ZHou Proof

Department for
Industrial Economics and Technology Management

Pricing Credit Derivatives

Harald Martin Myhre
Arve Ree
Amund Westbye

Trondheim, 8 December 2003
II
Foreword
We are students in the final year of our Master education in Industrial Economics and
Technology Management at Norwegian University of Science and Technology. This
paper is the result of our project during the autumn 2003, and it corresponds to half study
load for each of us. The main aim of the project is to form a strong foundation for a
Master Thesis within the same study field.

We have chosen to write about credit risk and credit derivatives because it is a relatively
new and rapidly growing area where consensus on standard models has yet to be reached.
Another attractive feature, especially with the structural model we concentrate on, is that
the principles are easily applicable to other areas within quantitative finance.

The CD at the back cover of this paper contains the model we have implemented.

We would like to thank our supervisor, Assistant Professor Sjur Westgaard at the
Department for Industrial Economics and Technology Management, and Professor
Hvard Rue at the Department of Mathematical Sciences, for useful help throughout the
process.

Trondheim, Norway, 8 December2003

Harald Martin Myhre Arve Ree Amund Westbye

Harald Martin Myhre: haraldma@stud.ntnu.no
Arve Ree: ree@stud.ntnu.no
Amund Westbye: westbye@stud.ntnu.no
III
Abstract
Defaults frequently result in unexpected and severe losses for parties involved with the
defaulted company, as has been demonstrated with the big financial scandals of Enron
and WorldCom recently. Traditionally the only means investors and banks have had to
elude the major consequences of bankruptcies has been to diversify their investments and
require collateral. In this aspect the rapid growth of credit derivatives has revolutionized
the financial industry and made it possible to both transfer credit risk to a third-party, as
well as taking on credit risk in otherwise unfunded positions.

The main purpose of this paper is to gain a thorough understanding of credit derivatives
through modeling the underlying credit risk. We try to incorporate both the financial
aspects and motivations behind the use of credit derivatives, as well as the more
fundamental mathematics behind the modeling and pricing of the underlying securities
and derivatives.

In particular, we investigate and implement a model proposed by Zhou (1997a), which
has its foundation in the classical models proposed by Merton and Black and Cox in the
1970s. The model allows for jumps in the market value of firms assets, and we will
show that this added complexity produces results that resemble observed features in the
market considerably better than the original framework. We do not assess empirical data,
but show how Markov chain Monte Carlo could be used to estimate the model. Finally,
we explain the implications of different assumptions and combinations of input
parameters.
IV
Table of Contents
FOREWORD........................................................................................................................................ II
ABSTRACT.......................................................................................................................................... III
CONTENTS ......................................................................................................................................... IV
FIGURES ...............................................................................................................................................V
1 INTRODUCTION......................................................................................................................... 1
1.1 OUTLINE ................................................................................................................................ 2
2 CREDIT DERIVATIVES OVERVIEW....................................................................................... 3
2.1 RAISON DTRE ...................................................................................................................... 3
2.2 THE MARKET ......................................................................................................................... 5
2.3 LEGISLATION AND CONTRACTS ............................................................................................... 6
2.3.1 Default .............................................................................................................................. 7
2.3.2 Recovery ........................................................................................................................... 7
2.4 PRODUCTS.............................................................................................................................. 8
2.4.1 Asset Swap ........................................................................................................................ 8
2.4.2 Total Return Swap............................................................................................................. 8
2.4.3 Credit Default Swap .........................................................................................................10
2.4.4 Credit-Linked Notes..........................................................................................................11
2.4.5 Collateralized Debt Obligation.........................................................................................11
2.4.6 Exotic Structures ..............................................................................................................12
3 CREDIT RISK MODELING.......................................................................................................13
3.1 STRUCTURAL MODELS...........................................................................................................13
3.1.1 The Merton Model (1974).................................................................................................14
3.1.2 The Black and Cox First Passage Time Model...................................................................16
3.1.3 Practical Implementations of Structural Models................................................................17
3.2 REDUCED FORM MODELS........................................................................................................19
3.2.1 Intensity Based Modeling..................................................................................................19
3.2.2 Recovery Modeling...........................................................................................................21
3.2.3 Credit Rating Modeling ....................................................................................................22
3.2.4 Jarrow and Turnbull (1995)..............................................................................................23
3.2.5 Jarrow, Lando and Turnbull (1997) ..................................................................................25
3.2.6 Duffie and Singleton (1999) ..............................................................................................27
3.3 HYBRID MODELS...................................................................................................................28
3.3.1 Giesecke (2001)................................................................................................................28
4 ZHOUS MODEL........................................................................................................................31
4.1 THE MODEL ..........................................................................................................................31
4.2 CLOSED-FORM SOLUTION ......................................................................................................33
4.3 IMPLEMENTATION..................................................................................................................37
4.3.1 Primer on Monte Carlo Methods.......................................................................................37
4.3.2 Solution............................................................................................................................38
4.3.3 Implementation Issues ......................................................................................................39
4.4 PRICING CREDIT DEFAULT SWAPS..........................................................................................41
4.5 RESULTS ...............................................................................................................................41
4.6 CALIBRATION........................................................................................................................47
4.6.1 Markov Chain Monte Carlo..............................................................................................49
4.6.2 Calibration Methodology..................................................................................................50
4.7 SUGGESTED EXTENSIONS.......................................................................................................52
5 CONCLUDING REMARKS .......................................................................................................55
6 BIBLIOGRAPHY........................................................................................................................56
A ZHOU.C DLL PRORAM CODE................................................................................................... I
B VBA PROGRAM CODE FOR CLOSED FORM SOLUTION................................................VII
C VBA CODE FOR CHANGING DIFFUSION VOLATILITY AND JUMP VOLATILITY..... IX
D MCMC RESULTS.....................................................................................................................XII
E R CODE FOR MCMC............................................................................................................. XIV
V
Figures
FIGURE 2.1 ASSET SWAP........................................................................................................................... 8
FIGURE 2.2 TOTAL RATE OF RETURN SWAP ................................................................................................ 9
FIGURE 2.3 CREDIT DEFAULT SWAP..........................................................................................................10
FIGURE 2.4 CREDIT LINKED NOTES............................................................................................................11
FIGURE 3.1 JARROW AND TURNBULL (1995).............................................................................................24
FIGURE 3.2 SHORT-TERM CREDIT SPREADS IN THE CASE OF INCOMPLETE INFORMATION (DUFFIE AND LANDO
2001) .............................................................................................................................................28
FIGURE 4.1 CONVERGENCE OF MONTE CARLO SIMULATION.......................................................................40
FIGURE 4.2 VARYING NUMBER OF TIME STEPS IN MONTE CARLO SIMULATION...........................................40
FIGURE 4.3 VARYING JUMP INTENSITY AND JUMP VOLATILITY WHILE KEEPING TOTAL ASSET VOLATILITY
CONSTANT......................................................................................................................................42
FIGURE 4.4 VARYING JUMP AND DIFFUSION VOLATILITY WHILE KEEPING TOTAL ASSET VOLATILITY AND JUMP
INTENSITY CONSTANT .....................................................................................................................43
FIGURE 4.5 VARYING THE INITIAL DEBT TO ASSET VALUE..........................................................................44
FIGURE 4.6 VARYING THE RISK-FREE INTEREST RATE ................................................................................44
FIGURE 4.7 CORRESPONDING SIMULATION AND CLOSED-FORM SOLUTION WHILE VARYING JUMP AND
DIFFUSION VOLATILITY AND KEEPING TOTAL ASSET VOLATILITY AND JUMP INTENSITY CONSTANT.....45
FIGURE 4.8 CORRESPONDING BOND SPREAD WITH CREDIT DEFAULT SWAP SPREAD .....................................46
FIGURE 4.9 CONVERGENCE OF ALGORITHM CALCULATING FIRM VALUE. INITIAL FIRM VALUE IS SET TO 5000,
FACE VALUE OF DEBT TO 1500, EQUITY TO 1550 AND OTHER PARAMETERS TO THE VALUES USED IN
CHAPTER 4.5...................................................................................................................................48
FIGURE 4.10 TRACE PLOT OF DRIFT DIFFUSION PARAMETER IN MCMC ESTIMATION ...................................52
Tables
TABLE 2.1 FUNDING EXAMPLE: INITIAL REVENUES..................................................................................... 4
TABLE 2.2 JP MORGAN REVENUES WITH CDS............................................................................................ 5
TABLE 2.3 EBC REVENUES WITH CDS...................................................................................................... 5
TABLE 2.4 MARKET SHARE BY INSTRUMENT (SCHNBUCHER 2003)........................................................... 6
TABLE 2.5 MARKET SHARE BY INSTITUTION (LEHMAN BROTHERS 2001).................................................... 6
TABLE 2.6 TOTAL RETURN SWAP: INDEPENDENT BANKS ............................................................................ 9
TABLE 2.7 TOTAL RETURN SWAP: BANKS ENGAGING IN A TRS................................................................... 9
TABLE 3.1 PAYOFFS FROM THE MERTON MODEL ......................................................................................14
TABLE 3.2 AVERAGE ONE-YEAR RATING MIGRATION RATES 1985-2001 (MOODYS 2002)...........................22
TABLE 3.3 INFORMATION WITH GIESECKE ................................................................................................29
TABLE 4.1 DESCRIPTION OF PARAMETERS USED IN ZHOU'S MODEL.............................................................31


1
1 Introduction
Financial risk is a measure of adverse changes in a financial position as a result of
changing environment. Broadly, financial risk may be divided into two categories:
Market risk stems from general economic variations while credit risk
1
is due to
uncertainty and changed perceptions of a debt issuers credit quality. (Giesecke 2001)
Credit risk may be further divided into components that have to be carefully considered
when deciding on the complexity and accuracy of credit models. Most obvious is the
arrival risk as a term for whether a default will occur or not, often measured by the
probability of default. Secondly, there is uncertainty about the timing of the default,
which expectation typically varies significantly between large, stable firms and risky
ventures. A third factor, which may often be just as important as the arrival and timing
risk, is the recovery risk. It is related to the uncertainty about the payoff that creditors will
receive after default. Finally, there is default correlation risk which is the risk of several
joint defaults in a time period. (Schnbucher 2003)

Although continuous-time finance has been used to value defaultable securities since the
initial proposal of Black and Scholes (1973) it is particularly in the last decade that this
area of research has received attention. Banks have devoted more attention to this task.
The European Monetary Union increased liquidity and competition and made credit risk
the key determinant of price differences. Furthermore, the historically low interest rates
have forced investors to accept more credit risk to maintain high yields. Most important
has been the transition where internal risk models are becoming increasingly accepted as
a basis for regulatory capital prescriptions. A system where internal risk models are
accepted as to determine adequate capital reserves will give significant advantages to
banks that have such models in place.

With the introduction of credit derivatives in the early 1990s investors have been able to
separate and trade credit risk according to their own risk profile. The rapid growth of
credit derivatives and the many still remaining challenges regarding standardization of
pricing models have motivated this paper with the objective summarized below.
This paper explores the vast area of credit derivatives and the literature
on modeling credit risk. We aim to gain insight in the financial as well as
the mathematical foundation of credit risk and derivatives, in particular
by investigating one specific model in depth.

1
Credit risk and default risk are used synonymously throughout this paper.

2
1.1 Outline
The rest of the paper is structured as follows: Chapter 2 gives to a general overview of
credit derivatives and explains the motivation for the use of credit derivatives with
background in market needs and regulations. Additionally, we give an in-depth analysis
of some of the most popular credit products available in market.

In chapter 3 we move on to discuss credit risk models proposed in the literature and
compare those to some models actually being used by credit rating agencies and
investment banks. The models can be separated into two main categories: structural
models and reduced form models. The structural models were initially proposed by Black
and Scholes (1973) and Merton (1974) and model the evolvement of firm value as a
diffusion process where default is occurring when the firm value goes below a barrier.
The reduced form approach contains models with a variety of foundations. However,
what they all have in common is that the link to the firms fundamentals is abandoned,
and the time of default is modeled as a totally inaccessible stopping time with an
exogenously given intensity.

Finally, in chapter 4 we analyze a jump-diffusion model proposed by Zhou (1997a),
which incorporates features from both approaches mentioned above and mitigates the
differences between model output and empirical data. For this particular model we derive
a closed-form solution for the pricing of equity in a special case. By building a model
with Monte Carlo in the C programming language we create a solid and flexible
framework that easily adapts to problems faced by real world situations. Though it may
be too detailed for practical purposes, we focus on understanding the exact way pricing of
securities and credit derivatives should be done, and incorporate it into our model. We
have focused on implementing the model in such a way that it in particular gives a
thorough analysis on how the model parameters influence the output. Estimating these
parameters is non-trivial because the value of a firms assets is neither observable in the
market, nor it can be retrieved directly from the balance sheet. We suggest methods for
estimating the firm value from equity data and thereby calibrating the model.

3
2 Credit Derivatives Overview
Credit derivatives were introduced to the market in the beginning of the 1990s. Despite
their short history their use have grown rapidly. Credit derivatives are now used not only
by banks, but also by various funds, insurance companies and even corporations, in order
to separate and trade credit risk.

The term credit derivatives is being used for a broad range of securities. Schnbucher
(2003) gives one definition of credit derivatives which covers the different securities
presented in this paper:
A credit derivative is a derivative security whose payoff is materially
affected by credit risk. (Schnbucher 2003, p 8)
In the following we will explain the motivation for the use of credit derivatives and give
an overview of the market. Moreover, we briefly explain the most general legislation
governing the use of credit derivatives, before we in depth explain the most important
credit derivatives currently available.
2.1 Raison dtre
The primary purpose of credit derivatives is to enable investors to transfer and repackage
credit risk (Das 2000). Basic credit derivatives provide an efficient way to replicate in a
derivative form the credit risk that would otherwise be present in a standard cash
instrument. More exotic credit derivatives enable the credit profile of a particular asset or
group of assets to be split up and redistributed into a more concentrated or diluted form
according to the various risk preferences of investors. (Lehman Brothers 2001) By
exploiting these features, investors can use credit derivatives to quantify their risk and
highlight credit risk concentrations (SunGard 2002). Other motives for entering into
credit derivatives contracts spans from hedging, speculation and diversifying, to taxation
issues (Schnbucher 2003; Tavakoli 2001).

In general there exists no equivalent replication of a credit derivative
2
. In other words,
there are no good alternatives if one wants to trade credit risk. The lack of equivalents
adds to the difficulty of pricing credit derivatives, because otherwise one could have used
replication pricing. But lack of alternatives makes the market expand, since people have
to use them if they want to obtain the benefits from credit derivatives. Below we explain
why regulation also has induced a need for credit derivatives.

2
Some features may be obtained through the use of e.g. special purpose vehicles, see chapter 2.4.5.

4
Most countries with internationally active banks have adapted their legislation to an
international standard, the July 1988 Basel Capital Accord, Basel I. This is introduced by
the Basel Committee on Banking Supervision, also called the BIS group after the Bank
for International Settlements which facilitates the meetings. (Bank for International
Settlements 2003)

Through this accord banks are required to hold capital to protect against unexpected
losses, and the agreement links the amount of capital to the risk involved for the lending
banks. However, the categories of risk are only differentiated in broad categories: banks
have to set aside as much capital against a loan to Microsoft as to a Hungarian dotcom, as
much against a loan to the US as one to South Korea (Economist 2002). Credit
derivatives provide a way to ease the capital requirements for banks. If a bank has lent to
the private sector, it would normally need to set aside 100% BIS risk weight times 8% of
the face value. However, if it buys protection from an OECD bank in form of a credit
derivative, it only needs to hold 20% * 8% = 1.6%. (Lehman Brothers 2001; Tavakoli
2001)

Let us look at an example. Liberty Corporation, rated AA-, is seeking $100 mill of
financing over five years priced at LIBOR + 75 bps. Both JP Morgan Chase (JPMC) and
European Banking Corporation (EBC), rated A+, have been approached to be a single
lender in the transaction. JPMC is able to fund the loan at LIBOR 10 bps, while EBCs
cost of capital is LIBOR + 25 bps. Both banks are subject to the requirements from BIS,
namely they must hold 8% against risk-weighted assets times 100% weight for
corporations. Assume LIBOR is currently 3%.

Each bank may fund Liberty with the below indicated return on capitals.

JP Morgan Chase European Banking Corp.
Interest Income 3,750,000 $ 3,750,000 $
Interest Expense -2,900,000 $ -3,250,000 $
Net Interest Margin (1-2) 850,000 $ 500,000 $
Return on Capital 10.63% 6.25%
Table 2.1 Funding example: Initial revenues
Now suppose JPMC enters into a credit default swap (CDS)
3
with EBC, where EBC
takes on the credit risk, and thus need to hold 100% of 8% of notional, in return for a 60
bps spread. Now the spreadsheets become:

3
In effect a credit risk insurance, discussed in chapter 2.4.3.

5
Table 2.2 JP Morgan revenues with CDS
JPMC
Interest Income 3,750,000 $
Interest Expense -2,900,000 $
Net Interest Margin 850,000 $
CDS Fee -600,000 $
Net Income 250,000 $
Return on Capital 15.63%
Table 2.3 EBC Revenues with CDS
EBC
CDS Fee 600 $
Return on Capital 7.5%
Both banks are better off with the credit default swap, and the customer was not affected.
This simple principle illustrates how the BIS regulations have helped to propel the use of
credit derivatives.

The above example also illustrates another point. Todays regulations are not
sophisticated enough since they treat totally different risks the same way. Therefore, the
BIS group is working on an upgrade, the Basel II. The changes will not be effective until
January 2007, but enough is ready to know that the impact will be substantial. The Basel
II proposals are not clarifications or amendments, but complete revision of the Capital
Accord of 1988 (Kesdee 2003). The Basel II will decrease the capital requirements for
many classes of credit risk, offset by a totally new charge for operational risk. That leaves
the overall minimum regulatory capital in the banking system about the same as now.
The details for credit derivatives usage await the final agreement. (Economist 2003)

The numerical example above also illustrates another motivation for the swap. This may
arise when the bank for some reason consider itself obliged to give a loan, for instance to
a loyal and important customer. In this case the bank may provide even a risky loan. The
bank can then insure itself from the risk by entering an insurance position like the one
exemplified above, without the knowledge of the customer and the according
implications on the customer relationship.
2.2 The Market
The credit derivatives market was created in London and New York in the early 1990s,
propelled by the Based I accord. After the first half of 2003 the notional outstanding
amounted to 2.69 trillion dollars, and trade grew by 25% in the first six months of 2003
(ISDA 2003).
4
The market share by instrument type as of 2002:

4
Beware that the statistics as trade is mainly OTC, and the statistics based on surveys.

6
Instrument Share
Credit default swaps (including First to Default swap) 67%
Synthetic balance sheet Collateralized Loan Obligations 12%
Tranched portfolio default swaps 9%
Credit-linked notes, asset repackaging, asset swaps 7%
Credit spread options 2%
Managed synthetic Collateralized Debt Obligations 2%
Total return swaps 1%
Hybrid credit derivatives 0.2%
Table 2.4 Market share by instrument (Schnbucher 2003)
We see that credit default swaps are by far the most traded credit derivative. This is a
relatively plain structure, and other varieties may gain market share as the market
matures.

Below are the market shares based on trade participants
5
.

Counterparty Protection Buyer Protection Seller
Banks 63% 47%
Securities firms 18% 16%
Insurance companies 7% 23%
Corporations 6% 3%
Hedge funds 3% 5%
Mutual funds 1% 2%
Pension funds 1% 3%
Government/Export credit agencies 1% 1%
Table 2.5 Market share by institution (Lehman Brothers 2001)
The variety of applications attracts a variety of market participants. One obstacle for a
even broader use is the difficulties regarding pricing. Since there is no widespread pricing
standard for credit derivatives of either type, the actors need to be of a certain size to have
the needed competence and resources. (Tavakoli 2001)
2.3 Legislation and Contracts
There has been a debate over whether credit derivatives are to be considered insurance,
and the legislation for them is in general incompletely developed. (Das 2000) In order to
facilitate over the counter (OTC) trade, the International Swaps and Derivatives
Association Inc., ISDA, has authored the Sovereign Master Credit Derivatives
Confirmation Agreement. This is a template for contracts that parties may voluntarily
refer to in their contracts. Today, 95% of credit derivative contracts are based on the
standardized ISDA framework (Schler 2001).

5
Again caution must be taken since the numbers are based on surveys, but they provide a general picture.

7
2.3.1 Default
Normally in the credit derivative contract, a credit event will trigger the equivalent to the
payment of an insurance payment. Therefore the terminology used is crucial. Even the
meaning of default is not fully straightforward. ISDA define the notion of default to be
one of failure to pay or deliver, breach of agreement, or credit support default (ISDA
1992). Schnbucher (2003) gives some more other examples of credit events that may
trigger a credit derivative payment, including restructuring and rating downgrades to or
below a specified level. Though the practitioner needs to be aware of the exact juridical
definitions, we take the notion of default as given in this paper. However, in calibrating
the model, one will need to pay attention if one uses market default data.
2.3.2 Recovery
After a default another important issue arise, namely recovery. Here is recovery used as
the percentage of notional a creditor receives after a credit event. If the recovery had been
100%, a default had been no issue at all. In practice this value depends on many factors,
thus making it very difficult to predict, even for similar companies. How to model
recovery in a credit risk framework will be further discussed in chapter 3.2.2.

In fact, recovery after bankruptcy may be a lengthy process involving many parties and,
thus, major costs. As far as credit derivatives are concerned, they are normally settled
within weeks of default, giving the protection buyer the contingent payoff within a
predictable time period. However, difficulties arise as uncertainty about recovery still
remains after the default. Most often the credit derivative contract should contain
information on how to deal with these problems, but there are also situations where a
third-party will enter into the process and give a fair value of recovery.

Now consider a company currently in, or on the way into, financial distress, resulting in a
low bond value. Suppose one of the bondholders buy credit protection on the asset
through a CDS contract. This bondholder will now have incentive to force the firm into
bankruptcy, for example by refusing to extend or renegotiate the loan contract in
situations he would have otherwise done it to avoid immediate default. The reason lies in
the fact that the CDS will pay off according to the face value of the debt and not on the
difference between the real value of the bond and the recovery. Additionally he will also
get away without paying the fixed contingent payments until the maturity of the CDS.
Obviously this may raise ethical questions and give incentives to declare otherwise
evitable bankruptcies.

Debt is issued with different seniorities giving the creditor varying security in the case of
default. After a default the most senior bonds are supposed to pay off completely before
more junior debts are paid off. In practice there are many legal aspects of a bankruptcy
resulting in that this absolute priority rule is not upheld. Nevertheless, expected recovery
for senior debt is still substantially higher than for junior debt.

8
2.4 Products
This chapter is devoted to an overview of some important credit derivatives products. By
far credit default swaps are the mostly traded credit derivative. As shown in the market
chapter they constitute approximately two thirds of the market when we include basket
and first-to-default derivatives. Still there exist a large variety of products accommodated
to certain situations and we introduce here some of the basic structures.
2.4.1 Asset Swap
The asset swap package converts the payoffs of a defaultable (coupon) bond into a payoff
stream of LIBOR plus a spread. This spread is chosen to give the whole package a value
of par. Thus, we see that the contract is a combination of a defaultable bond and an
interest-rate swap contract.

Asset swap buyer
Defaultable bond
Asset swap seller
Coupon payments
LIBOR +asset swap spread
Coupon payments

Figure 2.1 Asset swap
In its strictest form the asset swap is not a credit derivative. The reason is that the
payments of the swap will continue even if a credit event occurs. Hence, it is merely a
way of securing cash flows than separating and trading credit risk.

The market for asset swaps is very liquid and frequently it is the underlying asset for
other derivatives, such as asset swaptions. There exist many additional features that can
be included in the contract to give it the desired payoffs.
2.4.2 Total Return Swap
The total return swap (TRS), also called total rate of return swap (TRORS), adds another
perspective to the asset swap. With the TRS the counterparties agree to exchange all cash
flows from two different investments, normally one of them defaultable and the other one
considered not to be. Basically this structure does an exchange of the payoffs without
legally transferring the ownership of the asset.

9

TRS payer (A)
Defaultable reference
asset
TRS receiver (B)
Total rate of return
Asset payments
LIBOR +TRS spread
Mark-to-market

Figure 2.2 Total rate of return swap
The TRS payer gives all the payments of a defaultable reference asset to the TRS
receiver, at the same time as he receives LIBOR + TRS-spread. Additionally the contract
is marked to market at regular intervals. If the reference asset has appreciated, party A
must pay the difference to B, and if the asset has depreciated B must pay the difference to
A. The default event is handled by the latter case, as that will correspond to a depreciated
value of the asset, and the TRS receiver must pay A the difference between the pre-
default value and the recovery value.

As an example of a TRS we may take the following situation described in Tavakoli
(2001). Consider two banks with different ratings and funding costs that want to invest in
a BBB rated asset Table 2.6 gives an overview of the costs, payments and earnings for
these two actors when the purchase is committed independently. Both banks have a 100%
BIS risk weight.

Bank Rating Purchased Asset Risk Weight Coupon Payments Funding Cost Net Spread
Payer: AA BBB 100% LIBOR + 65 bps LIBOR 15 bps 80 bps
Receiver: A- BBB 100% LIBOR + 65 bps LIBOR + 30 bps 35 bps
Table 2.6 Total return swap: Independent banks
By entering a TRS with the A- bank, the AA bank is able to give the A- bank a more
favorable funding cost. Contemporaneously AA is hedging both its credit and market
risk, and can reduce its capital charge on the transaction from 100% to 20%
6
while still
remaining the legal owner of the asset. For many investors this might be the only
possibility to attain a short position in specific assets due to the fact that directly shorting
of defaultable bonds often is impossible. Hence, the TRS receiver benefits from lower
funding costs. The TRS payer benefits from lower joint default probability between the
receiver and the reference asset, thus reducing its risk weight in the position. (Table 2.7)

Bank Rating Purchased Asset Risk Weight Coupon Payments Funding Cost Net Spread
Payer: AA A+ 20% LIBOR + 15 bps LIBOR 15 bps 30 bps
Receiver: A- BBB 100% LIBOR + 65 bps LIBOR + 15 bps 30 bps
Table 2.7 Total return swap: Banks engaging in a TRS

6
Assuming that both banks are OECD banks.

10
The TRS can be used to defer losses because of the off-balance sheet feature of the
contract. An investor with an unrecognized loss in a bond position can defer the loss by
entering a TRS and then recognize the loss at maturity of the contract. (Tavakoli 2001)

For the TRS receiver, a TRS might be applied in a variety of ways in practice, but the
primarily use is that of financing. There are no initial investments, with the exception of a
collateral, which makes it possible for party B to take advantage of leverage. Some total
return swaps can be regarded as a new asset with specific maturity and features currently
not available in the market. In other cases the contract might be the only way B can invest
in the reference asset due to legal restrictions. (Tavakoli 2001)
2.4.3 Credit Default Swap
A credit default swap (CDS) is an exchange of a fee for a payment in case a credit default
event occurs on a reference asset. Other names for the CDS include credit default options
and default swaps. A CDS will make it possible to separate and trade purely in the credit
risk of different underlying assets, e.g. loans, bonds and receivables. With the aid of this
contract it is possible for a company to hedge its credit risk exposure at the same time as
an investor might speculate in the reference asset's future.

The contract consists of two parties; one protection buyer and one protection seller. The
buyer is paying a fee up front or a fee amortized over the life of the security. In the case
of a credit event the seller will have to pay a contingent payment. This payment can take
many forms and will need a thorough specification in the contract. The most common
alternatives include physical delivery of the reference asset against repayment at par, cash
settlement based on post-default market value and a pre-agreed fixed payoff, more
commonly called a digital default swap. Neither of the parties need to have a funded
position in the reference asset, but still the physical settlement is widely used as default
payment.

Protection buyer
Reference asset
Protection seller
Fee
Contingent payment upon default

Figure 2.3 Credit default swap
Notice the importance of the correlation between the protection seller and the reference
asset. If these two parties are perfectly correlated, the CDS will be of reduced value, as a
default on the reference asset will imply a default on the seller as well. Preferably we
would like a protection seller with zero or negative correlation with the reference asset
and as high credit rating as possible. In practice these criteria would have to be relaxed in
order to make the contract affordable.


11
The variations between different CDS contracts might be large and the specifications of
the contract might be adapted to the needs of the counterparties. ISDA has introduced a
set of common specifications that serve as a standard for credit default swaps, and there
exists a relatively liquid market for these instruments. Most credit default swaps are
quoted with five years to maturity.
2.4.4 Credit-Linked Notes
Credit-linked notes (CLN) are a fixed-income security with an embedded credit
derivative. There are several versions of credit-linked notes, but we will here focus on the
family of credit-linked structured notes, and more specifically on credit default linked
notes (Das 2000), which embed a CDS with the fixed-income security.

CDS
Issuer
Derivatives dealer
Investor
Principal/ recovery rate
Coupon
Investment

Figure 2.4 Credit linked notes
The issuer of the CLN is normally a high rated trust (AA or AAA). It is paying the
investor a fixed or floating coupon based on the chosen reference asset. If there are no
credit events the investor receives the principal of maturity, otherwise he will receive the
recovery rate or any other pre-agreed default payment version. At the same time the
issuer sells credit protection to a derivatives dealer through a CDS contract. The
payments from the CDS contract are further used to create coupon payments to the
investor. Hence, the issuer has created a synthetic bond, which gives the investor the
opportunity to obtain a position without making the direct investment in the reference
asset.

Das (2000) lists some of the main difficulties of direct investment, especially in high-
yield and emerging markets. Firstly there are regulatory issues and high transaction costs
related to foreign investments. Secondly there is often a lack of underlying securities and
liquid markets corresponding to the investors preferences. The credit-linked notes open
the market to a broader range of investors. Moreover, as buying a CLN is similar to a
fully collateralized sale of credit protection, the number of investors that can sell default
protection is significantly increased.
2.4.5 Collateralized Debt Obligation
Collateralized debt obligations (CDO) are used to securitize portfolios of defaultable
assets. The portfolio of assets is transferred into a specially created company, a special
purpose vehicle (SPV). The SPV is then issuing notes with loss layers, which are used in

12
the case of a default during the existence of the CDO. When the CDO is liquidated the
senior tranches are paid off first, and further the other notes according to their ranking.
The tranches of lower seniority serve as protection for the tranches of higher seniority.

CDOs allow investors to invest in notes that they would otherwise not be allowed to
invest in and adapt the desired risk profile. The two most common forms of collateralised
debt obligations are arbitrage CDOs and balance sheet CDOs. The former aims to
arbitrage the price between the components of the underlying portfolio with the sale price
of the CDO-notes while the latter tries to free up regulatory capital tied up in the
underlying loan
7
or bond
8
portfolio.
2.4.6 Exotic Structures
New credit derivatives have popped up regularly with more advanced features than the
plain vanilla structures introduced so far. The easiest of these include options and
forwards on a credit default swap. Furthermore we have asset swap switches, different
knock-in-options and credit spread options. One of the most promising and interesting
structures is the purchase of protection on a basket of credits. For a portfolio manager this
means that he can purchase a first-to-default structure on several assets instead of buying
them separately. Such a diversified basket will obviously be cheaper than a CDS on each
of the reference assets. We refer to Tavakoli (2001) and Das (2000) for a more thorough
presentation and analysis of these exotic structures.

7
Collateralized Loan Obligation (CLO)
8
Collateralized Bond Obligation (CBO)

13
3 Credit Risk Modeling
There are two main approaches to credit risk modeling. The first approach includes the
firm-value or structural models initially proposed by Black and Scholes (1973) and
Merton (1974). In these models the value of the firm follows a diffusion process. Default
is modeled as when the value of the firm goes below a boundary, usually a function of the
capitalization of the company. The main advantage of the structural approach is that it is
based on solid economic arguments and it models default in terms of fundamental firm
variables. Critics against this approach point out that time of default will be a predictable
stopping time because of the continuity of the process. Therefore, as time to maturity
goes to zero, credit spreads should also approach zero which again is not consistent with
empirical evidence.

The second approach is the intensity based or reduced form models. In this approach the
link to the firms fundamentals is abandoned, and the time of default is modeled as a
totally inaccessible stopping time with a default intensity. The advantage of this modeling
approach is in particular its tractability and empirically good performance in that the
models are easier calibrated to fit observed data than structural models are. Reduced form
models, however, do not formulate economic arguments about why a firm defaults. One
rather takes the default event and its stochastic structure, and frequently the recovery
process as well, as exogenously given.

The basic underlying assumption for the models described in the following is that there
exists a unique equivalent martingale measure
P making the markets for default-free and

risky debt complete and arbitrage-free. By using this technique pricing is done by
discounting expected values of payoffs under
P leading to greatly simplified valuation

problems. (Neftci 2000; Johannes and Polson 2002)

In the following we will present the basic theory behind the two approaches. In particular,
we will investigate the theoretical structural models and correspond them to the way
models are actually implemented in practice. Further we describe reduced form models
and how default intensities and recovery rates are modeled, before we discuss the most
referenced models. Finally, we give a brief overview of hybrid models, which incorporate
features from both approaches.
3.1 Structural Models
In a structural model the starting point is the value of the assets, V, of a firm that has
issued bonds. This value is assumed to change stochastically, usually following a
lognormal diffusion process

14
dV Vdt VdW +
where is the expected drift, s is the firms volatility and dW is a standard Brownian
motion. The value of the firm is then used to value all claims on the firm as derivative
securities with the assets as the underlying.

A default can either be triggered only at maturity of the debt if V is insufficient to repay
the debt, as in the Merton model, or it can be modeled as a down-and-out barrier, as in
equity options allowing the firm to default as soon as V hits a pre-specified barrier.
Below we investigate the two most famous models describing these events, namely the
Merton model and the Black and Cox model.

Although this kind of models has proven very useful in addressing the quantitatively
important principles of pricing credit risk, they have been less successful in practical
applications. This owes to the difficulty of modeling realistic boundary conditions. The
boundaries include both the conditions under which default occurs, and in the event of
default, the division of the value of the firm among claimants.
3.1.1 The Merton Model (1974)
In the classical Merton approach a firm is financed by a zero-coupon bond with face
value K and maturity date T. Default can only occur at maturity of the debt.

Assets Bonds Equity
No Default
T
V K K
T
V K
Default
T
V K <
T
V 0
Table 3.1 Payoffs from the Merton model
According to Black and Scholes (1973) by issuing debt, equity holders sell the firms
assets while keeping a European call option on the firms value. The strike is equal to the
face value of the debt. Similarly, the debt can be viewed as a default free bond and a short
European put on the firms value, with strike equal to the face value of the debt. Thus, we
have the following payoffs to the firms liabilities at time T, where ( ) , , B V t T is the price
of the risky bond at time t and maturity T, and S(T) the value of equity:
( , , ) min( , ( )) max(0, ( )) B V T T K V T K K V T
( ) max(0, ( ) ) S T V T K
Although V is not a traded asset, the stock as a derivative of V is. Assuming no arbitrage
opportunities Merton shows that the valuation of derivatives of V will be independent of
investors risk preferences. It therefore exists an equivalent probability measure
P under
which V is a martingale if discounted at the continuously compounded riskless rate r(t).If
we assume that interest rates are constant r(t)=r>0 we can use the classic Black-Scholes
formula to derive the price of the equity and the risky bond:

15
1 2
(0) ( , , (0, ), , (0)) (0) ( ) (0, ) ( )
c v
S BS T B T r V V N d B T N d
1 2
(0, ) (0, ) ( , , , , (0)) (0) ( ) (0, ) ( )
p v
B T B T BS T K r V V N d B T N d +
where
(0, )
rT
B T Ke
,
2
1
(0) 1
ln( ) ( )
2
V
V
V
r T
K
d
T
+ +
and
2 1 V
d d T
From the equation for the risky bond B and the risk free bond B we can easily compute
the bond and credit spread:
1 1
(0, ) (0)
( ) ( )
(0, )
v rT
B T V
N d T N d
B T Ke

1 1
(0)
ln( ( ) ( )
( )
V rT
V
N d T N d
Ke
CS T
T

+

We further find the actual and risk-neutral probabilities of default by solving the initial
SDE
( )
( )
( )
v
dV t
dt dW t
V t
+ which has a unique solution
2
1
( ) ( )
2
( ) (0)
v v
t W t
V t V e
+

[ ]
2
2
1
( ) ( )
2
1
ln ( )
(0) 2
( ) (0) ( )
v v
v
t W t
V
K
T
V
P V T K P V e K P W T

+
1 _

1
1
,
1
< < <
1
1
]
1
]

As W(t) is normally distributed with mean zero and variance T we get
[ ]
2
1
ln ( )
(0) 2
( )
v
V
K
T
V
P V T K N
T

_ _

,

<

,

Similarly to obtain the risk-neutral probabilities we substitute with r to get
[ ]
2
( ) ( ) P V T K N d <
To estimate the default barrier, K, one can use balance sheet data. In the classical Merton
model, K is assumed to be equal to the face value of debt. r can be estimated from
default-free bonds (Treasury bonds), though some models give better empirical results
when interbank rates are used (Schnbucher 2003). (0) V and
V
can be estimated by

16
first observing the equity value (0) S and the volatility
S
. We can then find the
parameters (0) V and
V
by solving two equations. The first is obtained from the equity
pricing formula in Mertons model above, and the second by applying Itos formula to
the equity value:
1
(0) ( ) (0)
S v
S N d V
Extensions to the Merton Model
In the original Merton setting, all bonds are considered to be pure discount bonds. A
simple extension to the Merton model is to treat coupon bonds as a portfolio of zero
coupon bonds each of which can be priced with the original Merton model. A more
precise treatment of risky coupon bonds would be the Geske (1977) model. Geske
assumes that equity holders have a compound option on the firms value, and that they
decide to pay the coupon or not. If the equity holders do not pay the coupon, the firm will
default and the default boundary is thus modeled endogenously. The value of the option
equity holders have can be expressed in terms of multivariate normal distribution with
dimensions depending on the number of coupon payments. The Merton model has also
been extended to callable and convertible bonds, variable rate bonds and bonds with
different seniorities. In JP Morgans E2C model, which we will treat later, they use the
original and simple Merton model and extend it with a variable default barrier. The KMV
model is also loosely based on the original Merton framework.
3.1.2 The Black and Cox First Passage Time Model
The Merton model described in the previous section does not allow bankruptcy before the
maturity of the bond. In order to allow this Black and Cox (1976) modified the Merton
model to allow bondholders to force bankruptcy during the lifetime of the security. These
types of models are often referred to as first time passage or first passage time models.

In addition to allowing the debt to default before maturity Black and Cox also assume
that the shareholders receive a continuous dividend payment proportional to the firm
value V , which gives the following SDE under

P :
( ) ( )( ) ( ) ( ) dV t V t r dt V t dW t +
0 > represents the dividend rate. Interest rate risk was not included in the original
Black and Cox model so interest rates are constant, r(t)=r>0.

Also, unlike the Merton model, the default barrier is made time dependent. For the firm
to default before maturity of the debt, the firms value must hit a default barrier
( ) T t
Ke

.
Otherwise the firm can default at maturity if the firms value is less or equal to the face
value of the debt, ( ) V T D . The default time is modeled by
inf{ 0 : ( ) ( )} t V t K t >
In the case with constant interest rates, Black and Cox provide a closed-form solution.

17

Black and Cox also provide an argument that allows capital structures with different
seniority of the debt. The price of a junior bond can be derived from the price of a senior
bond by assuming that the junior bondholders will only be paid off after the senior
bondholders have been paid, similar to the waterfall in securitization. If D=S+J is the
total debt owed at maturity to senior bondholders, S, and junior bondholders, J,
combined, then the price of the senior bond can be found as if there were no junior debt
( , 0, ) B S T . Then the price of the junior debt will be the difference between the bonds
considering the total debt and the bonds only considering the senior debt
( , , ) ( , , ) ( , , ) B J t T B D t T B S t T
This priority assumption holds for for instance CDOs and securitization products. But as,
mentioned in previous chapters, absolute priority is not upheld in corporate bankruptcy
proceedings this assumption does not hold for corporate bonds. All other first time
passage models as described below therefore assign a specific function for recovery with
different parameters for different seniorities.
Extensions to the Black and Cox model
Longstaff and Schwartz (1995) extend the Black and Cox model to allow interest rates to
be stochastic, with dynamics as proposed by Vasicek (1977). Moreover, they do not
require the recovery rate to be the boundary value, but rather use an exogenously given
recovery rate w. They can then price different seniorities by assigning different recovery
rates to the different seniorities (Cossin and Pirotte 2001). Coupon bonds are modeled as
a portfolio of zero coupon bonds. As in the illustration above by Giesecke they assume
the default boundary to be constant and exogenously given. In their implementation they
set the barrier equal to the par value of the bonds. The Longstaff and Schwartz model has
great advantages over the Merton model as it relaxes the assumptions made. It has great
flexibility and allows correlation between the Brownian motion of the firm value and the
risk-free interest rate in Vasicek. However, there is no explicit solution available, and
Longstaff and Schwartz only provide an approximation to the solution. Another problem
with the model is that it allows a payout upon default that might be greater than the firm
value at default. Ammann (1999) gives a good summary of possible extensions to fist
time passage models. These extensions include models with different interest rate
modeling like Cox-Ingersoll-Ross (1985), different default boundaries and models with
jump-processes like the Zhou (1997a) model we implement in this paper.
3.1.3 Practical Implementations of Structural Models
Below we present two models implementations that are being used by leading companies
providing credit risk analysis and pricing of credit derivatives; Moodys KMV model and
JP Morgans E2C model.

18
Moodys KMV
The most successful commercial variant of structural models is the KMV model that
loosely uses the original Merton framework (Soudaram 2001). Like the Merton model
KMV uses Black and Scholes to compute the asset volatility from the equity volatility.
For the default barrier, K, KMVs approach is based on the empirical observation that
default tends to occur when the market value of the firms assets falls below a point that
typically lies below the face value of all debt and above the book value of short-term
liabilities. The default barrier used is the sum of the short-term liabilities plus 50% of the
long-term liabilities. Given the firm value, the asset volatility and the default barrier,
KMV calculates how many standard deviation moves that would result in the firm value
falling below the default barrier. The result is called distance to default, d. The distance
to default is used to look up in a database to identify the proportion of firms with the
same d that actually defaulted within a year. The result is the expected default frequency,
which is the main output of the KMV model.

The main reason for using the distance to default to look up in a database, instead of just
calculating the default probability directly assuming V is normally distributed, is because
the distribution of V in reality has excess kurtosis.
9
Since a default usually occurs when
the firm value drops significantly kurtosis becomes very important and the normal
distribution is thus a poor approximation. (Soudaram 2001)
JP Morgans E2C Model
Another successful practical implementation of the Merton model is the Credit Grades
E2C
10
model, developed by JP Morgan. The model is currently being used to manage
high yield credit derivative inventory and monitor the investment derivative book. As in
the Merton model the asset volatility is derived from the volatility of equity. Further the
asset value is assumed to have zero drift, i.e. 0 , as one assumes that debt would be
issued to keep the leverage level steady over time. The default barrier is assumed to
follow a lognormal distribution with percentage standard deviation . The values of K
and are determined by an empirical study and set to 0.5 and 0.3 respectively, for the
general case. The probability of no default can then be calculated as
ln( ) ( ) ln( ) ( )
(0, )
( ) 2 ( ) 2
d A T d A T
B T N d N
A T A T
_ _

, ,

where,
2 2 2
( )
V
A T T + ,
2 (0) V
d e
K

and

9
The distribution of the asset values is fat-tailed compared to the normal distribution.
10
Equity-to-Credit

19
(0) (0) V S K +
The credit spread is calculated by integrating the default probability density and applying
a constant asset specific recovery
V
(Lardy 2002):
0
0
( , ) 1 (0, )
( ) (1 )
( , )
T
rt
V T
rt
e d B t T B T
CS T
e B t T dt

The power of this model lies in the simple expressions above combined with robust
approximations leaving us with only observable parameters. Compared to historical
results the model performs well and it is widely used in practice. The E2C-model
resembles the economical intuition behind the model proposed by Giesecke (2001),
which will be discussed in chapter 3.3.1.
3.2 Reduced form models
Das (2001) classifies the reduced form models into three broad categories. Default
models use stochastic processes to model default directly. Spread models decompose risk
into credit and recovery risk and solely model the spread without reference to its
components. Credit rating models depict the evolution of a firm as changes in its credit
rating.

We start this chapter by giving an introduction to the basic theory behind default
intensities and recovery rates. Further we move on to a general discussion on credit rating
modeling. In particular we will discuss the Markov chain models proposed by Jarrow and
Turnbull (1995) and Jarrow, Lando and Turnbull (1997) and the framework developed by
Duffie and Singleton (1999).

Of other contributors, we mention Schnbucher (1998) who presents a tree model for
defaultable bond prices. This model serves as an implementation framework for a wide
range of intensity based models and accommodates for comparison of these models
properties as well as several different recovery setups. There exists a wide range of
published papers on reduced form models and we show to Bohn (2000) for a more
thorough study of the literature available.
3.2.1 Intensity Based Modeling
Schnbucher (2003) and Giesecke (2002) give an overview of different properties and
limitations of the exogenous intensity process.

The most common approach is to model the intensity as a Poisson process, which is a
discrete statistical process used in a great variety of ways. Most commonly Poisson

20
processes are used in connection with a counting process ( ) { }
, 0 N t t having an
intensity rate 0 > . The number of events in any interval of length t is then Poisson
distributed with mean t . (Ross 2000)
( ) ( ) { }
( )
n=0,1,... s,t 0
!
n
t t
P N s t N s n
n
e

+
As for the simulation of default processes, we are normally concerned with the first jump,
interpreting it as default. Hence, the time of default, t , is exponentially distributed with
parameter (Walpole et al. 1998) and the intensity rate is the conditional default arrival
rate given no default (Giesecke 2002). However, this setup can easily be extended when
working on more advanced products like for example a portfolio of securities where we
might be interested in the structure of several defaults.

Constant intensity implies flat term structure of credit spreads. In order to develop more
sophisticated models allowing for time-varying intensity, we will need an
inhomogeneous Poisson process (as opposed to the homogeneous process described
above) with time-dependent intensity ( ) t . If the counting process N(t) has independent
increments for s<t we have (Ross 2000):
( ) ( ) { }
( )
1
, 0
!
( )
s t
s
y dy
n
s t
P N s t N s n n
n
s
y dy e

+
+
_

,

Giesecke (2002) also extends this intensity to be a function of the current information, ,
available in the market

A further generalization is applicable through the Cox process. The probability of a jump
in the interval [s, s + t] is proportional to the realized value of ( ) t at time t:
( ) ( ) { }
1 ( )
t
P N s t N s X t +
The intensity is now allowed to be random under the restriction that conditional on the
actual realization of , N is an inhomogeneous Poisson process. Accordingly, the
random intensity process can be a function of some state variables, like for example spot
interest rates and credit ratings. The Cox process is not measurable in any way, but it has
a background process, which determines the long-term average of the Cox process.
(Lando 1998)

This far we have only considered modeling of single firms. Often one may be interested
in modeling default correlations due to firms dependence on each other and economic
factors. A natural way of incorporating this feature is to allow the firms intensity
processes to be correlated through time. Alternatively it is possible to introduce jumps in
the intensities, either as a consequence of default of other firms, or as a result of joint

21
credit events. We will not cover default correlations in this paper, but we refer to
Giesecke (2002) and Schnbucher (2003) for a more in-depth treatment.
3.2.2 Recovery Modeling
In most structural form models recovery is given endogenously. However, in reduced
form models we need to make an assumption on the behavior of the recovery process.
Modeling of recovery is in many cases just as important as finding the time of default and
can have a significant effect on the pricing of credit derivatives (Schnbucher 2003). It is
obviously difficult to model the real recoveries that will happen in a bankruptcy process,
but there exist approaches that give an approximation of the real world.

In its simplest form it is assumed that recovery will be zero. Even though zero recovery
(ZR) is not realistic in most cases, it is an easy approach and often serves as a benchmark
for other recovery models. Houwelingen (2003) gives a summary of three recovery
models found in the literature; recovery of treasury, recovery of market value and
recovery of face value.

With recovery of treasury (RT) a bond at default will be replaced by a fraction of a
similar, but default-free bond. Obviously RT requires that the price of default-free bond
can be determined, but for coupon bonds it is normally not a problem. The RT model is
computational easy and is a convenient way of improving from zero recovery. The main
drawback with the model is that it gives unrealistic spread curves for low credit qualities
(Schnbucher 2003).

With recovery of market value (RMV) the bond is assumed to loose a fractional amount,
q, of its market value. In the case of default the value of our claim has been reduced to
(1 ) q of its original market value without any other default effects.

The last model considered is recovery of face value (RFV), or recovery of par (RP) as
denoted by Schnbucher. At default the claimant recovers a fraction of the face value of
the bond. This model closely correlates to what is happening in the case of bankruptcy.
According to the priority rule, loans with higher seniority will be paid off before those
with lower. However, since absolutely priority seldom holds in reality, seniority
claimants will receive a payoff of a certain fraction of par (not par), while the junior
claims will be paid off with a smaller fraction afterwards.

RMV and RFV are considered the best models for recovery. When it comes to pricing,
RMV is the easiest to handle, though the extra work still might have to be carried out
when pricing certain credit derivatives. Schnbucher (2003) gives an extensive
discussion on the pros and cons of the recovery models described and how they are
mathematically priced.

22
3.2.3 Credit Rating Modeling
There exist several agencies that evaluate and give credit ratings to defaultable
companies. The most familiar ones include Standard and Poors, Moodys KMV and
Fitch. Companies are classified according to their credit quality and these ratings are
updated as new market information is revealed. However, there is normally a time lag,
sometimes months, from the credit information reaches the market until the rating
agencies update the ratings. Even though the delay in updating a companys rating makes
them less suited for practical purposes, they are frequently used as a first classification of
the obligor.

Most credit rating models are based on discrete or continuous-time Markov chains with
transition matrices equal to historical average rating transition frequencies as published
by the rating agencies. Below is an example of a transition matrix based on average
ratings between 1985 and 2001.

Europe Rating to:
Aaa Aa A Baa Ba B Caa-C Default WR
Aaa 86,34% 8,21% 0,19% 0,00% 0,00% 0,00% 0,00% 0,00% 5,26%
Aa 0,76% 86,71% 9,13% 0,10% 0,00% 0,00% 0,00% 0,00% 3,30%
A 0,00% 5,05% 84,80% 3,63% 0,10% 0,02% 0,00% 0,02% 6,39%
Baa 0,74% 0,25% 4,82% 78,83% 2,86% 1,16% 0,04% 0,00% 11,31%
Ba 0,00% 0,00% 0,64% 10,52% 71,40% 9,29% 0,68% 0,25% 7,22%
B 0,00% 0,00% 0,33% 1,03% 9,40% 65,52% 8,28% 3,29% 12,17%
Rating
from:
Caa-C 0,00% 0,00% 0,00% 0,00% 0,00% 22,41% 48,58% 14,53% 14,47%
Table 3.2 Average one-year rating migration rates 1985-2001 (Moodys 2002)
In general the transition probability matrix Q(t,T) for the time interval [t,T] is given as
11 12 1
21 22 2
1 2
( , ) ( , ) ... ( , )
( , ) ( , ) ... ( , )
( , )
( , ) ( , ) ... ( , )
K
K
K K KK
q t T q t T q t T
q t T q t T q t T
Q t T
q t T q t T q t T
1
1
1
1
1
]
M M O M

The element q
ij
(t,T) is the probability that the rating process is in state j at time T given
that it was in state i at time t. Formally we can write
[ ] { } ( , ) ( ) | ( ) , 1,..., ,
ij
q t T P R T j R t i i j K t T
where R(t) is the credit rating at time t.

Schnbucher (2003) assumes that the credit spreads are constant throughout time and that
they are the same for all bonds in the same credit class. Specifically, a bond can only
change its spread through a rating transition. The setup rendered below further assumes
zero recovery, but Schnbucher also shows how this assumption can be relaxed to cover
RT or RP. Given zero recovery and an initial rating of R(0) by time 0, the price of the
defaultable bond will be

23
(0)
(0)
(0, ) (0, )(1 (0, )) R
R K
B T B T q T
Hence, the price of the risky bond is given by the default-free bond times the survival
probability.

Despite their apparent simplicity credit rating models have several major drawbacks
which make them a limited reflection of the real world. Firstly, the transition matrix is
based on historical (backward looking) average, though it has been verified that the
implied historical default rate overstates the true default rate. Furthermore empirical
studies show that especially downgrading of companies is followed by an increase in the
credit spread (Cossin and Pirotte 2001). As a result the credit agencies might be reluctant
to change the credit rating of a company, even though it could be from a low status in
class A to a very strong status in BBB. If there existed a continuous scale instead, the
consequences of minor updates in the credit status would not imply a jump in the credit
spread.

Finally, Schnbucher (2003) points out that calibration is the hardest problem related to
these models. There is a big difference between the historical transition matrices
published by the agencies and the credit spreads actually observed in the market. Partly
the problem is a consequence of the fact that all bonds of a credit rating class are assumed
to have the same credit term structure and partly because the ratings take little account of
recoveries, which might differ significantly between firms.

Reduced form models are built on the assumption that default hits the market by surprise.
The probability of this event is a hazard rate and is frequently called the intensity rate
( ) , t , which is a function of the time t and the state of the world . That means that
the default is not an effect of the firms assets or liabilities, but rather the result of an
exogenous process governed by selected state variables. The economic interpretation of
this kind of models is not as good as for the structural default models when it comes to
explaining why a default occurs. Nevertheless the tractability is good and the models
provide better fit to historical data. The reason lies in the fact that reduced form models
are calibrated to data from rating agencies and financial time series. (Cossin and Pirotte
2001)
3.2.4 Jarrow and Turnbull (1995)
Jarrow and Turnbull (1995) assume the capital structure of a firm to be irrelevant for the
default process, which in turn is based on an exogenous intensity model. They identify
two classes of underlying zero-coupon bonds needed in their model. Firstly, there exist
default-free zero-coupon bonds for all maturities, where ( , ) B t T is the time t value of the
bond paying a dollar at maturity, T. Secondly, we have defaultable zero-coupon bonds of
all maturities for a company XYZ where ( ) , B t T is the time t value of the bond
promising a dollar at maturity T.


24
They further give a foreign currency analogy in pricing the risky debt of XYZ. The
payments from XYZ, ( ) ,
foreign
B t T
11
, are denoted in a hypothetic XYZ-currency which is
linked to the dollar by a hypothetic exchange rate, e(t). Putting these expressions together
we have an intuitive decomposition of the XYZ zero-coupon bond:
( ) ( ) , , ( )
f
B t T B t T e t (0.1)
As long as company XYZ is not in default, e(t) will be 1 as each promised XYZ-dollar is
worth exactly 1 dollar. In the case of default, e(t) reduces to the recovery rate. As a result
of this analogy to foreign currency, techniques from foreign currency option pricing can
be applied in order to price credit derivatives involving credit risk.

Jarrow and Turnbull first present their model in two-periodic discrete time with
{ } 0,1, 2 t . At time t
i
the probability for default under the equivalent martingale measure
in the period [t
i
,t
i+1
] is
i
P
%
leaving us with the possibilities that default has occurred
and that it has not, respectively state b and n. In addition they give the default-free spot
interest rate a stochastic evolution with two possible states in the next time period: u and
d.

1
B
f
(0,2)

1
B
f
d,n(1,2)

d
d B
f
d,b(1,2)

1
B
f
u,n(1,2)

d
dB
f
u,b(1,2)

1
1 p
1
p
1
p

d
d
d
d
1
1
0 2 1
1
1 p
t

0
0
(1 )(1 ) p

0
0
(1 ) p

0
0
(1 ) p

0
0
p

Figure 3.1 Jarrow and Turnbull (1995)
The risk-neutral probability of an up-move is denoted by
i
and accordingly for a down-
move 1-

i
. For simplicity they assume that the spot interest rate process and the
bankruptcy process are independent under the equivalent measure. Finally they construct

11
This amount is default-free in the hypothetical currency. Further in the text we will shorten the
denotation to B
f
(t,T)

25
an exogenously given recovery process, which pays of a fraction, d, of notional. This
corresponds to the recovery of treasury model. To show the result of their model Jarrow
and Turnbull construct a two-periodic binomial tree and extends it to cover more periods
and finally the continuous case which we render here to show the intuition behind the
model.

As mentioned we model two processes which each have two possible outcomes. The
result is four different states at time 1, depending on the up-down move of the interest
rate and whether bankruptcy has been declared or not. The upper part of the boxes in the
figure gives the exchange rate and the lower part expresses the value ( ) , B t T as given by
equation (0.1). Note that when default has occurred there is no uncertainty about the
future as the bond in any case only pays out the recovery rate at maturity and the short-
term rate is given for the next period. Using risk-adjusted pricing techniques we obtain
the following bond prices at time 1:
( ) ( ) ( )
, ,
1, 2 1, 2 1, 2
f
u b u b u
B B B
( ) ( )

, , 1 1
1, 2 1, 2 (1 ) (1, 2)
f
u n u n u
B B p p B
1
+
]

( ) ( ) ( )
, ,
1, 2 1, 2 1, 2
f
d b d b d
B B B
( ) ( )

, , 1 1
1, 2 1, 2 (1 ) (1, 2)
f
d n d n d
B B p p B
1
+
]

where B
u
(1,2) and B
d
(1,2) are the spot prices of bonds after an up or down move in the
interest rate. Applying the same technique we obtain the defaultable bond price at time 0:
( )

( )

( )

( )

( )
0 0
, , 0 0
0 0
, , 0 0
0, 2 1, 2 (1 ) 1, 2
(1 ) 1, 2 (1 )(1 ) 1, 2
u b u n
d b d n
B p B p B
p B p B

+ +
+

Further in their article Jarrow and Turnbull relax some of the assumptions outlined above.
The suggested relaxations include introduction of correlation between the bankruptcy-
and default-free-free interest rate processes, inclusion of traded equity on company XYZ
and finally estimating the continuous-time limit of their model.
3.2.5 Jarrow, Lando and Turnbull (1997)
The model of Jarrow et al. (1997) is a Markov model for the term structure of credit
spreads and is an extension of Jarrow and Turnbulls paper from 1995. The life of a
company is seen as a journey through possible credit rating states with one absorbing
state, identical to bankruptcy. These transition parameters can be observed from credit
rating agencies and are assumed to give a good measure of the creditworthiness of the
company. Even though ratings are known to be rather slow in reacting to new market
information they are accepted to be a major variable, on which credit spreads can be
based. (Cossin and Pirotte 2001)

26

The K x K discrete, finite-state Markov chain ( ) : 0
t
t transition matrix is specified
as
11 12 1
21 22 2
1,1 1,2 1,
...
...
= ...
...
0 0 ... 1
K
K
K K K K
q q q
q q q
q q q

1
1
1
1
1
1
1
]
Q=
.
where 0
ij
q for all i, j, i j . The (i,j)th entry is the true probability of going from state i
to state j in one time step. The highest credit rating is defined to correspond to state 1 and
the lowest credit rating is in state K-1. State K is the absorbing state that corresponds to
default. Now historical transition matrices can be used as Q.

Under the assumption of no arbitrage opportunities we have to calculate a similar
transition matrix under the equivalent probability measure such that the credit rating
process is given by
%
( ) , 1 ( )
i ij ij
q t t t q i j + in which ( )
i
t is a deterministic
function of time. The risk-adjusted probability of default after time T is
( )
%
( ) Q * , 1 ( , )
i
t iK ij
j K
T q t T q t T
>

where { } * inf :
s
s t K represents the first time of bankruptcy. Furthermore
( )

, ( , )( (1 ) ( * ))
i
i
t
B t T B t T Q T + >
is the value of a zero-coupon bond issued by a firm currently rated in credit class i. The
term structure of the default-free spot rate can be modeled with any appropriate model,
like for example Black, Derman and Toy (1990), Cox, Ingersoll and Ross (1985) or
Heath, Jarrow and Morton (1992), as it is assumed to be independent of the credit rating
process. Jarrow, Lando and Turnbull also extends the credit ratings and default
probabilities to apply in the continuous case.

Recovery is modeled as a fraction of face value and is considered to be an exogenously
given constant, . Jarrow et al. suggest that this constant can be regulated for different
seniority debts in order to incorporate recovery differences into their model. As in Jarrow
and Turnbull (1995) they assume the recovery rate to be independent from the stochastic
process driving the term structures, however, this is relaxed by Das and Tufano (1995).
Das (2001) gives a step-by-step example on how the model can be used and calibrated on
the basis of real data.

The models simplicity also gives rise to some of its major drawbacks. The assumption
that recovery rates are constant is in most cases not applicable in the real world. The

27
same also applies to the modeling of intensities, which are set to be constant over time
and also equal for all bonds in the same rating class. (Schmidt 2001)

Lando (1998) makes a further generalization by allowing default intensities and
transitions to be governed by a Cox-process and the hazard rate to be dependent on the
risk-free short interest rate. The result is a more flexible model, which can depend on
different state variables of interest, but infers a computational cost.
3.2.6 Duffie and Singleton (1999)
Duffie and Singleton follow another approach than the two models already mentioned.
Their model has its foundation in Madan and Unal (1994).

Under the risk-neutral probability measure they denote the hazard rate, h
t
, as an
exogenous process governing the default process and L
t
the expected fractional loss
conditional on information available up to time t. Over a small, discrete interval t , a
default will then occur with the risk-neutral probability
t
h t . The recovery to creditors
stochastic and is assumed to follow a recovery of market value (RMV) model, though
Duffie and Singleton also give a discussion on other recovery models. RMV allows for
correlation between the hazard rate, loss process and the term structure of the default-free
interest rate.

Their objective is to be able to describe the price of a defaultable bond given with the
following equation
( )

0
0
T
u
R du
P
B T Xe
1
1
]
E
where X is the face value of the bond. In the following we will use X=1 for simplicity
and R is a default-adjusted short-rate process, replacing the usual short-term interest rate.
Cossin and Pirotte (2001) gives an intuitive explanation to the relation between these
interest rate processes and shows that it must satisfy
[ ]
1 1
(1 ) (1 )
1 1
h t L h t
R r
+
+ +

If we take the limit and let 0 t , we get R R hL + . Under certain circumstances in
the continuous case this representation of R is precise. (Cossin and Pirotte 2001) Duffie
and Singleton also propose another factor in this equation, denoted as liquidity effects, l ,
in order to account for carrying cost. The expression for the default-adjusted interest rate
then becomes R r hL + + l .

The processes h
t
and L
t
are often difficult to identify separately using defaultable bond
yields. However, Duffie and Singleton state that for some derivatives this does not imply

28
any problems. As an example they explain how bond option prices can be used to
separate these parameters and how this can be used to price a credit-spread put option.
3.3 Hybrid Models
Recently there have been proposed models incorporating features from both the structural
and the reduced form methodology. In chapter 4 we will consider the model proposed by
Zhou (1997a) that can be regarded as a hybrid model. First, however, we will give a brief
overview of a hybrid model proposed by Giesecke (2001) with a slightly different
foundation.
3.3.1 Giesecke (2001)
Gieseckes model (2001), like Zhou (1997a), has its foundation in the classical structural
approach. The structural model face especially two difficulties that are not consistent
with the real market. Firstly, short-term credit spreads above zero is not possible. As one
knows there is no default today, one is not interested in paying a default premium for
bonds with short maturities. Secondly, in these models bond prices converge
continuously to their recovery value, though in the market price jumps are frequently
observed in the case of default. Giesecke (2001) presents an extended model, which faces
these problems through the assumption of incomplete information.

Duffie and Lando (2001) established a model where firms assets cannot be observed
perfectly by investors as they only get incomplete and noisy reports on the firms true
assets. This way unexpected default can happen, as in reduced form models. However,
they still assume that investors have complete information about the default barrier, so
that a company will be forced into bankruptcy when the barrier is triggered:

Figure 3.2 Short-term credit spreads in the case of incomplete information (Duffie and Lando 2001)

29
Giesecke (2001) generalizes the model further as he allows for uncertainty about the
barrier level as well as the firms assets. Surprisingly, though, a different kind of
incomplete information results in a different term-structure. The reason is that we observe
the historical low of the assets as time passes by and accordingly get an upper bound on
the barrier. The table below summarizes the different kind of and combinations of
uncertainties about firm asset and the barrier, and shows the resulting term-structure and
short-term credit spreads.

Information
Incomplete
Barrier

Complete
Assets
At Historical
Low
Above
Historical
Low
Both
Term
Structure
Hump-
shaped
Hump-
shaped
Decreasing Hump-
shaped
Decreasing
Short
Spreads
Zero Non-zero Non-zero Zero Non-zero
Table 3.3 Information with Giesecke
Giesecke (2001) illustrates the Black and Cox model and provides a simpler solution by
considering the default barrier constant ( ) K t K as in the Merton model.

Giesecke defines a running minimum log-asset process ( ) for 0 M t t
2
1
( ) min(( ) )
2
v v s
s t
M t s W
+
where M is the historical low of the log-asset value. Then the default probability becomes
[ ] ( ) ln
(0)
K
P T P M t
V
1 _
<
1
, ]
.
To derive this probability he uses the fact that ( ) M t is inverse Gaussian and obtains
[ ]
2
2
2 2 2( )ln
(0) ( ) ln ln ( )
(0) (0)
1
v
v
K
V
v v
v v
K K
T T
V V
P T N e N
T T

,
_ _ _ _
+

, ,

+

, ,

Giesecke develops a reduced form framework that can be applied to all credit models
predicated on unexpected default. This does not only apply to the intensity based models,
but also Gieseckes own model. Through this framework he is able to derive closed-form
solutions under specific conditions and also to value credit spreads by making use of
compensators. Compensators may be explained by the Doob-Meyer decomposition

30
theorem. (Neftci 2000) It states that a continuous-time process can be decomposed into a
martingale and a drifting process, which in our case is the compensator. Hence, the
compensator simply equals the expected intensity of the indicator process.

The model is economically intuitive as it reflects the true world in which investors
seldom have complete information. It also makes sense when we take into account the
recent surprising bankruptcies at Enron and WorldCom where arrival of new information
gave unexpected news about firms assets and liabilities. The model requires the
following input: daily equity prices, equity volatility forecasts, reported liabilities and
interest rates. These data may be obtained in the market, thus making the model easy to
calibrate. Other variables, like firm volatility and initial firm value derives from the initial
parameters with option pricing formulas.

31
4 Zhous Model
There are some major drawbacks with the structural and reduced form approaches to
modeling credit risk. This chapter will be devoted to present a model proposed by Zhou
(1997a; 2001), which tries to incorporate the advantages of both approaches described
above. The model can just as well be categorized as a second-generation structural model
as a hybrid model, due to its foundation built on the economically sound structural
framework. At the same time it allows for better fit to real data by allowing jumps in the
firm value to occur.

We will first present the foundation of the model and then move on to a closed-form
solution valid under a certain assumption. Further, we give a Monte Carlo solution for the
general case as proposed by Zhou. Next, we present and correspond the results from the
two solution approaches and use the implementation to price a credit default swap.
Finally, we suggest how the model can be calibrated by the use of Markov chain Monte
Carlo and possible further extensions in order to make it fit real market data.
4.1 The Model
The model is described by a jump-diffusion process predicting the market value, V, of a
firms total assets according to the following process
1
( ) ( 1)
t
t
dV
dt dZ dY
V
+ +
where the variables have the interpretation stated in the table below. It is assumed that the
firm is funded solely by equity and zero-coupon bonds.

t
V Total market value of firm, i.e. assets + debt
Z
1
A standard Brownian motion
dY A Poisson process (allowing one jump)
Intensity parameter for the Poisson process dY
Volatility of the firms assets

Jump amplitude with ( ) E = 1 + and
2
ln( ) ~ ( ) N

+

2
2
: [ 1] exp( ) 1 E

+

Expected drift in the value of the firm
dt An infinitesimal time step
Table 4.1 Description of parameters used in Zhou's model

32
dZ
1
, dY and are mutually independent.

This process can be interpreted as follows; The firm value is subject to a standard
Brownian motion for all infinitesimal time steps dt. Jumps occur with an intensity and
expected amplitude . Given that is the expected drift in the diffusion process, we need
to compensate the asset process by subtracting to obtain the desired expected drift.

The firm defaults on all its obligations immediately if its value falls on or below a time-
dependent positive threshold value K
t
. Zhou assumes that K is a deterministic function
that follows
0
t
K e K
. By doing so, can be interpreted as the growth rate of the

firms liabilities (Zhou 2001). However, we will in the following set 0 implying that
K is a constant equal to K
t
.

Now, define X
t
as the ratio of the firms value to the barrier:
t
t
V
t K
X implying default
when 1
t
X . Given a default at time t , the bondholder receives a fraction 1 ( ) w X
of
face value at maturity time, T, where w(X
t
) is a function determining the write-down of
the bond. Normally w(X
t
) is a non-increasing function of X, so that the lower the firm
value at default, the less the bondholder receives. Zhou uses the function w(X)=w
0
-w
1
X
with w
0
and w
1
as non-negative constants to determine the write-down. In order to make
the model more realistic, one may use different write-down functions for different grades
of seniority. In reality, the write-down functions will differ for even the same seniority
class of bonds and for different time periods (Zhou 2001).

According to Zhou (1997a) the capital asset pricing model (CAPM) is still assumed to
hold for equilibrium security returns while the jump component is purely firm specific
and uncorrelated with the market. Consequently the for the jumps must equal 0, again
implying that there is no risk premium for the jump risk.
12

Giesecke and Goldberg (2003) claim that Zhous jump-diffusion model adds too much
complexity into the problem in order to incorporate the capability of positive short-term
spreads. One of the major drawbacks with Zhous model is its complexity when it comes
to calibrating the model to market data. We will come back to this issue in section 4.6, in
which we describe in detail how it may be done. Another problem discussed by Giesecke
(2001) is that jumps do not necessarily imply default as is the case with reduced form
models. Actually the jump may not be large enough to force the firm value down to or
beyond the barrier making it hard to correlate default rates actually observed in the
market with the jump intensity in Zhous model.

On the other side the model has several nice features. First of all it has the flexibility
needed in order to accommodate for the variety of term structures observed in the market,
including above zero short-term credit spreads. Also the model is consistent with the fact

12
According to CAPM the risk premium = (r
j
-r
f
)

33
that there is often observed a jump in the bond value at the moment of default. Finally,
the model differs from the structural models as it correlates the recovery rate to the firm
value at default, whereby generating a random recovery rate endogenously.
4.2 Closed-Form Solution
Zhou (1997a) provides a closed-form solution for valuing debt under certain conditions.
The main difference from the general model is that the debt cannot default before
maturity. As all debt has to be paid at maturity, the default barrier has to be equal to the
face value of debt.

In the following we derive a closed-form solution for valuing equity under the same
assumptions. Analogous to Zhou when calculating the risk-neutral price of a bond we
assume that investors do not receive a risk premium for the jump risk. Under these
assumptions the put-call parity holds if there are no bankruptcy costs, i.e. the value of the
firm is equal to the risk-neutral value of the equity and debt. When there are bankruptcy
costs there will still not exist arbitrage opportunities since the firms assets are not traded.

We want to calculate the expected value of the equity under the risk-neutral probability
measure

P . We know that the firm will default if the firms value is below the face value
of debt at maturity. As we define / X V K we know that the firm will default if
1
T
X and from a risk-neutrality argument we have for K = 1 = face value of debt:
( )

( )
( )
, max( , 0)
rT P
T
S X T e E KX K

Under
P the original SDE

( ) ( ) ln 1
dX
r dt dZ dY
X
+ +
can be rewritten as:
2
ln( ) ( / 2 ) ln( ) d X r dt dZ dY + +

We further define
( )

| : ( | ) (ln( ) ln( ) | )
P
T T T
G X P X X P X X
which is the probability that
T
X given the current X. In the following 1 due to
the discussion above.

To find the probability
( ) 1|
P
T
G X we first find the distribution of ln( ) X for a given
number of jumps. Given a certain number of jumps, i, before maturity, ln( )
T
X will be
normally distributed with the mean,
nd
, including the drift of the jump component

34
multiplied by the number of jumps. Similarly the variance,
2
nd
, will be the total variance
combining the diffusion and jump component:
2
2 2 2
ln( )
2
nd
nd
X r T i
T i

_
+ +

,
+

We let
T
Y be the number of jumps from time 0 to T to get
2
2 2
ln( | , ) ln( ) ( ) ,
2
T T
X X Y i N X r T i T i

_
+ + +

,
:
The probability of a certain number of jumps is given by the Poisson distribution
( )
( )
!
T i
T
e T
P Y i
i

We can then multiply the probabilities of a certain number of jumps with the normal
distribution given the number of jumps and sum up.

( )

( )
( )

( )
0
2 2
0
1| ln( ) ln(1) | ln( )
ln( ) ln(1| ,
ln(1) ln( ) ( )
( )
2
!
P
T T
T T T
i
T i
T
i
G X P X X
P Y i P X X Y i
X r T i
e T
N
i
T i

_

+

,
(0.2)

Merton (1974) consider equity as an option on the firms asset with strike equal to the
total debt. Similarly, the

P
T
G we have calculated represents the probability that the option
(equity) is in the money at expiration. It is important to notice, however, that this
probability also includes the cases were the firms value falls below the barrier in the
interval 0 to T and rises again to be above the barrier at time T.

To calculate the value of equity at time 0 we need to find the expected value of the equity
at time T. We find this by calculating the equity value given no default at time T and
multiplying it with the probability of no default.
( )

| 1 (1| )
P P
T T T
E X X G X

35
Above we found that ln( ) X is normally distributed with
nd
and
nd
. For each of the
possible number of jumps i we define

( )

( )
( )
2
2
2
2
2
2
ln( ) 0
2 2
ln( ) 0
2 2
2
2
2
2
| 1, 1|
2
2
2
1
nd nd
T
nd nd
nd
T
nd
nd
nd nd
nd
nd
nd
nd
P P
i T T T T
z
z
X
z
X
z
z
nd nd
nd
nd nd
nd
r
h E X X Y i G X
dz
e e
dz
e e
dz
e e
e N
e N
Xe

+
+
1 _

1

1
, ]
_
+

,
2 2
2
2 2
2
2 2
ln( )
2
T i
T i
X r T i T i
N
T i

_
+
+ +

,
_ _
+ + + +

,

+

,

We will then sum for the possible number of jumps to get the expected value up to n
jumps.
( )

2 2 2
2
2 2
2 2
2 2
0
| 1 (1| )
ln( )
2 ( )
!
V v
P P
T T T
V
T i
v
T i r T i
i
v
E X X G X
X r T i T i
e T
Xe N
i
T i

_
+
+ +

,

_ _
+ + + +

,

+

,
(0.3)


36
We can now combine (0.2) and (0.3) to calculate the risk-neutral value of the equity:
( )

( )
( )
( ) ( ) ( )

( )
( )
( )

( )

( )

( )
( )
( )

( )

( )
( )
2 2 2
2
2 2
2 2
2
, max( ,0)
| 1 1| 1 1|
| 1 1| 1| 1 1|
| 1 1| 1|
ln( )
2 ( )
!
rT P
T
rT P P
T T T T
rT P P P P
T T T T T
rT P P P
T T T T
T i
T i r T i
rT
S X T e E KX K
e E X X X G X
e E X X G X E X G X
e E X X G X G X
X r T i T i
e T
Xe N
e i
T
_ +
+ +

,

> >
> >
>
_
+ + + +

,
+
2 2 2
2
0
2 2
0
2
2 2
ln(1) ln( ) ( )
( )
2
!
ln( )
2 ( )
!
i
T i
T
i
T i
T i r T i
rT
i
X r T i
e T
N
i
T i
X r T i
e T
Xe N
i
e
_ +
+ +

,
_ _

, ,
_ _ _

+

, , ,
_
+ +

,
2 2
2 2
0
2 2
ln( ) ln( ) ( )
2
i
T
T i
T i
X r T i
N
T i
_ _ _
+ +

+

, ,

_

+

, ,

We see that our result is somewhat similar to the result in the E2C model described
previously. In the E2C model the firms assets follow a diffusion process and the default
barrier jumps at maturity. This is pretty much the same as in our result where a jump will
trigger a default in the same way. However, in our result there can be several jumps,
which will be significant if is large (Craine et al. 2000). Similarly to the E2C model we
could also have expanded our result to calculate an equity hedge for credit.

With our result the equity can be viewed as a European call option on the firms assets as
in Mertons model. In the general framework, in which the firm can default before the
maturity of the debt, the analogous option is a path dependent option or more specifically
a down-and-out call option. It is obvious that with the same strikes a European call will
be more valuable than a down-and out call. Similarly the debt will be less valuable in the
closed-form solution provided by Zhou.

It is evident that both equity and debt as options on the firms value is in fact path
dependent and that fist passage time modeling, as done by Black and Cox (1976) for a

37
pure diffusion process, provides a more realistic representation of reality than the Merton
model. However, when the underlying path includes jumps, the options become so
complex that they most often do not have a closed-form solution. In our simulation
described in the next section we will include the calculation of equity prices under risk-
neutrality as suggested by Zhou (1997b).
4.3 Implementation
We implement Zhous model by the use of Monte Carlo methods. Below we will give a
primer on Monte Carlo methods before moving on to a discussion on the solution of the
model and the algorithm we use in the implementation. Finally, we discuss some of the
practical concerns related to the implementation, including the choice of programming
language and parameter values.
4.3.1 Primer on Monte Carlo Methods
Monte Carlo (MC) simulation is a powerful and straightforward technique frequently
used to solve financial problems. The most common applications include high
dimensions integration and path-dependent problems.

Let X be a discrete random vector with a set of possible values x
j
and a probability
density function given by
{ }
, 1
j
P j X x . Now suppose that we are interested in
calculating
( ) { }
1
( )
j j
j
E h h P
1
]

X x X x
for some specified function h. Frequently h(x) is hard to calculate exactly and we have to
turn to approximation. Simulation is then a powerful approach. (Ross 2000)

By the central limit theorem and the law of large numbers (Glasserman 2003) one is
ensured that an estimate based on several simulations is approaching its correct value as
the number of simulations is increasing to infinity. Let X
1
, X
2
, , X
n
be independent and
identical distributed random vectors from the given mass function. Our Monte Carlo
estimate is then given by
$
1
1
lim ( )
n
i
n
i
h
n

X
and the standard deviation of the estimate is inversely proportional to the square root of
the number of simulations:
1
n
n


38
In order to reduce the error by half, the number of simulations has to be increased four
times. As a result, Monte Carlo is a computational intensive method, though it enjoys
great flexibility making it applicable to a wide range of problems.

One of the major strengths of Monte Carlo methods is related to path-dependent
problems. Through the simulation of stochastic processes, one does not only get the final
result but also the development path for the process. For instance is MC a common way
of valuing Asian options as they are dependent on the path a stock takes as well as its
final value. Also when it comes to modeling credit risk MC forms an intuitive method in
simulating the value of a firms assets, as will be described in the next chapter.
4.3.2 Solution
Though there under certain conditions exist closed-form solutions for credit risk models,
one often needs numerical methods in order to find approximate solutions. Monte Carlo
methods serve as a powerful option despite its computational disadvantage. In most cases
it is an easy way of implementing and experimenting with models, and the addition of
new features to the model is normally straightforward. Zhou (2001) describes a Monte
Carlo approach to valuing a defaultable bond and we here render the most important steps
of the algorithm:

1) Divide the time interval [0,T] into n equal sub periods.
2) Let
0
*
t
X X .
For j=1 to m
- Generate the independent random variables x
i
p,
i
, y
i
according to the
following distributions:
2
2
,
2
i
T T
x N r
n n

1 _

1
, ]
:
( )
2
,
i
N

:
T
0 with prob. 1-
n
T
1 with prob.
n
i
y

'
:
- Calculate
* *
1
ln( ) ln( ) for i=1,...,n
i i
t t i i i
X X x y
+ +
- Find the smallest integer i n such that 1
i
t
X . If such an i exists, let
the write-down
*
( )
i
j t
W w X
13
. Otherwise W
j
=0.
3) Calculate the value of the defaultable bond according to the expression:
( )
1
1
( , ) 1
M
rT
j
j
B X T e w
M
_

,
.

13
We base our write-down function on Zhous suggestion, but impose a liability restriction in order to
avoid recovery rates below 0. Our write-down function is therefore w(X)=min{w
0
-w
1
X,1}.

39
The complete computer code can be found in the appendix and run from the
accompanying CD.
4.3.3 Implementation Issues
Due to the fact that Monte Carlo methods are computer intensive we choose to implement
the model in the C programming language and Visual Basic for Applications (VBA).
With our speed test it turns out that the same algorithm runs 97% faster in C than in
VBA. This number varies with the implementation and tuning, but it hints to the speed
potential of C. The choice of C for another compiled language
14
is due to the fact that C
is a wide spread industry standard.
15
Therefore, we implement the number crunching
algorithms using the C compiler Bloodshed Dev-C++ version 4.9.8.3 Beta (Bloodshed
2003) on the Windows platform. In order to create user-interfaces and display graphics,
we chose the industry standard spreadsheet, Microsoft Excel with VBA. VBA includes an
option to link external programs with it by using dynamically linked libraries (dll).

One of the major concerns regarding C is that the built-in random number generator for
uniform deviates has poor performance (Press et al. 2002). The uniform deviates are
essential in Monte Carlo simulation and stochastic modeling. As a consequence we use
Mersenne Twister to implement a uniform random number generator (Mersenne Twister
2002). The Zhou-algorithm also makes use of standard normal distributed numbers, so in
order to sample from this distribution we use the Box-Muller algorithm described in
Glasserman (2003).

Our first runs of the model are designed to verify that the principle of Monte Carlo works
as expected. Figure 4.1 shows how the bond value converges towards the correct number
as the number of simulations increases.

In general the uncertainty is given by the standard error /
M M
M (Jckel 2001)
16
.
M
is normally uncertain and could be subject to further analysis
17
. We normally use
50000 in this paper, as a trade-off between accuracy and speed.

14
The code is pre-compiled to binary file format, in contrast to VBA code, which runs as it is typed.
15
C is a broadly speaking a subset of C++ programming language, where the main advantage of the latter is
its support for object orientation a feature we did not need.
16
Jckel uses N, but in order to follow Zhou (1997a) and our code we use M for number of simulations.
17
See for example Ross (2000) for a quick method on how to estimate
N
and the number of simulations.

40
0,577
0,579
0,581
0,583
0,585
0,587
0,589
0 10 000 20 000 30 000 40 000 50 000 60 000 70 000 80 000 90 000 100 000
Number of simulations m
B
o
n
d
v
a
lu
e

The maturity time is split into n intervals to emulate continuous time. We are interested in
keeping this number as low as possible because the Monte Carlo simulation time is close
to linearly dependent of n. Our variation of time steps yields Figure 4.2.

0,250
0,300
0,350
0,400
0,450
0,500
0,550
0,600
0,650
0 50 100 150 200 250 300 350 400 450 500
No of t i me st eps

X =
V/K
r
f
Mat. T Time
step
No.
MC
w_0 w_1

2
. Diff

X

2.0 5% Varying Varying 50000 1.4 1.0 1.0 0.0 0.25 0.0225 0.035
Figure 4.2 Varying number of time steps in Monte Carlo simulation
With a low number of time steps the bond value is a few percents higher than at the
converged part of the graph. The intuition comes as the value process is not the regular
Brownian motion looking path, but rather a straight line. As such, the implementation is
embedded in just one step and we include those paths that would have otherwise passed
the default barrier in situations with higher resolution. When n passes approximately 50,
the graph converges. We use n=100 further in this paper.
X =
V/K
r
f
Mat. T Time
step
No. MC w_0 w_1

2
. Diff

X

2.0 5% 10 100 Varying 1.4 1.0 0.05 0.0 0.25 0.0225 0.035
Figure 4.1 Convergence of Monte Carlo simulation

41
4.4 Pricing Credit Default Swaps
As previously described, the credit default swap involves the exchange of regular
payments at fixed intervals for a contingent payment in the case of a credit event. Zhou
(1997a) sets the first steps into pricing a CDS and we use it as a basis when we derive an
expression for the credit default spread, s
CDS
.

Let G(X
t
) be the default contingent paid at default. If we let A(t ,T) be the present value
of G(X
t
) it follows that it will be equal to the discounted expected write-down under the
risk-adjusted probability measure:
( )
( )

( ) ,
r P
T
A T e E G X I
1
]

By using the results from the Monte Carlo simulation we know the time of default, t , and
the write-down, ( ) w X
of the bond. W(X

T
)=0 when no default occurs so the present
value of the contingent payment turns into:
( )
( )
( )
( )
1 (1 )
r T r
A X e e w X

As we know t , it is also trivial to calculate the number of contingent legs that are paid on
the CDS before default or maturity;
{ }
( , , ) min , C T p T p p 1
1
where p is the number
of contingent payments per year. For simplicity we assume that the protection buyer must
pay the periodic fee for the whole outstanding period if default happens between two
payment dates. The present value of the default payment must equal the present value of
all regular payments. In order to calculate the periodic payment, K, we use the well
known amortization formula:
( )
( ) ( )
( )
( , , )
, 1
, ,
1
C T p
A T d
K C d A
d d

where d is the periodic discount factor given by
r
p
d e
. Finally, we calculate the CDS

spread by using amortization on all payments during a year:
( ) ( )
( )
( )
( )
( , , )
, , 1 1
( , )
1 1
p p
CDS
C T p
d K C d A d d
s A T
d d

In the following chapter we will make use of this formula during the simulation in order
to estimate the credit spread.
4.5 Results
Here we present some findings that shed light of the mechanics of the model. In general
we use Zhous (2001) values as reasonable initial values and keep the short-term risk-free

42
interest rate constant. Below we give an analysis of the output when varying a limited
number of parameters simultaneously.

By keeping the volatility of the assets and diffusion process constant over time,
2 2 2
X
+ , we may investigate the importance of varying
and
2
while
keeping their product constant.

0
5
10
15
20
25
30
35
40
45
50
0 2 4 6 8 10
Maturity
B
o
n
d

s
p
r
e
a
d Jump_vol=1
Jump_intensity=0,0125
Jump_vol=0,25
Jump_intensity=0,05
Jump_vol=0,00325
Jump_intensity=3,8461
5384615385

X =
V/K
r
f
Mat. T Time
step
No. MC W
0
W
1

2
. Diff

2
X

2.0 5% Varying 100 50000 1.4 1.0 Varying 0.0 Varying 2,25% 3,5%
Figure 4.3 Varying jump intensity and jump volatility while keeping total asset volatility constant
For the graph with high default intensity and low jump volatility we clearly see the
reason behind the criticism of the structural models. At short maturities, the jumps are too
small to cause sudden default, thereby leading to zero short-term bond spread. For longer
maturities however, the frequent jumps add to the diffusion volatility and increase the
spread to the other curves.

Jump intensity equal to 0.05 creates a high spread throughout the period. This is because
the balance of frequent jumps with considerable amplitude creates jumps that are large
enough to cause a default, while they also happen sufficiently frequently.

If we instead hold lambda constant and distribute the still fixed total volatility between
the jump volatility and Brownian motion component we get the graph represented below.


43
0
10
20
30
40
50
60
70
80
90
0 1 2 3 4 5 6 7 8 9 10
Maturity
B
o
n
d

s
p
r
e
a
d
Diffusion_vol=0,001
Jump_vol=0,68
Diffusion_vol=0,0175
Jump_vol=0,35
Diffusion_vol=0,034
Jump_vol=0,02

X =
V/K
r
f
Mat. T Time
step
No. MC W
0
W
1

2
. Diff

2
X

2.0 5% Varying 100 50000 1.4 1.0 0.05 0.0 Varying Varying 3,5%
Figure 4.4 Varying jump and diffusion volatility while keeping total asset volatility and jump intensity
constant
Again we notice that when the variance of the jump component is close to zero the short-
term bond spread is close to zero. Due to the relationship between firm value at default
and the write-down it is of major importance whether the default is caused by a jump or
by the diffusion process. Values below the barrier may only be reached by jumps and
lead to high write-down and low recovery. Under the observations made here, short-
maturity bonds have usually lower expected recovery rates than bonds with longer
maturity (Zhou 2001).

At the other extreme we have maximum variance in the jump component. Here the
Brownian motion is weak, and we observe a declining bond spread. In most textbooks the
term structure of interest rates are considered to be upward sloping. However, as Zhou
(2001) points out, there exist hump-shaped credit curves and even downward sloping
term structures. Especially bonds with high credit risk and low credit rating (B-rated or
BB-rated) may have a curve following the latter structure. The reason lies in the way
default risk is regarded for these firms. They are rated as junk bonds today, leaving them
with a high bond spread and a high default probability. If they still survive in the long run
it may be interpreted as a sign that the company has sound performance and thereby
deserves a lower bond spread given that they survive the first critical years.


44
0
50
100
150
200
250
300
350
400
0 2 4 6 8 10
Time
B
o
n
d

s
p
r
e
a
d
X_0=1.3
X_0=1.70
X_0=2.50

X =
V/K
r
f
Mat. T Time
step
No. MC W
0
W
1

2
. Diff

2
X

Varying 5% Varying 100 50000 1.4 1.0 0.05 0.0 25% 2.25% 3.5%
Figure 4.5 Varying the initial debt to asset value
X
0
is the firm value to default threshold ratio. Varying the initial ratio has substantial
impact on the bond spread. Higher X
0
yields lower bond spreads as the initial probability
of default is lower. When the debt ratio is high, i.e. X
0
is low, the bond spread decreases
again if the maturity is longer than approximately two years. This is because the
probability of default is large in the beginning of the period. If the firm is still alive after
a few years, it is more likely that it will make it till the maturity of the bond. These results
are sensitive to the choice of the other parameters and thus we cannot draw any
conclusions based on the results presented here.

0
10
20
30
40
50
60
70
80
90
100
0 1 2 3 4 5 6 7 8 9 10
Maturity
B
o
n
d
s
p
r
e
a
d
r= 0,02
r= 0,06
r= 0,15

X =
V/K
r
f
Mat. T Time
step
No.
MC
W
0
W
1

2
. Diff

2
X

2.0 Varying Varying 100 50000 1.4 1.0 0.05 0.0 25% 2,25% 3,5%
Figure 4.6 Varying the risk-free interest rate

45
Now we vary the risk-free interest rate. At short maturities the rate has little impact and
all graphs give approximately similar bond spreads. The reason is that at short maturity,
the spread is mainly caused by the jump term of the process making the spread less
dependent on the influence of r on the process drift.

More interesting is what happens at longer maturities. With high interest rates, the
expected drift of the assets is high. This drift will diminish the possibilities for default,
leading to low bonds spreads at longer maturities. For low rates we get the opposite
effect, where the long maturity augments the possibility for default and thus increases the
bond spread. It may also be inviting to think that the spread is influenced by the discount
rate of the bond, but the following calculation shows that this assumption is not correct:
[ ] ( )
[ ] ( ) ( )
( ) [ ] ( )
[ ] ( )
ln( )
, 1
ln 1
ln ln 1

ln 1

rT
rT
rT
B
s r B e E W
T
e E W
s r
T
e E W
r
T T
E W
T

where s is the bond spread and E[W] is the simulated write-down.

In order to illustrate the difference between the simulation and the closed-form solution
we have derived, we review Figure 4.4 again with the closed-form solutions included.
-5
15
35
55
75
95
115
0 1 2 3 4 5 6 7 8 9 10
Maturity
B
o
n
d

s
p
r
e
a
d
Diffusion_vol=0,001
Jump_vol=0,68
Jump_vol=0,35
Diffusion_vol=0,034
Jump_vol=0,02
Closed_f. Diff.vol=0,001
J.vol=0,68
Closed_f. Diff.vol.=0,0175
J.vol=0,35
Closed_f. Diff.vol.=0,034
J.vol=0,02

X =
V/K
r
f
Mat. T Time
step
No.
MC
W
0
W
1

2
. Diff

2
X

2.0 5% Varying 100 50000 1.4 1.0 0.05 0.0 Varying Varying 3,5%
Figure 4.7 Corresponding simulation and closed-form solution while varying jump and diffusion volatility
and keeping total asset volatility and jump intensity constant.

46
The closed-form solution is similar to Merton in that it only accepts default at the end of
the period; T . Economically this means that the shareholders option on the firm
value has become a regular call option, in contrast to a knock-out barrier option in the
simulation. The uppermost graphs are calculated from a large
, displaying a large error

compared to the simulated bond spread. This illustrates one of the advantages from a first
passage time model compared to one of Merton type. The middle closed-form graph
seems to fit quite well with the simulation, while the one with a higher diffusion build up
error with maturity. The large errors produced illustrate the effect of erroneously
modeling equity as a European option.

There are two sources of error when using the closed-form. The lack of path dependency
results in fewer defaults and a lower bond spread. This is also shown for the two data sets
with lowest jump volatility. However, raising the jump volatility yields higher bond
spreads with closed-form solution than with simulation. Firms are allowed to continue
past the default barrier until maturity of the bond, resulting in higher write-downs than
would have otherwise occurred. With higher jump volatility, the asset values are allowed
to drift well below the barrier and thereby increasing the bond spread above the simulated
spread.

0
50
100
150
200
250
300
0 1 2 3 4 5 6 7 8 9 10
Maturity
B
o
n
d

s
p
r
e
a
d
/

C
D
S

s
p
r
e
a
d
Bond_spread:
Diffusion_vol=0,001
Jump_vol=0,68
CDS_spread:
Diffusion_vol=0,001
Jump_vol=0,68
Bond_spread:
Jump_vol=0,35
CDS_spread:
Jump_vol=0,35
Bond_spread:
Diffusion_vol=0,034
Jump_vol=0,02
CDS_spread:
Diffusion_vol=0,034
Jump_vol=0,02

X =
V/K
r
f
Mat. T Time
step
No.
MC
w_0 w_1

2
. Diff

X

CDS
year

K
2.0 5% Varying 100 50000 1.4 1.0 0.05 0.0 Varying Varying 3,5% 2 1
Figure 4.8 Corresponding bond spread with credit default swap spread
From Figure 4.8 we see that the credit spread is always higher than the bond spread. This
result is as expected and follows by a closer investigation of what happens at default. A
bondholder will receive the discounted recovery value from the maturity of the bond.
With the CDS the situation is slightly different as the protection buyer is paid off
immediately. As shown in section 4.4 he will receive the difference between notional and

47
the recovery value at maturity, thus leaving him with a higher payoff than the bond
obligor.
4.6 Calibration
There are several unknown parameters that need to be estimated in the model, but we will
focus on the estimation of five of them { } , , , ,

. To our knowledge, no attempt
has been made to estimate the parameters of the Zhou model and test them empirically.

As opposed to the jump-diffusion models proposed by Merton (1976) and Ball and
Torous (1985), Zhou is modeling the drift of firms assets and not the stock value. This
issue raises the complexity of the calibration problem significantly. The reason lies in the
fact that asset values cannot be inferred from the firms balance sheet and neither be
observed in the market. As such, there exist no market data that can be applied directly to
the calibration problem though there are still a great number of securities that can be
used, of which we will discuss four; bond prices, CDS-prices, option prices, and stock
prices.

Bond prices and CDS-prices are only quoted for a few time intervals, thus providing
limited information. The most probable solution is that several sets of parameters match
our scarce data, but that none matches well with the data in general. Next, option prices
are not quoted for long enough maturities, giving no information on what the market
expects in the whole maturity period of the our simulation. If we instead of calibrating
our model to the forward-implied volatility calibrate it to historical volatility, we can use
the above data as well as equity prices to an accurate calibration. It is important to notice,
however, that the parameters will change over time. Analogous to equity derivatives it
would be more correct to use the implied volatility than the historical one.

Schnbucher (2003) describes how firm value and firm volatility can be inferred from
information on share prices, implied volatility on shares and the capitalization of the firm.
He applies the Black and Scholes option pricing framework, though requiring that assets
are purely governed by a diffusion process and that there is no path dependence. These
assumptions could have been relaxed separately and solved respectively by the Merton
(1976) or Black and Cox (1976) frameworks. However, the Zhou model comprises of
jumps as well as path dependency when it comes to default detection. Consequently, the
link between equity and assets are invalidated and more advanced techniques are
required.

Instead several of the data sources described above have to be combined. It is beyond the
scope of this paper to go into detail, but such a model could be built by defining graphical

48
models specifying conditional independence. BUGS
18
is a widely used program for
analyzing and solving complex statistical problems without an exact solution.

In order to estimate our model we choose an approximation. Still we make use of equity
data, but we run the Zhou algorithm iteratively to find the real initial asset value. The
algorithm is like the following:

1. Choose an arbitrary initial value for the assets, V
0

2. Do until error is less than pre-specified level
- Run Zhou-simulation with V
i
as firm value
- Calculate new value by using the bond value from the simulation
V
i+1
= BondValue * FaceValueDebt + Equity

Figure 4.9 Convergence of algorithm calculating firm value. Initial firm value is set to 5000, face value of
debt to 1500, equity to 1550 and other parameters to the values used in chapter 4.5.
This algorithm converges within a few simulations regardless of the initial choice of V
0
.
(See Figure 4.9) Using this technique we have the market value of the debt today and can
use it to calculate time series with asset values from historical share prices. The firm
value will still change over time and it may be necessary to run the algorithm on fixed
intervals. We assume that debt is continuously renewed everything equal and as such,
that market value of debt is constant.

Under the assumption that we have a time series of asset values, there exist a variety of
ways the model could be calibrated. Traditionally jump-diffusion models have been
estimated by using the maximum likelihood (ML) approach [Walpole et al. (1998);
Johnson and Wichern (2002)]. Still ML is largely criticized and proved erroneous by
empirical tests made by Honor (1998), Craine et al. (2000) and Lin and Huang (2001).
The problem lies in their likelihood function, which under certain conditions is

18
Bayesian Inference using Gibbs Sampling
0
1000
2000
3000
4000
5000
6000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Number of iterations
F
i
r
m
v
a
lu
e

49
unbounded. Results show that the estimator ignores the jump process and splits the data
into high and low volatility regimes for the jump and diffusion process respectively. The
jump probability therefore turns out to be overestimated while the jump volatility is
underestimated. Honor (1998) suggests limiting the parameter set to avoid unbounded
situations or to add a second jump term with constant jump volatility. In the following we
will, however, turn to the use of Bayesian inference statistics and Markov Chain Monte
Carlo simulation.
4.6.1 Markov Chain Monte Carlo
Let us turn to an estimation method with a wide number of possible applications within
finance as well as for other statistical purposes. By the use of Markov chain Monte Carlo
(MCMC) one calculates a time-reversible Markov chain with a transition kernel ( ) | q y x
for proposing the new candidate y when the chain is in state x. In the following we
present the underlying concepts of MCMC for the continuous case and show how these
can be applied in estimating our model. (Johannes and Polson 2002)

Under the assumption that the Markov chain is irreducible, aperiodic and recurrent it can
be shown that the stationary distribution of the Markov chain converges to the posterior
distribution for a set of unknown, stochastic parameters, T. The first step is to estimate a
prior distribution for these parameters. Usually there is no certain information on these
distributions, but for the most general MCMC algorithms it hardly imposes any problems.
Nevertheless, as we will show, there exist certain conditions on the knowledge of priors
that make a more efficient algorithm called the Gibbs sampler possible.

Next, we need the likelihood function for T observed data points, V, given a set of
parameters:
( ) ( )
1
| |
T
i
i
V V
l .
Finally, we need an expression for the density function for all unknown parameters given
a set of observed data:
( )
( ) ( )
( ) ( )
|
|
|
V
V
V d
l
l
.
The integral, however, is generally non-trivial to solve analytically. As a consequence we
use the Bayesian framework to calculate a proportional expression for the posterior
distribution. Due to the fact that ( ) V is unknown and not dependent on T we form a
posterior distribution that is not normalized, but which we could sample from:
( )
( ) ( )
( )
( ) ( )
|
| |
V
V V
V

l
l


50
In cases where the likelihood function and the priors give a conditional posterior
distribution that can be expressed in closed-form, one can make use of the Gibbs sampler
due to the phenomenon of conditional conjugacy. A more general algorithm available is
the Metropolis-Hastings algorithm, which does not require knowledge of all the
conditional distributions in advance. The algorithm will be discussed in the following as
we propose our estimation model.
4.6.2 Calibration Methodology
In this section we suggest a methodology for estimating the model under the assumption
that we have historical realization of the firm values. Based on Lin and Huang (2001) we
assume the priors to have the distributions motivated below:
The drift parameters are given normal distribution:
( )
2
0
, N : and
( )
2
1
, N

:
The standard deviation is always greater than zero. In the literature it is normal to
use the inverse of the variance to estimate the prior distribution. We suggest using
a chi-squared distribution with grades of freedom:
2 2
s h

: and
2 2
s h

: where
2
1
h
and
2
1
h

The probability of a jump through a certain time period is and must be between
0 and 1. A proper choice for prior distribution will then be: ( )
0 0
, Beta :
In order to avoid under- and overflow during the simulation it is a convenient standard to
use log-scale when calculating the priors. There are a number of parameters in the prior
distributions that are assumed to be predefined. These parameters have to be tuned to get
an acceptance rate boosting the performance of the algorithm. Normally tuning the rate is
not critical, but a recommended acceptance rate lies between 25 and 50%.

Further, we assume mutually independence between the prior distributions. Hence, the
prior density kernel can be described as:
( ) ( ) ( ) ( ) ( ) ( )

In addition we need a proposal distribution
( )
( 1) ( )
|
g g
q
+
in order to make use of
Metropolis-Hastings, which serves as a basis for our algorithm rendered below.


51
1. Set the initial parameters { }
(0)
, , , ,

and the parameters of the prior
distributions
{ }
2 2
0 1 0 0
, , , , , , , s h s h

.
2. Initialize a matrix, X, to store from each step in the simulation.
3. For g = 1 to G + burn-in period
- Draw { } ( )
1, 2,..., 5 i Uniform : to decide which variable to update.
- Set
( ) ( 1) g g
i i

where i indicates all variables but the one that is
updated.
- Calculate the asset values from the algorithm presented in the previous
section.
- Generate a candidate value
( )
* ( )
|
g
i i i i
q

: .
- Calculate
( )
( ) ( )
( ) ( )
* ( 1) *
( 1) *
( 1) * ( 1)
|
, exp min log , 0
|
g
g
g g
q
q

1 _

1
' ;

1

, ]
.
- Draw ( ) 0,1 U Uniform :
- If
( )
( 1) *
,
g
U

< then set
( ) * g
i i
.
- Else set
( ) ( 1) g g
i i

.
- If g > burn-in periods then update matrix X.
4. Calculate average and standard deviation for each parameter using X.

By choosing symmetric proposal distributions the acceptance probability turns into
( )
( )
( )
*
( 1) *
( 1)
, min log , 0
g
g
e

1 _

1
' ;

1

, ]
as
( ) ( )
( 1) * * ( 1)
| |
g g
q q

, greatly
simplifying the algorithm. For some of the distributions we are able to take advantage of
conditional conjugacy and derive the posterior distribution and hence, use the Gibbs
sampler. Lin and Huang (2001) show how the prior distributions above multiplied with
the likelihood function can be turned into specific posterior distributions which can be
drawn from directly.

We implement an MCMC with the background in Lin and Huang (2001) by using the
software package and programming language R.19 It is a basically a free, limited version
of S-PLUS, which is widely used for statistical purposes due to its simplicity when it
comes to data manipulation and graphical display. (Venables and Smith 2003) The full
code and results from the estimation is found in the appendix (D).

19
R can be regarded as an implementation of the S language.

52
0 2 0 0 4 0 0 6 0 0 8 0 0 1 0 0 0
0
.
1
0
.
2
0
.
3
0
.
4
0
.
5
0
.
6
0
.
7
T r a c e p lo t d r i f t d i f f u s i o n
Iter at i ons

Figure 4.10 Trace plot of drift diffusion parameter in MCMC estimation
A trace plot of each of the simulated parameters is a necessary tool to determine the
efficiency and state of the process. If we consider the trace plot of the jump drift variable
we see that the plot oscillates around an equilibrium, which is a typical pattern for this
kind of plots. Still there is no guarantee of convergence
20
and in many situations a large
data sample is needed. It is normal to remove the first steps, the burn-in period, of the
MCMC process to be able to calculate the empirical expectation and variance practically
independent of the initial parameters.
4.7 Suggested Extensions
Even though the results presented in previous chapters capture the basic features
observed in the market, there is a long way until we can conclude that the model
represents the real world in a satisfactory manner. First of all, as we have also discussed
in the previous chapter, the model has to be calibrated to historical and current data.
Estimating the parameters is a lengthy process and involves computer intensive methods
and a large amount of data. Moreover, as also mentioned by Zhou, the model lacks
empirical testing with statistical data. The flexible framework makes the model suitable
for most observed features regarding default probabilities, recovery rates and credit
spreads. However, we await extensive research on how well the models adapt to the
mentioned features.

20
It is impossible to prove convergence, but the suggested techniques give an idea of the state of the
process.

53
When it comes to the model itself, Zhou (1997a) suggests relaxing the assumption on
constant short-term interest rate with a stochastic interest rate. The risk-free interest rate
is then assumed to follow a diffusion process following the stochastic differential
equation proposed by Vasicek (1977)
21

( )
2
dr r dt dZ +
where dZ
2
is a standard Brownian motion correlated with dZ
1
in the jump-diffusion
process of the assets by dt . Both dZ
1
and dZ
2
are still independent of dY and . As a
result Zhou presents the following Feynman-Kac
22
solution to the bond value under the
risk-adjusted probability measure:
~
0
( , , ) exp ( (1 ( ))
T
rdt
P
T T
B X r T E I w X I

_

,
>
1
1
+
1
1
]

Using stochastic interest rate would add another parameter to our model, as there would
be a need to estimate the correlation between the asset diffusion process and the interest
rate.

There are also reasons to believe that the volatility will be time-dependent and not
constant as assumed in our implementation. A firm which is on the way to enter, or is
already entering financial distress, is likely to have a different volatility than it had
previous to this situation. Hand and Jacka (1998) give an overview of candidate models
for stochastic volatility. Clearly such an extension will add further complexity to the
model, but Craine et al. (2000) shows that the calibration of such a stochastic volatility
jump-diffusion model yields substantially improved results compared to the estimation of
a jump-diffusion process. Such a model will allow for volatility clustering without
producing non-reasonable values of the jump intensity. Research shows that the jump
intensity tends be overestimated in jump-diffusion models without a stochastic volatility
term as data is arranged into a high and low volatility regimes (Craine et at. 2000; Honor
1998; Lin and Huang 2001).

Another possible extension to the model is to alter the write-down function we are
currently using, in order to capture more advanced and realistic features that occur at
default. It may also be necessary to use different functions depending on the seniority of
the debt that is priced.

Taking these suggested extensions into account it is obvious that the flexibility and power
of Monte Carlo simulations make the inclusion of new features in the model easy. It may
also be interesting to price other securities and derivatives than the zero-coupon bonds

21
Other interest rate processes could also have been used.
22 The Feynman-Kac theorem connects a class of conditional expectations of future payoffs under risk-
neutral measure and certain partial differential equations. (Jckel 2001)

54
and credit default swaps we implement, and they only demand minor changes in the code
in order to incorporate the desired structures. For example it would be trivial to use the
framework to price coupon bonds or exotic structures with the same underlying.

Another issue that is eligible for further investigation is the risk premium of jump risk,
which is assumed to be zero due to the non-systematic nature of jumps. In the market this
is not necessarily reality and Zhous framework may be a way to observe the actual risk
premium demanded in the market. Even though the jumps are completely asset-specific
and uncorrelated with the market, Merton (1976) claims that it in practice will be
impossible to diversify the jump risk. Given that investors will not be able to diversify the
jump risk there will be a premium for this risk. As jumps in asset value are not directly
observable we propose to estimate this premium via the equity markets and equity option
markets. If we use down-and-out equity options when estimating this risk premium it
would be analogous to the equity with regard to the asset value. We would assume that
the risk premium for jump risk with equity as the underlying would be equal to the jump
risk when the asset value is the underlying.

55
5 Concluding Remarks
This paper gives an overview of the vast area of credit derivatives and different products
in the market. There exist two main approaches used to model credit risk; the structural
approach based on the diffusion of a firms value and the reduced form approach based
on the hazard rate. The two approaches are theorized by presenting models proposed in
the literature.

In particular, we discuss and implement a jump-diffusion model proposed by Zhou
(1997a). Through a closed-form solution and a Monte Carlo simulation we price
defaultable bonds and credit default swaps. From the results obtained by varying input
parameters we see that the model is capable of resembling observed features of bond and
credit spreads, with declining, increasing and hump-shaped term structures. We introduce
methodology for estimating the model, but still the model awaits empirical testing and the
investigation of possible extensions by including stochastic volatility and short-term
interest rate.

In this paper we only model individual securities. To model and price the more advanced
credit derivatives briefly presented, the modeler have to accommodate for default
correlation and default dependency. Though demanding, these are necessary features in
order to develop models suitable for credit analysis and risk management.

Reduced form models have a clear advantage over structural models in that they easily
can be calibrated to prevent arbitrage opportunities. The advantages of structural models
on the other hand are that they are based on solid economic arguments and provide a link
between equity and credit. This link is exploited in models like the E2C model where
equity is used as a proxy hedge for credit. The use of equity to hedge credit or vice versa
called capital structure arbitrage (Euromoney 2002) is an area that has received much
attention the last couple of years. However, as we point out, a caveat with the models
currently used for this purpose is that they model equity as an in-the-money European
call while it should be modeled as a down-and-out barrier option. The general model by
Zhou, though significantly more difficult to calibrate and more computer intensive,
would mitigate the problem of defaults before maturity in current models. For further
research we therefore also suggest empirical tests to verify Zhous model with regard to
capital structure arbitrage.


56
6 Bibliography
Ammann, Manuel (1999), Pricing Derivative Credit Risk, Berlin: Springer-Verlag

Ball, Clifford A. and Walter N. Torous (1985), On Jumps in Common Stock Prices and
Their Impact on Call Option Pricing, Journal of Finance, 40, 155-173

Bank for International Settlements (2003), http://www.bis.org/bcbs/aboutbcbs.htm

Black, Fischer and John C. Cox (1976), Valuing Corporate Securities: Some Effects of
Bond Indenture Provisions, Journal of Finance, 31, 361-367

Black, Fischer and Myron Scholes (1973), The Pricing of Options and Corporate
Liabilities, The Journal of Political Economy, 81, 637-654

Black, Fischer, Emanuel Derman and William Toy (1990), A One-Factor Model of
Interest Rates and Its Application to Treasury Bond Options, Financial Analysts
Journal, 46(11)

Bloodshed Software (2003), http://www.bloodshed.net

Bohn, Jeffrey R. (2000), A Survey of Contingent-Claims Approaches to Risky Debt
Valuation, The Journal of Risk Finance, Spring 2000, Vol 1, p. 53-71

Cossin, Didier and Hugues Pirotte (2001), Advanced Credit Risk Analysis, Chichester,
UK: John Wiley & Sons Ltd.

Cox, John C., Jonathan E. Ingersoll Jr. and Stephen A. Ross (1985), A Theory of the
Term Structure of Interest Rates, Econometrica, 53, 385-408

Craine, Roger, Lars A. Lochstoer and Knut Syrtveit (2000), Estimation of a Stochastic-
Volatility Jump-Diffusion Model, Working Paper, University of California at
Berkeley

Das, Sanjiv Ranjan (2001), Pricing Credit Derivatives, Working Paper, Santa Clara
University

Das, Sanjiv Ranjan, and Peter Tufano (1995), Pricing Credit-Sensitive Debt When
Interest Rates, Credit Ratings and Credit Spreads are Stochastic, Working paper,
Harvard University

Das, Satyajit (2000), Credit Derivatives and Credit Linked Notes, Singapore: John
Wiley & Sons Ltd.

Duffie, Darrell and David Lando (2001), Term structures of credit spreads with
incomplete accounting information, Econometrica, 69, 633-664


57
Duffie, Darrell and Kenneth J. Singleton (1999), Modeling Term Structures of
Defaultable Bonds, The Review of Financial Studies, 12, 687-720

Economist (2002), The regulator who isn't there, The Economist, May 16
th
2002

Economist (2003), Deep impact, The Economist, May 8
h
2002

Euromoney (2002), And Now for Capital Structure Arbitrage, Euromoney, December
2002

Geske, Robert L. (1977), The Valuation of Corporate Liabilities as Compound Options,
Journal of Financial and Quantitative Analysis, 12, 541-552.

Giesecke, Kay (2001), Default and information, Working Paper, Cornell University

Giesecke, Kay (2002), Credit risk modeling and valuation: an introduction, Working
paper, Cornell University

Giesecke, Kay and Lisa R. Goldberg (2003), Forecasting default in the face of
uncertainty, Working paper, Cornell University & Barra, Inc.

Glasserman, Paul (2003), Monte Carlo Methods in Financial Engineering, New York:
Springer-Verlag

Hand, David J. and Saul D. Jacka (1998), Statistics in Finance, London: Arnold
Applications of Statistics Series

Heath, David, Robert Jarrow and Andrew Morton (1992), Bond Pricing and the Term
Structure of Interest Rates: A New Methodology for Contingent Claims Valuation,
Econometrica, 60, 77-105

Honor, Peter (1998), Pitfalls in Estimating Jump-Diffusion Models, Working Paper,
The Aarhus School of Business

Houwelingen, Patrick (2003), Empirical studies on Credit Markets, PhD thesis.
Erasmus University Rotterdam

ISDA (1992), http://www.isda.org/publications/1992masterlc.pdf

ISDA (2003), http://www.isda.org/statistics/recent.html

Jarrow, Robert A. and Stuart M. Turnbull (1995), Pricing Derivatives on Financial
Securities Subject to Credit Risk, The Journal of Finance 50, No. 1 (Summer), 53-
85

Jarrow, Robert A., David Lando and Stuart M. Turnbull (1997), A Markov Model for
the Term Structure of Credit Risk Spreads, The Review of Financial Studies, 10,
481-523


58
Johannes, Michael and Nicholas Polson (2002), MCMC Methods for Financial
Econometrics, Working paper, To appear in Handbook of Financial Econometrics

Johnson, Richard A. and Dean W. Wichern (2002), Applied Multivariate Statistical
Analysis, Upper Saddle River, USA: Prentice Hall, Inc., 5
th
edition

Jckel, Peter (2001), Monte Carlo Methods in Finance, Chichester, UK: John Wiley &
Sons Ltd

Kesdee (2003), http://www.kesdee.com/html/cobasel.html

Lando, David (1998), On Cox processes and credit risky securities, Working paper,
University of Copenhagen

Lardy, Jean-Pierre (2000), E2C: A Simple Model to Assess Default Probabilities from
Equity Markets, JP Morgan Credit Derivatives Conference, January 16, 2002

Lehman Brothers (2001), Credit Derivatives Explained Market, Products and
Regulations, Structured Credit Research

Lin, Shinn-Juh and Ming-Tui Huang (2001), Estimating Jump-Diffusion Models using
the MCMC Simulation, Working Paper, National Tsing Hua University, Taiwan

Longstaff, Francis A. and Eduardo S. Schwartz (1995), A Simple Approach to Valuing
Risky Fixed and Floating Rate Debt, The Journal of Finance, 50, 789-819

Madan, Dilip B. and Haluk Unal (1994), Pricing the Risks of Default, Working paper,
University of Maryland

Mersenne Twister (2002), http://www.math.keio.ac.jp/matumoto/emt.html

Merton, Robert C. (1974), On the pricing of corporate debt: The risk structure of interest
rates, Journal of Finance, 29, 449-470

Merton, Robert C. (1976), Option Pricing when Underlying Stock Returns Are
Discontinuous, Journal of Financial Economics, 3, 125-144

Moodys (2002), Default & Recovery Rates of European Corporate Bond Issuers, 1985-
2001, Moodys Investor Service, Global Credit Research, July 2002

Neftci, Salih N. (2000), An Introduction to the Mathematics of Financial Derivatives,
San Diego: Academic Press

Press, William H., Saul A. Teukolsky, William T. Vetterling, Brian P. Flannery (2002),
Numerical Recipes in C++, Cambridge: Cambridge University Press

Ross, Sheldon M. (2000), Introduction to Probability Models, San Diego: Harcourt/
Academic Press


59
Schmidt, Thorsten (2001), Credit Risk An Introduction to Credit Risk Modelling with
respect to Credit Derivatives Pricing, Talk presented in the Frankfurt Mathfinance
Colloquium, Goethe University

Schnbucher, Philipp J. (1998), A Tree Implementation of a Credit Spread Model for
Credit Derivatives, Department of Statistics, Bonn University

Schnbucher, Philipp J. (2003), Credit Derivatives Pricing Models, Chichester, UK:
John Wiley & Sons Ltd

Schler, Marcus (2001), Credit Derivatives The Driving Force in Credit Markets, JP
Morgan presentation

Soudaram, Rangarajan K. (2001), The Merton/KMV Approach to Pricing Credit Risk,
Extra Credit

SunGard (2003), http://www.sungard.com/products_and_services/stars/panorama/
information/albournedec2001.pdf

Tavakoli, Janet M. (2001), Credit Derivatives & Synthetic Structures, New York: John
Wiley & Sons Ltd

Vasicek, Oldrich (1977), An equilibrium characterization of the term structure, Journal
of Financial Economics, 5, 117-161

Venables, W. N. and D. M. Smith (2003), An Introduction to R, http://www.r-
project.org

Walpole, Ronald E., Myers, Raymond H., Myers, Sharon L. (1998), Probability and
Statistics for Engineers and Scientists, New Jersey, USA: Prentice Hall
International, Inc, 6
th
edition

Zhou, Chunsheng (1997a), A Jump-Diffusion Approach to Modeling Credit Risk and
Valuing Defaultable Securities, Board of Governors of the Federal Reserve System

Zhou, Chunsheng (1997b), Path-Dependent Option Valuation When the underlying path
is Discontinuous, Board of Governors of the Federal Reserve System

Zhou, Chunsheng (2001), The term structure of credit spreads with jump risk, Journal
of Banking & Finance, 25, 2015-2040


I
I
A Zhou.c dll proram code
#include "zhou.h"
#include <windows.h>
#include <stdio.h>
#include <stdlib.h>
#include <math.h>

/* This C code file is developed with the free Bloodshed Dev-C++ (and C) compiler
under Microsoft Windows XP Pro*/

/* Possible improvements for the C code:
- The NAG C library is available to students at NTNU.
This library is heavily tested and used. Thus we could have
sped up CPU time the quality assured the code by using the NAG library.

- The random function we used was the best we found, and we have tested it both
for accuracy (uniform distribution) and stress (10E9). It has passed all tests.
Still there might be faster algorithms available, and possibly with less digits
(which are not crucial for an MC). We chose not to test further since the
performance is acceptable anyway. Our random() returns in the interval [0,1],
which is out of range for one func. We solved this with an if (while) for
simplicity, but we could also have modified the random() slightly.
*/

/* x(i) function defined in Zhou (2001): */
double x_i (
double discountRate,
double sigmaDiff_2,
double lambda,
double nu,
int n,
double timeStop)
{
double mu_i, sigmaDiff_2_i;

mu_i = (discountRate - (sigmaDiff_2 / 2) - (lambda * nu)) * timeStop / n;
sigmaDiff_2_i = sigmaDiff_2 * timeStop / n;

return dllRandNorm (mu_i, sqrt (sigmaDiff_2_i));
}

/* Calculate y = 0 or y * pi directly: */
double yPi (double lambda, double timeStop, int n, double mu_i, double sigma_pi_2)
{
if (genrand() > lambda * timeStop / n) {
return 0;
} else { // Return y_ * pi_i:
return dllRandNorm (mu_i, sqrt(sigma_pi_2));
}
}

/* Simple write-down function as described in Zhou (2001): */
double writedown (double w_0, double w_1, double x)
{
return min (w_0 - w_1 * x, 1);
}

DLLIMPORT double _stdcall dllPlotting(
double discountRate,
double sigmaDiff_2,
double lambda,
double nu,
double timeStop,

II
II
double mu_pi,
double sigma_pi_2,
double X_0,
double w_0,
double w_1,
int n, // for LnX
int m, // for W
double cdsPayments,
double LnX[],
double W[],
double Out[],
double K)
{
int i, j, k;
double x, pi, y, X_t;
double yPi_i = 0;
double npvW_j, cds_j, defaultTime_j;

// Initialize:
double bondValue = 0;
double bondSpread = 0;
double sumW = 0;
double sumNpvW = 0;
double sumCds = 0;
double sumCdsSpread = 0;
double cdsValue = 0;
double cdsAmt = 0;
double payments = 0;
double spreadCds = 0;
double dblN = 0;
double npvCdsW_j = 0;
double equity_j = 0;
double equity = 0;

double rateFactor = exp(-discountRate / cdsPayments);

// Be sure nu is correct:
nu = exp(mu_pi + sigma_pi_2/2) - 1;

for (j = 0; j < m; j++) { // Run from 0 to j-1

// Initialize:
LnX[0] = log(X_0);
i = 1;

while (i < n && LnX[i - 1] > 0) // Run from 0 to n-1
{
x = x_i (discountRate, sigmaDiff_2, lambda, nu, n, timeStop);
// Merge Zhou's y and pi in one go:
yPi_i = yPi(lambda, timeStop, n, mu_pi, sigma_pi_2);

LnX[i] = LnX[i - 1] + x + yPi_i;

i++;
}
// Mark end of this run with flag for use in VBA:
LnX[i] = -999;

// If default:
if (LnX[i-1] <= 0 ) {
W[j] = writedown(w_0, w_1, exp(LnX[i - 1]));

// Calculate equity:
equity_j = 0;
} else { // Not deafult:
W[j] = 0;

// Calculate equity:
equity_j = exp(LnX[i-1]) * K - 1;
}
sumW += W[j];

III
III

dblN = n;
defaultTime_j = (timeStop / dblN) * (i - 1); // Years

npvW_j = exp (-discountRate * timeStop) * W[j];
sumNpvW += npvW_j; // Paid at T

npvCdsW_j = (1 - exp(-discountRate * (timeStop - defaultTime_j))
* (1-W[j]))
* exp(-discountRate * defaultTime_j);

payments = min (cdsPayments * timeStop, ceil (defaultTime_j * cdsPayments));
// Test for error values (if zero, we get div 0 later):
if (payments < 0.0001) { // Seems to work with 0.0001
payments = 0.0001;
}

cdsAmt = ((1/rateFactor) * npvCdsW_j * (1 - rateFactor))
/ (1.0 - pow(rateFactor, payments));

spreadCds = cdsAmt * rateFactor * (1 - pow(rateFactor, cdsPayments))
/ (1.0 - rateFactor);

cdsValue += cdsAmt;
sumCdsSpread += spreadCds;
equity += equity_j;
}

//bondValue = exp (-discountRate * timeStop) - sumNpvW / m; // See dllBondValue
bondValue = exp (-discountRate * timeStop) * (1 - sumW / m);

bondSpread = -log(bondValue) / timeStop - discountRate;
bondSpread *= 10000;

// Find average:
sumCdsSpread /= m;
sumCdsSpread *= 10000;

cdsValue /= m;

equity /= m;
// Discount:
equity *= exp(-discountRate * timeStop);

// Add to vector for vba to read:
Out[0] = bondSpread;
Out[1] = cdsValue;
Out[2] = sumCdsSpread;
Out[3] = equity;

return bondValue;
}

/* Box-Muller algorithm to draw from Normal distribution: */
/* Returns: random ~ N(0, 1) */
/* Uses: math.h */

/* Vars for use with dllRandSNorm and dllRandNorm: */
double randSNorm2 = 0;
int rand1 = 1;


IV
IV
const float PI = 3.1415926535;
double dllRandSNorm()
{
/*******************/
// Box-Muller:
/*******************/
double u1 = genrand();
// Avoid div zero. (This happens extremely rarely):
while (u1 <= 0.0 || u1 >= 1.0) {
double u1 = genrand();
}
double r = -2 * log (u1);
double v = 2 * PI * genrand(); // This accepts randoms of 0 or 1

/* Calculate two vars at one go, and save the other: */
randSNorm2 = sqrt(r) * sin (v);
return sqrt(r) * cos(v);
}

// Returns: random ~ N(mu, sigma)
double dllRandNorm (double mu, double sigma)
{
if (rand1 == 1) {
rand1 = 0; // False
return (mu + dllRandSNorm() * sigma);
} else {
rand1 = 1; // True
return (mu + randSNorm2 * sigma);
}
}

/***********************************************/
/* S t a n d a r d d l l f u n c t i o n : */
/***********************************************/

/* Necessary method for the dll: */
BOOL APIENTRY DllMain (HINSTANCE hInst /* Library instance handle. */ ,
DWORD reason /* Reason this function is being called. */ ,
LPVOID reserved /* Not used. */ )
{
switch (reason)
{
case DLL_PROCESS_ATTACH:
break;

case DLL_PROCESS_DETACH:
break;

case DLL_THREAD_ATTACH:
break;

case DLL_THREAD_DETACH:
break;
}

/* Returns TRUE on success, FALSE on failure */
return TRUE;
}

/*************************************/
/* R A N D O M G E N E R A T O R : */
/*************************************/

/* A C-program for MT19937: Real number version */
/* genrand() generates one pseudorandom real number (double) */

V
V
/* which is uniformly distributed on [0,1]-interval, for each */
/* call. sgenrand(seed) set initial values to the working area */
/* of 624 words. Before genrand(), sgenrand(seed) must be */
/* called once. (seed is any 32-bit integer except for 0). */
/* Integer generator is obtained by modifying two lines. */
/* Coded by Takuji Nishimura, considering the suggestions by */
/* Topher Cooper and Marc Rieffel in July-Aug. 1997. */

/* This library is free software; you can redistribute it and/or */
/* modify it under the terms of the GNU Library General Public */
/* License as published by the Free Software Foundation; either */
/* version 2 of the License, or (at your option) any later */
/* version. */
/* This library is distributed in the hope that it will be useful, */
/* but WITHOUT ANY WARRANTY; without even the implied warranty of */
/* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. */
/* See the GNU Library General Public License for more details. */
/* You should have received a copy of the GNU Library General */
/* Public License along with this library; if not, write to the */
/* Free Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA */
/* 02111-1307 USA */

/* Copyright (C) 1997 Makoto Matsumoto and Takuji Nishimura. */
/* Any feedback is very welcome. For any question, comments, */
/* see http://www.math.keio.ac.jp/matumoto/emt.html or email */
/* matumoto@math.keio.ac.jp */

//#include "mt19937.h"
//#include <windows.h>
//#include <stdlib.h>
//#include <math.h>
//#include <stdio.h>

/* Period parameters */
#define N 624
#define M 397
#define MATRIX_A 0x9908b0df /* constant vector a */
#define UPPER_MASK 0x80000000 /* most significant w-r bits */
#define LOWER_MASK 0x7fffffff /* least significant r bits */

/* Tempering parameters */
#define TEMPERING_MASK_B 0x9d2c5680
#define TEMPERING_MASK_C 0xefc60000
#define TEMPERING_SHIFT_U(y) (y >> 11)
#define TEMPERING_SHIFT_S(y) (y << 7)
#define TEMPERING_SHIFT_T(y) (y << 15)
#define TEMPERING_SHIFT_L(y) (y >> 18)

static unsigned long mt[N]; /* the array for the state vector */
static int mti=N+1; /* mti==N+1 means mt[N] is not initialized */

/* initializing the array with a NONZERO seed */
void sgenrand(seed)
unsigned long seed;
{
/* setting initial seeds to mt[N] using */
/* the generator Line 25 of Table 1 in */
/* [KNUTH 1981, The Art of Computer Programming */
/* Vol. 2 (2nd Ed.), pp102] */
mt[0]= seed & 0xffffffff;
for (mti=1; mti<N; mti++)
mt[mti] = (69069 * mt[mti-1]) & 0xffffffff;
}

/* generating reals */ /* unsigned long */ /* for integer generation */
DLLIMPORT double _stdcall genrand()
{
unsigned long y;
static unsigned long mag01[2]={0x0, MATRIX_A};
/* mag01[x] = x * MATRIX_A for x=0,1 */

VI
VI

if (mti >= N) { /* generate N words at one time */
int kk;

if (mti == N+1) /* if sgenrand() has not been called, */
sgenrand(4357); /* a default initial seed is used */

for (kk=0;kk<N-M;kk++) {
y = (mt[kk]&UPPER_MASK)|(mt[kk+1]&LOWER_MASK);
mt[kk] = mt[kk+M] ^ (y >> 1) ^ mag01[y & 0x1];
}
for (;kk<N-1;kk++) {
y = (mt[kk]&UPPER_MASK)|(mt[kk+1]&LOWER_MASK);
mt[kk] = mt[kk+(M-N)] ^ (y >> 1) ^ mag01[y & 0x1];
}
y = (mt[N-1]&UPPER_MASK)|(mt[0]&LOWER_MASK);
mt[N-1] = mt[M-1] ^ (y >> 1) ^ mag01[y & 0x1];

mti = 0;
}

y = mt[mti++];
y ^= TEMPERING_SHIFT_U(y);
y ^= TEMPERING_SHIFT_S(y) & TEMPERING_MASK_B;
y ^= TEMPERING_SHIFT_T(y) & TEMPERING_MASK_C;
y ^= TEMPERING_SHIFT_L(y);

return ( (double)y / (unsigned long)0xffffffff ); /* reals */
/* return y; */ /* for integer generation */
}


VII
VII
B VBA Program Code for Closed Form Solution
Public Sub calcClosedForm(discountRate, _
sigmaDiff_2, lambda, timeStop, _
mu_pi, sigma_pi_2, _
X_0, w_0, w_1, n, m, cdsPayments)

' Check values:
If sigmaDiff_2 <= 0 Then 'Div by zero
MsgBox "Sigma diffusion must be above zero"
Range("d18").Activate
Exit Sub
End If

'Vars are set as parts of the complete expressions in
'order to make the computation simpler.
'Refer to the paper for a decription of the algorithm.
Dim mu_nd As Double ' subscript _nd: Normal Distributed, see report
Dim sigma_nd_2 As Double ' subscript _2: squared
Dim jumpProb As Double
Dim driftPart As Double

Dim norm_n As Double
Dim norm_n_G As Double
Dim normProb As Double
Dim normProb_G As Double
Dim expValue_S As Double
Dim expValue_B As Double
Dim survProb As Double
Dim stockValue As Double
Dim bondValue As Double
Dim bondSpread As Double

Dim norm_n_br As Double
Dim normProb_br As Double
Dim normProb_F_br As Double
Dim norm_n_F_br As Double
Dim survProbF_br As Double

Dim nu As Double

'Initialize:
expValue_B = 0
expValue_S = 0
survProb = 0
survProbF_br = 0

pubBondValue = 0
pubBondSpread = 0
pubStockValue = 0

nu = Exp(mu_pi + sigma_pi_2 / 2) - 1

Dim i As Long
Dim infinity As Long
infinity = 150 ' Max is 170
For i = 0 To infinity
mu_nd = (discountRate - (sigmaDiff_2 / 2) - lambda * nu) * timeStop _
+ i * mu_pi ' Ln(X) later
sigma_nd_2 = sigmaDiff_2 * timeStop + i * sigma_pi_2

jumpProb = (Exp(-lambda * timeStop) * (lambda * timeStop) ^ i) _
/ WorksheetFunction.Fact(i)

driftPart = X_0 * Exp(mu_nd + (sigma_nd_2 / 2))


VIII
VIII
norm_n = (Log(X_0) + (mu_nd + sigma_nd_2)) / Sqr(sigma_nd_2)
norm_n_F_br = (Log((w_0 - 1) / w_1) - Log(X_0) - (mu_nd)) / Sqr(sigma_nd_2)
norm_n_G = (Log(X_0) + (mu_nd)) / Sqr(sigma_nd_2)
norm_n_br = (-Log((w_0 - 1) / w_1) + (Log(X_0) + (mu_nd))) / Sqr(sigma_nd_2)

normProb = WorksheetFunction.NormSDist(norm_n)
normProb_br = WorksheetFunction.NormSDist(norm_n_br)
normProb_G = WorksheetFunction.NormSDist(norm_n_G)
normProb_F_br = WorksheetFunction.NormSDist(norm_n_F_br)

expValue_S = expValue_S + driftPart * jumpProb * normProb
expValue_B = expValue_B + driftPart * jumpProb * (normProb_br - normProb)

survProb = survProb + jumpProb * normProb_G
survProbF_br = survProbF_br + jumpProb * normProb_F_br
Next i

stockValue = Exp(-discountRate * timeStop) * (expValue_S - survProb)

bondValue = Exp(-discountRate * timeStop) * _
(survProb + (1 - w_0) * ((1 - survProb) - survProbF_br) + w_1 * expValue_B)

'Test for erronous values:
If bondValue < 0 Then 'Log of neg. number
MsgBox "Closed form solution produces non-real result " & vbCrLf & _
"Please decrease w_0 or increase w_1."
Range("d13").Activate
Exit Sub
End If

'Times 10000 to get bps:
bondSpread = 10000 * ((-Log(bondValue) / timeStop) - discountRate)

'Uodate findings to public vars:
pubBondValue = bondValue
pubBondSpread = bondSpread
pubStockValue = stockValue

End Sub


IX
IX
C VBA Code for Changing Diffusion Volatility and Jump
Volatility
Public Sub runSigmaDiff_SigmaPi()

'Start timer:
Dim startTime As Double
startTime = Timer

'Read public variables from first sheet:
If readInput = False Then Exit Sub

'Declare variables and arrays:
Dim nu As Double
Dim dllBVRes As Double
Dim cdsPremium As Double
Dim bondSpread As Double ' = credit spread
Dim cdsSpread As Double
Dim equity As Double
'dll more stable if we do not redim (dynamic memory allocation):
Dim ArrLnX(500) As Double
Dim ArrW(100000) As Double
Dim ArrOut(0 To 5) As Double

pubBondValue = 0
pubBondSpread = 0
pubStockValue = 0

'Start working on these graphs:
Worksheets("SigmaDiff-SigmaPi").Activate

'''''''''''''''''''''''''''''''''''
'Read changing vars:
'''''''''''''''''''''''''''''''''''
Dim timeStop As Double
Dim stepLength As Double
Dim stepNumber As Long
Dim timeStopLast As Double

timeStop = Range("d23").Value
stepLength = Range("d25")
stepNumber = Range("d27")
timeStopLast = timeStop + stepLength * stepNumber

Dim i As Long
Dim j As Long
j = 0

'Sigma_diff:
Dim ArrSigmaDiff_2(0 To 2) As Double
ArrSigmaDiff_2(0) = Range("g23").Value

'Update lambda and sigma_pi:
Dim ArrSigma_pi_2(0 To 2) As Double
Dim ArrLambda(0 To 2) As Double
For i = 0 To 2
If ArrSigmaDiff_2(i) >= pubSigmaTotal_2 Then
MsgBox "sigmaDiff must be less than SigmaTotal"
Exit Sub
End If

X
X

ArrLambda(i) = pubLambda 'Constant for now
ArrSigma_pi_2(i) = (pubSigmaTotal_2 - ArrSigmaDiff_2(i)) / ArrLambda(i)
Next i

'''''''''''''''''''''''''''''''''''''''''
'Set info:
'''''''''''''''''''''''''''''''''''''''''
Range("c32").Value = ArrSigmaDiff_2(0)
Range("d32").Value = ArrSigmaDiff_2(1)
Range("e32").Value = ArrSigmaDiff_2(2)
Range("i31").Value = ArrSigmaDiff_2(0)
Range("s31").Value = ArrSigmaDiff_2(1)
Range("ac31").Value = ArrSigmaDiff_2(2)

Range("k31").Value = ArrSigma_pi_2(0)
Range("u31").Value = ArrSigma_pi_2(1)
Range("ae31").Value = ArrSigma_pi_2(2)

''''''''''''''''''''''''''''''''''''
'Run model:
''''''''''''''''''''''''''''''''''''

'Clear for new run:
Rows("33:126").ClearContents
Range("k27").ClearContents

'Loop for every time step:
Do While timeStop < timeStopLast

'Print maturity:
Range("b33").Offset(j, 0).Value = timeStop

'Loop for every sigma/lambda pair:
For i = 0 To 2

nu = Exp(pubMu_pi + ArrSigma_pi_2(i) / 2) - 1

'Print input:
printInput pubX_0, pubDiscountRate, timeStop, pubN, pubM, pubW_0, _
pubW_1, ArrLambda(i), _
pubMu_pi, ArrSigma_pi_2(i), ArrSigmaDiff_2(i), _
pubSigmaTotal_2, pubCdsPayments, pubK

'Run dll:
dllBVRes = dllPlotting(pubDiscountRate, _
ArrSigmaDiff_2(i), _
ArrLambda(i), nu, _
timeStop, pubMu_pi, _
ArrSigma_pi_2(i), _
pubX_0, pubW_0, pubW_1, pubN, pubM, pubCdsPayments, _
ArrLnX(0), ArrW(0), ArrOut(0), pubK)

'Closed form:
calcClosedForm pubDiscountRate, _
ArrSigmaDiff_2(i), _
ArrLambda(i), _
timeStop, _
pubMu_pi, _
ArrSigma_pi_2(i), _
pubX_0, pubW_0, pubW_1, _
pubN, pubM, pubCdsPayments

'Print results:
bondSpread = ArrOut(0)
cdsPremium = ArrOut(1)
cdsSpread = ArrOut(2)

XI
XI
equity = ArrOut(3)

Range("c33").Offset(j, i).Value = bondSpread

Range("g33").Offset(j, i * 10).Value = timeStop
Range("h33").Offset(j, i * 10).Value = dllBVRes
Range("i33").Offset(j, i * 10).Value = bondSpread
Range("j33").Offset(j, i * 10).Value = cdsPremium
Range("k33").Offset(j, i * 10).Value = cdsSpread
Range("l33").Offset(j, i * 10).Value = equity

' Print output from closed form:
Range("m33").Offset(j, i * 10).Value = pubBondValue
Range("n33").Offset(j, i * 10).Value = pubBondSpread
Range("o33").Offset(j, i * 10).Value = pubStockValue

Next i

j = j + 1
timeStop = timeStop + stepLength
Loop

'Print time spent:
Range("h17").Value = Timer - startTime

End Sub


XII
XII
D MCMC Results
We do not link the MCMC implementation to our model in VBA and C. The reason is
that there are too many assumptions implied so that it would not give correct results. All
initial parameters are set inside the file gibbs.r and asset data is read from file. We put
approximately 600 stock observations of Statoil in the file and estimate the parameters
from these data. The algorithm could, of course, have been extended to calculate the asset
value as described in chapter 4.6, but our purpose is here to show the methodology in use.

The graphs on the left side show the density distributions on each of the parameters while
the right side shows a trace plot of the corresponding parameter. The parameters we use
can be found in the code below.

0.2 0.3 0.4 0.5 0.6 0.7 0.8
0
3
Density drift diffusion
D
e
n
s
i
t
y
0 200 400 600 800 1000
0
.
1
Trace plot drift diffusion
Iterations
-0.006 -0.005 -0.004 -0.003 -0.002 -0.001
0
6
0
0
Density drift jump
D
e
n
s
i
t
y
0 200 400 600 800 1000
-
0
.
0
0
6
Trace plot drift jump
Iterations
110 120 130 140 150 160 170
0
.
0
0
Density jump intensity
D
e
n
s
i
t
y
0 200 400 600 800 1000
4
0
Trace plot jump intensity
Iterations
0.006 0.008 0.010 0.012
0
4
0
0
Density drift vol
D
e
n
s
i
t
y
0 200 400 600 800 1000
0
.
0
0
5
Trace plot drift vol
Iterations
0.0020 0.0025 0.0030 0.0035
0
Density jump vol
D
e
n
s
i
t
y
0 200 400 600 800 1000
0
.
0
0
2
Trace plot jump vol
Iterations


XIII
XIII
Variable Mean Standard deviation
Diffusion drift 0.4578826 0.08269648
Jump drift -0.003356848 0.0006683849
Jump intensity 138.6237 7.971066
Diffusion volatility 0.007638973 0.0598364
Jump volatility 0.002546043 0.02818999
Figure A.1 Results from parameter estimation using MCMC

The plots show that the parameters use the first iterations to approach a seemingly
equilibrium before they oscilliate around this axis. The standard deviation is high for
some of the parameteres, which is most likely due to a limited amount of data. We have
run 1000 simulations with a burn-in period on 100, which is not counted in the
calculation of the mean and standard deviation.


XIV
XIV
E R Code for MCMC
setLogReturns <- function() {
size <- length(assets)
tmp <- seq(1,size - 1,1)
for(i in 2:size) {
tmp[i-1] <- log(assets[i]/assets[i-1])
}
return(tmp)
}

calcProb <- function() {
tmp1 <- logReturn[i] - muDiff * delta
tmp2 <- delta + hDiff / hJump

oddsRatio <- (1 - intensity * delta) * sqrt(tmp2 ) /
(intensity * delta * sqrt(delta)) *
exp((tmp1 - muJump) ^ 2 / (2 * tmp2 / hDiff ) - tmp1 ^ 2 / (2 * delta /
hDiff ))
return(1 / (1 + oddsRatio) )
}

# Initial parameters in prior
muDiff <- 0.12335
muJump <- -0.00095
intensity <- 2.7
hDiff <- 80
hJump <- 500
h <- c(25,0,0,100)
h <- matrix(h,2,2)
bUnder <- c(muDiff, muJump)

# Prior variables
s2Diff <- 0.1338
vDiff <- 8
s2Jump <- 0.87
vJump <- 4.22
alfa0 <- 0.5
beta0 <- 1

# Fetch asset data from specified file and calculate log-return
assets <- scan("statoil.dat")
logReturn <- setLogReturns()
T <- length(logReturn)
delta <- 0.004 # Time between asset observations
G <- 1000 # Number of simulations
burnIn <- 100 # Number of burn-in periods

# Run simulation with Gibbs sampler and Metropolis-Hastings
xx <- matrix(0,G,5)
for(g in 1:G) {
# Initialize matrices
Bernoulli <- c(matrix(0,T,1))
R <- matrix(0,T,1)
X <- matrix(0,T,2)
s2HDiff = 0
s2HJump = 0

# Calculate the posterior distribution of the jump times
for(i in 1:T) {
if(runif(1) <= calcProb()) {
Bernoulli[i] = 1
}

tmp = sqrt(delta + hDiff* Bernoulli[i] / hJump)
R[i,1]= logReturn[i] / tmp
X[i,1] = delta / tmp
X[i,2] = Bernoulli[i] / tmp

XV
XV
}

# Draw jump intensity
intensity = rbeta(1, alfa0 +sum(Bernoulli), beta0 + T - sum(Bernoulli)) / delta

# Draw drift parameters
hUpper <- hUnder + hDiff * (t(X) %*% X)
bUpper <- solve(hUpper, (hUnder %*% bUnder + hDiff * (t(X) %*% R)))
muDiff <- rnorm(1,bUpper[1], solve(hUpper)[1,1])
muJump <- rnorm(1,bUpper[2], solve(hUpper)[2,2])
bUnder <- c(muDiff, muJump)

# Generate proposal of 1/variance of diffusion process and run Metropolis-Hastings
for (j in 1:T) {
tmp = logReturn[i] - muDiff * delta
s2HDiff = s2HDiff + (tmp - muJump * Bernoulli[i] ) ^ 2 / (delta + hDiff /
hJump * Bernoulli[i] )
}
tmpV <- T + vDiff
tmpS <- s2Diff + s2HDiff
proposalH <- rchisq(1, tmpV ) / tmpS
prob <- 1
tmp <- sqrt(1 + (hDiff - proposalH )/(hJump*delta + proposalH))
for (j in 1:T) {
if (Bernoulli[i]==1) {
prob <- prob * tmp
}
}
if(runif(1) <= prob ){
hDiff = proposalH
}

# Generate proposal of 1/variance of jump process and run Metropolis-Hastings
for (j in 1:T) {
tmp = logReturn[i] - muDiff * delta
s2HJump = s2HJump + Bernoulli[i] * (tmp - muJump ) ^ 2 / (delta * hJump /
hDiff + 1)
}
tmpV <- sum(Bernoulli) + vJump
tmpS <- s2Jump + s2HJump
proposalH <- rchisq(1, tmpV ) / tmpS
prob = ((delta * proposalH + hDiff) / (delta * hJump + hDiff)) ^ (- sum(Bernoulli)
/ 2)
if(runif(1) <= prob ){
hJump = proposalH
}

xx[g,] = c(muDiff, muJump, intensity, hDiff, hJump)
}

# Print and plot results
print("Average values: (diff drift, jump drift, intensity, diff vol, jump vol)")
print(mean(xx[-(1:burnIn),1]))
print(1 / mean(xx[-(1:burnIn),4]))
print(1 / mean(xx[-(1:burnIn),5]))

print("Standard deviations: (diff drift, jump drift, intensity, diff vol, jump vol)")
print(sd(xx[-(1:burnIn),1]))
print(1 / sd(xx[-(1:burnIn),4]))
print(1 / sd(xx[-(1:burnIn),5]))

par(mfrow = c(5,2))

XVI
XVI
plot(density(xx[-(1:burnIn),1]), main="Density drift diffusion", xlab="")
plot(xx[,1], type="l", main="Trace plot drift diffusion", xlab="Iterations", ylab="")
plot(density(xx[-(1:burnIn),2]), main="Density drift jump", xlab="")
plot(xx[,2], type="l", main="Trace plot drift jump", xlab="Iterations", ylab="")
plot(density(xx[-(1:burnIn),3]), main="Density jump intensity", xlab="")
plot(xx[,3], type="l", main="Trace plot jump intensity", xlab="Iterations", ylab="")
plot(density(1/xx[-(1:burnIn),4]), main="Density drift vol", xlab="")
plot(1/xx[,4], type="l", main="Trace plot drift vol", xlab="Iterations", ylab="")
plot(density(1/xx[-(1:burnIn),5]), main="Density jump vol", xlab="")
plot(1/xx[,5], type="l", main="Trace plot jump vol", xlab="Iterations", ylab="")

ZHou Proof

Încărcat de

Informații document

Descriere originală:

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

ZHou Proof

Încărcat de

Drepturi de autor:

Formate disponibile

Department for

Industrial Economics and Technology Management

P making the markets for default-free and

P leading to greatly simplified valuation

. By doing so, can be interpreted as the growth rate of the

P the original SDE

of the bond. W(X

. Finally, we calculate the CDS

, displaying a large error

S-ar putea să vă placă și