The Swaption Cube

Anders B. Trolle
Ecole Polytechnique Fédérale de Lausanne and Swiss Finance Institute

Eduardo S. Schwartz
UCLA Anderson School of Management and NBER

Abstract

We use a comprehensive database of inter-dealer quotes to conduct the first empirical anal-
ysis of the swaption cube. Using a model independent approach, we first establish new
stylized facts regarding the variation in conditional volatility and skewness of swap rates.
In particular, we show that skewness is stochastic, largely unspanned by rates and volatil-
ity, and sometimes changes sign. We then incorporate these stylized facts into a dynamic
term structure model and provide new evidence on the pricing of volatility and skewness
risk in addition to interest rate risk. Finally, we relate volatility and skewness to perceived
macroeconomic uncertainty. We show that GDP growth uncertainty is the most important
determinant in the USD market, while inflation uncertainty is more important in the EUR
market. This is consistent with differences in monetary policy objectives.

JEL Classification: E43, G12, G13

CONTAINS ONLINE APPENDIX

This version: January 2012

———————————
We are grateful to Michael Brennan, Damir Filipovic, and Jukub Jurek for extensive discussions, Peter Honoré
and Kasper Lorenzen for valuable insights into the swaption market, and Michael Brandt, Patrick Cheridito,
Pierre Collin-Dufresne, Rudiger Fahlenbrach, Peter Feldhutter, Eric Ghysels, Harrison Hong, Julien Hugonier,
Markus Leippold, Loriano Mancini, Kristian Miltersen, Erwan Morellec, Lasse Pedersen, Pedro Santa-Clara,
Olivier Scaillet, Fabio Trojani, Rossen Valkanov, and seminar participants at the 2010 FINRISK meeting,
the 2011 European Finance Association meeting, the 2011 Princeton-Lausanne Workshop on Quantitative
Finance, the University of Lugano, and the University of Zurich for comments. Anders Trolle: Quartier UNIL-
Dorigny, Extranef 216, CH-1015 Lausanne, Switzerland. E-mail: anders.trolle@epfl.ch. Eduardo Schwartz:
UCLA Anderson School of Management, 110 Westwood Plaza, Los Angeles, CA 90095-1481. E-mail: ed-
uardo.schwartz@anderson.ucla.edu. Trolle gratefully acknowledges research support from NCCR FINRISK of
the Swiss National Science Foundation.

Electronic copy available at: http://ssrn.com/abstract=1698910

1 Introduction

A vast term structure literature has investigated the dynamics of interest rates and risk pre-
mia; see, e.g., the survey of Dai and Singleton (2003). More recently, starting with Ang and
Piazzesi (2003), a significant research effort has been directed at providing a link to macroeco-
nomic fundamentals. However, despite many advances, we still know surprisingly little about
the dynamics of higher-order moments of interest rates, their associated risk premia, and their
macroeconomic determinants. Such knowledge is critical for the trading, pricing, hedging,
and risk management of interest rate derivatives – particularly those that are in or out of the
money.1 In this paper, we use a new database of options on interest rate swaps, so-called
swaptions, to provide new insights on higher-order interest rate moments. Our database is
uniquely suited for this purpose, since it is the first to provide extensive information about
out-of-the-money swaptions.
Specifically, our paper has three related objectives: First, using a model independent
approach, we establish a set of stylized facts regarding the cross-sectional and time-series
variation in conditional moments of swap rates. Our focus is on volatility and skewness, since
we show that kurtosis displays less systematic variation. Second, we develop and estimate
a dynamic term structure model that is consistent with these stylized facts, and use it to
study the pricing of volatility and skewness risk in addition to interest rate risk. Third, we
investigate the fundamental drivers of volatility and skewness with a focus on the impact of
perceived uncertainty about the macroeconomy.
Apart from deepening our understanding of interest rate dynamics and investor risk prefer-
ences, our analysis has several practical applications: First, the swaption market is very large
with a daily trading volume in excess of USD 150 billion in terms of notional value. Since fi-
nancial institutions hold significant portfolios of swaptions, a better understanding of swaption
market dynamics is important for the stability of the financial system. Indeed, several large
losses in financial institutions in recent years have been blamed on deficient risk management
of swaptions.2 Second, the swaption market is important for the wider fixed-income market
as many standard fixed income securities, such as fixed-rate mortgage-backed securities and
1
Few outstanding interest rate derivatives are at the money. Even those that are at the money at the time of
issuance quickly become in or out of the money. Given the long-dated nature of many interest rate derivatives,
legacy trades are often deeply in or out of the money.
2
For instance, Stein, Tudor, and Fukase (2011) report that the majority of a USD 961 million loss incurred by
a Morgan Stanley and Mitsubishi UFJ joint venture was due to inadequate hedging of JPY swaption positions.

Electronic copy available at: http://ssrn.com/abstract=1698910

callable agency securities, imbed swaption-like options. This is also the case for a wide variety
of structured products. Such imbedded options are rarely, if ever, at the money. Third, many
large corporations are active in the swaption market either directly or indirectly (through the
issuance and swapping of callable debt), and our analysis may, therefore, have implications
for corporate finance decisions.3 Fourth, while new regulation will demand central clearing
and more transparent trading of interest rate swaps, regulation of swaptions is still being de-
bated. Our study helps inform this debate by providing more information about an otherwise
relatively opaque market.4
Our paper takes advantage of a unique data set from the largest inter-dealer broker in the
interest rate derivatives market, which records prices of swaptions along three dimensions: the
maturities of the underlying swaps, the expiries of the options, and the option strikes. This
three-dimensional grid of prices is known, among market participants, as the swaption cube.
The range along all three dimensions is wide with swap maturities from 2 years to 30 years,
option expiries from 1 month to 10 years, and strike intervals up to 800 basis points. The data
covers a period from 2001 to 2010 spanning two recessions, the financial crisis, and times of
inflation and deflation fears. Moreover, the data contains both USD and EUR denominated
swaptions. As such, it is arguably the most comprehensive data set on interest rate derivatives
available.
Our first contribution is a model independent analysis of the swaption cube. For a given
swap maturity and option expiry, we compute the conditional moments (under the appropriate
pricing measure) of the underlying swap rate at a time horizon equal to the option expiry. This
is done by suitably integrating over prices of swaptions with different strikes. For instance,
supposing we consider the 1-year option on the 10-year swap, then the strike-dimension of
swaption prices gives us the conditional moments of the distribution of the 10-year swap rate
in 1-year’s time.
We investigate how the conditional moments vary in the cross-section (with option expiry
and swap maturity) and over time. The aim is to establish a set of stylized facts that a
3
A 2009 survey by the International Swaps and Derivatives Association found that 88.3 percent of the
Fortune Global 500 companies use interest rate derivatives, such as swaps and swaptions, to manage interest
rate risk.
4
Among the main obstacles to central clearing of swaptions, is the ability of clearing houses to assess the
risks in swaption positions. Our paper identifies the main risk factors in the swaption market, and our model
framework can be used to determine appropriate initial and variation margins on swaption positions.

2
well-specified dynamic term structure model should be able to replicate. Results concerning
conditional volatility are largely consistent with existing studies.5 In particular, we find that
conditional volatility varies significantly over time, and that volatility contains a systematic
component that is unspanned by the term structure. Results regarding conditional skewness
and kurtosis are novel. We show that conditional skewness exhibits significant variation over
time to the extent that conditional skewness sometimes changes sign. Moreover, changes
in conditional skewness are largely unrelated to changes in swap rates and volatility, and
display strong common variation across swap maturities and option expiries. This strongly
indicates that skewness risk contains a systematic component that is unspanned by both the
term structure and volatility. Conditional swap rate distributions always have fat tails, but
conditional kurtosis exhibits less systematic variation than is the case for conditional skewness.
For this reason, we mainly focus on conditional volatility and skewness in the paper.
Our second contribution is the development and estimation of a new dynamic term struc-
ture model that is consistent with the established variation in conditional moments. Our
model features any number of term structure factors. Volatility depends on two additional
factors, both of which may contain a component that is unspanned by the term structure.
Under certain parameter restrictions, skewness contains a component that is unspanned by
both the term structure and volatility. The framework is flexible and highly tractable with
swaptions priced via a fast and accurate Fourier-based pricing formula. By specifying shocks
to the term structure judiciously (a specification that encompasses “level”, “slope”, and “cur-
vature” shocks), the model reduces to a particular case of an affine term structure model. Via
a parsimonious market price of risk specification, the model provides insights on the compen-
sation that market participants require for bearing exposures to variance and skewness risk in
addition to interest rate risk. The model is estimated by maximum-likelihood on a panel data
set, which includes all the swaption prices and underlying swap rates in the swaption cubes.
We find that a specification with three term structure factors, partially unspanned volatility
factors, and parameter restrictions to ensure unspanned stochastic skewness is able to capture
most of the cross-sectional and time-series variation in conditional volatility and skewness of
swap rates. In particular, it matches the degree of unspanned stochastic volatility and skewness
observed in the data. Importantly and in contrast to Joslin (2007), who argues that imposing
unspanned stochastic volatility in affine term structure models is detrimental to their ability to
5
Note, however, that existing studies rely on at-the-money implied volatility obtained from a pricing model,
while here we compute conditional volatility model-independently.

3
fit interest rates, our model preserves a good fit to the term structure. This suggests that our
framework is more flexible in terms of accommodating unspanned stochastic volatility (and
skewness).6
In terms of risk premia, our novel results relate to the pricing of skewness risk and the im-
pact that skewness has on the pricing of interest rate and variance risk. In our estimated model,
bond risk premia are positive, and Sharpe ratios on bonds increase with interest rate volatility
and skewness.7 Since both volatility and skewness of interest rates contain a component that
is unspanned by the term structure, bond risk premia also contain an unspanned component.
This is similar to recent findings in Cochrane and Piazzesi (2008), Ludvigson and Ng (2009),
Joslin, Priebsch, and Singleton (2010), Duffee (2011), and Buraschi and Whelan (2011), al-
though the source of the unspanned component is different (unspanned higher-order moments
instead of unspanned latent factors or macro risks).
Variance risk premia are negative suggesting that investors dislike states with high interest
rate (and bond return) variance. This is in line with Duarte, Longstaff, and Yu (2007) and
Joslin (2007) as well as a large number of papers that document negative variance risk premia
in equity markets. We find that the Sharpe ratio on a security with a pure negative exposure
to interest rate variance increases with interest rate volatility and skewness.
Skewness risk premia are also negative suggesting that investors dislike states with more
positively skewed interest rates distributions. Since bond return distributions are more neg-
atively skewed in those states, the result is consistent with Arditti (1967) and Scott and
Horvath (1980), who show that, under very general assumptions, investors have a preference
for positive skewness in return distributions.8 The Sharpe ratio on a security with a pure
negative exposure to interest rate skewness increases with interest rate volatility but is non-
6
Intuitively, the affine representation of our model has several locally deterministic state variables, which
cancel the convexity effects arising from stochastic volatility. Such additional state variables are not considered
by Joslin (2007).
7
Many empirical studies have shown that bond risk premia also depend on the shape of the term structure;
see, e.g., Campbell and Shiller (1991) and Cochrane and Piazzesi (2005). In principle, our model could incorpo-
rate such dependence. In practice, the length of our sample period severely restricts the number of market price
of risk parameters that can be reliably estimated. Also, since volatility and skewness are largely unspanned by
the term structure, our results are likely not very sensitive to letting bond risk premia also depend on the term
structure factors.
8
Building on these results, Kraus and Litzenberger (1976) and Harvey and Siddique (2000) have developed
asset pricing models, where assets’ coskewness with the market portfolio is priced.

4
monotonic in interest rate skewness. Robustness checks show that these results hold true in
more flexible models.
Our third objective is understanding the fundamental drivers of volatility and skewness.
A recent literature has investigated the link between the term structure and macroeconomic
variables such as output growth and inflation.9 Our hypothesis is that volatility and skewness
are related to the perceived uncertainty about these macroeconomic variables. To quantify
uncertainty, we infer probability distributions for 1-year ahead real GDP growth and inflation
from the survey of professional forecasters conducted in both the US and the Eurozone.10
Then, we regress interest rate volatility and skewness on the dispersion and skewness of the
probability distributions for future real GDP growth and inflation. We also control for other
factors that may have an effect on swap rate distributions, including volatility and skewness
of equity index returns, market-wide liquidity, and refinancing activity.
Consistent with our hypothesis, we find that volatility tends to increase when the macroe-
conomic outlook becomes more uncertain, while skewness tends to increase when the macroe-
conomic outlook becomes more skewed towards higher growth and inflation rates. At the same
time, we also show that beliefs about GDP growth are the most important determinants in the
USD market, while beliefs about inflation are more important in the EUR market. One likely
reason for this is that the primary policy goal of the European Central Bank is to maintain
price stability, whereas the Federal Reserve has a dual mandate of maximum employment
and price stability, leading it to place relatively more emphasis on expectations for real GDP
growth when setting interest rates.
The importance of the swaption market has spurred a number of empirical studies over the
past decade.11 Our paper differs from these in two important ways. First, all existing studies
rely on at-the-money swaption data, which provide no information about third or fourth
9
See, among others, Ang and Piazzesi (2003), Gallmeyer, Hollifield, and Zin (2005), Ang, Piazzesi, and
Wei (2006), Smith and Taylor (2009), Bekaert, Cho, and Moreno (2010), Bikbov and Chernov (2010),
Chun (2010), and Joslin, Priebsch, and Singleton (2010).
10
This survey is unique, because individual participants are asked to assign probability distributions to their
forecasts. Therefore, the aggregate probability distributions take both individual uncertainty and disagreement
among survey participants into account making it a better measure of uncertainty than commonly used measures
based purely on disagreement, see Zarnowitz and Lambros (1987).
11
See Longstaff, Santa-Clara, and Schwartz (2001a, 2001b), Driessen, Klaassen, and Melenberg (2003), Fan,
Gupta, and Ritchken (2003), de Jong, Driessen, and Pelsser (2004), Han (2007), Joslin (2007), Duarte (2008),
Trolle and Schwartz (2009), and Carr, Garbaix, and Wu (2009).

5
moments of interest rates. Second, existing studies are mostly concerned with the pricing of
swaptions, while a key objective of this paper is to also understand risk compensation and the
fundamental drivers of interest rate moments.
Information about higher-order interest rate moments may, in principle, also be inferred
from other fixed-income option markets for which implied volatility smiles exist, such as options
on Eurodollar and Treasury futures as well as interest rate caps/floors. The swaption cube data
is much more extensive, in that it covers a wider range of interest rate maturities and option
expiries.12 Swaptions are also, in many ways, simpler and more transparent instruments.13
In terms of the link to fundamentals, related papers include Beber and Brandt (2006), who
show that the reduction in macroeconomic uncertainty associated with the release of macroe-
conomic statistics reduces the implied volatility of Treasury bond futures options, and Cieslak
and Povala (2011), who find that macro and liquidity factors are important determinants of
realized Treasury market volatility.
The rest of the paper is organized as follows: Section 2 describes the swaption cube data.
Section 3 uses a model independent approach to establish several stylized facts about con-
ditional swap rate moments under the pricing measure. Section 4 describes and evaluates a
dynamic term structure model for swap rates and studies the pricing of risk. Section 5 in-
vestigates the economic determinants of conditional swap rate moments. Section 6 considers
a variety of robustness checks. Section 7 concludes. Several appendices, including an online
appendix, contain derivations and additional tables and figures.
12
Eurodollar futures options are essentially short-term options on 3M LIBOR rates, caps/floors are portfo-
lios of long-term options on 3M LIBOR rates, and Treasury futures options are essentially short-term options
on Treasuries. Of these three markets, most papers have analyzed volatility smiles in caps/floors; see, e.g.,
Gupta and Subrahmanyam (2005), Li and Zhao (2006), Jarrow, Li, and Zhao (2007), Deuskar, Gupta, and
Subrahmanyam (2008), Trolle and Schwartz (2009), Li and Zhao (2009), and Deuskar, Gupta, and Subrah-
manyam (2010).
13
Eurodollar and Treasury futures options are American style. Treasury futures come with the additional
cheapest-to-deliver option. In the cap/floor market, the individual LIBOR options (caplets) are not quoted
directly, but need to be stripped from quoted cap/floor prices, which is a non-trivial exercise.

6
2 The swaption market

2.1 The swaption cube data

A standard European swaption is an option to enter into a fixed versus floating forward starting
interest rate swap at a predetermined rate on the fixed leg. A receiver swaption gives the right
to enter a swap, receiving the fixed leg and paying the floating leg, while a payer swaption gives
the right to enter a swap, paying the fixed leg and receiving the floating leg.14 For instance,
a two-year into ten-year, five percent payer swaption is the option to pay a fixed rate of five
percent on a ten-year swap, starting two years from today.
The swaption cube is an object that shows how swaption prices vary along three dimensions:
the maturities of the underlying swaps (the swap tenors), the expiries of the options, and the
option strikes. In the swaption cube data provided to us, prices are quoted for 5 different swap
tenors (2, 5, 10, 20, and 30 years), 8 different option expiries (1, 3, 6, 9 months and 1, 2, 5,
and 10 years), and up to 15 different strikes given by fixed distances to the forward swap rate
(± 400, ± 300, ± 200, ± 150, ± 100, ± 50, ± 25, and 0 basis points), although swaptions
with negative strikes are obviously not quoted.15 Hence, the swaption cube gives an extremely
detailed view of the swaption market.
Swaptions trade over the counter (OTC) and typically via inter-dealer brokers. These act
as intermediaries; they facilitate price discovery and transparency by communicating dealer
interests and transactions, enhance liquidity, and allow financial institutions anonymity in
terms of their trading activities. Our swaption cube data is from ICAP plc., which is the
largest inter-dealer broker in the interest rate derivatives market, and as such provides the
most accurate quotations. In addition to swaption prices, ICAP also provided the underlying
forward swap rates as well as spot swap rates with maturities up to 40 years. Swaption prices
and swap rates reflect aggregates of dealer mid-quotes between 16:00 and 16:30 GMT.
We consider swaptions denominated in both USD and EUR, which are by far the most liquid
markets (see below). Although the data is available daily, we use weekly (Wednesday) data
14
USD swaps exchange a fixed rate for a floating 3-month LIBOR rate, with fixed-leg payments made semi-
annually, and floating-leg payments made quarterly. EUR swaps exchange a fixed rate for a floating 6-month
EURIBOR rate, with fixed-leg payments made annually, and floating-leg payments made semi-annually. In
both currencies, the daycount convention is 30/360 on the fixed leg, and Actual/360 on the floating leg.
15
Price quotes are for at- or out-of-the-money swaptions, i.e. receiver swaptions, when the strike is less than
the forward swap rate, and payer swaptions, when the strike is higher than the forward swap rate.

7
to avoid potential weekday effects, and to ease the computational burden of the estimation.
For the USD market, the data is from December 19, 2001 to January 27, 2010 (419 weeks),
while for the EUR market, the data is from June 6, 2001 to January 27, 2010 (449 weeks).
We apply various filters to the data; we eliminate obvious mistakes in the quotations and only
consider options for which the price is larger than USD (EUR) 100 in case of a swap notional
of USD (EUR) 1,000,000 (since, according to market sources, quotes for extremely deep OTM
swaptions are less reliable).16 In total, we use 172,658 quotes in the USD market, and 172,500
quotes in the EUR market for our analyses.
While swaptions are quoted in terms of prices, it is often more convenient to represent prices
in terms of implied volatilities – either log-normal or normal.17 Most market participants think
in terms of normal implied volatilities, as these are more uniform across the swaption grid and
more stable over time than log-normal implied volatilities. Therefore, in this paper, implied
volatilities always refer to the normal type, unless otherwise stated.

2.2 Liquidity

Given its OTC nature, the swaption market is relatively opaque. However, since 2010, the
TriOptima trade repository, which records all trades for 14 of the largest OTC derivatives
dealers, has contributed to increased transparency. For instance, as of July 1st, 2011, TriOp-
tima reports 213,457 outstanding swaption contracts with a combined notional value of 35.026
trillion USD equivalent. While this covers all currencies, the majority of trades are in USD
and EUR.18
Using data from the same set of firms, the International Swaps and Derivatives Association
(ISDA) recently reported the volume of swaption trades during the period April to June,
2011.19 The average daily volume amounted to 988 trades with a combined notional of USD 151
16
This implies that for a given underlying swap maturity, the range of strikes will increase with option expiry.
For a given swap maturity and option expiry, the range of strikes will vary over time with the level of volatility
and the level of the underlying forward swap rate.
17
The log-normal (or percentage) implied volatility is the volatility parameter that, plugged into the log-
normal (or Black (1976)) pricing formula, matches a given price. The normal (or absolute or basis point)
implied volatility is the volatility parameter that, plugged into the normal pricing formula, matches a given
price.
18
TriOptima reports that as of July 1st, 2011, 37.5 and 33.0 percent of the notional value of all outstanding
OTC interest rate derivatives contracts were in USD and EUR, respectively.
19
This data only includes trading activity that involves a transfer of risk between market participants.

8
billion. This significantly understates the true trading volume as it does not cover trades where
none of the 14 dealers is a counterparty. To put these numbers into perspective, during the
same period, the average daily trading volume of all options on Treasury futures amounted to
a combined notional of USD 29 billion.20 As such, in terms of traded notional values, liquidity
in the swaption market compares favorably with liquidity in the exchange-traded segment of
the interest rate options market.
In terms of the distribution of liquidity across swap maturities and option expiries, Tri-
Optima reports statistics on outstanding swaption contracts in the following seven swaption
maturity (swap maturity plus option expiry) buckets: 0-2 yrs, 2-5 yrs, 5-10 yrs, 10-15 yrs,
15-20 yrs, 20-30 yrs, and 30+ years. As of July 1st, 2011, the distribution was 5.6 (13.3),
16.8 (23.6), 20.8 (21.8), 25.0 (21.6), 9.6 (7.9), 13.0 (7.2), and 9.3 (4.6) percent in terms of the
number of contracts (notional values), showing that even ultra-long swaptions trade frequently.
In terms of the distribution of liquidity across moneyness, there is no publicly available data.
While at-the-money (ATM) swaptions are most liquid, market sources indicate that there is
significant trading activity in out-of-the-money (OTM) swaptions. Part of the reason is that
embedded options in callable bonds and structured products, which are often hedged via swap-
tions, are typically out of the money. In Section 3.2, we use indicative bid-ask spreads from
major investment banks to gauge the impact of liquidity on our results.

3 A model independent analysis of the swaption cube

In this section, we analyze the swaption cube from a model independent perspective. For
a given swap maturity and option expiry, the conditional moments (under the appropriate
pricing measure) of the underlying swap rate at a time horizon equal to the option expiry can
be inferred from the implied volatility smile.21 We analyze how conditional moments vary in
the cross-section (with option expiry and swap maturity) and over time.
20
This includes options on 2 yr, 5 yr, and 10 yr Treasury Note futures as well as the Treasury Bond and Ultra
T-Bond futures.
21
In principle, using the insight from Breeden and Litzenberger (1978), we could obtain the entire conditional
density, rather than just conditional moments, of the swap rate from the implied volatility smile. Beber and
Brandt (2006) and Li and Zhao (2009) study option-implied densities of Treasury futures prices and LIBOR
rates, respectively. In practice, however, it is a non-trivial matter to obtain the conditional density from a finite
number of option prices, and results may be quite sensitive to the choice of numerical scheme. In contrast,
conditional moments can be recovered in a robust fashion.

9
3.1 Conditional moments of the swap rate distribution

Consider a fixed versus floating interest rate swap for the period Tm to Tn with a fixed rate
of K. At every time Tj in a pre-specified set of dates Tm+1 , ..., Tn , the fixed leg pays τj−1 K
and the floating leg pays τj−1 L(Tj , Tj−1 ), where τj−1 is the year-fraction between times Tj−1
and Tj and L(Tj , Tj−1 ) is the τj−1 -maturity LIBOR rate set at Tj−1 .22 From the perspective
of the fixed-rate payer, the value of the swap at time t < Tm (assuming a notional of one) is
given by
n
X  RT 
− t j r(s)ds
Vm,n (t) = τj−1 EtQ e (L(Tj , Tj−1 ) − K) , (1)
j=m+1

where r(t) is the short rate and Q denotes the risk-neutral measure. Assuming that r(t) incor-
porates the same credit and liquidity risk as LIBOR, the above formula simplifies considerably
to

Vm,n (t) = P (t, Tm ) − P (t, Tn ) − KAm,n (t), (2)

where
n
X
Am,n (t) = τj−1 P (t, Tj ), (3)
j=m+1

is the swap annuity factor, and P (t, T ) denotes the time-t price of a zero-coupon bond maturing
at time T . In reality, inter-dealer contracts are virtually always collateralized, and cash flows
should, in principle, be discounted using the rate that is paid on collateral making swap (and
swaption) valuation significantly more involved; see, e.g., Johannes and Sundaresan (2007)
and Filipovic and Trolle (2011). In the interest of simplicity and in line with most studies, we
use the traditional approach to valuing swaps (and swaptions).23
The time-t forward swap rate, Sm,n (t), is the rate on the fixed leg that makes the present
value of the swap equal to zero and is given by
P (t, Tm ) − P (t, Tn )
Sm,n (t) = . (4)
Am,n (t)
22
For expositional ease, we assume that the fixed and floating leg payments occur at the same frequency,
although this is typically not the case. In the final valuation equations, it is only the frequency of the fixed leg
payments that matter.
23
For collateralized contracts, it is generally agreed that one should use discount factors inferred from overnight
index swaps (OIS), which are swaps that exchange a compounded overnight rate against a fixed rate. In practice,
the choice of discount rate has only a small impact on swap rates. For a comparison of the term structures
of forward rates with OIS and LIBOR discounting, see Bloomberg (2011). Even during the credit crisis, the
spread between the two term structures rarely exceeded a few basis points.

10
The forward swap rate becomes the spot swap rate at time Tm .
A payer swaption is an option to enter into an interest rate swap, paying the fixed leg at
a predetermined rate and receiving the floating leg. Let Pm,n (t, K) denote the time-t value of
a European payer swaption expiring at Tm with strike K on a swap for the period Tm to Tn .
At expiration, the swaption has a payoff of24

Vm,n (Tm )+ = (1 − P (Tm , Tn ) − KAm,n (Tm ))+

= Am,n (Tm ) (Sm,n (Tm ) − K)+ . (5)

At time t < Tm , its price is given by

h R Tm i
Pm,n (t, K) = EtQ e− t r(s)ds Am,n (Tm ) (Sm,n (Tm ) − K)+
 
= Am,n (t)EtA (Sm,n (Tm ) − K)+ , (6)

where A denotes the annuity measure associated with using Am,n (t) as numeraire.25 The
corresponding receiver swaption is denoted by Rm,n (t, K), and has a time-t price of
 
Rm,n (t, K) = Am,n (t)EtA (K − Sm,n (Tm ))+ . (7)

From (6) and (7) it is apparent that a receiver swaption can be viewed as a put option on a
swap rate, whereas a payer swaption can be viewed as a call option on a swap rate.
Using the insights from Bakshi and Madan (2000), Carr and Madan (2001), and Bakshi,
Kapadia, and Madan (2003) it follows that for any fixed Z, we can write any twice continuously
differentiable function of Sm,n (Tm ), g(Sm,n (Tm )), as
Z ∞
g(Sm,n (Tm )) = g(Z) + g′ (Z)(Sm,n (Tm ) − Z) + g′′ (K)(Sm,n (Tm ) − K)+ dK
Z
Z Z
+ g′′ (K)(K − Sm,n (Tm ))+ dK. (8)
0

24
EUR swaptions are typically cash-settled and have a payoff given by
n
X 1
(Sm,n (Tm ) − K)+ τj−1 ,
j=m+1
(1 + Sm,n (Tm ))τm,j

where τm,j is the year-fraction between times Tm and Tj . The advantage of using this formula, rather than
(5) is that counterparties only have to agree upon a single swap rate, rather than a complete set of discount
factors, to compute the cash settlement value. In practice, the difference between the two payoff formulas is
very small, and in the paper we use (5) also for EUR swaptions; see, e.g., Andersen and Piterbarg (2010) for
further details.
25
For a discussion of the annuity measure; see, e.g., Jamshidian (1997). Note that the annuity measure
changes with m and n. To lighten notation, we have suppressed this dependence.

11
Taking expectations under the annuity measure and setting Z = Sm,n (t), we obtain an expres-
sion in terms of prices of OTM receiver and payer swaptions

1 Z ∞
A
Et [g(Sm,n (Tm ))] = g(Sm,n (t)) + g′′ (K)Pm,n (t, K)dK
Am,n (t) Sm,n (t)
Z Sm,n (t) 
+ g′′ (K)Rm,n (t, K)dK . (9)
0

We can use this result to compute conditional moments of the swap rate distribution at a
time horizon equal to the expiry of the option. First, by construction of the annuity measure,
the conditional mean of the future swap rate distribution is simply the current forward swap
rate:

µt ≡ EtA [Sm,n (Tm )] = Sm,n (t). (10)

Then, using (9), we get the following expressions for conditional variance, skewness, and
kurtosis of the future swap rate distribution:
h i
2
VarA
t (S (T
m,n m )) = E A
t (S (T
m,n m ) − µ t ) =
2 Z ∞ Z Sm,n (t) 
Pm,n (t, K)dK + Rm,n (t, K)dK (11)
Am,n (t) Sm,n (t) 0
h i
A 3
Et (Sm,n (Tm ) − µt )
SkewA t (Sm,n (Tm )) = 3/2
=
VarA t (Sm,n (t, Tm ))
R R Sm,n (t) 
6 ∞
Am,n (t) Sm,n (t) (K − S m,n (t)) P m,n (t, K)dK + 0 (K − S m,n (t)) R m,n (t, K)dK
(12)
Vart (Sm,n (t, Tm ))3/2
A
h i
EtA (Sm,n (Tm ) − µt )4
KurtAt (Sm,n (Tm )) = 2 =
VarA t (Sm,n (t, Tm ))
R R Sm,n (t) 
12 ∞ 2 2
Am,n (t) Sm,n (t) (K − S m,n (t)) P m,n (t, K)dK + 0 (K − S m,n (t)) R m,n (t, K)dK
(13)
Vart (Sm,n (t, Tm ))2
A

As discussed in Section 2.1, swaptions are only available for a finite set of strikes, while
the formulas presume the existence of a continuum of strikes. Swaption prices corresponding
to the required strikes in the scheme used for numerical integration are obtained by, first,
linearly interpolating between the available normal implied volatilities, and then converting
from implied volatilities to prices. For strikes below the lowest available strike, we use the
implied volatility at the lowest strike. Similarly, for strikes above the highest available strike,
we use the implied volatility at the highest strike. The approximation error caused by the

12
extrapolation of implied volatilities is very small, since swaption prices are very low in the
regions of strikes where extrapolation is necessary.26

3.2 Results

We now investigate how conditional swap rate moments vary in the cross-section (with option
expiry and swap maturity) and over time.27 The aim is to establish a set of stylized facts that
a well-specified dynamic term structure model should be able to replicate. Table 1 displays
results for conditional volatility (the annualized standard deviation, measured in basis points)
of the swap rate distribution for different swap maturities (tenors) and at different option
expiries. It reports the sample means and, in parentheses, the sample standard deviations.
At short option expiries, conditional volatility is, on average, a hump-shaped function of swap
maturity. For the shortest swap maturity, conditional volatility is, on average, a hump-shaped
function of option expiry. This pattern is also found in earlier studies using ATM implied
volatilities, not taking the implied volatility smile into account. It is suggestive of a model in
which innovations to the term structure of forward rates exhibit a hump-shape.28
Conditional volatility exhibits significant variation over time. For instance, Figure 1, Panel
A displays the volatility of the conditional 1-year ahead distribution of the USD 10-year swap
rate, which varies between 67 bp and 177 bp through the sample period. For a given swap
maturity, the variation in volatility declines with option expiry, suggestive of a model exhibiting
mean-reverting stochastic volatility. Moreover, results reported in the online appendix show
that changes in volatility are largely unrelated to changes in the term structure. This is the
“unspanned stochastic volatility” phenomenon, which is the subject of a number of recent
26
We have experimented with different interpolation/extrapolation schemes and find that the results are very
robust to the choice of scheme. The only exception is for swaptions with 10-year option expiries in times of high
volatility, where swaptions outside of the available strike range do have some (small) values making the results
slightly dependent on the extrapolation approach. The integrals are evaluated with the trapezoid scheme using
999 integration points for each integral. The first integral in each expression is truncated at Sm,n (t) + 0.10.
27
When comparing moments for different tenors and at different option expiries, one should keep in mind that
these numbers are computed under different measures. However, a simulation exercise based on the estimated
dynamic term structure model of Section 4, shows that moments computed under the common risk-neutral
measure and the (tenor,expiry)-specific annuity measures are rather similar.
28
Here, and below, we emphasize suggestive, since moments in the cross-section are not, strictly speaking,
directly comparable (see previous footnote).

13
papers.29 A principal component (PC) analysis reveals large common variation in conditional
volatility across the swaption matrix. For instance, in the USD market, the first PC explains 84
percent of the variation (while the second and third PCs explain 9 and 5 percent, respectively).
Table 2 displays results for conditional skewness of the swap rate distributions for different
swap maturities (tenors) and at different option expiries. Like the previous table, it reports
the sample means and, in parentheses, the sample standard deviations. Conditional skewness
tends to be positive on average, implying that for most points on the swaption matrix, OTM
payer swaptions are more expensive than equivalently OTM receiver swaptions. For a given
swap maturity, conditional skewness is an increasing and concave function of option expiry,
which is suggestive of volatility following a stochastic process with low to moderate degrees of
mean reversion.30
Like conditional volatility, conditional skewness also exhibits significant variation over time.
The variation is such that the sign of conditional skewness often changes. For example, Figure
1, Panel B displays the skewness of the conditional 1-year ahead distribution of the USD 10-
year swap rate, which varies between -0.38 and 0.87 through the sample period. A principal
component analysis shows that conditional skewness also exhibits large common variation
across the swaption matrix. In the case of the USD market, the first PC explains 85 percent
of the variation (while the second and third PCs explain 7 and 2 percent, respectively).
Similar to volatility being only partially spanned by the term structure, an interesting
question is the extent to which skewness is spanned by the term structure and/or volatility;
that is, the extent to which “skewness risk” represents a separate source of risk. To investigate
this issue, we initially extract the main PCs driving weekly changes in forward swap rates and
the main PCs driving weekly changes in the conditional volatility of the swap rate distributions.
We use all the PCs explaining more than one percent of the variation, ensuring that they
summarize virtually all of the information in interest rates and volatility. Next, for each
point on the swaption matrix, we regress changes in the conditional skewness of the swap
29
This line of research was initiated by Collin-Dufresne and Goldstein (2002) and further evidence has been
provided by Heidari and Wu (2003), Andersen and Benzoni (2010), Li and Zhao (2006, 2009), Trolle and
Schwartz (2009), and Collin-Dufresne, Goldstein, and Jones (2009), among others.
30
The term structure of conditional skewness (and kurtosis) in stochastic volatility models is explored in Das
and Sundaram (1999). In the Heston (1993) model, where variance follows a square-root process, the term
structure of conditional skewness exhibits a hump shape. For parameter values often encountered in practice,
the point of maximum conditional skewness occurs years into the future, implying that the term structure will
often be increasing and concave over the set of option maturities actually observed.

14
rate distribution on the interest rate and volatility PCs (we also include the squared PCs
in the regressions in an attempt to take non-linearities into account). The R2 s, which are
reported in Table 3, are relatively small; in the USD market, ranging from 0.03 to 0.30 with
an average of 0.14. Then, we factor analyze the covariance matrix of the 40 time series of
regression residuals. The PCs of the residuals are, by construction, independent of those of
interest rates and volatility. There is large common variation in the regression residuals, with
the first PC explaining 78 percent of the variation in the USD market. Taken together, this
strongly indicates that there is systematic variation in skewness which is largely independent
of variation in interest rates and volatility.31
Due to space constraints, we only briefly summarize the results for conditional kurtosis of
the swap rate distributions. The swap rate distributions always exhibit excess kurtosis. For a
given swap maturity, the term structure of conditional kurtosis is hump-shaped with a peak
between 5 and 10 years – again suggestive of a stochastic volatility model with low to moderate
degrees of mean reversion. While conditional kurtosis also exhibits variation over time, it is
less systematic across the swaption matrix. For instance, in the case of the USD market, the
first PC explains only 42 percent of the variation. For this reason, in the rest of the paper, we
will mainly focus on conditional volatility and skewness.
A natural question is how sensitive our results are to liquidity issues in the swaption market.
Taking the 1-year option on the 10-year swap rate as an example, we polled major investment
banks for indicative bid-ask spreads in 2010. Typical bid-ask spreads for ATM, 50 bp OTM,
and 100 bp OTM swaptions were 1.5, 2, and 5 percent, respectively, of the mid-price.32 We
then recompute the model-free measures of volatility and skewness of the conditional 1-year
ahead distribution of the 10-year swap rate using, respectively, bid prices and ask prices. In
doing so, we assume that bid-ask spreads are 1.5 percent for strikes up to 25 bp OTM, 2.0
percent for strikes between 25 bp and 75 bp OTM, and 5 percent for strike more than 75 bp
31
If the low R2 s in Table 3 were simply due to noisy data, we would not expect to find much common variation
in the residuals. We have run several other regressions to check the robustness of the result. For instance, to
take potential time-variation in the relationship between rates, volatility, and skewness into account, we also
perform the analysis using a rolling window of 52 observations. That is, for each window we extract the first
three PCs of rates and volatility, run the regressions, and factor analyze the residuals. The average R2 s are
now somewhat larger, but we continue to find large common variation in the regression residuals.
32
In the case of ATM swaptions, the indicative bid-ask spreads that we obtained are consistent with data
from Bloomberg, where the average bid, mid, and ask prices in 2010 for a 1Y-10Y ATM swaption with a USD
100 notional were USD 7.3711, 7.4276, and 7.4821, respectively. This also gives a 1.5 percent bid-ask spread.

15
OTM.33 Panel C in Figure 1, shows the difference between volatility computed from ask and
bid prices. The spread is always small when put in relation to the level of volatility, with
the average volatility computed from bid, mid, and ask prices being 114.0, 114.8, and 115.7
bp, respectively. Panel D in Figure 1 shows the difference between skewness computed from
ask and bid prices. The spread is very small compared to the level of skewness, with the
average skewness computed from bid, mid, and ask prices being 0.1592, 0.1595, and 0.1598,
respectively. The reason is that, although OTM swaptions with larger bid-ask spreads receive
more weight in the computation of skewness than in the computation of volatility, OTM
receiver and payer swaptions enter with opposite signs in the skewness formula.

4 A dynamic term structure model for swap rates

We now propose and estimate a dynamic term structure model, which is capable of matching
the stylized facts regarding the variation in conditional swap rate moments. We use this model
to study the pricing of variance and skewness risk in addition to interest rate risk.

4.1 The model

We first set up a general model under the risk-neutral measure and then find its dynamics
under the physical and annuity measures. Subsequently, we discuss model features and the
specifications that we estimate.

4.1.1 Dynamics under the risk-neutral measure

Let P (t, T ) denote the time-t price of a zero-coupon bond maturing at time T . We assume
the following general specification for the dynamics of zero-coupon bond prices

dP (t, T )
N
X p p 
Q
= r(t)dt + σP,i (t, T ) v1 (t)dWiQ (t) + v2 (t)dW i (t) (14)
P (t, T )
i=1
p
dv1 (t) = (η1 − κ1 v1 (t) − κ12 v2 (t))dt + v1 (t)dZ Q (t) (15)
p Q
dv2 (t) = (η2 − κ21 v1 (t) − κ2 v2 (t))dt + v2 (t)dZ (t), (16)

Q Q
where WiQ (t) and W i (t), i = 1, ..., N , and Z Q (t) and Z (t) denote Wiener processes under
the risk-neutral measure. We allow for correlations between Z Q (t) and WiQ (t), i = 1, ..., N ,
33
It is likely that bid-ask spreads were wider at the height of the financial crises, while they were tighter in
the middle of the decade. For simplicity, we apply these bid-ask spreads for the entire period.

16
 Q Q
with correlations denoted by ρi . Similarly, we allow for correlations between Z (t) and W i (t),
i = 1, ..., N , and denote these correlations by ρi . This is the most general correlation structure
that preserves the tractability of the model. It follows that, by construction, both of the
volatility factors, v1 (t) and v2 (t), contain a component that is unspanned by the term structure.
Now, applying Ito’s Lemma to (4) gives the dynamics of the forward swap rate under Q
N
!
X
dSm,n (t) = − σS,i (t, Tm , Tn )σA,i (t, Tm , Tn )(v1 (t) + v2 (t)) dt +
i=1
N
X p p 
Q
σS,i (t, Tm , Tn ) v1 (t)dWiQ (t) + v2 (t)dW i (t) , (17)
i=1

where
n
X
σS,i (t, Tm , Tn ) = ζj (t)σP,i (t, Tj ) (18)
j=m
Xn
σA,i (t, Tm , Tn ) = χj (t)σP,i (t, Tj ), (19)
j=m+1

and ζj (t) and χj (t) are (stochastic) weights that are given in Appendix A.

4.1.2 Dynamics under the physical measure

The dynamics under the physical probability measure P is obtained by specifying the market
prices of risk that link the Wiener processes under Q and P. We apply the following relatively
standard specifications
p P Q p
dWiP (t) = dWiQ (t) − λi v1 (t)dt, dW i (t) = dW i (t) − λi v2 (t)dt, i = 1, ..., N (20)

and
p P Q p
dZ P (t) = dZ Q (t) − ν v1 (t)dt, dZ (t) = dZ (t) − ν v2 (t)dt. (21)

An implication of (20) is that all variation in bond risk premia is driven by the volatility state
variables. Papers like Joslin (2007) and Almeida, Graveline, and Joslin (2011) make similar
assumptions and find that, during the last 10-15 years, a significant fraction of the variation
in bond risk premia is related to volatility. Of course, a large body of empirical studies have
shown that bond risk premia also depend on the shape of the term structure such as its slope
(Campbell and Shiller (1991)) or a particular linear combination of forward rates (Cochrane
and Piazzesi (2005)). In principle, such dependence could be incorporated into the model.

17
However, for two reasons, we do not pursue such extensions here: First, the length of our
sample period severely restricts the number of market price of risk parameters that can be
reliably estimated. Second, our focus is on the effect of volatility and skewness on conditional
bond risk premia (in addition, of course, to conditional risk premia associated with exposures
to volatility and skewness). Since volatility and skewness are largely unspanned by the term
structure, our results are likely not very sensitive to letting bond risk premia also depend on
the term structure factors.
We now have the following dynamics of the forward swap rate under P
N
!
X

dSm,n (t) = − σS,i (t, Tm , Tn ) (σA,i (t, Tm , Tn ) + λi )v1 (t) + (σA,i (t, Tm , Tn ) + λi )v2 (t) dt +
i=1
N
X p p 
P
σS,i (t, Tm , Tn ) v1 (t)dWiP (t) + v2 (t)dW i (t) , (22)
i=1

where
  p
dv1 (t) = η1 − κP1 v1 (t) − κ12 v2 (t) dt + v1 (t)dZ P (t) (23)
  p P
dv2 (t) = η2 − κ21 v1 (t) − κP2 v2 (t) dt + v2 (t)dZ (t), (24)

and κP1 = κ1 − νv and κP2 = κ2 − ν v .

4.1.3 Dynamics under the annuity measure

As discussed in Section 3, for pricing swaptions it is convenient to work under the annuity
measure, A. Straightforward computations give the following dynamics of the forward swap
rate under A
N
X p p 
A
dSm,n (t) = σS,i (t, Tm , Tn ) v1 (t)dWiA (t) + v2 (t)dW i (t) , (25)
i=1

where
  p
dv1 (t) = η1 − κA v
1 1 (t) − κ v
12 2 (t) dt + v1 (t)dZ A (t) (26)
  p A
dv2 (t) = η2 − κ21 v1 (t) − κA v
2 2 (t) dt + v2 (t)dZ (t), (27)

PN PN
and κA
1 = κ1 − i=1 ρi σA,i (t, Tm , Tn ) and κA
2 = κ2 − i=1 ρi σA,i (t, Tm , Tn ). This leads to a
fast and accurate Fourier-based pricing formula for swaptions derived in Appendix A.

18
4.1.4 Model features

The model has the potential to match the main stylized facts regarding conditional moments
of the swap rate distributions reported in Section 3. A specification with only one volatility
process could capture the cross-sectional variation in average conditional swap rate volatil-
ity and skewness. The former requires specifying the zero-coupon bond volatility functions,
σP,i (t, T ), such that the intermediate part of the term structure is the most volatile. The
latter is a consequence of a mean-reverting volatility processes combined with the possibility
of correlation between innovations to the term structure and volatility. Clearly, a specification
with only one volatility process would also be consistent with variation in conditional swap
rate volatility over time, including the unspanned stochastic volatility phenomenon. However,
it would not be consistent with two important properties of conditional swap rate skewness;
first, its significant variation over time – in particular the switch in the sign of skewness – and,
second, the fact that skewness is largely unspanned by rates and volatility. In order to match
these two properties, we need at least two volatility processes along with certain parameter
restrictions.
To illustrate how a specification with two volatility processes could be consistent with the
first property, note that the instantaneous variance of the forward swap rate, Sm,n (t), is given
by
N
!
1 X
2
V ar (dSm,n (t)) = σS,i v(t), (28)
dt
i=1

where

v(t) = v1 (t) + v2 (t). (29)

This implies that the instantaneous correlation between innovations to the forward swap rate
and its variance is given by
N
X σS,i
Ωm,n (t) = qP ((1 − w(t))ρi + w(t)ρi ) , (30)
N 2
i=1 i=1 σS,i

where
v2 (t)
w(t) = . (31)
v1 (t) + v2 (t)
In words, the instantaneous correlation is the sum of weighted averages of ρi and ρi . The
weight on ρi , w(t), is given by the ratio of the second volatility factor to the sum of the two

19
volatility factors and obviously lies between zero and one. The implication is that Ωm,n (t)
is stochastic and may switch sign if ρi and ρi have opposite signs. Consequently, conditional
swap rate skewness, which depends on Ωm,n (t), is also stochastic and subject to possible sign
change.34
In the unrestricted specification, however, skewness will be spanned by volatility. The
reason is that given dissimilar drift parameters, v1 (t) and v2 (t) will impact the term structure
of volatility differently, in which case one can back out v1 (t) and v2 (t), and therefore w(t),
by observing conditional volatility at different horizons. To generate unspanned stochastic
skewness, we impose that the drifts of the volatility processes are identical, so that the processes
only differ in terms of the their correlations with the term structure factors. In other words,
we impose η2 = η1 = η, κ2 = κ1 = κ, and κ21 = κ12 = 0, which in the following are called
the unspanned stochastic skewness (USS) restrictions. Then, shocks to v1 (t) and v2 (t) have
the same effect on conditional volatility and one cannot disentangle the two state variables by
observing conditional volatility alone.35

4.1.5 Level, slope, and curvature shocks

It is well known that the term structure of interest rates is driven by three factors, and,
accordingly, we set N = 3 in all specifications. For the bond price volatility functions in (14),
we note that we can equally well specify the volatility functions for instantaneous forward
RT
rates, since the two are related by σP,i (t, T ) = − t σf,i (t, u)du. As it is generally easier to
relate to interest rate volatility than bond price volatility, we show both. The specifications
we use are
α1


σf,1 (t, T ) = α1 e−ξ(T −t) ⇒ σP,1 (t, T ) = ξ e−ξ(T −t) − 1
α2


σf,2 (t, T ) = α2 e−γ(T −t) ⇒ σP,2 (t, T ) = γ e−γ(T −t) − 1
α3

α3
σf,3 (t, T ) = α3 (T − t)e−γ(T −t) ⇒ σP,3 (t, T ) = γ2
e−γ(T −t) − 1 + γ (T − t)e−γ(T −t) .

34
It is well known that in a stochastic volatility setting, conditional skewness depends on the correlation
between the underlying and its variance, see e.g. Das and Sundaram (1999).
35
This statement is true, when we are operating under the risk-neutral measure. However, the unspanned
stochastic skewness regressions, discussed in Section 3.2, were done with moment computed under the annuity
measure. Under this measure, the mean-reversion parameters of v1 (t) and v2 (t) differ, even if they are identical
under the risk-neutral measure. Nevertheless, for plausible parameter values, and certainly for the estimated
parameter values reported later, the difference between κA1 and κA2 is very small, and for all practical purposes,
unspanned stochastic skewness also exist under the annuity measure.

20
The second and third are the “slope” and “curvature” factor loadings proposed by Nelson and
Siegel (1987), while the first becomes their “level” factor loading in the limit ξ → 0. These
factor loadings are popular in the term structure literature as they parsimoniously capture
the predominant shocks to the term structure. The reason behind modifying the first factor
loading relative to Nelson and Siegel (1987) is that this allows us to express the dynamics
of the term structure in terms of a finite dimensional affine state vector, making it possible
to estimate the model with well-established techniques from the vast affine term structure
literature (in reality, ξ is estimated close to zero, which implies that the first factor still acts
much like a level factor). The affine representation of the model can be derived along the lines
of Trolle and Schwartz (2009) and is given in the online appendix.
In principle, the model is time-inhomogeneous and fits the initial term structure by con-
struction. For the purpose of estimation, it is more convenient to work with the model’s
time-homogeneous counterpart. We therefore assume that the initial forward rate curve is flat
and equal to a constant, ϕ, which is an additional parameter of the model. One can show that
ϕ equals the infinite-maturity forward rate.

4.1.6 Model specifications

In the following, we consider three model specifications. In the SV 1 specification, we allow

for only one volatility factor, v1 (t). In the SV 2U SS specification, we allow for both volatility
factors, v1 (t) and v2 (t), but we impose the unspanned stochastic skewness restrictions. Finally,
in the SV 2 specification, we put no restrictions on the two volatility factors. In the last spec-
ification, admissability requires κ12 ≤ 0 and κ21 ≤ 0, but preliminary estimations show that
these cross-terms are close to zero. We therefore restrict both to zero in the final estimation,
which has negligible impact on the pricing performance and the paths of the state variables.
Even with the parsimonious market price of risk specifications outlined in Section 4.1.2,
there are 8 market price of risk parameters to be estimated in the SV 2U SS and SV 2 specifica-
tions (7 in the SV 1 specification), which is a stretch given the length of the sample period. To
reduce the number of parameters further, we follow Cochrane and Piazzesi (2008) in assuming
that bond risk premia only arise from exposure to shocks in the “level” factor (we refer to
their paper for an indepth discussion of the evidence supporting this assumption). In other
words, we impose λ2 = λ2 = λ3 = λ3 = 0.
Then, SV 1, SV 2U SS , and SV 2 have 11 (2), 14 (4), and 16 (4) risk-neutral parameters

21
(market price of risk parameters), respectively. This is well within what is often encountered
in the empirical term structure literature, and given the vast amount data at our disposal,
even the most general specification is quite tightly parameterized.

4.1.7 Maximum-likelihood estimation

We estimate the three specifications on all available swap rates and swaptions using maximum-
likelihood in conjunction with Kalman filtering. Critical to estimating the model on all swap-
tions across time, and across all option expiries, swap maturities, and strikes, is the existence
of an efficient pricing formula. Due to the non-linearities in the relationship between obser-
vations and state variables, we apply the non-linear unscented Kalman filter.36 Details are
provided in Appendix C.

4.2 Results

4.2.1 Maximum-likelihood estimates

Table 4 displays parameter estimates with asymptotic standard errors are in parentheses. In
the SV 1 specification, the risk-neutral speed of mean-reversion in volatility is moderate, and
innovations to volatility and the term structure factors are positively correlated, which is
consistent with the mostly positive skewness in swap rate distributions on average.
In the SV 2U SS specification, ρi and ρi have opposite signs for all i, which is consistent with
stochastic skewness that may switch sign. When v1 (t) is high relative to v2 (t), the swap rate
distributions will be skewed towards lower interest rates, whereas when v1 (t) is low relative to
v2 (t), the swap rate distributions will be skewed towards higher interest rates. The risk-neutral
speed of mean-reversion in volatility is about the same as for the SV 1 specification.
Compared to SV 2U SS , in the SV 2 specification, v1 (t) has faster risk-neutral mean-reversion,
while v2 (t) has slower risk-neutral mean-reversion suggesting that the former captures rela-
tively transitory shocks to volatility, while the latter captures more persistent shocks to volatil-
ity. Also the correlation parameters ρi and ρi become more similar for all i, suggesting that
36
The unscented Kalman filter has gained popularity in recent years as an alternative to the more standard
extended Kalman filter. Christoffersen et al. (2009) perform an extensive Monte Carlo experiment, which
shows that the unscented Kalman filter significantly outperforms the extended Kalman filter in the context of
estimating dynamic term structure models with swap rates. Their results most likely carry over to our context,
where swaptions are also used in the estimation.

22
the SV 2 specification will exhibit less variation in skewness. This hints at a tension within
the model of matching volatility and skewness. We return to this issue below.
Almost all market price of risk parameters are negative and the majority are statistically
significant. The implications for risk premia are explored further in Section 4.2.3.

4.2.2 Specification analysis

For each specification, from the filtered state variables, we compute the fitted values of swap
rates and swaption prices. From the fitted swaption prices, we compute implied volatilities.
We also compute volatility, skewness, and kurtosis of the future swap rate distributions (under
A) using the fitted swaption prices and the same interpolation/extrapolation scheme as in
Section 3. For swap rates, implied swaption volatilities, and swap rate moments, we then
compute root mean squared errors (RMSEs) on each day, producing time series of RMSEs.
Means of RMSEs and comparisons between model specifications are displayed in Table 5.
The fit to swap rates is displayed in the first column of Table 5. The fit is very good (a
mean RMSE of about 5 bp in the USD market) and similar for the different model specifi-
cations, which is not surprising, since all specifications have three term structure factors and
similar estimates for the factor loadings. However, the results do run counter to Joslin (2007),
who argues that imposing unspanned stochastic volatility in affine term structure models is
detrimental to their ability to fit interest rates. For instance, in the case of his A1 (4) model, he
finds that imposing the USV restrictions causes in-sample RMSEs of interest rates to double
or triple. This occurs despite the fact that volatility is nearly unspanned in the correspond-
ing unrestricted model. Our SV 1 specification is conceptually similar to Joslin’s A1 (4)U SV
specification (both have three term structure factors and one unspanned volatility factor) yet
our specification retains a close fit to the term structure. This shows that our framework,
which has a different affine structure than Joslin’s, is more flexible in terms of accommodating
unspanned stochastic volatility (and skewness).
The fit to implied swaption volatilities is displayed in the second column of Table 5. The
SV 2U SS specification improves significantly on the SV 1 specification with the mean RMSE
decreasing from 5.44 bp to 4.58 bp in the USD market. The unrestricted SV 2 specification
performs even better with the mean RMSE decreasing further to 3.64 bp.
The last three columns of Table 5 shows the fit to swap rate moments and provides insights
into the sources of the differing pricing performances. Comparing the SV 1 and SV 2U SS speci-

23
fications shows that they have a very similar fit to volatility, but that the SV 2U SS specification
has a significantly better fit to skewness with the mean RMSE decreasing from 0.21 to 0.06 in
the USD market.37 This is not surprising since the SV 2U SS specification essentially has one
factor, v(t), driving volatility and one factor, w(t), driving skewness. Comparing the SV 2U SS
and SV 2 specifications shows that the SV 2U SS specification has a significantly better fit to
volatility with the mean RMSE decreasing from 4.83 bp to 3.04 bp, but a significantly worse
fit to skewness with the mean RMSE increasing from 0.06 to 0.16. This illustrates a tension
within the model between fitting volatility and skewness. In essence, swap rate volatility is
driven by multiple principal components and it appears that, from a pricing perspective, cap-
turing the second principal component of swap rate volatility is more important than capturing
the first principal component of swap rate skewness.
To visualize the fit to swap rate moments, Figure 2 displays time series of model-implied
conditional volatility and skewness of the 1-year ahead distribution of the USD 10-year swap
rate (again under A). While all model specifications have very similar fit to conditional volatil-
ity, they differ markedly in their fit to conditional skewness. The SV 1 specification has too
little variation in skewness, and model-implied skewness is in fact negatively correlated with
actual skewness. The SV 2 specification performs somewhat better, although actual and model-
implied skewness still deviate significantly. In contrast, skewness in the SV 2U SS specification
tracks actual skewness closely, including the switches between positive and negative skewness.
Figure 3 shows actual and fitted time series of the USD implied volatility smile of the 1-
year option on 10-year swap rate (one out of the 40 time series of swaption smiles in the data).
The pattern is similar to that of Figure 2. While the SV 1 and SV 2 specifications generate too
little variation in the slope of the implied volatility smile, the SV 2U SS specification matches
it quite closely.
The aggregate RMSEs reported in Table 5 have the potential to mask problems with
fitting individual swap rates, implied volatilities, or moments. In the online appendix, we
provide a series of tables with individual RMSEs for the SV 2U SS specification. In general, the
individual RMSEs are relatively close to the aggregate RMSEs. However, as for virtually all
term structure models, the fit to swap rates does deteriorate somewhat at both ends of the
maturity range. For instance, in the USD market, the RMSE is 12.1 bp for 6 month LIBOR
and 9.2 bp for the 40 yr swap rate. Similarly, the fit to implied volatilities and swap rate
37
Note that the mean kurtosis RMSE is also significantly lower for the SV 2U SS specification.

24
moments declines at extreme combinations of swap maturities and option expiries.
To see whether the specifications are able to generate the degree of unspanned stochastic
skewness reported in Table 3, we conduct the following limited simulation exercise for the
USD market (the EUR market gives similar results). For each specification, we simulate 100
samples of the same size as our original data set. For each sample, we then replicate the
unspanned stochastic skewness regressions described in Section 3.2. Table 6 shows the mean
(along with its standard error in parentheses) of the R2 s across the 100 samples. For the SV 1
and SV 2 specifications, the mean R2 s are much higher than the numbers reported in Table
3.38 In contrast, the SV 2U SS specification generates mean R2 s that are close to those observed
in the data. In fact, most of the R2 s in the data are within the 95 percent confidence intervals
of the model generated R2 s, strongly suggesting that the SV 2U SS specification does indeed
capture the degree of unspanned stochastic skewness observed in the data.39

4.2.3 Pricing of risk

The specification analysis reveals that for learning about overall variation in volatility and
skewness, the SV 2U SS specification is the most informative, even if the SV 2 specification has
better pricing performance. Consequently, we will use the SV 2U SS specification to study the
pricing of risk. We express the compensation that market participants require for bearing
interest rate, variance, and skewness risk in terms of v(t) and w(t), which drive variance and
skewness, respectively, as described in Section 4.1.4.40

Pricing of interest rate risk Consider a security with a pure exposure to interest rate
risk. Let f (St ) denote its price. Appendix B shows that the instantaneous Sharpe ratio is
38
Despite the fact that skewness is fully spanned in these two specifications, the R2 s remain less than one.
This is due to the fact that the regression does not sufficiently account for non-linearities in the relationship
between rates, volatility, and skewness.
39
A similar analysis shows that all the specifications more or less capture the degree of unspanned stochastic
volatility observed in the data.
40
While the relationship between v(t) and swap rate variance is clear, the relationship between w(t) and swap
rate skewness is more indirect. In the case of the SV 2U SS specification, w(t) has a correlation of 0.984 with the
first principal component of fitted conditional swap rate skewness across the swaption matrix. This confirms
that w(t) does indeed capture skewness risk. See the online appendix for further results.

25
given by
 
1 df (St ) σS,1 p

SRIR = sign qP v(t) (1 − w(t))λ1 + w(t)λ1 . (32)
f (St ) dSt N 2
i=1 σS,i

Table 4 shows that λ1 and λ1 are estimated to be negative. Since σS,1 > 0, and recalling that
w(t) lies between zero and one, the price of interest rate risk is negative. For bonds (or swaps
df (St )
where one receives the fixed rate and pays the floating rate) we have dSt < 0, and bond risk
premia are consequently positive.
For a given level of interest rate skewness, the Sharpe ratio on a bond is increasing in interest
rate volatility – a finding that has also been reported in other studies. Our contribution is to
investigate the effect of skewness. Since λ1 < λ1 < 0, for a given level of volatility, the Sharpe
ratio is increasing in skewness.
A number of paper including Cochrane and Piazzesi (2008), Ludvigson and Ng (2009),
Joslin, Priebsch, and Singleton (2010), Duffee (2011), and Buraschi and Whelan (2011) have
shown that a significant component of the compensation for bearing interest rate risk is un-
spanned by the term structure. In these papers, the unspanned component is treated as a
latent factor or is attributed to unspanned macro risks. In our model, there is also an un-
spanned component in bond risk premia, but the source is different. It arises because both
volatility and skewness contain a component that is unspanned by the term structure.41

Pricing of variance risk Consider a security with a pure exposure to variance risk, i.e.
v(t). This could be a swaption straddle, which is long both an ATM payer swaption and an
ATM receiver swaption, or a variance swap. Let f (vt ) denote the security’s price. Appendix B
shows that its instantaneous Sharpe ratio is given by
 
1 df (vt ) p
SRV ar = sign v(t) ((1 − w(t))ν + w(t)ν) . (33)
f (vt ) dvt
Table 4 shows that ν and ν are estimated to be negative, which implies that the price of
variance risk is negative. This suggests that investors dislike states with high interest rate
(and bond return) variance in that they are willing to accept a negative expected return on a
security that pays off in such states. This is consistent with results in Duarte, Longstaff, and
Yu (2007) and Joslin (2007) as well as a large number of papers on variance risk premia in
equity markets.
41
Joslin (2007) makes a similar observation in a model with (nearly unspanned) volatility risk but no distinct
skewness risk.

26
 For a given level of skewness, the Sharpe ratio becomes more negative as volatility increases,
which makes intuitive sense; as volatility increases, the volatility of variance (i.e. the quantity
of variance risk) also increases, and it is only natural that the compensation for bearing variance
risk increases. The more novel result is that, since ν < ν < 0, for a given level of volatility,
the Sharpe ratio becomes more negative as skewness increases.

Pricing of skewness risk This issue has not been explored in the fixed income literature.
Consider a security with a pure exposure to skewness risk, i.e. w(t). This could be a swaption
risk-reversal position, which is long an OTM payer swaption and short an OTM receiver
swaption. Let f (wt ) denote the security’s price. Appendix B shows that its instantaneous
Sharpe ratio is given by
 
1 df (wt ) p p p
SRSkew = sign v(t) 1 − w(t) w(t) (ν − ν) . (34)
f (wt ) dwt

The sign of SRSkew depends on the difference between the market prices of risk associated
with the volatility factors that generate positive and negative skewness, respectively (in a risk
reversal, one is essentially long v2 (t) and short v1 (t)). Since ν < ν < 0, the price of skewness
risk is negative. This suggests that investors dislike states with more positively skewed interest
rates distributions in that they are willing to accept a negative expected return on a security
that pays off in such states. Since those are states with more negatively skewed bond return
distributions, this result is consistent with Arditti (1967) and Scott and Horvath (1980), who
show that, under very general assumptions, investors have a preference for positive skewness
in return distributions
For a given level of skewness, the Sharpe ratio becomes more negative as volatility increases.
On the other hand, for a given level of volatility, the Sharpe ratio is a U-shaped function of
skewness; that is, the Sharpe ratio is zero for w(t) = 0 and w(t) = 1, and most negative
for w(t) = 1/2. This is quite intuitive; from the process for w(t) given in Appendix B, it is
apparent that the volatility of skewness (i.e. the quantity of skewness risk) is zero for w(t) = 0
and w(t) = 1, and at its maximum for w(t) = 1/2. It makes sense that the price of skewness
risk follows this pattern.

Magnitudes In Figure 4, the solid lines show how SRIR , SRV ar , and SRSkew vary with
p
volatility and skewness. The panels to the left show the dependence on v(t), when w(t)
equals its sample mean, while the panels to the right show the dependence on w(t), when

27
p
v(t) equals its sample mean. In the case of SRIR , we show the Sharpe ratio arising from
p
exposure to the 10 year swap rate. Clearly, both v(t) and w(t) have a significant impact on
Sharpe ratios.
Figure 5 shows how SRIR , SRV ar , and SRSkew vary over the sample. Their sample averages
are -0.34, -0.38, and -0.20, respectively. During most of the sample period, the Sharpe ratios
are positively correlated. However, in the crisis period, we observe an interesting divergence.
The Sharpe ratio on skewness risk becomes small (since w(t) is close to one), while the Sharpe
ratios on interest rate and variance risk become large in absolute terms and highly volatile
p
(since v(t) is high).

5 Fundamental drivers of the swap rate distributions

In the previous section, we treated volatility and skewness as latent variables. Now we in-
vestigate if there is a link to fundamentals. We are primarily interested in the effects of
perceived macroeconomic uncertainty, but we also investigate the influence of other factors
such as volatility and skewness of equity index returns, market-wide illiquidity, and refinanc-
ing activity in the MBS market. Again, we focus on the SV 2U SS specification.

5.1 Explanatory variables

Perceived macroeconomic uncertainty There is ample evidence that most central banks
set short-term interest rates in response to growth and inflation (see, e.g., the voluminous
literature on Taylor rules) although the relative importance of growth and inflation in central
banks’ objective functions differ. Long-term interest rates, being risk-adjusted expectations of
future short rates, therefore reflect expectations about future growth and inflation.42 Following
this reasoning, our hypothesis is that volatility and skewness of long-term interest rates is
related to perceived uncertainty about future growth and inflation. Specifically, we would
expect volatility to increase, when the outlook for future growth and inflation becomes more
uncertain. Similarly, we would expect skewness to increase, when the outlook for future growth
and inflation becomes more skewed towards higher growth and inflation rates.
42
Indeed, in recent years, there has been significant interest in linking the term structure to such macroe-
conomic variables; see, among others, Ang and Piazzesi (2003), Gallmeyer, Hollifield, and Zin (2005), Ang,
Piazzesi, and Wei (2006), Smith and Taylor (2009), Bekaert, Cho, and Moreno (2010), Bikbov and Cher-
nov (2010), Chun (2010), and Joslin, Priebsch, and Singleton (2010).

28
 In testing this hypothesis, we face the challenge of quantifying uncertainty. For this pur-
pose, we use the quarterly survey of professional forecasters (SPF) conducted in the US by the
Federal Reserve Bank of Philadelphia and in the Eurozone by the ECB. The SPF is unique,
because participants are asked to assign a probability distribution to their forecasts for real
GDP growth and inflation. We aggregate the probability distributions of the individual re-
spondents, and compute dispersion (i.e., standard deviation) and skewness of the aggregate
distributions of future real GDP growth and inflation.43 Since the aggregate probability dis-
tributions take both individual uncertainty and disagreement among survey participants into
account, it provides a better measure of uncertainty than commonly used measures based
purely on disagreement, see Zarnowitz and Lambros (1987).44

Moments of equity index returns Numerous papers document that equity and fixed-
income markets are interconnected. We, therefore, investigate the extent to which moments
of equity index returns have an impact on moments of swap rates. Specifically, we consider
the S&P 500 index in the USD market and the Eurostoxx 50 in the EUR market and compute
volatility and skewness of the risk-neutral return distributions in a model independent way,
using the formulas in Bakshi, Kapadia, and Madan (2003). As in Section 3.1, this involves
integrating over options with different strikes, although the formulas are slightly different. We
obtain risk-neutral moments for return horizons corresponding to the option expiries that are
traded, and we use the first principal component of volatility and skewness, respectively, in
the regressions.45 Our measure of volatility in the USD market has a very high correlation
(above 0.98) with the VIX index, which has been used in numerous studies as a proxy for
overall financial market volatility, or as a sentiment indicator.
43
Participants are asked to provide probability distributions for the current and following calendar year. We
follow Bekaert and Engstrom (2009) in weighting the probability distributions so as to maintain a 1-year-ahead
forecast horizon. Another issue is that in forming their probability distributions, respondents are asked to attach
probabilities to the outcome being in specific ranges. When computing moments of the aggregate distributions,
we assume that the probability for a given range relates to the mid-point of that range.
44
One can show that the variance of the aggregate distribution is equal to the average variance of the
individual distributions plus the variance of the distribution of individual point estimates (i.e. disagreement).
The skewness of the aggregate distribution depends in a less straightforward way on the average skewness of
the individual distributions and the skewness of the distribution of individual point estimates; see, e.g., Garcia
and Manzanares (2007).
45
For the S&P 500 index, the first principal component explains 98 (92) percent of the variation in volatility
(skewness) across the option expiries. For the Eurostoxx 50, the numbers are similar.

29
Market-wide illiquidity A number of papers show that liquidity affects derivatives prices.46
Unfortunately, our data set does not include bid-ask spreads or other measures, such as market
depth, that can be used to construct liquidity measures at the level of individual contracts.
Instead, we investigate the effect of liquidity at the market-wide level. As a proxy for market-
wide illiquidity, we use the spread between the 3-month overnight index swap (OIS) rate and
the 3-month Treasury bill rate (for the EUR market, we use the German counterpart to the
3-month Treasury bill). An OIS is a measure the expected average overnight rate during the
life of the swap and is virtually free of credit and counterparty risk. Treasury bills are arguably
the most liquid debt market instrument available. Therefore, the spread is a fairly clean proxy
for illiquidity, as also observed by Krishnamurthy (2010).47

Refinancing activity Several papers find that derivatives prices are affected by supply and
demand.48 In the swaption market, dealers absorb or redistribute supply and demand for
interest rate options. From a dealer perspective, much of the option supply emanates from
issuance of callable debt by financial institutions and large corporations. A significant part of
this debt is swapped into floating rate payments with the embedded optionality often passed
on to dealers. In the USD market, an important demand for options comes from investors in
MBSs, who actively hedge the negative convexity risk stemming from the prepayment options
embedded in fixed-rate mortgages.49 Since there is no quantitative information on the flows in
the swaption market, it is difficult to estimate the extent to which demand and supply affects
46
For instance, in the related market for caps and floors, Deuskar, Gupta, and Subrahmanyam (2010) find
that liquidity, as proxied by bid-ask spreads, impacts prices. Other studies include Cetin et al. (2006) and
Bongaerts, de Jong, and Driessen (2011), who provide theory and evidence in support for liquidity having an
impact on the pricing of stock options and credit default swaps, respectively.
47
In the USD market, an OIS is referenced to the overnight federal funds rate, while in the EUR market, it
is referenced to the euro overnight index average (EONIA) rate. The swap contract itself is fully collateralized,
making credit and counterparty risk negligible. As an alternative illiquidity proxy, we have used the spread
between yields on off-the-run and on-the-run government bonds of comparable maturities but found similar
results.
48
For instance, Bollen and Whaley (2004) and Gârleanu, Pedersen, and Poteshman (2009) find demand effects
in the pricing of equity options.
49
Active hedgers include Fannie Mae and Freddie Mac as well as mortgage hedge funds. Fannie Mae and
Freddie Mac were placed into conservatorship in September 2008. However, they were allowed to increase (and
actively hedge) their total portfolio of retained mortgages to USD 1.7tn by 2009. Therefore, they continue to
play a role in the interest rate derivatives market.

30
pricing. Like previous papers, we focus on the hedging of prepayment risk, using the Mortgage
Bankers Association’s Refinancing Index (a weekly measure of refinancing activity) as a proxy
for swaption demand by active hedgers.50

5.2 Regression specification

In specifying the regression model, we face the issue that our proxies for macroeconomic
uncertainty are only available at a quarterly frequency, while the remaining variables are
available at a weekly frequency. To make use of all the information in the data, we run
MIDAS-type regressions51 of the following form

yt = β0 + β1 f (θ, τ )GDP voltq + β2 f (θ, τ )GDP skewtq + β3 f (θ, τ )IN F voltq +

β4 f (θ, τ )IN F skewtq + β5 EQvolt + β6 EQskewt + β7 ILLIQt + β8 REF It + ǫt ,(35)

where τ = t − tq is the time between the weekly observation at t and the most recent quarterly
p
observation at tq , and yt is either v(t) or w(t), which drive instantaneous volatility and
skewness, respectively. The function f (θ, τ ) weighs the quarterly observations according to
their distance from t. We assume the following simple functional form f (θ, τ ) = exp(−θτ );
i.e., the weights are exponentially declining in τ .52 We also assume that the same weighing
function applies to all four quarterly series. The MIDAS regression model is estimated by
non-linear least squares.

5.3 Results

Tables 7 displays the results. Consider first the effect of perceived macroeconomic uncertainty.
In the USD market, interest rate volatility and skewness primarily depend on the characteristics
of the GDP probability distribution. Volatility depends significantly and positively on the
50
Li and Zhao (2009) use the same measure as a proxy for option demand in the cap/floor market, while
Duarte (2008) uses a more refined measure of the average refinancing incentive in the mortgage universe as a
proxy for swaption demand. Both papers find that hedging pressures have a significant impact on option prices.
51
MIxed DAta Sampling regressions have become popular in the econometrics literature following the initial
publications by Ghysels, Santa Clara, and Valkanov (2005, 2006). Typically, a low-frequency variable is regressed
on higher-frequency variables, but the methodology is equally applicable when, as in our case, a high-frequency
variable is regressed on lower-frequency variables.
52
We have experimented with other weighing functions suggested in the MIDAS literature, but found the
results quite robust to the choice of (sensible) weighing scheme.

31
dispersion of the GDP distribution, while skewness depends positively and significantly on the
skewness (and the dispersion) of the GDP distribution. That is, volatility tends to increase,
when the perceived uncertainty about future real GDP growth increases, while skewness tends
to increase, when the risk is perceived to be more skewed to the upside.
In contrast, in the EUR market, interest rate volatility and skewness are more related to
the characteristics of the inflation probability distribution. Volatility depends significantly and
positively on the dispersion of the inflation distribution (and, less strongly, on the skewness of
the GDP distribution), while skewness depends positively and significantly on the skewness of
the inflation (and GDP) distributions.
Overall, the results are consistent with our hypothesis about the effect of perceived macroe-
conomic uncertainty. However, it is striking that GDP growth beliefs are the most important
determinants in the USD market, while inflation beliefs are the most important determinants
in the EUR market. One likely explanation is differences in monetary policy objectives in the
two currency areas. The primary policy goal of the European Central Bank is to maintain
price stability, whereas the Federal Reserve has a dual mandate of maximum employment
and price stability, leading it to place relatively more emphasis on expectations for real GDP
growth when setting interest rates.53
Next, consider the effect of the characteristics of the equity index return distributions,
market-wide illiquidity, and refinancing activity. In both markets, equity return volatility has
a significantly positive effect on interest rate volatility, while equity return skewness has a
significantly positive effect on interest rate skewness. These results underscore the integration
between equity and fixed income markets, and is consistent with equities and interest rates
being positively correlated during the sample period.
Market-wide illiquidity has a significantly positive effect on volatility, but little effect on
skewness. Note, however, that there is a relatively high correlation between our illiquidity
proxy and equity market volatility. Leaving out equity market volatility from the regressions
increases the importance of market-wide illiquidity for interest rate volatility.
53
The different policy objectives of the two central banks were clearly illustrated on July 2, 2008, when
the ECB, despite signs of financial stress and weakening growth, raised its benchmark rate in a bid to attack
inflation. At that point, the Federal Reserve had already reduced its benchmark rate significantly to support
economic growth in the face of the financial crisis. A similar situation occurred in the Spring of 2011, after the
end of our sample period, when the ECB started tightening monetary policy in response to inflation, in spite
of the European sovereign debt crisis.

32
 Finally, in the USD market, refinancing activity has a significantly positive effect on volatil-
ity. Compared with the results in Duarte (2008) (and Li and Zhao (2009) for the cap/floor
market), the effect of refinancing activity is relatively modest. This may be due to differences
in data, sample period, and methodology, but may also, in part, reflect the Federal Reserve’s
USD 1.25 trillion program to purchase MBSs, initiated in late 2008. Since the Federal Reserve
does not engage in convexity hedging, its massive involvement in the MBS market reduces the
effect of refinancing activity on the swaption market.54 Refinancing activity, which is a U.S.
phenomenon, has virtually no effect on the swap rate distributions in the EUR market.

6 Robustness checks

Our results in Sections 4 and 5 are clearly model dependent. In this final section, we consider
the robustness of our results to alternative specifications for market prices of risk, the term
structure model, and the regression model.

6.1 Alternative market price of risk specification

We have experimented with the following, more flexible, market price of risk specification,
first suggested by Cheredito, Filipovic, and Kimmel (2007) and Collin-Dufresne, Goldstein,
and Jones (2009). If we continue to assume that bond risk premia only reflect compensation
for exposure to shocks in the “level” factor, (20) extends to

λ0,1 + λ1 v1 (t) P Q λ0,1 + λ1 v2 (t)

dW1P (t) = dW1Q (t) − p dt, dW 1 (t) = dW 1 (t) − p dt, (36)
v1 (t) v2 (t)

while (21) extends to

ν0 + νv1 (t) P Q ν 0 + νv2 (t)

dZ P (t) = dZ Q (t) − p dt, dZ (t) = dZ (t) − p dt. (37)
v1 (t) v2 (t)

We reestimate the SV 2U SS specification with these extended market prices of risk, and call
the resulting specification by SV 2U SS,EX .
The estimates of the new market price of risk parameters are relatively close to zero and
rather imprecisely estimated.55 λ0,1 and ν0 are estimated to be positive, while λ0,1 and ν 0 are
54
This appear to be a consensus view among market participants. For instance, in a recent research report,
Barclays Capital (2010) concludes that “...the Fed portfolio serves as a dampener on the option bid from MBS
portfolios”.
55
The parameter estimates for this specification are reported in Table F.6 in the online appendix.

33
estimated to be negative. The estimates of the remaining market price of risk parameters are
mostly smaller in absolute terms than in the SV 2U SS specification.56
To evaluate the effect on the pricing of risk, again let f (St ), f (vt ), and f (wt ) denote prices
of securities with pure exposures to interest rate, variance, and skewness risk, respectively.
With the extended market price of risk specification, their Sharpe ratios are
  !
1 df (St ) σS,1 λ0,1 + λ0,1 p

SRIR = sign qP p + v(t) (1 − w(t))λ1 + w(t)λ1
f (St ) dSt N 2 v(t)
i=1 σS,i
  !
1 df (vt ) ν0 + ν 0 p
SRV ar = sign p + v(t) ((1 − w(t))ν + w(t)ν)
f (vt ) dvt v(t)
  !
1 df (wt ) (1 − w(t))ν 0 − w(t)ν0 p p p
SRSkew = sign p p p + v(t) 1 − w(t) w(t) (ν − ν) .
f (wt ) dwt v(t) 1 − w(t) w(t)

The dotted lines in Figure 4 show that the sensitivities of the Sharpe ratios to variations in
volatility and skewness are generally slightly smaller than in the SV 2U SS specification, even
if the overall magnitude of the Sharpe ratios are roughly the same.

6.2 Alternative term structure model specification

The analysis in Section 4.2 reveals a tension between matching volatility and skewness. As a
way of alleviating this tension, we extend the SV 2U SS specification with an additional process,
which captures low-frequency variation in v1 (t) and v2 (t). Specifically, we set η1 = η2 = η(t),
where η(t) follows the stochastic process
p
dη(t) = (η − κη η(t)) dt + ση η(t)dZeQ (t), (38)

Q
and where ZeQ (t) is another Wiener process uncorrelated with Z Q (t) and Z (t) (as well as
Q
with WiQ (t) and W i (t), i = 1, ..., N ).57 We denote this specification by SV 3U SS .
The estimated parameters of the v1 (t) and v2 (t) processes, including the volatility parame-
ters, are similar to those in the SV 2U SS specification, except that the speed of mean-reversion
is somewhat faster, while the process for η(t) is estimated to have slower mean-reversion.
56
For (36) and (37) to be consistent with absence of arbitrage, we need to impose the Feller conditions under
P and Q. As these are typically binding, we experience a small deterioration in the fit to swaptions; see, Table
F.7 in the online appendix.
57
Via a straightforward extension of the results in Appendix A, swaptions can also be priced quasi-analytically
in this specification. To get the dynamics for η(t) under the physical probability measure P, we apply the
p
eP (t) = dZ
following market price of risk specification: dZ eQ (t) − νe η(t)dt.

34
However, the market price of risk parameters are roughly the same as those in the SV 2U SS
specification.58 As expected, the SV 3U SS specification succeeds in improving the fit to con-
ditional volatility, while preserving the fit to conditional skewness.59 It also preserves the
unspanned nature of skewness risk with the unspanned stochastic skewness regressions pro-
ducing R2 s only slightly larger than those of the SV 2U SS specification.
However, although the fit is better, the pricing of interest rate, variance, and skewness risk
is similar to that of the SV 2U SS specification; see the dashed-dotted lines in Figure 4.

6.3 Alternative regression specification

As an alternative to the MIDAS regressions, we run the regressions in quarterly differences,

i.e.

∆ytq = β0 + β1 ∆GDP voltq + β2 ∆GDP skewtq + β3 ∆IN F voltq + β4 ∆IN F skewtq +

β5 ∆EQvoltq + β6 ∆EQskewtq + β7 ∆ILLIQtq + β8 ∆REF Itq + ǫtq , (39)

p
where ∆ytq is the quarterly change in either v(t) or w(t). While this entails discarding
information, it may be more robust than the MIDAS specification. The results, which are
reported in Table 8, are generally consistent with those obtained from the MIDAS regressions.
In particular, we continue to find that volatility and skewness are related to the perceived
uncertainty about the macroeconomy, with GDP growth beliefs being the most important
factor in the USD market and inflation beliefs being the most important factor in the EUR
market.

7 Conclusion

In this paper, we use a comprehensive database of inter-dealer quotes to conduct the first
empirical analysis of the dynamics of the swaption cube.
58
The parameter estimates for this specification are also reported in Table F.6 in the online appendix. The
new market price of risk parameter, νe, is estimated to be negative. However, this parameter does not enter the
expressions for instantaneous Sharpe ratios in Section 4.2.3.
59
See Table F.7 in the online appendix. Specifically, compared to the SV 2U SS specification, in the USD
market, the mean RMSE of swaption implied volatilities decreases from 4.58 to 3.43, with the mean RMSE of
conditional swap rate volatilities decreasing from 4.83 to 3.20, and the mean RMSE of conditional swap rate
skewness essentially unchanged.

35
 We first analyze the swaption cube from a model independent perspective. We use the
fact that for a given swap maturity and option expiry, one can compute conditional swap rate
moments (under the appropriate pricing measure) at a time horizon equal to the option expiry
by suitably integrating over swaptions with different strikes. We establish new stylized facts
regarding the cross-sectional and time-series variation in conditional volatility and skewness
of swap rates. In particular, we show that skewness is stochastic, largely unspanned by rates
and volatility, and sometimes changes sign.
We then develop and estimate a dynamic term structure model that is consistent with
these stylized facts. The model provides new insights on the pricing of risk in fixed income
markets. Among our novel results are that the price of interest rate and variance risk depend
significantly on the level of skewness, and that skewness risk itself is priced.
Finally, we investigate the fundamental drivers of volatility and skewness with a focus
on the impact of perceived uncertainty about the macroeconomy. Consistent with intuition,
we find that volatility tends to be increase, when the macroeconomic outlook becomes more
uncertain, while skewness tends to increase, when the macroeconomic outlook becomes more
skewed towards higher growth and inflation rates. We also show that beliefs about GDP
growth are the most important determinants in the USD market, while beliefs about inflation
are more important in the EUR market. This finding is consistent with differences in monetary
policy objectives in the two currency areas.
The recent literature on incorporating macro factors into dynamic term structure models
is based almost exclusively on Gaussian models. Our analysis indicates that it will be fruitful
to incorporate uncertainty about macro factors into models with stochastic volatility and
skewness.60 We leave this issue for future research.

60
A number of equilibrium models for term structure dynamics do incorporate macroeconomic uncertainty;
see, e.g., Piazzesi and Schneider (2007), Bansal and Shaliastovich (2009), Le and Singleton (2010), and Xiong
and Yan (2010). However, their main focus is on the impact of uncertainty on interest rates and the pricing of
interest rate risk, rather than volatility and skewness.

36
A Swaption pricing

A.1 The weights in (18) and (19)

The weights ζj (t) in (18) are given by

P (t, Tm )
ζm (t) = (40)
A(t, Tm , Tn )
P (t, Tj )
ζj (t) = −τj−1 S(t, Tm , Tn ) , j = m + 1, ..., n − 1 (41)
A(t, Tm , Tn )
P (t, Tn )
ζn (t) = −(1 + τn−1 S(t, Tm , Tn )) , (42)
A(t, Tm , Tn )

while the weights χj (t) in (19) are given by

τj−1 P (t, Tj )
χj (t) = , j = m + 1, ..., n. (43)
A(t, Tm , Tn )

A.2 Fourier-based pricing formula for swaptions

The dynamics of the forward swap rate under A is not entirely affine, due to the stochastic
weights ζj (t) and χj (t). However, these are low variance martingales under A, and following
much of the literature on LIBOR market models, we may “freeze” these at their initial values
to obtain a truly affine model, in which case swaptions can be priced quasi-analytically.61
First, we find the characteristic function of Sm,n (Tm ) given by
h i
ψ(u, t, Tm , Tn ) = EtA eiuSm,n (Tm ) , (44)
√
where i = −1. This has an exponentially affine solution as demonstrated in the following
proposition:

Proposition 1 (44) is given by

P2
ψ(u, t, Tm , Tn ) = eM (Tm −t)+ j=1 Nj (Tm −t)vj (t)+iuSm,n (t)
, (45)
61
In a LIBOR market model setting, this “freezing” technique results in very small biases in swaptions prices.
Extensive simulations show that the biases are also very small in our context (these results are available upon
request).

37
where M (τ ), N1 (τ ), and N2 (τ ) solve the following system of ODEs

dM (τ )
= N1 (τ )η1 + N2 (τ )η2 (46)
dτ
N
!
dN1 (τ ) X 1
= −κA
1 + iu ρi σS,i (t, Tm , Tn ) N1 (τ ) − κ21 N2 (τ ) + N1 (τ )2
dτ 2
i=1
N
X
1
− u2 σS,i (t, Tm , Tn )2 (47)
2
i=1
N
!
dN2 (τ ) X 1
= −κA
2 + iu ρi σS,i (t, Tm , Tn ) N2 (τ ) − κ12 N1 (τ ) + N2 (τ )2
dτ 2
i=1
N
X
1
− u2 σS,i (t, Tm , Tn )2 (48)
2
i=1

subject to the boundary conditions M (0) = 0, N1 (0) = 0, and N2 (0) = 0.

Proof: See online appendix.

Next, we follow the general approach of Carr and Madan (1999) and Lee (2004) to price
swaptions. The idea is that the Fourier transform of the modified swaption price,

Pbm,n (t, K) = eαK Pm,n (t, K), (49)

can be expressed in terms of the characteristic function of Sm,n (Tm ).62 The swaption price is
then obtained by applying the Fourier inversion theorem. The result is given in the following
proposition:

Proposition 2 The time-t price of a European payer swaption expiring at Tm with strike K
on a swap for the period Tm to Tn , Pm,n (t, K), is given by
Z ∞  
e−αK e−iuK ψ(u − iα, t, Tm , Tn )
Pm,n (t, K) = Am,n (t) Re du. (50)
π 0 (α + iu)2

Proof: See online appendix.

62
The control parameter α must be chosen to ensure that the modified swaption price is L2 integrable, which
is a sufficient condition for its Fourier transform to exist.

38
B Pricing of risk

We derive the expressions (32), (33), and (34). We first state the dynamics of v(t) and w(t)
in the SV 2U SS specification under the risk-neutral measure. The dynamics of v(t) is given by
p p Q
dv(t) = (2η − κv(t)) dt + v1 (t)dZ Q (t) + v2 (t)dZ (t)
p
= (2η − κv(t)) dt + v(t)dZbQ (t), (51)

where
p p Q
bQ v1 (t)dZ Q (t) + v2 (t)dZ (t)
dZ (t) = p (52)
v(t)

is a Wiener process under the risk-neutral measure. This shows that v(t) itself follows a
square-root process. The dynamics of w(t) is given by
p p
η v1 (t) Q v2 (t) Q
dw(t) = (1 − 2w(t)) dt − w(t) dZ (t) + (1 − w(t)) dZ (t). (53)
v(t) v(t) v(t)

From Table 4 it is clear that the estimate of η is such that the Feller condition for the v(t)-
process is satisfied (4η > 1). Consequently, v(t) is strictly positive and the w(t)-process is
well-defined.
Let f (St ) denote the price of a security with pure exposure to interest rate risk. By Ito’s
Lemma, the dynamics of f under the risk-neutral measure is given by

df (St )
N
1 df (St ) X p p 
Q
= r(t)dt + σS,i v1 (t)dWiQ (t) + v2 (t)dW i (t) (54)
f (St ) f (St ) dSt
i=1

Applying the market price of risk specification (20) (and recalling that only exposure to the
first term structure factor is priced), the instantaneous expected excess return is given by
 
1 P df (St ) 1 df (St )

E − r(t) = σS,1 v1 (t)λ1 + v2 (t)λ1
dt f (St ) f (St ) dSt
1 df (St )

= σS,1 v(t) (1 − w(t))λ1 + w(t)λ1 . (55)
f (St ) dSt

The instantaneous return variance is given by

   2 X
N
1 df (St ) 1 df (St ) 2
V ar = σS,i v(t). (56)
dt f (St ) f (St ) dSt
i=1

Hence, the instantaneous Sharpe ratio is given by (32).

39
 Next, let f (vt ) denote the price of a security with pure exposure to variance risk. By Ito’s
Lemma, the dynamics of f under the risk-neutral measure is given by
df (vt ) 1 df (vt ) p
= r(t)dt + v(t)dZbQ (t) (57)
f (vt ) f (vt ) dvt
Applying the market price of risk specification (21), the instantaneous expected excess return
is given by
 
1 P df (vt ) 1 df (vt )
E − r(t) = (v1 (t)ν + v2 (t)ν)
dt f (vt ) f (vt ) dvt
1 df (vt )
= v(t) ((1 − w(t))ν + w(t)ν) . (58)
f (vt ) dvt
Since the instantaneous return variance is given by
   
1 df (vt ) 1 df (vt ) 2
V ar = v(t), (59)
dt f (vt ) f (vt ) dvt
the instantaneous Sharpe ratio is given by (33).
Finally, let f (wt ) denote the price of a security with pure exposure to skewness risk.
p p !
df (wt ) 1 df (wt ) v1 (t) Q v2 (t) Q
= r(t)dt + −w(t) dZ (t) + (1 − w(t)) dZ (t) (60)
f (wt ) f (wt ) dwt v(t) v(t)

Applying the market price of risk specification (21), the instantaneous expected excess return
is given by
 
1 P df (wt ) 1 df (wt )
E − r(t) = w(t)(1 − w(t)) (ν − ν) . (61)
dt f (wt ) f (wt ) dwt
The instantaneous return variance is given by
   
1 df (wt ) 1 df (wt ) 2 1
V ar = w(t)(1 − w(t)) . (62)
dt f (wt ) f (wt ) dwt v(t)
Hence, the instantaneous Sharpe ratio is given by (34).

C Maximum likelihood estimation

C.1 The state space form

We cast the model in state space form, which consists of a measurement equation and a
transition equation. The measurement equation describes the relationship between the state
variables and the prices of swaps and swaptions, while the transition equation describes the
discrete-time dynamics of the state variables.

40
 Let Xt denote the vector of state variables. While the transition density of Xt is unknown,
its conditional mean and variance is known in closed form, since Xt follows an affine diffusion
process. We approximate the transition density with a Gaussian density with identical first
and second moments, in which case the transition equation becomes

Xt = Φ0 + ΦX Xt−1 + wt , wt ∼ N (0, Qt ), (63)

where Qt = Q0 + Qv1 v1,t + Qv2 v2,t and Φ0 , ΦX , Q0 , Qv1 , and Qv2 are given in closed form.63
The measurement equation is given by

Zt = h(Xt ) + ut , ut ∼ N (0, Ω), (64)

where Zt is a vector consisting of all the swaptions and underlying swap rates in the time-t
swaption cube, h is the pricing function, and ut is a vector of iid. Gaussian pricing errors with
covariance matrix Ω.
Ideally, we would like to fit the model directly to normal implied volatilities, which are
more stable than prices (or log-normal implied volatilities) along the swap maturity, option
expiry, moneyness, and time-series dimensions. This is not practical, however, since computing
implied volatilities from prices requires a numerical inversion for each swaption, which would
add an extra layer of complexity to the estimation procedure. Instead, we fit the model to
option prices scaled by their normal vegas (i.e., the sensitivities of the swaption prices to
variations in volatilities in the normal pricing model).64
To reduce the number of parameters in Ω, we assume that the measurement errors are
2
cross-sectionally uncorrelated (that is, Ω is diagonal), and that one variance, σrates , applies
2
to all pricing errors for swap rates, and that another variance, σswaption , applies to all pricing
errors for scaled swaption prices.
63
Approximating the true transition density with a Gaussian, makes this a QML procedure. While QML
estimation has been shown to be consistent in many settings, it is in fact not consistent in a Kalman filter
setting, since the conditional covariance matrix Qt in the recursions depends on the Kalman filter estimates of
the volatility state variables rather than the true, but unobservable, values; see, e.g., Duan and Simonato (1999).
However, simulation results in several papers have shown this issue to be negligible in practice.
64
This essentially converts swaption pricing errors in terms of prices into swaption pricing errors in terms of
normal implied volatilities, via a linear approximation.

41
C.2 The unscented Kalman filter

If the pricing function were linear, h(Xt ) = h0 +HXt , the Kalman filter would provide efficient
estimates of the conditional mean and variance of the state vector. Let X̂t|t−1 = Et−1 [Xt ] and
Ẑt|t−1 = Et−1 [Zt ] denote the expectation of Xt and Zt , respectively, using information up to
and including time t − 1, and let Pt|t−1 and Ft|t−1 denote the corresponding error covariance
matrices. Furthermore, let X̂t = Et [Xt ] denote the expectation of Xt including information
at time t, and let Pt denote the corresponding error covariance matrix. The Kalman filter
consists of two steps: prediction and update. In the prediction step, X̂t|t−1 and Pt|t−1 are
given by

X̂t|t−1 = Φ0 + ΦX X̂t−1 (65)

Pt|t−1 = ΦX Pt−1 Φ′X + Qt , (66)

and Ẑt|t−1 and Ft|t−1 are in turn given by

Ẑt|t−1 = h(X̂t|t−1 ) (67)

Ft|t−1 = HPt|t−1 H ′ + Ω. (68)

In the update step, the state estimate is refined based on the difference between predicted and
observed swaps and swaptions, with X̂t and Pt given by

X̂t = X̂t|t−1 + Wt (Zt − Ẑt|t−1 ) (69)

Pt = Pt|t−1 − Wt Ft|t−1 Wt′ , (70)

where

−1
Wt = Pt|t−1 H ′ Ft|t−1 (71)

is the covariance between pricing and filtering errors.

In our setting, the pricing function is non-linear for both swaps and swaptions, and the
Kalman filter has to be modified. Non-linear state space systems have traditionally been han-
dled with the extended Kalman filter, which effectively linearizes the measure equation around
the predicted state. However, in recent years the unscented Kalman filter has emerged as an
attractive alternative. Rather than approximating the measurement equation, it uses the true
non-linear measurement equation and instead approximates the distribution of the state vector
with a deterministically chosen set of sample points, called “sigma points”, that completely

42
capture the true mean and covariance of the state vector. When propagated through the
non-linear pricing function, the sigma points capture the mean and covariance of swaps and
swaptions accurately to the 2nd order (3rd order for true Gaussian states) for any nonlinear-
ity.65
More specifically, a set of 2L+ 1 sigma points and associated weights are selected according
to the following scheme

0 κ
X̂t|t−1 = X̂t|t−1 w0 = L+κ
q 
i 1
X̂t|t−1 = X̂t|t−1 + (L + κ)Pt|t−1 wi = 2(L+κ) i = 1, ..., L (72)
q i
i 1
X̂t|t−1 = X̂t|t−1 − (L + κ)Pt|t−1 wi = 2(L+κ) i = L + 1, ..., 2L,
i

where L is the dimension of X̂t|t−1 , κ is a scaling parameter, wi is the weight associated with
q 
the i’th sigma-point, and (L + κ)Pt|t−1 is the i’th column of the matrix square root.
i
Then, in the prediction step, (67) and (68) are replaced by
2L
X
Ẑt|t−1 = wi h(Xˆt|t−1
i
) (73)
i=0
2L
X
Ft|t−1 = wi (h(Xˆt|t−1
i
) − Ẑt|t−1 )(h(Xˆt|t−1
i
) − Ẑt|t−1 )′ + Ω. (74)
i=0

The update step is still given by (69) and (70), but with Wt computed as
2L
X
Wt = wi (X̂t|t−1
i
− X̂t|t−1 )(h(Xˆt|t−1
i −1
) − Ẑt|t−1 )′ Ft|t−1 . (75)
i=0

Finally, the log-likelihood function is given by

X T T T
1 1X 1X
logL = − log2π Nt − log|Ft|t−1 | − (Zt − Ẑt|t−1 )′ Ft|t−1
−1
(Zt − Ẑt|t−1 ), (76)
2 2 2
t=1 t=1 t=1

where T is the number of observation dates, and Nt is the dimension of Zt .66

65
For comparison, the extended Kalman filter estimates the mean and covariance accurately to the 1st order.
Note that the computational costs of the extended Kalman filter and the unscented Kalman filter are of the
same order of magnitude.
66
To maximize the log-likelihood function, we initially apply the Nelder-Mead algorithm and later switch
to the gradient-based BFGS algorithm. The optimization is repeated with several different plausible initial
parameter guesses, to minimize the risk of not reaching the global optimum.

43
 Tenor Option expiry
1 mth 3 mths 6 mths 9 mths 1 yr 2 yrs 5 yrs 10 yrs
Panel A: USD market
2 yrs 110.1 112.0 114.4 117.4 120.7 123.6 116.9 98.2
(35.5) (31.5) (27.9) (26.9) (26.6) (24.6) (18.0) (12.0)

5 yrs 122.4 122.2 121.8 121.1 121.3 120.4 111.7 93.2

(37.4) (33.5) (29.4) (27.4) (26.3) (23.2) (16.5) (10.4)

10 yrs 115.9 115.8 115.4 114.7 114.4 113.3 104.7 86.8

(36.6) (32.7) (28.7) (26.5) (25.0) (22.0) (15.0) (9.2)

20 yrs 106.9 105.7 104.0 102.7 101.8 99.5 89.9 74.1

(36.1) (31.4) (26.7) (24.0) (21.8) (18.6) (12.3) (7.9)

30 yrs 103.5 101.7 99.9 98.6 97.5 95.1 85.4 69.6

(37.6) (31.5) (26.3) (23.4) (20.9) (17.1) (10.8) (6.3)

Panel B: EUR market

2 yrs 80.7 81.2 81.6 81.3 80.8 80.6 78.0 72.1
(29.2) (24.6) (20.2) (17.2) (15.3) (12.9) (8.7) (6.3)

5 yrs 83.6 82.4 80.9 79.6 78.6 77.1 74.2 69.1

(25.3) (21.1) (17.1) (14.8) (13.4) (11.5) (8.5) (7.2)

10 yrs 74.7 74.7 74.3 73.7 73.3 73.4 71.9 67.1

(23.1) (20.5) (17.6) (15.9) (15.0) (13.6) (9.9) (7.6)

20 yrs 72.3 72.0 71.2 70.4 70.0 69.3 67.2 61.9

(30.0) (26.4) (21.8) (19.3) (18.0) (15.4) (10.9) (8.2)

30 yrs 71.6 71.2 70.3 69.3 68.7 67.9 65.3 59.8

(36.8) (32.5) (26.7) (23.0) (20.8) (17.6) (12.9) (9.2)

Notes: The table shows average conditional volatilities (annualized and in basis points) of the future swap rate
distributions under the annuity measure A. Standard deviations of conditional volatilities are in parentheses.
In the USD market, each statistic is computed on the basis of 419 weekly observations from December 19, 2001
to January 27, 2010. In the EUR market, each statistic is computed on the basis of 449 weekly observations
from June 6, 2001 to January 27, 2010.

Table 1: Volatility (annualized) of swap rate distributions

 Tenor Option expiry
1 mth 3 mths 6 mths 9 mths 1 yr 2 yrs 5 yrs 10 yrs
Panel A: USD market
2 yrs 0.00 0.20 0.24 0.24 0.27 0.30 0.43 0.42
(0.30) (0.36) (0.43) (0.40) (0.42) (0.41) (0.33) (0.29)

5 yrs 0.01 0.17 0.19 0.20 0.21 0.23 0.33 0.37

(0.20) (0.23) (0.28) (0.25) (0.28) (0.34) (0.35) (0.30)

10 yrs −0.00 0.15 0.16 0.15 0.16 0.18 0.29 0.32

(0.17) (0.18) (0.22) (0.20) (0.23) (0.30) (0.34) (0.31)

20 yrs −0.03 0.12 0.13 0.12 0.13 0.18 0.27 0.30

(0.13) (0.13) (0.16) (0.16) (0.19) (0.25) (0.32) (0.36)

30 yrs −0.04 0.10 0.12 0.11 0.12 0.15 0.22 0.27

(0.14) (0.13) (0.15) (0.15) (0.17) (0.23) (0.30) (0.33)

Panel B: EUR market

2 yrs −0.00 0.17 0.27 0.31 0.36 0.47 0.60 0.62
(0.15) (0.21) (0.28) (0.27) (0.29) (0.35) (0.33) (0.38)

5 yrs −0.15 −0.03 0.04 0.09 0.12 0.22 0.39 0.46

(0.38) (0.41) (0.40) (0.33) (0.35) (0.30) (0.27) (0.25)

10 yrs −0.21 −0.11 −0.06 −0.00 0.02 0.12 0.29 0.36

(0.37) (0.39) (0.36) (0.29) (0.30) (0.25) (0.25) (0.25)

20 yrs −0.29 −0.16 −0.11 −0.05 −0.02 0.05 0.24 0.35

(0.28) (0.27) (0.24) (0.21) (0.22) (0.22) (0.24) (0.27)

30 yrs −0.32 −0.18 −0.13 −0.07 −0.04 0.00 0.22 0.33

(0.29) (0.27) (0.25) (0.22) (0.24) (0.24) (0.28) (0.34)

Notes: The table shows average conditional skewness of the future swap rate distributions under the annuity
measure A. Standard deviations of conditional skewness are in parentheses. In the USD market, each statistic
is computed on the basis of 419 weekly observations from December 19, 2001 to January 27, 2010. In the EUR
market, each statistic is computed on the basis of 449 weekly observations from June 6, 2001 to January 27,
2010.

Table 2: Skewness of swap rate distributions

 Tenor Option expiry
1 mth 3 mths 6 mths 9 mths 1 yr 2 yrs 5 yrs 10 yrs
Panel A: USD market
2 yrs 0.09 0.12 0.16 0.17 0.17 0.15 0.13 0.11
5 yrs 0.09 0.10 0.13 0.17 0.19 0.17 0.16 0.16
10 yrs 0.08 0.07 0.10 0.15 0.19 0.17 0.20 0.24
20 yrs 0.07 0.06 0.07 0.10 0.15 0.17 0.20 0.27
30 yrs 0.06 0.04 0.04 0.06 0.12 0.18 0.22 0.25
Panel B: EUR market
2 yrs 0.08 0.11 0.14 0.17 0.20 0.22 0.22 0.22
5 yrs 0.02 0.05 0.07 0.10 0.13 0.18 0.25 0.32
10 yrs 0.01 0.04 0.07 0.09 0.11 0.16 0.27 0.36
20 yrs 0.01 0.04 0.07 0.10 0.11 0.13 0.25 0.36
30 yrs 0.01 0.05 0.08 0.10 0.10 0.11 0.21 0.32
Notes: The table shows R2 s from regressing weekly changes in the conditional skewness of swap rate distributions
on the main principal components (PCs) of weekly changes in all forward swap rates and the main PCs of
weekly changes in conditional volatilities of all swap rate distributions. We also include the squared PCs in the
regressions. In the USD market, the time series consist of 419 weekly observations from December 19, 2001 to
January 27, 2010. In the EUR market, the time series consist of 449 weekly observations from June 6, 2001 to
January 27, 2010.

Table 3: Evidence for unspanned stochastic skewness

 USD market EUR market
SV 1 SV 2U SS SV 2 SV 1 SV 2U SS SV 2
α1 0.0072 0.0062 0.0084 0.0090 0.0067 0.0100
(0.0005) (0.0005) (0.0007) (0.0015) (0.0005) (0.0017)

α2 0.0021 0.0024 0.0001 0.0015 0.0058 0.0019

(0.0001) (0.0002) (0.0000) (0.0002) (0.0008) (0.0003)

α3 0.0049 0.0043 0.0055 0.0020 0.0021 0.0024

(0.0008) (0.0004) (0.0004) (0.0001) (0.0001) (0.0002)

ξ 0.0104 0.0090 0.0116 0.0115 0.0007 0.0123

(0.0007) (0.0017) (0.0022) (0.0014) (0.0001) (0.0010)

γ 0.2470 0.2447 0.2525 0.2837 0.1651 0.3888

(0.0218) (0.0181) (0.0269) (0.0261) (0.0237) (0.0271)

ϕ 0.0521 0.0464 0.0487 0.0435 0.0484 0.0476

(0.0068) (0.0070) (0.0058) (0.0043) (0.0049) (0.0077)

ρ1 0.2944 −0.3665 0.0372 0.2877 −0.2464 −0.0382

(0.0446) (0.0432) (0.0425) (0.0455) (0.0565) (0.0521)

ρ1 0.5141 0.3956 0.5773 0.2595

(0.0689) (0.0520) (0.0443) (0.0423)

ρ2 0.0881 −0.1627 −0.0099 0.2467 −0.2271 −0.1669

(0.0562) (0.0401) (0.0478) (0.0444) (0.0656) (0.0472)

ρ2 0.6395 0.2926 0.4677 0.2639

(0.0632) (0.0640) (0.0587) (0.0437)

ρ3 0.2397 −0.4027 −0.1235 0.3047 −0.2534 −0.1509

(0.0423) (0.0645) (0.0529) (0.0661) (0.0505) (0.0455)

ρ3 0.6412 0.5247 0.4014 0.4270

(0.0661) (0.0673) (0.0554) (0.0472)

κ1 0.5039 0.4153 0.1427 0.5295 0.7006 0.0715

(0.0346) (0.0344) (0.0118) (0.0506) (0.0494) (0.0088)

κ2 3.6804 4.3121
(0.0433) (0.0692)

η1 0.5953 0.3658 0.0996 0.3797 0.3165 0.1203

(0.0458) (0.0471) (0.0128) (0.0585) (0.0252) (0.0181)

η2 2.2244 1.0727
(0.0486) (0.0573)

Notes: Continued on next page.

Table 4: Maximum-likelihood estimates

47
 USD market EUR market
SV 1 SV 2U SS SV 2 SV 1 SV 2U SS SV 2
λ1 −0.2327 −0.1301 0.0307 −0.2029 −0.0874 −0.1543
(0.0522) (0.0824) (0.0872) (0.0684) (0.0965) (0.0723)

λ1 −0.3502 −0.3136 −0.2617 −0.4109

(0.0725) (0.0594) (0.0888) (0.0653)

ν −0.2852 −0.0940 −0.1432 −0.3542 −0.1485 −0.2660

(0.0825) (0.0774) (0.0843) (0.0890) (0.0743) (0.0754)

ν −0.4749 −0.3011 −0.3763 −0.1367

(0.0648) (0.0592) (0.0718) (0.0755)

σrates (bp) 6.3465 6.3676 6.2941 5.4623 5.4587 5.4786

(0.6885) (1.0664) (0.5449) (0.4074) (0.6200) (0.7181)

σswaptions (bp) 5.5125 4.5998 4.1243 4.8212 4.3356 3.9887

(0.3894) (0.3476) (0.6224) (0.6392) (0.4060) (0.3456)

Log-L ×104 -27.127 -25.923 -24.197 -23.696 -23.245 -21.155

Notes: Maximum-likelihood estimates of the SV 1, SV 2U SS , and SV 2 specifications. The SV 2U SS specification
imposes the unspanned stochastic skewness restrictions κ2 = κ1 and η2 = η1 . The sample period is December 19,
2001 to January 27, 2010 in the USD market and June 6, 2001 to January 27, 2010 in the EUR market. Outer-
product standard errors are in parentheses. σrates denotes the standard deviation of swap rate measurement
errors and σswaptions denotes the standard deviation of scaled swaption price measurement errors.

Table 4: Maximum-likelihood estimates (cont.)

48
 Swap rates swaption volatility skewness kurtosis
IV
Panel A: USD market
SV 1 6.23 5.44 4.96 0.21 0.34
SV 2U SS 6.07 4.58 4.83 0.06 0.18
SV 2 6.21 3.62 3.04 0.16 0.30
SV 2U SS − SV 1 −0.16 −0.86∗∗∗ −0.13∗ −0.15∗∗∗ −0.16∗∗∗
(−0.34) (−6.24) (−1.73) (−7.02) (−5.17)

SV 2 − SV 2U SS 0.14 −0.95∗∗∗ −1.78∗∗∗ 0.10 ∗∗∗ 0.12

(0.29) (−3.84) (−5.51) (5.39) (1.00)

Panel B: EUR market

SV 1 7.09 3.72 2.84 0.24 0.63
SV 2U SS 7.13 3.31 2.77 0.13 0.50
SV 2 7.16 2.30 1.36 0.22 0.40
SV 2U SS − SV 1 0.04 −0.41∗∗∗ −0.07 −0.12∗∗∗ −0.12∗
(0.37) (−4.88) (−1.11) (−9.74) (−1.85)

SV 2 − SV 2U SS 0.03 −1.01∗∗∗ −1.41∗∗∗ 0.09 ∗∗∗ −0.10

(0.42) (−4.83) (−4.65) (6.68) (−1.42)

Notes: The table compares the SV 1, SV 2U SS , and SV 2 specifications in terms of their ability to match swap
rates, normal implied volatilities (in basis points) as well as conditional volatility (annualized and in basis
points), skewness, and kurtosis of the future swap rate distributions under the annuity measure A. It reports
means of RMSE time series of swap rates, implied volatilities, and swap rate moments. It also reports mean
differences in RMSEs between model specifications. T -statistics, corrected for serial correlation up to 26 lags
(i.e., two quarters), are in parentheses. In the USD market, each statistic is computed on the basis of 419 weekly
observations from December 19, 2001 to January 27, 2010. In the EUR market, each statistic is computed on
the basis of 449 weekly observations from June 6, 2001 to January 27, 2010. *, **, and *** denote significance
at the 10, 5, and 1 percent levels, respectively.

Table 5: Comparison of model specifications

 Tenor Option expiry
1 mth 3 mths 6 mths 9 mths 1 yr 2 yrs 5 yrs 10 yrs
Panel A: SV 1 specification
2 yrs 0.22 0.26 0.28 0.41 0.56 0.64 0.71 0.71
(0.01) (0.04) (0.04) (0.06) (0.06) (0.05) (0.04) (0.04)

5 yrs 0.33 0.46 0.55 0.70 0.77 0.77 0.73 0.66

(0.01) (0.03) (0.05) (0.06) (0.06) (0.05) (0.04) (0.04)

10 yrs 0.37 0.51 0.74 0.89 0.91 0.84 0.74 0.63

(0.01) (0.03) (0.05) (0.06) (0.06) (0.05) (0.03) (0.04)

20 yrs 0.41 0.55 0.83 0.97 0.98 0.89 0.76 0.62

(0.01) (0.04) (0.05) (0.06) (0.06) (0.05) (0.03) (0.03)

30 yrs 0.48 0.62 0.88 0.98 0.98 0.92 0.78 0.64

(0.02) (0.04) (0.05) (0.06) (0.06) (0.05) (0.03) (0.03)

Panel B: SV 2U SS specification
2 yrs 0.09 0.16 0.20 0.24 0.24 0.22 0.21 0.21
(0.02) (0.03) (0.05) (0.06) (0.06) (0.08) (0.08) (0.04)

5 yrs 0.14 0.19 0.22 0.23 0.22 0.20 0.19 0.19

(0.02) (0.04) (0.05) (0.06) (0.07) (0.08) (0.08) (0.04)

10 yrs 0.15 0.19 0.22 0.22 0.21 0.18 0.17 0.18

(0.03) (0.04) (0.05) (0.06) (0.07) (0.08) (0.08) (0.04)

20 yrs 0.12 0.18 0.21 0.22 0.20 0.17 0.16 0.18

(0.03) (0.04) (0.05) (0.06) (0.07) (0.08) (0.08) (0.04)

30 yrs 0.09 0.17 0.21 0.21 0.19 0.17 0.16 0.20

(0.03) (0.04) (0.05) (0.06) (0.07) (0.08) (0.07) (0.04)

Panel C: SV 2 specification
2 yrs 0.21 0.36 0.42 0.47 0.51 0.56 0.68 0.87
(0.04) (0.04) (0.06) (0.06) (0.06) (0.06) (0.07) (0.06)

5 yrs 0.25 0.40 0.48 0.51 0.52 0.57 0.69 0.88

(0.03) (0.04) (0.05) (0.05) (0.06) (0.06) (0.07) (0.06)

10 yrs 0.25 0.41 0.51 0.53 0.54 0.58 0.71 0.89

(0.03) (0.04) (0.05) (0.05) (0.06) (0.06) (0.08) (0.06)

20 yrs 0.22 0.42 0.53 0.55 0.56 0.59 0.71 0.88

(0.04) (0.05) (0.06) (0.06) (0.06) (0.07) (0.08) (0.06)

30 yrs 0.20 0.42 0.55 0.57 0.57 0.60 0.71 0.88

(0.04) (0.05) (0.06) (0.06) (0.06) (0.07) (0.08) (0.06)

Notes: In this table, we repeat the analysis from Table 3 on 100 simulated samples of the same size as our
original data. We consider all three model specifications estimated on USD data. For each sample, we regress
weekly changes in the conditional skewness of swap rate distributions on the main principal components (PCs)
of weekly changes in all forward swap rates and the main PCs of weekly changes in conditional volatilities of all
swap rate distributions. We also include the squared PCs in the regressions. The table shows the mean (along
with its standard error in parentheses) of the R2 s across the 100 samples.

Table 6: Model-implied unspanned stochastic skewness

 GDP vol GDP skew IN F vol IN F skew EQvol EQskew ILLIQ REF I R2
Panel A: USD market
p
v(t) 0.350 ∗∗∗ 0.211 ∗∗∗ 0.526 −0.106 0.392
(4.642) (2.957) (0.850) (−0.654)
p
v(t) 0.309 ∗∗∗ 0.016 0.898 0.176 1.713 ∗∗∗ 1.449 ∗∗ 0.193 ∗∗ 0.037 ∗∗ 0.573
(5.877) (0.202) (1.369) (1.497) (5.594) (1.966) (2.326) (2.046)

w(t) 0.170 ∗∗ 0.150 ∗∗∗ −0.327 −0.047 0.540

(2.520) (2.587) (−1.099) (−0.529)

w(t) 0.121 ∗∗ 0.126 ∗∗ −0.329 0.093 0.441 ∗∗ 1.386 ∗∗∗ −0.113 0.012 0.621
(2.136) (2.374) (−0.930) (1.378) (2.226) (2.597) (−1.574) (1.024)

Panel B: EUR market

p
v(t) −0.034 0.066 ∗∗ 1.905 ∗∗∗ 0.179 ∗∗ 0.627
(−0.537) (2.024) (6.564) (2.020)
p
v(t) 0.053 0.048 ∗∗ 1.089 ∗∗∗ 0.057 0.486 ∗∗∗ 0.089 0.074 ∗ 0.007 0.697
(1.191) (2.315) (5.026) (0.783) (3.879) (0.249) (1.830) (0.762)

w(t) 0.061 ∗ 0.055 −0.269 0.153 ∗∗ 0.135

(1.746) (1.563) (−1.567) (2.129)

w(t) 0.057 ∗ 0.065 ∗∗ −0.232 0.163 ∗∗ −0.185 1.312 ∗∗∗ 0.037 0.005 0.291
(1.743) (1.986) (−1.394) (2.177) (−1.397) (3.566) (1.005) (0.613)
p
Notes: The table reports estimates of the MIDAS regression specification (35) in which the volatility and skewness state variables ( v(t) and w(t),
respectively) of the SV 2U SS specification are regressed on a constant, the dispersion and skewness of the SPF probability distributions for future real
GDP growth and inflation (GDP vol, GDP skew, IN F vol, and IN F skew), volatility and skewness of the risk-neutral equity index return distribution
(EQvol and EQskew), the spread between the 3-month OIS rate and the 3-month bill yield (ILLIQ), and the MBA Refinancing Index (REF I) divided by
1000. Estimation is by non-linear least squares. T -statistics, corrected for heteroscedasticity and serial correlation up to 26 lags (i.e., two quarters), are in
parentheses. The sample period is December 19, 2001 to January 27, 2010. *, **, and *** denote significance at the 10, 5, and 1 percent levels, respectively.

Table 7: Fundamental drivers of volatility and skewness

 ∆GDP vol ∆GDP skew ∆IN F vol ∆IN F skew ∆EQvol ∆EQskew ∆ILLIQ ∆REF I R2
Panel A: USD market
p
∆ v(t) 0.416 ∗∗ −0.020 0.297 0.130 0.170
(2.445) (−0.204) (0.466) (1.124)
p
∆ v(t) 0.376 ∗∗ 0.021 0.468 0.204 0.668 −1.456 0.217 ∗∗ 0.039 0.327
(2.214) (0.173) (0.747) (1.552) (1.188) (−0.876) (2.158) (1.188)

∆w(t) −0.006 0.195 ∗∗ 0.435 0.073 0.242

(−0.045) (2.478) (0.839) (0.776)

∆w(t) 0.044 0.221 ∗∗ 0.701 0.101 0.036 −1.833 −0.121 0.026 0.432
(0.314) (2.201) (1.361) (0.939) (0.078) (−1.343) (−1.457) (0.960)

Panel B: EUR market

p
∆ v(t) −0.014 0.111 ∗ 1.217 ∗∗ −0.037 0.155
(−0.159) (1.949) (2.163) (−0.325)
p
∆ v(t) −0.029 0.114 ∗ 1.194 ∗∗ −0.089 0.647 ∗ −0.696 0.106 ∗ −0.023 0.298
(−0.330) (1.791) (2.009) (−0.763) (1.809) (−0.721) (1.646) (−1.129)

∆w(t) 0.077 ∗ 0.029 0.517 ∗ 0.130 ∗∗ 0.310

(1.860) (1.055) (1.923) (2.393)

∆w(t) 0.086 ∗ 0.035 0.533 ∗ 0.149 ∗∗ −0.120 0.483 0.019 0.002 0.330
(1.876) (1.072) (1.733) (2.456) (−0.648) (0.967) (0.556) (0.140)
p
Notes: The table reports estimates of the regression specification (39) in which the quarterly change in the volatility and skewness state variables (∆ v(t)
U SS
and ∆w(t), respectively) of the SV 2 specification are regressed on a constant and the quarterly changes in the dispersion and skewness of the SPF
probability distributions for future real GDP growth and inflation (∆GDP vol, ∆GDP skew, ∆IN F vol, and ∆IN F skew), the volatility and skewness of the
risk-neutral equity index return distribution (∆EQvol and ∆EQskew), the spread between the 3-month OIS rate and the 3-month bill yield (∆ILLIQ),
and the MBA Refinancing Index (∆REF I) divided by 1000. Estimation is by ordinary least squares. T -statistics, corrected for heteroscedasticity and serial
correlation up to 2 lags (i.e., two quarters), are in parentheses. The sample period is January, 2002 to January, 2010. *, **, and *** denote significance at
the 10, 5, and 1 percent levels, respectively.

Table 8: Fundamental drivers of volatility and skewness, regression in differences

 Panel A: Volatility (mid) Panel B: Skewness (mid)
200 1

180

160
0.5
Basis points

140

120
0
100

80

60 -0.5
Jul01 Jul04 Jul07 Jul10 Jul01 Jul04 Jul07 Jul10

Panel C: Volatility (ask-bid) −3 Panel D: Skewness (ask-bid)

4 6 ×10
3.5
4
3
Basis points

2.5 2
2
1.5 0

1
-2
0.5
0 -4
Jul01 Jul04 Jul07 Jul10 Jul01 Jul04 Jul07 Jul10

Figure 1: Volatility and skewness of the conditional 1-year ahead distribution of the USD
10-year swap rate
Notes: Panels A shows the time series of volatility computed from quoted mid-prices. Panels B shows the time
series of skewness computed from quoted mid-prices. Panels C shows the time series of the difference between
volatility computed from indicative ask-prices and volatility computed from indicative bid-prices. Panels D
shows the time series of the difference between skewness computed from indicative ask-prices and skewness
computed from indicative bid-prices. Moments are computed under the annuity measure A. The time series
consist of 419 weekly observations from December 19th, 2001 to January 27th, 2010.

53
 Volatility Skewness
Panel A: SV 1 specification Panel B: SV 1 specification
200 1
180
160
Basis points

0.5
140
120
0
100
80
60 -0.5
Jul01 Jul04 Jul07 Jul10 Jul01 Jul04 Jul07 Jul10

Panel C: SV 2U SS specification Panel D: SV 2U SS specification

200 1
180
160
Basis points

0.5
140
120
0
100
80
60 -0.5
Jul01 Jul04 Jul07 Jul10 Jul01 Jul04 Jul07 Jul10

Panel E: SV 2 specification Panel F: SV 2 specification

200 1
180
160
Basis points

0.5
140
120
0
100
80
60 -0.5
Jul01 Jul04 Jul07 Jul10 Jul01 Jul04 Jul07 Jul10

Figure 2: Fitted volatility and skewness of the conditional 1-year ahead distribution of the
USD 10-year swap rate
Notes: Panels A, C, and E display the fit to conditional volatility for the SV 1, SV 2U SS , and SV 2 specifications,
respectively. Panels B, D, and F display the fit to conditional skewness for the SV 1, SV 2U SS , and SV 2
specifications, respectively. Volatility is measured in basis points. Grey lines refer to fitted moments, while
black lines refer to actual moments. Moments are computed under the annuity measure A. The time series
consist of 419 weekly observations from December 19th, 2001 to January 27th, 2010.

54
 Panel A: Data Panel B: SV 1 specification

40 40
30 30
20 20
Basis points

Basis points
10 10
0 0
-10 -10
-20 -20
-30 200 -30 200
Jul10 100 Jul10 100
Jul07 0 Jul07 0
Jul04 -100 Jul04 -100
Jul01 -200 Moneyness Jul01 -200 Moneyness

Panel C: SV 2U SS specification Panel D: SV 2 specification

40 40
30 30
20 20
Basis points
Basis points

10 10
0 0
-10 -10
-20 -20
-30 200 -30 200
Jul10 100 Jul10 100
Jul07 0 Jul07 0
Jul04 -100 Jul04 -100
Jul01 -200 Moneyness Jul01 -200 Moneyness

Figure 3: Time series of the USD normal implied volatility smile of the 1-year option on 10-
year swap rate
Notes: Panel A displays the data, Panel B displays the fitted smiles in the SV 1 specification, Panel C displays
the fitted smiles in the SV 2U SS specification, and Panel D displays the fitted smiles in the SV 2 specification.
The smiles are the differences between implied volatilities for different strikes and ATM implied volatilities.
Implied volatilities are measured in basis points. The time series consist of 419 weekly observations from
December 19th, 2001 to January 27th, 2010.

55
 Panel A: SRIR Panel B: SRIR
0 0
-0.1 -0.1
-0.2 -0.2
-0.3 -0.3
-0.4 -0.4
-0.5 -0.5
-0.6 -0.6
-0.7 -0.7
0.5 1 1.5 2 2.5 0 0.2 0.4 0.6 0.8 1

Panel C: SRV ar Panel D: SRV ar

0 0
-0.1
-0.2
-0.2
-0.3 -0.4

-0.4 -0.6
-0.5
-0.8
-0.6
-0.7 -1
0.5 1 1.5 2 2.5 0 0.2 0.4 0.6 0.8 1

Panel E: SRSkew Panel F: SRSkew

-0.1 -0.05
-0.1
-0.2
-0.15
-0.3 -0.2

-0.4 -0.25
-0.3
-0.5
-0.35
-0.6 -0.4
0.5 1 1.5 2 2.5 0 0.2 0.4 0.6 0.8 1
p
v(t) w(t)

Figure 4: Impact of volatility and skewness on Sharpe ratios in USD market

Notes: The figure shows how instantaneous Sharpe ratios associated with interest rate risk (SRIR ), variance
risk (SRV ar ), and skewness risk (SRSkew ) vary with the level of volatility and skewness. The panels to the left
p
(A, C, and E) show the dependence on v(t), when w(t) equals its sample mean, while the panels to the right
p
(B, D, and F) show the dependence on w(t), when v(t) equals its sample mean. In the case of SRIR , the
Sharpe ratio refers to exposure to the 10 year swap rate. Solid, dotted, and dashed-dotted lines refer to the
SV 2U SS , SV 2U SS,EX , and SV 3U SS specifications, respectively.

56
 0

-0.2

-0.4

-0.6

-0.8

-1

Jan02 Jan03 Jan04 Jan05 Jan06 Jan07 Jan08 Jan09 Jan10

Figure 5: Time series of Sharpe ratios in USD market

Notes: The figure shows instantaneous Sharpe ratios associated with interest rate risk (black line), variance
risk (light-grey line), and skewness risk (dark-grey line) implied by the SV 2U SS specification. In the case of
interest rate risk, the Sharpe ratio refers to exposure to the 10 year swap rate. The time series consist of 419
weekly observations from December 19th, 2001 to January 27th, 2010.

57
References
Almeida, C., J. Graveline, and S. Joslin (2011): “Do interest rate options contain information
about excess returns,” Journal of Econometrics, 164:35–44.

Andersen, L. and V. Piterbarg (2010): Interest Rate Modeling, Atlantic Financial Press.

Andersen, T. G. and L. Benzoni (2010): “Do bonds span volatility risk in the U.S. Treasury
market? A specification test for affine term structure models,” Journal of Finance, 65:603–653.

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.

Ang, A., M. Piazzesi, and M. Wei (2006): “What does the yield curve tell us about GDP growth?,”
Journal of Econometrics, 131:359–403.

Arditti, F. D. (1967): “Risk and the required return on equity,” Journal of Finance, 22:19–36.

Bakshi, G., N. Kapadia, and D. Madan (2003): “Stock return characteristics, skew laws, and the
differential pricing of individual equity options,” Review of Financial Studies, 16:101–143.

Bakshi, G. and D. Madan (2000): “Spanning and derivative-security valuation,” Journal of Finan-
cial Economics, 55:205–238.

Bansal, R. and I. Shaliastovich (2009): “A long-run risks explanation of predictability puzzles in

bond and currency markets,” Working paper, Duke University.

Barclays (2010): 2010 US Volatility Outlook,

Beber, A. and M. Brandt (2006): “The effects of macroeconomic news on beliefs and preferences:
Evidence from the options market,” Journal of Monetary Economics, pages 1997–2039.

Bekaert, G., S. Cho, and A. Moreno (2010): “New-Keynesian macroeconomics and the term
structure,” Journal of Money, Credit and Banking, 42:33–62.

Bekaert, G. and E. Engstrom (2009): “Asset return dynamics under bad environment-good
environment fundamentals,” Working paper, Columbia Business School.

Bikbov, R. and M. Chernov (2010): “No-arbitrage macroeconomic determinants of the yield curve,”
Journal of Econometrics, 159:166–182.

Black, F. (1976): “The pricing of commodity contracts,” Journal of Financial Economics, 3:167–179.

Bloomberg (2011): “Building the Bloomberg interest rate curve – Definitions and methodology,”
Working paper, Bloomberg Markets.

Bollen, N. and R. Whaley (2004): “Does net buying pressure affect the shape of implied volatility
functions?,” Journal of Finance, 59:711–753.

58
Bongaerts, D., F. de Jong, and J. Driessen (2011): “Derivative pricing with liquidity risk:
Theory and evidence from the credit default swap market,” Journal of Finance, 66:203–240.

Breeden, D. and R. Litzenberger (1978): “Prices of state-contingent claims implicit in option

prices,” Journal of Business, pages 621–651.

Buraschi, A. and P. Whelan (2011): “Macroeconomic uncertainty, difference in beliefs, and bond
risk premia,” Working paper, Imperial College.

Campbell, J. Y. and R. J. Shiller (1991): “Yield spreads and interest rate movements: a bird’s
eye view,” Review of Economic Studies, 58:495–514.

Carr, P., X. Garbaix, and L. Wu (2009): “Linearity-generating processes, unspanned stochastic

volatility, and interest-rate option pricing,” Working paper, New York University.

Carr, P. and D. Madan (1999): “Option valuation using the Fast Fourier Transform,” Journal of
Computational Finance, 2:61–73.

(2001): “Optimal positioning in derivatives securities,” Quantitative Finance, 1:19–37.

Cetin, U., R. Jarrow, P. Protter, and M. Warachka (2006): “Pricing options in an extended
Black Scholes economy with illiquidity: Theory and empirical evidence,” Review of Financial
Studies, 19:493–529.

Cheredito, P., D. Filipovic, and R. Kimmel (2007): “Market price of risk specifications for affine
models: Theory and evidence,” Journal of Financial Economics, 83:123–170.

Christoffersen, P., K. Jakobs, L. Karoui, and K. Mimouni (2009): “Non-linear filtering in

affine term structure models: Evidence from the term structure of swap rates,” Working paper,
McGill University.

Chun, A. L. (2010): “Expectations, bond yields, and monetary policy,” Working paper, Copenhagen
Business School, forthcoming Review of Financial Studies.

Cieslak, A. and P. Povala (2011): “Understanding the term structure of yield curve volatility,”
Working paper, Northwestern University.

Cochrane, J. and M. Piazzesi (2005): “Bond risk premia,” American Economic Review, 95:138–
160.

(2008): “Decomposing the yield curve,” Working paper, University of Chicago.

Collin-Dufresne, P. and R. Goldstein (2002): “Do bonds span the fixed income markets? Theory
and evidence for unspanned stochastic volatility,” Journal of Finance, 57:1685–1730.

Collin-Dufresne, P., R. Goldstein, and C. Jones (2009): “Can interest rate volatility be
extracted from the cross section of bond yields?,” Journal of Financial Economics, 94:47–66.

59
Dai, Q. and K. Singleton (2003): “Term structure dynamics in theory and reality,” Review of
Financial Studies, 16:631–678.

Das, S. R. and R. K. Sundaram (1999): “Of smiles and smirks: A term structure perspective,”
Journal of Financial and Quantitative Analysis, 34:211–239.

de Jong, F., J. Driessen, and A. Pelsser (2004): “On the information in the interest rate term
structure and option prices,” Review of Derivatives Research, 7:99–127.

Deuskar, P., A. Gupta, and M. Subrahmanyam (2008): “The economic determinants of interest
rate option smiles,” Journal of Banking and Finance, 32:714–728.

(2010): “Liquidity effects in OTC options markets: Premium or discount?,” Working

paper, New York University, forthcoming Journal of Financial Markets.

Driessen, J., P. Klaassen, and B. Melenberg (2003): “The performance of multi-factor term
structure models for pricing and hedging caps and swaptions,” Journal of Financial and Quanti-
tative Analysis, 38:635–672.

Duan, J.-C. and J.-G. Simonato (1999): “Estimating and testing exponential-affine term structure
models by Kalman filter,” Review of Quantitative Finance and Accounting, 13:111–135.

Duarte, J. (2008): “The causal effect of mortgage refinancing on interest rate volatility: Empirical
evidence and theoretical implications,” Review of Financial Studies, 21:1689–1731.

Duarte, J., F. Longstaff, and F. Yu (2007): “Risk and return in fixed income arbitrage: Nickels
in front of a steamroller,” Review of Financial Studies, 20:769–811.

Duffee, G. (2011): “Information in (and not in) the term structure,” Working paper, Johns Hopkins
University, forthcoming Review of Financial Studies.

Fan, R., A. Gupta, and P. Ritchken (2003): “Hedging in the possible presence of unspanned
stochastic volatility: Evidence from swaption markets,” Journal of Finance, 58:2219–2248.

Filipovic, D. and A. Trolle (2011): “The term structure of interbank risk,” Working paper, Ecole
Polytechnique Federale de Lausanne and Swiss Finance Institute.

Gallmeyer, M., B. Hollifield, and S. Zin (2005): “Taylor rules, McCallum rules, and the term
structure of interest rates,” Journal of Monetary Economics, 52:921–950.

Garcia, J. and A. Manzanares (2007): “What can probabilistic forecasts tell us about inflation
risks,” Working paper, European Central Bank.

Gârleanu, N., L. H. Pedersen, and A. M. Poteshman (2009): “Demand-based option pricing,”

Review of Financial Studies, 22:4259–4299.

60
Ghysels, E., P. Santa-Clara, and R. Valkanov (2005): “There is a risk-return trade-off after
all,” Journal of Financial Economics, 76:509–548.

(2006): “Predicting volatility: Getting the most out of return data sampled at different
frequencies,” Journal of Econometrics, 131:59–95.

Gupta, A. and M. Subrahmanyam (2005): “Pricing and hedging interest rate options: Evidence
from cap-floor markets,” Journal of Banking and Finance, 29:701–733.

Han, B. (2007): “Stochastic volatilities and correlations of bond yields,” Journal of Finance, 62:1491–
1524.

Harvey, C. and A. Siddique (2000): “Conditional skewness in asset pricing tests,” Journal of
Finance, 55:1263–1295.

Heath, D., R. Jarrow, and A. Morton (1992): “Bond pricing and the term structure of interest
rates: A new methodology for contingent claims valuation,” Econometrica, 60:77–105.

Heidari, M. and L. Wu (2003): “Are interest rate derivatives spanned by the term structure of
interest rates?,” Journal of Fixed Income, 13:75–86.

Heston, S. (1993): “A closed form solution for options with stochastic volatility,” Review of Financial
Studies, 6:327–343.

Jamshidian, F. (1997): “Libor and swap market models and measures,” Finance and Stochastics,
1:293–330.

Jarrow, R., H. Li, and F. Zhao (2007): “Interest rate caps ’smile’ too! But can the LIBOR market
models capture the smile?,” Journal of Finance, 62:345–382.

Johannes, M. and S. Sundaresan (2007): “The impact of collateralization on swap rates,” Journal
of Finance, 62:383–410.

Joslin, S. (2007): “Pricing and hedging volatility risk in fixed income markets,” Working paper, MIT.

Joslin, S., M. Priebsch, and K. Singleton (2010): “Risk premiums in dynamic term structure
models with unspanned macro risks,” Working paper, MIT and Stanford.

Kraus, A. and R. Litzenberger (1976): “Skewness preference and the valuation of risk assets,”
Journal of Finance, 31:1085–1100.

Krishnamurthy, A. (2010): “How debt markets have malfunctioned in the crisis,” Journal of
Economic Perspectives, 24:3–28.

Le, A. and K. Singleton (2010): “An equilibrium term structure model with recursive preferences,”
American Economic Review, Papers and Proceedings.

61
Lee, R. (2004): “Option pricing by transform methods: Extensions, unification, and error control,”
Journal of Computational Finance, 7:51–86.

Li, H. and F. Zhao (2006): “Unspanned stochastic volatlity: Evidence from hedging interest rate
drivatives,” Journal of Finance, 61:341–378.

(2009): “Nonparametric estimation of state-price densities implicit in interest rate cap

prices,” Review of Financial Studies, 22:4335–4376.

Longstaff, F., P. Santa-Clara, and E. Schwartz (2001a): “The relative valuation of caps and
swaptions: Theory and evidence,” Journal of Finance, 56:2067–2109.

(2001b): “Throwing away a billion dollars: the cost of suboptimal exercise strategies in
the swaptions market,” Journal of Financial Economics, 62:39–66.

Ludvigson, S. and S. Ng (2009): “Macro factors in bond risk premia,” Review of Financial Studies,
22:5027–5067.

Nelson, C. and A. Siegel (1987): “Parsimonious modeling of yield curves.,” Journal of Business,
60:473–489.

Piazzesi, M. and M. Schneider (2007): “Equilibrium yield curves,” in, Daron Acemoglu, Kenneth
Rogoff, and Michael Woodford, eds.: NBER Macroeconomics Annual 2006, MIT press, Cambridge
MA.

Scott, R. C. and P. A. Horvath (1980): “On the direction of preference for moments of higher
order than the variance,” Journal of Finance, 35:915–919.

Smith, J. and J. Taylor (2009): “The term structure of policy rules,” Journal of Monetary
Economics, 56:907–919.

Stein, P., A. Tudor, and A. Fukase (2011): “Bad rates bet was source of loss for Morgan Stanley
Japan venture,” Wall Street Journal, April 16.

Trolle, A. B. and E. S. Schwartz (2009): “A general stochastic volatility model for the pricing
of interest rate derivatives,” Review of Financial Studies, 22:2007–2057.

Xiong, W. and H. Yan (2010): “Heterogeneous expectations and bond markets,” Review of Financial
Studies, 23:1433–1466.

Zarnowitz, V. and L. A. Lambros (1987): “Consensus and uncertainty in economic prediction,”

Journal of Political Economy, 95:591–621.

62
 Online Appendix to “The Swaption Cube”

Anders B. Trolle
Ecole Polytechnique Fédérale de Lausanne and Swiss Finance Institute

Eduardo S. Schwartz
UCLA Anderson School of Management and NBER
D Proofs

D.1 Proof of Proposition 1

We can rewrite (44) as

Tm ,Tn
h Tm ,Tn
h ii
ψ(u, t, Tm , Tn ) = EtQ ETQm eiuS(Tm ,Tm ,Tn )
Tm ,Tn
= EtQ [ψ(u, Tm , Tm , Tn )] . (D.1)

Therefore, the proof consist of showing that the process

η(t) ≡ ψ(u, t, Tm , Tn ) (D.2)

is a martingale under QTm ,Tn . To this end we conjecture that ψ(u, t, Tm , Tn ) is of the form
(45). Applying Ito’s Lemma to η(t) and setting τ = Tm − t we obtain
 
dη(t) dM (τ ) dN1 (τ ) dN2 (τ )
= − − v1 (t) − v2 (t) dt + N1 (τ )dv1 (t) + N2 (τ )dv2 (t) + iudS(t, Tm , Tn )
η(t) dτ dτ dτ
1 1 1
+ N1 (τ )2 (dv1 (t))2 + N2 (τ )2 (dv2 (t))2 − u2 (dS(t, Tm , Tn ))2
2 2 2
+N1 (τ )N2 (τ )dv1 (t)dv2 (t) + N1 (τ )iudv1 (t)dS(t, Tm , Tn ) + N2 (τ )iudv2 (t)dS(t, Tm , Tn ). (D.3)

η(t) is a martingale if
 
1 QTm ,Tn dη(t)
0 = Et
dt η(t)
dM (τ ) dN1 (τ ) dN2 (τ )
= − − v1 (t) − v2 (t)
dτ dτ dτ
+N1 (τ ) (η1 − κ
e1 v1 (t) − κ12 v2 (t)) + N2 (τ ) (η2 − κ21 v1 (t) − κ
e2 v2 (t))
N
1 1 1 X
+ N1 (τ )2 v1 (t) + N2 (τ )2 v2 (t) − u2 σS,i (t, Tm , Tn )2 (v1 (t) + v2 (t))
2 2 2
i=1
N
X N
X
+N1 (τ )iu ρi σS,i (t, Tm , Tn )v1 (t) + N2 (τ )iu ρi σS,i (t, Tm , Tn )v2 (t)
i=1 i=1
dM (τ )
= − + N1 (τ )η1 + N2 (τ )η2 +
dτ
" N
!
dN1 (τ ) X 1
− + −e κ1 + iu ρi σS,i (t, Tm , Tn ) N1 (τ ) − κ21 N2 (τ ) + N1 (τ )2
dτ 2
i=1
N
#
1 2X
− u σS,i (t, Tm , Tn )2 v1 (t)
2
i=1
" N
!
dN2 (τ ) X 1
− + −e κ2 + iu ρi σS,i (t, Tm , Tn ) N2 (τ ) − κ12 N1 (τ ) + N2 (τ )2
dτ 2
i=1

64
 N
#
1 2X
− u σS,i (t, Tm , Tn )2 v2 (t) (D.4)
2
i=1

Hence, η(t) is a martingale, provided that M (τ ), N1 (τ ), and N2 (τ ) satisfy (46), (47), and (48).
Furthermore, we have that

ψ(u, Tm , Tm , Tn ) = eiuS(Tm ,Tm ,Tn ) , (D.5)

which is true, provided that M (0) = 0, N1 (0) = 0, and N2 (0) = 0.

D.2 Proof of Proposition 2

Let qTm (S) denote the density of S(Tm , Tm , Tn ) under QTm ,Tn . Then, the price of the payer
swaption is given by
Z ∞
P(t, Tm , Tn , K) = A(t, Tm , Tn ) (S − K)qTm (S)dS. (D.6)
K

The Fourier transform of the modified swaption price is given by

Z ∞
φ(t, Tm , Tn , u) = b Tm , Tn , K)dK
eiuK P(t,
−∞
Z ∞ Z ∞
iuK
= A(t, Tm , Tn ) e eαK (S − K)qTm (S)dSdK
−∞ K
Z ∞ Z S
= A(t, Tm , Tn ) qTm (S) e(α+iu)K (S − K)dKdS
−∞ −∞
R ∞ (α+iu)S
e qTm (S)dS
= A(t, Tm , Tn ) −∞
(α + iu)2
ψ(u − iα, t, Tm , Tn )
= A(t, Tm , Tn ) (D.7)
(α + iu)2
Now, from the Fourier inversion theorem we have
Z ∞
−αK 1
P(t, Tm , Tn , K) = e e−iuK φ(t, Tm , Tn , u)du
2π −∞
Z
1 ∞

= e−αK Re e−iuK φ(t, Tm , Tn , u) du, (D.8)
π 0
where the second equality follows from the fact that the real part of the integrand is even,
while the imaginary part of the integrand is odd.

E Affine term structure specification

Although the model is based on the Heath, Jarrow, and Morton (1992) framework, it may
be represented as an affine model with a finite-dimensional state vector. Building on results

65
in Trolle and Schwartz (2009), the time-t instantaneous forward interest rate for risk-free
borrowing and lending at time T , f (t, T ), is given by

f (t, T ) = f (0, T ) + α1 e−ξ(T −t) x1 (t) + α2 e−γ(T −t) x2 (t) + α3 τ e−γ(T −t) x3 (t)
X8
+ Bφj (T − t)φj (t), (E.1)
j=1

where

Bφ1 (τ ) = −α1 e−2ξτ (E.2)

Bφ2 (τ ) = −α2 e−2γτ (E.3)

Bφ3 (τ ) = α3 e−γτ (E.4)

 2
α3
Bφ4 (τ ) = τ e−γτ (E.5)
γ
 2
α3

Bφ5 (τ ) = − τ + γτ 2 e−2γτ (E.6)
γ
 2
α3
Bφ6 (τ ) = e−γτ (E.7)
γ
 2
α3
Bφ7 (τ ) = − (1 + 2γτ ) e−2γτ (E.8)
γ
α2
Bφ8 (τ ) = − 3 e−2γτ (E.9)
γ

and the state variables evolve, under the risk-neutral measure, according to
 
α1 p p Q
dx1 (t) = (v1 (t) + v2 (t)) − ξx1 (t) dt + v1 (t)dW1Q (t) + v2 (t)dW 1 (t) (E.10)
ξ
 
α2 p p Q
dx2 (t) = (v1 (t) + v2 (t)) − γx2 (t) dt + v1 (t)dW2Q (t) + v2 (t)dW 2 (t) (E.11)
γ
p p Q
dx3 (t) = −γx3 (t)dt + v1 (t)dW3Q (t) + v2 (t)dW 3 (t) (E.12)
 
α1
dφ1 (t) = (v1 (t) + v2 (t)) − 2ξφ1 (t) dt (E.13)
ξ
 
α2
dφ2 (t) = (v1 (t) + v2 (t)) − 2γφ2 (t) dt (E.14)
γ
dφ3 (t) = (x3 (t) − γφ3 (t))dt (E.15)

dφ4 (t) = (v1 (t) + v2 (t) − γφ4 (t))dt (E.16)

dφ5 (t) = (v1 (t) + v2 (t) − 2γφ5 (t))dt (E.17)

dφ6 (t) = (φ4 (t) − γφ6 (t))dt (E.18)

dφ7 (t) = (φ5 (t) − 2γφ7 (t))dt (E.19)

66
 dφ8 (t) = (2φ7 (t) − 2γφ8 (t))dt (E.20)

subject to xj (0) = 0 for j = 1, 2, 3 and φj (0) = 0 for j = 1, ..., 8.

In summary, the term structure of forward rates is an affine function of three term structure
factors xj (t), j = 1, 2, 3, and eight locally deterministic state variables, φj (t), j = 1, ..., 8. It
also follows from (E.1) that the term structure does not depend directly on the volatility
factors.

67
F Additional tables and figures

Evidence for unspanned stochastic volatility To investigate the extent of unspanned

stochastic volatility, we initially extract the main PCs driving weekly changes in forward swap
rates. We use all the PCs explaining more than one percent of the variation, ensuring that they
summarize virtually all of the information in the term structure. Next, for each point on the
swaption matrix, we regress changes in the conditional volatility of the swap rate distribution
on the term structure PCs (we also include the squared PCs in the regressions in an attempt
to take non-linearities into account). The R2 s, which are reported in Table F.1, are relatively
small; in the USD market, ranging from 0.10 to 0.22 with an average of 0.17. Then, we factor
analyze the covariance matrix of the 40 time series of regression residuals. The PCs of the
residuals are, by construction, independent of those of the term structure. There is large
common variation in the regression residuals, with the first PC explaining 73 percent of the
variation in the USD market. Taken together, this strongly indicates that there is systematic
variation in volatility that is largely orthogonal to variation in the term structure.

Details on fit of SV 2U SS specification Tables F.2-F.5 show the fit of the SV 2U SS specifi-
cation (in terms of RMSEs) to swap rates, implied swaption volatilities, conditional swap rate
volatilities, and conditional swap rate skewness, respectively. Figure F.1 shows the average of
actual and fitted volatility and skew surfaces in USD.

Interpretation of v(t) and w(t) in the SV 2U SS specification We confirm that v(t) and
w(t) does indeed capture variance and skewness risk, respectively, as described in Section
4.1.4. Figure F.2 provides a plot of v(t) against the first principal component (PC) of fitted
conditional swap rate variance across the swaption matrix and a plot of w(t) against the first
principal component of fitted conditional swap rate skewness across the swaption matrix. In
the former case, the correlation is one, while the in the latter case the correlation is 0.984.

Estimates of alternative specifications Table F.6 provides maximum-likelihood esti-

mates of the alternative SV 2U SS,EX and SV 3U SS specifications.

Analysis of alternative specifications Table F.7 compares the fit of the SV 2U SS,EX and
SV 3U SS specifications with that of the SV 2U SS specification.

68
 Tenor Option expiry
1 mth 3 mths 6 mths 9 mths 1 yr 2 yrs 5 yrs 10 yrs
Panel A: USD market
2 yrs 0.10 0.14 0.17 0.19 0.22 0.21 0.18 0.14
5 yrs 0.13 0.15 0.18 0.19 0.19 0.19 0.18 0.14
10 yrs 0.17 0.17 0.18 0.19 0.19 0.18 0.17 0.15
20 yrs 0.18 0.18 0.17 0.18 0.18 0.17 0.16 0.18
30 yrs 0.20 0.18 0.16 0.16 0.16 0.15 0.14 0.13
Panel B: EUR market
2 yrs 0.19 0.22 0.23 0.23 0.23 0.19 0.12 0.06
5 yrs 0.22 0.25 0.25 0.24 0.22 0.19 0.13 0.07
10 yrs 0.26 0.28 0.28 0.26 0.24 0.19 0.14 0.07
20 yrs 0.29 0.30 0.29 0.27 0.25 0.20 0.14 0.07
30 yrs 0.31 0.31 0.31 0.29 0.26 0.20 0.12 0.05
Notes: The table shows R2 s from regressing weekly changes in the conditional volatility of swap rate distributions
on the main principal components (PCs) of weekly changes in all forward swap rates. We also include the squared
PCs in the regressions. In the USD market, the time series consist of 419 weekly observations from December
19, 2001 to January 27, 2010. In the EUR market, the time series consist of 449 weekly observations from June
6, 2001 to January 27, 2010.

Table F.1: Evidence for unspanned stochastic volatility

 Maturity USD market EUR market
0.50 12.12 13.09
1 10.59 11.06
2 8.67 8.51
3 8.48 6.17
4 5.60 4.39
5 3.48 3.17
6 2.68 2.89
7 3.12 3.42
8 3.72 4.09
9 4.17 4.74
10 4.46 5.98
12 4.90 6.20
15 5.68 7.14
20 6.90 6.69
30 7.01 7.69
40 9.20 8.39
Notes: For each swap maturity, the table reports the RMSE of swap rates (in basis points). In the USD market,
the time series consist of 419 weekly observations from December 19, 2001 to January 27, 2010. In the EUR
market, the time series consist of 449 weekly observations from June 6, 2001 to January 27, 2010.

Table F.2: Fit to swap rates for SV 2U SS specification

 Tenor Option expiry
1 mth 3 mths 6 mths 9 mths 1 yr 2 yrs 5 yrs 10 yrs
Panel A: USD market
2 yrs 10.50 8.39 7.43 6.49 6.61 6.20 5.50 4.29
5 yrs 7.17 4.72 3.71 3.01 3.26 4.01 4.30 3.76
10 yrs 6.71 4.15 2.77 2.39 2.85 3.65 4.21 3.43
20 yrs 7.74 5.44 3.84 3.17 3.13 3.31 3.31 3.55
30 yrs 9.50 7.15 5.31 4.37 3.97 3.79 3.57 3.67
Panel B: EUR market
2 yrs 6.96 5.51 4.92 4.46 4.65 4.60 4.34 3.53
5 yrs 6.03 4.52 3.68 3.07 3.22 3.18 3.13 3.16
10 yrs 4.11 2.86 2.47 2.26 2.62 2.80 3.04 3.21
20 yrs 5.34 4.04 3.26 2.75 2.89 2.78 2.95 3.39
30 yrs 6.61 5.24 4.31 3.61 3.59 3.24 3.29 3.84
Notes: For each tenor – option expiry category, the table reports the RMSE of normal implied volatilities (in
basis points). In the USD market, the time series consist of 419 weekly observations from December 19, 2001
to January 27, 2010. In the EUR market, the time series consist of 449 weekly observations from June 6, 2001
to January 27, 2010.

Table F.3: Fit to implied swaption volatilities for SV 2U SS specification

 Tenor Option expiry
1 mth 3 mths 6 mths 9 mths 1 yr 2 yrs 5 yrs 10 yrs
Panel A: USD market
2 yrs 13.06 9.81 8.10 7.03 6.97 6.85 6.09 4.99
5 yrs 9.14 5.66 3.56 2.99 3.24 4.06 4.55 4.05
10 yrs 7.94 4.64 2.65 2.56 3.06 3.81 4.79 3.75
20 yrs 10.74 7.63 4.86 3.95 3.55 3.49 3.72 3.49
30 yrs 13.60 9.95 6.92 5.80 5.05 4.42 4.07 3.52
Panel B: EUR market
2 yrs 9.10 6.77 5.68 5.40 5.55 5.12 4.43 2.95
5 yrs 6.81 4.55 3.36 3.12 3.34 3.19 2.82 2.65
10 yrs 4.20 2.48 1.83 2.14 2.45 2.91 3.17 2.92
20 yrs 7.66 5.67 3.63 3.04 3.04 3.41 3.35 2.89
30 yrs 11.45 9.20 6.48 5.08 4.49 4.40 4.15 3.42
Notes: For each tenor – option expiry category, the table reports the RMSE of the conditional volatility
(annualized and in basis points) of future swap rate distributions under the annuity measure A. In the USD
market, the time series consist of 419 weekly observations from December 19, 2001 to January 27, 2010. In the
EUR market, the time series consist of 449 weekly observations from June 6, 2001 to January 27, 2010.

Table F.4: Fit to conditional swap rate volatilities for SV 2U SS specification

 Tenor Option expiry
1 mth 3 mths 6 mths 9 mths 1 yr 2 yrs 5 yrs 10 yrs
Panel A: USD market
2 yrs 0.12 0.14 0.14 0.12 0.12 0.10 0.08 0.09
5 yrs 0.07 0.08 0.08 0.06 0.06 0.07 0.07 0.09
10 yrs 0.06 0.06 0.06 0.04 0.04 0.06 0.08 0.10
20 yrs 0.04 0.04 0.04 0.04 0.04 0.05 0.07 0.12
30 yrs 0.05 0.04 0.05 0.04 0.05 0.06 0.08 0.11
Panel B: EUR market
2 yrs 0.06 0.08 0.11 0.11 0.11 0.14 0.17 0.22
5 yrs 0.18 0.17 0.16 0.12 0.12 0.10 0.12 0.13
10 yrs 0.18 0.16 0.14 0.10 0.10 0.07 0.09 0.11
20 yrs 0.16 0.12 0.10 0.07 0.07 0.07 0.09 0.12
30 yrs 0.17 0.13 0.11 0.09 0.09 0.09 0.11 0.15
Notes: For each tenor – option expiry category, the table reports the RMSE of the conditional skewness of the
future swap rate distributions under the annuity measure A. In the USD market, the time series consist of 419
weekly observations from December 19, 2001 to January 27, 2010. In the EUR market, the time series consist
of 449 weekly observations from June 6, 2001 to January 27, 2010.

Table F.5: Fit to conditional swap rate skewness for SV 2U SS specification

 USD market EUR market
SV 2U SS,EX SV 3U SS SV 2U SS,EX SV 3U SS
α1 0.0046 0.0067 0.0043 0.0098
(0.0004) (0.0005) (0.0004) (0.0007)

α2 0.0027 0.0060 0.0044 0.0094

(0.0002) (0.0009) (0.0004) (0.0009)

α3 0.0033 0.0052 0.0013 0.0028

(0.0003) (0.0005) (0.0001) (0.0003)

ξ 0.0060 0.0009 0.0010 0.0010

(0.0007) (0.0001) (0.0001) (0.0001)

γ 0.2373 0.2192 0.1377 0.1403

(0.0205) (0.0191) (0.0131) (0.0179)

ϕ 0.0507 0.0478 0.0465 0.0447

(0.0046) (0.0072) (0.0052) (0.0061)

ρ1 −0.2371 −0.3443 −0.2086 −0.2416

(0.0408) (0.0508) (0.0538) (0.0352)

ρ1 0.5812 0.5094 0.4971 0.5919

(0.0561) (0.0379) (0.0699) (0.0621)

ρ2 −0.3746 −0.2085 −0.2399 −0.2689

(0.0442) (0.0430) (0.0377) (0.0636)

ρ2 0.5827 0.5930 0.4284 0.4071

(0.0579) (0.0670) (0.0505) (0.0654)

ρ3 −0.2735 −0.3685 −0.3156 −0.2302

(0.0591) (0.0403) (0.0387) (0.0380)

ρ3 0.4195 0.5941 0.5831 0.5868

(0.0612) (0.0639) (0.0687) (0.0490)

κ1 0.4827 1.2399 0.7911 1.5714

(0.0404) (0.0841) (0.0554) (0.2792)

η1 0.5000 0.5000
(0.0612) (0.0588)

κη 0.3450 0.2533
(0.0768) (0.0402)

ση 0.6287 0.5733
(0.0579) (0.0536)

η 0.1925 0.0978
(0.0266) (0.0101)

Notes: Continued on next page.

Table F.6: Maximum-likelihood estimates of alternative model specifications

74
 USD market EUR market
SV 2U SS,EX SV 3U SS SV 2U SS,EX SV 3U SS
λ0,1 0.0476 0.0020
(0.0620) (0.0548)

λ1 −0.0970 −0.1695 −0.0400 −0.1600

(0.0562) (0.0669) (0.0566) (0.0522)

λ0,1 −0.1007 −0.1137

(0.0592) (0.0971)

λ1 −0.3859 −0.3765 −0.3079 −0.3094

(0.0620) (0.0685) (0.0978) (0.0825)

ν0 0.0168 −0.0436
(0.0709) (0.0788)

ν −0.1224 −0.1215 −0.1672 −0.1498

(0.0525) (0.0890) (0.0530) (0.0824)

ν0 −0.0635 −0.1309
(0.0951) (0.0617)

ν −0.4136 −0.5510 −0.4828 −0.5696

(0.0972) (0.0621) (0.0677) (0.0774)

νe −0.1042 −0.1848
(0.0702) (0.0648)

σrates (bp) 6.5735 6.2676 5.4387 5.4903

(0.7457) (0.5009) (0.5588) (1.0725)

σswaptions (bp) 4.7643 4.0239 4.5467 3.8990

(0.3765) (0.5329) (0.5339) (0.4658)

Log-L ×104 -26.0881 -23.9800 -23.5282 -21.0588

Notes: Maximum-likelihood estimates of the alternative SV 2U SS,EX and SV 3U SS specifications. Both specifi-
cations impose unspanned stochastic skewness restrictions. The sample period is December 19, 2001 to January
27, 2010 in the USD market and June 6, 2001 to January 27, 2010 in the EUR market. Outer-product standard
errors are in parentheses. σrates denotes the standard deviation of swap rate measurement errors and σswaptions
denotes the standard deviation of scaled swaption price measurement errors.

Table F.6: Maximum-likelihood estimates of alternative model specifications (cont.)

75
 Swap swaption volatility skewness kurtosis
rates IV
Panel A: USD market
SV 2U SS 6.07 4.58 4.83 0.06 0.18
SV 2U SS,EX 6.14 4.73 4.87 0.06 0.23
SV 3U SS 6.22 3.43 3.20 0.07 0.21
SV 2U SS,EX − SV 2U SS 0.07 0.15 ∗∗ 0.05 0.01 0.05
(0.17) (2.44) (0.79) (1.10) (1.34)

SV 3U SS − SV 2U SS 0.15 −1.15∗∗∗ −1.63∗∗∗ 0.01 0.03

(0.32) (−5.59) (−5.98) (1.25) (1.06)

Panel B: EUR market

SV 2U SS 7.09 3.31 2.77 0.13 0.50
SV 2U SS,EX 7.03 3.79 2.78 0.16 0.64
SV 3U SS 6.99 2.58 1.42 0.16 0.52
SV 2U SS,EX − SV 2U SS −0.06 0.48 ∗∗∗ 0.01 0.03 ∗ 0.14 ∗∗
(−0.34) (5.84) (0.14) (1.84) (2.42)

SV 3U SS − SV 2U SS −0.04 −0.73∗∗∗ −1.35∗∗∗ 0.04 ∗∗∗ 0.01

(−0.23) (−3.39) (−4.12) (2.72) (1.00)

Notes: The table compares the SV 2U SS,EX and SV 3U SS specifications with the SV 2U SS specification in
terms of their ability to match swap rates, normal implied volatilities (in basis points) as well as conditional
volatility (annualized and in basis points), skewness, and kurtosis of the future swap rate distributions under
the annuity measure A. It reports means of RMSE time series of swap rates, implied volatilities, and swap rate
moments. It also reports mean differences in RMSEs between model specifications. T -statistics, corrected for
serial correlation up to 26 lags (i.e., two quarters), are in parentheses. In the USD market, each statistic is
computed on the basis of 419 weekly observations from December 19, 2001 to January 27, 2010. In the EUR
market, each statistic is computed on the basis of 449 weekly observations from June 6, 2001 to January 27,
2010. *, **, and *** denote significance at the 10, 5, and 1 percent levels, respectively.

Table F.7: Comparison of alternative model specifications

 Panel A: Volatility surface, data Panel B: Volatility surface, SV 2U SS

140 140

120 120
Basis points

Basis points
100 100

80 80

60 60
25 25
10 10 10 10
20 5 20 5
Tenor 30 1 2 Option maturity Tenor 30 1 2 Option maturity

Panel C: Skewness surface, data Panel D: Skewness surface, SV 2U SS

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0

-0.1 -0.1
25 25
10 10 10 10
20 5 20 5
Tenor 30 1 2 Option maturity Tenor 30 1 2 Option maturity

Figure F.1: Actual vs. fitted volatility and skew surfaces in the SV 2U SS specification
Notes: Panel A and B display the average conditional volatilities, measured in basis points, in the data and in
the SV 2U SS specification, respectively. Panel C and D display the average conditional skewness in the data
and in the SV2 specification, respectively. The moments are computed for the USD market under the annuity
measure A. Averages are taken over 419 weekly observations from December 19th, 2001 to January 27th, 2010.

77
 Panel A: Variance and v(t) Panel B: Skewness and w(t)
7 1

0.9
6
0.8
5 0.7

0.6
4
w(t)
v(t)

0.5
3
0.4

2 0.3

0.2
1
0.1

0 0
0 5 10 15 20 -1 0 1 2 3 4 5
1st PC of swap rate variance 1st PC of swap rate skewness

Figure F.2: Interpretation of v(t) and w(t) in the SV 2U SS specification

Notes: Panel A is a scatter plot of v(t) against the first principal component (PC) of fitted conditional swap
rate variance. Panel B is a scatter plot of w(t) against the first PC of fitted conditional swap rate skewness.
The moments are computed for the USD market under the annuity measure A. Each plot contains 419 weekly
observations from December 19th, 2001 to January 27th, 2010.

78