981).
303
c
by
A. Mele
By the usual absence of arbitrage opportunities arguments, the following equation is satised
by the pre-default bond price P (r, 0, , T ) = P pre (r, , T ):

0=
+ L r P (r, 0, , T ) + v(r) [P (r, 1, , T ) P (r, 0, , T )]

=
+ L (r + v(r)) P (r, 0, , T ) + v(r)P ( ), [t, T ),
(11.40)
with the usual boundary condition P (r, 0, T, T ) = 1.

The solution for the pre-default bond price is:
P
pre

" T

(x, t, T ) = E exp
(r( ) + v(r( )))d
t
" T
"

exp
(r(u) + v(r(u)))du v(r( ))P ( )d ,
+E
where E [] is the expectation taken with reference to only the rst equation of system (11.39).
This coincides with Due and Singleton (1999, Eq. (10) p. 696) when we dene a percentage
loss process l in [0, 1] so as to have P = (1 l) P . Indeed, inserting P = (1 l) P into Eq.
(11.40) gives:
0=
+ L (r + l( )v(r)) P (r, 0, , T ), (r, ) R++ [t, T ),
with the usual boundary condition, the solution of which is:

P
pre
(x, t, T ) = E
" T

exp
(r( ) + l( ) v(r( )))d
.
t
To validate the claim that the bond price is decreasing with v, consider two economies A and B
in which the corresponding default-intensities are vA and v B , and assume that the coecients
of L dont depend on default-intensity. The pre-default bond price function in economy i is
P i (r, , T ), i = A, B, and satises:
0=
+ L r P i + vi (P i P i ), i = A, B,
with the usual boundary condition. Substracting these two equations and rearranging terms
reveals that the price dierence P (r, , T ) P A (r, , T ) P B (r, , T ) satises, (r, )
R++ [t, T ),
0=

A
+ L (r + v ) P (r, , T )+ v A v B P B ( ) P B (r, , T ) +v A P A ( ) P B ( ) ,
with P (r, T, T ) = 0, r R++ . Given the previous assumptions, the proof is complete by an
application of the maximum principle (see Appendix ? in Chapter 6).
304
c
by
A. Mele

11.3.7
Estimation strategies
Let r (t) be the short-term rate process, solution to the following stochastic dierential equation,
dr (t) = ( r (t)) dt +

v (t)r (t) dW (t) ,
t 0,
(11.41)
where W (t) is a standard Brownian motion, and ,

and are three positive constants. Suppose, also, that the instantaneous volatility process v (t)r (t) is such that v (t) is solution
to,

dv (t) = ( v (t)) dt + v (t) dW (t) + 1 2 dU (t) , t 0,
(11.42)
where U (t) is another standard Brownian motion; , , and are four positive constants,
and is a constant such that || < 1.
11.3.7.1 The level eect
Which empirical regularities would the short-term rate model in Eqs. (11.41)-(11.42) address?
Which sign of the correlation coecient would be consistent with historical episodes such as
the Monetary Experiment of the Federal Reserve System between October 1979 and October
1982?
The short-term rate model in Eqs. (11.41)-(11.42) would address two empirical regularities.
1) The volatility of the short-term rate is not constant over time. Rather, it seems to be
driven by an additional source of randomness. All in all, the short-term process seems to be
generated by the stochastic volatility model in Eqs. (11.41)-(11.42), in which the volatility
component v (t) is driven by a source of randomness only partially correlated with the source
of randomness driving the short-term rate process itself.
2) The volatility of the short-term rate is increasing in the level of the short-term rate.
This phenomenon is known as the level eect. Perhaps, periods of high interest rates arise
because of erratic liquidity. (Erratic liquidity would command a high risk-premium and so a
high LIBOR rate say.) But precisely because of erratic liquidity, interest rates are also very
volatile. The short-term rate model in Eqs. (11.41)-(11.42) is a very useful reduced form able
to capture these eects through the two parameters: and . If the parameter is greater than
zero, the instantaneous interest rate volatility increases with the level of the interest level. If the
correlation coecient > 0, the interest rate volatility is also partly related to the sources
of interest rate volatility not directly related to the level of the interest rate.
During the Monetary Experiment, the FED decided to target money supply, rather than
interest rates. So the high volatility of money demand mechanically translated to high interest
rate volatility as a result of the market clearing. Moreover, the quantity of monetary base was
kept deliberately low - to ght against ination. So the US experienced both high interest rate
volatility and high interest rates (see, for example, Andersen and Lund, 1997, for an empirical
study). Moreover, high nominal interest rates may be so because they might be compensating
for high inflation volatility, that is, not only high ination. There is no empirical study about the
issues related to the sign of the correlation coecient . Here is a suggestion. A rolling window
estimation suggestive that the level of changed a lot around the Monetary Experiment would
mean that the bulk of interest rate volatility was not entirely due to the mechanical eects
related to the FED behavior.
305
c
by
A. Mele

11.3.7.2 The simplest estimation case
Next, suppose we wish to estimate the parameter vector = [, , , , , , , ] of the model

in Eqs. (11.41)-(11.42). Under which circumstances would Maximum Likelihood be a feasible
estimation method?
The ML estimator would be feasible under two sets of conditions. First, the model in Eqs.
(11.41)-(11.42) should not have stochastic volatility at all, viz, = = 0; in this case, the
short-term rate would be solution to,
dr (t) = ( r (t)) dt +
r (t) dW (t) ,
t 0,
where is now a constant. Second, the value of the elasticity parameter is important. If = 0,
the short-term rate process is the Gaussian one proposed by Vasicek (1977). If = 12 , we obtain
the square-root process of Cox, Ingersoll and Ross (1985). In the Vasicek case, the transition
density of r is Gaussian, and in the CIR case, the transition density of r is a noncentral chisquare. So in both the Vasicek and CIR, we may write down the likelihood function of the
diusion process. Therefore, ML estimation is possible in these two cases.
In the more general case, we have to go for simulation methods described in Chapter 5.
11.3.7.3 More general models
Estimating the model in Eqs. (11.41)-(11.42) is certainly instructive. Yet a more important
question is to examine the term-structure implications of this model. More generally, how would
the estimation procedure outlined in the previous subsection change if the task is to estimate
a Markov model of the term-structure of interest rates? There are three steps.
Step 1
Collect data on the term structure of interest rates. We will need to use data on two maturities
(say a time series of riskless 6 months and 5 year interest rates).
Step 2
Let us consider a model of the entire term-structure of interest rates. By the fundamental
theorem of asset pricing, and the Markov property of the diusion, the price of a riskless bond
predicted by the model is,

#N

j

r(s)ds
j
P (r (t) , v (t)) P (r (t) , v (t) , Nj t) = E e t
(11.43)
r (t) , v (t) ,
where E (.) is the conditional expectation taken under the risk-neutral probability, and Nj is a
sequence of expiration dates. Naturally, the previous formula relies on some assumptions about
risk-aversion correction. Some of these assumptions may be of a reduced-form nature; others
may rely on the specication of preferences, beliefs, markets and technology. But we do not
need to be more precise at this level of generality. In turn, these assumptions entail that the
pricing formula in Eq. (11.43) depends on some additional risk-adjustment parameter vector,
say . Precisely, the Radon-Nykodim derivative
probability
with respect to
1 # of the2 risk-neutral
#

the physical probability is given by exp 2 (t) dt (t) dZ (t) , where Z = [W U ] ,
W and U are the two Brownian motions in Eqs. (11.41)-(11.42), and (t) is some process
adapted to Z, which is taken to be of the form (t) m (r (t) , v (t) ; ), for some vector
valued function m and some parameter vector . The function m makes risk-adjustment
306
c
by
A. Mele
11.4. No-arbitrage models
corrections dependent on the current value of the state vector (r (t) , v (t)), and thus makes the
model Markov.
So the estimation problem is actually one in which we have to estimate both the physical
parameter vector = [, , , , , , , ] and the risk-adjustment parameter vector .
Next, compute interest rates corresponding to two maturities,
Rj (r (t) , v (t) ; , ) =
1
ln P j (r (t) , v (t)) , j = 1, 2,
Nj
(11.44)
where the bond prices are computed through Eq. (11.43), and where the notation Rj (r, v; , )
emphasizes that the theoretical term-structure depends on the parameter vector (, ). We can
now use the data (Rj$ say) and the model predictions about the data (Rj ), create moment conditions, and proceed to estimate the parameter vector (, ) through some method of moments
(provided the moments are enough to make (, ) identiable). But there are two diculties.
First, the volatility process v (t) is not observable by the econometrician. Second, the bond
pricing formula in Eq. (11.43) does not generally admit a closed-form.
The rst diculty can be overcome through inference methods based on simulations. Here is
an outline of these methods that could be used here. Simulate the system in Eqs. (11.41)-(11.42)
for a given value of the parameter vector (, ). For each simulation, compute a time series of
interest rates Rj from Eq. (11.44). Use these simulated data to create moment conditions.
The parameter estimator is the value of (, ) which minimizes some norm of these moment
conditions obtained through the simulations, with any of the methods explained in Chapter 5.
The next step discusses how to address the second diculty.
Step 3
The use of affine models would considerably simplify the analysis. Ane models place restrictions on the data generating process in Eqs. (11.41)-(11.42) and in the risk-aversion corrections
in Eq. (11.43) in such a way that the term structure in Eq. (11.44) is,
Rj (r (t) , v (t) ; , ) = A (j; , ) + B (j; , ) y (t) , j = 1, 2,
where A (j; , ) and B (j; , ) are some functions of the maturity Nj (B is vector valued),
and generally depend on the parameter vector (, ); and nally the state vector y = [r v] .
(Namely, an ane model obtains once = 0, = 12 , and the function m is ane.) So once Eqs.
(11.41)-(11.42) are simulated, the computation of a time series of interest rates Rj is straight
forward.
11.4 No-arbitrage models

11.4.1
Fitting the yield-curve, perfectly
For derivative trading purposes, we do not really wish to explain the term structure. Rather, we
wish to take it as given. Consider, for instance, a European option written on a bond. We may
nd it unsatisfactory to have a model that only explains the bond price. A models mistake
on the bond price is likely to generate a huge option price mistake. How can we trust an option
pricing model that is not even able to pin down the value of the underlying asset price? To
illustrate these points, denote with P (r( ), , S) the rational price process of a zero coupon
bond maturing at some time S. What is the price of a European option written on this bond,
307
c
by
A. Mele
struck at K and expiring at T < S? By the FTAP, there are no arbitrage opportunities if and
only if the option price C b is:
!

+
b
tT r( )d
C (r(t), t, T, S) = E e
(P (r(T ), T, S) K) ,
where the symbol (a)+ denotes max (0, a). As an example, in ane models, P is lognormal
whenever r is normally distributed. This happens precisely for the Vasicek model. The intuition
developed for the Black and Scholes (1973) (BS) formula suggests that in this case, the previous
expectation is a nonlinear function of the current bond price P (r(t), t, T ). This claim can not
be shown with the simple risk-neutral
tools used to show the BS formula. One of the troubles

tT r( )d
is due to the presence of the e
term inside the brackets, which is obviously unknown
at the time of evaluation t. But the problem is tractable, thanks to the forward martingale
probability introduced in Section 11.2.4. Precisely, let 1ex be the indicator of all events s.t. the
option is exercized i.e., that P (r(T ), T, S) K. We have:
C b (r(t), t, T, S)
=E e
T
r( )d
T
r( )d
P (r(T ), T, S) 1ex K E e
1ex
.
.
S
T
e t r( )d
e t r( )d
= P (r(t), t, S) E
1ex KP (r(t), t, T ) E
1ex
P (r(t), t, S)
P (r(t), t, T )
t
= P (r(t), t, S) EQSF [1ex ] KP (r(t), t, T ) EQTF [1ex ]

= P (r(t), t, S) QSF [P (r(T ), T, S) K] KP (r(t), t, T ) QTF [P (r(T ), T, S) K] ,
(11.45)
where the rst term in the second equality follows by an argument nearly identical to that
produced in Section 11.1 (see footnote 2);13 QiF (i = T, S) is the i-forward probability; and
nally, EQiF [] is the expectation taken under the i-forward martingale probability (see Section
11.1 for more details).
Section 11.7 explains how the two probabilities in Eq. (11.45) are computed. The important
issue, now, is to emphasize that the bond option price does depend on the theoretical bond
prices P (r(t), t, T ) and P (r(t), t, S), which, in turn, cannot equal the current, observed market
prices. Theoretical prices are simply the output of a rational expectations model. This fact is not
a source of concern to those who wish to predict future term-structure movements with the help
of a few, key state variables, as in the multifactor models discussed earlier. However, a source of
concern to practitioners dealing with pricing a bond option is that the pricing model perfectly
matches the yield curve at the time of evaluation. The aim of this section is to introduce a class
of models that t the yield curve without errors, which we call perfectly tting models:: these
models predict prices of bonds expiring at some date S, which are of course random at time
T < S, but also exactly equal to the current market bond prices (at time t). Finally, these prices
must, of course, be arbitrage-free. As we show, these conditions can be met by augmenting the
models seen in the previous sections with a set of innite dimensional parameters. We do not
13 By
the Law of Iterated Expectations,

T

T
E e t r( )d P (r(T ), T, S)1ex = E e t
r()d

S

S

1ex E e T r( )d F (T ) = E e t r( )d 1ex .
308
c
by
A. Mele
develop a general model-building principle, however. Rather, we discuss two specic yet famous
such models: the Ho and Lee (1986) model, and one generalization of it, introduced by Hull
and White (1990).
A nal remark. In Section 11.7, we will show that at least for the Vasiceks model, Eq. (11.45)
does not explicitly depend on r because it only depends on P (r(t), t, T ) and P (r(t), t, S). So
why do we look for perfectly tting models in the rst place? Wouldnt it be enough, then,
to just replace the theoretical prices P (r(t), t, T ) and P (r(t), t, S) with the market values, say
P $ (t, T ) and P $ (t, S)? This way, the model is perfectly tting. Apart from being logically
inconsistent (you would have a model predicting something generically dierent from prices),
this way of proceeding also has practical drawbacks. Section 11.7 shows that option pricing
formulae for European options, might well agree in notation with those relating to perfectly
tting models. However, Section 11.7.3 explains that as we move towards more complex interest
rate derivatives products, such as options on coupon bonds and swaption contracts, the situation
gets dramatically dierent. Finally, it can be the case that some maturity dates are actually
not traded at some point in time. For example, it may happen that P $ (t, T ) is not observed
and that we could still be interested in pricing more exotic or less liquid bonds or options
on these bonds. An intuitive procedure to deal with this this diculty is to interpolate the
traded maturities. In fact, the objective of perfectly tting models is to allow for such an
interpolation while preserving absence of arbitrage opportunities.
11.4.2
Ho and Lee
The original Ho and Lee (1986) model is in discrete-time and is analyzed in the context of
Chapter 13, along with other models. The model below, represents the diusion limit of the
original Ho and Lee model:
( ) , t,
dr ( ) = ( ) d + dW
(11.46)
is a Q-Brownian motion, is a constant, and ( ) is an innite dimensional

where W
parameter introduced to pin down the initial, observed term structure. The time of evaluation
is t. The reason we refer ( ) to as innite dimensional parameter is that we ( ) is a function
of calendar time t. Crucially, then, we assume that this function is known at the time of
evaluation t.
Clearly, Eq. (11.46) gives rise to an ane model. Therefore, the bond price takes the following
form,
P (r ( ) , , T ) = eA( ,T )B( ,T )r( ) ,
(11.47)
for two functions A and B to be determined below. It is easy to show that,
" T
1
A ( , T ) =
(s) (s T ) ds + 2 (T )3 , B ( , T ) = T .
6
Let f$ (t, ) denote the instantaneous, observed forward rate. By matching the instantaneous
forward rate f ( , T ) predicted by the model to f$ ( , T ) yields:
" T
ln P (r ( ) , , T )
1
f$ ( , T ) = f ( , T ) =
=
(s) ds 2 (T )2 + r ( ) .
(11.48)
T
2
#

T
Because P (t, T ) = exp t f (t, ) d , the drift term (s) satisfying Eq. (11.48) also guarantees an exact t of the yield curve. By dierentiating the previous equation with respect to
309
c
by
A. Mele

T , leaves (T ) =
f
T $
( , T ) + 2 (T ), or:
( ) =
f$ (t, ) + 2 ( t) .
(11.49)
To check that is indeed the solution we were looking for, we replace Eq. (11.49) into Eq.
(11.48) and verify indeed that Eq. (11.48) holds as an identity.
[Develop connections with the HML approach introduced in Section 11.5 below]
11.4.3
Hull and White
Consider the model
( ),
dr( ) = (( ) r( )) d + dW
(11.50)
is a Q-Brownian motion, and , are constants. Clearly, this model generalizes both
where W
the Ho and Lee model (11.46) and the Vasicek model (11.30). In the original formulation of
Hull and White, and were both time-varying, but the main points of this model can be
learnt by working out this particular simple case.
Eq. (11.50) also gives rise to an ane model. Therefore, the solution for the bond price is
given by Eq. (11.47). It is easy to show that the functions A and B are given by
" T
"
1 2 T
2
B(s, T ) ds
(s)B(s, T )ds,
(11.51)
A( , T ) =
2
and

1
1 e(T ) .
(11.52)
By reiterating the same reasoning produced to show (11.49), one shows that the solution for
is:
2
( ) =
f$ (t, ) + f$ (t, ) +
1 e2( t) .
(11.53)
2
A proof of this result is in Appendix 5.
Why did we need to go for this more complex model? After all, the Ho & Lee model is
already able to pin down the entire yield curve. The answer is that in practice, investment
banks typically prices a large variety of derivatives. The yield curve is not the only thing to be
exactly t. Rather it is only the starting point. In general, the more exible a given perfectly
tting model is, the more successful it is to price more complex derivatives.
B( , T ) =
11.4.4
Critiques
Two important critiques to these models:

- As we shall see in Section 11.6, closed-form solution for options on bond prices are easy
to implement when the short-term rate is Gaussian. We will use the T -forward probability
machinery to show this. In principle, they could also be used to price caps, oors and swaptions.
But in general, no-closed form solutions are available that reproduce the standard market
practice. This diculty is overcome by a class of models known as market models that is
built upon the modelling principles of the HJM models examined in Section 11.4.
- Intertemporal inconsistencies: functions have to be re-calibrated every single day. (As Eq.
(11.49) demonstrates, at time t, ( ) depends on the slope of f$ which can change every day.)
This kind of problems is present in HJM-type models
- Stochastic string shocks models.
310
c
by
A. Mele
11.5. The Heath-Jarrow-Morton model
11.5 The Heath-Jarrow-Morton model

11.5.1
Motivation
The bond price representation in Eq. (11.11),

P ( , T ) = e
T
f ( ,)d
all [t, T ],
(11.54)
underlies the modeling approach started by Heath, Jarrow and Morton (1992) (HJM, henceforth). Given Eq. (11.54), this approach takes as a primitive the -stochastic evolution of the
entire structure of forward rates, not only the special case of the short-term rate, r (t) =
limt f (t, ) f (t, t). So given Eq. (11.54) and the initial, observed structure of forward rates
{f (t, )}[t,T ] , no-arbitrage cross-equations restrictions determine the stochastic behavior of
{f ( , )} (t,] for any [t, T ].
By construction, the HJM approach allows for a perfect fit of the initial term-structure. This
point can be illustrated quite simply, as the bond price P ( , T ) is,
T
P ( , T ) = e f ( ,)d
P (t, T ) P (t, ) T f ( ,)d
=
e
P (t, ) P (t, T )
P (t, T ) f (t,)d+ T f (t,)d T f ( ,)d
t
=
e t
P (t, )
P (t, T ) T f (t,)d T f ( ,)d
e
=
P (t, )
P (t, T ) T [f ( ,)f (t,)]d
e
=
.
P (t, )
The key point of the HJM methodology is to take the current forward rates structure f (t, ) as
given, and to model the future forward rate movements,
f ( , ) f (t, ).
Therefore, the HJM methodology takes the current term-structure as given and, hence, perfectly
tted, as we we observe both P (t, T ) and P (t, ). In contrast, the initial approach to interest
rate modeling is to model the current bond price P (t, T ) through a model for the short-term
rate (as illustrated in Section 11.3), which for this reason, does not t the initial term structure.
As explained in the previous section, tting the initial term-structure is an important issue when
the objective is to price interest-rate derivatives.
11.5.2
The model
11.5.2.1 Primitives
Because the primitive is still a Brownian information structure, once we want to model future
movements of {f ( , T )} [t,T ] , we also have to accept that for every T , {f ( , T )} [t,T ] is F ( )adapted. There thus exist functionals and such that, for a given T ,
d f ( , T ) = ( , T ) d + ( , T ) dW ( ) , (t, T ],
311
(11.55)
c
by
A. Mele

where f (t, T ) is given. The solution to Eq. (11.55) is:
"
"
f ( , T ) = f (t, T ) +
(s, T )ds +
(s, T )dW (s), (t, T ].
t
(11.56)
In other terms, W doesnt depend on T . In some sense, however, we may also want to index
W by T . The so-called stochastic string models are capable of doing that, and are discussed in
Section 11.7.
11.5.2.2 Arbitrage restrictions
The next step is to derive #restrictions on that are consistent with absence of arbitrage opT
portunities. Let X( ) f ( , )d. We have
dX( ) = f ( , )d
"
where
I
( , T )

(d f ( , )) d = r( ) I ( , T ) d I ( , T )dW ( ),
"
( , )d,
( , T )
"
( , )d.
By Eq. (11.54), P = eX . By Itos lemma,

:2
1:
d P ( , T )
I
I
= r( ) ( , T ) + : ( , T ): d I ( , T )dW ( ).
P ( , T )
2
By the FTAP, there are no arbitrage opportunties if and only if

:2
1:
d P ( , T )
I
I
I
:
:
( ),
= r( ) ( , T ) +
( , T ) + ( , T )( ) d I ( , T )dW
P ( , T )
2
#
( ) = W ( ) + (s)ds is a Q-Brownian motion, and satises:
where W
t
:
1:
: I ( , T ):2 + I ( , T )( ).
2
I ( , T ) =
(11.57)
By dierentiating the previous relation with respect to T gives us the arbitrage restriction that
we were looking for:
" T
( , T ) = ( , T )
( , ) d + ( , T )( ).
(11.58)
11.5.3
The dynamics of the short-term rate
By Eq. (11.56), the short-term rate satises:

"
"
r( ) f( , ) = f (t, ) +
(s, )ds +
t
Dierentiating with respect to yields

"
"
dr( ) = f2 (t, ) + ( , )( ) +
2 (s, )ds +
t
312
(s, )dW (s), (t, T ].
(11.59)

2 (s, )dW (s) d + ( , )dW ( ),
c
by
A. Mele

where
2 (s, ) = 2 (s, )
"
(s, ) d + (s, )(s, ) + 2 (s, )(s).
As is clear, the short-term rate is in general non-Markov. However, the short-term rate can be
risk-neutralized, and used to price exotics through simulations. A special case of Eq. (11.59)
is the Ho and Lee model, where
(s, ) = , a constant, such that, by Eq. (11.58), (s, ) = ( s) + ( ).
11.5.4
Embedding
At rst glance, it might be guessed that HJM models are quite distinct from the models of the
short-term rate introduced in Section 11.3. However, there exist embeddability conditions
turning HJM into short-term rate models, and viceversa, a property known as universality
of HJM models.
11.5.4.1 Markovianity
One natural question to ask is whether there are conditions under which HJM-type models
predict the short-term rate to be a Markov process. The question is natural insofar as it relates to the early literature in which the entire term-structure was driven by a scalar Markov
process representing the dynamics of the short-term rate. The answer to this question is in
the contribution of Carverhill (1994). Another important contribution in this area is due to
Ritchken and Sankarasubramanian (1995), who studied conditions under which it is possible to
enlarge the original state vector in such a manner that the resulting augmented state vector
is Markov and at the same time, includes that short-term rate as a component. The resulting
model resembles a lot some of the short-term rate models surveyed in Section 11.3. In these
models, the short-term rate is not Markov, yet it is part of a system that is Markov. Here we
only consider the simple Markov scalar case.
Assume the forward-rate volatility structure is deterministic and takes the following form:
(t, T ) = g1 (t)g2 (T ) all t, T .
(11.60)
By Eq. (11.59), r is then:

r( ) = f (t, ) +
"
(s, )ds + g2 ( )
"
g1 (s)dW (s), (t, T ],
Also, r is solution to:

"
dr( ) = f2 (t, ) + ( , )( ) +

"
= f2 (t, ) + ( , )( ) +
2 (s, )ds + g2 ( )
2 (s, )ds +

"
= f2 (t, ) + ( , )( ) +
+ ( , )dW ( ).
2 (s, )ds +
g2 ( )
g2 ( )
"

g1 (s)dW (s) d + ( , )dW ( )
g2 ( )
"
g1 (s)dW (s) d + ( , )dW ( )

"
g2 ( )
r( ) f(t, )
(s, )ds d
g2 ( )
t
313
c
by
A. Mele
11.6. Stochastic string shocks models
Done. This is Markov. Precisely, the condition in Eq. (11.60) ensures the HJM model predicts
the short-term rate is Markov. Mean reversion, then, obtains assuming that g2 (T ) < 0 for all
T . For example, take to be a constant, and:
g1 (t) = et ,
g2 (t) = et ,
> 0,
0.
This is the Hull-White model discussed in Section 11.3, and of course, the Ho and Lee model
obtains in the special case = 0.
11.5.4.2 Short-term rate reductions
We prove everything in the Markov case. Let the short-term rate be solution to:
( ),
dr( ) = b( , r( ))d + a( , r( ))dW
is a Q-Brownian motion, and b is some risk-neutralized drift function. The rational
where W
!

tT r( )d
. The forward rate implied by this model
bond price function is P (r(t), t, T ) = E e
is:
f (r(t), t, T ) =
ln P (r(t), t, T ).
T
By Itos lemma,

1 2
.
df =
f + bfr + a frr d + afr dW
t
2
But for f(r, t, T ) to be consistent with the solution to Eq. (11.56), it must be the case that
1
f (r, t, T ) + b(t, r)fr (r, t, T ) + a(t, r)2 frr (r, t, T )

t
2
= a(t, r)fr (t, r)
(t, T ) (t, T )(t) =

(t, T )
(11.61)
and
f (t, T ) = f(r, t, T ).
(11.62)
In particular, the last condition can only be satised if the short-term rate model under consideration is of the perfectly tting type.
11.6 Stochastic string shocks models

The rst papers are Kennedy (1994, 1997), Goldstein (2000) and Santa-Clara and Sornette
(2001). Heaney and Cheng (1984) are also very useful to read.
11.6.1
Addressing stochastic singularity
Let ( , T ) = [ 1 ( , T ) , , N ( , T )] in Eq. (11.55). For any T1 < T2 ,

E [df ( , T1 ) df ( , T2 )] =
N
i ( , T1 ) i ( , T2 ) d ,
i=1
and,
c ( , T1 , T2 ) corr [df ( , T1 ) df ( , T2 )] =
314
N
i ( , T1 ) i ( , T2 )
.
( , T1 ) ( , T2 )
i=1
(11.63)
c
by
A. Mele
11.6. Stochastic string shocks models

By replacing this result into Eq. (11.58),
( , T ) =
"
( , T ) ( , ) d + ( , T )( )
"
( , ) ( , T ) c ( , , T ) d + ( , T )( ).
One drawback of this model is that the correlation matrix of any (N + M )-dimensional vector
of forward rates is degenerate for M 1. Stochastic string models overcome this diculty by
modeling in an independent way the correlation structure c ( , 1 , 2 ) for all 1 and 2 rather
than implying it from a given N-factor model (as in Eq. (11.63)). In other terms, the HJM
methodology uses functions i to accommodate both volatility and correlation structure of
forward rates. This is unlikely to be a good model in practice. As we will now see, stochastic
string models have two separate functions with which to model volatility and correlation.
The starting point is a model in which the forward rate is solution to,
d f ( , T ) = ( , T ) d + ( , T ) d Z ( , T ) ,
where the string Z satises the following ve properties:
(i) For all , Z ( , T ) is continuous in T ;
(ii) For all T , Z ( , T ) is continuous in ;
(iii) Z ( , T ) is a -martingale and, hence, a local martingale i.e. E [d Z ( , T )] = 0;
(iv) var [d Z ( , T )] = d ;
(v) cov [d Z ( , T1 ) d Z ( , T2 )] = (T1 , T2 ) (say).
Properties (iii), (iv) and (v) make Z Markovian. The functional form for is crucially important to guarantee this property. Given the previous properties, we can deduce a key property
of the forward rates. We have,

var [df ( , T )] = ( , T )
c ( , T1 , T2 ) corr [df ( , T1 ) df ( , T2 )] =
( , T1 ) ( , T2 ) (T1 , T2 )
= (T1 , T2 )
( , T1 ) ( , T2 )
As claimed before, we now have two separate functions with which to model volatility and
correlation.
11.6.2
No-arbitrage restrictions
Similarly as in the HJM-Brownian case, let X ( )

dX ( ) = f ( , ) d
"
#T
f ( , ) d. We have,

d f ( , ) d = r ( ) d I ( , T ) d
315
"
[ ( , ) d Z ( , )] d,
c
by
A. Mele
11.7. Interest rate derivatives

where as usual, I ( , T )
#T
( , ) d. But P ( , T ) = exp (X ( )). Therefore,
dP ( , T )
1
= dX ( ) + var [dX ( )]
P ( , T )
2

" "
1 T T
I
( , 1 ) ( , 2 ) (1 , 2 ) d1 d2 d
= r ( ) ( , T ) +
2
" T
[ ( , ) d Z ( , )] d.
Next, suppose that the pricing kernel satises:

"
d ( )
= r ( ) d ( , T ) d Z ( , T ) dT,
( )
T
where T denotes the set of all risks spanned by the string Z, and is the corresponding
family of unit risk-premia.
By absence of arbitrage opportunities,

dP
d
dP d
0 = E [d (P )] = E P drift
+ drift
+ cov
,
.
P
P
By exploiting the dynamics of P and ,

" "
1 T T
dP d
I
,
,
( , T ) =
( , 1 ) ( , 2 ) (1 , 2 ) d1 d2 + cov
2
P
where
cov
dP d
,
P
= E
"
"
"
( , S) d Z ( , S) dS
"

( , ) d Z ( , ) d
( , S) ( , ) (S, ) dSd.
By dierentiating I with respect to T we obtain,

" T
"
( , ) ( , T ) (, T ) d + ( , T ) ( , S) (S, T ) dS.
( , T ) =
(11.64)
A proof of Eq. (11.64) is in the Appendix.
11.7 Interest rate derivatives

11.7.1
Introduction
Options on bonds, caps and swaptions are the main interest rate derivatives traded in the
market. The purpose of this section is to price these assets. In principle, the pricing problem
could be solved very elegantly. Let w denote the value of any of such instrument, and be the
instantaneous payo process paid by it. Consider any model of the short-term rate considered
in Section 11.3. To simplify, assume that d = 1, and that all uncertainty is subsumed by the
316
c
by
A. Mele
short-term rate process in Eq. (11.24). By the FTAP, w is then the solution to the following
partial dierential equation:
w
1
+ bwr + a2 wrr + rw, for all (r, ) R++ [t, T )
(11.65)
2
subject to some appropriate boundary conditions. In the previous PDE, b is some risk-neutralized
drift function of the short-term rate. The additional term arises because to the average instantaneous increase rate of the derivative, viz w
+ bwr + 12 a2 wrr , one has to add its payo . The
sum of these two terms must equal rw to avoid arbitrage opportunities. In many applications
considered below, the payo can be approximated by a function of the short-term rate itself
(r). However, such an approximation is at odds with standard practice. Market participants
dene the payos of interest-rate derivatives in terms of LIBOR discretely-compounded rates.
The aim of this section is to present more models that are more realistic than those emananating
from Eq. (11.65).
The next section introduces notation to cope expeditiously with the pricing of these interest
rate derivatives. Section 11.7.3 shows how to price options within the Gaussian models discussed
in Section 11.3. Section 11.7.4 provides precise denitions of the remaining most important
xed-income instruments: xed coupon bonds, oating rate bonds, interest rate swaps, caps,
oors and swaptions. It also provides exact solutions based on short-term rate models. Finally,
Section 11.7.5 presents the market model, which is a HJM-style model intensively used by
practitioners.
0=
11.7.2
Notation
We introduce notation that will prove useful to price interest rate derivatives. For a given
non-decreasing sequence of dates {Ti }i=0,1, , we set,
Fi ( ) F ( , Ti , Ti+1 ).
(11.66)
That is, Fi ( ) is solution to:

P ( , Ti+1 )
1
=
.
P ( , Ti )
1 + i Fi ( )
An obvious but important relation is
Fi (Ti ) = L(Ti ).
11.7.3
The put-call parity in fixed income markets
Consider the identity,

[K P (T, S)]+ [P (T, S) K]+ + K P (T, S) .
Taking risk-neutral, discounted expectations of both sides of this equation leaves,
#T

+
r( )d
E e t
(K P (T, S))
+
#T

#T
r( )d
r( )d
=E e t
(P (T, S) K) + P (t, T ) K E e t
P (T, S)
#T
+
r( )d
=E e t
(P (T, S) K) + P (t, T ) K P (t, S) ,
317
(11.67)
(11.68)
c
by
A. Mele
where the last equality follows by the same argument leading to Eq. (11.45). Therefore, we have
the put-call parity relation:
Put (t, T ; P (t, S) , K) = Call (t, T ; P (t, S) , K) + P (t, T ) K P (t, S) ,
(11.69)
where Put (t, T ; P (t, S) , K) is the price of a European put written on a zero expiring at time
S, expiring at time T < S, and struck at K, and Call () denotes the corresponding call price.
11.7.4
European options on bonds
Let T be the expiration date of a European call option on a bond and S > T be the expiration
date of the bond. We consider a simple model of the short-term rate with d = 1, and a rational
bond pricing function of the form P ( ) P (r, , S). We also consider a rational option price
function C b ( ) C b (r, , T, S). By the FTAP, there are no arbitrage opportunities if and only
if,
!
T
C b (t) = E e t r( )d (P (r(T ), T, S) K)+ ,
(11.70)
where K is the strike of the option. In terms of PDEs, C b is solution to Eq. (11.65) with 0
and boundary condition C b (r, T, T, S) = (P (r, T, S) K)+ , where P (r, , S) is also the solution
to Eq. (11.65) with 0, but with boundary condition P (r, S, S) = 1. In terms of PDEs, the
situation seems hopeless. As we show below, the problem can considerably be simplied with
the help of the T -forward martingale probability introduced in Section 11.1. In fact, we shall
show that under the assumption that the short-term rate is a Gaussian process, Eq. (11.70) has
a closed-form expression. We now present two models enabling this. The rst one was developed
in a seminal paper by Jamshidian (1989), and the second one is, simply, its perfectly tting
extension.
11.7.4.1 Jamshidian & Vasicek
Suppose that the short-term rate is solution to the Vasiceks model considered in Section 11.3
(see Eq. (11.30)):
,
dr( ) = ( r( ))d + dW
is a Q-Brownian motion and . As shown in Section 11.3, Eq. (11.32), the
where W
bond price is:
P (r( ), , S) = eA( ,S)B( ,S)r( ) ,

for some function A, and for B(t, T ) = 1 1 e(T t) (see Eq. (11.52)).
In Section 11.3, Eq. (11.45), it was also shown that
!
T
E e t r( )d (P (r(T ), T, S) K)+
= P (r(t), t, S) QSF [P (r(T ), T, S) K] KP (r(t), t, T ) QTF [P (r(T ), T, S) K] ,
(11.71)
where QTF denotes the T -forward martingale probability (see Section 11.1.4).
In Appendix 8, we show that the two probabilities in Eq. (11.71) can be evaluated by the
changes of numeraire described in Section 11.1.4, such that the solution for P (r, T, S) is:
P (r, T, S) 1 2 T [B( ,S)B( ,T )]2 d T [B( ,S)B( ,T )]dW QTF ( )
t
e 2 t
under QTF
P (r, t, T )
P (r, T, S) 1 2 T [B( ,S)B( ,T )]2 d T [B( ,S)B( ,T )]dW QSF ( )
t
P (r, T, S) =
e2 t
under QSF
P (r, t, T )
(11.72)
318
P (r, T, S) =
c
by
A. Mele

T
where W QF is a Brownian motion under the forward probability QTF . Therefore, simple algebra
now reveals that:
!
P (r(t),t,S)
ln KP (r(t),t,T ) + 12 v 2
S
T
QF [P (T, S) K] = (d1 ) , QF [P (T, S) K] = (d1 v) , d1 =
,
v
where
2
v =
"
[B( , S) B( , T )]2 d = 2
1 e2(T t)
B(T, S)2 .
2
(11.73)
11.7.4.2 Perfectly tting extension
We now consider the perfectly tting extension of the previous results. Namely, we consider
model (11.50) in Section 11.3, viz
( ),
dr( ) = (( ) r( ))d + dW
where ( ) is now the innite dimensional parameter that is used to invert the term-structure.
The solution to Eq. (11.70) is the same as in the previous section. However, in Section 11.7.3
we shall argue that the advantage of using such a perfectly tting extension arises as soon as
one is concerned with the evaluation of more complex options on xed coupon bonds.
11.7.4.3 Bond price volatility and the persistence of the short-term rate
The implied vol on options on bonds is typically very large, in fact comparable to that on
stocks. Why is it that this implied vol is so large, when in fact, the volatility of the short-term
rate is one order of magnitude less than that on stock markets? The answer is that the shortterm rate is very persistent, and it is a risk for the long-run, pretty much in the same spirit
of the explanations attempting to explain the equity premium puzzle, reviewed in Chapter
7. To make this point precise, dene, rst, the term-structure of volatility. It is the function,
Vol (R ( )), where R ( ) is the spot rate for the maturity , and Vol (R ( )) is the standard
deviation of this spot-rate. By the denition of R ( ), the term-structure of volatility can also
be written as the function

1
Vol ln P ( ) ,
where P ( ) is the price of a zero with maturity equal to . It is instructive to see what this
volatility looks like, for a concrete model. Consider again the Vasicek model. This model assumes
that the short-term rate is solution to,
drt = ( rt ) dt + dWt ,
where Wt is a Brownian motion, and , and are three positive constants. By previous results
given in this chapter, we know that for this model,
R ( ) =
A ( ) 1
+ B ( ) r,
B ( ) =
1 e
.
for some function A ( ). Therefore, we have that,

Vol [R ( )] =
1
B ( ) Vol (r) ,
319
(11.74)
c
by
A. Mele

where Vol (r) is the ergodic volatility of the short-term rate, dened as, Vol (r) = 2 /2.
For example, if = 0.2 and = 0.03, then Vol (r) 4.7%. Given the previous values for
and , the picture below depicts the term-structure of volatility, i.e. Eq. (11.74).
Vol(R)
0.045
0.040
0.035
0.030
0
Maturity (years)
As we can see, the term-structure of volatility is decreasing in the maturity of the zero, and
attains its maximum at Vol (r) 4.7%.
Despite this, the volatility of bond returns can be much higher, as we now illustrate. We need
to gure out what the dynamics of the bond price are, for the Vasicek model. By Itos lemma,
dP ( )
= [ ] dt + [ B ( )] dWt
P ( )
Therefore, the volatility of bond returns is,

dP
Vol
P
= B ( ) .
(11.75)
Compare Eq. (11.75) with Eq. (11.74). The main dierence between the two equations is
that the right hand side of Eq. (11.74) is divided by , which makes Vol [R ( )] decreasing in .
(Otherwise, Vol (r) and have roughly the same order of magnitude.) The point is, indeed,
that the yield, R ( ), is simply an average return which we obtain were we to decide not to sell
the bond until its expiry. This average return is, of course, progressively less volatile as time to
is, instead, measuring
maturity gets large and it becomes a constant, eventually. The return dP
P
the capital gains we may obtain by trading the bond, and tends to be more and more volatile
as time
to maturity gets large. Indeed, even if is very small, the volatility
dP of bond return,
Vol dP
,
can
be
quite
high.
For
example,
if
is
close
to
zero,
then,
Vol
, which
P
P
is 15% for a 5Y zero. This fact is illustrated by the next picture, which depicts Eq. (11.75),
evaluated at the previous parameter values, = 0.2 and = 0.03.
320
c
by
A. Mele
Vol(dP/P)
0.12
0.10
0.08
0.06
0.04
0.02
0.00
0
Maturity (years)
Intuitively, it is the high persistence of the short-term rate (measured by the low value of )
to make the bond price so volatile in correspondence of large maturity dates. High persistence
in the short-term rate means that a shock in the short-term rate, is permanently embedded
in the future path of the short-term rate, or it has persistent consequences. This makes the
short-term rate very volatile in the long-run, which makes the value of the long maturity zero
very volatile as well. These facts can be seen at work analyzing the option-based
volatility in
Eq. (11.73). In that case, we have that as gets small, v tends to T t (S T ), so

that the implied vol equals (S T ), which increases with the bonds time to maturity left
at its expiration, S T .
The previous reasoning does, of course, still hold in the more realistic case of a three-factor
model, such as that in Eqs. (11.36). In that case, as explained, r is large and r is small:
the short-term rate is quite persistent because it mean-reverts, quickly, to a persistent process,
which we denoted as r ( ). Naturally, in such as a three-factor model, Eq. (11.75) does not hold
anymore, as we should add two more volatility components, related to stochastic volatility,
v ( ), and the persistent process r ( ). However, the bond return volatility would be boosted
by the high persistence of r ( ).
11.7.5
Related pricing problems
11.7.5.1 Fixed coupon bonds
Given a set of dates {Ti }ni=0 , a xed coupon bond pays o a xed coupon ci at Ti , i = 1, , n
and one unit of numeraire at time Tn . Ideally, one generic coupon at time Ti pays o for the
time-interval Ti Ti1 . It is assumed that the various coupons are known at time t < T0 . By
the FTAP, the value of a xed coupon bond is
Pfcb (t, Tn ) = P (t, Tn ) +
n

i=1
321
ci P (t, Ti ) .
c
by
A. Mele

11.7.5.2 Floating rate bonds
A oating rate bond works as a xed coupon bond, with the important exception that the
coupon payments are dened as:
1
ci = i1 L (Ti1 ) =
1,
(11.76)
P (Ti1 , Ti )
where i Ti+1 Ti , and where the second equality is the denition of the simply-compounded
LIBOR rates introduced in Section 11.1 (see Eq. (11.2)). By the FTAP, the price pfrb as of time
t of a oating rate bond is:
n
!

T
t i r( )d
pfrb (t) = P (t, Tn ) +
E e
i1 L(Ti1 )
= P (t, Tn ) +
= P (t, Tn ) +
i=1
n

i=1
n

i=1
= P (t, T0 ).
.
Ti
n

e t r( )d
E
P (t, Ti )
P (Ti1 , Ti )
i=1
P (t, Ti1 )
n
P (t, Ti )
i=1
where the second line follows from Eq. (11.76) and the third line from Eq. (11.8) given in Section
11.1.
The same result can be obtained by assuming an economy in which the oating rates continuously pay o the instantaneous short-term rate r. Let T0 = t for simplicity. In this case, pfrb
is solution to the PDE (11.65), with (r) = r, and boundary condition pfrb (T ) = 1. As it can
veried, pfrb = 1, all r and , is indeed solution to the PDE (11.65).
11.7.5.3 Options on xed coupon bonds
The payo of an option maturing at T0 on a xed coupon bond paying o at dates T1 , , Tn

is given by:
.+
n

[Pfcb (T0 , Tn ) K]+ = P (T0 , Tn ) +
ci P (T0 , Ti ) K .
(11.77)
i=1
At rst glance, the expectation of the payo in Eq. (11.77) seems very dicult to evaluate.
Indeed, even if we end up with a model that predicts bond prices at time T0 , P (T0 , Ti ), to be
lognormal, we know that the sum of lognormals is not lognormal. This issue can be dealt with
in an elegant manner. Suppose we wish to model the bond price P (t, T ) through any one of
the models of the short-term rate reviewed in Section 11.3. In this case, the pricing function is
obviously P (t, T ) = P (r, t, T ). Assume, further, that
P (r, t, T )
< 0,
r
For all t, T,
(11.78)
and that
For all t, T, lim P (r, t, T ) > K and
r0
lim P (r, t, T ) = 0.
(11.79)
Under conditions (11.78) and (11.79), there is one and only one value of r, say r , that solves
the following equation:
n
P (r , T0 , Tn ) +
ci P (r , T0 , Ti ) = K.
(11.80)
i=1
322
c
by
A. Mele

Then, the payo in Eq. (11.77) can be written as:
- n

i=1
ci P (r(T0 ), T0 , Ti ) K
.+
n

i=1
ci (P (r(T0 ), T0 , Ti ) P (r , T0 , Ti ))
.+
where ci = ci , i = 1, , n 1, and cn = 1 + cn .
Next, note that by condition (11.78), the terms P (r(T0 ), T0 , Ti ) P (r , T0 , Ti ) have the same
sign for all i.14 Therefore, the payo in Eq. (11.77) is,
-
n

i=1
ci P (r(T0 ), T0 , Ti ) K
.+
n

i=1
ci [P (r(T0 ), T0 , Ti ) P (r , T0 , Ti )]+ .
(11.81)
Next, note that each term of the sum in Eq. (11.81) can be evaluated as an option on a
pure discount bond with strike price equal to P (r , T0 , Ti ). Typically, the threshold r must be
found with some numerical method. The device to reduce the problem of an option on a xed
coupon bond to a problem involving the sum of options on zero coupon bonds was invented by
Jamshidian (1989).15 The price of the call on the xed coupon bond is, therefore,
Call (t, T0 ; Pfcb (t, Tn ) , K, v) =
n
ci Call (t, T0 ; P (t, Ti ) , P (r , T0 , Ti ) , vi ) ,
(11.82)
i=1
where r solves Eq. (11.80), and,
d1,i
Call (t, T0 ; Pi , Pi , vi ) = Pi (d1,i ) Pi P (t, T0 ) (d1,i vi ) ,

Pi
ln P P (t,T
+ 1 v2

0) 2 i
1e2(T0 t)
i
=
,
v
=
B (T0 , Ti ) , B (t, T ) = 1 1 e(T t) .

i
vi
2
Why are perfectly tting models so important, in practice? Suppose that in Eq. (11.80), the
critical value r is computed by means of the Vasiceks model. This assumption is attractive
because it allows to evaluate the payo in Eq. (11.81) with the Jamshidians formula of Section
11.7.2. However, this way to proceed does not ensure that the yield curve is perfectly tted.
The natural alternative is to use the corresponding perfectly tting extension, as in Eq. (11.82).
However, such a perfectly tting extension gives rise to a zero-coupon bond option price that
is perfectly equal to the one that can be obtained through the Jamshidians formula. However,
things dier as far as options on zero coupon bonds are concerned. Indeed, by using the perfectly
tting model (11.50), one obtains bond prices such that the solution r in Eq. (11.80) is radically
dierent from the one obtained when bond prices are obtained with the simple Vasiceks model.
11.7.5.4 Interest rate swaps
An interest rate swap is an exchange of interest rate payments. Typically, one counterparty
exchanges a xed against a oating interest rate payment. For example, the counterparty receiving a oating interest rate payment has good (or only) access to markets for variable
14 Suppose
that P (r(T0 ), T0 , T1 ) > P (r , T0 , T1 ). By Eq. (11.78), r(T0 ) < r . Hence P (r(T0 ), T0 , T2 ) > P (r , T0 , T2 ), etc.
conditions in Eqs. (11.78) and (11.79) hold, within the Vasiceks model that Jamshidian considered in his paper. In fact,
the condition in Eq. (11.78) holds for all one-factor stationary, Markov models of the short-term rate. However, the condition in
Eq. (11.78) is not a general property of bond prices in multi-factor models (see Mele (2003)).
15 The
323
c
by
A. Mele
interest rates, but wishes to pay xed interest rates. And viceversa. The counterparty receiving
a oating interest rate payment and paying a xed interest rate Kirs has a payo equal to,
i1 [L(Ti1 ) Kirs ]
at time Ti , i = 1, , n. It is a FRA. By convention, we say that the swap payer is the
counterparty who pays the xed interest rate Kirs , and that the swap receiver is the counterparty
who receives the xed interest rate Kirs .
By the FTAP, the value pirs as of time t of an interest rate swap payer is:
pirs (t) =
n
E e
i=1
Ti
t
r( )d
i1 (L(Ti1 ) Kirs ) =
n
IRS(t, Ti1 , Ti ; Kirs ),
(11.83)
i=1
where IRS is the value of a forward-rate agreement and is, by Eq. (11.9) in Section 11.1,
IRS (t, Ti1 , Ti ; Kirs ) = i1 [F (t, Ti1 , Ti ) Kirs ] P (t, Ti ) .
The forward swap rate Rswap is the value of Kirs such that pirs (t) = 0. Simple computations
yield:
n
i1 F (t, Ti1 , Ti )P (t, Ti )
P (t, T0 ) P (t, Tn )
Rswap (t) = i=1 n
= n
,
(11.84)
i=1 i1 P (t, Ti )
i=1 i1 P (t, Ti )
where the last equality is due to the denition of F (t, Ti1 , Ti ) given in Section 11.1, Eq.
(11.67).16 This expression is quite similar to the par coupon rate in Eq. (11.4).
So by plugging the expression for the forward swap rate into Eq. (11.83), we obtain the
following intuitive expression for the swap payer:
pirs (t) =
=
n

i=1
n

i=1
i1 F (t, Ti1 , Ti ) P (t, Ti ) Kirs
n
i1 P (t, Ti )
i=1
i1 P (t, Ti ) (Rswap (t) Kirs )
PVBPt (T1 , , Tn ) (Rswap (t) Kirs ) ,
(11.85)
where PVBPT (T1 , , Tn ) is the so-called swaps Present Value of the Basis Point (see, e.g.,
Brigo and Mercurio, 2006), i.e. the present value impact of one basis point move in the forward
swap rate at T .
11.7.5.5 Caps & oors
A cap works as an interest rate swap, with the important exception that the exchange of interest
rates payments takes place only if actual interest rates are higher than K. A cap protects against
upward movements of the interest rates. Therefore, we have that the payo as of time Ti is
i1 [L (Ti1 ) K]+ ,
i = 1, , n.
Such a payo is usually referred to as the caplet.

16 To cast this problem in terms of continuous time swap exchanges and, then, PDEs, we set p (T ) 0 as a boundary condition,
irs
and (r) = r k, where k plays the same role as Kirs above. Then, if the bond price P ( ) is solution to Eq. (11.65), the following
T
function, pirs ( ) = 1 P ( ) k P (s)ds, does also satisfy Eq. (11.65), with (r) = r k.
324
c
by
A. Mele

Floors are dened in a similar way, with a single floorlet paying o,
i1 [K L(Ti1 )]+
at time Ti , i = 1, , n.
We will only focus on caps. By the FTAP, the value pcap of a cap as of time t is:
pcap (t) =
n
E e
i=1
Ti
t
r( )d
!
i1 (L(Ti1 ) K)+ .
(11.86)
We can develop explicit solutions to this problem, relying upon models of the short-term
rate. First, we use the standard denition of simply compounded rates given in Section 11.1
1
(see formula (11.2)), viz i1 L(Ti1 ) = P (Ti1
1, and rewrite the caplet payo as follows:
,Ti )
[ i1 L(Ti1 ) i1 K]+ =
1
P (Ti1 , Ti )
[1 (1 + i1 K)P (Ti1 , Ti )]+ .
We have,
.
Ti
e t r( )d
pcap (t) =
E
(1 (1 + i1 K)P (Ti1 , Ti ))+
P
(T
,
T
)
i1
i
i=1

n

T
+
t i1 r( )d 1
E e
(Ki P (Ti1 , Ti )) , Ki = (1 + i1 K)1 ,
=
K
i
i=1
n
(11.87)
where the last equality follows by a simple computation.17 If bond prices are as in Jamshidian
or in Hull and White, the cap price in Eq. (11.87) can be expressed in closed-form. Indeed, as
Eq. (11.87) makes clear, a cap is a basket of puts on zero coupon bonds, with strikes Ki , and
can be priced through the models in Sections 11.7.4.1 and 11.7.4.2. We have:
n

1
pcap (t) =
Put (t, Ti1 ; P (t, Ti ) , Ki , v) ,
Ki
i=1
(11.88)
where Put () satises the put-call parity in Eq. (11.69), and, by the pricing formulae in Section
11.7.4.1,
d1,i
17 By
Call (t, Ti1 ; P (t, Ti ) , Ki , v) = P (t, Ti ) (d1,i ) Ki P (t, Ti1 ) (d1,i v) ,

P (t,Ti )
ln
+ 1 v2

Ki1 P (t,Ti1 ) 2
1e2(Ti1 t)
=
,
v
=
B (Ti1 , Ti ) , B (t, T ) = 1 1 e(T t) .

v
2
(11.89)
the law of iterated expectations,
T
T

t i r( )d
t i r( )d

e
e
+
+
E
[1 (1 + i1 K)P (Ti1 , Ti )] = E E
(1 (1 + i1 K)P (Ti1 , Ti )) F (Ti )
P (Ti1 , Ti )
P (Ti1 , Ti )

T

T
i
+
t i r( )d Ti1 r( )d
=E E e
e
(1 (1 + i1 K)P (Ti1 , Ti )) F (Ti )

Ti1

r()d
= E E e t
(1 (1 + i1 K)P (Ti1 , Ti ))+ F (Ti )
T
i1
r()d
= E e t
(1 (1 + i1 K)P (Ti1 , Ti ))+
325
c
by
A. Mele
Naturally, caps on interest rates, which are nothing but baskets of calls, are portfolios of puts
on xed coupon bonds, due to the inverse relation between prices and interest rates.
Finally, note that we could also price caps and oors through the PDE (11.65), after setting
(r) = (r k)+ (caps) and (r) = (k r)+ (oors), and where k plays the same role played
by K above. However, this type of contracts, where payos are continuous, is highly stylized,
and does not obviously exist in the markets.
11.7.5.6 Swaptions
Swaptions are options to enter a swap contract on a future date. Let the maturity date of this
option be T0 . Then, at time T0 , the payo for a payer swaption is the maximum between zero
and the value of a payer interest rate swap at T0 , pirs (T0 ), viz
- n
.+ - n
.+

IRS (T0 , Ti1 , Ti ; Kirs ) =
i1 (F (T0 , Ti1 , Ti ) Kirs ) P (T0 , Ti ) .
(pirs (T0 ))+ =
i=1
i=1
(11.90)
By the FTAP, the value of the payer swaption at time t is:

$ n
%+ .
T0

pswaption (t) = E e t r( )d
i1 (F (T0 , Ti1 , Ti ) Kirs ) P (T0 , Ti )
-
= E e
T0
t
r( )d
i=1
1 P (T0 , Tn )
n
ci P (T0 , Ti )
i=1
%+ .
(11.91)
i1 )
where ci i1 Kirs , and where we used the relation i1 F (T0 , Ti1 , Ti ) = PP(T(T0 ,T
1.
0 ,Ti )
Eq. (11.91) is the expression of the price of a put option on a fixed coupon bond struck at one.
Therefore, we can price this contract in closed-form, through the models in Section 11.7.4.1 and
11.7.4.2, similarly to that we did in the previous section for caps pricing. We have,
pswaption (t) = Put (t, T0 ; Pfcb (t, Tn ) , Kirs , v) ,

where Put () satises the put-call parity in Eq. (11.69). By the pricing formulae in Section
11.7.4.1,
n

Call (t, T0 ; Pfcb (t, Tn ) , Kirs , v) =
ci Call (t, T0 ; P (t, Ti ) , Pi , v) ,
i=1
where ci = i1 Kirs , for i = 1, , n 1 and cn = 1 + n1 Kirs , and Call (t, T0 ; P (t, Ti ) , Pi , v)

is as in Eq. (11.89), with Pi = P (r , T0 , Ti ), and r solution to Eq. (11.80) for K = 1.
11.7.6
Market models
11.7.6.1 Models and market practice
As illustrated in the previous sections, models of the short-term rate can be used to obtain
closed-form solutions of virtually every important interest rate derivative product. The typical
examples are the Vasiceks model and its perfectly tting extension. Yet practitioners evaluate
caps through the Blacks (1976) formula. The assumption underlying the market practice is that
the simply-compounded forward rate is lognormally distributed. As it turns out, the analytically
tractable (Gaussian) short-term rate models are not consistent with this assumption. Clearly,
326
c
by
A. Mele
the (Gaussian) Vasiceks model does not predict that the simply-compounded forward rates are
Geometric Brownian motions.18
Can a non-Markovian HJM model address this problem? Yes. However, a practical diculty
arising with the HJM approach is that instantaneous forward rates are not observed. Does this
compromise the practical appeal of the HJM methodology to the pricing of caps and oors
- which constitute an important portion of the interest rate derivative markets? No. Brace,
Gatarek and Musiela (1997), Jamshidian (1997) and Miltersen, Sandmann and Sondermann
(1997) observed that the general HJM framework can be somehow forced to address some of
the previous diculties.
The key feature of the models identied by these authors is the emphasis on the dynamics of
the simply-compounded forward rates. An additional, and technical, assumption is that these
simply-compounded forward rates are lognormal under the risk-neutral probability Q. That is,
given a non-decreasing sequence of reset times {Ti }i=0,1, , each simply-compounded rate, Fi , is
solution to the following stochastic dierential equation:19
dFi ( )
( ), [t, Ti ] , i = 0, , n 1,
= mi ( )d + i ( )dW
Fi ( )
(11.92)
where Fi ( ) F ( , Ti , Ti+1 ), and mi and i are some deterministic functions of time ( i is

vector valued). On a mathematical point of view, that assumption that Fi follows Eq. (11.92)
is innocous.20
As we shall show, this simple framework can be used to use the simple Blacks (1976) formula
to price caps and oors. However, we need to emphasize that there is nothing wrong with the
short-term rate models analyzed in previous sections. The real advance of the so-called market
model is to give a rigorous foundation to the standard market practice to price caps and oors
by means of the Blacks (1976) formula.
11.7.6.2 Simply-compounded forward rate dynamics, and no-arb restrictions
By the denition of the simply-compounded forward rates in Eq. (11.67),

P ( , Ti )
ln
= ln [1 + i Fi ( )] .
P ( , Ti+1 )
(11.93)
The logic we follow, now, is the same as that underlying the HJM representation of Section
11.4. We wish to express the volatility of bond prices in terms of the volatility of forward rates.
To achieve this task, we rst assume that bond prices are driven by Brownian motions and
expand the l.h.s. of Eq. (11.93) (step 1). Then, we expand the r.h.s. of Eq. (11.93) (step 2).
Finally, we identify the two diusion terms derived from the previous two steps (step 3).
Step 1: Let Pi P ( , Ti ), and assume that under the risk-neutral probability Q, Pi is solution
to:
18 Indeed,
dPi
.
= rd + bi dW
Pi
1 + i Fi ( ) =
P ( ,Ti )
P (,Ti+1 )
= exp [Ai ( ) Bi ( ) r ( )], where Ai ( ) = A ( , Ti ) A ( , Ti+1 ), and Bi ( ) =
B ( , Ti ) B ( , Ti+1 ). Hence, Fi ( ) is not a Geometric Brownian motion, despite the fact that the short-term rate r is Gaussian
and, hence, the bond price is log-normal. Black 76 can not be applied in this context.
19 Brace, Gatarek and Musiela (1997) derived their model by specifying the dynamics of the spot simply-compounded Libor
interest rates. Since Fi (Ti ) = L(Ti ) (see Eq. (11.68)), the two derivations are essentially the same.
20 It is well-known that lognormal instantaneous forward rates create mathematical problems to the money market account (see, for
example, Sandmann and Sondermann (1997) for a succinct overview on how this problem is easily handled with simply-compounded
forward rates).
327
c
by
A. Mele

In terms of the HJM framework in Section 11.4,
I
bi ( ) = ( , Ti ) =
"
Ti
( , )d,
(11.94)
where ( , ) is the instantaneous volatility of the instantaneous -forward rate as of time

. By Itos lemma,

1
P ( , Ti )
.
= bi 2 b,i+1 2 d + ( bi b,i+1 ) dW
d ln
P ( , Ti+1 )
2

(11.95)
Step 2: Applying It
os lemma to ln [1 + i Fi ( )], and using Eq. (11.92), yields:
i
1
2i
dFi
(dFi )2
2
1 + i Fi
2 (1 + i Fi )
.
i mi Fi
1 2i Fi2 i 2
i Fi
.
=
d
+
dW
1 + i Fi 2 (1 + i Fi )2
1 + i Fi i
d ln [1 + i Fi ( )] =
(11.96)
Step 3: By Eq. (11.93), the diusion terms in Eqs. (11.95) and (11.96) have to be the same.
Therefore,
bi ( ) b,i+1 ( ) =
i Fi ( )
( ), [t, Ti ] .
1 + i Fi ( ) i
By summing over i, we get the following no-arbitrage restriction applying to the volatility
of the bond prices:
bi ( ) b,0 ( ) =
i1

j=0
j Fj ( )
( ).
1 + j Fj ( ) j
(11.97)
As is clear, Eq. (11.97) is merely a restriction to the general HJM framework. In other
words, assume the instantaneous forward rates are as in Eq. (11.55) of Section 11.4. As we
demonstrated in Section 11.4, then, the bond prices volatility is given by Eq. (11.94). But if we
also assume that simply-compounded forward rates are solution to Eq. (11.92), then, the bond
prices volatility is also equal to Eq. (11.97). Comparing Eq. (11.94) with Eq. (11.97) produces,
"
Ti
T0
( , )d =
i1

j=0
j Fj ( )
( ).
1 + j Fj ( ) j
The practical interest to restrict the forward-rate volatility dynamics in this way lies in the
possibility to obtain closed-form solutions for some of the interest rates derivatives surveyed in
Section 11.7.3.
328
c
by
A. Mele

11.7.6.3 Pricing formulae
Caps & floors
We provide the solution for caps only. We have:

pcap (t) =
n
E e
Ti
E e
Ti
i=1
n

i=1
n

i=1
r( )d
i1 (L(Ti1 ) K)+
r( )d
i1 (F (Ti1 , Ti1 , Ti ) K)+
i1 P (t, Ti ) EQTi [F (Ti1 , Ti1 , Ti ) K]+ ,
(11.98)
where EQTi [] denotes, as usual, the expectation taken under the Ti -forward martingale probaF
bility QTFi ; the rst equality is Eq. (11.86); and the second equality has been obtained through
the usual change of probability technique introduced Section 11.1.4.
The key point is that
Fi1 ( ) Fi1 ( , Ti1 , Ti ), [t, Ti1 ], is a martingale under QTFi .
A proof of this statement is in Section 11.1. By Eq. (11.92), this means that Fi1 ( ) is solution
to:
Ti
dFi1 ( )
= i1 ( ) dW QF ( ) , [t, Ti1 ] , i = 1, , n,
Fi1 ( )
under QTFi . Therefore, the cap price in Eq. (11.98) reduces to that of Black (1976), once we
assume is deterministic:
EQTi [F (Ti1 , Ti1 , Ti ) K]+ = Fi1 (t) (d1,i1 ) K (d1,i1 si ) ,
(11.99)
where
d1,i1
(t)
+ 12 s2i
ln Fi1
K
=
,
si
s2i
"
Ti1
i1 ( )2 d .
For sake of completeness, Appendix 8 provides the derivation of the Blacks formula.
Swaptions
By Eq. (11.85), the payo of a payer swaption expiring at time T0 is:

+
[pirs (T0 )] = PVBPT0 (T1 , , Tn ) (Rswap (T0 ) Kirs ) , PVBPT0 (T1 , , Tn ) =
n

i=1
Therefore, by the FTAP, and a change of measure,

.
n
T

+
t 0 r( )d
pswaption (t) = E e
i1 P (T0 , Ti ) (Rswap (T0 ) Kirs )

= P (t, T0 ) EQT0
F
- i=1
n

i=1
329
i1 P (T0 , Ti ).
c
by
A. Mele
This can be dealt with through the so-called forward swap probability. Dene the forward
swap probability Qswap by:
n
n
i1 P (T0 , Ti )
i1 P (T0 , Ti )
dQswap
i=1
n
=
= P (t, T0 ) i=1
,
n
T0
EQT0 [ i=1 i1 P (T0 , Ti )]
dQF
i=1 i1 P (t, Ti )
F
where
last equality
follows from the following elementary facts: P (t, Ti )
#the

T0
r( )d
P (T0 , Ti ) = P (t, T0 )EQT0 [P (T0 , Ti )]. (The rst equality is the FTAP, the sec=E e t
F
ond is a change of measure.)

Therefore,
pswaption (t) = P (t, T0 ) EQT0

F
- n

i=1
- n

i=1
i1 P (t, Ti ) EQswap (Rswap (T0 ) Kirs )+
PVBPt (T1 , , Tn ) EQswap (Rswap (T0 ) Kirs )+
(11.100)
Furthermore, the forward swap rate Rswap is a Qswap -martingale.21 And naturally, it is positive.
Therefore, it must satisfy:
dRswap ( )
= swap ( ) dWswap ( ) , [t, T0 ] ,
Rswap ( )
(11.101)
where Wswap is a Qswap -Brownian motion, and swap ( ) is adapted.

If the volatility swap ( ) in Eq. (11.101) is deterministic, we can use Black 76 to price the
payer swaption in Eq. (11.100) in closed-form. We have:

pswaption (t) = PVBPt (T1 , , Tn ) Black76(Rswap (t) ; T0 , Kirs , V ),
(11.102)
where Black76 () is given by Blacks (1976) formula:
Black76(Rswap (t) ; T0 , Kirs , V ) = Rswap (t) (dt ) Kirs (dt V ),

dt =
ln
Rswap (t) 1
+2V
Kirs
, V =
Inconsistencies
# T0
t
swap ( )2 d .
An issue is that if F is solution to Eq. (11.92), swap can not be deterministic. And as you may
easily conjecture, if you assume that forward swap rates are lognormal, then you dont end
up with Eq. (11.92). Therefore, you may use Black 76 to price either caps or swaptions, not
both. This limits considerably the importance of market models. A couple of tricks that seem to
work in practice. The best known is based on a suggestion by Rebonato (1998), to replace the
21 By
Eq. (11.84), and one change of measure,

P (t, )EQF [P ( , T0 ) P ( , Tn )]
P ( , T0 ) P ( , Tn )
P (t, T0 ) P (t, Tn )
EQswap [Rswap ( )] = EQswap n
=
n
= n
= Rswap (t).
P
(
,
T
)
P
(t,
T
)
i1
i
i1
i
i=1
i=1
i=1 i1 P (t, Ti )
330
c
by
A. Mele
true pricing problem with an approximating pricing problem in which swap is deterministic.
That works in practice, but in a world with stochastic volatility, we should expect that trick to
generate unstable things in periods experiencing highly volatile volatility. See, also, Rebonato
(1999) for an essay on related issues. The next section suggests to use numerical approximation
based on Montecarlo techniques.
11.7.6.4 Numerical approximations
Suppose forward rates are lognormal. Then you price caps with Black 76. As regards swaptions,
you may wish to implement Montecarlo integration as follows.
By a change of measure,
$ n
%+ .
T0

i1 (F (T0 , Ti1 , Ti ) K) P (T0 , Ti )
pswaption (t) = E e t r( )d
i=1
= P (t, T0 )EQT0
F
- n

i=1
i1 (F (T0 , Ti1 , Ti ) K) P (T0 , Ti )
.+
where F (T0 , Ti1 , Ti ), i = 1, , n, can be simulated under QTF0 .

Details are as follows. We know that
Ti
dFi1 ( )
= i1 ( )dW QF ( ).
Fi1 ( )
(11.103)
By results in Appendix 3, we also know that:

Ti
T0
dW QF ( ) = dW QF ( ) [ bi ( ) b0 ( )] d
i1

T0
j Fj ( )
= dW QF ( ) +
j ( )d ,
1
+
F
(
)
j
j
j=0
where the second line follows from Eq. (11.97) in the main text. Replacing this into Eq. (11.103)
leaves:
i1

T0
dFi1 ( )
j Fj ( )
= i1 ( )
j ( )d + i1 ( )dW QF ( ), i = 1, , n.
Fi1 ( )
1 + j Fj ( )
j=0
These can easily be simulated with the methods described in any standard textbook such as
Kloeden and Platen (1992).
11.7.6.5 Volatility surfaces
Caps & floors
The market practice relies on the models of this section, rather than those of Sections 11.7.4.111.7.4.2, in providing volatility surfaces. In the models of Sections 11.7.4.1-11.7.4.2, volatility
surfaces might be produced, but only indirectly, after calibration of the two parameters and
, as Eq. (11.88) indicates. It is easier, however, to provide volatility surfaces in the rst place,
through the models of this section. Quite simply, traders use Eq. (11.99) and quote volatilities
such that the market price of a cap equals to the value predicted by Eq. (11.99) using the
desired implied volatility si . In Eq. (11.99),

si = Ti1 t (i) ,
331
c
by
A. Mele

for some (i), although traders simply quote the value of i that satises:
n :
p$cap
(t; n) =
n

i=1
i1 P (t, Ti ) Black76 (Fi1 (t) ; K, si,n ) ,
where p$cap (t; n) is the market price of the cap, and:

Black76 (Fi1 (t) ; K, si,n ) = Fi1 (t) dn1,i1 K dn1,i1 si,n ,
dn1,i1 =
ln
Fi1 (t) 1 2
+ 2 si,n
K
si,n
, si,n =
Ti1 t n
Given n, we can bootstrap (i), i.e. we can recursively solve for (i), as follows:
0=
n

i=1
i1 P (t, Ti ) [Black76 (Fi1 (t) ; K, si,n ) Black76 (Fi1 (t) ; K, si )] ,
n = 1, , N,
where N is the latest available maturity, and si = Ti1 t (i). The values of (i) constitute
what is known as the term structure of caps volatilities.
Swaptions
As for swaptions, the situation is much simpler. The market practice is to quote swaptions
through standard implied vols, i.e. those vols IVt such that, once inserted into Eq. (11.102),
delivers the swaption market price:
pswaption (t) = PVBPt (T1 , , Tn ) Black76 (Rswap (t) ; T0 , Kirs , IVt ) .
332
c
by
A. Mele
11.8. Appendix 1: The FTAP for bond prices
11.8 Appendix 1: The FTAP for bond prices

Suppose there exist m pure discount bond prices {{Pi P ( , Ti )}m
i=1 } [t,T ] satisfying:
dPi
= bi d + bi dW, i = 1, , m,
Pi
(11A.1)
where W is a Brownian motion in Rd , and bi and bi are progressively F( )-measurable functions

guaranteeing the existence of a strong solution to the previous system ( bi is vector-valued). The value
process V of a self-nancing portfolio in these m bonds and a money market technology satises:
!
dV = (b 1m r) + rV d + b dW,
where is some portfolio, 1m is a m-dimensional vector of ones, and
b = (b1 , , b2 ) ,
b = ( b1 , , b2 ) .
Next, suppose that there exists a portfolio such that b = 0. This is an arbitrage opportunity if
there exist events for which at some time, b 1m r = 0 (use when b 1m r > 0, and when
m 1d r < 0: the drift of V will then be appreciating at a deterministic rate that is strictly greater
than r). Therefore, arbitrage opportunities are ruled out if:
(b 1m r) = 0 whenever b = 0.
In other terms, arbitrage opportunities are ruled out when every vector in the null space of b is
orthogonal to b 1m r, or when there exists a taking values in Rd satisfying some basic integrability
conditions, and such that
b 1m r = b
or,
bi r = bi , i = 1, , m.
(11A.2)
In this case,
dPi
= (r + bi ) d + bi dW, i = 1, , m.
Pi
#
#
#
= W + d , dQ = exp( T dW 1 T 2 d ). The Q-martingale property of
Now dene W
dP
2 t
t
the normalized bond price processes now easily follows by Girsanovs theorem. Indeed, dene for a
generic i, P ( , T ) P ( , Ti ) Pi , and:
g( ) e

t
r(u)du
P ( , T ),
[t, T ] .
By Girsanovs theorem, and an application of Itos lemma,

dg
, under Q.
= bi dW
g
Therefore, for all [t, T ], g( ) = E [g(T )], implying that:
g( ) e
or

t
r(u)du
P ( , T ) = E [g(T )] = E[e
P ( , T ) = e
r(u)du
E e
T
t
r(u)du
T
t
r(u)du
= E e
333
P (T, T ) ] = E e

=1
T
r(u)du
T
t
r(u)du
, all [t, T ],
11.8. Appendix 1: The FTAP for bond prices
c
by
A. Mele
which is Eq. (11.5).

Notice that no assumption has been made on m. The previous result holds for all m, be they less
or greater than d. Suppose, for example, that there are no other traded assets in the economy. Then,
if m < d, there exists an innite number of risk-neutral proabilities Q. If m = d, there exists one and
only one risk-neutral probability Q. If m > d, there exists one and only one risk-neutral probability
but then, the various bond prices have to satisfy some basic no-arbitrage restrictions. As an example,
take m = 2 and d = 1. Eq. (11A.2) then becomes
r
b1 r
= = b2
.
b1
b2
In other terms, the Sharpe ratio of any two bonds must be identical. Relation (11A.2) will be used
several times in this chapter.
In Section 11.3, it is assumed that the primitive of the economy is the short-term rate, solution
of a multidimensional diusion process, and bi and bi will be derived via Itos lemma.
In Section 11.4, bi and bi are restricted through a model describing the evolution of the forward
rates.
334
11.9. Appendix 2: Certainty equivalent interpretation of forward prices
c
by
A. Mele
11.9 Appendix 2: Certainty equivalent interpretation of forward prices

Multiply both sides of the bond pricing equation (11.5) by the amount S(T ):
!
T
P (t, T ) S(T ) = E e t r( )d S(T ).
Suppose momentarily that S(T ) is known at T . In this case, we have:
!
T
P (t, T ) S(T ) = E e t r( )d S(T ) .
But in the applications we have in mind, S(T ) is random. Dene then its certainty equivalent by the
number S(T ) that solves:
!
T
P (t, T ) S(T ) = E e t r( )d S(T ) ,
or
S(T ) = E [T (T ) S(T )] ,
(11A.3)
where T (T ) has been dened in (11.16).

Comparing Eq. (11A.3) with Eq. (11.15) reveals that forward prices can be interpreted in terms of
the previously dened certainty equivalent.
335
11.10. Appendix 3: Additional results on T -forward martingale probabilities
c
by
A. Mele
11.10 Appendix 3: Additional results on T -forward martingale probabilities

Eq. (11.16) denes T (T ) as:
T (T ) =
T
t
E e
r( )d
T
t
More generally, we can dene a density process as:

e
T ( )

t
r(u)du
E e
T
t
1
!
r( )d
P ( , T )
! ,
r( )d
[t, T ] .
#
By the FTAP, {exp( t r(u)du) P ( , T )} [t,T ] is a Q-martingale (see Appendix 1 to this chapter).

dQT
Therefore, E[ dQF F ] = E[ T (T )| F ] = T ( ) all [t, T ], and in particular, T (t) = 1. We now
show that this works. And at the same time, we show this by deriving a representation of T ( ) that
can be used to nd forward premia.
We begin with the dynamic representation (11A.1) given for a generic bond price # i, P ( , T )
P ( , Ti ) Pi :
dP
= d + dW,
P
where we have dened bi and bi .
Under the risk-neutral probability Q,
dP
,
= r d + dW
P
#
= W + is a Q-Brownian motion.
where W
By Itos lemma,
d T ( )
( ), T (t) = 1.
= [( , T )] dW
T ( )
The solution is:

"
"
1
2
(u, T ) du
((u, T )) dW (u) .
T ( ) = exp
2 t
t
Under the usual integrability conditions, we can now use the Girsanovs theorem and conclude that
"

QT
F
W ( ) W ( ) +
(u, T ) du
(11A.4)
t
is a Brownian motion under the T -forward martingale probability QTF .

Finally, note that for all integers i and non decreasing sequences of dates {Ti }i=0,1,, ,
"

Ti
( ) +
W QF ( ) = W
(u, Ti ) du, i = 0, 1, .
t
Therefore,
W
QFi
( ) = W
QFi1
( )
"
!
(u, Ti ) (u, Ti1 ) du, i = 1, 2, ,
(11A.5)
is a Brownian motion under the Ti -forward martingale probability QTFi . Eqs. (11A.5) and (11A.4) are
used in Section 11.7 on interest rate derivatives.
336
11.11. Appendix 4: Principal components analysis
c
by
A. Mele
11.11 Appendix 4: Principal components analysis

Principal component analysis transforms the original data into a set of uncorrelated variables, the
principal components, with variances arranged in descending order. Consider the following program,
max [var (Y1 )] s.t. C1 C1 = 1,
C1
where var (Y1 ) = C1 C1 , and the constraint is an identication constraint. The rst order conditions
lead to,
( I) C1 = 0,
where is a Lagrange multiplier. The previous condition tells us that must be one eigenvalue of
the matrix , and that C1 must be the corresponding eigenvector. Moreover, we have var (Y1 ) =
C1 C1 = which is clearly maximized by the largest eigenvalue. Suppose that the eigenvalues of
are distinct, and let us arrange them in descending order, i.e. 1 > > p . Then,
var (Y1 ) = 1 .

, where C1 is the eigenvector corresponding
Therefore, the rst principal component is Y1 = C1 R R
to the largest eigenvalue, 1 .
Next, consider the second principal component. The program is, now,
max [var (Y2 )] s.t. C2 C2 = 1 and C2 C1 = 0,
C2
where var (Y2 ) = C2 C2 . The rst constraint, C2 C2 = 1, is the usual identication constraint. The
second constraint, C2 C1 = 0, is needed to ensure that Y1 and Y2 are orthogonal, i.e. E (Y1 Y2 ) = 0.
The rst order conditions for this problem are,
0 = C2 C2 C1
where is the Lagrange multiplier associated with the rst constraint, and is the Lagrange multiplier
associated with the second constraint. By pre-multiplying the rst order conditions by C1 ,
0 = C1 C2 ,
where we have used the two constraints C1 C2 = 0 and C1 C1 = 1. Post-multiplying the previous
expression by C1 , one obtains, 0 = C1 C2 C1 C1 = C1 , where the last equality follows by
C1 C2 = 0. Hence, = 0. So the rst order conditions can be rewritten as,
( I) C2 = 0.
The solution is now 2 , and C2 is the eigenvector corresponding
to 2 . (Indeed, this time we can not

, implying that E (Y1 Y2 ) = 0.) It follows
choose 1 as this choice would imply that Y2 = C1 R R
that var (Y2 ) = 2 .
In general, we have,
var (Yi ) = i ,
i = 1, , p.
Let be the diagonal matrix with the eigenvalues i on the diagonal. By the spectral decomposition
of , = CC , and by the orthonormality of C, C C = I, we have that C C = and, hence,

p
var
(R
)
=
Tr
()
=
Tr
CC
=
Tr
C
C
= Tr () .
i
i=1
Hence, Eq. (11.22) follows.
337
11.12. Appendix 5: A few analytics for the Hull and White model
c
by
A. Mele
11.12 Appendix 5: A few analytics for the Hull and White model
As in the Ho and Lee model, the instantaneous forward rate f( , T ) predicted by the Hull and White
model is as in Eq. (11.48), where functions A2 and B2 can be easily computed from Eqs. (11.51) and
(11.52) as:
A2 ( , T ) = 2
"
B(s, T )B2 (s, T )ds
"
(s)B2 (s, T )ds, B2 ( , T ) = e(T ) .
Therefore, the instantaneous forward rate f ( , T ) predicted by the Hull and White model is obtained
by replacing the previous equations in Eq. (11.48). The result is then equated to the observed forward
rate f$ (t, ) so as to obtain:
!2 "
2
(s)e( s) ds + e( t) r(t).
f$ (t, ) = 2 1 e( t) +
2
t
By dierentiating the previous equation with respect to , and rearranging terms,
"

2
( ) =
f$ (t, ) +
1 e( t) e( t) +
(s)e( s) ds + e( t) r(t)
t

2
2
2
( t)
( t)
( t)
e
f (t, ) +
1e
+ f$ (t, ) + 2 1 e
=
,
$
2
which reduces to Eq. (11.53) after using simple algebra.
338
11.13. Appendix 6: Expectation theory and embedding in selected models
c
by
A. Mele
11.13 Appendix 6: Expectation theory and embedding in selected models

A.
Expectation theory
Suppose that
(, ) = and () = ,
(11A.6)
where and are constants. We derive the dynamics of r and compare them with f to deduce
something about the expectation theory. We have:
"
r( ) = f(t, ) +
(s, )ds + (W ( ) W (t)) ,
t
where
( , T ) = ( , T )
"
Hence,
"
Finally,
( , )d + ( , T )( ) = 2 (T ) + .
1
(s, )ds = 2 ( t)2 + ( t).
2
1
r( ) = f(t, ) + 2 ( t)2 + ( t) + (W ( ) W (t)) ,
2
and since E ( W ( )| F(t)) = W (t),
1
E [ r( )| F(t)] = f(t, ) + 2 ( t)2 + ( t).
2
Even with < 0, this model is not able to always generate E[ r( )| F(t)] < f (t, ). As shown in the
following exercise, this is due to the nonstationary nature of the volatility function. Indeed, suppose,
next, that instead of Eq. (11A.6), we have that
(t, T ) = exp((T t)) and () = ,
where , and are constants. In this case, we have:
"
"
r( ) = f(t, ) +
(s, )ds +
t
e( s) dW (s),
where
2 ( s)
(s, ) = e
"
e(s) d + e( s) =
!
2 ( s)
e
e2( s) + e( s) .
Finally,
E [ r( )| F(t)] = f(t, ) +
"

( t)
( t)
(s, )ds = f(t, ) +
1e
1e
+ .
Therefore, it is sucient to have a risk-premium such that >
2 ,
to generate the prediction that:
E [ r( )| F(t)] < f(t, ) for any .

In other words, < 0 is a necessary condition, not sucient. Notice that when = 0, it always holds
that E ( r( )| F(t)) > f(t, ).
339
11.13. Appendix 6: Expectation theory and embedding in selected models
c
by
A. Mele
B. Embedding
We now embed the Ho and Lee model in Section 11.5.2 in the HJM format. In the Ho and Lee model,
( ),
dr( ) = ( )d + dW
is a Q-Brownian motion. By Eq. (11.48) in Section 11.4,
where W
where A2 (t, T ) =
#T
t
f (r, t, T ) = A2 (t, T ) + B2 (t, T )r,

(s)ds 12 2 (T t)2 and B2 (t, T ) = 1. Therefore, by Eqs. (11.61),
(t, T ) = B2 (t, T ) = ,
(t, T ) (t, T )(t) = A12 (t, T ) + B12 (t, T )r + B2 (t, T )(t) = 2 (T t).
Next, we embed the Vasicek model in Section 11.4 in the HJM format. The Vasicek model is:
( ),
dr( ) = ( r( ))d + dW
is a Q-Brownian motion. Results from Section 11.3 imply that:
where W
f (r, t, T ) = A2 (t, T ) + B2 (t, T )r,
#T
#T
where A2 (t, T ) = 2 t B(s, T )B2 (s, T )ds + t B2 (s, T )ds, B2 (t, T ) = e(T t) and B(t, T )

= 1 1 e(T t) . By Eqs. (11.61),
(t, T ) = B2 (t, T ) = e(T t) ;
(t, T ) (t, T )(t) = A12 (t, T ) + B12 (t, T )r + ( r)B2 (t, T ) =
!
2
1 e(T t) e(T t) .
Naturally, this model can never be embedded within a HJM model because it is not of the perfectly
tting type. In practice, condition (11.62) can never hold in the simple Vasicek model. However, the
model is embeddable once is turned into an innite dimensional parameter a
` la Hull and White (see
Section 11.3).
340
11.14. Appendix 7: Additional results on string models
c
by
A. Mele
11.14 Appendix 7: Additional results on string models

Here we prove Eq. (11.64). We have, I ( , T ) =
g ( , T, 2 )
"
1
2
#T
d
g ( , T, 2 ) d2 + cov( dP
P , ), where
( , 1 ) ( , 2 ) (1 , 2 ) d1 .
Dierentiation of the cov term is straight forward. Moreover,

" T
" T
g ( , T, 2 )
g ( , T, 2 ) d2 = g ( , T, T ) +
d2
T
T
" T

= ( , T )
( , x) [ (x, T ) + (T, x)] dx
" T

= 2 ( , T )
( , x) (x, T ) dx .
341
c
by
A. Mele
11.15. Appendix 8: Changes of numeraire
11.15 Appendix 8: Changes of numeraire

A.
Jamshidian (1989)
Consider the following change-of-numeraire result. Let

dA
= A d + A dW,
A
and consider a similar process B with coecients B and B . We have:

d(A/B)
= A B + 2B A B d + (A B ) dW.
A/B
(11A.7)
( ,S)
S
T
We apply this result to the process y( , S) PP (
,T ) , under QF as well as under QF . The objective
is to obtain the solution as of time T of y ( , S) viz
P (T, S)
= P (T, S) under QSF as well as under QTF .
P (T, T )
y(T, S)
This allows us to calculate the two probabilities in Eq. (11.71).

By Itos lemma, the PDE (11.31) and the fact that Pr = BP ,
dP ( , x)
( ), x T.
= rd B( , x)dW
P ( , x)
By applying Eq. (11A.7) to y( , S),

dy( , S)
( ).
= 2 B( , T )2 B( , T )B( , S) d [B( , S) B( , T )] dW
y( , S)
(11A.8)
All we need to do now is to change measure with the tools of Appendix 3. We have that:
x
( ) + B( , x)d
dW QF ( ) = dW
x
is a Brownian motion under the x-forward martingale probability. Replace then W QF into Eq. (11A.8),
then integrate, and obtain:
y(T, S)
y(t, S)
y(T, S)
y(t, S)
T

QT
2
1 2 T
P (t, T )
= e 2 t [B( ,S)B(,T )] d t [B( ,S)B( ,T )]dW F ( ) ,
P (t, S)

T
QS
2
1 2 T
P (t, T )
= P (T, S)
= e 2 t [B( ,S)B( ,T )] d t [B( ,S)B( ,T )]dW F ( ) ,
P (t, S)
= P (T, S)
Rearranging terms gives Eqs. (11.72) in the main text.
B. Black (1976)
To prove Eq. (11.99), we need to evaluate the following expectation:
E [x(T ) K]+ ,
where
x(T ) = x(t)e 2
T
t

( )
( )2 d + tT ( )dW
342
(11A.9)
c
by
A. Mele

Let 1ex be the indicator of all events s.t. x(T ) K. We have
E [x(T ) K]+ = E [x(T ) 1ex ] K E [1ex ]

x(T )
= x(t) E
1ex K E [1ex ]
x(t)
= x(t) EQx [1ex ] K E [1ex ]
= x(t) Qx (x(T ) K) K Q (x(T ) K) .
where the probability Qx is dened as:

T
1 T
x(T )
2
dQx
=
= e 2 t ( ) d + t ( )dW ( ) ,
dQ
x(t)
a Q-martingale starting at one. Under Qx ,

( ) d
dW x ( ) = dW
is a Brownian motion, and x in Eq. (11A.9) can be written as:
1
x (T ) = x (t) e 2
T
t

( )2 d + tT ( )dW x ( )
It is straightforward that Q (x(T ) K) = (d2 ) and Qx (x(T ) K) = (d1 ), where

#T
ln x(t)
12 t ( )2 d
K
#
d2/1 =
.
T
2 d
(
)
t
Applying this to EQTi [Fi1 (Ti1 ) K]+ gives the formulae of the text.
F
343
c
by
A. Mele
References
At-Sahalia, Y. (1996): Testing Continuous-Time Models of the Spot Interest Rate. Review
of Financial Studies 9, 385-426.
Andersen, T. G., and J. Lund (1997): Estimating Continuous-Time Stochastic Volatility
Models of the Short-Term Interest Rate. Journal of Econometrics 77, 343-377.
Ang, A. and M. Piazzesi (2003): A No-Arbitrage Vector Autoregression of Term Structure
Dynamics with Macroeconomic and Latent Variables. Journal of Monetary Economics
50, 745-787.
Balduzzi, P., S. R. Das, S. Foresi and R. K. Sundaram (1996): A Simple Approach to Three
Factor Ane Term Structure Models. Journal of Fixed Income 6, 43-53.
Bernanke, B. S. and A. Blinder (1992): The Federal Funds Rate and the Channels of Monetary
Transmission. American Economic Review 82, 901-921.
Black, F. (1976): The Pricing of Commodity Contracts. Journal of Financial Economics 3,
167-179.
Brace, A., D. Gatarek and M. Musiela (1997): The Market Model of Interest Rate Dynamics.
Mathematical Finance 7, 127-155.
Bremaud, P. (1981): Point Processes and Queues: Martingale Dynamics. Berlin: Springer Verlag.
Brigo, D. and F. Mercurio (2006): Interest Rate ModelsTheory and Practice, with Smile,
Inflation and Credit. Springer Verlag (2nd Edition).
Brunnermeier, M. (2009): Deciphering the Liquidity and Credit Crunch 2007-08. Journal of
Economic Perspectives 23, 77-100.
Carverhill, A. (1994): When is the Short-Rate Markovian? Mathematical Finance 4, 305-312.
Cochrane, J. H. and M. Piazzesi (2005): Bond Risk Premia. American Economic Review 95,
138-160.
Conley, T. G., L. P. Hansen, E. G. J. Luttmer and J. A. Scheinkman (1997): Short-Term
Interest Rates as Subordinated Diusions. Review of Financial Studies 10, 525-577.
Cox, J. C., J. E. Ingersoll and S. A. Ross (1985): A Theory of the Term Structure of Interest
Rates. Econometrica 53, 385-407.
Dai, Q. and K. J. Singleton (2000): Specication Analysis of Ane Term Structure Models.
Due, D. and R. Kan (1996): A Yield-Factor Model of Interest Rates. Mathematical Finance
6, 379-406.
344
c
by
A. Mele
Due, D. and K. J. Singleton (1999): Modeling Term Structures of Defaultable Bonds.

Review of Financial Studies 12, 687720.
Fama, E. F. and R. R. Bliss (1987): The Information in Long-Maturity Forward Rates.
American Economic Review 77, 680-692.
Fong, H. G. and O. A. Vasicek (1991): Fixed Income Volatility Management. The Journal
of Portfolio Management (Summer), 41-46.
Geman, H. (1989): The Importance of the Forward Neutral Probability in a Stochastic Approach to Interest Rates. Unpublished working paper, ESSEC.
Geman H., N. El Karoui and J. C. Rochet (1995): Changes of Numeraire, Changes of Probability Measures and Pricing of Options. Journal of Applied Probability 32, 443-458.
Goldstein, R. S. (2000): The Term Structure of Interest Rates as a Random Field. Review
Heaney, W. J. and P. L. Cheng (1984): Continuous Maturity Diversication of Default-Free
Bond Portfolios and a Generalization of Ecient Diversication. Journal of Finance 39,
1101-1117.
Heath, D., R. Jarrow and A. Morton (1992): Bond Pricing and the Term-Structure of Interest
Rates: a New Methodology for Contingent Claim Valuation. Econometrica 60, 77-105.
Ho, T. S. Y. and S.-B. Lee (1986): Term Structure Movements and the Pricing of Interest
Rate Contingent Claims. Journal of Finance 41, 1011-1029.
Hull, J. C. (2003): Options, Futures, and Other Derivatives. Prentice Hall. 5th edition (International Edition).
Hull, J. C. and A. White (1990): Pricing Interest Rate Derivative Securities. Review of
Financial Studies 3, 573-592.
Jagannathan, R. (1984): Call Options and the Risk of Underlying Securities. Journal of
Jamshidian, F. (1989): An Exact Bond Option Pricing Formula. Journal of Finance 44,
205-209.
Jamshidian, F. (1997): Libor and Swap Market Models and Measures. Finance and Stochastics 1, 293-330.
Joreskog, K. G. (1967): Some Contributions to Maximum Likelihood Factor Analysis. Psychometrica 32, 443-482.
Karlin, S. and H. M. Taylor (1981): A Second Course in Stochastic Processes. San Diego:
Academic Press.
Kennedy, D. P. (1994): The Term Structure of Interest Rates as a Gaussian Random Field.
345
c
by
A. Mele
Kennedy, D. P. (1997): Characterizing Gaussian Models of the Term Structure of Interest

Rates. Mathematical Finance 7, 107-118.
Kloeden, P. and E. Platen (1992): Numeric Solutions of Stochastic Differential Equations.
Berlin: Springer Verlag.
Knez, P. J., R. Litterman and J. Scheinkman (1994): Explorations into Factors Explaining
Money Market Returns. Journal of Finance 49, 1861-1882.
Langetieg, T. (1980): A Multivariate Model of the Term Structure of Interest Rates. Journal
of Finance 35, 71-97.
Litterman, R. and J. Scheinkman (1991): Common Factors Aecting Bond Returns. Journal
of Fixed Income 1, 54-61.
Litterman, R., J. Scheinkman, and L. Weiss (1991): Volatility and the Yield Curve. Journal
of Fixed Income 1, 49-53.
Longsta, F. A. and E. S. Schwartz (1992): Interest Rate Volatility and the Term Structure:
A Two-Factor General Equilibrium Model. Journal of Finance 47, 1259-1282.
Mele, A. and F. Fornari (2000): Stochastic Volatility in Financial Markets: Crossing the Bridge
to Continuous Time. Boston: Kluwer Academic Publishers.
Merton, R. C. (1973): Theory of Rational Option Pricing. Bell Journal of Economics and
Management Science 4, 141-183.
Miltersen, K., K. Sandmann and D. Sondermann (1997): Closed Form Solutions for Term
Structure Derivatives with Lognormal Interest Rate. Journal of Finance 52, 409-430.
Rebonato, R. (1998): Interest Rate Option Models. Wiley.
Rebonato, R. (1999): Volatility and Correlation. Wiley.
Ritchken, P. and L. Sankarasubramanian (1995): Volatility Structure of Forward Rates and
the Dynamics of the Term Structure. Mathematical Finance 5, 55-72.
Rothschild, M. and J. E. Stiglitz (1970): Increasing Risk: I. A Denition. Journal of Economic Theory 2, 225-243.
Sandmann, K. and D. Sondermann (1997): A Note on the Stability of Lognormal Interest
Rate Models and the Pricing of Eurodollar Futures. Mathematical Finance 7, 119-125.
Santa-Clara, P. and D. Sornette (2001): The Dynamics of the Forward Interest Rate Curve
with Stochastic String Shocks. Review of Financial Studies 14, 149-185.
Stanton, R. (1997): A Nonparametric Model of Term Structure Dynamics and the Market
Price of Interest Rate Risk. Journal of Finance 52, 1973-2002.
346
c
by
A. Mele
Stock, J. H. and M. W. Watson (2003): Forecasting Output and Ination: The Role of Asset
Prices, Journal of Economic Literature 41, 788-829.
Vasicek, O. (1977): An Equilibrium Characterization of the Term Structure. Journal of
347
12
Risky debt and credit derivatives
12.1 Introduction
12.2 Conceptual approaches to valuation of defaultable securities
12.2.1
Firms value, or structural, approaches
Relies on the structure of the rm. Shares and bonds as derivatives on the rms assets.
Stylized balance sheet
Equity (E)
(Shares)
Assets (A)
Debt (D)
(Bonds)
Therefore, we have the accounting identity: Assets = Equity + Debt, or
A = E + D.
At the time of debt expiration, debtholders receive the minimum between the debt nominal value
and the value of the assets the rm can liquidate to honour the debt obligation. Debtholders are
senior claimants. Equity holders are residual claimants to the rms assets > Junior claimants
We can use these basic insights to illustrate the rst model about the risk-structure of interest
rates, the Merton - KMV approach. Equity is like a European call option written on the rms
assets, with expiration equal to the debt expiration, and strike equal to the nominal value of
debt. Current value of debt equals the value of the assets minus the value of equity, i.e. the
value of a risk-free discount bond minus the value of a put option on the rm with strike price
equal to the nominal value of debt, as shown by Eq. (12.3) below.
Merton (1974) uses the Black and Scholes (1973) formula to derive the price of debt. The
main assumption underlying this model is that the assets of the rm can be traded, and that
c
by
A. Mele
12.2. Conceptual approaches to valuation of defaultable securities

their value At satises,1
dAt
t,
= rdt + dW
At
(12.1)
t is a Brownian motion under the risk-neutral probability, is the instantaneous

where W
standard deviation, and r is the short-term rate on riskless bonds.
Let N be the nominal value of debt, T be time of expiration of debt; Dt the debt value as
of at time t T . As argued earlier, shareholders have long a European call option, and the
bondholders are residual claimants. Mathematically,
+
AT , if the rm defaults, i.e. AT < N
DT =
N, if the rm is solvent, i.e. AT N
We can decompose the assets value at time T , into the sum of the value of equity and the value
of debt, at time T ,
(12.2)
DT = min {AT , N } = AT max {AT N, 0} .
Equity at T
Note, also, that,

DT = min {AT , N} = N max {N AT , 0} .
(12.3)
Put on the rm
That is, credit risk raises the cost of capital.
1.2
1.0
0.8
0.6
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
2.2
A_T
FIGURE 12.1. Dashed line: the value of equity at the debt maturity, T , max {AT N, 0}, plotted as a
function of the value of assets, AT . Solid line: the value of debt at maturity, min {AT , N} as a function
of AT . Nominal value of debt is xed to N = 1.
A word on convexity, and risk-taking behavior. Convexity: Managers have incentives to invest
in risky assets, as the terminal payo to them is increasing in the assets volatility, . Concavity:
The value of debt, instead, is decreasing in the assets volatility.
t , where t is the instantaneous cash ow to the
(12.1) could be generalized to one in which dA
t = (rAt t) dt + At dW
rm. This would make the rm value equal to A0 = E 0 ert t dt . For example, one could take t to be a geometric Brownian
motion with parameters g and , in which case At = (r g)1 t , forever, but were just ignoring this complication.
1 Eq.
349
c
by
A. Mele

12.2.1.1 Merton
The current value of the bonds equals the current value of the assets, A0 , minus the current
value of equity. The current value of equity can obtained through the Black & Scholes formula,
as equity is a European call option on the rm, struck at N. By Eq. (12.2), and standard
risk-neutral evaluation, then, the current value of debt, D0 , is,

1 2
ln
(
A
T
/
N)
+
r
+
0
2
D0 = A0 (d1 ) + N erT (d2 ), d1 =

, d2 = d1 T , (12.4)
T
where () denotes the distribution function of a standard normal variable.2
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
A_0
FIGURE 12.2. Solid line: the no-arbitrage bound, min {A0 , N}, depicted as a function of A0 , when
the nominal value of debt is xed to N = 1. Dashed line: the bond value predicted by the Mertons
model when T = 1, r = 3% and = 20%, annualized. Dotted line: same as the dashed line, but with
a larger asset volatility, = 40%.
Bond prices are decreasing in the asset volatility as bad outcomes are exaggerated on the
downside, due to the concavity properties depicted in Figure 12.1.
The risk-structure of interest rates is obtained with the standard formula for continuously
compounded interest rates as,

1
D0
R = ln
= r + Spread,
T
N
where the term-spread or, simply, the spread, is

1
A0 rT
Spread = ln
e (d1 ) + (d2 ) .
T
N
2 For
(12.5)
the details, note that D0 = erT E [ DT | A0 ] and, then, by Eq. (12.2),

D0 = erT E ( AT | A0 ) erT E [ max {AT N, 0}| A0 ] = A0 A0 (d1 ) Ner (d2 ) ,
where the last equality follows by the Black & Scholes formula. Eq. (12.4) follows after rearranging terms in the previous equation.
350
c
by
A. Mele

Figure 12.3 depicts the spread predicted by this model.
Spread
0.03
0.02
0.01
0.00
0
Time to maturity
FIGURE 12.3. The term structure of spreads, s0 , obtained with initial asset values A0 = 1.1 (solid
line), A0 = 1.2 (dashed line), and A0 = 1.3 (dotted line). The short-term rate, r = 3%, and asset
volatility is = 0.20. Nominal debt N = 1.
We can introduce a useful summary statistics for credity quality: distance-to-default (under
Q). We can use the previous model to estimate the likelihood of default for a given rm. First,
we develop Eq. (12.2),
DT = min {AT , N } = AT I{AT <N} + N I{AT N} ,
where I{E} is the indicator function, i.e. I{E} = 1 if the event E is true and I{E} = 0 if the event
E is false. Second, we have,
D0 = erT E (DT )

= erT E AT I{AT <N} + N E I{AT N}
= erT [E ( AT | Default) Q (Default) + N Q (Survival)] ,
(12.6)
where E ( AT | Default) is the expected asset value given the event of default, Q (Default) is the
probability of default, and Q (Survival) = 1 Q (Default) is the probability the rms does not
default.
Comparing Eq. (12.6) with Eq. (12.4) reveals that for the Mertons model,
Q (Survival) = (d2 ) .
351
c
by
A. Mele
Pr(surv)
1.0
0.9
0.8
0.7
0.6
0.5
0.0
0.1
0.2
0.3
0.4
sigma
FIGURE 12.4. Probability of survival for a given rm predicted by the Mertons model, (d2 ), depicted
as a function of the asset volatility, . Assets value is xed at A0 = 1.1, and plotted are survival
probabilities for bonds maturing at T = 0.5 years (solid line), T = 1 year (dashed line) and T = 2
years (dotted line). The short-term rate, r = 3%. Nominal debt N = 1.
The probability of survival, (i) decreases with debt maturity and (ii) the asset volatility.
Property (i) is not a general property, though. With lower values of A0 , the relation between
maturity and probability of survival can be increasing or decreasing, according to the values of
, as shown in Figure 12.5. Intuitively, when A0 N, the probability of survival is:

ln AN0 + r 12 2 T
r 12 2
T,
Q (Survival) = (d2 ) , with d2 =
T
such that the survival probability decreases in T for large although then it increases in T for
small . The intuition underlying this property is that for large , the probability the asset value
will end up below N from
A0 N can only increase with time to maturity, T . Analytically,
E ( ln AT | A0 ) = ln A0 + r 12 2 T ln N + r 12 2 T , such that the probability the assets
value will be above N is, indeed, approximately Q (Survival).
352
c
by
A. Mele
Pr(surv)
1.0
0.9
0.8
0.7
0.6
0.5
0.0
0.1
0.2
0.3
0.4
sigma
FIGURE 12.5. Probability of survival for a given rm predicted by the Mertons model, (d2 ), depicted
as a function of the asset volatility, . Assets value is xed at A0 = 1.01, and plotted are survival
probabilities for bonds maturing at T = 0.5 years (solid line), T = 1 year (dashed line) and T = 2
years (dotted line). The short-term rate, r = 3%. Nominal debt N = 1.
The summary statistics, distance-to-default, is dened as,

ln ( A0 / N ) + r 12 2 T
d2 =
.
T
(12.7)
It is a very intuitive measure of how far the rm is from defaulting, as we know the (riskadjusted) probability of surviving is (d2 ), which is increasing in d2 . The larger the current
asset value A0 is, the less likely it is the rm will default at T . Distance-to-default is decreasing
in the assets volatility , as illustrated earlier by Figure 12.4.
By Eq. (12.1), we have that E ( ln AT | A0 ) = ln A0 + (r 12 2 )T , so Eq. (12.7) tells us that
distance-to-default is simply the dierence E ( ln AT | A0 ) ln N, normalized by the standard
deviation of the assets over the life of debt.
Some prefer to use the slightly dierent formula,
Distance-to-default =
Mkt value of Assets Default value

.
Mkt value of Assets Asset volatility
Another useful concept is Loss-given-default, under Q. Comparing Eq. (12.6) with Eq. (12.4)
reveals another property of the Mertons model,
E ( AT | Default) =
A0 erT (d1 )
(d1 )
(d1 )
= A0 erT
= E (AT )
E (AT ) .
Q (Default)
(d2 )
(d2 )
Recovery rates are dened as the fraction of the bond value the bondholders expect to obtain
in the event of default, at maturity:
Recovery Rate =
A0 rT (d1 )
E ( AT | Default)
=
e
.
N
N
(d2 )
353
c
by
A. Mele
Loss-given-default is dened as the fraction of the bond value the bondholders expect to lose
in the event of default, at maturity:
Loss-given-default = 1 Recovery Rate.
Finally, by Eq. (12.5), we can write,

A0 rT
1
s0 = ln
e (d1 ) + (d2 )
T
N
1
= ln (Recovery Rate Q (Default) + Q (Survival))
T
1
[Loss-given-default Q (Default)] .
T
This is actually a general formula, which goes through beyond the Mertons model). It can
easily be obtained through Eq. (12.6).
What is the asymptotic behavior of the survival probabilities, and then the spreads? If r >
1 2
, then, as T , (d2 ) 1, that is, the assets value is expected to grow so much that
2
default will never occur, such that thr bond becomes riskless and s0 T1 ln (d2 ) 0.
An important note. Survival probabilities, distance-to-default, loss-given-default, were previously dened under the risk-adjusted probability Q. To calculate the same objects under the
true probability P, we replace r with the asset growth rate under the physycal probability, , in
the formulae for the survival probabilities, (d2 ), distance-to-default, d2 , and loss-given-default.
However, it is hard to estimate for many single names. Moodys KMV EDFTM are based
on dynamic structural models like these, although the details are not publicly known. Finally,
we could use historical data about default frequencies to estimate the probability that a given
single name within a certain industry will default. These frequencies are based on samples of
rms that have defaulted in the past, with similar characteristics to those of the rm under
evaluation (in terms, for instance, of distance-to-default).
How to estimate At and ? One algorithm is to start with some equal to the volatility
(0)
of equity returns, say (0) , and use Mertons formula for equity, to extract At for each date
(0)
t {1, , T }, where T is N the sample size. Then, use At to compute the standard deviation
(0)
(0)
of ln(At /At1 ). This gives say (1) , which can be used as the new input to the Mertons formula
(i) T
(i)
to extract say A(1)
t . We obtain a sequence of (At )t=1 and , and we stop for i suciently
large, according to some criterion.
12.2.1.2 One example
Assume the assets value of a given rm is A0 = 110, and that the instantaneous volatility of the
assets value growth is = 30%, annualized. The safe interest rate is r = 2%, annualized, and
the expected growth rate of the assets value is = 5%, annualized. The rm has outstanding
debt with nominal value N = 100, which expires in two years.
First, we compute the distance-to-default implied by the Mertons model, which is,

ln AN0 + r 12 2 T
ln (1.1) + 0.02 12 0.32 2
D-t-D =
=
= 0.10680.
0.3 2
T
Accordingly, the probability of default is,
1 (0.10680) = 1 0.54253 = 0.45747.
354
c
by
A. Mele
We can compute the same probability, under the physical measure, by simply replacing
r = 2% with = 5%, in the formula for D-t-D. We have,

ln (1.1) + 0.05 12 0.32 2
ln AN0 + 12 2 T
=
D-t-Dphysical =
= 0.24822.
0.3 2
T
Therefore, the probability of default under the physical distribution is,
1 physical (0.24822) = 1 0.59802 = 0.40198.
It is, of course, lower under the physical probability than under the risk-neutral probability,
due to the larger asset growth rate, > r.
Finally, we can compute the spread on this bond, which is given by:

1
A0 rT
Spread = ln
e (d1 ) + (d2 ) ,
T
N
where d2 = D-t-D, and d1 = d2 + T . So we have,

1
Spread = ln 1.1e0.022 0.10680 + 0.30 2 + (0.10680)
2

1
= ln 1.1e0.022 0.29769 + 0.54253
2
= 6.20%.
12.2.1.3 First passage
The timing of default can be triggered by some exogeneously specied events. For example,
default occurs if the value of the assets hits some exogenously lower bound even before the
expiration of debt. These models are known as rst passage models, because they rely on
mathematical techniques that solve for the probability the rst time the asset value hit some
exogenous barrier, as in Black and Cox (1976).
12.2.1.4 Strategic defaulting
The timing of default can be endogenous. Managers choose the defaulting barrier (i.e. the
asset value that triggers bankruptcy) so as to maximize the rms value. Naturally, strategic
defaulting works if the assumptions underlying the Modigliani-Miller theorem do not hold. The
mechanism is the following: on the one hand, debt is a tax-shielding device. On the other hand,
issuing too much debt increases the likelihood of default, which triggers bankruptcy costs. The
rst eect raises the value of the rm while the second, decreases the value of the rm. Equity
holders choose the value of the asset that triggers bankruptcy to maximize the value of equity.
Leland (1994): Long-term debt. Leland and Toft (1996): Extension to nite maturity debt.
Anderson and Sundaresan (1996): Debt re-negotiation.
The Leland model works as follows. First, debt is innitely lived in that, it pays o an
instantaneous coupon equal to Cdt. In the absence of default risk, debt would simply equal
C/r. Second, tax benets are assumed to be proportional to the coupon, Cdt. Third, there are
bankruptcy costs: if the rm defaults at V = VB , recovery is (1 ) VB . Equity holders choose
VB . Naturally, VB < V0 .
355
c
by
A. Mele
The value of debt is a function of V , say D (V ). Under the risk-neutral probability, it satises:

d
+
C = rD (Vt ) .
E [D (VT )| V0 ]
dT
T =t

=coupon
=Expected capital gains
By Itos lemma, this is an ordinary dierential equation, subject to the following boundary
conditions. First, at bankruptcy, D (VB ) = (1 ) VB . Second, for large V , debt is substantially
riskless, i.e. limV D (V ) = Cr .
The solution to this is,
D (V ) = (1 pB (V ))
where
C
+ pB (V ) [(1 ) VB ] ,
r
2r
VB 2
pB (V )
.
V
Note, we may interpret pB (V ) as the present value of $1, contingent on future bankruptcy.
Accordingly, the total benets arising from tax shielding are,

T B (V ) = (1 pB (V ))
C
.
r
and the present value of bankruptcy costs is,

BC (V ) = pB (V ) VB .
We have,
Value of the rm = Equity + Debt
= Value of Assets (V ) + T B (V ) BC (V ) .
Summing up,
C
pB (V ) VB .
r
Equity equals (i) the value of the assets, V ; minus (ii) the present value of debt contingent
on no-bankruptcy net of tax benets, (1 pB (V )) (1 ) Cr ; minus (iii) the present value of
debt contingent on bankruptcy net of bankruptcy costs, pB (V ) VB . The second term decreases
with the default boundary, VB or, equivalently, pB (V ). The third term, instead, increases with
VB . So the time equityholder wait before declaring bankruptcy, which is inversely related to VB ,
aects in opposite ways the last two terms. Equityholders choose VB to maximize the value of
equity. Their solution is a default boundary, VB , such that the value of equity does not change
for small changes in the value of the assets V around VB , or VB : E (V )|V =VB = 0, a smooth
pasting condition. The result is,
E (V ) Equity = V (1 pB (V )) (1 )
VB = (1 )
C
.
r + 12 2
How is it that tax shielding does not seem to aect the existence of a solution to this problem?
That is, the default boundary, VB , still exists, even with = 0. This issue is easily resolved.
If = 0, there are no reasons to issue debt in the rst place, as no tax benets would ow
to the rm, thereby increasing its value! In fact, when > 0, there is a value of leverage that
maximizes the value of the rm, according to simulations reported in Leland (1994).
356
c
by
A. Mele
12.2.1.5 Pros and cons of structural approaches to risky debt assessment
Pros. First, they allow to think about more complicated structures or instruments easily (e.g.,
convertibles). Second, they lead to simple yet consistent relations between dierent securities
issued by the same name. Structural approaches were very useful for theoretical research in the
1990s.
Cons. The rms asset value and asset volatility are not observed. Must rely on calibration/estimation methods. Bond prices generated by the model = market prices. These models
are a bit dicult to use in practice, for trading or hedging purposes, as we know that in this
case we need theoretical prices that exactly match market prices. Finally, how do we go for
sovereign issuers?
Most important. Structural models predict unrealistically low short-term spreads: see, e.g.,
Figure 12.3. The intuition is that diusion processes are smooth: the probability of default tends
to zero as time to maturity approaches zero, because default cannot just jump in an unexpected
way. This is not what we exactly observe. Jumps seem to be a more realistic device to modeling
spreads.
12.2.2
Reduced form approaches: rare events, or intensity, models
Default often displays a few strong characteristics. It arrives unexpectedly, it is rare, and causes
causes discontinuous price changes. The structural models in the previous section do not accommodate for these features diusion processes are continuous: passage times are known,
locally. This feature is responsible for the low short-term spreads.
12.2.2.1 Poisson-driven defaults
We model the arrival of defaults through the Poisson processes introduced in Chapter 4, as
follows. Suppose to count the number of times some event happens. Denote with Nt the
corresponding counting process.
Nt
D efault
t0
t2
t1
FIGURE 12.6.
357
t3
c
by
A. Mele
Default time is simply dened as t0 , i.e. the rst time Nt jumps, as in Figure 12.6. So
assume we chop a given interval [0, T ] in n pieces, and consider each resulting interval t = Tn .
Assume jump probability over each of these small intervals of time t is proportional to t,
with proportionality factor equal to ,
p Pr {One jump over t} = t.
(12.8)
Assume the number of jumps over the n intervals follows a binomial distribution:

n k
T
Pr {k jumps over [0, T ]} =
p (1 p)nk , where p = .
k
n
For n large (or, equivalently, for small intervals t),

(T )k
T nk (T )k T
Pr {k jumps over [0, T ]}
1
e
.
k!
n
k!
We can use the previous basic computations to come up with a few fundamental properties
for the distribution of default. We have,
Pr {Survival} = Pr {0 jumps over [0, T ]} = eT
Pr {Default} = 1 Pr {Survival} = 1 eT
Pr {Default occurs at some t} = et dt
1
Expected Time-to-Default =
We can now use these probabilities to assess the value of debt subject to default risk. Consider
Eq. (12.6):
D0 = erT [Rec Q (Default) + N Q (Survival)],

B0
where Rec is the expected recovery value of the asset. Using the probabilities predicted by the
Poisson model, we obtain:

B0 = Rec 1 eT + N eT .
(12.9)
The Appendix supplies an alternative derivation of Eq. (12.9).
12.2.2.2 Predicted spreads
The implications for spreads, for small maturities T , are easily seen, after some innocous approximations,

1
B0
1 B0
1
Spread = ln
1 =
1 eT Loss-given-default.
T
N
T N
T
Note, for T small, the spread is not zero, as in the previous structural models. Rather, it
is given by the expected default loss per period, dened as the instantaneous probability of
default times loss-given-default,
Short-Term Spread = Loss-given-default.
358
c
by
A. Mele
Figure 12.7 depicts the behavior of the spread predicted by the model at all maturities, given
by,

B0
1
Rec
1
T
T
Spread = ln
= ln
1e
.
+e
T
N
T
N
Spread
240
239
238
237
236
235
234
233
232
231
0
Time to maturity
FIGURE 12.7. The term structure of bond spreads (in basis points) implied by an intensity model,
with recovery rate equal to 40% and intensity equal to = 0.04, implying an expected time-to-default
equal to 1 = 25 years.
Its a decreasing function in time-to-maturity. Eventually, as time to maturity gets large, the
bond becomes, so to speak, certain to default, with the unusual feature to deliver, for sure,
some recovery rate at some pointthe bond is certain to deliver the recovery rate. Indeed,
in Appendix 1, we show that if the recovery value of the bond is not constant, but shrinks
exponentially to zero as RecT R eT , for two constants R and , then, asymptotically, the
spread is
+
, if
lim s (T ) =
(12.10)
, if
T
A few additional issues. is the risk-neutral instantaneous probability of default, not the
physical probability of default, say. The ratio / is generally larger than one. Its inverse,
/, is an indicator of the risk-appetite in the credit market. Similarly, loss-given-default is
an expectation under the risk-neutral probability, and should contain useful indications about
market participants risk appetite.
Assume that under the risk-neutral probability, the instantaneous intensity of default for a
given rm is = 4%, annualized, and that under the physical probability, the instantaneous
probability of default for the same rm is = 2%, annualized. From here, we can compute the
probability of survival of the rm within 5 years, under both probabilities. They are:
e5 = e50.04 = 0.81873,
e5 = e50.02 = 0.90484.
Naturally, the probability of survival is lower under the risk-neutral measure.

359
c
by
A. Mele
Next, assume that the spread on a 5 year bond with face value N = 1, equals 3%. What is
the implied expected recovery rate from this spread? We have,
D0 = erT [Rec Q (Default) + N Q (Survival)] = erT [Rec (1 0.81873) + 1 0.81873] .
The spread is,

1
Rec (1 0.81873) + 1 0.81873
s0 = 3% = ln
.
5
1
Solving for Rec, gives, Rec = 23.16%.

12.2.3
Ratings
From
In practice, corporate debt is rated by rating agencies, such as Moodys and Standard and Poors.
Depending on the rating, corporate debt may be either investment grade or non-investment
grade (junk). Moodys ratings range from Aaa to C. Standard and Poors range from Aaa to
D. One can compute the probability of migrations based on past experience Transition
probabilities. Consider, for example, the following table:
One year rating transition probabilities (%), S&P's 1981-1991
To
AAA
AA
A
BBB
BB
B
CCC
AAA
89.1
9.63
0.78
0.19
0.3
0
0
AA
0.86
90.1
7.47
0.99
0.29
0.29
0
A
0.09
2.91
88.94
6.49
1.01
0.45
0
BBB
0.06
0.43
6.56
84.27
6.44
1.6
0.18
BB
0.04
0.22
0.79
7.19
77.64
10.43
1.27
B
0
0.19
0.31
0.66
5.17
82.46
4.35
CCC
0
0
1.16
1.16
2.03
7.54
64.93
D
0
0
0
0
0
0
0
D
0
0
0.09
0.45
2.41
6.85
23.19
100
TABLE 11.1
12.2.3.1 Foundations
A natural approach, then, is to assess credit risk by making reference to probabilities of default
built up on transition probabilities like those in Table 12.1.
Such an approch, also known as a migration approach, is somewhat less drastic than that
based on rare events, and hopefully more realistic. However, it is also technically more complex
than the intensity approach of the previous section. We provide the most foundational issues
related to this approach, leaving some details in the Appendix.
At time t, there exists several rating classes, N say, denoted as Ratt ,
Ratt {1, 2, , N} .
Transition probabilities of rating from time t to time T are,
P (T t)ij Pr ( RatT = j| Ratt = i) ,
i, j N.
We can build a Markov chain from here, by assuming that P (T t)ij only depends on T t.
Finally, we must have that,
N
P (T t)ij 0 and
P (T t)ij = 1.
j=1
360
c
by
A. Mele
For example, the probability of transition from rating Ratt = i to rating Ratt+1 = j in one
year is, P (1)ij . Table 12.1 contains one possible example of P (1)ij . The probability of transition
from rating Ratt = i to rating Ratt+2 = j in two years is P (2)ij , and is obtained as follows,
P (2)ij =
N

k=1
P (1)ik

Pr(transition from i to k in one year)
P (1)kj

Pr(transition from k to j in one further year)
More generally, we have, P (T ) = P (1)T , where P (T ) is the matrix with elements {P (T )ij }.
For example, the probability transition matrix P in Table 12.1 is,
89.1 9.63 0.78 0.19

0.3
0
0
0
0.86 90.1 7.47 0.99 0.29 0.29
0
0
0.09 2.91 88.94 6.49 1.01 0.45
0
0.09
0.06 0.43 6.56 84.27 6.44

1.6
0.18 0.45
P =
0.04
0.22
0.79
7.19
77.64
10.43
1.27
2.41
0 0.19 0.31 0.66 5.17 82.46 4.35 6.85
0
0
1.16 1.16 2.03 7.54 64.93 23.19
0
0
0
0
0
0
0
100
The 15 year transition matrix
20.01
3.38
1.17
0.64
P (15)
0.33
0.14
0
0
12.2.3.2 Evaluation
is:
35.82
30.28
13.12
6.76
3.22
1.65
1.08
0
23.91 9.92
4.05 3.06 0.43 2.66
32.71 15.91 6.38 5.11 0.77 5.34
34.21 21.93 9.69 8.01 1.29 10.33
22.21 22.40 12.42 11.93 2.09 21.39
10.71 13.616 11.36 14.68 2.78 43.16
5.01
6.75
7.48 13.17 2.64 63.04
3.54
3.90
3.51 5.60 1.22 81.02
0
0
0
0
0
100
The previous probabilities, {P (T )ij }, are meant to be taken under the physical world, not
the risk-neutral world. They can be used for risk-management purposes, but certainly not for
pricing. Indeed, historical default rates are too low to explain the price of defaultable securities.
A natural explanation relies on the presence of risk-premia. To use migration data for pricing,
it is vital to implement a number of steps.
First, clean up the data smoothing. For example, it might well be that downgrades from
class i to class i + 2 are more frequent than downgrades from class i to class i + 1. Moreover,
remove zero entries: although some rating event did not happen in the past, they might well
occur in the future. Second, add positive risk-premia to the previous smoothed data so as to
obtain realistic asset prices.
As regards pricing, according to the migration model, there are N classes of assets. Each
single asset may migrate from one class to another. Because evaluation is a dynamic business,
we cannot evaluate defaultable securities within a given asset class without simultaneously
evaluate the defaultable securities in the remaining classes. For example, there could be a
chance that a given asset will mutate into a dierent one in the next year. Given this, the
price of this asset, today, must reect the price of the asset in the other classes where it can
361
c
by
A. Mele
possibly migrate. Hence, we must simultaneously solve for all the asset prices in all the rating
classes. This approach, developed by Jarrow, Lando and Turnbull (1997), is quite complex and
is given a succinct account in the Appendix.
Consider the simplest case, which arises when the expected recovery rate is zero. In this case,
by Eq. (12.6),
D0,i
= erT (1 Qi (T t)) ,
N
where Qi (T t) is the risk-neutral probability the rm defaults, by time T , given it belongs
to rating i at time T .
More generally, by Eq. (12.6),

D0,i
rT Rec
=e
Qi (T t) + (1 Qi (T t)) .
N
N
The risk neutral probabilities, Qi (T t), must be found using migration frequencies such as
those in Table 12.1, which we must clean up and corrct with appropriate risk-premia, as
discussed.
Consider the following transition matrix:
From
A
B
Def
A
0.9
0.15
0
To
B
0.07
0.75
0
Def
0.03
0.10
1
where Def denotes the state of default. What is the probability that a name A will remain name
A in two years? What is the probability that a name A will default in two years?
Consider the following two year transition matrix:
0.90 0.07 0.03

0.90 0.07 0.03
Q (2) = 0.15 0.75 0.10 0.15 0.75 0.10 ,
0
0
1
0
0
1

Q(1)
Q(1)
such that:
Pr {A is A in 2 years} =
0.90
0.90 + (0.07) (0.15) + 0.03

0

AAA
ABA
= 0.8205,
ADef A
and
Pr {A defaults in 2 years} =
0.90
0.03 + (0.07) (0.10) +

AADef
= 0.064.
362
ABDef
0.03
1
ADef Def
c
by
A. Mele

In general, we have that:
Q (2)ij =
3

k=1
and for any T ,
Q (1)ik Q (1)kj ,
T
0.90 0.07 0.03
Q (T ) = Q (1)T = 0.15 0.75 0.10 .
0
0
1
Next, consider the following transition matrix, under the risk-neutral probability:
From
A
B
Def
A
0.80
0.15
0
To
B
0.20
0.75
0
Def
0
0.10
1
From here, we may easily compute, again, the (risk-neutral) probability A will default in two
years, and the probability B will default in two years. We have,

0.80 0.20 0
0.80 0.20 0
Q (2) = 0.15 0.75 0.10 0.15 0.75 0.10 ,
0
0
1
0
0
1

Q(1)
Q(1)
such that:
Pr {A defaults in 2 years} = Q (2)13

= 0.80
(0.10) +
0 + (0.20)

AADef
ABDef
= 0.02.
0
1
ADef Def
(multiply rst row by the third column), and,

Pr {B defaults in 2 years} = Q (2)23
= 0.15
(0.10) +
0 + (0.75)

BADef
= 0.175.
BBDef
0.10
1
BDef Def
(multiply second row by the third column).

Finally, suppose that the bonds issued by both A and B mature in two years. Furthermore,
assume that if these two bonds default, they pay o the same recovery rate, equal to 30%, and
only at the end of the second period. From here, we can compute the credit spreads for the two
bonds. We have,
erT Price A = (0.30) (0.02) + (1 0.02) = 0.986
1
Spread A = ln (0.986) = 7.0495 103 .
2
and,
erT Price B = (0.30) (0.175) + (1 0.175) = 0.8775
1
Spread B = ln (0.8775) = 6.5339 102 .
2
363
c
by
A. Mele
12.3. Derivatives on corporate assets
12.3 Derivatives on corporate assets

12.3.1
Callable and puttable bonds
Callable bonds are corporate bonds that give the issuer the right to buy them back at certain
times and predetermined prices. Puttable bonds, on the other hand, give the investor the right
to sell them back to the issuer at a certain strike price. Let us focus attention on callable
bonds. The next chapter, illustrates how to price these assets through trees. In this section, we
illustrate some useful properties of callable bonds, with the help of a few simple points. For
simplicity, we consider a callable, non-defaultable bond.
Let K be the strike price of the callable bond maturing at time S, and suppose that the date
of exercise, if any, is some future time T < S. Then, the payo of the callable bond at time T
is worth min {K, P }, where P is the price of a non-callable bond. Indeed, if K < P , then, the
issuer can buy its bonds back at K and re-issue the same bond at better market conditions, P .
The dierence, P K, is just a net gain for the issuer. So the callable bond is worth just K
when K < P . Instead, if K > P , the issuer does not have any incentives to exercise and, then,
the value of the callable bond is just that of a non-callable bond. therefore, the callable bond
is worth P when K > P .
It easy to see that,
min {P, K} = P max {P K, 0} .
Therefore, we see that the price of a callable bond with maturity date S, equals the price of
a non-callable bond with the same maturity date S, minus the value to call the bond, which
equals the price of an hypothetical option on the non-callable bond, struck at K.
We can apply these insights to price a callable option in a concrete example. Consider, for
example, the short-term rate in the Vasiceks model discussed in Chapter 11. Then, if the
short-term rate is r (t) at time t, the value as of time t of the non-defaultable zero coupon bond
maturing at time S, callable at time T < S, at a strike price equal to K, is,
P callable (r (t) , t, T, S) = P (r (t) , t, S) C b (r(t), t, T, S) ,
(12.11)
where P (r (t) , t, S) is the value of the non-callable zero maturing at time S, and C b (r(t), t, T, S)
is the value of a call option on the non-callable S-zero, maturing at time T and having a strike
price equal to K.
Eq. (12.11) shows that the presence of the option to call the bond raises the cost of capital
for the issuer.
In the context of the Vasiceks model, the solution to C b (r(t), t, T, S) in Eq. (12.11) is given
by the Jamshidians (1989) formula given in Chapter 11, which we now use below. Figure 12.8
below depicts the behavior of the price of the callable bond in Eq. (12.11), P callable (r, 0, T, S),
as a function of the short-term rate, r, when the exercise price, K = 0.65, and S = 10, T = 0.5,
= 0.2, = 0.03, = 0.06 , where , the unit risk-premium, equals 1.7146 102 .3
3 To
evaluate Eq. (12.11), we make use of the closed-form solution for the bond price, given by P (r, t, T ) = eA(T t)B(T t)r ,

2
(T t)
2
(T t) 2 , r
where the functions A and B are given by A (T t) = (T 1e
)
r 4
= 1
12
and
3 1e

B (T t) = 1 1 e(T t) .
364
c
by
A. Mele
0.70
0.65
0.60
0.55
0.50
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.10
short-term rate
FIGURE 12.8. Negative convexity. Solid line: the price of a callable bond. Dashed line: the price
of a non-callable bond. The price of a callable bond exhibits negative convexity with respect to the
short-term rate.
As Figure 12.8 illustrates, the convexity of the non-callable bond price is destroyed by the
convexity of the price of the option embedded in the callable bond. Intuitively, as the short-term
rate gets small, callable and non-callable bond prices increase. However, the price of callable
bonds increases less because as the short-term rate decreases, bond prices increase and then, the
probability the issuer will exercise the option, at maturity, increases. This makes the risk-neutral
distribution of the callable bond price markedly shifted towards the value of the strike price,
K = 0.65, which entails a progressively lower decay rate for the bond price as the short-term
rate gets small.
12.3.2
Convertibles
We only consider corporate convertible bonds. Convertible bonds oer bondholders the option
to convert their bonds into shares of the rm. The option to convert can be exercized at any
time up to maturity. How do they work? By denition, the face value of the convertible is,
Face value = $1 CR CP,
(12.12)
where CR is the conversion ratio, i.e. the number of shares this face value converts into, and
CP is the conversion price, i.e. the stock price implicitly dened by Eq. (12.12).
Typically, the bond is any like other xed income instrument, with coupon payments, callable
features, credit risk, etc. Callable features are almost invariably embedded into this type of
contracts. The parity, or conversion value, is the value of the bond if the bondholders decide
to convert. It is dened as,
CV = CR S,
where S is the price of the common share. Not only is the convertible bond price aected by
interest rates, credit risk, timing risk, etc. This price is also aected by the movements of the
underlying stock price. This is quite natural as there is a positive probability that the bond
365
c
by
A. Mele
will become a share in the future. To emphasize this, we also say that convertible bonds
are hybrid instruments. The embedded option oers the bondholders the possibility to obtain
equity returns (not just bond returns) in good times, while oering protection against the
downside. As mentioned earlier, convertible bonds are almost always callable. The holder can
always convert the bond, once it has been called. The rationale behind callability is to induce
the bondholder to convert the bond earlier.
The pricing problem of convertible bonds has been intensively studied, theoretically. Ingersoll (1977) provides the rst theoretical article which lays down the foundations to rational
evaluation of convertible callable bonds. Let us dene the dilution factor, denoted as , as the
fraction of common equity that would be held by the convertible bondowners if the entire issue
was converted. If there are nout shares outstanding, and the convertible bond can be exchanged
for n shares, then, in aggregate,
n
=
.
n + nout
Let V be the market value of the rm and B conv (V, ; N ) the aggregate value of the convertible
bond with time to maturity and balloon payment N. To simplify the presentation, we do not
consider callability issues. However, we shall provide some intuition about this issue later. Let us
assume that the stocks and the convertible bonds are the only two claims in the capital structure
of the rm. Since, after conversion, only the stocks will remain, then, the post-conversion value
of the convertible bonds is simply the conversion value of the convertible, i.e. V . Moreover,
we have, for any 0,
(12.13)
V B conv (V, ; N) V.
The rst inequality in (12.13) is simple to understand. Indeed, suppose that B conv (V, ; N) <
V . Then, we can purchase the convertibles, convert them into shares and, nally, sell the shares
for V . The second inequality follows by limited liability equity holders, and the ModiglianiMiller theorem.
At maturity, we have that,
B conv (V, 0; N) = min {V, max {N, V }} .
(12.14)
max {N, V } is the value of the convertible, in case of no-default. Then, min{V, B}
Indeed, B
in case of no-default.
is what the rm will pay, to the bondholders: V in case of default, and B
It is possible to show that it is never optimal to exercize the option to convert before maturity.
Therefore, to price the convertible bond, we only need to be concerned with the risk-neutral
evaluation of the terminal payo in Eq. (12.14).
We can re-express the terminal payo in Eq. (12.14) in a manner that allows a better understanding of the issues underlying the exercise of the convertibles. In particular, we have
that,
B conv (V, 0; N ) = min {V, max {N, V }} = max {V, min {V, N }} .
(12.15)
min{V, N}, which is what the rm is ready to pay, to the bondholders, if the
Indeed, let B
is obviously the payo
bondholders do not exercise the option to convert. Then, max{V, B}
prole to the bondholders.
The terminal payo in Eq. (12.15) illustrates very clearly that convertible bonds embed an
option to convert - on top of the plain vanilla non-convertible bond. Intuitively, at maturity, a
non-convertible bond is worth min {V, N }, and the option to convert is either worthless (in case
of non conversion) or V N (in case of conversion), i.e. it is max {V N, 0}. This intuition
366
c
by
A. Mele

is conrmed, mathematically, as we have that:
max {V, min {V, N }} = min {V, N} + max {V N, 0} .
Therefore, the value of the terminal payo is, by Eq. (12.15),
B conv (V, 0; N) = min {V, N } + max {V N/, 0} .
(12.16)
B conv (V, ; N) = B (V, ; N ) + W (V, ; N/) .
(12.17)
Eq. (12.16) shows that the current value of the convertible bond is the sum of the value of
a straight bond plus the value of options on the rm with strike price equal to N/.
Accordingly, let B (V, ; N ) and W (V, ; N/) be the prices of the straight bond and the option
on the rm. We have,
We may use the Mertons (1974) model to nd the price of the straight bond, B (V, ; N). By
the results in Section 12.2, it is:

1 2

ln
(
V
/
N
)
+
r
+
B (V, ; N ) = V (d1 ) + Ner d1 , d1 =

,
(12.18)

where is the instantaneous volatility of the asset value, r is the (constant) instantaneous
short-term rate, and is the cumulative distribution of a standard normal. Similarly, we may
use the Black-Scholes model to compute the function W .
Eq. (12.18) reveals the intuitive property that as V gets large, B (V, ; N ) N er : the
probability of default gets extremely tiny as the value of the assets gets large. Moreover, the
Black-Scholes model suggests that W (V, ; N/) V er N/ as V gets large. Therefore, by
Eq. (12.17), we have that, for large V , B conv (V, ; N) Ner + (V er N/) = V . Eq.
(12.18) also shows that for small values of V , B conv (V, ; N) 0.
To sum-up, the value of the convertible bond is less than the value of the rm, V , and larger
than the conversion value, V . Moreover, it approaches V , as the value of the rm gets large.
Figure 12.9 below illustrates the shape of the convertible bond price, as a function of the value
of the rm.
V
Convertible bond value
Ne r
Straight bond value
FIGURE 12.9. The value of a convertible bond
The value of a callable convertible bond is between the value of the straight bond and the
value of the convertible bond.
367
12.4. (Credit-) risk shifting derivatives and structured products
c
by
A. Mele
12.4 (Credit-) risk shifting derivatives and structured products

12.4.1
Securitization, and a brief history of credit risk and financial innovation
Securitization is a process by which some illiquid assets are transformed into a package of securities backed by these assets, through packaging, credit and liquidity enhancements. Two leading
examples are: (i) mortgages and (ii) receivables. Financial institutions nd the securitization
process attractive, as they can carve out certain items in their balance sheet, thus boosting
their return on investments or simply because by securitizing assets, less capital is needed to
meet capital-requirements standards. For example, the accounts receivables of a corporation
may be used to back the issue of commercial paper known as asset-backed commercial paper.
Securitization is a way (not the only way) to trade/transfer credit risk, in principle.
What are the origins of credit derivatives and nancial innovation? The rst interest rates
derivatives were created around the mid 80s. In the late 80s the business proliferated and
become fairly complex. But nancial innovation is easy to imitate, which led banks to become
increasingly creative, so as to exploit their initial competitive advantage as longer as possible.
During the early 1990s, just after the 1991 recession, interest rates were quite low, and the
volatility of capital markets extraordinarily low. Derivatives, then, could be used as devices
to boost investors returns. JPMorgan introduced many structures such as, LIBOR squared,
inverse oaters, power options, convexity forwards, etc.
However, after the 1994 nancial turmoil, the interest rate climate suddenly changed, and
many of the derivatives contracts produced large losses. There was a call for regulation by public
opinion and certain policy makers. At the same time, the International Swaps and Derivatives
Association argued that more regulation would destroy markets creativity. These regulatory
pressures vanished by the mid 1990s.
During the mid 1990s, the markets started to innovate, again, with a view that risks could
be assessed, and controlled, through market discipline, rather than through regulation. Swaps
markets recovered. They did do slowly though, as these products were already in the end of the
innovation cycle. They had been massively imitated, and margins for prots had consequently
been eroded. They had become, so to speak, a mass-product. The markets, then, were ready
for a new major innovation wave. The natural innovation had to with credit. Market players
such as JPMorgan, Credit Suisse, Bankers Trust soon realized that borrower defaulting was a
source of substantial risk, which could be conveniently reallocated through the use of dedicated
derivatives. Credit risks could be transferred in pretty much the same way as market risks
can be transferred through the underwriting of options written on stocks, or on interest rates.
JPMorgan had serious motivations to innovate, as its books contained vast pools of loans, which
could be used as practical material to experiment with. Importantly, these loans required too
many reserves and were consequently expensive.
The main idea, then, was to repackage the loans into derivatives, in a way that default risk
and/or part of the securitized loans, or both, could be sold to outside investors. In a sense,
then, credit derivatives were also a regulatory mitigation device, partly useful as a response to
regulation. The idea was simple: turn loans into derivatives that could be sold, and/or create
new insurance product such as credit default swaps. At the very beginning, derivatives were just
designed to have single loans as underlying. Afterwards, the idea emerged to create structures
organized in derivatives bundles, with cash ow indexed in some way to baskets of loansthe
ancestors to collateralized debt obligations. For example, JPMorgan created Bistro (Broad
Index Secured Trust Oering), a structure relying on a variety of assets, ranging from corporate
368
c
by
A. Mele
debt to student loans. ABN - Amro created similar structures, Heineken and Amstel. But
then, competition increased and prot margins fell, which triggered the need for new innovation.
As explained in Section 12.4.7, these products were channeled through o-balance-sheet vehicles
escaping national supervision, a sort of shadow banking system. Basel II was not yet in place.
The response to increased competition was the creation of structured products having riskier
and riskier assets. For example, in the mid 1990s, derivatives teams begun to interact with
teams managing loans extended to borrowers with poor credit historysubprime mortgages. As
a result, subprime loans begun to be securitized and then structured into CDOs. JPMorgan was
not the leader in the creation of these products, compared to other institutions such as Merrill
Lynch or UBS. Ironically, JPMorgan itself bought Bear Stearns during the 2008 springtime.
The subprime turmoil arose out of mechanics that are by now well-understood. First, there
was a boom, sustained by (i) low interest rates and house price appreciation and (ii) a business
model that changes from buy to hold to originate and distribute, as explained earlier.
After the boom, the burst, caused by increasing interest rates and falling housing prices.
Evaluation models, if any, relied on the assumption delinquencies would remain the same, and
small risk-aversion adjustments to the calculation of expected actuarial losses were made (if
any). The picture below shows this wasnt true and that in fact, the pieces of information
emanating from those simple pictures could have helped predict the crisis. Finally, correlation
issues were simply ignored or, at best, badly calibrated.
Section 12.4.7 provides a more systematic analysis of these issues, but it is instructive to
discuss since now, some of the causes leading to the burst and the 2007 crisis. One of them
is certainly related to model misspecication, or an inappropriate rating mapping system,
by which rating agencies used to tend to transplant the rating system for corporation to
structured products relyin on MBS. A second diculty was the presence of a true shadow
banking system escaping the attention of the ocial nancial community. The dynamics of
the crisis were a sharp liquidity dry-up, then a credit crunch, followed by a drop in the real
economic activity, which further fed the credit crunch, etc. In that context, it is quite dicult
369
c
by
A. Mele
to draw the line between liquidity squeezes and solvency issues.
Mortgage Delinquencies by Vintage Year (60+ day delinquencies, in percent of balance).

Source: IMF, Global Financial Stability Report, April 2008.
370
c
by
A. Mele
Left hand side panel: U.S. and European House Price Changes. Right hand side panel:
U.S. Mortgage-Related Securities Prices. Source: IMF, Global Financial Stability Report,
April 2008.
12.4.2
Total Return Swaps (TRS)
In a total return swap, or TRS, one party (who owns some asset, the asset underlying the
TRS) receives from the counterparty payments based on a mutually agreed rate, either xed or
variable, and makes payments to the counterparty based on the return of the underlying asset,
which includes both the income it generates and any capital gains. The underlying asset can
be a loan, a bond, an equity index, or a basket of assets. The interest payments are typically
based on the LIBOR plus a spread. Consider the following example. Party A receives LIBOR
+ xed spread equal to 3%. Party B receives the total return of the S&P 500 on a principal
amount of $1 million. If the LIBOR is 7% and the S&P 500 is up by 12%, A pays B 12% and
B pays A 7% + 3%. By netting, A pays B $20,000, i.e. $1 million (12% 10%).
While TRS are usually categorized as credit derivatives, they combine both market risk and
credit risk. The benets from longing a TRS relate to the fact that the party with the asset
on the balance sheet buys protection against loss in value. On the other hand, shorting a TRS
allows the counterparty to receive the payos guaranteed by the asset without necessarily having
to put it in the balance sheet. Hedge funds nd it quite convenient to short a TRS, as this allows
them to have views with limited collateral upfront. The market for TRS is over-the-counter and
market participants include institutions only.
371

12.4.3
c
by
A. Mele
Spread Options (SOs)
In general, SOs are options written on the dierence between two indexes. For example, let
S1 (T ) and S2 (T ) be the prices of two assets at time T . The payo promised by a SO entered
at some time t < T , might be max {S1 (T ) S2 (T ) K, 0}, where K is the strike of the SO.
A SO can be written on the spread between two rates of returns too. Importantly, a SO can be
written on the spread between the yield of a corporate bond and the yield of a Treasury bond.
Examples include: (i) NOB spread (notes - bonds), which are spreads between maturities; (ii)
Spreads between quality levels, such as the TED spread (treasury bills Eurodollars); (iii)
MOB spreads, i.e. the dierence between municipal bonds and treasury bonds. More generally,
the denition of a SO has now been extended to include payos written as a linear combination
of indexes, interest rates and yields.
12.4.4
Credit spread options (CSOs)
In a CSO, the payo is the dierence between (i) the spread between two reference securities
(say Italian Government bonds and US Government bonds having the same maturity, or the
spread between the share on xyz and LIBOR, or two credit instruments), and (ii) a given strike
spread, for a certain maturity date. It may be an American or European option. So CSOs
allow to hedge against, or take specic views about, changes in credit spreads. For example,
an investor, while bullish on Italian bonds, might hedge against the uncertain outcome of a
political election, which could trigger a widening of short-term spreads of Italians versus US.
The investor, then, may long a CSO, with time to maturity around the days of the political
election, where the underlying are the Italian and Government bonds expiring in ten years, say.
A possible payo to the CSO holder can be proportional to, (ITA/US K), where ITA/US is
the ten year Italian-US spread in three months, and K is the strike spread.
12.4.5
Credit Default Swaps (CDS)
12.4.5.1 CDS on single names
CDS dier from TRS insofar as they provide protection against a credit event. TRS, instead,
provide protection against a loss in asset value, which could be triggered by both market or
credit riskm although it is obviously more often market risk than credit to kick in.
The premium, assumed to be paid quarterly, on a CDS contract at time t, is obtained by
equating the expected discounted value of the premium paid over the life of the contract, i.e.
at dates t < t1 < t2 < < tM , where ti = t + 4i , M = 4 N, and N is the number of years the
CDS refers to,
4N

Premiumt =
er(ti t) CDSt (N ) Pr {Survival at ti } ,
i=1
with the expected discounted value of the protection,

Protectiont =
4N

i=1
er(ti t) LGD (ti ) Pr {Default (ti1 , ti )} ,
where r is the (constant) risk free rate, CDSt (N ) is the premium paid every quarter, prevailing
at time t, and LGD (ti ) is the Loss-Given-Default at time ti , which for simplicity is assumed to
be constant, i.e. known at time t.
372

Equating Premiumt and Protectiont , and solving for CDSt , leaves:
4N r(ti t)
e
LGD (ti ) Pr {Default (ti1 , ti )}
CDSt (N ) = i=1 4N
.
r(ti t) Pr {Survival at t }
e
i
i=1
c
by
A. Mele
(12.19)
At rst glance, the previous derivation might look like actuarial, although it is not, actually.
The reason is that the probabilities in Eq. (12.19) are risk-neutral probabilities. As such, they
are, obviously, the same as those we use to price the bonds underlying the CDS contract.
Therefore, there must be no-arb relations linking bond prices to CDS premiums, which shall
be emphasize later on (see Section 12.4.5.4). This point illustrates in a remarkable way one key
dierence between nance and insurance. Even if in insurance, one may end up pricing some
products through risk-adjusted probabilities, nance is where we typically end up having many
more traded risks than in insurance, and these risks are tightly related to no-arb restrictions.
Eq. (12.19) is a general formula we can use, once we have a model determining the risknuutral probability of default. In this chapter, we implement Eq. (12.19) through a reducedform approach, which will allows us then to nd the quarterly premium (or spread) CDSt (N)
quite easily, as follows.
We have, denoting again with the instantaneous probability of default, that Pr{Survival at
ti } = e(ti t) , and that Pr{Default at any z (ti1 , ti )} = e(ti1 t) e(ti t) . Intuitively, if
the name survives at ti (event Ei ), it must necessarily have survived at ti1 (event Ei1 ), but
the converse is not true: Ei Ei1 , and the complement of Ei to Ei1 is nothing but the event
of default between ti1 and ti .4 Substituting the previous probabilities into Eq. (12.19), we nd
that:

4N r(ti t)
LGD (ti ) e(ti1 t) e(ti t)
i=1 e
CDSt (N ) =
.
(12.20)
4N (r+)(t t)
i
i=1 e
For example, if LGD (ti ) is constant and equal to LGD for each ti , then, for t = ti ti1 = 14 ,
CDSt (N ) LGD t (expected losses per unit of time) t,
(12.21)
where the approximation is obtained by making e(ti1 t) e(ti t) e(ti t) t. Naturally,

is the risk-neutral instantaneous probability of default for the security.
Note, Eq. (12.21) shows that the CDS premium is approximately the same as the instantaneous spread of a defaultable bonds, as explained in Section 12.2. This property is to be
expected, so to speak, as a purchase of a defaultable bond and protection on it is nothing
but a synthetic default-free bond. Therefore, there must be a no-arbitrage relation between
CDS spreads and defaultable bond spreads, as we anticipated earlier. However, in general, Eq.
(12.21) does not hold, as the assumptions made to achieve it ( is constant, LGD is constant,
r is constant, etc.) are quite unrealistic. On the contrary, we often observe CDS spreads curves
that increase with maturity, as we shall explain in more analytical detail in Section 12.4.5.4.
Indeed, we may take interesting views. For example, buying CDS for 2Y and sell CDS for 3Y
is a view that default will not occur between the second and the third year from now.
12.4.5.2 CDS on indexes
A CDS index is a basket of credit entities in which the protection buyer, pays the same premium, called the fixed rate, on all the names in the index. Credit events are typically bound
4 Mathematically,
e(zt) dz.
we have that Pr{Default at any z (ti1 , ti )} =
373
ti
ti1
Pr{Default at z}dz, where Pr{Default at z} =
c
by
A. Mele
to bankruptcy or delinquencies. After a credit event, the entity is removed from the index and
the contract goes through, although with a reduced notional amount, until expiration. While
CDS on single names are over-the-counter, CDS indexes are completely standardized and can
be more liquid, as historical data on bid-ask spreads show. In fact, it can be cheaper to hedge
a portfolio of CDS or bonds with a CDS index than it would be to buy many CDS to achieve
a similar eect. There exist two main indices: (i) CDX index, which contains North American
and Emerging Market companies; and (ii) iTraxx index, which contains companies from the
rest of the world
12.4.5.3 Disentangling default probability from risk-aversion
The following picture, taken from Fender and Hordahl (2007), illustrates the behavior of the
credit market risk appetite before the 2007 credit market turmoil.
FIGURE 12.10. Antonio Mele does not claim any copyright on this picture, which is taken from Fender
and Hordahl (2007). The picture has been put here for illustrative purposes only, and permission to
the authors shall be duly asked before the book will be published.
How did the authors estimate the price of risk? Consider the expected losses under the
actuarial, or physical probability for a given security. The counterpart to Eq. (12.21), under the
physical probability, is:
Expected LossesP P LGD t,
where P is the physical instantaneous probability of default for a given security. Assume that
LGD is constant, to simplify. If the investors require compensation for the default event, then,
the actuarial losses should be less than the CDS spread, i.e. Expected LossesP < CDS, or,
> P .
374
c
by
A. Mele
The risk-premium is dened as the dierence between the actuarial losses, Expected LossesP ,
and the CDS premium,

Risk-Premium = P LGD t.
The price of risk in Figure 12.10 is dened as the ratio of the CDS spread over Expected LossesP ,
Price-of-Risk =
.
P
Early references to estimation methods are Due et al. (2005) and Amato (2005). Typically,
Expected LossesP are proxied by Moodys KMVs Expected Default Frequencies (EDFsTM ),
obtained through fully specied structural models for credit risk. The next pictures are taken
from Amato (2005). As we can see, during the 2003-2005 period, credit spreads were so low,
and this in turn gave incentives to CDO issuers to look for illiquid and relatively more complex
assets to put as collateral, which led to the issuance of CDO relying on ABS such as MBS, or
CDO2 , explained below.
FIGURE 12.11. Antonio Mele does not claim any copyright on this picture, which is taken from Amato
(2005). The picture has been put here for illustrative purposes only, and permission to the author shall
be duly asked before the book will be published.
375
c
by
A. Mele
FIGURE 12.12. Antonio Mele does not claim any copyright on this picture, which is taken from Amato
(2005). The picture has been put here for illustrative purposes only, and permission to the author shall
be duly asked before the book will be published.
The following picture illustrates the behavior of CDS indexes during approximately 20 years
before the 2007-2009 credit market turmoil.
FIGURE 12.13. Valuation of Financial Instruments Based on Implied Probability of Default. Antonio
Mele does not claim any copyright on this picture, which is taken from IMF (2008). The picture has
376
c
by
A. Mele
been put here for illustrative purposes only, and permission to the authors shall be duly asked before
the book will be published.
12.4.5.4 Continuous time
We may relax the assumption the instantaneous intensity of default, , is constant. This intensity is dened under the risk-neutral probability and can change either because the intensity
of default under the physical probability changes or because risk-appetite changes, or both.
We aim to examine the asset pricing implications of time-varying intensities, by rst exploring
how probabilities of survival change in a simple setting, where we do not single out the reasons
leading to variations in .
First, we assume the instantaneous probability of default can only change at discrete times,
giving rise to random intensities t , meaning that t is the intensity of default in the time interval
[t 1, t]. Let Ft be the information set as of time t. We assume that t is Ft -measurable. What
is, then, the probability of survival of any given name in this case? We have, by Bayess theorem,
Pr { Surv at t| Surv at t 1} =
Pr {Surv at t}
.
Pr {Surv at t 1}
(12.22)
By a repeated use of Eq. (12.22),

Pr {Surv at t} = Pr { Surv at t| Surv at t 1} Pr {Surv at t 1}
=
=
t
1
n=1
Pr { Surv at n| Surv at n 1} .
(12.23)
So we are left with nding Pr { Surv at n| Surv at n 1}. Consider the following arguments.
n , then, Pr {Surv at n| Surv at n 1} = en .
If n was not random and xed at some
When n is random, en is the probability of survival, conditioned upon some particular
value the intensity could possibly take. Heuristically, then, Pr { Surv at n| Surv at n 1} =

n (s)
Pr {s}, where n (s) is, so to speak, the value n would take in state s, Pr {s}
sS e
is the likelihood that state s occurs
and, nally,
S is the set of all possible states. Therefore,
Pr { Surv at n| Surv at n 1} = E en Fn1 , where E denotes the expectation taken under
the risk-neutral probability. Inserting this result into Eq. (12.23), and using the Law of Iterated
Expectations, leaves:
t

n
n=1
Pr {Surv at t} = E e
.
Under regularity conditions, we can easily extend the previous result to a continuous time
setting. For example, we may assume that the risk-neutral default intensity, (t), is solution
to:

(t) dt + (t)dW (t) , (0) = .
d (t) =
(12.24)
and are
where W is a standard Brownian motion under the risk-neutral probability, and ,
three positive constants. This is the same as the Cox, Ingersoll and Ross (1985) (CIR) model
of the short-term rate reviewed in Chapter 11. Therefore, under the parameter restrictions in
Chapter 11, (t) is always positive, and
#t

(s)ds
Psurv (, t) Pr {Surv at t} = E e 0
.
(12.25)
377
c
by
A. Mele
Eq. (12.25) is, formally, the same as the Feynman-Kac representation of a solution to a PDE,
solved by a bond price in the CIR model. In other words, the model for the survival probability
in Eqs. (12.24)-(12.25) has the same mathematical structure as that leading to the price of a
bond in the CIR model. Therefore, a closed-form solution is available for Psurv (, t). It is given
by:
(N) =
Psurv (, N) = (N ) eB(N) ,
% 22

1
2 eN 1
2e 2 (+)N
,
, B (N ) =
( + ) (eN 1) + 2
( + ) (eN 1) + 2

2 + 2 2 .
(12.26)
More generally, we can build up a whole family of models with a closed-form solution, the
ane class reviewed in Chapter 11, by just assuming that:
(t) = 0 + 1 y (t) ,
(12.27)
where 0 is a constant, 1 is a vector of constants, and y is a multivariate jump-diusion process,

with drift and diusion terms as in Section 11.3.6 of Chapter 11. This model is interesting, as
we can judiciously choose the components of y (t) which we suppose may aect the default
intensity. For example, some of them could be unobservable, and others could be observable,
and relate for example to the business cycle or even the structure of the rm.
So given any solution for the survival probability predicted by any of these ane models
when y (0) = y, Psurv (y, t) say, we can easily compute
Pr{Default (ti1 , ti )} = Psurv (y, ti1 ) Psurv (y, ti ) .
(12.28)
We can then look at the bond spreads and the CDS spreads implied by this modeling choice.
In Appendix 3, we show the price of a defaultable pure discount bond expiring in N years is:
" N
rN
P (y, N ) = e
Psurv (y, N ) +
ert Pr{Default dt}Rec (t) dt,
(12.29)
0
where Rec (t) denotes the recovery value in case of default, supposed to be known. This evaluation result is, naturally, consistent with a similar derivation provided in Section 11.3.7 of
Chapter 11, although in this chapter we are emphasizing more survival arguments.
As for the CDS spreads, we have, by Eq. (12.19),
4N r(ti t)
e
LGD (ti ) [Psurv (y, ti1 ) Psurv (y, ti )]
CDSt (N) = i=1
,
4N r(t t)
i
Psurv (y, ti )
i=1 e
where N is, again, the number of years the CDS refers to, and ti = t + 4i .
Assume the short-term rate, r, is zero, and that loss-given-default is constant and equal to
LGD. Then, as shown in Appendix 3, the price of a defaultable pure discount bond, P (, N ),
and the current CDS premium, CDS0 (N ), are given by:
P (, N) = 1 LGD (1 Psurv (, N )) ,
1 Psurv (, N )
CDS0 (N) = LGD 4N
.
P
(,
t
)
surv
i
i=1
(12.30)
Figure 12.14 depicts the bonds spread, 1

ln P (, N ), and the annualized credit default
N
= 0.04 and
spreads, 4 CDS0 (N), when the parameters in Eq. (12.24) are = 0.25,
378
c
by
A. Mele
with loss-given-default LGD = 0.60, and two values of the current intensity: = ,
= ,
and = 0.02. Assuming that LGD is constant is not a plausible, empirically. Instead, we know
LGD moves countercyclically for most names, although it does not exhibit strong business cycle
features, for sovereigns. For sovereigns, the size of the country and debt distribution seem to
be by far more important.
Spreads, in basis points, for average default intensity
Spreads, in basis points, for low default intensity
240
180
bond spreads
CDS spreads, annualized
235
170
bond spreads
CDS spreads, annualized
230
225
160
220
150
215
210
140
205
130
200
195
120
10
years
10
years
FIGURE 12.14. Spreads on bonds and CDS predicted by the ane model (12.24). The left panel
= 0.04. The
depicts the spreads when the current default intensity equals the long-run mean, =
right panel depicts the spreads in good times, i.e., when the current intensity of default takes a low
value, = 0.02. In each case the recovery rate equals 40%.
The mechanism is that good times are followed by bad times, and so when = 0.02, we
expect default rates to rise in the future. As a consequence, spreads are increasing in maturity.
Moreover, we easily see that bond spreads are approximately equal to CDS spreads at short
maturities. At longer maturities, the two spreads diverge, with CDS spreads, 4 CDS0 (N ),
dominating bonds spreads, 1
ln P (, N). Moreover, we have that the two curves are decreasing
N
in time to maturity even when the current value of the intensity equals the long-run one, .
379
c
by
A. Mele
Where do these two properties originate from? The rst one follows because we have, approximately,
1
1
ln P (, N) =
ln [1 LGD (1 Psurv (, N ))]
N
N
1 Psurv (, N )
LGD
N
1 Psurv (, N)
LGD 1 4N
i=1 Psurv (, ti )
4
= 4 CDS0 (N ) .
As regards the second property, its a convexity eect. We can tackle this issue using arguments similar to those we made for another
topic in Chapter 11, Section 11.3.4. For the bond
+ es
, we have, approximately,
spreads, since E ( (s)) =
1
1
ln P (, N) =
ln [1 LGD (1 Psurv (, N ))]
N
N

#N
1
(s)ds
ln 1 LGD 1 E e 0
=
N

#N
1
E((s))ds
ln 1 LGD 1 e 0
N
#N
1 e 0 E((s))ds
LGD
N
N
N
() 1e
1e
,
= LGD
N
then, bond spreads are bounded away by a decreasing function (in N ).

so that even if = ,
Of course, it doesnt necessarily mean that bond spreads have to be decreasing as well, but that
function helps this happening. As for the CDS spreads, we have, approximately:
4 CDS0 (N ) = LGD
1
4
1 Psurv (, N )
1 Psurv (, N)
1
LGD
ln Psurv (, N ) ,
4N
N Psurv (, N )
N
i=1 Psurv (, ti )
CDS0 (N ) is bounded away by a decreasing function (in N ), for the same

such that for = ,
arguments made as regards the bond spreads, N1 ln P (, N).
12.4.5.5 A trading strategy
Bond prices and CDS spreads are driven by the same state variable, the default intensity, and
so they are restricted to lie on some space, to be consistent with no-arbitrage. To illustrate,
consider, rst, the simple case where the default intensity is constant, such that CDS spreads
are given by Eq. (12.20). Given this model, we can look at the market data for CDS spreads,
380
c
by
A. Mele

and infer the risk-neutral intensity, as in the picture below.
Inferring riskneutral intensity from CDS market data
CDS spreads, modelbased, in basis points
350
300
250
200
150
100
50
0.01
0.02
0.03
Default intensity
0.04
0.05
In this picture, the CDS spreads predicted by Eq. (12.20) are depicted as a function of the
risk-neutral intensity, , assuming N = 5 years, LGD = 0.60 and the short-term rate r is zero.
For example, if we had to observe a CDS equal to 200 basis points, we would infer a value of
= 0.033. The key point is this very same
should be pricing the
approximately equal to
zero as well, such that for N = 5,
P (N ) = 1 LGD 1 eN = 0.90874,
and so we might go long (short) the zero if its market price is lower (higher) than 0.90874.
Naturally, this example is based on the unrealistic assumption that the default intensity is
constant. But the same strategy can be used in the more general case where default intensities
are stochastic. In this case, bond prices and CDS spreads should also be restricted, by noarbitrage. The picture below shows the restrictions between bond spreads and CDS spreads,
obtained with the same parameter values as those used to produce Figure 12.14, and values of
381
c
by
A. Mele

current default intensities ranging from nearly zero to up to 0.05.
Noarb restrictions between bond spreads and CDS spreads
Bond spreads, modelbased, in basis points
260
240
220
200
180
160
140
120
100
80
80
100
120
140
160
180
200
220
CDS spreads, modelbased, in basis points
240
260
12.4.5.6 Hazard rates
In a pricing context, the relevant probabilities of survival are obviously conditioned upon the
time of evaluation, time 0 say. For example, the probability of default in Eq. (12.28) is only conditioned to the information we have at time zero. More generally, the probability of defaulting
in the interval of time (ti1 , ti ), conditional upon survival at time t < ti1 , is:
Pr{Default (ti1 , ti )| Survival at t} =
Psurv (y, ti1 ) Psurv (y, ti )

.
Psurv (y, t)
(12.31)
For example, for t = ti1 , and (ti1 , ti ) small, and deterministic, a simple approximation to
this conditional probability can be,
P
t surv
(y, t)
(ti ti1 )
Psurv (y, t)
pdefault (y, t)
(ti ti1 )
1 Pdefault (y, t)
= (t) (ti ti1 ) ,
Pr{ Default (ti1 , ti )| Survival at t}
with straight forward notation. The previous expressions are known as hazard rates. They coincide with (t) dt, when (t) is deterministic. If (t) is not deterministic, simple computations
lead to:
Pr{ Default (t, t + dt)| Survival at t} = EQ [ (t)] dt,
(12.32)
where Q is a new probability, with Radon-Nikodym derivative given by:
#t

(s)ds

dQ
e 0
=
.
dQ F0 Psurv (, t)
382
(12.33)
c
by
A. Mele
Accordingly, under Q , the state variables in Eq. (12.27) follow a diusion process, with a drift
process tilted, due to this change of measure. For example, in the simple setting of Eq. (12.24),
we have that, for a xed t,

d (s) = (B0 B1t (s) (s)) ds + (s)dW (s) , s (0, t] , (0) = ,
(12.34)
Bt (s) = + B (t s) 2 , B () as in Eq. (12.26),
B0 = ,
1
where W is a Brownian motion under Q . Therefore, by Eq. (12.32), and computations,

" t
#x
t
G (s)
Pr{ Default (t, t + dt)| Survival at t} =

+ B0
ds dt, G (x) e 0 B1 (u)du .
G (t)
0 G (t)
Appendix 4 provides a proof of these results, which to the best of our knowledge, are developed
here for the rst time.
12.4.5.7 Extracting probabilities of default from market data
Market data obviously convey information about probabilities of default, which might be extracted from these data, under a number of assumptions. To illustrate this possibility in a simple
case, assume that the recovery rate is zero, and that the short-term rate and the instantaneous
probability of default are both continuous time Markov and independent of each other. Then,
the price of a defaultable zero is: Pdef (, N ) = P (N) Psurv (, N), where Pdef (, N ) is the price
of a defaultable zero and P (N ) is the price of a non-defaultable bond. Therefore, we can read
the risk-neutral probability of survival from the defaultable/non-defaultable bond price ratio:
Psurv (, N ) =
Pdef (, N )
.
P (N )
(12.35)
Naturally, surviving until some time N2 means having survived until some time N1 < N2 and
having survived from N1 to N2 . Therefore, Psurv (, N2 ) = Psurv (, N1 ) Psurv (, N1 , N2 ), where
Psurv (, N1 , N2 ) is the risk-neutral probability of survival between N1 and N2 . Using Eq. (12.35),
then, we can extract this probability, as follows:
Psurv (, N1 , N2 ) =
Pdef (, N2 ) P (N1 )
.
Pdef (, N1 ) P (N2 )
The previous example relies on the simplifying assumption of a zero recovery rate, but it can
be generalized to the case where the recovery rate is nonzero. But in this case, bond prices would
convey information about both probabilities of default and recovery rates, an identication issue
to be dealt with.
12.4.6
Collateralized Debt Obligations (CDOs)
12.4.6.1 A crash description
CDOs are securitized shares in pools of assets. Collateral assets include loans or debt instruments. A CDO may be a collateralized loan obligation (CLO) or collateralized bond obligation
(CBO) according to whether it relies only on loans or bonds, respectively. CDO investors bear
the credit risk of the collateral. Multiple tranches of securities are issued by the CDO, oering
investors various maturity and credit risk characteristics. Tranches are categorized as senior,
mezzanine, and subordinated, or junior, or equity, according to their degree of credit risk. If
383
c
by
A. Mele
there are defaults or the CDOs collateral otherwise underperforms, scheduled payments to
senior tranches take precedence over those of mezzanine tranches, and scheduled payments to
mezzanine tranches take precedence over those to junior tranches. Typically, senior tranches
are rated, with ratings of A to AAA. Mezzanine are also rated, typically with ratings of B to
BBB. In principle, these ratings should reect both the credit quality of the collateral and the
protection a given tranche is given by the tranches subordinating to it. CDOs are part of a
more general securitization process, and can also include mortgages, as in the stylized example
below.
(i) In a rst step, subprime mortgages are securitized, as illustrated below:
Subprime
Mortgage
Monthly
payments
Subprime
Mortgage
Subprime
Mortgage
Asset
Backed
Security
(ABS)
Monthly
payments
ABS
investor
ABS
investor
ABS
investor
(ii) In a second step, a CDO is created, out of the secutized subprime mortgages and additional
Asset Backed Secutities (ABS):
Subprime
ABS
ABS relating to other
forms of collateral
(e.g. corporate debt)
Collateralized
Debt Obligation
(CDO)
Subprime
ABS
CDO
Investors
CDO
Investors
CDO
Investors
(iii) In a third and nal step, the structuring process involves creating seniority rules.
Investors in CDOs senior tranches include banks and pension funds, which might benet from
the expertise of the asset managers, and the risk-return proles dicult to nd in the market.
Investors in junior tranches are hedge funds searching for highly risky investment opportunities
that at the same time, are quite rewarding and certainly unavailable in the market. Additional
investors in junior tranches were dedicated o-balance-sheets entities such as SIV, conduits,
and SIV-lites, which will be reviewed in Section 12.4.7.
Underwriters of CDOs are investment banks, typically. They work closely with the asset
manager and create the right debt/equity ratio and perform collateral quality tests. They
liase with law rms and create the special purpose vehicle (possibly in some tax heaven system)
384
c
by
A. Mele
that will purchase the assets and issue the tranches, price the various tranches, and obviously
nd the investors. Fees to underwriters are very generous, due to the complexity of the CDOs.
According to Thomson Financial, top underwriters in 2006 were: Bear Sterns, Merrill Lynch,
Wachovia, Citigroup, Deutsche Bank, and Bank of America Securities.
Involved in the structuring process are also (i) trustee and collateral administrator, who distribute noteholder reports, check compliance and execute priority of payments; (ii) accountants,
who perform due diligence on the CDOs collateral pool, verifying for example credit ratings for
each asset; and (iii) rating agencies, which we shall discuss in the next subsection.
The economics behind structured nance is quite interesting. An originator may have private
information about the quality of certain assets and/or a comparative advantage in evaluating
these assets relative to other market participants. If the originator wishes to sell some of its
assets, an adverse selection problem will arise: because investors do not know the true quality of
the assets, they will demand a premium to purchase them or even worse, a market might fail to
arise. Structured nance helps originators mitigate this problem. First, by pooling the assets,
diversication benets can be achieved. Second, tranching allows relatively poorly informed
investors to access senior tranches, and be relatively protected from default. In the process, the
originator or arranger may retain subordinated exposure to alleviate investors concerns about
incentive compatibility. The following scheme summarizes the structuring process.
Source: Committee on the Global Financial System: The role of ratings in structured
nance: issues and implications, January 2005.
12.4.6.2 The role of rating agencies
Structured nance has always been a rated market. Issuers of structured instruments had
a natural appetite for a rating to occur at a scale comparable to that available for bonds.
385
c
by
A. Mele
The main reason was this would facilitate the sale of these products to investors bound by
ratings-based constraints dened by their investment mandates.
However, the involvement of rating agencies into the delivery of their opinion about credit
risk diers from that related to traditional bonds. As regards traditional instruments, rating
agencies simply aim to assess the risk of default as given, which they take as given. As regards
structured nance transactions, rating agencies play a much more ex-ante, reverse engineering
role. A tranche rating reects a view about both the credit risk of the asset pool and the extent
of credit support to be provided. These two elements are organized to reverse engineer the
tranche rating targeted by the deals arrangers. Deal origination thus involves rating agencies
in the structuring process.
12.4.6.3 Types of CDOs
In practice, CDOs are considerably more complex than the stylized examples outlined earlier.
We have a number of cases. We say that a CDO is static, if it holds the same set of assets.
Insetad, a CDO is managed, if the asset manager is allowed to change the composition of assets.
If the claims to the CDO arise from the cash ows originated by the assets, we have a cashflow CDO. If the claims to the CDO arise from the cash ows originated by the assets and/or
active asset management, we have a market-value CDO. CDOs can also be created to carve out
balance sheets, in which case we have balance-sheet CDOs. Moreover, and interestingly, CDOs
can be created (i) to achieve investment grade bonds through a pool of noninvestment grade
bonds, and (ii) to create riskier securities than those in the asset pool. In these cases, we have
arbitrage CDOs. Naturally, arbitrage CDOs do not give rise to any arbitrage opportunity.
These instruments merely reshue risk and returns of the assets in the pool, as we shall see
in the next section. Typically, then, arbitrage CDOs dier from balance sheet CDOs, because of
course, issuers of arbitrage CDOs do not necessarily hold the underlying collateral in advance,
which is obviously the case for issuers of balance-sheet CDOs. Therefore, the assets to be put
into the an arbitrage CDO pool have to be reasonably liquid.
Furthermore, we have synthetic CDOs, which are exposed to a pool of assets that are not
strictly owned or in the asset pool, typically through CDS underwriting. Like a cash-ow CDO,
the vehicle receives payments (the premium), which is then transferred to the tranche holders.
Naturally, there can be default events, which are also passed through to the investors, according
to the prespecied seniority rules. A synthetic CDO is funded, if the relevant tranche holders are
to pay for in the case of a credit event related to the assets the CDO is exposed to. Typically,
some funding is made available at the very time of investment. At maturity, the investor receives
a payo equal to the funding minus the realized losses. Junior tranches are typically funded,
and senior are typically not. However, senior tranches investors might have to make payments
in the unlikely event losses had ever to erode their tranches.
Finally, we have hybrid CDOs, which are partly cash-ow CDOs and partly synthetic CDOs.
In a single-tranche CDO, the entire CDO is structured to accommodate the specic needs of a
small group of investors, with some remaining tranche held by the dealer. And we have CDO 2 ,
where a large portion of the assets in the pool are tranches from other CDOs; or more generally,
CDO n .
12.4.6.4 Pricing
CDOs repackage cash ows from a set of assets. We provide simple examples to show how to
price this repackaging process. We begin with a simple example, taken from McDonald (2006, p.
583), which we further elaborate. Suppose we have three one-year bonds with face value = 100.
386
c
by
A. Mele
For each of these bonds, the risk-neutral probabilities of default equal 10% and the recovery
rates are 40. The safe interest rate for one year is 6%. So each bond price equals,
b = e0.06 (
0.10
Def. Prob
40 +
0.90

Surv. Prob
100) = 88.526.
b
= 12.19%.
The yield is, naturally, ln 100
A CDO can restructure the payments promised by the three bonds in a way that transforms
the riskiness and attractiveness of the initial assets. Consider the following example:
Senior tranche = 140

Face Value
= 300
Mezzanine tranche = 90
Junior tranche = 70
Asset Pool
CDO claims
In this example, each tranche receives the minimum between (i) the nominal value claimed by
the tranche and (ii) what is left available to the tranche after having satised the other tranches
by order of seniority.
Let Ni be the nominal values claimed by the tranches, so that N1 = 140, N2 = 90 and
N3 = 70. Let
be the realized payo of the asset pool, dened as,
= No. of Defaults 40 + (3 No. of Defaults) 100.

No. of surviving bonds
Naturally,
is random because the number of defaults is random. At the expiration,
(i) the senior tranches receives the minimum between N1 and
. For example, if only one
bond defaults,
= 240, and the senior tranche receives 140. If, however, three bonds
default, then,
= 120, which is less than the senior tranch nominal value, and the senior
tranche then receives 120. So a quite severe loss is needed to erode the senior tranche
claims.
(ii) The mezzanine tranche receives the minimum between N2 and the left-over from the
senior tranche.
(iii) Finally, at the expiration, the junior tranche reveives the minimum between N3 and the
left-over from the senior and mezzanine tranches.
More generally, tranche no. i receives,
i = min {Left-over from previous tranches up to tranche i 1 , Ni } ,
where
+
,
i1

Left-over from previous tranches up to tranche i 1 = max
k , 0 .
k=1
387
c
by
A. Mele

Synthetically,
i = min max
i1
k , 0 , Ni .
k=1
All we need, now, is to model the risk-neutral probability of default for each rm. Initially, we
assume the default events are independent across rms. Assume binomial distribution,

3 k
Pr (No. of Defaults = k) =
p (1 p)nk , p = 10%, k {1, 2, 3} .
k
We can then derive the following payo structure
Payoffs to CDO tranches, and prices: with independent defaults
Defaults Pr(Defaults) : pool payoff (1)
1: Senior
2: Mezzanine
3: Junior
0
0.729
300
140
90
70
1
2
3
0.243
0.027
0.001
240
180
120
140
140
120
131.8281994
0.060142867
90
40
0
83.40266709
0.076129382
10
0
0
50.34673197
0.329561531
Price
Yield
(1)
: pool payoff = Def*40+(3-Def)*100

N1 = 140
N2 = 90
N3 = 70
The price of each tranche is computed as the tranche payo, averaged across states, discounted
at the safe interest rate. For example, the price of the mezzanine tranche is,
Price Mezzanine = e0.06 (0.729 90 + 0.243 90 + 0.027 40 + 0.001 0) = 83.403.
Its yield is, Yield Mezzanine = ln 83.403
= 7.61%. Naturally, the sum of the three bond prices,
90
88.5263 = 265.58, is equal to the total value of the three tranches, 131.828+83.403+50.347 =
265.58. As anticipated, a CDO is a mere re-packaging device. It doesnt add or destroy value.
It merely redistributes risks (and returns).
The assumption defaults among names are uncorrelated is unrealistic, as argued in Section
12.5.4. We now remove this assumption. First, what happens in the special case where default
events are perfectly correlated ? In this case, either the three rms all default (with probability
0.10) or none defaults (with probability 0.90), and we have the situation summarized below,
Payoffs to CDO tranches, and prices: with perfectly correlated defaults
Defaults Pr(Defaults) : pool payoff (1)
1: Senior
2: Mezzanine
3: Junior
0
0.9
300
140
90
70
1
2
3
0
0
0.1
NA
NA
120
NA
NA
120
129.9635056
0.074388737
NA
NA
0
76.28292722
0.165360516
NA
NA
0
59.33116562
0.165360516
Price
Yield
(1)
: pool payoff = Def*40+(3-Def)*100

N1 = 140
N2 = 90
N3 = 70
388
c
by
A. Mele
Note, now, that mezzanine and junior tranches yield the same as they each pay o either their
nominal value or zero in exactly the same states of nature.
The previous cases (with independent and perfectly correlated defaults) are extreme. It is
by far more relevant to see what happens when defaults are only imperfectly correlated. When
defaults are imperfectly correlated, there are no simple tables to use to come up with tranche
pricing. Instead, one might make use of simulations, described succinctly in the Appendix.
Figure 12.14 below, obtained through Monte Carlo simulations, illustrates how the yield on
each tranche changes as a result of a change in the default correlation underlying the assets in
the CDO.
Yields on CDO tranches

0.4
Junior
Mezzanine
Senior
0.35
0.3
Yield
0.25
0.2
0.15
0.1
0.05
0.2
0.4
0.6
default correlation
0.8
FIGURE 12.15. Yields on the three CDO tranches, as functions of the default correlation among the
assets in the structure, with probability of default for each name equal to 20%. The thick, horizontal,
line is the yield on each securitized asset.
Arbitrage CDOs
Figure 12.15 illustrates how arbitrage CDOs work. The CDO has three assets yielding the same,
12.19% (the horizontal line in the picture). However, by restructuring the asset base through a
CDO, we can create claims (Senior and Mezzanine tranches) that yield less than 12.19%, as they
are considerably less risky than the asset base. Such an excess return, (12.19% Yieldtranche ),
with Yieldtranche {Senior, Mezzanine}, is made available to the Junior tranche/equity holders, once we account for management fees and expenses. Note, the previous redistribution of
risk always works when the default correlation is relatively low. As the default correlation in
the asset base increases, the situation may change dramatically, as we now illustrate. Figure
12.16 below makes some comparative statics: with p = 20%, instead of p = 10%. The yields are
obviously larger for each tranche, and the three assets now yield 18.78%, reecting the highr p.
389
c
by
A. Mele

Correlation assumptions
In Figures 12.15 and 12.16, the yield on the junior tranche decreases with default correlation.
This happens because we are assuming that the probability of default is xed at p = 10% for
each default correlation (say). As increases, the probability of clustering events increases,
which makes the Senior and Mezzanine tranches relatively less valuable and, correspondingly,
the Junior tranches more valuable. A more appropriate model is one in which p increases as
increases, to capture the fact that in bad times, both default correlation and probability of
defaults increase as these two things are intimately related (by, e.g., some common business
cycle factor).

0.7
Junior
Mezzanine
Senior
0.6
0.5
Yield
0.4
0.3
0.2
0.1
0.2
0.4
0.6
default correlation
0.8
assets in the structure, with probability of default for each name equal to 20%. The thick, horizontal,
line is the yield on each securitized asset.
Addressing the correlation assumption
Relax the assumption that the probability of default, p, and the default correlation, are
independent. For simplicity, assume that = 3.8 ln (p + 1), and let p vary from 0.10 to 0.30,
such that then, varies from 0 to 1. The situation, then, changes dramatically. Figure 12.17
depicts the results, which show how modeling might substantially aect eective pricing. First,
and naturally, the yield on each securitized asset is increasing in because is itself increasing
in the probability of default. Second, the Junior tranche has a yield that increases over a wide
390
c
by
A. Mele

spectrum of values for the default correlation, .

0.5
0.45
Junior
Mezzanine
Senior
0.4
Yield
0.35
0.3
0.25
0.2
Yield on each securitized asset
0.15
0.1
0.05
0.4
0.5
0.6
0.7
default correlation
0.8
0.9
assets in the structure, with probability of default and default correlation related by = 3.8ln (p + 1).
The thick curve line depicts the yield on each securitized asset.
12.4.6.5 Nth to default
In this contract, the owner of the 1st to default bears the risk of the rst default that occurs in
the asset pool:
Payo = Pr(No. of Defaults 1) 40 + Pr(No. of Defaults < 1) 100.
Likewise, the owner of the 2nd to default bears the risk of the second default that occurs in the
asset pool:
Payo = Pr(No. of Defaults 2) 40 + Pr(No. of Defaults < 2) 100.
Finally, the owner of the 3rd to default bears the risk of the third default that occurs in the
asset pool:
Payo = Pr(No. of Defaults = 3) 40 + Pr(No. of Defaults < 3) 100.
Let us assume that default correlation is zero for simplicity. We have previously computed
the previous probabilities as:
Pr(No. of Defaults 1) = 0.243 + 0.027 + 0.001 = 0.271
Pr(No. of Defaults 2) = 0.027 + 0.001 = 0.028
Pr(No. of Defaults = 3) = 0.001
391
c
by
A. Mele

Thus, we have the following prices,
Price1st -to-default = e0.06 [0.271 40 + (1 0.271) 100] = 78.863

Price2nd -to-default = e0.06 [0.028 40 + (1 0.028) 100] = 92.594
Price3rd -to-default = e0.06 [0.001 40 + (1 0.001) 100] = 94.120
From here, we can compute the yields as follows, Yield1st -to-def = ln (78.863/100) = 23.74%,
Yield2nd -to-def = ln (92.594/100) = 7.69%, and Yield3rd -to-def = ln (94.120/100) = 6.06%.
12.4.7
One stylized numerical example of a structured product
A. Defaultable bonds
Suppose we observe the following risk-structure of spreads, related to two bonds maturing in
two years:
SpreadA (2 years) = 1.5%, SpreadB (2 years) = 2.5%,
where A and B denote the rating classes the bond issuers belong to. Assume that the one-year
transition rating matrix, dened under the risk-neutral probability, is:
A
A
0.7
From
B
0.3
Def 0
To
B
0.3
0.5
0
Def
0
0.2
1
where Def denotes default. We assume that in the event of default, the recovery value of the
bond is paid o at the end of the second period. We want to determine the expected recovery
rates for the two bonds, and which expected recovery rate is the largest. We have:

Reci
rT D0,i
e
=
Qi (2) + (1 Qi (2)) , i {A, B} .
N
N
Therefore,

1
RecA
SpreadA (2 years) = 1.5% = ln
QA (2) + (1 QA (2))
2
N

1
RecB
SpreadB (2 years) = 2.5% = ln
QB (2) + (1 QB (2))
2
N
We have to nd QA (2) and QB (2). The transition matrix for two years is,
0.7 0.3 0
0.7 0.3 0
Q (2) = 0.3 0.5 0.2 0.3 0.5 0.2 ,
0
0
1
0
0
1
such that,
Pr {A defaults in 2 years} = QA (2)

= 0.70
0.20 +
0 + 0.30

AADef
= 0.06
392
ABDef
0
1
ADef Def
(12.36)
(12.37)
c
by
A. Mele
Pr {B defaults in 2 years} = QB (2)

=
0.20
1
BDef Def
+ 0.50
0.20 + 0.30

0
BBDef
BADef
= 0.20 + 0.10 = 0.30.
Hence, using Eqs. (12.36)-(12.37), we have

1
RecA
SpreadA (2 years) = 1.5% = ln
0.06 + (1 0.06)
2
N

RecB
1
SpreadB (2 years) = 2.5% = ln
0.30 + (1 0.30)
2
N
Solving, yields,
RecB
RecA
= 50.7%,
= 83.7%.
N
N
The expected recovery rate for the second bond is the largest. This is because the probability
rm B defaults is much larger than the probability rm A defaults and yet the two spreads are
relatively close to each other. So to rationalize the two spreads, we need a large recovery rate
for the second bond.
What would happen to the two credit spreads, then, once we assume that the recovery rates
are the same, and equal to 50%? This question sheds additional light to the previous ndings.
If the recovery rates are the same and both equal 50%,
1
SpreadA (2 years) = ln [0.50QA (2) + (1 QA (2))]
2
1
SpreadB (2 years) = ln [0.50QB (2) + (1 QB (2))]
2
Then, using the previously computed transition probabilities for two years, we obtain:
SpreadA (2 years) = 1.52%,
SpreadB (2 years) = 8.12%.
When the recovery rates are the same, the spread on the second bond diverges substantially
from that on the rst bond.
B. Collateralized debt obligations
Let us keep on using the same framework as before, but use dierent gures, so as to gure out
the implications for CDOs pricing. Consider the following one year transition matrix, under the
risk-neutral probability:
A
A
0.7
From
B
0.1
Def 0
To
B
0.3
0.6
0
Def
0
0.3
1
where Def denotes default. Consider (i) 1 one-year bond issued by a company rated A, and
(ii) 3 one-year bonds issued by a company rated B. Both bonds have face value equal to 100.
393
c
by
A. Mele
We assume that the recovery values in case of default of all these bonds are the same, and equal
to 50. Finally, we assume the safe interest rate is taken to be equal to zero.
Consider a collateralized debt obligation (CDO, in the sequel), which gathers the previous
four bonds. Therefore, the CDO has nominal value of 400, and pays o in one year. The CDO
has (i) a senior tranche, with nominal value equal to 150; (ii) a mezzanine tranche, with nominal
value equal to N1 ; and (iii) a junior tranche, with nominal value equal to N2 . We assume that
the structure is such that N1 > 100.
First, we determine the price and yields on all the four bonds. Since the safe interest rate is
zero, and the company rated A is safe, up to the next year, the price of the A bond is 100, and
its yield is zero. As for the three bonds rated B, we have:
P B = 50 0.3 + 100 0.7 = 85.0, Y ield B = ln 0.85 = 16.25%.
Second, we determine the yield on the junior tranche, and derive the yield on the mezzanine,
as a function of its nominal value N1 . To determine the yield on the tranches, we need to gure
out the following table:
0
1
2
No Def Pr
0
0.7 400 150 N1 N2
1
0
NA NA NA NA
2
0
NA NA NA NA
3
0.3 250 150 100 0
4
4
NA NA NA NA
where No Def denotes the number of defaults, Pr is the probability of No Def, is the pool
payo, dened as,
= No Def 50 + (4 No Def) 100,
and, nally: 0 is the payo to the senior tranche, 1 is the payo to the mezzanine tranche,
and, 2 is the payo to the junior tranche. Therefore, we have:
price mezzanine = 0.70 N1 + 0.30 100, price junior = 0.70 N2 ,
such that:
Yield mezzanine
Yield junior

0.70 N1 + 0.30 100
100
= ln
= ln 0.70 + 0.30
N1
N1

0.70 N2
= ln
= 35.67%.
N2
Naturally, we need to have that Yield mezzanine < Yield junior. It is simple to show this
relation: it suces to note that,

100
Yield junior = ln (0.70) > ln 0.70 + 0.30
= Yield mezzanine.
N1
A reverse enginnering question is, now, to determine which nominal value of the mezzanine
tranche N1 is needed, to ensure that the yield on the mezzanine tranche is equal to or greater
than the yields on the bonds issued by the company with credit rating B? The answer is
N1 = 200, for in this case, the mezzanine tranche would have the same payo structure as the
bond rated B: it would deliver (i) the face value, in the event the company rated B does not
default; and (ii) half of its nominal value, 100, in the event the company rated B does default.
394
c
by
A. Mele
Finally, we ask which nominal value of the mezzanine tranche N1 is needed, to ensure that
the yield on the mezzanine is equal to 18%? And what is the corresponding nominal value of
the junior tranche, N2 ? To address these issues, we rst want that:

0.70 N1 + 0.30 100

Yield mezzanine = ln
N1
= 18%.
Solving for N1 yields, N1 = 221.78. Therefore, N2 = 400 Nominal value senior N1 =

400 150 221.78 = 28.22.
12.4.7.1 The 2007 subprime crisis

Issuance data
European and U.S. Structured Credit Insurance. Source: IMF, Global Financial Stability
Report, April 2008.
395
c
by
A. Mele
Outstanding U.S. Subprime issuance. Source: IMF, Global Financial Stability Report,
April 2008.
Off-balance-sheet entities: SIV, conduits, and SIV-lites
[b. circa 1985]

On the funding side, a typical SIV (Structured Investment Vehicle) issues long-maturity
notes. On the asset side, a SIV typically relies on assets that are more complex than those
conduits rely on. SIVs tended to be more leveraged than conduits. Please remember: SPV =
Special Purpose Vehicle, i.e. a vehicle that organizes securitization of assets; SIV = Structured
Investment Vehicle, i.e. a fund that manages asset backed securities. In a sense, SIV were virtual
baks, as they used to borrow through low-interest securities and invest the money in longer term
securities yielding large rewards (and risk), as we discuss below. SIVs and conduits typically
had an open-ended lifespan.
SIV-lites are less conservatively managed, and structured with greater leverage. Their portfolios are not much diversied, and are much smaller in size than SIVs. SIV-lites had a nite
lifespan, with a one-o issuance vehicle. They were greatly exposed to the U.S. subprime market,
more so than SIVs.
O-balance-sheet entities borrow in the shorter term, typically through commercial paper or
auction rate securities with average maturity of 90 days, as well as medium term notes with
average maturity of a year. They purchase long-maturity debt, such as nancial corporate bonds
or asset-backed securities, which is high-yielding. Naturally, the prots made by these entities
are paid to the capital note holders, and the investment managers. The capital note holders
are, of course, the rst-loss investors.
The obvious risk incurred by these entities relates to solvency, which happens when longterm asset values fall below the value of short-term liabilities. This risk has great chance to
materialize when the pricing of the assets is informal, as argued below. A second risk relates
to funding liquidity, which is the risk related to duration mismatch: renancing occurs shortterm, but if the short-term market conditions are bad, the entities need to sell the assets into a
396
c
by
A. Mele
depressed market. To cope with this risk, the sponsoring banks would grant credit lines. Typical
sponsors were: Citibank ($100bn), JP Morgan Chase ($77bn), Bank of America ($60bn). In the
European Union: HBOS ($42bn), ABN Amro ($40bn), HSBC ($32bn).

The 2007 meltdown
The rst obvious issue to think about relates to pricing and the role played by credit ratings.
Being illiquid, the pricing of structured credit products used to rely on that of similarly rated
comparable products for which quotations were available. For example, the price of AAA ABX
subindices would be used to estimate the values of AAA-rated tranches of MBS. Or, the price of
BBB subindices would be used to value BBB-rated MBS tranches. This is the mapping role
credit ratings played for the pricing of customized or illiquid structured credit products. However, it is well-known that the risk prole of structured products diers from that of corporate
bonds. Even if a tranche has the same expected loss as an otherwise similar corporate bond,
unexpected loss or tail risk can be much larger than that for corporate bonds. Therefore, it is
397
c
by
A. Mele
misleading to extrapolate structured products ratings from corporate bonds ratings. Typically,
ratings used to capture only the rst moments of the distribution. Moreover, credit rating inertia for bonds does not necessarily work for structured products, as illustrtated in the picture
below.
Two additional fundamental aspects contributing to the meltdown. First, there was an erosion in lending standards: statistical models were based on historically low mortgage default
and delinquency rates that arose in a credit environment with tight credit standards. Second,
there were correlation issues: past data suggested a quite weak correlation between regional
mortgages, which made investors perceive a sense of diversication. However, the housing
downturn turned up to be a nation-wide phenomenon.
The mechanics of the crisis started with fears of contagion from the rising level of defaults
in subprime underlying instruments, many of which were incorporated in complex products.
The fears of contagion related to safer tranches as well. They came from the investors understanding the pricing models were misspecied, and their lack of trust vis-à-vis the rating
agencies. Banks, on the other side, were aected for a number of reasons: (i) they had invested
in subprime securities directly; (ii) they had provided credit lines to SIV (indebted through
commercial paper) and conduits that held these securities, thereby creating a shadow banking system, which escaped accounting and supervision rules; and (iii) this very same shadow
system generated banks loss of condence in the ability of their counterparties to meet their
contractual obligations. So the Asset Backed Commercial Paper market dried up, triggering
credit lines. The result was a sell-o of anything related to structured nance, from junk to
AAA, which led to a complete liquidity black hole, and a severe reappraisal of structured
nance.
In turn, the reappraisal of structured nance determined severe writedowns, arising in part
through the liquidity black hole, i.e. by the market participants expectations. Repricing was
dicult indeed. In the absence of a liquid market, writedowns have to rely on marking to model.
But investors did not trust the models and the rating process leading to them! Meanwhile,
credit agencies proceeded to severe downgrades, conrming the investors beliefs that ratings
were not entirely appropriate, a quite self-reinforcing mechanism. These events escalated to
a complete dry up in September-October 2008, partly restored by painful bank bail-outs and
398
12.5. A few hints on the risk-management practice
c
by
A. Mele
recapitalizations.
12.5 A few hints on the risk-management practice

12.5.1
Value at Risk (VaR)
We need to review Value at Risk (VaR), in general. VaR is a method of assessing risk that uses
statistical techniques. Useful for supervision and management of nancial risks. Origins: reaction
to nancial disasters in the early 1990s involving Orange County, Barings, Metallgesellschaft,
Daiwa, etc.
Definition I: VaR measures the worst expected loss over a given horizon under normal market
conditions at a given confidence level.
Definition II: We are (1 p)% certain that a given portfolio will not suffer of a loss larger
than $W over the next N weeks, Pr (Loss < W ) = p. That is, $VaRp = $W .
399
c
by
A. Mele
Equivalently, note that

Loss
V
=
= portfolio return
V0
V0
where V denotes the change in value of the portfolio over the next N days, and $V0 is the
current value of the portfolio. Hence,
p = Pr (Loss < VaRp ) = Pr
V
VaRp
<
V0
V0
This formulation leads us to the following alternative denition:

Definition III: We are (1 p)% certain that a given portfolio will not experience a relative
p
loss larger than VaR
over the next N weeks.
V0
So in practice, we shall have to nd the relative loss, p , for a given condence p, as follows:
p = Pr

V
< p ,
V0
where p =
VaRp
.
V0
The corresponding VaRp is just

VaRp = p V0
For example, suppose that the portfolio return over the next 2 weeks, V
, is normally disV0
V
tributed with mean zero and unit variance. We know that 0.01 = Pr( V0 < 2.32), hence,
VaRp = 2.32 V0 .
400
c
by
A. Mele
0.4
0.35
0.3
0.25
0.2
0.15
0.1
1%
VaR/V
0.05
0
3
We are 99% certain that our portfolio will not suer of a loss larger than 2.32 times its
current value over the next 2 weeks. We are 99% certain that our portfolio will not experience
a relative loss larger than 2.32 over the next 2 weeks.
As a second example note that the previous assumption about the portfolio return was
, is normally distributed
extreme. Assume, instead, the porfolio return over the next 2 weeks, V
V0
2 2
2
2
with mean zero and variance = 52 year , where year is the annualized variance. We assume
that 2year = 0.152 . We have to re-scale the previous formulas, as follows. First, we introduce a
variable N (0, 1), i.e. is normally distributed with mean zero and variance = 1. So we can
write,

V d
= N 0, 2 ,
V0
and, hence,
0.01 = Pr ( < 2.32) = Pr (V < 2.32 V0 ) ,
whence, VaRp = 2.32 V0 . We know the annualized variance, 2year = 0.152 , from which we
p
2 2
can derive the two-week standard deviation, 2 = 52
year 0.032 , and, hence, VaR
= 2.32 =
V0
2.32 0.03 7%. Thata is, we are 99% certain that our portfolio will not suer of a loss larger
than 7% times its current value over the next 2 weeks. We are 99% certain that our portfolio
will not experience a relative loss larger than 7% over the next 2 weeks.
More generally, we may assume the porfolio return over the next 2 weeks, V
, is normally
V0
2
distributed with mean and variance . In this case,
and, hence,

V d
= + N , 2 ,
V0
0.01 = Pr ( < 2.32) = Pr (V < V0 (2.32 ))
whence, VaRp = V0 (2.32 ). In practice, is very small if the horizon is as short as two
weeks.
401
c
by
A. Mele
12.5.1.1 Challenges to VaR
Challenges related to distributional assumptions, nonlinearities, or conceptual difficulties.

Distributional assumptions
The assumption that data are generated by a normal distribution does not describe asset
returns well. Chapters 10 and 11 explain that we need ARCH eects, stochastic volatility and
multifactor models. More generally, data can exhibit changes in regimes, nonlinearities and fat
tails. Fat tails are particularly important to understand, since this is what were interested
in after all. More in general, it is quite challenging to understand what the data generating
process is, especially in so far as we consider portfolios of assets. Asset returns and volatilities
are typically correlated, with correlation rising in bad timescorrelation is stochastic.
We may make distributional assumptions but, then, these assumptions have to be carefully
assessed through, for example, backtesting (to be explained below). We may proceed with
nonparametric methods, and this is indeed a promising avenue, but with its caveats.
How do nonparametric methods work? These methods rely on an old and idea, which is to
estimate the data distribution through histograms. These histograms, then, can be readily used
to compute VaR. This approach is nonparametric in nature, as it does not rely on any model.
A more rened method replaces rough histograms with smoothed histograms, as follows.
Suppose to have access to a time series of data xn , which are drawn from a certain probability
law, with density f (x). We may dene the following estimate of the density f (x),

N
1 1
x xn
fN (x) =
K
,
N n=1
where N is the sample size, and K is some symmetric function integrating to one. We may
think of fN (x) as a smoothed histogram, with window bin equal to . It is possible to show
that as N goes to innity and goes to zero at a certain rate, fN (x) converges in probability
to f (x), for all x. But we are not done, since there are not obvious rules to choose and K?
The choice of is notoriously dicult. Unfortunately, the bias, fN (x) f (x), tends to be
large exactly on the tails of f (x), which do represent the region were interested in. In general,
we can use Montecarlo simulations out of a smoothed density like this to compute VaR.
Nonlinearities
Finally, portfolios of assets can behave in a nonlinear fashion, especially when the portfolio
contains derivatives. In general, the value of a portfolio including M assets is,
P =
M
i Si ,
j=1
where i is the number of the i-th asset in the portfolio, and Si is the price of the i-th asset
in the portfolio. Holding i constant, the variation on the portfolio return is simply a weigthed
average of all the asset returns,

M
M

i Si Si
P
PT PT Pt =
i Si
=
,
P
P
S
i
j=1
j=1
where the variations relate to any time interval. Often, the prices Si are rational functions of
the state variables, or are interlinked through arbitrage restrictions. Use factors to determine
402
c
by
A. Mele
the risk associated with xed income securities. When the horizon of the VaR is large, it is
unlikely that i is constant. Typically, we shall need to go for numerical methods, based, for
example, on Monte Carlo simulations. So all in all, we need to have a careful understanding of
the derivatives in the book, and proceed with back testing and stress testing.
VaR as an appropriate measure of risk
There are technical diculties related to the very denition of VaR. VaR suers from some
statistic-theoretic foundation. VaR tells us that 1% of the time, losses will exceed the VaR gure,
but it does not tell us the entity of the loss. So we need to compute the expected shortfall. Any
risk measure should enjoy a number of sensible properties. Artzner et al. (1999) have noted a
number of properties, and showed that VaR does not enjoy the so-called subadditivity property,
according to which the sum of the risk measures for any two portfolios should be larger than the
risk measure for the sum of the two portfolios. VaR doesnt satisfy the subadditivity property,
but expected shortfall does satisfy the subadditivity property.
12.5.2
Backtesting
How well the VaR estimate would have performed in the past? How often the loss in a given
sample exceeded the reference-period 99% VaR? If the exceptions occur more than 1% of the
time, there is evidence that the models leading to VaR estimates are misspecieda nice
word for saying bad models.
The mechanics of backtesting is as follows. Suppose the models leading to the VaR are
good. By construction, the probability the VaR number is exceeded in any reference period
is p, where p is the coverage rate for the VaR. Next, we go to our sample, which we assume
it comprises N days, and let M be the number of days the VaR is exceeded. We wish to test
whether the number of exceptions we observe in the sample conforms to the expected number
of exceptions based on the VaR. For example, it might be that the number of exceptions we
have observed, M, is larger than the expected number of exceptions, p N . We want to make
sure this circumstance arose due to sample variability, rather than model misspecication. A
simple one-tail test is described below.
Let us compute the probability that in N days, the VaR is exceeded for M or more days.
Assuming exceptions are binomially distributed, this probability is,
p =
N
k=M
N!
pk (1 p)Nk .
k! (N k)!
Then, we can say the following. If p 5% (say), we reject the hypothesis that the probability
of exceptions is p at the 5% levelthe models were using are misspecied. If p > 5% (say),
we cannot reject the hypothesis that the probability of exceptions is p at the 5% levelwe cant
say the models were using are misspecied. This test is reviewed in more detail by Hull (2007,
p. 208). Other tests are reviewed by Christoersen (2003, p. 184).
12.5.3
Stress testing
Stress testing is a technique through which we generate articial data from a range of possibles
scenarios. Stress scenarios help cover a range of factors that can create extraordinary losses
or gains in trading portfolios, or make the control of risk in those portfolios very dicult.
These factors include low-probability events in all major types of risks, including the various
403
c
by
A. Mele
components of credit, market, and operational risks. Stress scenarios need to shed light on the
impact of such events on positions that display both linear and nonlinear price characteristics
(i.e. options and instruments that have options-like characteristics).
Possible scenarios include simulating (i) shocks that although rare or even absent from the
historical database at hand, are likely to happen anyway; and (ii) shocks leading to structural
breaks and/or smooth transition in the data generating mechanism. One possible example is to
set the percentage changes in all market variables in the portfolio equal to the worst percentage
changes having occured in ten days in a row during the subprime crisis 2007-2008.
This example on the subprime crisis is related to the historical simulation approach to generate scenarios. This approach consists can be explained through a single formula. Let vt the value
of some market variable i in day t in our sample, where t = 0, , T (say). We can generate T
scenarios for the next day, T + 1, as follows.
(i) The rst scenario is that in which each variable grows by the same amount it grew at
time 1,
v1
vT +1 = vT .
v0
(ii) The second scenario is that in which each variable grows by the same amount it grew at
time 2,
v2
vT +1 = vT .
v1
(iii)
(iv) The T -th scenario is that in which each variable grows by the same amount it grew at
time T ,
vT
.
vT +1 = vT
vT 1
(v) The T scenarios are generated for all the market variables, which would give us an articial
multivariate sample of T observations. We can use this sample for many things, including
VaR.
12.5.4
Credit risk and VaR
We can use the tools in Section 12.2 to assess the likelihood of default for a given name. The
important thing to do is to use the physical probability of default, not the risk neutral one. The
risk neutral probability of default is likely to be larger than the physical one. Therefore, using
the risk neutral probability leads to too conservative estimates.
VaR for credit risks pose delicate issues as well. The key issue is the presence of default
correlation. In practice, defaults among names or loans are likely to be correlated, for many
reasons. First, there might be direct relationships or, more generally, network eects, among
names. Second, rms performance could be driven by common economic conditions, as in the
one factor model which we now describe. This one factor model, developed by Vasicek (1987),
is at the heart of Basel II. In the appendix, we provide additional technical details about how
this model is related to a modeling tool known as copulae functions. We now proceed to develop
this model in an intuitive manner. Let us dene the following variable:

zi = F + 1 i ,
(12.38)
404
c
by
A. Mele
where F is a common factor among the names in the portfolio, i is an idiosynchratic term,
and F N (0, 1), i N (0, 1). As we explain in the Appendix, 0 is meant to capture the
default correlation among the names.
Next, assume that the physical probability each rm defaults, by T , say P (T ), is the same
for each rm within the same class of risk, and given by,
P (T ) = ( PD ) PD,
where PD is the probability of default, and is the cumulative distribution of a standard
normal variable. That is, by time T , each rm defaults any time that,
zi < PD 1 (PD) ,
where 1 denotes the inverse of . One economic interpretation of Eq. (12.38) is that zi is the
value of a rm and, then, the rm defaults whenever this value hits some exogenously given
barrier PD .
Conditionally upon the realization of the macroeconomic factor F , the probability of default
for each rm is,
1

(PD) F
p (F ) Pr ( Default| F ) =
.
(12.39)
1
By the law of large numbers, this is quite a good approximation to the default rate for a portfolio
of a large number of assets falling within the same class of risk.
We see that this conditional probability is decreasing in F : the larger the level of the common
macroeconomic factor, the smaller the probability each rm defaults. Hence, we can x a value
of F such that Pr ( Default| F ) = Default rate is what we want. Note, the probability F is larger
than 1 (x) is just x! Formally,

Pr F > 1 (x) = Pr F < 1 (x) = 1 (x) = x.
Then, with probability x, the default rate will not exceed
1

(PD) + 1 (x)
VaRCredit Risk (x) =

.
1
It is easy to see that VaRCredit Risk (x) increases with . Basel II sets x = 0.999 and, accordingly,
it imposes a capital requirement equal to,
Loss-given-default [VaRCredit
Risk
(0.999) PD] Maturity adjustment.
The reason Basel II requires the term VaRCredit Risk (0.999)PD, rather than just VaRCredit Risk ,
is that what is really needed here is the capital in excess of the 99.9% worst case loss over the
expected idiosyncratic loss, PD. Well functioning capital markets should already discount the
idiosyncratic losses.
Finally, Basel II requires banks to compute through a formula in which is inversely related
to PD. The formula is based on empirical research (see Lopez, 2004): for a rm which becomes
less creditworthy, the PD increases and its probability of default becomes less aected by market
conditions. Basel II requires banks to compute a maturity adjustment factor that takes into
account that the longer the maturity the more likely it is a given name might eventually migrate
towards a more risky asset class.
405
c
by
A. Mele
The previous model can be further elaborated. We ask: (i) What is the unconditional probability of defaults, and (ii) what is the density function of the fraction of defaulting loans?
First, note that conditionally upon the realization of the macroeconomic factor F , defaults
are obviosly independent, being then driven by the idiosyncratic terms i in Eq. (12.38). Given
N loans, and the realization of the macroeconomic factor F , these defaults are binomially
distributed as:

N
Pr (No of defaults = n| F ) =
p (F )n (1 p (F ))Nn ,
n
where p (F ) is as in Eq. (12.39). Therefore, the unconditional probability of n defaults is:
"
Pr (No of defaults = n) =
Pr ( No of defaults = n| F ) (F ) dF,
where denotes the standard normal density. This formula provides a valuable tool analysis in
risk-management. It can be shown that VaR levels increase with the correlation .
Next, let denote the fraction of defaulting loans. For a large portfolio of loans, = p (F ),
such that:
"
"
Pr ( x) =
Pr ( x| F ) (F ) dF =
Ip(F )x (F ) dF = (F ) ,
(12.40)
where
the indicator function, and F satises, by Eq. (12.39), F : x = p (F ) =
1I denotes

(PD)+ F
. Solving for F leaves:

1
1 1 (x) 1 (PD)
F =
.
It is the threshold value taken by the macroeconomic factor that guarantees a frequency of defaults less than x. Replacing F into Eq. (12.40) delivers the cumulative distribution function
for . The density function f (x) for the frequency of defaults is then:
A
1 12 1 (x) 21 (11 (x)1 (PD))
e
f (x) =
.
406
c
by
A. Mele
12.6. Appendix 1: Proof of selected results
12.6 Appendix 1: Proof of selected results

Alternative derivation of Eq. (12.9). Under the risk-neutral probability, the expected change
of any bond price must equal zero when the safe short-term rate is zero,
B (t)
+ (Rec B (t)) = rB = 0,
t
with B (T ) = N,
where the rst term, B(t)

t , reects the change in the bond price arising from the mere passage of time,
and (Rec B (t)) is the expected change in the bond price, arising from the event of default, i.e. the
probability of a sudden default arrival, , times the consequent jump in the bond price, Rec B (t).
The solution to the previous equation is,
" T
t
B (0) =
Rec
dt + NeT ,
e
0
=Pr{Default at t}
Proof of Eq. (12.10). The spread is given by:

$
%

RecT 1 eT + NeT
1
s (T ) = ln
.
T
N
With N = 1, and RecT = R eT , we have,

1
1
s (T ) = ln ReT 1 eT + eT = ln Re()T 1 eT + 1 ,
T
T
or equivalently,
s (T ) =

1 T
1
1 eT + eT = ln R 1 eT + e()T ,
ln Re
T
T
Therefore, if , then limT s (T ) = , and if , limT s (T ) = .
407
12.7. Appendix 2: Details on transition probability matrixes and pricing
c
by
A. Mele
12.7 Appendix 2: Details on transition probability matrixes and pricing

Consider the matrix P (T t) for T t t, P (t), and write,
+
i=j
1 + ij t,
P (t)ij
ij t,
i = j
(12A.1)
We are dening the constants ij as they were the counterparts of the intensity of the Poisson process
in Eq. (12.8). Accordingly, these constants are simply interpreted as the instantaneous probabilities
of migration from rating i to rating j over the time interval t. Naturally, for each i, we have that

N
j=1 P (t)ij = 1, and using into Eq. (12A.1), we obtain,
ii =
N
ij .
(12A.2)
j=1,j=i
The matrix containing the elements ij dened in Eqs. (12A.1) and (12A.2) is called the generating
matrix.
Next, let us rewrite Eq. (12A.1) in matrix form,
P (t) = I + t.
Suppose we have a time interval [0, T ], which we chop into n pieces, so to have t =

T n
P (T ) = P (t)n = I +
.
n
T
n.
We have,
For large n,
P (T ) = exp (T ) ,
(12A.3)
(T )n
the matrix exponential, dened as, exp (T ) n=0 n! .
To evaluate derivatives written on states, we proceed as follows. Suppose Fi is the price of derivative in state i {1, , N}. Suppose the Markov chain is the only source of uncertainty relevant for
the evaluation of this derivative. Then,
dFi =
Fi
dt + [FR Fi ],
t
{1, , N}, with the usual conditional probabilities. In words, the instantaneous change in
where R
i
the derivative value, dFi , is the sum of two components: one, F
t dt, related to the mere passage of
time, and the other, [FR Fi ], related to the discrete change arising from a change in the rating.
Suppose that r = 0. Then,
N
E (dFi )
Fi
Fi
rFi = 0 =
=
+
ij [Fj Fi ] =
+
ij [Fj Fi ] ,
dt
t
t
j=1
j=i
with the appropriate boundary conditions.

As an example, consider defaultable bonds. In this case, we may be looking for pricing functions
having the following form,
Fi (T t) = xQi (T t) + 1 Qi (T t) ,
and then solve for Qi (T t), for all i {1, , N}. Naturally, we have

0 = xQi Qi +
ij [x (Qj Qi ) (Qj Qi )]
=x
Qi
j=i
j=i
!

ij (Qj Qi ) Qi +
j=i
408
!
ij (Qj Qi ) ,
12.7. Appendix 2: Details on transition probability matrixes and pricing

which holds if and only if,

Qi =
ij (Qj Qi ) =
ij Qj +
ij Qi =
j=i
j=i
j=i
j=i
c
by
A. Mele
!
ij Qj + ii Qi .
That is, Q = Q, which solved through the appropriate boundary conditions, yields precisely Eq.
(12A.3).
409
c
12.8. Appendix 3: Derivation of bond spreads with stochastic default intensity by
A. Mele
12.8 Appendix 3: Derivation of bond spreads with stochastic default intensity

We derive Eq. (12.29), by relying on the pricing formulae of Chapter 11. If the short-term is constant,
the price of a defaultable bond derived in Section 11.3.7 of Chapter 11 can easily be extended to, with
the notation of the present chapter,
#N
" N

#t
0 (t)dt
0 (u)du
rN
rt
E e
+
e E (t) e
Rec (t) dt.
(12A.4)
P (y, N) = e
0

=Pr{Default(t,t+dt)}
The term indicated inside the integral of the second term, is indeed the density of default time at t,
because,
#t

Pdefault by time t () = 1 E e 0 (s)ds ,
such that by dierentiating with respect to t, yields, under the appropriate regularity conditions, that
Pr{Default (t, t + dt)} is just the term indicated in Eq. (12A.4). So Eq. (12.29) follows. Naturally,
Pr{Default (t, t + dt)} =
Replacing this into Eq. (12A.4),
#N

"
P (y, N) = erN E e 0 (t)dt + Rec
Psurv (, t) .
t

ert Psurv (, t) dt
t
0
" N

= 1 LGD 1 erN Psurv (, N) (1 LGD)
rert Psurv (, t) dt,
N
where the second equality follows by integration by parts and the assumption of constant recovery
rates. Setting r = 0, produces Eq. (12.30).
410
12.9. Appendix 4: Conditional probabilities of survival
c
by
A. Mele
12.9 Appendix 4: Conditional probabilities of survival

We prove Eqs. (12.32)-(12.34). First, for (ti1 , ti ) small, the numerator in Eq. (12.31) can be replaced
by

#t
0 (s)ds
Psurv (, t) E (t) e
,
t
and rescaled by dt. Regularity conditions under which we can perform this dierentiation can be found
in a related context developed in Mele (2003). Eqs. (12.32)-(12.33) follow.
As for Eq. (12.34), the proof follows the same lines of reasoning as that in Appendix 3 of Chapter
11. That is, we can dene a density process,
#

#T

e 0 (s)ds Psurv ( ( ) , , T )
(s)ds
#T

T ( ) =
, Psurv ( ( ) , , T ) E e
F .
0 (s)ds
E e
It is easy to show that the drift of Psurv is ( ) d , such that by Itos lemma,
dT ( )
= [Vol (Psurv ( ( ) , , T ))] dW ( ) ,
T ( )
where,

Psurv ( ( ) , , T )
( ) = B (T ) ( ),
Vol (Psurv ( ( ) , , T ))
Psurv ( ( ) , , T )
where the second line follows by the closed-form expression of Psurv in Eq. (12.26). Therefore, W ( )
is a Brownian motion under Q , where

dW ( ) = dW ( ) + B (T ) ( )d ,
and Eq. (12.34) follows.
411
12.10. Appendix 5: Modeling correlation with copulae functions
c
by
A. Mele
12.10 Appendix 5: Modeling correlation with copulae functions

A.
Statitical independence and correlation
Two random variables are always uncorrelated when they are independently distributed. Yet there
might be situations where two random variables are not correlated and still exhibit statistical dependence. As an example, suppose a random variable y relates to another random variable x through y =
kx3 , for some constant k, and x can take on 2N+1 values,
x {xN , xN1 ,
, x1 , 0, x1 , , xN1 , xN },
N
1
3+
3
. Then, we have that Cov (x, y) = N
(x
)
x
and Pr {xj } = 2N+1
j
j
j=1
j=1 (xj ) xj = 0 and yet, y
and x are obviously dependent. This example might be interpreted, economically, as one where y and
x are two returns on two asset classes. These two returns are not correlated, overall. Yet the comove in
the same direction in both very bad and in very good times. This appendix is a succinct introduction
to copuale, which are an important tool to cope with these issues.
Consider two random variables Y1 and Y2 . We may relate Y1 to another random variable Z1 and we
may relate Y2 to a second random variable Z2 , on a percentile-to-percentile basis, viz
Fi (yi ) = Gi (zi ) , i = 1, 2,
(12.41)
where Fi are the cumulative marginal distributions of Yi , and Gi are the cumulative marginal distributions of Zi . That is, for each yi , we look for the value of zi such that the percentiles arising through the
mapping in Eq. (12.41) are the same. Then, we may assume that Z1 and Z2 have a joint distribution
and model the correlation between Y1 and Y2 through the correlation between Z1 and Z2 . This indirect
way to model the correlation between Y1 and Y2 is particularly helpful. It might be used to model the
correlation of default times, as in the main text of this chapter.
B. Copulae functions
We begin with the simple case of two random variables, This simple case shall be generalized to the
multivariate one with a mere change in notation. Given two uniform random variables U1 and U2 ,
consider the function C (u1 , u2 ) = Pr (U1 u1 , U2 u2 ), which is the joint cumulative distribution of
the two uniforms. A copula function, then, is any such function C, with the property of being capable
to aggregate the marginals Fi into a summary of them, in the following natural way:
C (F1 (y1 ) , F2 (y2 )) = F (y1 , y2 ) ,
(12A.5)
where F (y1 , y2 ) is the joint distribution of (y1 , y2 ). Thus, a copula function is simply a cumulative
bivariate distribution function, as F (Y1 ) and F (Y2 ) are obviously uniformly distributed. To prove Eq.
(12A.5), note that
C (F1 (y1 ) , F2 (y2 )) = Pr (U1 F1 (y1 ) , U2 F2 (y2 ))

= Pr F11 (U1 ) y1 , F21 (U2 ) y2
= Pr (Y1 y1 , Y2 y2 )
= F (y1 , y2 ) .
(12A.6)
That is, a copula function evaluated at the marginals F1 (y1 ) and F2 (y2 ) returns the joint density
F (y1 , y2 ). In fact, Sklar (1959) proves that, conversely, any multivariate distribution function F can
be represented through some copula function.
The most known copula function is the Gaussian copula, which has the following form:

1
C (u1 , u2 ) = 1
1 (u1 ) , 2 (u2 ) ,
where denotes the joint cumulative Normal distribution, and i denotes marginal cumulative Normal
distributions. So we have,

1
F (y1 , y2 ) = C (F1 (y1 ) , F2 (y2 )) = 1
(12A.7)
1 (F2 (y2 )) , 2 (F2 (y2 )) ,
412
12.10. Appendix 5: Modeling correlation with copulae functions
c
by
A. Mele
where the rst equality follows by Eq. (12A.6) and the second equality follows by Eq. (12A.7).
As an example, we may interpret Y1 and Y2 as the times by which two names default. A simple
assumption, then, is to set:
Fi (yi ) = i (zi ) , i = 1, 2,
(12A.8)
for two random variables Zi that are stretched as explained in Part A of this appendix. By replacing
Eq. (12A.8) into Eq. (12A.7),
F (y1 , y2 ) = (z1 , z2 ) .
This reasoning can be easily generalized to the N-dimensional case, where:
F (y1 , , yN ) = C (F1 (y1 ) , , FN (yN )) = (z1 , , zN ) ,
where
zi : Fi (yi ) = i (zi ) .
We use this approach to model default correlation among names, as explained in the main text.
413
12.11. Appendix 6: Details on CDO pricing with imperfect correlation
c
by
A. Mele
12.11 Appendix 6: Details on CDO pricing with imperfect correlation

We follow the copula approach to price the stylized CDOs in the main text of this chapter. For each
name, create the following random variable,

zi = F + 1 i , i = 1, 2, 3,
(12A.9)
where F is a common factor among the three names, i is an idiosynchratic term, and F N (0, 1),
i N (0, 1). Finally, 0 is meant to capture the default correlation among the names, as follows.
Assume that the risk-neutral probability each rm defaults, by T , is given by,
Qi (T ) = ( 0.10 ) 10%,
where is the cumulative distribution of a standard normal variable. That is, by time T , each rm
defaults any time that,
zi < 0.10 1 (10%) .
Therefore, is the default correlation among the assets in the CDO.
We can now simulate Eq. (12A.9), build up payos for each simulation, and price the tranches by
just averaging over the simulations, as explained below. Naturally, the same simulation technique can
be used to price tranches on CDOs with an arbitrary number of assets. Precisely, simulate Eq. (12A.9),
and obtain values zi,s , s = 1, , S, where S is the number of simulations and i = 1, 2, 3. At simulation
no s, we have
z1,s , z2,s , z3,s , s {1, , S} .
We use the previously simulated values as follows:
For each simulation s, count the number of defaults across the three names, dened as the
number of times that zi,s < 0.10 , for i = 1, 2, 3. Denote the number of defaults as of simulation
s with Def s .
For each simulation s, compute the total realized payo of the asset pool, dened as,
s = Def s 40 + (3 Def s ) 100.

For each simulation s, compute recursively the payos to each tranche, i,s ,
+
+
,
,
i1

i,s = min max
s
k,s , 0 , Ni ,
k=1
where Ni is the nominal value of each tranche (N1 = 140, N2 = 90, N3 = 70).
Estimate the price of each tranche by averaging across the simulations,

r
Price Senior = e
S
S
S

1
r 1
r 1
1,s , Price Mezzanine = e
2,s , Price Junior = e
3,s .
S s=1
S s=1
S s=1
Note, the previous computations have to be performed under the risk-neutral probability Q. Using
the probability P in the previous algorithm can only be lead to something useful for risk-management
and VaR calculations at best
Note, this model, can be generalized to a multifactor model where,

zi = i1 F1 + + id Fd + 1 i1 id i ,
with obvious notation.
414
c
by
A. Mele
References
Amato, J. D. (2005): Risk Aversion and Risk Premia in the CDS Market. BIS Quarterly
Review, September, 55-68.
Anderson, R. W. and S. Sundaresan (1996): Design and Valuation of Debt Contracts. Review
Artzner, P., F. Delbaen, J.-M. Eber, and D. Heath (1999): Coherent Measures of Risk.
Berndt, A., R. Douglas, D. Due, M. Ferguson and D. Schranz (2005): Measuring Default
Risk-Premia from Default Swap Rates and EDFs. BIS Working Papers no. 173.
Black, F. and J. Cox (1976): Valuing Corporate Securities: Some Eect of Bond Indenture
Provisions. Journal of Finance 31, 351-367.
Christoersen, P. F. (2003): Elements of Financial Risk Management. Academic Press.
Cox, J. C., J. E. Ingersoll and S. A. Ross (1985): A Theory of the Term Structure of Interest
Rates. Econometrica 53, 385-407.
Fender, I. and P. Hordahl (2007): Overview: Credit Retrenchement Triggers Liquidity Squeeze.
BIS Quarterly Review (September), 1-16.
Hull, J. C. (2007): Risk Management and Financial Institutions. Pearson Education International.
Ingersoll, J. E. (1977): A Contingent-Claims Valuation of Convertible Securities. Journal of
International Monetary Fund, (2008): Global Financial Stability Report. April 2008.
Jamshidian, F. (1989): An Exact Bond Option Pricing Formula. Journal of Finance 44,
205-209.
Jarrow, R. A., D. Lando and S. M. Turnbull (1997): A Markov Model for the Term-Structure
of Credit Risk Spreads. Review of Financial Studies 10, 481-523.
Jorion, Ph. (2008): Value at Risk. New York: McGraw Hill.
Leland, H. E. (1994): Corporate Debt Value, Bond Covenants and Optimal Capital Structure. Journal of Finance 49, 1213-1252.
Leland, H. E. and K. B. Toft (1994): Optimal Capital Structure, Endogenous Bankruptcy,
and the Term Structure of Credit Spreads. Journal of Finance 51, 987-1019.
Lopez, J. (2004): The Empirical Relationship Between Average Asset Correlation, Firm Probability of Default and Asset Size. Journal of Financial Intermediation 13, 265-283.
415
c
by
A. Mele
McDonald, R. L. (2006): Derivatives Markets, Boston: Pearson International Edition.

Merton, R. C. (1974): On the Pricing of Corporate Debt: The Risk-Structure of Interest
Rates. Journal of Finance 29, 449-470.
Sklar, A. (1959): Fonction de Repartition à N dimensions et Leurs Marges. Publications de
lInstitut Statistique de lUniversite de Paris 8: 229-231.
Vasicek, O. (1987): Probability of Loss on Loan Portfolio. Working paper KMV, published
in: Risk (December 2002) under the title Loan Portfolio Value.
416
13
Financial engineering and fixed income securities
13.1 Introduction
13.1.1
Relative pricing in fixed income markets
This chapter lies down foundational issues relating to nancial engineering for xed income
securities. Fixed income securities can be particularly complex, as outlined in the previous
two chapters. Many instruments in the xed income markets dier substantially from those in
the remaining portions of the capital markets. For example, a simple instrument such a pure
discount bond is very dicult to price. Intuitively, the price of a pure discount bond reects
the time value for money. It is related to the intertemporal preferences and beliefs of the
market participants, which are unobservable. The situation is dierent in the case of traditional
relative pricing, i.e. when we price a number of assets given the price of some other assets,
while ensuring that there are no arbitrage opportunities left on the table. In this case, we
can evaluate derivatives without reference to any preferences or beliefs. The Black & Scholes
formula, for example, is a preference free formula, although this type of formula or reasoning
cannot exactly be applied to evaluate xed income securities, as explained below.
13.1.2
Complexity of fixed income securities
The rapid growth in the xed income markets was also led by many new instruments that
are substantially more complex than the traditional plain vanilla bonds (i.e. default-free, noncallable bonds, defaultable bonds), or other instruments related to credit risk transfers, or
baskets of xed income instruments or callable bonds, where the borrower can call the contract
to anticipate the payment of the principal, as we have seen in the previous chapter.
We have seen that the standard tools of asset evaluation are unlikely to work in this context.
For example, we cannot even hope to adapt such models as the Black & Scholes model to
price interest rate derivatives. Indeed, the Black & Scholes model relies on the assumption
of a constant volatility of the asset price underlying the contract. In the context of interest
rate derivatives, instead, the volatility of the underlying asset price depends on the maturity
of the underlying (tends to zero as the maturity goes to zero). More generally, pricing and
hedging interest rate derivatives requires a model that describes the evolution of the entire term
c
by
A. Mele
13.2. Bootstrapping and curve tting
structure of interest rates. Academics and practitioners have proposed a variety of solutions to
this problem, from the mid 80s to the beginning of the 90s. Today, dozens of new methods are
available to price xed income products. The general principles underlying the APT are still
the same, though.
13.1.3
Many evaluation paradigms
While dozens of new methods are available to price xed income products, we do not see the
emergence of a single model to price all of the extant xed income products! Market participants use dierent models to price interest rate derivatives. Typically, a single investment bank
has a battery of dierent models with which to ght in the market. Pieces of this battery
may ght for dierent goals. For example, an investment bank might display a preference for a
certain type of models as a result of (i) its culture and history (see, e.g., the intellectual legacy
of Fisher Black and Emanuel Derman in Goldman), (ii) the particular business the bank is
pursuing. For example, we have seen that to price options on interest rates such as caps, we
may use the market model, which relies on the Black 76 formula. However, using this model
implies that we do not have a closed-form solution for the price of swaptions, which can only
be solved through numerical methods. If the swaptions business is not important for the bank
then, we may safely adopt the market model.
This chapter presents the main challenges to solve for complicated models, while ensuring
that all the products in the books are perfectly tted.
13.2 Bootstrapping and curve tting

We start with a standard denition. The yield to maturity y (YTM, henceforth) on a bond is
its rate of return. It is the discount rate that would equate the present value of the stream of
payos with its market price,
y : B (T ) =
n

i=1
1
Cti
.
ti +
(1 + y)
(1 + y)T
(13.1)

C
1
This formula diers from the price formula B (T ) = ni=1 [1+r(tti )]ti + [1+r(T
, as Eq. (13.1)
)]T
i
uses the same discount rate y to discount the future payements. Clearly, for zeros we have,
y = R (T ).
13.2.1
Extracting zeros from bond prices
In principle, the zeros can be extracted from the market price of the bonds, provided there is
a sucient spread of bonds across maturities. As an example, consider three bonds. The rst
bond pays o at T1 , the second bond pays o at T1 , T2 , the third bond pays o at T1 , T2 , T3 . By
no-arbitrage,
B (T1 )
C11 + 1
0
0
P (T1 )
B (T2 ) = C21
P (T2 ) ,
C22 + 1
0
B (T3 )
C31
C32
C33 + 1
P (T3 )
418
c
by
A. Mele
for some coupons Cij . Therefore, we can use the observed prices B (t, Ti ) and the payments Cij
to calculate the zeros P (t, Ti ) as,
P (T1 )
C11 + 1
0
0
B (T1 )
P (T2 ) = C21
B (T2 ) .
C22 + 1
0
P (T3 )
C31
C32
C33 + 1
B (T3 )
(13.2)
The previous procedure can be generalized to the case in which some maturity is missing.
The resulting algorithm is known as the bootstrap, which is described next.
13.2.2
Bootstrapping
Bootstrapping proceeds as follows. Let Bi be the price of a bond paying o coupons at the
sequence of dates t1 , t2 , , ti and a principal of $1 at ti . Let Pi be the price of the zero
maturing at ti . Then,
(i) The equation B1 = (C11 + 1) P1 implies that we can extract the zero P1 as follows,
B1
P1 = 1+C
.
11
(ii) Given the equation (C22 + 1) P2 + C21 P1 = B2 , and the previously computed P1 , we
21 P1
proceed to extract the zero P2 as follows, P2 = B2CC
.
22 +1
(iii) In general, we extract the zero Pn as follows, Pn =
n1
Bn
Cni Pi
i=1
.
Cnn +1
(iv) The previous steps work if we have an ordered number of bonds and all of the maturity
dates. Indeed, the previous procedure boils down to the computation of the solution of
Eq. (13.2). When some of the maturity dates are not available, we replace the required
coupon rate Cni at time ti with a linear interpolation Cni between the coupon Cn,i1 at
time ti1 and Cn,i+1 at time ti+1 , as follows,
ti+1 ti
ti ti1
Cni =
Cn,i1 +
Cn,i+1 .
ti+1 ti1
ti+1 ti1
The eects of the interpolation should be visible near the missing maturitites.
Consider a sequence of coupon bearing bonds maturing at n with xed coupon streams Cn .
Then, as explained in Chapter 11, let us dene the par yield curve as the sequence of Cn such
that the price Bn is forced to equal 100%. Therefore, we can extract zeros and, then, the
yield curve, from step (iii) above, by just using the the recursive formula,

Bn Cn n1
i=1 Pi
,
Pn =
Cn + 1
419
(13.3)
c
by
A. Mele

where Bn = 100%. The following table provides a numerical example.
n
Coupon
Maturity, n
Zero price
i=1 Pi
6.00%
1
0.9434
0.9434
7.00%
2
0.8728
1.8162
8.00%
3
0.7914
2.6076
9.50%
4
0.6870
3.2946
9.00%
5
0.6454
3.9400
10.50%
6
0.5306
4.4706
11.00%
7
0.4579
4.9285
11.25%
8
0.4005
5.3290
11.50%
9
0.3472
5.6762
11.75%
10
0.2980
not useful
Discretely compounded
13.2.3
Yield curve
6.00%
7.04%
8.11%
9.84%
9.15%
11.14%
11.81%
12.12%
12.47%
12.87%
Curve fitting
We may use statistical techniques alternative to bootstrapping, to cope with situations in which
the number of bonds does not equal the number of maturity dates. Suppose we observe N bonds,
where the i-th bond entitles to receive the coupons Cij , for j = 1, , Mi . We assume that the
bond prices are observed with errors, or
B (Mi ) =
Mi

j=1
Cij P (tj ) + P (tMi ) + i , i = 1, , N,
where i is the measurement error for the i-th bond.

We aim to nd the curve T P (T ) that minimizes the errors, in some statistical sense. The
natural device is to parametrize the function P (T ), with a number of k parameters, where
k < N . To parametrize the function P (tj ) for a generic tj , we can use polynomials, as originally
suggested by McCulloch (1971, 1975),
P (tj ) = 1 + a1 tj + a2 t2j + + ak tkj ,
where the ai are the parameters. Cubic splines are polynomials up to the third order, and
are verypopular. The parameters ai can be estimated by minimizing the sum of the squared
2
errors, N
i=1 i . A well-known pitfall of polynomials is that a high k might imply that while the
polynomial approximation works reasonably well near the observed maturities, it may exhibit
an erratic behavior in between. To avoid this problem, we can use local polynomials, which are
low-order polynomials (typically splines) tted to non-overlapping subintervals.
Naturally, we may also want to parametrize the spot rates, R (T ), as polynomials. Alternatively, Nelson and Siegel (1987) propose the following parametrization,

1 eT
1 eT
T
R (T ) = 1 + 2
+ 3
e
,
T
T
where i and are the parameters. These coecients may be given an interpretation, in terms
of the factors driving the yield curve, reviewed in Chapter 11. The coecient 1 governs the
level of the yield curve. The coecient 2 relates to the slope, as an increase in this coecient
420
13.3. Duration, convexity and asset liability management
c
by
A. Mele
increases short yields more than long yields. The coecient 3 shapes the curvature, as an
increase in this coecient has little eect on very short and very long yields, but increases the
middle of the yield curve. Moreover, the coecient controls the exponential decay of the yield
curve: small values of translate to slow decay and can better t the curve at long maturities;
large values of , instead, lead to a fast decay, which helps t the short-end of the yield curve.
Finally, determines where the loading on 3 achieves its maximum. Diebold and Li (2006)
have used this setting to estimate i for each date, and then used these estimated time series
of i to forecast future values of i through vector autoregressions and, then, the future yield
curve.
13.3 Duration, convexity and asset liability management

The risk of longing a default-free bond is that the future bond price is uncertain, due to the
possibility that the spot interest rates could change in the future. Synthetically, we can say that
the risk of a bond is related to the changes in the required bond return, or the YTM. Consider
the denition of the YTM y in Eq. (13.1). Next, consider the following function B (y; T ),
B (y; T ) =
n

i=1
Cti
1
.
ti +
(1 + y)
(1 + y)T
This function aims to mimic how the market price B (T ) would behave if the YTM y changed
to some value y. Naturally,
B (
y ; T ) = B (T ) .
Motivated by the previous remarks, we can dene a measure of risk of the bond based on
the sensitivity of the bond price with respect to changes in y. Economically, we are trying
to answer the following question: What happens to the bond price once we perturb the one
rate y that discounts all the payos? Mathematically, this sensitivity is the rst partial of the
bond-pricing formula B (y; T ) with respect to y,
- n
.

1
T 1
ti Cti
By (y; T ) =
ti +
1 + y i=1 (1 + y)
(1 + y)T
B (y; T ). Graphically, this

where the subscript denotes a partial derivative, i.e. By (y; T ) = y
sensitivity measure By (y; T ) is the tangent to the price-yield relation, as shown in Figure 13.1
below.
13.3.1
Duration
We dene the Macaulay duration as,

n
DMac

By (y; T )
(1 + y) =
ti ti +
T T,
B (y; T )
i=1
where
ti =
Cti / (1 + y)ti
1/ (1 + y)T
,
T =
.
B (y; T )
B (y; T )
421
c
by
A. Mele
B o n d p ric e
2 nd o rd e r a p p ro x im a tio n
1
st
o rd e r a p p ro x im a tio n
YTM
FIGURE 13.1. The bond price-yield relation (solid line), its rst-order approximation (duration) and
its second-order approximation (convexity).
In words, the Macaulay duration is a weighted average of the payment dates. The weights ti
are the discounted coupons at the various payment dates, Cti / (1 + y)ti , related to the current
market value of these coupons, i.e. the bond price B (y; T ) when the YTM is y. That is, the
weights are the proportionsof the bonds present value that is attributable to the payo at
date t. The weights satisfy ni=1 ti +
T = 1. Therefore, DMac T . The Macaulay duration is
a measure of how far in the future the bond pays o. For zeros, DMac = T .
For small y, DMac (y) is simply the semi-elasticity of the bond price with respect to the YTM.
This semi-elasticity is also referred to as modied duration:
D
DMac
By
=
.
B
1+y
2
Byy
By
A simple computation reveals that the modied duration, D, satises: D
=
+
.
y
B
B
Therefore, the modied duration is decreasing in the YTM when the bond price is suciently
convex in the YTM, which is surely the case for long-term maturity dates.
Interestingly, the modied duration is increasing in the YTM when the bond price is concave
in the YTM, a property that arises for callable bonds and mortgage-backed securities (MBS,
henceforth), as explained in Chapters 11 and 12. Intuitively, the incentives to proceed to early
repayments kick in as the YTM decreases, which makes the duration of the MBS decrease.
The Macaulay duration for continuously compounded rates is even simpler to compute. First,
dene the continuously compounded YTM as the single number x such that
B(
x; T ) =
n
cti exti + exT ,
i=1
where B(
x; T ) is the market price of a bond paying o the principal of one at maturity and the
stream of payos cti . Next, consider, the function x B (x; T ). Compute the semi-elasticity of
the bond price B (x; T ) with respect to the continuously compounded YTM x,
Bx (x; T )
=
B (x; T )
n

+ T exT
=
wti ti + wT T,
B (x; T )
i=1
422
xti
i=1 cti ti e

c exti
c
by
A. Mele
xT
)
ti
e
where Bx (x; T ) = B(x;T
, wti = B(x;T
and wT = B(x;T
. Note, the weights are such that
x
)
)
n
T = 1. Therefore, the Macaulay duration for continuously compounded rates
i=1 wti + w
is equal to the semi-elasticity of the bond price with respect to the continuously compounded
YTM x.1 This result may simplify some calculations.
13.3.2
Convexity
Convexity measures how the sensitivity, By , changes with y. Mathematically, convexity is related
to the second partial of the bond price with respect to y, Byy . If the second partial, Byy , is
positive, then, the interest rate sensitivity declines as y increases (see Figure 13.1). This is
because y
(By ) = Byy < 0. Formally, convexity is dened as,
C
Byy
.
B
We may, then, consider the following expansion of the bond price:

B
1
D y + C (y)2 .
B
2
That is, for very convex securities, duration may not be a safe measure of return, as also
shown in Figure 13.1
13.3.3
Duration and asset-liability management
13.3.3.1 Introductory issues
We can use duration to assess how exposed a bond portfolio is to movements in the interest
rates. We can then immunize a portfolio of bonds to changes in the interest rates. Duration
is relevant for asset-liability management. For example, pension funds have known streams of
liabilities that must be matched by the assets they hold. In words, the duration of the assets
must equal the duration of the liabilities. In the UK, pension funds must mark-to-market the
liabilities. Therefore, one objective of these funds is to immunize their liabilities against
movements in the interest rates.
Alternatively, consider the following basic example. A bank borrows $100 at 2% for a year
and lends this money at 4% for 5 years, where the higher rate compensates for many things
such as risk, the banks market power, etc. Assuming that the banks borrower does not default,
in the rst year, the bank generates prots equal to $(4% 2%) 100 = 2, according to its
books. However, the right computation to make should not relate to past market (interest rate)
conditions, but to the current ones. Suppose for example that in one year, the interest rate
for borrowing raises from 2% to 5%. This is of course a bit unrealistic, but it gives the idea
5
of where the action is. In this case, The market value of the assets is: 1001.04
= 100.09. The
1.054
market value of the liabilities is, of course, 100 1.02 = 102. The banks problem is, of course,
a duration mismatch.
Let us consider a more substantive example, based on asset-liability management for pension
funds. We consider the following extreme example. In 30 years from now, a pension fund is due
1 Mathematically, we could have obtained this result in a straightforward manner, as follows. Dene the bond price function as
B (y (x)), where by denition, y (x) = ex 1. Hence, Bx (y (x)) = By (y (x)) y (x) = By (y (x)) ex = By (y (x)) (1 + y). It follows
that DMac =
By (1+y)
B
Bx
.
B
423
c
by
A. Mele
to deliver $100,000 to some future retiree. Suppose the current market situation is such that
the yield curve is at at 4%, such that the market value of this liability is $100, 000(1.04)30 =
$30, 832. Accordingly, the would-be retiree invests $30.832 in the pension fund. So we have
the following situation:
Cash
Pensions
$30, 832 $30, 832
Suppose, now, that the pension fund does not invest this cash. This is of course inecient, but
it is precisely the point of this simple exercise to see why the strategy is inecient.
Consider two extreme cases, occurring under two scenarios underlying developments in the
xed income market. In one week,
(i) Scenario : the yield curve shifts up parallely to 5%. Accordingly, the value of the liability
for the pension fund is: $100, 000 (1.05)30 = 23, 138.
Cash
$30, 832
Prot
$7, 694
Pensions
$23, 138
(ii) Scenario : the yield curve shifts down parallely to 3%. Accordingly, the value of the
liability for the pension fund is: $100, 000 (1.03)30 = 41, 199.
Cash
Loss
$30, 832 $10, 367
Pensions
$41, 199
Therefore, a drop in the yield curve results in a loss for the pension fund: when interest rates
go down, the pension fund faces a challenging situation as it has to honour its obligations in
30 years, but the nancial market yields less than one week ago.
Naturally, the pension fund would face the opposite situation were interest rates to go up.
In some countries, we do not like pension funds to experience volatility. The previous volatility
arises simply because the pension fund, receives $30, 832, and then it just puts this money
under the pillow. The most ecient way to kill volatility is, of course, to invest $30, 832 in
a 30 bond as soon as we receive this moneyat the market conditions of 4%. This is perfect
hedging! But, we do not necessarily have access to such a bond. How do we proceed, then?
We now develop examples that illustrate how to deal systematically with issues relating to
asset-liability management.
13.3.3.2 Hedging
Let us consider a portfolio of two bonds with dierent durations. Its value is given by,
V = B1 (
y1 ) 1 + B2 (
y2 ) 2 ,
424
c
by
A. Mele
where B1 (
y1 ) and B2 (
y2 ) are the market value of the bonds, y1 and y2 are the YTM on the
bonds and, nally, 1 and 2 are the quantities of bonds in the portfolio. Let us consider a small
change in the two YTM y1 and y2 . We have,
dV = [D (
y1 ) B1 (
y1 ) 1 d
y1 + D (
y2 ) B2 (
y2 ) 2 d
y2 ] .
The question is: How should we choose 1 and 2 so as to make the value of the portfolio remain
constant after a change in y1 and y2 ?
Let us assume a parallel shift in the term structure of interest rates. In this case, d
y1 = d
y2 .
The portfolio is said to be immunized if its value V does not change as y1 and y2 change, i.e.
dV = 0, which is true when,
y2 )
D (
y2 ) B2 (
1 =
2 .
(13.4)
D (
y1 ) B1 (
y1 )
A useful interpretation of this portfolio is that we may be holding a bond with some duration,
say we hold 2 units of the second bond. Given these holdings, we may wish to sell another
bond, possibly with a lower duration, to hedge against movements in the price of the bond we
hold.
Alternatively, we can think of the second asset as a liability the value of which uctuates after
a change in the interest rates. Then, we may wish to purchase some asset to hedge against the
liability. Mathematically, 2 < 0 and 1 > 0. Moreover, Eq. (13.4) reveals that the number of
assets to hold to hedge against the liability is high if the ratio of the two durations of the assets,
D (
y2 )/ D (
y1 ), is large. In this case, the hedging position is obviously inecient. Asset-liability
management, and immunization, is costly when we hedge high-duration liabilities with low
duration assets. We now illustrate these cases through a few basic examples.
13.3.3.3 A rst example: hedging zeros with zeros
Suppose that we hold one bond, a zero with maturity equal to 5 years. We want to hedge the
risk of this bond through another bond, a zero with maturity equal to 1 year. Let us assume
that the term-structure is at at 5%, discretely compounded. Then,
1
1
DMac (
1
y1 )
=
= 0.95238, D (
y1 ) =
=
= 0.95238
1 + y1
1 + 0.05
1 + y1
1 + 0.05
1
1
DMac (
y2 )
5
B2 (
y2 ) =
D (
y2 ) =
=
= 4.7619
5 =
5 = 0.78353,
1 + y2
1 + 0.05
(1 + y2 )
(1 + 0.05)
B1 (
y1 ) =
and:
1 =
D (
y2 ) B2 (
y2 )
4.7619 0.78353
2 =
1 = 4.1135.
D (
y1 ) B1 (
y1 )
0.95238 0.95238
That is, to hedge the 5Y zero, we need to short-sell approximately four 1Y zeros. The balance
of this hedging position is,
B1 (
y1 ) 1 + B2 (
y2 ) 2 = (4.1135) 0.95238 + 0.78353 = 3.1341,
(13.5)
a quite inecient hedge.

The reason why this is inecient is clear. Hedging high maturity bonds with short maturity
ones implies we should rebalance too often. Moreover, as time goes on, the sensitivity of the
short-term bonds to changes in the YTM is very small (at the extreme, the price equals face
425
c
by
A. Mele
value plus coupon, at maturity), compared to that of long-term bonds. Therefore, rebalancing
becomes increasingly severe as time unfolds.
Next, we study how the value of this portfolio changes after large changes in the YTM.
By the assumption that the initial term-structure is at at 5%, y1 = y2 = 5%. Moreover, by
rearranging Eq. (13.5),
B2 (y = 5%) = 4.1135 B1 (y = 5%) 3.1341.
The left hand side of this equation is the price of the 5Y bond. The right hand side is the value
of the replicating portfolio, which consists of (i) approximately 4 units of the 1Y bond, and
(ii) the balance of the hedging position.
When y = 5%, the previous relation can only approximately hold,
B2 (y) 4.1135 B1 (y) 3.1341.
Figure 13.2 below plots the left hand side and the right hand side of this relation.
1.0
0.9
0.8
0.7
0.6
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.10
YTM
1
FIGURE 13.2. Dotted line (top): The price of the 5Y zero, B2 (y) = (1+y)
5 , where y is
the YTM. Solid line (bottom): The value of the replicating portfolio consisting of (i)
4.1135 units of the 1Y zero, and (ii) the balance of the hedging position, which is equal
1
is the 1Y zero price.
to $3.1341, i.e. 4.1135 B1 (y) 3.1341, where B1 (y) = 1+y
What is going on? We are hedging the 5Y zero by selling approximately four 1Y zeros. In a
neighborhood of y = 5%, the value of the synthetic 5Y zero we sold, 4.1135 B1 (y) 3.1341,
behaves as B2 (y). However, the 5Y zero displays more convexity than the synthetic bond.
This larger convexity implies that:
If the interest rates go down, the price of the 5Y zero bond we hold increases more than
the value of the synthetic bond we sold. As a result, we make prots.
If the interest rates go up, the price of the 5Y zero bond we hold decreases less than the
value of the synthetic bond we sold. As a result, we make prots.
426
c
by
A. Mele
In all cases, we make prots.2 However, this is not an arbitrage opportunity! The previous
reasoning hinges on the assumption of a parallel shift in the term-structure of interest rates,
that is d
y1 = d
y2 , where y1 = spot rate for 1 year, and y2 = spot rate for 5 years. While parallel
shifts in the term-structure seem empirically relevant, they are not the only shifts that are likely
to occur, as we explained in Chapter 11.
13.3.3.4 Fixed income arbitrage
Swap spread arbitrage is a popular strategy. It was responsible for leading LTCM to a loss of
about $1.6 billion in 1997. The strategy works as follows: (i) enter a swap paying the oating
(ii) short a par Treasury with the same maturity as
LIBOR, Lt , and receiving a xed rate C;
the swap, thus paying the xed coupon rate CT , and invest the proceeds at the repo rate rt .
Thus, the payo of the strategy is the xed spread to be received, F = C CT , and the oating
spread to be paid, St = Lt rt . So we go long or short this strategy according to whether we
view F to be larger or smaller than the average oating spread St over the strategy horizon.
Historically, the spread St has certainly been volatile, but quite stable, so it is a reasonable
strategy. The problem occasionally, though, St can reach quite large levels.
A second strategy is yield curve arbitrage. For example, go long ve year bonds, and short
two- and ten-year bonds, an implicit view that short term interest raise will raise and medium
term interest rates will lower. This buttery strategy is also known as barbell trading, and
will be further illustrated in the next subsection. This might be quite cheap, intellectually, and
not necessarily rewarding.
More sophisticated strategies rely on models, which identify which points of the yield curve
are misaligned from those predicted by the model. The strategy, substantially, is: buy the poor
and short the model-based rich, where the model-based rich is replicated through a portfolio
with cash and the bonds that are well-priced by the model, weighted with the model-based
delta.
13.3.3.5 Duration trading: Barbell and bullet hedges
As a second example of duration hedging, consider the barbell trading strategy, which is a
way to hedge some liability (a bullet) with duration D2 through two assets with durations
D1 and D3 , where D1 < D2 < D3 . This trading strategy is expected to work when we expect
the yield curve to atten, with its short-end part not going too much high. Moreover, investing
in the short-term segment of the yield curve, allows one to invest elsewhere relatively rapidly
once the rst asset expires, were the bond market to go down.
Let us consider the previous example, and suppose there is another bond available for trading,
a zero with maturity equal to 10 years. We aim to hedge against movements in the price of the
5Y zero with a portfolio consisting of (i) one 1Y zero and (ii) the 10Y zero. We continue to
2 Mathematically, we buy 1 unit of the 5Y zero at B and sell units of the 1Y zero at B , thereby cashing in B B = 3.1341.
2
1
1
1 1
2
Then, in one one month (say), consider what would happen if we had to reverse the position in Eq. (13.5), i.e. sell the 5Y zero and
buy back the 1Y zeros we sold. We consider three scenarios: (i) The yield curve will be the same as today. In this case, reversing the
position in Eq. (13.5) implies that we shall simply have to pay 3.1341 (assuming the change in value of the two bonds due to the
mere passage of time is small enough). (ii) The yield curve will experience a positive parallel shift. In this case, the prices of the two
zeros will be B1 B1 and B2 B2 , where B1 and B2 are both positive. Therefore, we shall obtain, 3.1341 +1 B1 B2 ,
where 1 B1 B2 is positive because by convexity, the value of the portfolio decreases more than the price of the 5Y zero, thus
yielding a prot. (iii) The yield curve will experience a negative parallel shift. In this case, the prices of the two zeros will be
B1 + B1 and B2 + B2 , where B1 and B2 are both negative. Therefore, we shall obtain, 3.1341 + B2 1 B1 , where
B2 1 B1 is positive because by convexity, the value of the portfolio increases less than the price of the 5Y zero, thus yielding
a prot.
427
c
by
A. Mele
assume that the yield-curve is at at 5%, and only consider parallel shifts in the term-structure
of interest rates.
Such a buttery trade can be implemented as follows. We look for a portfolio of the 1Y and
10Y zero with the following properties: (i) the market value of the portfolio equals the market
price of the 5Y zero,
B2 (
y2 ) = B1 (
y1 ) 1 + B3 (
y3 ) 3 ;
(13.6)
and (ii) the duration of the portfolio equals the duration of the 5Y zero,
D (
y2 ) B2 (
y2 ) = D (
y1 ) B1 (
y1 ) 1 + D (
y3 ) B3 (
y3 ) 3 .
(13.7)
The solution to Eqs. (13.6) and (13.7) is given by,

1 =
y2 ) B2 (
y2 )
D (
y3 ) D (
,
D (
y3 ) D (
y1 ) B1 (
y1 )
3 =
y1 ) B2 (
y2 )
D (
y2 ) D (
.
D (
y3 ) D (
y1 ) B3 (
y3 )
(13.8)
By the same computations made in the previous example, we have that B3 (

y3 ) = 0.61391
and D (
y3 ) = 9.5238. By using the gures in the previous example, we compute 1 and 3 in
Eqs. (13.8) to be
1 =
9.5238 4.7619 0.78353

= 0.45706,
9.5238 0.95238 0.95238
3 =
4.7619 0.95238 0.78353

= 0.56724.
9.5238 0.95238 0.61391
Figure 13.2 depicts the behavior of the bullet price and the market value of the barbell as
we change the YTM. Several comments are in order. First, note that the barbell portfolio is
now more convex than the bullet! Now, large movements in the YTM lead to prots, provided
we maintain the assumption of parallel shifts in the term-structure of interest rates. Second,
the barbell trade is self-nanced. By construction, the value of the bullet we sell equals the
value of the barbell portfolio. However, the barbell is clearly not an arbitrage opportunity. The
scenario underlying Figure 13.3 relies on the assumption of a parallel shift in the term structure
of interest rates. As we explained in Chapter 11, it is not realistic to simultaneously assume
large and parallel movements in the term-structure of interest rates. Historically, large interest
rate shifts (that is, typically, shifts occurring over large horizons of time) are accompanied by
the occurrence of a variety of shape modications.
1.0
0.9
0.8
0.7
0.00
0.01
0.02
0.03
0.04
0.05
428
0.06
0.07
0.08
0.09
0.10
YTM
c
by
A. Mele
FIGURE 13.3. Barbell trading. Dotted line (bottom): The price of the 5Y zero, B2 (y) =
1
, where y is the YTM. Solid line (top): The value of the barbell portfolio consisting
(1+y)5
of (i) 0.45706 units of the 1Y zero and (ii) 0.56724 of the 10Y zero, i.e. B1 (y1 ) 0.45706 +
1
1
B3 (y3 ) 0.56724, where B1 (y) = 1+y
is the 1Y zero price and B3 (y) = (1+y)
10 is the 10Y
zero price.
Table 13.1 considers the case of non-parallel shifts in the term-structure. We assume that
the initial term-structure is not at. Then, we consider two scenarios: (i) A twist in the
term-structure, i.e. long-term rates lower than short-term rates; (ii) a steepening of the termstructure.
TABLE 13.1.
YTM
Initial term-structure
1Y
y1 = 4%
5Y
y2 = 5%
10Y
y3 = 6%
Bullet price
Mod. dur.
Barbell value =
1 B1 (
y1 ) + 3 B3 (
y3 )
B1 (
y1 ) = 0.961 D (
y1 ) = 0.961
B2 (
y2 ) = 0.783 D (
y2 ) = 4.762
B3 (
y3 ) = 0.558 D (
y3 ) = 9.434
Barbell value = 0.783
Twist
1Y
5Y
10Y
y1 = 6%
y2 = 5%
y3 = 4%
B1 (
y1 ) = 0.943 D (
y1 ) = 0.943
B2 (
y2 ) = 0.783 D (
y2 ) = 4.762
B3 (
y3 ) = 0.675 D (
y3 ) = 9.615
Steepening
1Y
5Y
10Y
y1 = 4%
y2 = 5%
y3 = 7%
B1 (
y1 ) = 0.961 D (
y1 ) = 0.961
B2 (
y2 ) = 0.783 D (
y2 ) = 4.762
B3 (
y3 ) = 0.508 D (
y3 ) = 9.346
We use the portfolio in Eq. (13.8), and nd that in correspondence of the initial term-structure
(
y1 = 4%, y2 = 5%, y3 = 6%), 1 = 0.449 and 3 = 0.629. We keep this portfolio xed, and
compute the barbell value, 1 B1 (
y1 ) + 3 B3 (
y3 ), occurring at the two scenarios twist and
steepening. The convexity of the barbell trade is in fact a bet on long-term bonds and leads
to a prot in the twist scenario (since B2 (
y2 ) = 0.783 in all cases). That is, by convexity, the
price B3 varies more than the price of shorter maturity zeros, thus leading to prots. However,
note that the barbell bet leads to losses in the steepening scenario.
A caveat. The previous computations should be interpreted with some care, as the value of
the zeros changes over time. Notably, the value of the zeros changes over the horizon after which
we are designing scenarios, even without any changes in the yield curve. However, this eect
is usually minor when the horizon is suciently small and, generally, can be factored into the
analysis.
To sumup, duration hedging is a useful tool, although it has some limitations. It is only a
rst-order approximation to the price of bond. A conventional bond is typically strictly convex
in the YTM. Therefore, for large changes in the YTM, we should update the duration-based
429
c
by
A. Mele
13.4. Foundational issues on interest rate modeling
hedging ratios. Re-adjustments are in order anyway, since xed income securities have duration
that decreases over time.
13.3.3.6 Negative convexity
What happens when bond prices have negative convexity? In Chapter 12, we saw that the
value of a callable bond can be concave in the short-term rate. A similar feature is displayed
by mortgage-backed-securities (MBS, henceforth), which can now be concave in the YTM! The
reason for this negative convexity is that early repayments are likely to occur as the YTM
decreases, which entails two inextricable consequences: (i) the price of the MBS increases less
than a conventional bond price after a decline in the YTM, especially when the YTM is low;
(ii) the duration of the MBS decreases as the YTM decreases.
MBS may be responsible of nancial turmoil. The mechanism is well-known. Institutions
that hold MBS typically short conventional bonds for hedging purposes. But the MBS dura= Convexity.
tion increases as interest rate increase, due to the negative convexity: Duration
r
Therefore, an interest rate increase can lead these institutions to short additional conventional
bonds, which worsens liquidity and leads to a further increase in the interest rates, thereby
feeding a vicious circle. Perli and Sack (2003) estimate that in 2002 and 2003, this mechanism
may have amplied the volatility of the long-term US rates by a factor between 15% and 30%.
13.4 Foundational issues on interest rate modeling

In principle, the classical ideas underlying contingent claim analysis through binomial trees,
can be put at work in the context of xed income instruments. In this context, however, we
need to revise quite a few methodological details. Let us illustrate the general issues. First, let
us review how binomial trees are constructed, in general:
(i) We begin with a probabilistic representation of how the price develops over time, using a
tree-like information structure.
(ii) For example, at the time of evaluation, we observe the state. In the next period, there
can be two mutually exclusive states of the world: (a) the state up, occurring with
probability p; and (b) the state down, occurring with probability 1 p.
(iii) After two periods, there can be three mutually exclusive states of the world, as in the
following diagram. We label the tree in this diagram a recombining tree, to emphasize
that the up & down and the down & up nodes are the same.
p
u p
state
1 -p
Today
p
1 -p
u p , u p
u p , d o w n
d o w n , u p
d o w n
state
1 -p
d o w n , d o w n
F irst p erio d
430
S eco n d p erio d
c
by
A. Mele
The previous diagram can be used to price options written on stocks. The stock price unfolds
through the branches of the tree. Then, we gure out the no-arbitrage movements of the option
price along the tree. Suppose, however, we wish to price an option written on a zero, a 3 Year
zero say. Can we apply the same methodology to price the option? The answer is no, and the
reason is that we cannot exogenously track the movements of the prices of the zero, as in the
case of the stock price. Instead, after one year, the 3 Year zero becomes a 2 Year zero, i.e. quite
a dierent asset.
The trick, here, is to model the movements of the yield curve. There are two approaches. In
the rst approach, we model the dynamics of the short-term rate, dened as the interest rate
on a loan with maturity equal to the time intervals in the tree. The resulting model, which in
Chapter 11 we called model of the short-term rate, has implications in terms of the movements of
the entire term-structure. This approach gives rise to evaluation formulae in which the current
prices of the zeros predicted by the model are not necessarily equal to the market prices. We
develop this approach in this section. In a second approach, which in Chapter 11 we called
no-arbitrage, or calibration, approach, we model the dynamics of the entire term-structure.
This approach gives rise to option evaluation formulae in which the current prices of the zeros
predicted by the model are equal to the market prices. We develop this approach in the last
sections of this chapter.
13.4.1
Tree representation of the short-term rate
13.4.1.1 Recursive evaluation
Consider a two-period and two-state tree in which the current short-term rate is r. The development of the short-term rate is uncertain. That is, the future short-term rate, r, is random,
and can take two values: either r+ with probability p, or r with probability 1 p. We assume
that r+ > r . We emphasize p is the physical probability:
p
1p
r+ P (r+ , T )
r P (r , T )
Suppose, also, that two zeros with distinct maturities are available for trading. A money market accounting technology is also available (MMA, in the sequel). Investing $1 in the MMA
generates $1(1 + r) in the second period. We aim to derive an evaluation formula for the zero
based on the previous probabilistic model for the short-term rate dynamics. The general idea
is to build up a portfolio that contains one zero and the MMA. We shall make the value of this
portfolio in the second period replicate the value of the zero we wish to price. By no-arbitrage,
then, the value of the portfolio in the rst period must equal the value of the zero we wish
to price, and we shall be done. The appendix develops the arguments, and shows that in the
absence of arbitrage, there is a constant , such that the following relation holds true:
Ep [P (
r, T )] (1 + r) P (r, T ) =
P (
r, T )
Vol (
r r)
r

= volatility of the price

(13.9)
= unit risk premium
where Vol(
r r) = |r+ r |, and Ep [P (
r, T )] denotes the expectation of the bond price under
the probability p.
431
c
by
A. Mele
As we explained in previous chapters, Eq. (13.9) is an APT relation. It says that the excess
return on the zero equals the volatility of its price multiplied by the unit price of risk. We call
the term,
P (
r, T )
Vol (
r r) ,
r
price volatility because it measures the amplitude of the price variation due to changes in the
(
r,T )
short-term rate in the future, P
, i.e. the price-sensitivity, where this price sensitivity is
r
normalized by the volatility of the short-term rate, Vol(
r r).
Eq. (13.9) can now be cast in a format that we can use to make it more operational. After
rearranging terms, we obtain:
Eq [P (
(p ) P (r+ , T ) + [1 (p )] P (r , T )
r, T )]
=
(13.10)
1+r
1+r
where q p is the risk-neutral probability.
A few considerations. We expect that < 0 because bond prices are decreasing in the
short-term rate here. Then, q p > p.3 Hence, the risk-neutral probability of an upward
movement of the short-term rate, q, is higher than the true probability, p. An investor who longs
a bond, is concerned by an increase of the short-term rate in the future and, hence, corrects
the true probability p by assigning a higher risk-adjusted probability to the upward state.
P (r, T ) =
Assume the current short-term rate equals 10%. We know that with (physical ) probability p,
the short-term rate as of the next year will increase by 2 percentage points, and with probability
1 p, it will decrease by 2 percentage points. Finally, with the same probability p, the shortterm rate prevaling from the next year to two years time, will increase by 2 further percentage
points from its previous value in one year time. Suppose that the probability of an upward
movement is 20% and that the the absolute value of the Sharpe ratio is 30%.
Risk-neutral probability
These data suce to provide an estimate of the risk-neutral probability of an upward movement
of the short-term rate. We simply use the formula, q = p , and obtain q = 20% (30%) =
50%.
Pricing zeros
Next, we can price, say, a zero maturing in two years. We can set up the following tree:
14%
q= 1/2
12%
10%
r = 10%
8%
6%
1 Year
2 Years
3 To be able to interpret q as a probability, we must have that (i) q p > 0 > p and q p < 1 < 1 p.
That is, (p, 1 p)
432
c
by
A. Mele
We can use Eq. (13.10) to ll-in each node of the tree. We start from the end of the tree,
where the price of the two years zero is $1, and then use Eq. (13.10) to ll every node, as
illustrated below.
P = 1
q = 1 /2
0 .8 9 2 8
= 1 /1 .1 2
P = 1
0 .8 2 6 7
0 .9 2 5 9
= 1 /1 .0 8
P = 1
1 Y ear
2 Y ears
The price of the zero, in one year, is simply one divided by the interest relevant at the beginning
of the year, next year. The price we are looking for is obtained by applying Eq. (13.10) yielding,
r, 2)]
Eq [P (
qP (r+ , 2) + (1 q) P (r , 2)
=
=
1+r
1+r
1
2
(0.8928) + 12 (0.9259)
= 0.8267.
1.10
Convexity effects
What is the discretely compounded two-years spot rate? Does it equal 10%? Why or why not?
The two-year spot rate, r (0, 2), satises,
1
0.8266 =
r (0, 2) =
[1 + r (0, 2)]2
1
1 = 9.98%.
0.8266
Even though r = 10% and Eq (

r) = 10%, we have that two years spot rate equals, 9.98%. That
is,

1
1
1
1
Eq
= 0.8264.
0.8266 =
>
1+r
1 + r
1 + r 1 + Eq (
r)
Prices increase after activation of uncertainty. Its a convexity eect similar to that we have
explained in Chapter 11 (Section 11.3.5.1, Figure 11.4).
13.4.2
Tree pricing
We can simply generalize the tree to a multiperiod case. We use Eq. (13.10) to evaluate zeros
at all nodes of the tree and maturities. Given q, which can be estimated once we estimate p
and , we use recursively Eq. (13.10). Then, we may price options on zeros. The weakness of
the approach is that the initial term structure is predicted with error! Let us illustrate this
approach with a concrete numerical example. Consider the following tree, in which the current
433
c
by
A. Mele

short-term rate for one year is r = 4%.
P =1
r = 6%
P = 1 / 1 .0 6 = 0 .9 4 3 3
P =1
r = 5%
r = 4%
r = 4%
P = 1 / 1 .0 4 = 0 .9 6 1 5
P =1
r = 3%
r = 2%
P = 1 / 1 .0 2 = 0 .9 8 0 4
P =1
t =0
t =1
t =2
t =3
FIGURE 13.5. The dynamics of the short-term rate
At time t = 1, the short-term rate is either 5%, with probability p (the true probability) or
3%, with probability 1 p. At time t = 2, the short-term rate behaves as follows:
If at time t = 1, r = 5%
If at time t = 1, r = 3%
then,
then,
at time t = 2, r =
at time t = 2, r =
6% with probability p
4% with probability 1 p
4% with probability p
2% with probability 1 p
Also shown in the previous diagram is the price of a hypothetical 3 Year zero, P , at time
t = 3 and at time t = 2. At time t = 3, the expiration date, P = 1 in all states of nature. At
r, T )]/ (1 + r) = 1/ (1 + r), for r = 6%, 4% and 2%.
time t = 2, the price P is P (r, T ) = Eq [P (
The issue, now, is how to compute the price of the zero in correspondence of the remaining
nodes. We should use the formula, P (r, T ) = Eq [P (
r, T )]/ (1 + r) to populate the tree, but
we do not know p, , and q. Suppose we estimate p and . In this case, we compute q as
q = p , as in Eq. (13.10). (For example, p = 20% and = 30%, so that q = 50%.) Suppose
that we come up with q = 12 . Then, the following diagram gives the price of the zero at all the
434
c
by
A. Mele

nodes as of time t = 1, and at the evaluation time t = 0.
q=
P = 0.9433
1
2
P =1
r = 5%
P = ( 12 0.9433 + 12 0.9615) / 1.05 = 0.9070

q=
P=
r = 4%
1
2
P =1
1
2
P = 0 .9615
0 . 9070 + 12 0 . 9427 ) / 1 . 04 = 0 . 8893

q=
1
2
P =1
r = 3%
P=
( 12 0 .9615 + 12 0 . 9804 ) / 1 .03 = 0 .9427

P = 0.9804
t=0
P =1
t =3
t=2
t =1
So the price of the 3 Year zero equals 0.8893. Next, consider a European call option written
on the 3 Year zero, with expiration date equal to 2 and strike price K = 0.95. The following
diagram gives the value of the option predicted by the model at each node of the tree.
q=
P = 0.9433, K = 0.9500
C = max{P K ,0} = 0
1
2
r = 5%
C = ( 12 0 + 12 0.0115) / 1.05 = 0.0055

q=
1
2
r = 4%
C = ( 12 0 . 0055 + 12 0 . 0203 ) / 1 .04 = 0 .0124
P = 0.9615, K = 0.9500
q=
1
2
C = max{P K ,0} = 0.0115
r = 3%
C = ( 12 0 .0115 + 12 0 . 0304 ) / 1 . 03 = 0 . 0203
P = 0.9804, K = 0.9500
C = max{P K,0} = 0.0304
t =0
t =1
t =2
The model predicts that the current price of the call option is 0.0124.
13.4.2.1 Calibration
The model we are dealing with predicts that the price of the 3 Year zero is equal to 0.8893.
However, there is no guarantee that this model-implied price equals the market price of the 3
Year zero. Suppose, instead, that the market price of the 3 Year zero, P$ say, equals 0.8700.
What should we do to make the model-implied price of the 3 Year zero equal to the market
price? The question is important: how can we trust an option pricing model that is not even
able to pin down the initial market value of the underlying zero?
435
c
by
A. Mele
To make the model-implied price of the 3 Year zero equal to the market price, P$ = 0.8700,
we cannot take the risk-neutral probability q as given, i.e. independent of the observed price
P$ = 0.8700, as we did before. Rather, we should calibrate the probability q, as follows,
P$ = 0.8700 =
1
[q P1 (5%) + (1 q) P1 (3%)]
1.04
(13.11)
where P1 (5%) and P1 (3%) are the prices of the zero at time t = 1, in the events that the
short-term rate is up to 5% or down to 3%.
The previous equation follows, again, by Eq. (13.10). But here, the unknown is not the
price, which is instead given by the market price. Rather, we are looking for, or calibrating,
the probability q that makes the RHS of Eq. (13.11) equal to its LHS. Naturally, we need to
compute the prices of the zeros P1 (5%) and P1 (3%). These prices can be found by another
application of Eq. (13.10), as follows,
P1 (5%) =
q 0.9433 + (1 q) 0.9615
,
1.05
P1 (3%) =
q 0.9615 + (1 q) 0.9804
.
1.03
By replacing the previous expressions for P1 (5%) and P1 (3%) into Eq. (13.11), we obtain,

1
q 0.9433 + (1 q) 0.9615
q 0.9615 + (1 q) 0.9804
P$ = 0.8700 =
q
+ (1 q)
.
1.04
1.05
1.03
This is a nonlinear equation in q, that we can easily solve, to obtain, q = 0.8779. Hence, we
nd:
P1 (5%) = 0.9005 and P1 (3%) = 0.9357.
The next diagram depicts the implied binomial tree, i.e. the tree that results after matching
the model-implied price of the 3 Year zero to the market price, P$ = 0.8700.
q = 0.8779
r = 5%
P1 (5 % ) = [q 0 .9433 + (1 q )0.9615 ] / 1 .05 = 0.9005
P = 0.9433
P =1
P =1
q = 0.8779
r = 4%
P$ = 0.8700 = [qP1 (5% ) + (1 q )P1 (3% )] / 1.04
P = 0.9615
q = 0.8779
P =1
r = 3%
P1 (3 % ) = [q 0 .9615 + (1 q )0 .9804 ] / 1.03 = 0 .9357
P = 0.9804
t =0
t =1
t =2
P =1
t =3
Note how dierent P1 (5%) and P1 (3%) are from those we found earlier by imposing that
q = 12 . In the implied tree, they are smaller than those obtained with q = 12 , state by state.
This is because in the implied tree, q = 0.8779. The implied tree puts more weight on those
436
c
by
A. Mele
states of nature in which the short-term rate is high or, equivalently, bond prices are low. We
expect that the price of the option on the implied binomial tree to be dierent (lower) from
that we found earlier.
So lets do the computations by utilizing the implied binomial tree:
P = 0.9433, K = 0.9500
C = max{P K,0} = 0
q = 0.8779
r = 5%
C = [q 0 + (1 q)0.0115] / 1.05 = 0.0013

q = 0.8779
r = 4%
C = (q 0.0013 + (1 q )0.0134 ) / 1.04 = 0.0026
P = 0.9615, K = 0.9500
C = max{P K,0} = 0.0115
q = 0.8779
r = 3%
C = [q 0.0115 + (1 q )0.0304 ] / 1.03 = 0 .0134
P = 0.9804, K = 0.9500
C = max{P K,0} = 0.0304
t =0
t =1
t =2
The computations in the previous diagram reveal that the option price predicted by the
implied binomial tree is 0.0026, which is one order of magnitude less than the option price
we nd earlier, 0.0124! The interpretation for this result is, again, related to the implied riskneutral probability, which is much larger than q = 12 . The implied tree puts a relatively large
weight on the events in which the short-term rate is high or bond prices are low, which makes
the option price relatively so small.
13.4.2.2 Another zero
We are not done. Let us go back to the zero pricing problem, and suppose that we observe the
price of a 2 Year zero, and that this price equals 0.9200, a quite reasonable gure. Is there any
chance that the inputs to the pricing problem related to the 3 Year zero are such that we can
t the 2 Year zero as well? The answer is, of course, not. There are no reasons for which the
inputs utilized to t the price of the 3 Year zero could also lead to t the price of the 2 Year
zero. The 2 Year zero is quite a dierent asset! Indeed, in the next diagram, we use the inputs
to the pricing problem related to the 3 Year zero, and Eq. (13.10), and nd that the price of
the 2 Year zero implied by the price of the 3 Year zero is equal to 0.9178. Unless the market
price happens, by chance, to equal 0.9178, we cannot simultaneously t the price of the 3 Year
437
c
by
A. Mele

and the 2 Year zeros.
P =1
q = 0.8779
r = 5%
P1 ,1 (5 % ) = 1 / 1 . 05 = 0 . 95 2 3
P =1
P = [q 0 . 95 2 3 + (1 q )0 . 9 7 09 ] / 1 . 04 = 0 .91 7 8
r = 3%
P1,1 (3 % ) = 1 / 1 . 03 = 0 . 97 0 9
P =1
t=0
t =1
t=2
To simultaneously t the price of the 3 Year and the 2 Year zeros, we should implement at
least one of the two strategies: (i) To make the probabilities q time-varying; (ii) To calibrate the
entire structure of the short-term movements in Figure 13.5 and t the initial term-structure
of market prices. We implement the rst of these two strategies in the next subsection. We
develop the second strategy in Section 13.4.
13.4.2.3 Implementing implied binomial trees
We now build up the implied binomial tree in the general case, i.e. when we have several bond
prices to match. Suppose the time interval is six months, so that the short-term rate is for six
months. The current short-term rate is 3.99%, annualized. It can change to either 4.50% or to
4.00%, with equal (physical) probability. Suppose that two zeros are available for trading: a 6M
zero and a 1Y zero, where the current price of the 1Y zero is 0.95974. What is the risk-neutral
probability implied by this tree? This probability must be such that, the price of all the zeros
are matched exactly.
The tree we face is depicted below.
p=
r =
1
2
r =
4 .5 0 %
2
3 .9 9 %
2
r =
t=0
4 .0 0 %
2
t = 0 .5
FIGURE 13.6. The dynamics of the short-term rate: high interest rate scenario
In this tree, p = 12 denotes the physical

probability. Naturally, the price of a 6M zero at

t = 0, equals, P$ (0, 0.5) = 1/ 1 + 0.0399
=
0.9804.
This price is actually observed. That is, the
2
438
c
by
A. Mele
current short-term rate, 3.99%, is a mere denition. Next, we proceed to nd the no-arbitrage
movements of the 1Y zero, which are displayed below.
1
p = 12
r=
3.99 %
2
r=
4.50 %
2
P (0.5,1) = 1 / (1 +
0 .045
2
) = 0.9779
P$ (0,1) = 0.95974
r=
4 .00 %
2
P (0.5,1) = 1 / (1 +
0 .040
2
) = 0.9804
1
t = 0.5
t =0
t =1
Note, the current market price, P$ (0, 1) = 0.95974, is less than the expected price to prevail
tomorrow, discounted at the current interest rate,

1
1
1
1
Ep [P (0.5, 1)] =
0.9779 + 0.9804 = 0.9599.
1+r
2
2
1 + 0.0399
2
Hence, p = 12 cannot be the risk-neutral probability. To nd out the risk-neutral probability,
we proceed as follows. In the absence of arbitrage opportunities,
P$ (0, 1) = 0.95974
1
=
[qPup (0.5, 1) + (1 q) Pdown (0.5, 1)]
1+r
1
=
[q 0.9779 + (1 q) 0.9804]
1 + 0.0399
2
with obvious notation. This is one equation with one unknown, q. The solution for q is, q = 0.605.
We may now proceed with pricing derivatives. Consider a call option on the 1Y zero, with
expiration date in six months and exercise price equal to 0.9785. Its payo is as depicted below:
1
q = 0 .6 0 5
r =
P (0 . 5 ,1 ) = 0 . 9 7 7 9
C = m a x {P (0 . 5 ,1 ) K , 0 } = 0
3 .9 9 %
2
C = ?
1
P (0 . 5 ,1 ) = 0 . 9 8 0 4
C = m a x {P (0 . 5 ,1 ) K , 0 } = 0 . 0 0 1 9
t=0
t = 0 .5
439
t =1
c
by
A. Mele
So the option price is, by risk-neutral evaluation,

C=
1
1+
0.0399
2
[q 0 + (1 q) 0.0019] = 0.9804 [0.395 0.0019] = 7.3579 104 .
(13.12)
What happens when the short-term rate does not evolve as in the diagram of Figure 13.6
but, instead, as in Figure 13.7?
r =
r=
4 .4 1 54 %
2
3.99 %
2
r =
t =0
4 .0 0 %
2
t = 0.5
FIGURE 13.7. The dynamics of the short-term rate: low interest rate scenario
The previous tree is one in which the short-term in the upper state of the world equal to
r = 4.4154%, not 4.50%, as in Figure 13.6. It implies that:
Pup (0.5, 1) =
1
1+
r
2
1
1+
4.4154%
2
= 0.9784.
Then, the risk-neutral probability, q, solves the following pricing equation,

P$ (0, 1) = 0.95974
1
=
[qPup (0.5, 1) + (1 q) Pdown (0.5, 1)]
1+r
1
[q 0.9784 + (1 q) 0.9804] .
=
1 + 0.0399
2
The solution is, q = 0.756, which is larger than the solution we found earlier using the tree in
Figure 13.6 (i.e., q = 0.605). The option price is, now,
C=
1
1+
0.0399
2
[q 0 + (1 q) 0.0019] = 0.9804 [0.244 0.0019] = 4.5451 104 .
Why is this price smaller than that computed in Eq. (13.12)? In the tree of Figure 13.7, the
up-state of the world is, so to speak, less severe than the up-state of the world in the tree of
Figure 13.6. To be able to match the initial price P$ (0, 1) = 0.95974, the model in Figure 13.7
must put more weight on the up-state of the world, i.e. a larger implied risk-neutral probability.
This implies a larger risk-neutral probability that low bond prices will arise in the future and,
hence, a lower option price.
In a segmented market, two investment banks might have dierent views about the evolution
of the short-term rate (the view in Figure 13.6 and the view in Figure 13.7). The rst bank
440
c
by
A. Mele
favours a high interest rate scenario, but it is not too risk-averse to that scenario (rup = 4.5%,
q = 0.605). The second bank favours a mild interest rate scenario, but it is more too riskaverse to that scenario (rup = 4.4154%, q = 0.9784). But then, naturally, both institutions
need to agree on the initial bond price, P$ (0, 1) = 0.95974. The segmentation could arise, for
example, because the clientèle of the rst bank and that of the second bank are unlikely to
meet and, the prices charged by the banks are not publicly known. In the absence of market
imperfections (and arbitrage), however, the investment banks should agree on the option price
too.
Next, let us another period to the diagram in Figure 13.6, assuming that the short-term rate
is as in the following diagram:
q1 = ?
q 0 = 0 .6 0 5
r =
r =
t =0
4 .9 0 %
2
4 .5 0 %
2
3 .9 9 %
2
r =
r =
r =
4 .3 0 %
2
r =
3 .9 0 %
2
4 .0 0 %
2
t = 0 .5
t =1
FIGURE 13.8.
In this tree, q0 is the risk-neutral probability for the rst period, and q1 is the risk-neutral
probability for the second period.
We already know that q0 = 0.605. The probability q1 is the risk-neutral probability for the
time-period (0.5, 1), and can be dierent from q0 . Suppose, also, that an additional zero is
available for trading, a 1.5Y zero. The current price of the 1.5Y zero is P$ (0, 1.5) = 0.9382.
To derive the the risk-neutral probability q1 , we proceed as follows. First, we consider the tree
441
c
by
A. Mele

below.
1
r=
q1 = ?
r=
q0 = 0.605
r=
P (1,1.5) = 1 / (1 + 0.049
) = 0.9761
2
4.50 %
2
PU (0.5,1.5) = ?
3.99 %
2
P$ (0,1.5) = 0.9382
4. 90 %
2
q1 = ?
r=
4.00 %
2
r=
4. 30 %
2
) = 0.9789
P (1,1.5) = 1 / (1 + 0.043
2
PD (0.5,1.5) = ?
r=
3.90 %
2
) = 0.9808
P (1,1.5) = 1 / (1 + 0.039
2
1
t =0
t = 0.5
t =1
t =1.5
We need to compute the prices PU (0.5, 1.5) and PD (0.5, 1.5). Once we compute these prices,
we shall use the no-arbitrage property of the zero, and the previously computed q0 = 0.605, to
recover q1 . By the usual no-arbitrage property of the zero, we have that:
PU (0.5, 1.5) =
PD (0.5, 1.5) =
1
1+
0.045
2
1
1+
0.040
2
[q1 0.9761 + (1 q1 ) 0.9789]
(13.13)
[q1 0.9789 + (1 q1 ) 0.9808]
(13.14)
The problem, q1 is not known. Therefore, Eqs. (13.13)-(13.14) do not allow us to pin down
the prices PU (0.5, 1.5) and PD (0.5, 1.5). But here is where calibration comes in! We know the
current price of the 1.5Y zero, which is, P$ (0, 1.5) = 0.9382. In the absence of arbitrage,
P$ (0, 1.5) = 0.9382 =
1
1+
0.0399
2
[q0 PU (0.5, 1.5) + (1 q0 ) PD (0.5, 1.5)] ,
where PU (0.5, 1.5) and PD (0.5, 1.5) are as in Eqs. (13.13)-(13.14), and where q0 = 0.605. So we
have,
1
0.9382 =
[0.605 PU (0.5, 1.5) + 0.395 PD (0.5, 1.5)] ,
(13.15)
1 + 0.0399
2
where PU (0.5, 1.5) and PD (0.5, 1.5) are as in Eqs. (13.13)-(13.14). Hence, by replacing Eqs.
(13.13)-(13.14) into Eq. (13.15) leaves one equation with exactly one unknown, q1 . Solving,
yields, q1 = 0.8412, which implies that,
PU (0.5, 1.5) = 0.9549, PD (0.5, 1.5) = 0.9600.
442
c
by
A. Mele

So, to sum up, we have the tree below.
1
q1 = 0.8418
r=
4 .90 %
2
P (1,1 .5 ) = 0 .9761
r = 4.502 %
q0 = 0.605 P (0 .5,1 .5 ) = 0 .9549
U
r=
3 . 99 %
2
1
r=
P$ (0,1 .5 ) = 0 .9382
r=
4 . 30 %
2
P (1,1 .5 ) = 0 .9789
4.00 %
2
PD (0 . 5,1 . 5 ) = 0 .9600
1
r=
3 .90 %
2
P (1,1 .5 ) = 0 . 9808
t = 0.5
t =0
t =1
t = 1.5
We are now ready to compute the no-arbitrage price of a call option on the 1.5Y zero, with
expiration date in 1Y and exercise price equal to 0.9800. The price of the option at time t = 0.5,
is C = 0.00012, as illustrated below.
P (1 ,1 . 5 ) = 0 . 9 7 6 1
C = m a x {P (1 ,1 . 5 ) K , 0 } = 0
q1 = 0 .8 4 1 8
C = 0
P (1 ,1 . 5 ) = 0 . 9 7 8 9
C = [q 1 0 + (1 q 1 ) 0 . 0 0 0 8 ] / (1 +
0 .0 4
2
C = m a x {P (1 ,1 . 5 ) K , 0 } = 0
= 0 .0 0 0 1 2
P (1 ,1 . 5 ) = 0 . 9 8 0 8
C = m a x {P (1 ,1 . 5 ) K , 0 } = 0 . 0 0 0 8
t = 0 .5
t =1
We can now calculate the no-arbitrage price of the 1Y call option on the 1.5Y zero, struck at
K = 0.9800. It is,
C=
1
1+
0.0399
2
[0 q0 + 0.00012 (1 q0 )] = 0.9804 [0.00012 (1 0.605)] = 4.647 105 .
We can use Figure 13.8 to price derivatives, such as, say, a call option on the 1.5Y zero, with
expiration date in six months, and exercise price equal to 0.9580. We have the following tree.
443
c
by
A. Mele
q 0 = 0 .6 0 5
r=
P U (0 . 5 ,1 . 5 ) = 0 . 9 5 4 9
C = m ax{P U (0 . 5 ,1 . 5 ) K , 0 } = 0
3.9 9 %
2
C =?
PD (0 . 5 ,1 . 5 ) = 0 . 9 6 0 0
C = m ax{PD (0 . 5 ,1 . 5 ) K , 0 } = 0 . 0 0 2 0
t =0
t = 0.5
Therefore, the no-arbitrage price of the option is,
C=
1
1+
0.039
2
[q0 0 + (1 q0 ) 0.0020] = 0.9804 [0.395 0.0020] = 7.745 104 .
13.4.2.4 Summing up
So lets sum up what weve done. Given is the evolution of the short-term rate in Figure 13.8,
which we use to recover the two risk-neutral probabilities q0 (for the time span (0, 0.5)) and q1
(for the time span (0.5, 1)), starting from the knowledge of the market prices of two zeros, the
1Y zero and the 1.5Y zero. Precisely, given P$ (0, 1), the price of the 1Y zero, we recover q0 , as
illustrated below:
1
q0
PU (0 . 5 ,1 )
P$ (0 ,1 )
1
P D (0 . 5 ,1 )
t=0
t = 0 .5
t =1
This is possible as PU (0.5, 1) and PD (0.5, 1) do not depend on q0 and so they are obtained
in a straightforward manner. Given q0 , then, we compute q1 , using P$ (0, 1.5), the price of the
444
c
by
A. Mele

1.5Y zero, as illustrated below:
1
q1
q 0
PU U (1 ,1 . 5 )
PU (0 . 5 ,1 . 5 )
1
PU D (1 ,1 . 5 )
P$ (0 ,1 . 5 )
PD (0 . 5 ,1 . 5 )
1
PD D (1 ,1 . 5 )
t=0
t = 0 .5
t =1
t = 1 .5
Again, the risk-neutral probability, q1 , can be recovered because PU U (1, 1.5), PU D (1, 1.5) and
PDD (1, 1.5) do not depend on q1 , and are easily obtained. So, given PU U (1, 1.5), PUD (1, 1.5)
and PDD (1, 1.5), we can express PU (0.5, 1.5) and PD (0.5, 1.5) as two (linear) functions of q1 .
Finally, we impose the no-arbitrage property to P$ (0, 1.5), which makes the observed price,
P$ (0, 1.5), a (linear) function of PU (0.5, 1.5) and PD (0.5, 1.5) and, hence, q1 , thereby allowing
us to recover q1 .
We can continue, and consider an additional time period, as in the tree in Figure 13.9 below.
We can recover q2 , once we are given the market price of a 2Y zero, P$ (0, 2), as follows:
The prices of the 2Y zero at time t = 1.5 (the lled nodes in Figure 13.9) (say P (1.5, 2))
are easily computed, given an assumption about the numerical values of the short-term
rate in those nodes.
Then, given the prices P (1.5, 2) at time t = 1.5, and the previously calibrated probabilities
q0 and q1 , we can express the current market price P$ (0, 2) as a (linear) function of q2 .
Then, we solve for q2 .
1
q2
q 1
q 0
P $ (0 , 2
1
t =0
t =1
t = 0 .5
445
t = 1 .5
t = 2
c
by
A. Mele
13.5. The Ho and Lee model

FIGURE 13.9.
The calibration can continue. We extend the tree to one period more. Then, we use the
price of one additional zero to recover time varying risk-neutral probabilities. An alternative
procedure consists in: (i) fixing the risk-neutral probabilities q to some value at all times (e.g.,
q = 12 ), and (ii) guring out the implied values for the short-term rate in each node of the
tree. The next section develops a systematic approach for implementing this procedure.
13.5 The Ho and Lee model

Ho and Lee (1986) introduced a revolutionary approach to modeling yield curve movements.
Their model is not about an economic theory of determination of the yield curve. Rather,
their approach is to take the yield curve as given, and then to model the movements of the
entire yield curve in order to price interest rate derivatives. As explained in Chapter 11 and in
the previous section, we need to match prices, to avoid having derivatives with underlyings
deviating from market prices. In Chapter 11, we discussed the Ho and Lee model in continuous
time, as this allowed to illustrate the general methodology underlying the HJM approach. The
original formulation of the model was, however, in discrete time. In this section, we present the
discrete time version of the model and some of its extensions, as well as the general philosophy
underlying matching the initial yield curve within a discrete time framework, which represents
indeed the industry practice.
The main idea underlying the Ho and Lee model is to model the movements of the yield
curve along a binomial tree, much in the spirit of the Cox, Ross and Rubenstein (1979) tree
representation of the Black and Scholes (1973) model. The main issues can be summarized as
follows. In Black and Scholes (1973) and Cox, Ross and Rubenstein (1979), the asset underlying
the option contract is a traded risk. So the underlying asset price satises the martingale
condition. Interest rate derivatives, instead, generally depend on non-traded risks. The mere
presence of boundary conditions induce bond return volatility to be time-varying.
13.5.1
The tree
The price of any zero evolves randomly over time, according to a binomial tree. Let Pj (t, T )
be the price of a pure discount bond as of time t, with time to maturity T t, after j upstate
price movements. Let j B (t, q), a binomial random variable,
V ar (j) = tq (1 q) ,
E (j) = tq,
where q is the risk-neutral probability of a single upstate movement. Therefore we have,

Pj+1 (t + 1, T )
q
Pj (t, T )
1q
Pj (t + 1, T )
That is, if at time t, the number of upstate movements is equal to j then, at time t + 1, the
number of upstate movements can either jump to j + 1, with probability q, or stay at j, with
446
c
by
A. Mele
probability 1 q. Note also that after one period, the price of any zero is one period closer to
maturity. At maturity, t = T , the price of any zero is worth one unit of numeraire, viz
Pj (T, T ) = 1,
for all j and T.
Note, in the previous tree, it shall not necessarily hold that Pj (t + 1, T ) < Pj (t, T ). On
the contrary, we would expect that especially when the maturity approaches, Pj (t + 1, T ) >
Pj (t, T ), as the price of the zero needs to converge to par.
13.5.2
The price movements and the martingale restriction
In the absence of arbitrage opportunities, the expected return on the zero at t must equal the
short-term rate, viz Pj (t, T ) = erj (t) Eq (P (t + 1, T )), or
Pj (t, T ) = Pj (t, t + 1) [qPj+1 (t + 1, T ) + (1 q) Pj (t + 1, T )] ,
(13.16)
where Pj (t, t + 1) = erj (t) , and rj (t) is the continuously compounded short-term rate at time
t after j upward movements. We call this condition the martingale restriction.
Let us introduce notation for the movements of the price of any zero along the tree,
Pj+1 (t + 1, T )
1
Pj (t + 1, T )
1
= u (T t)
and
= d (T t)
. (13.17)
Pj (t, T )
Pj (t, t + 1)
Pj (t, T )
Pj (t, t + 1)

up at t
down at t
The two functions u () and d (), also called perturbation functions, capture the fact that
in the case of uncertainty, the price of the zero can either go up or down with respect to the
risk-free of return. In other words, Eqs. (13.17) tell us that the discounted gross return from
going long a bond is:
u (T t) with probability q
P (t + 1, T )
P (t, t + 1) =

P (t, T )
d (T t) with probability 1 q

Discount
Gross return
where the two functions u (T t) and d (T t) have to be determined endogeneously. If there

was no uncertainty, we would have u (T t) = d (T t) = 1, for all t T . In general, we have
that d (T t) 1 u (T t), as we shall now demonstrate.
One period before the expiration date, i.e. at t = T 1, our price is certain to jump to one,
with jump size equal to the short-term rate rj (t). Hence, the following boundary condition for
the two functions u () and d () holds:
u (1) = d (1) = 1.
(13.18)
In terms of the two functions u () and d (), the martingale restriction in Eq. (13.16) is,
1 = qu (T t) + (1 q) d (T t) ,
t T.
(13.19)
This relation is quite familiar as it matches the standard risk-neutral relation for stock prices
in which the short-term rate is tied down to the up and down movements of the stock price.
However, in this context the up and down movements of the zero price depend on the maturity
of the price itself through the two functions u (T t) and d (T t), which makes the evaluation
problem more intricate.
447
c
by
A. Mele

13.5.3
The recombining condition
Ho and Lee consider a recombining tree: the price Pj (t, T ) we are looking for depends only
on j, not on the exact sequence of up and down movements leading to j upstate movements.
To summarize, we are looking for two functions u (T t) and d (T t) such that (i) the noarbitrage condition in Eq. (13.19) holds true and (ii) the tree is recombining. We now elaborate
the arguments that lead to the recombining property of the tree.
Pj+2 (t + 2, T )
Pj+1 (t + 1, T )
Pj (t, T )
Pj+1 (t + 2, T )
Pj (t + 1, T )
Pj (t + 2, T )
The recombining property of the tree implies that the bond price at time t + 2 in the event
of j + 1 jumps, i.e. Pj+1 (t + 2, T ), can be generated by one of the two paths:
(i) The path Pj (t, T ) Pj+1 (t + 1, T ) Pj+1 (t + 2, T )
up & down
(ii) The path Pj (t, T ) Pj (t + 1, T ) Pj+1 (t + 2, T )
down & up
We can use the two relations in Eqs. (13.17), to gure out the two paths leading to the bond
price at time t + 2 in the event of j + 1 jumps, i.e. Pj+1 (t + 2, T ). We have that along the rst
path,
Pj+1 (t + 1, T )
1
= u (T t)
,
Pj (t, T )
Pj (t, t + 1)

up at t
and along the second path,
Pj+1 (t + 2, T )
1
= d (T t 1)
,
Pj+1 (t + 1, T )
Pj+1 (t + 1, t + 2)

down at t+1
Pj (t + 1, T )
1
Pj+1 (t + 2, T )
1
= d (T t)
,
= u (T t 1)
.
Pj (t, T )
Pj (t, t + 1)
Pj (t + 1, T )
Pj (t + 1, t + 2)

up at t+1
down at t
To sum up:
Pj+1 (t+1,T )

1
1
Pj+1 (t + 2, T ) = d (T t 1)
u (T t)
Pj (t, T ) (up & down)
Pj+1 (t + 1, t + 2)
Pj (t, t + 1)
1
1
Pj+1 (t + 2, T ) = u (T t 1)
d (T t)
Pj (t, T )
(down & up)
Pj (t + 1, t + 2)
Pj (t, t + 1)

Pj (t+1,T )
By equating the previous two equations, we obtain,
u (T t 1) Pj+1 (t + 1, t + 2)
u (T t)
=
d (T t)
d (T t 1) Pj (t + 1, t + 2)
448
(13.20)
c
by
A. Mele

By evaluating Eq. (13.20) at T = t + 2,
u (2)
u (1) Pj+1 (t + 1, t + 2)
Pj+1 (t + 1, t + 2)
=
=
1 ,
d (2)
d (1) Pj (t + 1, t + 2)
Pj (t + 1, t + 2)
where we assume that is constant. Clearly, 0 1. Substituting back into Eq. (13.20),
u (T t)
u (T t 1) 1
=
.
d (T t)
d (T t 1)
Therefore, given that u (1) = d (1) = 1,
u (T t)
= (T t1) .
d (T t)
(13.21)
Eq. (13.21) gives us the condition under which the tree is recombining. To rule out arbitrage
opportunities, the martingale restriction in Eq. (13.19) must also hold true. Therefore, we have
to solve the following system of two equations (Eq. (13.21) and Eq. (13.19)) with two unknowns
(u () and d ()),
+
u (T t) = (T t1) d (T t)
qu (T t) + (1 q) d (T t) = 1
The solution to this system is,
u (T t) =
1
,
q + (1 q) T t1
d (T t) =
T t1
.
q + (1 q) T t1
(13.22)
So we have solved the problem. We know how to populate the tree. Suppose we know how
to assign values to q and . Given q and , and an initial bond price P (t, T ), we can use Eqs.
(13.17) to populate the tree, using the solution for u (T t) and d (T t) given in Eqs. (13.22).
In this way, we can gure out the exact bond prices to insert in each node of the tree. Once
we have computed the bond prices in each node, we can price interest rate derivatives, i.e. the
asset the payo of which depend on the particular value taken by the bond price on a given set
of nodes. Below, we provide the closed-form solution for the bond price in this model.
(t+1,t+2)
What is the interpretation of ? We have dened to be, 1 PPj+1
, or,
j (t+1,t+2)
ln
Pj+1 (t + 1, t + 2)
= ln
Pj (t + 1, t + 2)
= [rj+1 (t + 1) rj (t + 1)] .
(13.23)
But we know that conditionally upon time t and (price) jumps equal to j t, the short-term
rate is binomially distributed, and can take on two values: (i) rj+1 (t + 1) with probability q
and rj (t + 1) with probability 1 q. Then, the conditional variance of the short-term rate is,
vart [
r (t + 1)] = q (1 q) [rj+1 (t + 1) rj (t + 1)]2 ,
where vart [
r (t + 1)] is the conditional variance at time t, of the short-term rate one-period
ahead. Then, we may use Eq. (13.23), and the previous equation, to obtain,

vart [
r (t + 1)] = q (1 q) ln 1 .
449
c
by
A. Mele
That is, is a parameter related to the volatility of the short-term rate, which in this basic
model, is constant. In general, could be time-varying, although it is then dicult to nd
closed-form solutions for the model.
The Appendix shows that the solution to the Ho and Lee model (i.e. with xed ), is:
T 1
P (0, T ) (T t)(tj) 1 q + (1 q) St
Pj (t, T ) =
.
S
P (0, t)
q
+
(1
q)
S=t
(13.24)
From the perspective of time 0, the price of the zero at t, is only a function of the initial yield
curve, the volatility parameter , and of course the risk-neutral probability q.
13.5.4
Calibration of the model
We need to estimate the value of . We can proceed as follows. Consider Eq. (13.24), and let
T = t + 1. We have,
1
P (0, t + 1) tj
Pj (t, t + 1) =
.
P (0, t)
q + (1 q) t
The continuously compounded short-term rate predicted by the model is,

rj (t) ln Pj (t, t + 1) = Ft (0) + ln q + (1 q) t (t j) ln , j t,
(13.25)
where Ft (0) ln P (0, t) ln P (0, t + 1). We also have,
rj (1) r (0) = F1 (0) F0 (0) + ln (q + (1 q) ) + ln 1 (1 j) .

Hence, the parameter can be chosen so that the volatility of the short-term rate predicted by
the model matches exactly the volatility
of the short-term rate that we see in the data. Concretely, we can take = exp( Std (r)/ q (1 q)), where Std(r) is the standard deviation
of the short-term rate in the data.
Note, then, the interesting feature of the model. The Ho and Lee model doesnt take any
a priori stance on the dynamics of the short-term rate. Rather, it imposes: (i) the martingale
restriction on bond prices, an economic restriction, Eq. (13.19); and (ii) the simplifying assumption the tree is recombining, a technical condition, Eq. (13.17). These two conditions suce to
to tell what to expect from the dynamics of the short-term rate. While deliberately simple, the
Ho and Lee model is quite powerful. The modern approach to interest rate modeling simply
aims to make the Ho and Lee methodology more accurate for practical purposes.
13.5.5
An example
Assume that three zero coupon bonds are available for trading, with current market prices: (i)
P$ (0, 1) = 0.9851 (the price of a 6M zero), (ii) P$ (0, 2) = 0.9685 (the price of a 1Y zero), and
(iii) P$ (0, 3) = 0.9445 (the price of the 1.5Y zero). We know that the price of one-period zero
at time t, in the event of j upward price-jumps from the current date to t, is:
Pj (t, t + 1) =
P$ (0, t + 1) tj
1
,
P$ (0, t)
q + (1 q) t
j t,
(13.26)
where P$ (0, t) is the current market price of a zero expiring at time t, with t equal to six
months, one year and eighteen months, in this example. We assume that q = 12 and = 0.9802.
450
c
by
A. Mele

13.5.5.1 The dynamics of the short-term rate
We want to determine the evolution of the short-term rate on a recombining tree for as many
periods as we can, given the market price of the zeros we observe. We use Eq. (13.26) to nd
the one-period zeros in each node.
t = 0. We have, trivially, P (0, 1) = P$ (0, 1) = 0.9851.
t = 1. We have three cases:
j = 0: P0 (1, 2) = 2 PP$$ (0,2)
1 = 0.9733
(0,1) 1+
1
j = 1: P1 (1, 2) = 2 PP$$ (0,2)
= 0.9930
(0,1) 1+
t = 2. We have two cases:

1
j = 0: P0 (2, 3) = 2 PP$$ (0,3)
2 1+
2 = 0.9557
(0,2)
1 = 0.9750
j = 1: P1 (2, 3) = 2 PP$$ (0,3)
(0,2) 1+ 2
1
j = 2: P2 (2, 3) = 2 PP$$ (0,3)
= 0.9947
(0,2) 1+2
So we face the tree below.

P (2 ,3 ) = 0 . 9 5 5 7
q=
1
2
P (1, 2 ) = 0 . 9 7 3 3
q=
1
2
P (2 ,3 ) = 0 . 9 7 5 0
P (0 ,1 ) = 0 . 9 8 5 1
P (1, 2 ) = 0 . 9 9 3 0
P (2 , 3 ) = 0 . 9 9 4 7
t =0
t =1
t=2
13.5.5.2 Pricing a coupon bearing bond
Suppose, now, that we want to nd the price of some additional bond, e.g., a 1.5Y bond which
pays (semiannually) coupons at 3% of the principal of $1. First, we need to nd the value of
this bond in each node of the tree. Note, at each node, we have to gure out (i) the discounted
expectation of its future value (including coupons), and (ii) the current coupons, as illustrated
in the tree below. That is, the convention, here, is that the bond purchased at time t doesnt
give the owner the right to receive any coupon at time t, only from time t + 1 onwards.
451
c
by
A. Mele
1 .0 3
PU U (2 , 3 ) = 0 .9 5 5 7
0 .0 3 + PU U (2 ,3 ) 1 . 0 3 = 1 .0 1 4
q=
1
2
P (1, 2 ) = 0 . 9 7 3 3
0 .0 3 + P (1, 2 )( 12 1 .0 1 4 + 12 1 . 0 3 4) = 1 . 0 2 6 7
q = 12
1 .0 3
PU D (2 ,3 ) = 0 .9 7 5 0
0 .0 3 + PU D (2 ,3 ) 1 . 0 3 = 1 . 0 3 4
P (0 ,1 ) = 0 . 9 8 5 1
P (1,2 ) = 0 . 9 9 3 0
0 .0 3 + P (1,2 )( 12 1 .0 3 4 + 12 1 . 0 5 4) = 1 .0 6 6 7
1 .0 3
PD D (2 ,3 ) = 0 .9 9 4 7
0 .0 3 + PD D (2 ,3 ) 1 . 0 3 = 1 . 0 5 4
1 .0 3
t =1
t =0
t =2
t =3
Naturally, the bond does not pay coupons at time zero. Therefore, the current price is,
1
1
P = P (0, 1)
1.0267 + 1.0667
2
2
1
1
= 0.9851
1.0267 + 1.0667
2
2
= 1.0311.
Note, naturally, the answer to this question could have been obtained by just adding [P$ (0, 1) + P$ (0, 2) +
0.03 + P$ (0, 3), although the results in the tree above are going to matter while pricing derivatives written on the coupon bearing bond.
13.5.5.3 Pricing European options
Next, we wish to nd the price of options, say the price of two call options on the 1.5Y bond
considered in the previous subsection, when the strike price is $1 and the maturities of the
options are 6 months and 1 year. Again, we need to gure out the no-arbitrage movements of
the ex-coupons bond price. (This is because if we purchase the bond today, we are not entitled
to receive any coupon, today. The ow of coupons we are entitled to receive starts from the
next period.) We easily obtain the tree below. We must just subtract the coupon, 0.03, from
452
c
by
A. Mele

each cum-coupons price in each node of the tree. Then, we obtain:
P = 1.01 4 0 .0 3 = 0 .98 4
q=
1
2
P = 1.0267 0 .03 = 0.99 7
q = 12
P = 1.034 0 .0 3 = 1 .004
P (0,1) = 0 .98 51
P = 1.06 67 0.03 = 1 .0 367
P = 1.05 4 0 .0 3 = 1.0 24
t =0
t =1
t =2
We are ready to price the two options. As for the call option on the 1.5Y bond, with 6 months
maturity, and strike price K = $1, we have the following tree:
P = 0.997
q = 12
C = max{P K ,0} = 0
P(0,1) = 0.9851
C =?
P = 1.0367
C = max{P K ,0} = 0.0367
t =0
t =1
Therefore,
C = 0.9851

1
1
0 + 0.0367 = 1.808 102 .
2
2
453
c
by
A. Mele
13.6. Beyond Ho and Lee: Calibration
The call option on the 1.5Y bond with 1 year maturity, and strike price K = $1, is dealt
with similarly. We have the following tree:
q = 12
P = 0.984
C = m ax{P K ,0} = 0
P(1, 2 ) = 0.9733
C = P (1,2 )( 12 0 + 12 0.004) = 0 .0019

P (0,1) = 0.9851
q = 12
P = 1.004
C = m ax{P K ,0} = 0.004
C=?
P (1,2 ) = 0.9930
C = P (1,2 )( 0.004 + 12 0.024) = 0.014

1
2
P = 1.024
C = m ax{P K ,0} = 0.02 4
t =0
t =1
t =2
Therefore, the price of the option is,

1
1
0.0019 + 0.014
C = P (0, 1)
2
2
= 0.9851
1
1
0.0019 + 0.014
2
2
= 7.831 103 .
13.6 Beyond Ho and Lee: Calibration

The modeling approach in the previous sections imposes no-arbitrage conditions on the price
of the zeros, thereby determining the implied stochastic process for the short-term rate. In this
section, we show how to implement this approach by looking for, or fitting, the right short-term
rate process in the rst place. Practitioners might prefer to view the Ho and Lee model by
the same calibration perspective we develop in this section. To illustrate how the calibration
works, we develop three points. First, we review how Arrow-Debreu securities can be put at
work in the very applied context of xed income security evaluation. We shall see Arrow-Debreu
securities are conceptually very useful here, as they allow us to turn the martingale restriction
of the previous sections to a set of analytically simpler conditions. Second, we use these ArrowDebreu securities to implement a general algorithm to populate the short-term rate tree,
while ensuring that the initial term-structure is perfectly tted. Finally, we apply the previous
algorithm to illustrate how to solve two models, in practice: (i) the Ho and Lee model, and (ii)
the model developed by Black, Derman and Toy (1990).
13.6.1
Arrow-Debreu securities
We know from Chapter 2, that an Arrow-Debreu security is a security that promises to pay $1
in some prespecied state of the nature, and zero otherwise. Consider, for example, the diagram
454
c
by
A. Mele

below.
0
q
1 q
s,
0,0
s 1,
Arrow-Debreu security
s, + 1
q
1 q
FIGURE 13.10. In the binomial tree of this section, an Arrow-Debreu security for state s
at time + 1 is a security that pays $1 at time + 1 in state s, and zero otherwise. This
section aims to show how to recover Arrow-Debreu prices from the price of xed income
securities.
In this diagram, q is the risk-neutral probability of an upward movement of the short-term

rate. A generic pair (s, ) at each node tracks the number of upward movements of the shortterm rate, s, and calendar time, , where of course s (since there can only be one possible
short-term rate movement in each period). From now on, let us focus attention on the ArrowDebreu security for the state s at time + 1.
Let ps ( ) denote the current price of an Arrow-Debreu security that pays o $1 in state s
at time , and zero otherwise. Then, the current market price of a zero that matures at time T
is necessarily,
T

P$ (0, T ) =
ps (T ) .
s=0
More generally, consider a derivative that pays o Ds ( ) in node (s, ), meaning a dividend
equal to D1 ( ) in state s = 1, equal to D2 ( ) in state s = 2, , and equal to D ( ) in state
s = . The price of this asset, denoted as C$ (0, T ), is given by,
C$ (0, T ) =
T
ps ( ) Ds ( ) .
(13.27)
=1 s=0
Our objective, now, is to recover the price of the Arrow-Debreu securities ps ( ) for all s
and , where {1, , T }, from the observation of the initial term-structure of interest rates.
Consider the Arrow-Debreu security that promises to pay $1 in node (s, + 1) (see Figure
13.7). Let its value at time in state j (j ) be denoted as j, [s, + 1]. What is this
value at time in all states? The key observation, here, is that in this binomial tree, the node
455
c
by
A. Mele
(s, + 1) (the lled circle) can only be accessed to through the nodes (s, ) and the nodes
(s 1, ) occurring at time (the two empty circles in Figure 13.7). For this reason, at time
, the value j, [s, + 1] is zero in all the nodes (j, ) that are distinct from the empty circles
(s, ) and (s 1, ). This is because starting from any node dierent from these empty circles,
it is impossible to reach the node (s, + 1) (the lled circle) where the Arrow-Debreu security
pays o.
So, we are left with nding the values j, [s, + 1] in the nodes corresponding to the empty
circles (s, ) and (s 1, ), i.e. s, [s, + 1] and s1, [s, + 1]. Let rs ( ) be the continuously
compounded short-term rate in node (s, ). Consider the upper node (s, ). We have,
s, [s, + 1] = ers ( ) [0 q + 1 (1 q)] = ers ( ) (1 q) .
Similarly, in the lower node, (s 1, ),
s1, [s, + 1] = ers1 ( ) [1 q + 0 (1 q)] = ers1 ( ) q.
We can think of our Arrow-Debreu security for (s, + 1) as a derivative that at time ,
delivers the following payos
s, [s, + 1] = ers ( ) (1 q)
(13.28)
[s, + 1] = ers1 ( ) q
s1,
j, [s, + 1] = 0, for all j < s
These payos are simply the market value of the Arrow-Debreu security for (s, + 1), in the
various states occurring at time , i.e. the money the holder can make by selling the asset at
time , in the various states. Therefore, we can apply Eq. (13.27) to obtain,
ps ( + 1) =
pj ( ) j, [s, + 1]
j=0
= ps ( ) s, [s, + 1] + ps1 ( ) s1, [s, + 1] .

By replacing the Arrow-Debreu prices in (13.28) into the previous equation, we obtain the
so-called forward equation for the Arrow-Debreu prices,
ps ( + 1) = ps ( ) ers ( ) (1 q) + ps1 ( ) ers1 ( ) q
13.6.2
(13.29)
The algorithm in two examples
The algorithm aims to populate the interest rate tree by making a repeated use of the forward
equation (13.29) and the zero pricing equation
P$ (0, + 1) =
ps ( ) ers ( ) .
s=0
The input is, of course, a number of zeros equal to the largest maturity date the tree extends
to. We describe how the algorithm works by developing two concrete examples.
We start with Ho and Lee. We assume continuous compounding, for analytical reasons claried below. We assume that the continuously compounding short-term rate is solution to,
rs ( ) = r0 ( ) + ln 1 s,
456
(13.30)
c
by
A. Mele
where rs ( ) is the short-term rate at time , in the event of s upward movements of the
short-term rate, and is a volatility parameter, i.e. such that
Std (r)
ln 1 =
,
q (1 q)
where Std (r) is the standard deviation of the short-term rate in the data.4 At time zero, the
price of a zero maturing at time + 1 is:

P$ (0, + 1) =
ps ( ) ers ( ) = er0 ( )
s ps ( ) ,
s=0
s=0
where the second equality follows by the assumption that the short-term rate is solution to Eq.
(13.30).
By rearranging terms in the previous equation, we obtain a closed-form expression for the
future short-term rate at time , in the event of zero upward movements,

s
s=0 ps ( )
r0 ( ) = ln
.
(13.31)
P$ (0, + 1)
We use Eq. (13.31) and the forward equation (13.29) to populate the interest rate tree, under
the assumption that q = 12 . Precisely, the algorithm proceeds as follows:
(i) Given the boundary condition for the Arrow-Debreu price, p0 (0) = 1, compute the initial
value of the short-term rate, r0 (0), using Eq. (13.31), as r0 (0) = ln( 1/ P$ (0, 1)).
(ii) Suppose we know the future value of the short-term rate at time 1, in the event of
no upward movements, i.e. r0 ( 1). Then, given the value of r0 ( 1), and the price
of the Arrow-Debreu securities ps ( 1) for s 1, compute ps ( ) for s , through
the forward equation (13.29),
1
ps ( ) = ps ( 1) s er0 ( 1) (1 q) + ps1 ( 1) s1 er0 ( 1) q, q = ,
2
where the last equation follows by plugging Eq. (13.30) into Eq. (13.29).
(iii) Given the Arrow-Debreu prices ps ( ) for s , use Eq. (13.31) to compute the future
value of the short-term rate at time , in the event of no upward movements, i.e. r0 ( ).
(iv) If = T , stop. Otherwise, go to (ii).
As a second example, consider the Black, Derman and Toy (1990) model. In this model, the
short-term rate is solution to,
rs ( ) = s r0 ( ) ,
(13.32)
where is, once again, a volatility parameter.5 For computational convenience, this model
assumes that the short-term rate in Eq. (13.32) is discretely compounded. Accordingly, we
rewrite the forward equation (13.29) in terms of discretely compounded rates,
1
1
ps ( + 1) = ps ( )
(1 q) + ps1 ( )
q.
(13.33)
1 + rs ( )
1 + rs1 ( )
The algorithm is as follows:
4 Hence,
the short-term rate movements that we shall derive do depend on the value of the risk-neutral probability q that we
choose.
5 In its most general form, this model assumes that r ( ) = s r ( ), where is a volatility parameter that varies determiniss
0
tically over time. This more general formulation leads to more exibility, which is useful to t the term structure of volatility.
457
c
by
A. Mele

(i) Compute the initial value of the short-term rate, r0 (0), as the solution to,
P$ (0, 1) =
1
.
1 + r0 (0)
(ii) Suppose we know the future value of the short-term rate at time 1, in the event of
no upward movements, i.e. r0 ( 1). Then, given the value of r0 ( 1), and the price
of the Arrow-Debreu securities ps ( 1) for s 1, compute ps ( ) for s , through
the forward equation (13.33),
ps ( ) = ps ( 1)
1
s
1 + r0 ( 1)
(1 q) + ps1 ( 1)
1+
s1
1
1
q, q = ,
2
r0 ( 1)
where the last equation follows by plugging Eq. (13.32) into Eq. (13.33).
(iii) Given the boundary condition p0 (0) = 1, and the Arrow-Debreu prices, ps ( ) for s ,
use the pricing equation for the zero,
P$ (0, + 1) =

s=0
ps ( )
1
,
1 + r0 ( )
s
to solve, numerically, for the future value of the short-term rate at time , in the event
of no upward movements, i.e. r0 ( ). Note, we did not need this additional step for the
solution of the Ho and Lee model, as the short-term rate r0 ( ) is known in closed form
in the Ho and Lee model (see Eq. (13.31)).
(iv) If = T , stop. Otherwise, go to (ii).
13.6.2.1 A numerical example
Consider, again the Ho and Lee example in Section 13.5.5, where three zeros were traded: (i)
one zero maturing in 6 months, (ii) one zero maturing in 1 year, and (iii) one zero maturing in
1.5 years, with market prices P$ (0, 1) = 0.9851, P$ (0, 2) = 0.9685, P$ (0, 3) = 0.9445. The Ho
and Lee model assumes that,

rs ( ) = r0 ( ) + ln 1 s.
(13.34)
We now want to use this equation to nd the values of the short-term rate rs ( ) in each
node, under the assumption that q = 12 , and that the standard deviation of the short-term rate
is 0.014, annualized.
Remarks on notation. By Eq. (13.25), the short-term rate predicted by the Ho and Lee
model is:
(13.35)
rj ( ) = F (0) + ln (q + (1 q) ) ( j) ln .
where F (0) is the continuously compounded forward rate, at time zero, for maturity [ , + 1],
and j is the number of upward movements of the bond prices. Naturally, then, s (t j) is
the number of downward movements of the bond prices or, equivalently, the number of upward
movements of the short-term rate. Hence, we may equivalently index the short-term rate by s,
instead than by j, and rewrite Eq. (13.35) as follows:

rs ( ) = F (0) + ln (q + (1 q) ) + ln 1 s,

= r0 ( )
458
c
by
A. Mele

Std(r)
To nd , we make reference to the relation, ln 1 =
, where q =
q(1q)
1
2
and Std(r)
is the standard deviation of the short-term rate, which equals Std(r) = 0.014, annualized.
/ 1 = 0.02 or = 0.9802.
Therefore ln 1 = 0.014
2
2
For the Ho & Lee model, we know the closed-form expression for r0 ( ),

s
s=0 ps ( )
r0 ( ) = ln
,
(13.36)
P$ (0, + 1)
where ps ( ) denotes the price of an Arrow-Debreu security which pays of $1 in state s at time
, and zero otherwise. Given the term-structure of prices P$ (0, + 1), = 0, 1, 2, we populate
the tree using Eq. (13.36) and the forward equation for the Arrow-Debreu prices developed in
the lecture notes,

1
ps ( ) = er0 ( 1) s ps ( 1) + s1 ps1 ( 1) ,
2
(13.37)
with the appropriate boundary conditions.

So we have to compute interest rates and Arrow-Debreu prices for = 0, 1, 2.
= 0. Eq. (13.36) is trivial. It leads to,

r0 (0) = ln
1
P$ (0, 1)
= 0.015.
The forward equation for the Arrow-Debreu prices, Eq. (13.37), is also trivial,
p0 (0) = 1.
= 1. Let us use Eq. (13.37), the forward equation for the Arrow-Debreu prices, to nd
p0 (1) and p1 (1). We have two cases:
s = 0. We have:
1
1
p0 (1) = er0 (0) [p0 (0) + 0] = er0 (0) = 0.4925.
2
2
The previous relation holds because p0 (1) is the current price of the Arrow-Debreu
security which pays o $1 in state 0 at time 1, as illustrated by the tree in the Figure
1 below,
q =
1
2
s =1
s = 0
1
= 0
=1
459
c
by
A. Mele

s = 1. By a similar reasoning,
1
1
p1 (1) = er0 (0) [0 + p0 (0)] = er0 (0) = 0.4925.
2
2
Eq. (13.36) is, now,

0.4925 (1 + 0.9802)
p0 (1) + p1 (1)
r0 (1) = ln
= ln
= 0.0069.
P$ (0, 2)
0.9685
Hence, by Eq. (13.34),

r1 (1) = r0 (1) + ln 1 = 0.0069 + 0.02 = 0.0270.
So, to sum up, we have the tree below,
q=
r1 (1) = 0.027
1
2
r0 (0) = 0.015
r0 (1) = 0.0069
=0
=1
where p0 (1) = p1 (1) = 0.4925.

We now proceed to compute the values of the short-term rate for one further period.
= 2. By Eq. (13.37), the forward equation for the Arrow-Debreu prices, we have the
following three cases:
(s = 0)
(s = 1)
(s = 2)
p0 (2) = 12 er0 (1) [p0 (1) + 0] = 0.2446

p1 (2) = 12 er0 (1) [p1 (1) + p0 (1)] = 0.4843
p2 (2) = 12 er0 (1) [0 + p1 (1)] = 0.2397
The tree below further illustrates how to obtain these prices.

s = 2
s =1
s =1
s = 0
s = 0
= 0
=1
460
= 2
c
by
A. Mele
13.7. Copying with credit risk
Consider, for example, p0 (2). It is the price of the Arrow-Debreu security for time 2,
under two consecutive downward movements of the short-term rate. This state can only
be accessed to through the state s = 0 at time = 1. But at state s = 0 at time = 1, the
value of the Arrow-Debreu asset is 12 er0 (1) . Hence, p0 (2) = p0 (1) 12 er0 (1) . By a similar
reasoning, we have that p2 (2) = p1 (1) 12 er1 (1) = p1 (1) 12 er0 (1) .
We can now compute the values of the short-term rate for each node. Eq. (13.36) is, now,

p0 (2) + p1 (2) + 2 p2 (2)
r0 (2) = ln
P$ (0, 3)
$
%
0.2446 + 0.9802 0.4843 + (0.9802)2 0.2397
= ln
= 0.0054.
0.9445

Hence, by Eq. (13.34),

rs (2) = r0 (2) + ln 1 s = 0.0054 + 0.02 s,
s = 0, 1, 2.
This yields the following values for the short-term rate: r0 (2) = 0.0054, r1 (2) = 0.0253,
and r2 (2) = 0.0452.
To summarize, the implied tree for the short-term rate is given by Figure 4 below.
r2 (2 ) = 0 . 0 4 5 2
q=
1
2
r1 (1) = 0 . 0 2 7
q=
1
2
r1 (2 ) = 0 . 0 2 5 3
r0 (0 ) = 0 . 0 1 5
r0 (1) = 0 . 0 0 6 9
r0 (2 ) = 0 . 0 0 5 4
=0
=1
=2
Naturally, the prices P = er in the nodes of the previous tree match those calculated in
Section 13.5.5, apart from discrepancies arising due to rounding errors.
13.7 Copying with credit risk

Two examples: callable and convertible bonds.
461

13.7.1
c
by
A. Mele
Callable bonds
For example, to evaluate callable bonds through trees, we follow that methodology described
in this chapter, that we correct for the presence of credit risk.
(i) First, we populate a short-term rate tree through one of the models described in this
chapter (say, for example, through the Black, Derman and Toy model).
(ii) Second, we use this tree to nd the value of some coupon bearing bond of interest, by
just using the short-term rate process of the previous step.
(iii) Third, we use the results obtained in the second step and build up a tree for the callable
bond. In each node immediately preceding the maturity, we compare the strike price with
the non-callable coupon bearing bond price (ex-coupon) and take the minimum of the
two. We add the coupon to this minimum and nd, then, the payo of the non-callable
bond at the relevant node. This gives us V = min{K, B rolled-back (ex-coupon)} + coupon,
where K is the call price, and B rolled-back (ex-coupon) is the ex-coupon bond price, which
is found from the values of the bond V in the next nodes (by using, as usual, recursive,
backward solution, i.e. the risk-neutral expectation of the future payos).
(iv) Fourth, we go backward, discounting the values obtained in the previous steps, V say,
obtaining, for each node, V = min{K, V } + coupon, etc. Hence, we nd the price. If the
callable bond is not subject to default risk, we stop. Otherwise, we proceed to the next
step.
(v) Fifth, we correct for credit risk. The price we found in the fourth step is typically dierent
than the market price. One issue is that the market price reects the credit risk of the
rm, and should be typically less than the price obtained in the fourth step. The trick,
here, is to search for an additional spread to add to the short-term rate process obtained
in the rst step, such that the theoretical bond price equals the market price of the bond.
This is done numerically, and of course alters the results obtained in steps 3 and 4.
At this point, we may price options written on callable bonds. Ho and Lee (2004) (Chapter
8, Section 8.3 p. 274-278) develop a number of useful exercises on the pricing of options on
callable bonds, through tree methods.
13.7.2
Convertible bonds
Recall from Chapter 12, that the parity, or conversion value, is CV = CR S, where S is the
price of the common share. The evaluation tree features the following three steps.
(i) First, we set the life of the tree equal to the life of the callable convertible bond.
(ii) Second, we assess the evolution of the stock price along the tree, under the risk-neutral
probability. (This is done according to the usual Cox, Ross and Rubinstein (1979) approach.)
(iii) Third, in each node, we compute the value of the bond as max{CV, min{B, K}}, where
CV is the conversion value, K is the call value, and B is the value of the bond which is
rolled-back from the values of the bond in the next nodes (by using, as usual, recursive, backward solution, i.e. the risk-neutral expectation of the future payos). That is,
462
c
by
A. Mele
assuming the bondholder does not convert, the value is B = min {B, K}, where B is the
rolled-back value of the bond. Then, the value is max{CV, B }.
Note, this procedure leads to ll in the nodes, once we know the appropriate interest rate. If
the rm was not subject to default risk, we would simply use the riskless interest rate. However,
the rm is obviously subject to default risk. In practice, we proceed as follows. In each node, the
value of the bond is decomposed in two parts. One part, related to the pure debt component,
which is discounted at the defaultable interest rate; and one part related to the pure equity
component, which is discounted at the default-free interest rate. Exercise 25.7 in Hull (2003)
(p. 653-654) illustrates a specic example.
463
c
by
A. Mele
13.8. Appendix 1: Proof of Eq. (13.9)

The derivation in this appendix is the discrete-time counterpart to that in Section 11.3.2.2 in Chapter
11.
Let P (r, Ti ) denote the price of a zero with maturity Ti , i = 1, 2, when the interest rate is equal
to r. We wish to replicate a zero with maturity T2 by means of a portfolio that includes a zero with
maturity T1 . Consider the following portfolio: (i) Go long zeros with maturity T1 and (ii) invest
M in the MMA. Let V0 be the current value of this portfolio. V0 is clearly a function of the current
short-term rate r, and equals,
V0 (r) = P (r, T1 ) + M.
In the second period, the value of the portfolio is random, as it depends on the development of the
short-term rate r. Precisely, the value of the portfolio in the second period, is
+
V (r+ ) = P (r+ , T1 ) + M (1 + r), with probability p
V (
r) =
V (r ) = P (r , T1 ) + M (1 + r), wit probability 1 p
We also know that in the second period, the value of the second zero is,
+
with probability p
P (r+ , T2 )
P (
r, T2 ) =
P (r , T2 )
with probability 1 p
Next, we select and M to make the value of the portfolio equal the value of the second zero, in each
state of nature, viz
V (
r) = P (
r, T2 ) , in each state.
Mathematically, this is tantamount to solving the following system of two equations with two unknowns
( and M),
+
V (r+ ) = P (r+ , T1 ) + M (1 + r) = P (r+ , T2 )
(13A.1)
V (r ) = P (r , T1 ) + M (1 + r) = P (r , T2 )
The solution is,
= P (r , T2 ) P (r , T2 ) ,
P (r+ , T1 ) P (r , T1 )
+
+
= P (r , T2 )P (r , T1 ) P (r , T2 )P (r , T1 ) .
M
[P (r+ , T1 ) P (r , T1 )] (1 + r)
M
), replicates the value of the second zero in the second
By construction, the previous portfolio, (,
period. But if two assets (the portfolio, and the second zero) yield the same payos in each state of
the nature, they must be worth the same, in the absence of arbitrage. Therefore, we must have,
V0 (r)|=,M=
= P (r, T1 ) + M = P (r, T2 ) ,
M
or,
= (1 + r) P (r, T2 ) (1 + r)
P (r, T1 ) .
(1 + r) M
(13A.2)
Next, let us gure out the prediction of the model in terms of the expected return it generates for
M
). To do this, multiply the rst equation in
the price of the bond maturing at T1 , when (, M) = (,
M =M
(13A.1) by p, and multiply the second equation in (13A.1) by 1 p. Add the result for = ,
to obtain,

pP (r+ , T1 ) + (1 p) P (r , T1 ) + M
(1 + r) = pP (r+ , T2 ) + (1 p)P (r , T2 ).
Replacing (13A.2) into the previous equation yields,

pP (r+ , T1 ) + (1 p) P (r , T1 ) (1 + r)P (r, T1 )

= pP (r+ , T2 ) + (1 p)P (r , T2 ) (1 + r)P (r, T2 ) .
464
c
by
A. Mele

into the previous equation leaves,
Finally, replacing the solution for
[pP (r+ , T1 ) + (1 p)P (r , T1 )] (1 + r)P (r, T1 )

[pP (r+ , T2 ) + (1 p) P (r , T2 )] (1 + r) P (r, T2 )
=
.
P (r+ , T1 ) P (r , T1 )
P (r+ , T2 ) P (r , T2 )
The previous equation is easy to interpret. The numerators are the expected excess returns from
holding the assets. They equal Ep [P (
r, Ti )] (1 + r) P (r, Ti ), where Ep [P (
r, Ti )] is what the investors
expect to receive, the next period, by investing P (r, Ti ) today, in the bond; and (1 + r) P (r, Ti )
is what the investors expect to receive, the next period, by investing P (r, Ti ) today, in the MMA.
The denominators constitute a measure of volatility related to holding the assets. Then, the previous
equation tells us that the Sharpe ratios, or the unit risk premiums, on the two zeros agree.
Let the Sharpe ratio on any zero be equal to some function of the short-term rate r only (and
possibly of calendar time). This function, , does not clearly depend on the maturity of the zeros.
Then, we have,

pP (r+ , T1 ) + (1 p)P (r , T1 ) (1 + r) P (r, T1 ) = P (r+ , T1 ) P (r , T1 )
=
P (r+ , T1 ) P (r , T1 )
[(r+ r )]. (13A.3)
r + r
We can interpret (r+ r ) as a measure of interest rate volatility, and dene Vol(
r r) (r+ r ).
Eq. (13.9) follows by rewriting Eq. (13A.3) for a generic maturity date T > 2.
465
c
by
A. Mele

Consider the equation dening the discretely compounded forward rate (see Chapter 11):
1
1+FT ( ) , where FT ( ) F ( , T, T + 1). Iterating this equation leaves:
P ( , T ) =
T1
1
S=
P ( ,T +1)
P ( ,T )
T 1
P (t, T ) P (t, ) 1
1
1
=
.
1 + FS ( )
P (t, ) P (t, T )
1 + FS ( )
S=
Therefore, at any instant of time t : t < < T , we have that,

T 1
P (t, T ) 1 1 + FS (t)
P ( , T ) =
.
P (t, )
1 + FS ( )
(13A.4)
S=
Eq. (13A.4) gives us the price of the bond at a future date . It reveals that the price P ( , T ) as of
time can be expressed as a function of the current bond prices P (t, T ) and P (t, ), and how forward
rates will change from the current time t to the time at which the derivative payo will be paid,
1+FS (t)
, for S = , , T 1. Hence, once we model the evolution of forward rates, we also have
i.e. 1+F
S ( )
a model of the future bond price movements, P ( , T ), which we can use to price, at the evaluation
time t, interest rate derivatives, with payos depending on the realization of the bond price P ( , T )
at time .
To normalize the time-line, we now set t = 0. Redening = t, Eq. (13A.4) then reduces to,
P (t, T ) =
T 1
P (0, T ) 1 1 + FS (0)
.
P (0, t)
1 + FS (t)
(13A.5)
S=
It is quite natural, at this juncture, to search for the models predictions about the evolution of
future forward rates. Not only is this task theoretically important, it is also relevant as a matter of
the practical implementation of the model. Indeed, if the models predictions about the evolution of
future forward rates yields a closed-form solution, the bond price at the future date t, P (t, T ), could
be expressed in a closed-form, which might facilitate the implementation details of the model.
Let us introduce some further notation. Let FSj (t) be the forward rate as of time t after the occurrence of j upward movements in the bond price, and let the continuously compounded forward rate
FSj (t) be dened as,

F j (t) ln 1 + F j (t) , j t.
S
By Eq. (13A.5), then,
T 1
1 j
P (0, T ) 1 1 + FS (0)
P (0, T ) TS=t
(FS (t)FS (0)) .
Pj (t, T ) =
=
e
j
P (0, t)
P (0, t)
S=t 1 + FS (t)
(13A.6)
We have the following important result, which we shall prove later on:
u (S + 1 t)
(t j) ln ,
FSj (t) = FS (0) + ln
u (S + 1)
j t.
(13A.7)
By replacing Eq. (13A.7) into Eq. (13A.6), and using the solution for the perturbation function u ()
in Eqs. (13.22), we get Eq. (13.24).
So we are left with proving Eq. (13A.7). The proof proceeds by induction. Eq. (13A.7) holds true
for t = 0. Next, suppose that it holds at time t. We wish to show that in this case, Eq. (13A.7) would
also hold at time t + 1. At time t + 1, we have two cases.
466
c
by
A. Mele

Case 1 : A positive price jump occurs between time t and time t + 1. In this case,
Pj+1 (t + 1, S)
Pj+1 (t + 1, S + 1)

Pj (t, S)
Pj (t, S + 1)
= ln u (S t)
ln u (S + 1 t)
Pj (t, t + 1)
Pj (t, t + 1)
u (S t)
+ FSj (t)
= ln
u (S + 1 t)
u (S + 1 (t + 1))
= ln
+ FS (0) [(t + 1) (j + 1)] ln ,
u (S + 1)
FSj+1 (t + 1) = ln
where the rst equality and the third follow by the denition of FSj+1 (t), the second equality holds by
the denition of the jump in Eq. (13.17), the fourth equality follows by using Eq. (13A.7). Hence, Eq.
(13A.7) holds at time t + 1 in the occurrence of a positive price jump between time t and time t + 1.
Case 2 : A negative price jump occurs between time t and time t + 1. In this case,
Pj (t + 1, S)
Pj (t + 1, S + 1)

Pj (t, S)
Pj (t, S + 1)
= ln d (S t)
ln d (S + 1 t)
Pj (t, t + 1)
Pj (t, t + 1)
d (S t)
= ln
+ FSj (t)
d (S + 1 t)
FSj (t + 1) = ln
u (S + 1 t)
d (S t) (St)+1
1 + FS (0) + ln
(t j) ln
(S+1t)+1
u (S + 1)
d (S + 1 t)
u (S t)
= ln
+ FS (0) [(t + 1) j] ln ,
u (S + 1)
= ln
where the rst four equalities follow by the same arguments produced in Case 1, the fth equality
holds by the relation u (T ) = d (T ) (T 1) in Eq. (13.21) and the last equality follows by rearranging
terms. Hence, Eq. (13A.7) holds at time t + 1 in the occurrence of a negative price jump between time
t and time t + 1.
These two cases reveal that if Eq. (13A.7) holds at time t for any j t, it also holds at time t + 1,
in each state of nature. By induction, Eq. (13A.7) is therefore true.
467
c
by
A. Mele
References
Black, F., E. Derman and W. Toy (1990): A One Factor Model of Interest Rates and its
Application to Treasury Bond Options. Financial Analysts Journal (January-February),
33-39.
Cox, J. C., S. A. Ross and M. Rubinstein (1979): Option Pricing: A Simplied Approach.
Journal of Financial Economics 7, 229-263.
Diebold, F. X. and C. Li (2006): Forecasting the Term Structure of Government Bond Yields.
Ho, T. S. Y. and S.-B. Lee (1986): Term Structure Movements and the Pricing of Interest
Rate Contingent Claims. Journal of Finance 41, 1011-1029.
Ho, T. S. Y. and S.-B. Lee (2004): The Oxford Guide to Financial Modeling. Oxford University
Press.
Hull, J. C. (2003): Options, Futures, and Other Derivatives. Prentice Hall. 5th edition (International Edition).
Hull, J. C. and A. White (1990): Pricing Interest Rate Derivative Securities. Review of
Financial Studies 3, 573-592.
McCulloch, J. (1971): Measuring the Term Structure of Interest Rates. Journal of Business
44, 19-31.
McCulloch, J. (1975): The Tax-Adjusted Yield Curve. Journal of Finance 30, 811-830.
Nelson, C.R. and A.F. Siegel (1987): Parsimonious Modeling of Yield Curves. Journal of
Business 60, 473-489.
Vasicek, O. (1977): An Equilibrium Characterization of the Term Structure. Journal of
468

bf4a9329-ba54-4241-a6a3-4ec51aba7767

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

bf4a9329-ba54-4241-a6a3-4ec51aba7767

Încărcat de

Drepturi de autor:

Formate disponibile

Lecture Notes in Financial Economics

1 The classic capital asset pricing model

2 The CAPM in general equilibrium

The static general equilibrium in a nutshell . . .

3 Infinite horizon economies

4 Continuous time models

5 Taking models to data

Asset pricing and reality

6 On kernels and puzzles

9 Information and other market frictions

Applied asset pricing theory

10 Options and volatility

12 Risky debt and credit derivatives

12.4.4 Credit spread options (CSOs) . . . . . . . . . . . . . . . . . . . .

Fischer Black, 1995, p. 4, Exploring General Equilibrium, The MIT Press.

1.1 Portfolio selection

The wealth constraint

1.1. Portfolio selection

The rst order conditions for [1.P1] are,

1.1. Portfolio selection

Without the safe asset

his current wealth is w = m

(vp ) = arg max E w+ ()

1.1. Portfolio selection

1.1. Portfolio selection

Hence, we see that

The market portfolio

1.1. Portfolio selection

Let M be the market portfolio. To identify M , we note that it belongs to AM C if

Therefore, we must be looking for the value vM that solves

1.2. The CAPM

1.2 The CAPM

1.2. The CAPM

1.3. The APT

The certainty equivalent C is dened as:

1.3 The APT

1.3. The APT

This is simply the SML in Eq. (1.18).

The APT with idiosyncratic risk and a large number of assets

[Ross (1976), and Connor (1984), Huberman (1983).]

1.3. The APT

1.3. The APT

Fama-MacBeth. Economic forces driving asset returns.

1.4. Appendix 1: Some analytical details for portfolio choice

1.4 Appendix 1: Some analytical details for portfolio choice

The primal program

Given Eq. (1.8), the Lagrangian function associated to [P2] is,

Using the rst and the third conditions, we obtain,

We can solve for 2 , obtaining,

Next, we derive the value of the program [P2]. We have,

It is easy to check that

Let us gather Eqs. (1A.3) and (1A.4),

1.4. Appendix 1: Some analytical details for portfolio choice

We can use this expression for 1 to express

The dual program

We now solve the dual program, dened as follows,

= arg minm var

for some constant Ep . The rst order conditions are

By replacing the rst condition in (A14) into the third one,

By Eqs. (1A.8) and (1A.9), the solutions for 1 and 2 are,

1.4. Appendix 1: Some analytical details for portfolio choice

Therefore, the solution for the portfolio in Eq. (A14) is,

1.5. Appendix 2: The market portfolio

1.5 Appendix 2: The market portfolio

Finally, it is easy to show that a) b).

= 2 , , m . The converse is straight forward.

and, nally, the null space of W ,

FIGURE 2.3. Complete markets, V = R2 .