You are on page 1of 166

UNIVERSITY OF CALIFORNIA

Santa Barbara

Interacting Particle Systems for Pricing Credit


Derivatives

A Dissertation submitted in partial satisfaction


of the requirements for the degree of

Doctor of Philosophy

in

Statistics and Applied Probability

by

Douglas W. Vestal

Committee in Charge:
Professor Jean-Pierre Fouque, Chair
Professor Guilluame Bonnet
Professor John Hsu

June 2008
3316500

3316500
2008
The Dissertation of
Douglas W. Vestal is approved:

Professor Guilluame Bonnet

Professor John Hsu

Professor Jean-Pierre Fouque, Committee Chairperson

June 2008
Interacting Particle Systems for Pricing Credit Derivatives

Copyright 
c 2008

by

Douglas W. Vestal

iii
To my grandfather, William Addison Vestal,

who taught me two great qualities:

Patience and Persistence.

iv
Acknowledgements

My graduate career has been a journey, both figuratively and literally. Be-

hind my dissertation is a long list of individuals whose support, advice, and

encouragement were critical. First and foremost, I would like to think my ad-

visor Dr. Jean-Pierre Fouque. It was Dr. Fouque’s Financial Mathematics class

at N.C. State that convinced me to study financial mathematics, a conversation

with him at Mitch’s Tavern in Raleigh, N.C., that convinced me to do a Ph.D.

with him, and his move to UCSB that has allowed me to experience the won-

derful city of Santa Barbara and of course, wine country. He has provided me

with many opportunities that I know I would not have experienced otherwise. I

am immensely grateful for his instruction and support throughout my graduate

years.

I would also like to thank Dr. Guilluame Bonnet and Dr. John Hsu for

serving on my committee and providing me with good advice. In addition, this

dissertation benefited from a special topics class I took with Dr. Bonnet, for

which I’m grateful.

I’d also like to thank our co-author, René Carmona for his feedback and

advice on our research topic. The participants of our NSF FRG research group,

namely Dr. Ronnie Sircar and Dr. Thaleia Zariphopoulou and graduate stu-

dents, provided a wonderful forum for which to present our research and obtain

v
an outside viewpoint. In the Fall of 2007, I had the wonderful opportunity

to visit Chuan-Hsiang (Sean) Han at the National Tsing Hua University in

Hsinchu, Taiwan for one month. This time was extremely valuable for further

exploring my research and my writing. The daily interaction I had with Dr.

Han was extremely beneficial and enlightening. In addition, I’m much appre-

ciative of his family’s hospitality while I was visiting. I’d also like to thank the

participants of the probability seminar at Academia Sinica in Taipei, Taiwan

for their support and feedback during my week-long stay. I hope to make it

back to Taiwan in the future. I really miss the night markets!

Along my academic career there have been many wonderful teachers and

professors that have influenced me and contributed to my desire to pursue the

mathematical sciences. At Apex High School, Mrs. Whitaker and Mr. Trezona

were both first-rate instructors and kept me challenged. At N.C. State, I had the

great fortune to take classes with Dr. Paur, Dr. Hartwig, Dr. Bob Martin, and

Dr. Silverstein. Special thanks to Dr. Bob Martin and Dr. Silverstein for their

instrumental role in getting me into graduate school. In addition, I benefited

greatly from the encouragement and conversations with Dr. Silverstein over

roasted chicken dinners and Film Noir at his house.

I would also like to thank the department of Statistics and Applied Proba-

bility at UCSB for their work in making my transfer as smooth as possible. In

vi
particular, Dr. Feldman, Dr. Hsu, Dr. Carter, Dr. Bonnet, and staff have been

wonderful.

Along the way I’ve had the great fortune to have a couple of internships. At

Progress Energy, I worked with some wonderful people, namely Yan Gao, John

Daily, Albert Hopping, and Leo Kang. At UBS Investment Bank, Peter Aerni,

Mark Nyfeler, Anders Wulf-Andersen, Oliver Rochet, and Alan Pitts provided

many interesting problems and questions about the practical implementation of

financial models.

Thanks also goes to Peter Cotton at Julius Finance who read an early draft

version of my dissertation and provided good advice.

My family has been very supportive throughout my undergraduate and grad-

uate career. My brothers, David and Dennis Vestal, and their families, have

provided delightful breaks from the stresses of graduate school. My mother and

stepfather, Marsha and Frank O’Neal, have often provided much needed finan-

cial and emotional support. I just stumbled upon the first budget that they sat

me down with to explain the value of hedging myself financially. It is advice

I still use to this day. Thanks also go to my father Randy Vestal who always

emphasized discipline and attention to detail.

Graduate school wouldn’t have been nearly as enjoyable without the creative

outlet of cooking. For this I thank Thomas Keller and his two cookbooks,

vii
The French Laundry Cookbook and Bouchon, which provided the source of

inspiration for never having to eat ramen noodles during graduate school. In

addition, daily walks of our dog Maple to the beach provided welcome relief

during the months I was sequestered at home furiously writing this dissertation.

Last but not least, I couldn’t have done this without the support of my won-

derful wife, Lindsey Vestal. She has been amazingly encouraging and supportive

of me and the different twists and turns that accompany graduate school and

finding a job. In addition, she has been a wonderful editor. She has always

provided the much needed balance and perspective after spending so much time

writing equations. I could not have picked a more wonderful person to share

my life with.

viii
Curriculum Vitæ
Douglas W. Vestal

Education

2005 Masters of Science, Financial Mathematics, North Carolina

State University.

2003 Bachelor of Science, Mathematics, North Carolina State Uni-

versity.

Experience

2007 Credit Risk Intern, UBS Investment Bank.

2006 Quantitative Analyst Intern, Progress Energy.

2005 – 2006 Graduate Fellow, SAMSI.

2004 – 2005 Technical Student, SAS Institute.

Selected Publications

D. Vestal, R. Carmona, and J.P. Fouque. Interacting Particle Systems for the
Computation of CDO Tranche Spreads with Rare Defaults. Submitted, Jan.
2008.

SAS
R
OpRisk VaR 2.5 User’s Guide, SAS Publishing, 2004.

ix
Abstract
Interacting Particle Systems for Pricing Credit Derivatives

Douglas W. Vestal

In this dissertation, we consider the numerical challenge of pricing credit

derivatives that depend on rare event probabilities. In particular, much of our

attention will be devoted to Collateralized Debt Obligations (CDOs). CDOs are

highly dimensional credit derivatives for which many computational challenges

exist. One of the main difficulties in pricing CDOs is the high-dimensional

nature of the problem. This fact usually precludes analytical solutions and so

numerical techniques, such as Monte Carlo, are employed. Due to the slow

converge of Monte Carlo techniques, variance reduction becomes critical and

importance sampling is the main variance reduction technique used. However,

the change of measure that is required to perform importance sampling can be

difficult to find and to simulate.

Therefore, we are interested in the intersection of two events; a complicated

model that doesn’t lend itself to importance sampling, and the computation

of rare probabilities under such a model. We develop the methodology to use

Feynman-Kac path measures and their interacting particle system interpreta-

tion to the field of credit derivatives pricing under the first passage model. Our

x
main result is the derivation of the asymptotic variance of the interacting par-

ticle system for a particular choice of weight function that naturally arises in a

credit derivatives pricing context. We show that the variance of using interact-

ing particle systems for this class of models is substantially less than traditional

MC. In addition, we derive many different probability densities on the path to

computing the variance. We also develop and show how to apply interacting

particle systems to obtain Vaue-at-Risk estimates and compute expected short-

fall. We also will present several numerical results that confirm our theoretical

results.

xi
Contents

Acknowledgements v

Curriculum Vitæ ix

Abstract x

List of Figures xiv

1 Introduction 1

2 Credit Derivatives Pricing Methods 8


2.1 Some Common Credit Derivatives . . . . . . . . . . . . . . . . . 9
2.1.1 Defaultable Bond . . . . . . . . . . . . . . . . . . . . . . 9
2.1.2 Credit Default Swap . . . . . . . . . . . . . . . . . . . . 13
2.1.3 Collateralized Debt Obligations . . . . . . . . . . . . . . 15
2.2 Structural-Based Models . . . . . . . . . . . . . . . . . . . . . . 26
2.3 Intensity-Based Models . . . . . . . . . . . . . . . . . . . . . . . 30
2.4 Copula-Based Models . . . . . . . . . . . . . . . . . . . . . . . . 36

3 Feynman-Kac Path Measures and Interacting Particle Systems 40


3.1 General Class of Feynman-Kac Path Measures . . . . . . . . . . 40
3.2 Useful Results of Feynman-Kac Path Measures . . . . . . . . . . 46
3.2.1 Note on Feynman-Kac Change of Measure . . . . . . . . 48
3.3 Interacting Particle System Interpretation of Feynman-Kac Models 51
3.3.1 Interacting Particle System Algorithm . . . . . . . . . . 55
3.4 Some Convergence Results for Interacting Particle Systems for
Rare Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

xii
4 Credit Derivatives Pricing with Interacting Particle Systems 65
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2.1 Multi-Name Model . . . . . . . . . . . . . . . . . . . . . 70
4.3 Feynman-Kac Path Measures and Interacting Particle Systems . 72
4.3.1 Interacting Particle System Interpretation and General
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.4 Pricing Using Interacting Particle Systems . . . . . . . . . . . . 78
4.4.1 Note on Choice of Weight Function . . . . . . . . . . . . 80
4.4.2 Algorithm on the Structural Model . . . . . . . . . . . . 82
4.4.3 Note on Algorithm for Pricing Collateralized Debt Obli-
gations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.5 Variance Analysis: Single-Name Case . . . . . . . . . . . . . . . 85

5 Numerical Implementation and Results 104


5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.2 Single Name . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.3 Multi-Name . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

6 Other Applications of Interacting Particle Systems 113


6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.2 Portfolio Value at Risk . . . . . . . . . . . . . . . . . . . . . . . 113
6.3 Review of Monte Carlo for VaR Calculation . . . . . . . . . . . 115
6.4 Interacting Particle System Interpretation of Value at Risk . . . 117
6.5 Expected Shortfall Calculation . . . . . . . . . . . . . . . . . . . 119
6.5.1 Interacting Particle Systems’ Answer to Expected Short-
fall Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.6.1 Further Analysis of Interacting Particle Systems . . . . . 125
6.6.2 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.7 Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . 128

7 Conclusion 133

A Proofs 135

Bibliography 148

xiii
List of Figures

2.1 An example of a credit spread curve with parameters S0 = 90,


K = 50, r = .05, σ = .25, η = 0 . . . . . . . . . . . . . . . . . . . . . 30

5.1 Default probabilities for different barrier levels for IPS and MC 106
5.2 Variances for different barrier levels for IPS and MC . . . . . . . 107
5.3 Standard Deviation to Probability ratio for MC and IPS . . . . 109
5.4 Probability mass function of the loss function described in equa-
tion (6.3) with N = 25 . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.5 Probability mass function of the loss function described in equa-
tion (6.3) with N = 25 in log scale . . . . . . . . . . . . . . . . . . . 112

6.1 Estimated Expected Shortfall Contributions . . . . . . . . . . . 130


6.2 Estimated Expected Shortfall Contributions with Rare Proba-
bilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

xiv
Chapter 1

Introduction

The field of option pricing theory can be traced back to the early 1900’s.

Louis Bachelier, often called the father of mathematical finance, defended his

thesis Théorie de la Spéculation on March 29th, 1900. His thesis started the pro-

cess of quantitatively describing the financial markets by beginning the theory

of Brownian motion and applying it to option pricing for the French financial

markets [8]. Among other things, Bachelier’s thesis postulated two critical in-

gredients that are still used today. The first is that asset prices should be valued

under a martingale measure and the second is that the price evolves as a Markov

process.

The past few years have seen an explosive growth in the credit derivatives

market. Meanwhile, the field of credit risk and credit derivatives research has

substantially increased. As the credit derivative products have grown in com-

1
Chapter 1. Introduction

plexity, so has the need for fast and accurate numerical methods to accurately

price such derivatives.

In this dissertation, we consider the pricing of credit derivatives, with an eye

toward Collateralized Debt Obligations (CDOs), under the first passage model

where default is a rare event. A CDO is a credit derivative that pools together

many different firms (125 comprises a typical contract) and sells exposure to

different default levels of the portfolio; the so-called tranches. This segmentation

of the risk in a portfolio enables the buyer to purchase only the tranches that

they deem appropriate for their hedging positions. Since the tranche levels are

fairly standardized, there are also new products called bespoke CDOs that sell

customized default levels.

The main difficulty in pricing CDOs is the high-dimensional nature of the

problem. To accurately price the CDO, the distribution of the joint default for

many names is needed. Even if the distribution of joint defaults is found explic-

itly, there is no guarantee that the expectations that are then needed to compute

tranche spreads can be found analytically. Therefore, the user has to rely on

numerical schemes to compute the CDO prices. Due to the high-dimensional

nature of the problem, PDE-based methods are ruled out and Monte Carlo

(MC) methods are heavily relied upon. While MC methods are easy to imple-

ment in high-dimensional problems, they do suffer from slow convergence. In

2
Chapter 1. Introduction

addition, due to the rare nature of joint defaults, the computational problem is

exacerbated since many MC simulations are needed to observe the joint default

of multiple names. Therefore, variance reduction and efficiency become very

important for these MC methods.

The main variance reduction technique used is importance sampling. There

have been many successful applications of importance sampling towards credit

risk [2, 20]. However, most authors have concentrated on multifactor Gaussian

copula models or the reduced-form model for credit risk. However, we have

not found any reference in the literature that applies importance sampling to

the first passage model. In addition, the difficulty with implementing impor-

tance sampling is computing the change of measure under which one simulates

the random variables. More information about importance sampling and the

difficulty required can be found in [18] and the references therein.

Therefore, in this dissertation, we are concerned with first passage models

where it is impractical, if not impossible, to compute the importance sampling

change of measure explicitly. Examples we have in mind are first passage models

with stochastic volatility and/or regime switching models with many factors.

In this situation, our solution is to use an Interacting Particle System (IPS)

technique that can briefly be described as applying importance sampling tech-

niques in path space. That is, we never compute a change of measure explicitly

3
Chapter 1. Introduction

but rather decide on a judicious choice of a potential function and a selec-

tion/mutation algorithm under the original measure such that the measure in

path space converges to the desired “twisted” measure. This “twisted” measure

will be exactly the change of measure that one would use in an importance

sampling scheme. The advantage is that the random variables are always simu-

lated under the original measure, but through the selection/mutation algorithm

converge to the desired importance sampling measure. Many of the the IPS

techniques and theoretical support we use to design our algorithm can be found

in [11, 12].

In this dissertation, we are interested in the intersection of two events; a

complicated model that doesn’t lend itself to importance sampling, and the

computation of rare probabilities under such a model. We develop the method-

ology to use Feynman-Kac path measures and their IPS interpretation to the

field of credit derivatives pricing under the first passage model. Our main result

is the derivation of the asymptotic variance of the interacting particle system for

a particular choice of weight function that naturally arises in a credit derivatives

pricing context. We show that the variance of using interacting particle systems

for this class of models is substantially less than traditional MC. In addition,

we derive many different probability densities on the path to computing the

variance. We also develop and show how to apply interacting particle systems

4
Chapter 1. Introduction

to obtain Vaue-at-Risk estimates and compute expected shortfall. We also will

present several numerical results that confirm our theoretical results.

The rest of this dissertation is organized as follows:

Chapter 2: Credit Derivatives Pricing Methods. This chapter in-

troduces the basics of pricing credit derivatives and provides a review of three

separate credit derivatives. The first is a defaultable bond, then a credit default

swap, and finally Collateralized Debt Obligations (CDOs). By analyzing the

cash flows, the no-arbitrage price of the derivatives are presented in a general

framework without assuming the explicit dynamics of the default times. After

presenting the credit derivatives, we focus our attention on three different clas-

sical pricing methods. First we present the structural, or first passage model,

as this will be the pricing formulation that is used throughout the dissertation.

We end this chapter by briefly presenting the intensity-based model and the

copula approach.

Chapter 3: Feynman-Kac Path Measures and Interacting Parti-

cle Systems. This chapter leaves finance behind and introduces the reader to

Feynman-Kac path measures. A general class of Feynman-Kac path measures

are introduced in the non-homogeneous Markov chain framework. This class

is analyzed and properties are explored that will be significant for developing

5
Chapter 1. Introduction

the results of the dissertation. We then introduce the IPS interpretation of the

Feynman-Kac models and present a general algorithm for changing the mea-

sure in path space. That is, by implementing a selection and mutation scheme

inside a standard MC program, we show how the Feynman-Kac model can be

interpreted as a change of measure in path space. Finally, we present some of

the theoretical results of interacting particle systems applied to calculating rare

event probabilities.

Chapter 4: Credit Derivatives Pricing with Interacting Particle

Systems. This chapter lies at the intersection of the two previous chapters;

the pricing of credit derivatives using interacting particle systems. We introduce

the model we use for pricing multi-dimensional credit derivatives. In addition,

we formulate the CDO pricing problem in the structural model under a homo-

geneous portfolio as reducing to computing a probability density function of

the losses the portfolio experiences. We formulate this rare event probability

estimation in terms of an interacting particle system with a specific choice of a

weight function. We then formulate and compare the theoretical variance of our

IPS estimator to traditional MC. Our main theorem and results are presented

and proved.

Chapter 5: Numerical Implementation and Results. This chapter

presents the numerical results for applying our interacting particle system to

6
Chapter 1. Introduction

pricing CDOs. In addition, we compare the empirical variance between IPS and

traditional MC. We demonstrate empirically the variance reduction achieved by

applying our IPS formulation to the first passage model.

Chapter 6: Other Applications of Interacting Particle Systems.

We present some other applications for using interacting particle systems in

finance. Specifically, we develop the methodology for using IPS for assessing

portfolio Value-at-Risk (VaR). In addition, we develop and present an algorithm

for using IPS to compute expected shortfall for a portfolio of credit risky assets.

We compare the IPS method for computed expected shortfall to traditional MC

and find substantial computational accuracy on a toy portfolio.

Chapter 7: Conclusion. In this chapter, we present the conclusions of

the dissertation. In addition, we present some topics for future research studies.

7
Chapter 2

Credit Derivatives Pricing


Methods

The past few years have seen an explosion in the field of credit derivatives.

According to the International Swaps and Derivatives Association (ISDA), in

the first six months of 2007, the notional amount underlying credit derivatives

grew from $34.42 trillion to $45.46 trillion; a 32% increase over 2006. This

increase follows on the heels of a 75% increase in 2006 as well. In addition, the

ISDA’s definition (for the sake of the survey) of credit derivatives only includes

credit default swaps on single names, baskets, and portfolios and as such only

underestimates the true notional amount of credit derivatives transacted.

Credit derivatives enable companies to trade the default risk of different

companies. In short, credit derivatives provide a way for companies to hedge

the default risk of their assets. Typically, when discussing the default risk of a

company, one thinks of a company not repaying its loan. Through no arbitrage

8
Chapter 2. Credit Derivatives Pricing Methods

arguments, the price of this risk must be factored into the price of the bonds

issued by the company. That is, the holder of the defaultable instrument must

be compensated for bearing the risk of default.

The fundamental building block for pricing credit derivatives is to have a

way to quantify the time of default for a company. While there are many differ-

ent legal definitions of default, we mention that they can include not repaying

the notional, missing a coupon payment, bankruptcy, restructuring, etc. [25].

For the remainder of this dissertation, we take the more simple view that default

happens when a company doesn’t repay it’s obligation. While the term obliga-

tion covers defaultable bond payments, it also covers any other asset that the

company owes money on, such as a deep in the money interest rate derivative.

2.1 Some Common Credit Derivatives

2.1.1 Defaultable Bond

The ability to price non-coupon bearing defaultable bonds is one of the

crucial components of building more complex credit derivative products. We as-

sume that all pricing occurs under a risk-neutral probability measure (Ω, F , P, (Ft )0≤t<∞ )

satisfying the usual conditions. In addition, we recall the definition of a stopping

time.

9
Chapter 2. Credit Derivatives Pricing Methods

Definition 2.1.1. A random time τ is a stopping time of the filtration, if the

event {τ ≤ t} belongs to the σ-field Ft , for every t ≥ 0.

Assuming the risk-free interest rate, r, is constant, then the price of a zero-

coupon non-defaultable unit notional bond at time t, maturing at time T is

given by

Γ(t, T ) = e−r(T −t)

This is the price of a contract, at time t, that pays $1 at maturity time T

when there is no chance of default.

To price defaultable bonds, one naturally needs to have a model for how a

company defaults. In the Sections 2.2-2.4, we detail some of the different ways

in which one can model how a company will default. For the moment, we just

assume that the time of default for a company is given by a stopping time τ .

That is, according to the above definition, τ is a random time representing the

default time of the company.

Naturally, the payout of the defaultable bond at time T is the following





⎨1 if τ > T ,
f (T ) =



⎩R if τ ≤ T .

10
Chapter 2. Credit Derivatives Pricing Methods

where R is called the recovery rate. That is, R is the amount that the bond

holder will receive if the company defaults before time T . We assume that R

is paid at time T . Because of debt holder negotiations and restructuring in

the event of default, most bond holders can expect to receive some nonzero

amount if the company does default. Typically, R is assumed to be anywhere

from 0 − 40% of the face value of the bond. However, there is a whole separate

literature that deals with modeling the recovery rate as a stochastic process

itself. See [1] for a fairly complete review of the different methods for recovery

rate modeling.

Through no arbitrage arguments, the price of a zero-coupon defaultable bond

at time t, maturing at time T is given by

Γ̄(t, T ) = E[e−r(T −t) f (T )|Ft ]

= e−r(T −t) 1τ >t E[1τ >T + R1t≤τ ≤T |Ft ]

= (Γ(t, T )E[1τ >T |Ft ] + Γ(t, T )RE[1t≤τ ≤T |Ft ])1τ >t

Therefore, at time 0 the price of the defaultable bond is

11
Chapter 2. Credit Derivatives Pricing Methods

Γ̄(0, T ) = Γ(0, T )P(τ > T ) + Γ(0, T )RP(τ ≤ T )

= Γ(0, T ) (R + P(τ > T )(1 − R))

The yield spread on defaultable bonds is an important quantity in credit

derivatives pricing. The yield spread measures the rate of return, over the fixed

risk-free rate, that the defaultable bond earns. This is interpreted to be the

amount of money that bond holders are compensated for bearing the risk of

default.

The yield spread of the defaultable bond is defined as


 
Γ̄(t,T )
ln Γ(t,T )
Y (t, T ) = −
T −t

Empirical evidence shows that as the time to maturity decreases, the yield

spread converges to a non zero value [23]. That is,

lim Y (0, T ) → c > 0


T →0

While there are many empirical reasons for this fact, Duffie and Lando at-

tribute it to the investor having incomplete information. Therefore, even for

very short maturities, they want to be compensated for the risk of holding the

bond [14].

12
Chapter 2. Credit Derivatives Pricing Methods

2.1.2 Credit Default Swap

The most heavily traded credit derivative is the credit default swap (CDS).

Here we will only consider a single name CDS. Essentially, a single name CDS

allows one party to trade the default risk of holding defaultable bonds to a

second party. That is, party A holds defaultable bonds issued by company

C. A would like to offset some of the risk of holding the defaultable bonds by

buying default protection on company C from party B. In the event of default,

party B would pay party A the amount of money that is lost (that is 1 − R

in the set up above). If company C doesn’t default, then party B doesn’t owe

party A anything.

Obviously, party A has to pay party B for this default protection. The

default protection can be thought of as an insurance premium and the amount

of money that party A pays to party B is called the credit default swap spread.

The goal is to figure out what the fair price of the CDS spread should be.

While there are many different formulations for the payoffs for CDS’, for the

ease of exposition and to introduce the flavor of CDS’, we will take a simple

view. The first thing to do is to look at the cash payments for both party A

and party B. For simplicity we assume that the notional amount is equal to $1.

13
Chapter 2. Credit Derivatives Pricing Methods

We assume that the two parties have agreed that T1 , . . . , Tn are the payment

dates that party A will pay the spread r  to party B. Once default happens,

party A stops paying the spread r  . This is called the “premium” payment leg.

For the “protection” leg, party B pays party A the amount LGD = 1 − R

when default happens. At time 0, the value of the payments that party A makes

is


n
VA (r  ) = r  Γ(0, Ti)1τ >Ti
i=1

Since party B only makes a payment at the default time if it occurs before

time Tn , the value of its payment is

VB = LGD Γ(0, τ )1τ ≤Tn



n
= LGD Γ(0, τ )1Ti−1 <τ ≤Ti
i=1

Through no arbitrage arguments, the spread r  is the value that makes the

above contract fair. That is, we have to solve for r  such that

E[VA (r  )] = E[VB ]

Hence, taking expectations and doing a little algebra, we arrive at

14
Chapter 2. Credit Derivatives Pricing Methods

n
E[e−rτ 1Ti−1 <τ ≤Ti ]
r = LGD ni=1


i=1 Γ(0, Ti )P(τ > Ti )

If we assume that the protection payment happens the first date Ti after

default, then the above simplifies to

n
Γ(0, Ti )P(Ti−1 < τ ≤ Ti )

r = LGD n
i=1

i=1 Γ(0, Ti )P(τ > Ti )

A general, fairly non-technical, overview of credit default swap mechanics

is given in [13]. In [21], the authors present a two-step procedure for valuing

credit default swaps. The method involves estimating the probability of default

from yields on corporate bonds and then applying those probabilities to value

the CDS by taking an expectation of the cash flows. This paper was extended

in [22] to include the possibility of counterparty default and develops a model

to include correlated defaults among names.

2.1.3 Collateralized Debt Obligations

The previous section reviewed the basics of CDS’ for single names. The

single name CDS allows a party to transfer the default risk they are exposed

to for a single name. However, often times parties have an exposure to many

different names and want to package this risk together. If a company has a

large portfolio comprising of many names, they might not be concerned about

15
Chapter 2. Credit Derivatives Pricing Methods

the individual default of each name, but rather that a small collection of names

in the portfolio default. In many cases, it wouldn’t be financially prudent to buy

protection on each single name in the portfolio. But, if the company could buy

protection on just a slice of the portfolio, then they might be properly hedged

for their own internal risk management purposes.

One solution that is becoming more and more popular is the Collateralized

Debt Obligation (CDO). In a CDO, the credit risk associated with a portfolio

is broken up and sold in different tranches. A tranche is simply a proportion

of the portfolio amount. Each tranche is characterized by its upper and lower

attachment point. The following table shows a typical structure of a CDO and

its tranches.
Attachment points(%)
Tranche number Tranche name Lower KL Upper KU
1 Equity 0 3
2 Mezzanine 1 3 7
3 Mezzanine 2 7 10
4 Mezzanine 3 10 15
5 Senior 15 30

The rest of this section borrows directly from [15] which provides a very

comprehensive overview of CDO mechanics. We adopt the same notation and

for completeness, present many of the same things.

The protection in the event of a default goes in order of the tranches. That

is, if a total of 6% of the value of the portfolio defaults, then the holder of

16
Chapter 2. Credit Derivatives Pricing Methods

the Equity tranche (0 − 3%) pays the first 3% of the losses and the holder of

the first Mezzanine tranche (3 − 7%) pays the next 3 to 6% of the losses. The

other tranches are not touched. In this way, the holder of tranche [KL , KU ]

is only subject to paying for (KU − KL ) percentage of losses if the amount of

losses exceed KL percent. Therefore, each protection seller is only liable for a

maximum of (KU − KL ) percent of losses over the life of the portfolio.

We present a common structure for a CDO. Like a CDS, a CDO has two

payment legs: a fixed leg that the tranche holders receive for selling protection

and a floating leg the tranche holders pay if the losses hit that tranche. The

fixed leg is known as the tranche spread and is represented by sj for the j

different tranches in a CDO.

In other words, a CDO is typically structured as follows. The holders of the

tranche receive payments at a given frequency during the life of the CDO. Once

the tranche that they hold is hit with a loss, they have to pay the amount of

the loss to the protection buyer at the next payment date. The holder of the

tranche stops receiving the protection premium payments once the amount of

losses exceed KU for that tranche.

The mathematical details of a CDO structure are as follows. We assume

that the CDO has payment dates {t1 , . . . , tK } where tK is the maturity and

η = tk+1 − tk ; η = .25, corresponding to quarterly payments is the norm. In

17
Chapter 2. Credit Derivatives Pricing Methods

addition, suppose that the face value of the assets underlying the CDO at time

t0 = 0 is M.

Let t denote the time that has passed since the CDO was sold. In addition,

we denote by Zt the fraction of losses (of M) that have occurred up until time

t. Therefore, the total loss to the CDO is given by MZt . We denote by Zj,t

the percentage of losses that has struck the j th tranche up until time t. Recall

that the j th tranche is only responsible for losses that exceed KLj percent of the

notional but are less than KUj percent of the notional. In fact, we can write

Zj,t as

Zj,t = min(Zt , KUj ) − min(Zt , KLj )

Since min(Zt , KUj ) = KUj + min(Zt − KUj , 0), the above equation can be

written as

Zj,t = (KUj − KLj ) + (min(Zt − KUj , 0) − min(Zt − KLj , 0))

Therefore, at time t, the amount of loss that the j th tranche has suffered is

MZj,t .

Now, if a loss occurs between two payment dates, i.e. tk < t ≤ tk+1 , then at

time tk+1 the holder of the tranches will have to pay the fraction Zj,tk+1 −Zj,tk of

the notional M to the protection buyers. The protection buyer keeps paying the

18
Chapter 2. Credit Derivatives Pricing Methods

spread sj to the protection seller until the j th tranche has suffered its maximum

loss.

However, the actual amount of money that the protection buyer pays doesn’t

remain constant because while he/she pays a constant spread sj , the amount

of money that needs protecting is decreasing as losses are suffered. That is,

the holder of the tranche receive the periodic payment of sj of the outstanding

amount of tranche j. At time t, the outstanding notional of tranche j is given by

its initial amount ((KUj − KLj )M) minus the amount of losses that the tranche

has suffered up to time t. That is, representing Φj,t as the amount of notional

corresponding to tranche j at time t, we see that

Φj,t = (KUj − KLj )M − Zj,tM

= (KUj − KLj − Zj,t )M

19
Chapter 2. Credit Derivatives Pricing Methods

By using the definition of Zj,t , we see that Φj,t simplifies to

Φj,t = (min(Zt − KLj , 0) − min(Zt − KUj , 0))M






⎪ (KUj − KLj )M, if Zt < KLj ,





= (KUj − Zt )M, if KLj ≤ Zt ≤ KUj .








⎩0, if Z > K t Uj

We are now in a position to state exactly what the cash flows will be to each

tranche holder at each payment date.

Recall that the CDO is struck at time t0 = 0 and that {t1 , . . . , tK } are the

agreed upon payment dates. For simplicity we have assumed that η = ti+1 − ti ,

∀ti ∈ {t1 , . . . , tK } is a constant. Then, at each payment date, t ∈ {t1 , . . . , tK },

the holder of tranche j has the following cash flow

• They receive the amount

sj ηΦj,t

and

• They pay the amount

(Zj,t − Zj,t−η )M

It should be emphasized that the spread, sj , they receive doesn’t change

over the life of the CDO. However, the net amount they receive changes based

20
Chapter 2. Credit Derivatives Pricing Methods

on how Φj,t changes. It is easy to see that Φj,t is a decreasing function in Zt M.

That is, as the notional amount of losses increases, the effective amount that

tranche holder j receives decreases. Intuitively, this is reasonable because the

protection buyer is paying to protect his/her losses. If some losses have already

occurred, he/she shouldn’t have to pay as much for protection since the amount

that needs protection is lower. This can be seen mathematically by taking the

partial of Φj,t with respect to Zt M:








0, if Zt < KLj ,


∂Φj,t ⎨
= −1, if KL ≤ Zt ≤ KU ,
∂Zt M ⎪
⎪ j j






⎩0, if Zt > KUj ,

CDOs are similar to CDS’ in the sense that neither require an initial invest-

ment and each has two payment legs. CDO’s two payment legs are the fixed

and floating leg. The fixed leg is given by the spread the tranche holders receive

and the floating leg is given by the amount of loss protection they have to pay.

The purpose of pricing a CDO is to find the tranche spread sj that makes

the CDO fair. That is, we need to find the spread sj that makes the discounted

expected value of the two payment streams equal at time t0 = 0.

Taking expectations and discounting the cash flows back to time t0 = 0, we

see that the value of the fixed payment leg is given by

21
Chapter 2. Credit Derivatives Pricing Methods


K
Vf ixed = Γ(0, ti )sj ηE[Φj,ti ]
i=1
K
= Γ(0, ti )sj ηME[(KUj − KLj − Zj,ti )]
i=1

K


= sj ηM Γ(0, ti) E[min(Zti − KLj , 0)] − E[min(Zti − KUj , 0)]
i=1

Again, taking expectations and discounting the cash flows back to time

t0 = 0, we see that the value of the floating payment leg is given by


K
Vf loating = Γ(0, ti )E[(Zj,ti − Zj,ti−1 )M]
i=1

K


= M Γ(0, ti) E[min(Zti − KUj , 0)] − E[min(Zti−1 − KUj , 0)]
i=1
K


+M Γ(0, ti) E[min(Zti−1 − KLj , 0)] − min(Zti − KLj , 0)]
i=1

Therefore, setting Vf ixed = Vf loating and solving for sj , we find that

K

Γ(0, ti ) E[min(Z ti
− K U j
, 0)] − E[min(Z ti−1
− K U j
, 0)]
sj = i=1
K

η i=1 Γ(0, ti ) E[min(Zti − KLj , 0)] − E[min(Zti − KUj , 0)]

K

i=1 Γ(0, ti ) E[min(Z ti−1
− K Lj
, 0)] − E[min(Z ti
− K Lj
, 0)]
+

η Ki=1 Γ(0, t i ) E[min(Z ti − K Lj , 0)] − E[min(Z ti − K Uj , 0)]

22
Chapter 2. Credit Derivatives Pricing Methods

The key component of the above expression is that it relies on being able to

compute

E[min(Zt − Kj , 0)]

for all t ∈ {t1 , . . . , tK } and each of the attachment points Kj . Recall that

Zt represents the fraction of losses that have occurred up until time t.

Note on Homogeneous Portfolio

So far, we haven’t explicitly described the portfolio underlying the CDO and

as such, haven’t described Zt in much detail. There are many different types

of CDOs traded and one way of distinguishing them is by the assets that are

in the portfolio. A cash CDO is when the underlying portfolio is comprised of

credits, loans, or debt instruments. A synthetic CDO is when the underlying

portfolio is composed of CDS’.

Suppose we consider a CDO where there are N different names underlying

the portfolio. We assume that each of the N names has a default time τi , 1 ≤

i ≤ N. Note that the time of default among the names can and will usually be

correlated.

For simplicity, we assume that each of the individual names underlying the

portfolio are homogeneous. That is, the losses given default are the same for

23
Chapter 2. Credit Derivatives Pricing Methods

each name in the portfolio. Let c denote the amount of loss that occurs when

a firm defaults. Let M = Nc be the total value of the portfolio at time t0 = 0.

Therefore, we can write the fraction of losses experienced up to time t, Zt , as

N
i=1 c1τi ≤t
Zt =
M
1
= L(t)
N

where


N
L(t) = 1τi ≤t
i=1

is the number of firms that have defaulted before time t.

Recall that the spread sj depends on being able to compute

E[min(Zt − Kj , 0)]

We can write the above expression as,

1
E[min(Zt − Kj , 0)] = E[min( L(t) − Kj , 0)]
N
1
= E[min(L(t) − NKj , 0)]
N

24
Chapter 2. Credit Derivatives Pricing Methods

Therefore, the spread sj depends explicitly on the joint distribution of de-

faults given by L(t). L(t) is a discrete probability distribution taking values in

the set {0, 1, . . . , N}. Therefore, we can write the expectation of interest as

1
E[min(Zt − Kj , 0)] = E[min(L(t) − NKj , 0)]
N
1 
N
= P(L(t) = i) min(i − NKj , 0)]
N i=0

1 
N
= P(L(t) = i)(i − NKj )
N i:i>N Kj

where P is the risk-neutral probability measure of the joint default of the firms

in the portfolio. Foreshadowing later chapters, one can see that computing the

quantities P(L(t) = i), either explicitly or numerically, are critical to pricing

CDOs.

So far, we have introduced a few common credit derivative products in a

general setting. Each credit derivative price, or spread, depends on the mod-

eling of the default time τ . In general, there are three different approaches to

modeling the default time. The first approach, called the structural model, links

the default time to the behavior of the company’s assets in relation to its debt.

The second approach, the intensity based model, models the default time as

25
Chapter 2. Credit Derivatives Pricing Methods

the first jump time of a stochastic Poisson process. The third approach, copula

models, attempt to emphasize the effect of correlation.

The next few sections introduce these different approaches to pricing credit

derivatives. Although later chapters will focus primarily on structural based

approaches, for completeness, we present all three methodologies.

2.2 Structural-Based Models

Default in the firm value model has a very nice economical appeal. The firm

value approach, or structural model, models the total value of the firm that has

issued the defaultable bonds. Typically, the value of the firm includes the value

of equity (shares) and the debt (bonds) of the firm [27]. There are two main

approaches to modeling default in the firm value approach: one is that default

can only happen at the maturity of the bond and the second is that default can

occur any time before maturity.

We follow both [26, 3] and assume that the value of the firm, St , follows

geometric Brownian motion. However, more generally we can assume that the

value of the firm follows any diffusion process. We also assume that interest

rates are constant. Under the risk-neutral probability measure P, we have,

dS(t) = rS(t)dt + σS(t)dW (t) (2.1)

26
Chapter 2. Credit Derivatives Pricing Methods

where r is the risk-free interest rate, and σ is constant volatility. At any time

t ≤ T , the price of the nondefaultable bond is Γ(t, T ) = e−r(T −t) .

We also assume that at time 0, the firm issued non-coupon corporate bonds

expiring at time T . The price of the defaultable bond at time t ≤ T is denoted

Γ̄(t, T ).

In [26] default was assumed to only occur at the expiration date T . Fur-

thermore, default at time T is triggered if the value of the firm is below some

default threshold B; that is if ST ≤ B.

Therefore, assuming zero recovery, the price of the defaultable bond satisfies,


Γ̄(t, T ) = E e−r(T −t) 1ST >B |St

= Γ(t, T )P(ST > B|St )

= Γ(t, T )N(d2 )

where N(·) is the standard cumulative normal distribution function and,



St
ln + (r − 12 σ 2 )(T − t)
d2 = B

σ T −t

The next model, the Black-Cox model, is also known as the first passage

approach. Developed in [3], default can occur anytime before the expiration of

the bond and the barrier level, B(t), is some deterministic function of time. In

27
Chapter 2. Credit Derivatives Pricing Methods

[3], they assume that the default barrier is given by the function B(t) = Keηt

(exponentially increasing barrier) with K > 0 and η ≥ 0. The default time τ is

defined by

τ = inf {t : St ≤ B(t)}

That is, default happens the first time the value of the firm passes below the

default barrier. This has the very appealing economical interpretation that

default occurs when the value of the firm falls below its debt level, thereby not

allowing the firm to pay off its debt and honor its obligations.

Assuming zero recovery, we can price the defaultable bond by pricing a

barrier option. Therefore,


Γ̄(t, T ) = 1τ >t E e−r(T −t) 1τ >T |St

= 1τ >t Γ(t, T )P(τ > T |St )


  p 
St −
= 1τ >t Γ(t, T ) N(d2 ) −
+
N(d2 )
B(t)

where
 
± ln + (r − η − 12 σ 2 )(T − t)
St
B(t)

2 = √
σ T −t
2(r − η)
p = 1−
σ2

where we used a Girsanov change of measure above and results for the run-

ning minimum for Brownian motion.

28
Chapter 2. Credit Derivatives Pricing Methods

We can also compute the cumulative distribution function for τ . In fact,

  p 
S0
P(τ ≤ t) = 1 − N(d+
2) − N(d−
2)
B(0)

The yield spread of the defaultable bond is defined as


 
Γ̄(t,T )
ln Γ(t,T )
Y (t, T ) = −
T −t

We remark that in the first passage model above, and all firm value models,

the yield spread for very short maturities goes to 0 in contrast to the empirical

evidence [16]. This is because the default time τ defined above is a predictable

stopping time. However, by incorporating a fast mean-reverting stochastic pro-

cess for σ, [17] were able to raise the yield spread for short maturities.

The graph below shows an example of a credit spread curve in the structural

model:

29
Chapter 2. Credit Derivatives Pricing Methods

Credit Spread: Structural Model


6

4
Spead (in percent)

0
0 1 2 3 4 5 6
Time to Maturity (years)

Figure 2.1: An example of a credit spread curve with parameters S0 = 90,


K = 50, r = .05, σ = .25, η = 0

2.3 Intensity-Based Models

In contrast to the structural model, the intensity-based model doesn’t have

a clear economic interpretation of default. In addition, the intensity approach

requires more a mathematical set up than the structural based model that only

relies on diffusion theory.

The mathematical background needed for intensity-based modeling is be-

yond the scope of this dissertation. We encourage the reader to look at [27]

for many of the building blocks needed. Though we don’t include the basics of

30
Chapter 2. Credit Derivatives Pricing Methods

intensity based modeling, we do present some of the most important concepts.

For the purpose of this section, we will only present Cox processes.

Before introducing Cox processes, we recall the following definition,

Definition 2.3.1. An inhomogeneous Poisson process with intensity function

λ(t) > 0 is a non-decreasing, integer-valued process with initial value N(0) = 0

whose increments are independent and satisfy,

 T n   T 
1
P(N(T ) − N(t) = n) = λ(s)ds exp − λ(s)ds
n! t t

The intensity λ(t) is a non-negative function of time only.

The process N(t) can be considered a counting process that starts at zero.

In fact, default in the intensity-based framework is defined by the first jump

time of the process N(t). That is,

τ = inf{t : N(t) > 0}

Therefore, to price defaultable bonds in the intensity based model, we have

to calculate

31
Chapter 2. Credit Derivatives Pricing Methods

Γ̄(0, T ) = E[e−rT 1τ >T ]

= e−rT E[1N (T )=0 ]

= Γ(0, T )P(N(T ) = 0)
RT
= Γ(0, T )e− 0
λ(s)ds

Since λ(s) is deterministic, we can fit exactly the initial term structure of

defaultable bond prices. This is done by assuming λ(s) is a piecewise constant

function. However, the inhomogeneous Poisson framework doesn’t solve every

problem. Since λ(s) is deterministic, we can’t correlate the spread with other

random variables, such as interest rates or stochastic volatility. Therefore, most

attention has been paid to developing stochastic intensity rate models. This

leads us to Cox processes.

Essentially, a Cox process is a Poisson process with stochastic intensity. That

is, we assume that the process λ(t) follows the diffusion,

dλ(t) = μ(t, λ(t))dt + σ(t, λ(t))dW (t)

where the drift and volatility processes above are chosen such that λ(t) is

always non-negative. We then have the following definition of a Cox process.

32
Chapter 2. Credit Derivatives Pricing Methods

Definition 2.3.2. A point process N(t) with intensity process λ(t) is a Cox pro-

cess if, conditional on the background information F , N(t) is a inhomogeneous

Poisson process with intensity λ(t).

This means that to compute things with this new process N(t), we will need

to use conditional expectations and the results presented above for inhomoge-

neous Poisson processes. So for Cox processes we have,

P(N(T ) − N(t) = n) = E[P(N(T ) − N(t) = n|F )]

= E[P(N(T ) − N(t) = n|λ)]


  T n   T 
1
= E λ(s)ds exp − λ(s)ds
n! t t

Therefore, to price defaultable bonds with stochastic intensity, we have

Γ̄(0, T ) = E[e−rT 1τ >T ]

= e−rT E[1N (T )=0 ]

= Γ(0, T )P(N(T ) = 0)
RT
= Γ(0, T )E[e− 0 λ(s)ds
]

33
Chapter 2. Credit Derivatives Pricing Methods

One important item to discuss is the numerical simulation of the default

times. That is, being able to simulate

τ = inf{t : N(t) > 0}

will allow us to price many different expectations that are functions of the

default time. In fact, it is very straightforward to simulate different default

times by the following algorithm.

1. Draw N independent U(0, 1) random variables.

t
2. Simulate the values 0
λi (s)ds, where λi (s) denotes N iid copies of the

original process λ.

3. Define the N default times to be

Rt
τi = min{t > 0 : e− 0 λi (s)ds
≤ Ui }

Then, we have set of N iid default times (τi )1≤i≤N that can be used in a MC

routine for pricing.

In fact, it is straightforward to see that the simulated default time, τi , comes

from the correct distribution by the calculation

34
Chapter 2. Credit Derivatives Pricing Methods

RT
P(τi ≤ T ) = P(e− 0
λi (s)ds
≤ Ui )
RT
= E[P(e− 0
λi (s)ds
≤ Ui |Ui )]
RT
= E[e− 0 λi (s)ds
]
RT
= E[e− 0
λ(s)ds
]

The type of diffusion λ(t) to use has been left intentionally vague. From the

definition, one only has to specify a non-negative diffusion process. However, it

is clear that not every non-negative diffusion process would give rise to behavior

that is exhibited in the credit market. As such, there is a great deal of research

on specifying different models for λ(t) that try to capture the observed market

conditions.

For example, one such model is the Cox-Ingersoll-Ross (CIR) model that

stipulates the following dynamics for λ(t):


dλ(t) = κ(θ − λ(t))dt + σ λ(t)dWt

The CIR model is mean reverting with θ being the mean-level and κ the

rate of mean-reversion. The λ(t) ensures that the intensity process stays

non-negative. This model was originally described in [9].

35
Chapter 2. Credit Derivatives Pricing Methods

2.4 Copula-Based Models

An extremely important component of credit derivatives pricing is the ability

to create correlation between default times. Correlation plays an important role

in capturing observed market behavior and the ability to price multi-dimensional

credit derivatives under the dependence of defaults assumption. There are many

ways to create correlation. In both the structural and intensity-based models,

one can create correlation by correlating the driving Brownian motions between

the asset prices or intensity dynamics, respectively. However, there is another

way that is extremely popular in credit derivatives pricing: the copula.

At it’s foundation, a copula provides a method for introducing dependencies

between random variables while retaining their original marginal distributions.

We have the following formal definition of a copula.

Definition 2.4.1. A function C : [0, 1]n → [0, 1] is a copula if:

1. There are random variables U1 , . . . , Un taking values in [0, 1] such that C

is their distribution function;

2. C has uniform marginal distributions, i.e., for all i ≤ n, ui ∈ [0, 1]

C(1, . . . , 1, ui, 1, . . . , 1) = ui

36
Chapter 2. Credit Derivatives Pricing Methods

One of the most important theorems for copulas is Sklar’s theorem. Sklar’s

theorem enables one to separate a multivariate density into the marginals and

a dependency structure given by the copula. Specifically,

Theorem 2.5. (Sklar) Let X1 , . . . , Xn be random variables with marginal dis-

tribution functions F1 , F2 , . . . , Fn and joint distribution function F . Then there

exits an n-dimensional copula C such that for all x ∈ Rn :

F (x) = C(F1 (x1 ), F2 (x2 ), . . . , Fn (xn ))

i.e. C is the distribution function of (F1 (x1 ), . . . , Fn (xn )). If F1 , . . . , Fn

are continuous, then C is unique. Otherwise, C is uniquely determined on

Range(F1 ) × · · · × Range(Fn ), where Range(Fi ) denotes the range of Fi for

i = 1, . . . , n.

The critical interpretation of the above is that we can create any multi-

dimensional distribution by using the marginal distributions and a judicious

choice of a copula function. Therefore, in principal, we can create any type

of correlation structure. While the list of copulas is endless, we focus on one

particular type of copula that has attracted a lot of attention and use in financial

modeling.

37
Chapter 2. Credit Derivatives Pricing Methods

The Gaussian copula is the copula of choice for many financial applications.

It became popular based on Li’s paper on default correlation modeling [10]. The

Gaussian copula is the following:

Definition 2.5.1. (Gaussian copula) Let X1 , . . . , Xn be normally distributed

random variable with means μ1 , . . . , μn , standard deviations σ1 , . . . , σn and cor-

relation matrix R. Then the distribution function CR (u1 , . . . , un ) of the random

variables

 
Xi − μ i
Ui := Φ , i≤n
σi

is a copula and it is called the Gaussian copula to the correlation matrix R

(where Φ(·) denotes the cumulative univariate standard normal distribution

function).

The Gaussian copula is popular mainly due to its analytical tractability.

That is, it is very easy to simulate from and the dependency structure is more

clear than other families of copulas. To simulate from the Gaussian copula, one

can follow the algorithm presented in [27]. That is,

1. Generate X1 , . . . , Xn with are jointly Φ0,R (·) normal distributed with mean

zero, standard deviation one and correlation matrix R.

38
Chapter 2. Credit Derivatives Pricing Methods

2. Let

Ui = Φ(Xi ), ∀i ≤ n

where Φ(·) is the univariate cumulative normal distribution function. Then

U1 , . . . , Un are CR distributed.

Then these uniform random variables can be used, for example, to generate

default times in the intensity modeling framework introduced in the previous

section. That is, we replace the CR correlated random variables Ui with the

independent U[0, 1] random variables in the case of multi-dimensional credit

derivatives.

39
Chapter 3

Feynman-Kac Path Measures


and Interacting Particle Systems

3.1 General Class of Feynman-Kac Path Mea-

sures

In this section, we introduce the reader to Feynman-Kac path measures.

These path measures will be used throughout the rest of this dissertation and

will form the foundation for our analysis of financial models presented in later

chapters.

Feynman-Kac path measures are formulated with regard to two objects.

First, a Markov chain that will be associated to a reference probability mea-

sure. Second, a collection of potential functions that will provide a way to form

Boltzman-Gibbs measures.

40
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

We first introduce some fundamental notation that will be needed for our

analysis. We suppose that we have a non-homogenuous Markov chain, Xn ,

n ∈ N. The chain Xn takes values in some measurable state space (En , En ) with

Markov transitions Kn (xn−1 , dxn ).

We write,

 

Ω= En , F = (Fn )n∈N , X = (Xn )n∈N , Pμ
n≥0

as the full representation of the nonhomogenous Markov chain Xn . This

chain takes values at each time n on En and has the elementary transitions Kn

with initial distribution μ, where μ is chosen from the space of all probability

measures on E0 . Let Bb (E) denote the space of bounded, measurable functions

with the uniform norm on some measurable space (E, E).

To compute expectations of functions of the Markov chain Xn we let Fn ∈

Bb (E[0,n] ). Then we have,


Eμ (Fn (X0 , . . . , Xn )) = Fn (x0 , . . . , xn )Pμ,n (d(x0 , . . . , xn ))
E[0,n]

where the distribution Pμ,n on E[0,n] is given by

Pμ,n ((d(x0 , . . . , xn )) = μ(dx0 )K1 (x0 , dx1 ) · · · Kn (xn−1 , dxn )

41
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

which follows directly from the fact that Xn is Markov with transitions given

by Kn .

We now turn our attention to the potential functions that will also be needed.

Let Gn : En → [0, ∞) be a collection of bounded and measurable functions such

that, for any n ∈ N,

 

n
Eμ Gp (Xp ) >0
p=0

Next, we have the following definition.

Definition 3.1.1. The Feynman-Kac prediction and updated path models asso-

ciated with the pair (Gn , Kn ) (and the initial distribution μ) are the sequence of

path measures defined respectively by

n−1 
1 
Qμ,n (d(x0 , . . . , xn )) = Gp (xp ) Pμ,n (d(x0 , . . . , xn )) (3.1)
Zn p=0

 

n
 μ,n (d(x0 , . . . , xn )) = 1
Q Gp (xp ) Pμ,n (d(x0 , . . . , xn ))
Zn p=0

for any n ∈ N. The normalizing constants,


n−1   n 
 
Zn = Eμ Gp (Xp ) and Zn = Zn+1 = Eμ Gp (Xp )
p=0 p=0

are also often called the partition functions.

42
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

We also have the alternative definition for the measures Qμ,n that’s given

by,

 
1 
n−1
Qμ,n (Fn ) = Eμ Fn (X0 , . . . , Xn ) Gp (Xp )
Zn p=0

for any test function Fn ∈ B(F[0,n] ).

We digress briefly to note that we have the following property,

 

n−1
Eμ (Fn ) = Zn Qμ,n Fn G−
p (3.2)
p=0

i.e., the measure is applied pointwise to the product of the functions where

the inverse of Gp is defined as

G−
p = 1/Gp

Equation (3.2) is important because it enables us to express an expectation

under the original measure in terms of a new expectation under a different path

measure and how to correct for it to ensure the two are equal. This relationship

will play a pivotal role in determining the change of measure to use later.

In addition, to simplify the presentation and to be more conducive to the

analysis we will be using, we introduce the following notation.

43
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

Definition 3.1.2. The sequence of bounded nonnegative measures γn and γn on

En defined for any fn ∈ B(En ) by


 

n−1
γn (fn ) = Eμ fn (Xn ) Gp (Xp )
p=0

and
 

n
γn (fn ) = Eμ fn (Xn ) Gp (Xp )
p=0

are respectively called the unnormalized prediction and the updated Feynman-

Kac model associated with the pair (Gn , Kn ). The sequence of distributions ηn

and ηn defined for any fn ∈ B(En ) as

ηn (fn ) = γn (fn )/γn (1) and ηn (fn ) = γn (fn )/
γn (1) (3.3)

are respectively called the normalized prediction and updated Feynman-Kac

model associated with the pair (Gn , Kn ).

In fact, there are several properties and connections between the various

equations above that are worth exploring. Firstly, note that for n = 0, we have

trivially that γ0 = η0 = μ. Secondly, it is easy to check the following relationship

 

n
Zn = Zn+1 = 
γn (1) = γn+1 (1) = Eμ Gp (Xp )
p=0

In addition,

44
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

 

n−1
γn (fn Gn ) = Eμ fn (Xn )Gn (Xn ) Gp (Xp ) =
γn (fn )
p=0

It is easy to check that we also have,

γn (fn Gn ) γn (fn Gn )/γn (1) ηn (fn Gn )


ηn (fn ) = = =
γn (Gn ) γn (Gn )/γn (1) ηn (Gn )

We are led to the easily proven proposition:

Proposition 3.1.1. For any n ∈ N and for any fn ∈ Bb (En ), we have


n−1 
n
γn (fn ) = ηn (fn ) ηp (Gp ) and 
γn (fn ) = ηn (fn ) ηp (Gp )
p=0 p=0

As we develop the interacting particle interpretation of the Feynman-Kac

path measures, we will be using Boltzmann-Gibbs measures frequently. Let

P(E) denote the space of probability measures on a given measure space (E, E).

We now present the formal definition for the Boltzmann-Gibbs transformation.

Definition 3.1.3. The Boltzmann-Gibbs transformation associated with the po-

tential function Gn on (En , En ) is the mapping

Ψn : η ∈ P(En ) → Ψn (η) ∈ P(En )

from the subset Ψn (En ) = {η ∈ P(En ); η(Gn ) > 0} into itself and defined for

any η ∈ P(En ) by the Boltzmann-Gibbs measure

1
Ψn (η)(dxn ) = Gn (xn )η(dxn )
η(Gn )

45
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

3.2 Useful Results of Feynman-Kac Path Mea-

sures

Section 3.1 presented the notation and general class of Feynman-Kac models

that will form the foundation for the interacting particle system we will develop

in later chapters. In this section, we extend the previous section by presenting

the results of Feynman-Kac path measure theory that are instrumental for our

model development.

The results presented in Section 3.1 were presented for a general non-homogenous

Markov chain Xn , n ∈ N. For the development of our model, we will find it

more convenient to work on the historical path of the process Xn .

Let
def. def.
Yn = (X0 , . . . , Xn ) ∈ Fn = (E0 × · · · × En )

Let Mn (yn−1 , dyn ) be the Markov transitions associated with the chain Yn .

Then, given any fn ∈ Bb (Fn ), and the pair of potentials/transitions (Gn , Mn ),

we have the following Feynman-Kac measure defined by


 

γn (fn ) = E fn (Yn ) Gk (Yk ) (3.4)
1≤k<n

46
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

We denote by ηn (·) the normalized measure defined as



E fn (Yn ) 1≤k<n Gk (Yk )
ηn (fn ) =  = γn (fn )/γn (1) (3.5)
E 1≤k<n Gk (Yk )

In addition, in [12] they assume that the potential functions are chosen such

that

sup Gn (yn )/Gn (ȳn ) < ∞


(yn ,ȳn )∈Fn2

However, the authors note that this condition can be relaxed by considering

traditional cut-off techniques, among other techniques (see [12, 11] for more

details).

Proposition 3.2.1 ([11]). The normalized prediction and updated Feynman-

Kac distributions ηn and ηn associated with the pair (Gn , Mn ) satisfy the non-

 n (
linear recursive equations ηn = Φn (ηn−1 ) and ηn = Φ ηn−1 ) with the mappings

 n from P(Fn−1 ) into P(Fn ) defined for any ηn−1 ∈ P(Fn−1 ) by


Φn and Φ

 n (η) = Ψn (ηMn ) = Ψ
Φn (η) = Ψn−1 (η)Mn and Φ  n−1 (η)M
n

 n denote the Boltzmann-Gibbs transformations on P(Fn ) given by


Ψn and Ψ

Gn (xn ) 
Ψn (μ)(dxn ) =  n (μ)(dxn ) = Gn (xn ) μ(dxn )
μ(dxn ) and Ψ
μ(Gn ) μ(Gn)

47
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

n , M
The pair potentials/kernels (G n ) are defined for any fn ∈ Bb (Fn ) by the

formula

n = Mn+1 (Gn+1 ) and M


G n (fn ) = Mn (fn Gn )/Mn (Gn )

The results in Proposition 3.2.1 can also be written in the equivalent integral

form

def.
ηn = Φn (ηn−1 ) = ηn−1 (dyn−1 )Gn−1 (yn−1)Mn (yn−1 , ·)/ηn−1(Gn−1 )
Fn−1

starting from η1 = M1 (x0 , ·).

In fact, by direct inspection of the above proposition, we see that ηn =

Φn (ηn−1 ) is actually a two step process given by the following updating/prediction

model

ηn−1 ∈ P(Fn−1 ) −→ ηn−1 ∈ P(Fn−1 ) −→ ηn ∈ P(Fn )


updating prediction

where “updating” occurs by applying the Botzmann-Gibbs transformation

Ψn−1 (ηn−1 )(dxn ) to the known measure ηn−1 followed by the “prediction” that’s

given by applying the known Markov kernel Mn .

3.2.1 Note on Feynman-Kac Change of Measure

It is important to emphasize the impact and implication of having this re-

cursive relationship. Let n be a fixed time. Suppose we start with γ0 = η0 = μ,

48
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

which is known. In addition, the chain Yn has Markov transitions given by the

kernel Mn . Then, following the notation in Section 3.1 and using the Markov

property we have

Pμ,n (d(x0 , y1 , . . . , yn )) = μ(dx0 )M1 (x0 , dy1) . . . Mn (yn−1 , dyn) (3.6)

We call the measure in Equation 3.6 the original measure because we have

assumed that the kernels are known explicitly (or at least how to sample from

them).

Recall that the normalized distributions associated with the pair (Gn , Mn )

is given by

n−1 
1 
ηn (d(x0 , y1, . . . , yn )) = Gp (yp ) Pμ,n (d(x0 , y1 , . . . , yn ))
γn (1) p=0

ηn is a change of measure that is sometimes called the “twisted” distribution

in the importance sampling literature. Given any fn ∈ Bb (Fn ) we can use

Proposition 3.1.1 to express the expectation under the original measure as an

expectation under the change of measure. We find that,

49
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

 

n−1 
n−1
Eμ,n (fn (Yn )) = Eμ,n fn (Yn ) G−
p (Yp ) Gp (Yp )
p=0 p=0


n−1
= γn (fn G−
p)
p=0


n−1 
n−1
= ηn (fn G−
p) ηp (Gp )
p=0 p=0

We can write things more explicitly as


n−1 
n−1  
n−1
ηn (fn G−
p) = Eηn (fn (Yn ) G−
p (Yp )) = fn (yn ) Gp (yp )ηn (d(x0 , y1, . . . , yn ))
p=0 p=0 Fn p=0

ηp (Gp ) = Eηp (Gp (Yp )) = Gp (yp )ηp (d(x0 , y1 , . . . , yp))
Fp

Now the power of the recursive relationship in Proposition 3.2.1 is clear. To

compute the expectations that appear by using a Feynman-Kac type change of

measure, we need to be able to either explicitly find the new measures ηp or be

able to sample from them for a numerical scheme. However, Proposition 3.2.1

tells us exactly the prescription that’s needed. The power lies in the fact that

at each time step n the recursive relationship only depends on ηn−1 , which is

known and Mn , which is also known. In addition, given that η0 = μ which is

also known, the circle is complete and we have a clear way to proceed.

50
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

While the recursive relationship in Proposition 3.2.1 is very useful when

we are able to compute things explicitly, its real strength will be shown when

we aren’t able to solve the recursive equation explicitly and have to rely on

numerical approximations to the measures ηp .

3.3 Interacting Particle System Interpretation

of Feynman-Kac Models

In this section, we present the IPS interpretation of Feynman-Kac models.

We are specifically interested in the situation discussed in the previous section

where the changes of measure can’t be explicitly computed. The idea can be

tersely described as appending a standard MC routine with a judiciously formed

selection/mutation algorithm such that we can form the normalized Feynman-

Kac measures empirically. After such a development, we will analyze the rate of

convergence of the empirically formed normalized Feynman-Kac path measures

to their true distribution.

To begin, we recall Proposition 3.2.1 that states the measure ηn associated

with the pair (Gn , Mn ) satisfies the nonlinear recursive equation ηn = Φn (ηn−1 ),

where

51
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

Φn (η) = Ψn−1 (η)Mn

and

Gn (xn )
Ψn (μ)(dxn ) = μ(dxn )
μ(Gn )

Using this proposition, we can write the transition ηn → ηn+1 as

ηn+1 = Ψn (ηn )Mn+1 (3.7)

where,
Gn (xn )
Ψn (ηn ) = ηn (dxn )
ηn (Gn )

In fact, it will be more convenient to write Equation (3.7) in the more

symbolic form

ηn+1 = ηn Kn+1,ηn (3.8)

where

Kn+1,η = Sn,η Mn+1

Sn,η (xn , ·) = Ψn (η)

52
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

We note that in this notation, Kn+1,ηn represents the Markov transition for

ηn ∈ P(Fn ) → ηn+1 ∈ P(Fn+1 ).

We can now explicitly define the interacting particle model.

Definition 3.3.1. The interacting particle model associated with a collection

of Markov transitions Kn,η , η ∈ P(Fn ), n ≥ 1, and with an initial distribution

η0 ∈ P(F0 ) is a sequence of nonhomogeneous Markov chains

 

Ω(N ) = FnN , F N = (FnN )n∈N , ξ (N ) = (ξn(N ) )n∈N , PN
η0
n≥0

taking values at each time n ∈ N in the product space FnN , that is

ξn(N ) = (ξn(N,1) , . . . , ξn(N,N ) ) ∈ FnN = Fn × · · · × Fn


 ! "
N times

(N )
The initial configuration ξ0 consist of N independent and identically dis-

tributed random variables with common law η0 . Its elementary transitions from

N
Fn−1 into FnN are given in a symbolic integral form by

  
N
(N ) (N,p)
PN
η0 ξn ∈ dxn |ξn−1 =
(N )
Kn, 1 PN (ξn−1 , dxpn ) (3.9)
N i=1 δξ(N,i)
p=1 n−1

where dxn = dx1n × · · · × xN


n is an infinitesimal neighborhood of a point xn =

n ) ∈ Fn
(x1n , . . . , xN N

53
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

Using the above definition as a springboard, we write

1 
N
m(xn ) = δxi
N i=1 n

for each xn = (xn )i1≤i≤N ∈ FnN . Then, we can write Equation (3.9) in the

more compact notation

  N
(N ) p
PN
η0 ξ (N )
n ∈ dx |ξ
n n−1 = Kn,m(ξn−1 ) (ξn−1 , dxpn )
p=1

Recalling the definition of Kn,η for our problem, we can write the above

equation as

  
N
(N ) p
PN
η0 ξn ∈ dxn |ξn−1 =
(N )
(Sn,m(ξn ) )Mn+1 (ξn−1 , dxpn )
p=1

Which leads directly to

  
(N )
PN
η0 ξn(N ) ∈ dxn |ξn−1 = Sn (ξn , dxn )Mn+1 (xn , dxn+1 ) (3.10)
FnN

where,


N
Sn (ξn , dxn ) = Sn,m(ξn ) (ξnp , dxpn )
p=1


N
Mn+1 (xn , dxn+1 ) = Mn+1 (xpn , dxpn+1 )
p=1

54
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

is the Boltzman-Gibbs transformation from FnN into itself and the mutation

transition from FnN to Fn+1


N
, respectively.

3.3.1 Interacting Particle System Algorithm

Interpreting the above definitions and equations is critical to form the correct

IPS algorithm. By inspecting the integral decomposition above, we see that we

can replace

Sn,η Mn+1
ηn ∈ P(Fn ) −→n ηn = ηn Sn,ηn ∈ P(Fn ) −→ ηn+1 = ηn+1 Mn+1

with the two step selection/mutation transition in product spaces

ξn ∈ FnN −→ ξn ∈ FnN −→ ξn+1 ∈ Fn+1


selection mutation N

where the selection and mutation transition is given by the following algo-

rithm:

• Selection: Assume we are at time n with particles ξn = (ξni )1≤i≤N ∈ FnN .

Selection only depends on the particles ξn and the potential function Gn .

We see that from the definition of Sn (ξn , ·) we randomly select N particles

from (ξni )1≤i≤N with distribution

Sn,m(ξn ) (ξni , ·) = Ψn (m(ξn ))

55
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

More precisely, we select N particles ξn = (ξni )1≤i≤N with distribution


N
Gn (ξni )
Ψn (m(ξn )) = n j δξn
i

i=1 j=1 G n (ξ n )

Note that selecting each particle ξni is done independently of one another

and the sampling is done with replacement. Therefore, the number of

particles stays the same throughout the algorithm.

• Mutation: The mutation stage is very straightforward. Once we have

selected the new N particles, ξn = (ξni )1≤i≤N ∈ FnN , we mutate each par-

ticle ξni independently of one another according to the Markov transition

i
prescribed by Mn+1 . That is, we sample independently N particles, ξn+1 ,

with respective distribution Mn+1 (ξni , ·). The new set of particles is de-

i
noted by ξn+1 = (ξn+1 )1≤i≤N ∈ Fn+1
N
.

Building from previous sections, we recall that we have the following repre-

sentation connecting the un-normalized measures to the normalized measures,


n−1
γn (fn ) = ηn (fn ) ηp (Gp )
p=0

Given the above algorithm, the particle approximation measures for (γn , ηn )

are defined by

56
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

n−1 
 1 
N
γnN (·) = ηpN (Gp ) × ηnN (·) with ηnN = δ i i

p=0
N i=1 (ξ0,n ,...,ξn,n )

3.4 Some Convergence Results for Interacting

Particle Systems for Rare Events

In this section, we present some important convergence results developed in

[11] and expanded upon in [12]. Specifically, we find it important to concentrate

on the results in [12] as this article directly analyzes interacting particle systems

for rare events. These convergence results, including the proof of the unbiased-

ness for IPS estimators, will be used directly in later chapters to analyze our

IPS interpretation of credit derivatives pricing.

As mentioned in the previous paragraph, [12] analyzes the use of IPS for

calculating rare event probabilities. Therefore, we present many of the results

contained in [12] for completeness as they will be pivotal for the results we

present later.

To analyze the convergence of the IPS scheme, we will focus on analyzing

the local sampling errors of the particle approximation steps. As such, we define

the random field

57
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems


WnN (fn ) = N[ηnN − Φn (ηn−1
N
)](fn )

for any test function fn ∈ Bb (Fn ). Then we have the following central limit

theorem that is presented in [11] and simplified in [12].

Theorem 3.5. For any fixed time horizon n ≥ 1, the sequence (WpN )1≤p≤n

converges in law, as N goes to infinity, to a sequence of n independent, Gaussian

and centered random fields (Wp )1≤p≤n ; with, for any fp , gp ∈ Bb (Fp ) and 1 ≤

p ≤ n,

E[Wp (fp )Wp (gp )] = ηp ([fp − ηp (fp )][gp − ηp (gp )])

Proof. See Theorem 9.3.1, page 295 in [11].

In addition, define Qp,n , 1 ≤ p ≤ n by the equation

 

Qp,n (fn )(yp ) = E fn (Yn ) Gk (Yk )|Yp = yp
p≤k<n

for any fn ∈ Bb (Fn ) and any path sequence yp = (x0 , . . . , xp ) ∈ Fp with the

convention Qn,n = I. Then, Qp,n is the Feynman-Kac semigroup associated

with the measures (γp )1≤p≤n .

The following lemma, presented in [12], is instrumental. We include the

proof for completeness.

58
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

Lemma 3.5.1. The semigroup Qp,n satisfies, ∀1 ≤ p ≤ n

γn = γp Qp,n

Proof. First, note that for any fn ∈ Bb (Fn ),


 

γn (fn ) = E fn (Yn ) Gk (Yk )
1≤k<n
   
 
= E Gk (Yk ) × E fn (Yn ) Gk (Yk )|(Y0, . . . , Yp )
1≤k<p p≤k<n

Now, using the Markov property,

   
 
γn (fn ) = E Gk (Yk ) × E fn (Yn ) Gk (Yk )|Yp
1≤k<p p≤k<n
  

= E Gk (Yk ) Qp,n (fn )(Yp ) = γp Qp,n (fn )
1≤k<p

The reason for having the above lemma is to make the telescopic decompo-

sition


n
γnN − γn = [γpN Qp,n − γp−1
N
Qp−1,n ]
p=1

59
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

Recall that for p = 1, we have η0N = δx0 and γ1 = η1 = M1 (x0 , ·). This

implies that η0N Q0,n = γ1 Q1,n = γn . In addition, we also have

N
γp−1 N
Qp−1,p = ηp−1 (Gp−1 ) × Φp−1 (ηp−1
N N
) and γp−1 (Gp−1 ) = γpN (1)

This implies that

def. √  n
Wnγ,N (fn ) = N [γnN − γn ](fn ) = γpN (1)WpN (Qp,n fn ) (3.11)
p=1

Then we readily have the following lemma from [12].

Lemma 3.5.2. γnN is an unbiased estimate for γn , in the sense that for any

p ≥ 1 and fn ∈ Bb (Fn ), with ||fn || ≤ 1, we have


E(γnN (fn )) = γn (fn ) and sup N E[|γnN (fn ) − γn (fn )|p ]1/p ≤ cp (n)
N ≥1

for some constant cp (n) < ∞ whose value does not depend on the function fn .

Proof. See [12] for the complete proof. We include the first part for illustrative

purposes.

First, notice that (γnN (1))n≥1 is a predictable sequence, in the sense that

n−1 
 
n−1
E(γnN (1)|ξ0 , . . . , ξn−1) =E ηpN (Gp )|ξ0 , . . . , ξn−1 = ηpN (Gp ) = γnN (1)
p=1 p=1

60
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

In addition, we also have,

E(WpN (Qp,n fn )|ξ0, . . . , ξp−1)



= N E([ηpN − Φp (ηp−1
N
)](Qp,n fn )|ξ0 , . . . , ξp−1) = 0

by the central limit theorem presented earlier in this chapter. Therefore,

combining these two observations, we have

E(γpN (1)WpN (Qp,n fn )|ξ0 , . . . , ξp−1) = 0

Hence, E(γpN (1)WpN (Qp,n fn )) = 0.

Therefore, recalling that

def. √  n
Wnγ,N (fn ) = N[γnN − γn ](fn ) = γpN (1)WpN (Qp,n fn )
p=1

we find

 n 
1 
E([γnN − γn ](fn )) = √ E γpN (1)WpN (Qp,n fn )
N p=1

1 
n
= √ E(γpN (1)WpN (Qp,n fn ))
N p=1
= 0

Hence γnN is unbiased.

61
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

Lemma 3.5.2 is important because it shows that as N goes to infinity, the ran-

dom sequence (γpN (1))1≤p≤n converges to the deterministic sequence (γp (1))1≤p≤n

in probability. Then, a quick application of Slutsky’s lemma implies that the

random fields Wnγ,N converge in law, as N tends to infinity, to the Gaussian

random fields Wnγ defined for any fn ∈ Bb (Fn ) by


n
Wnγ (fn ) = γp (1)Wp (Qp,n fn ) (3.12)
p=1

We end this section by presenting the major theorem that was developed in

[12] for analyzing the asymptotic variance of the IPS estimators for rare event

probabilities. First, we introduce some new notation.

Suppose we would like to estimate the probability

Pn (a) = P(Vn (Xn ) ≥ a)

where Vn ∈ En → R is some function and the value a is chosen such that the

above probability is extremely small (i.e. Pn (a) ∈ [10−12 , 10−6 ]). Recall that

Xn is a non-homogeneous Markov chain taking values in some space En .

We are interested in the situation where we have no analytical formulas

for Pn (a) so a numerical routine has to be used. In addition, because Pn (a)

is so small, using traditional MC would require an extremely large number of

62
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems

trajectories to reach the rare-level set. Therefore, one possible solution is to

develop an IPS routine to estimate Pn (a) and analyze the resulting variance.

Recall that we set Yk = (X0 , . . . , Xk ). Then, for a given set of potential

functions (Gk )1≤k≤n that satisfy the conditions stated in Section 3.2 we can

rewrite the probability Pn (a) in terms of a Feynman-Kac formula,

Pn (a) = E(1Vn (Xn )≥a )


 
 
= E 1Vn (Xn )≥a G−
k (Yk ) Gk (Yk )
1≤k<n 1≤k<n
 

= γn 1Vn (Xn )≥a G−
k (Yk )
1≤k<n

= γn (Tn(a) (1))

= ηn (Tn(a) (1)) ηk (Gk )
1≤k<n

(a)
where we define Tn for any path yn = (x0 , . . . , xn ) by


Tn(a) (1)(yn ) = Tn(a) (1)(x0 , . . . , xn ) = 1Vn (xn )≥a G−
k (x0 , . . . , xk )
1≤k<n

Therefore, by following the selection and mutation algorithm presented in

Section 3.3.1, we form the estimator of Pn (a) by replacing the normalized

Feynman-Kac measures above by their particle approximations. That is, we

set our estimator of Pn (a) to be

63
Chapter 3. Feynman-Kac Path Measures and Interacting Particle Systems


PnN (a) = ηnN (Tn(a) (1)) ηkN (Gk )
1≤k<n
 
1 
N  
= 1Vn (Xni )≥a G− i i i
k (X0 , X1 , . . . , Xk ) ηkN (Gk )
N i=1 1≤k<n 1≤k<n

We are now in a position to state the theorem.

Theorem 3.6. The estimator PnN (a) is unbiased, and it satisfies the central

limit theorem

√ N →∞
N [PnN (a) − Pn (a)] → N(0, σnγ (a)2 ) (3.13)

with the asymptotic variance


n
σnγ (a)2 = [γp (1)γp− (1)ηp− (Pp,n (a)2 ) − Pn (a)2 ]
p=1
n
= [γp (1)γp− (Pp,n (a)2 ) − Pn (a)2 ] (3.14)
p=1

and the collection of functions Pp,n (a) defined by

xp ∈ Ep
→ Pp,n (a)(xp ) = E[1Vn (Xn )≥a |Xp = xp ] ∈ [0, 1] (3.15)

Proof. See [12].

64
Chapter 4

Credit Derivatives Pricing with


Interacting Particle Systems

4.1 Introduction

In contrast to intensity-based approaches in credit risk where default is given

by an exogenously defined process, default in the firm value model has a very

nice economical appeal. The firm value approach, or structural model, models

the total value of the firm that has issued the defaultable bonds. Typically, the

value of the firm includes the value of equity (shares) and the debt (bonds) of the

firm [27]. There are two main approaches to modeling default in the firm value

approach: one is that default can only happen at the maturity of the bond, and

the second is that default can occur any time before maturity. We have chosen

to focus on the firm value approach, and specifically the first passage approach,

due to a feature unique in applying interacting particle systems to rare event

65
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

probabilities. As will be discussed later, to capture the most variance reduction

by applying IPS, one has to choose potential functions (the Gk ’s) that closely

mimics the behavior of the random variable that precedes the rare event. As

will be shown, the first passage model admits a very natural choice of weight

function to use.

We are specifically interested in the intersection of two events; a rare event

probability calculation in the context of credit derivatives pricing with a sub-

stantial reduction in variance when compared to Monte Carlo. As such, we

have chosen to formulate our approach in the context of a first passage model.

Our major concern is developing a method to estimate these small probabilities

and not on the calibration of the model. As such, we assume that our model

has been calibrated to market data and that the pricing that is needed exhibit

a strong dependence on the small probabilities that we wish to calculate. We

encourage the reader to see [5] for information regarding the calibration of a

structural model to CDS rates.

It should be emphasized that the method and simulations presented in later

sections occur under the original measure. That is, to achieve the variance

reduction, one does not have to compute a change of measure explicitly. Rather,

the change of measure is performed in path space through the IPS algorithm

of selection and mutation. The important point is that the simulation of the

66
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

trajectories is performed under the original measure. This is in direct contrast to

one of the most powerful and popular variance reduction techniques, importance

sampling. Therefore, our method has a distinct advantage over importance

sampling in the framework where a change of measure is either not available or

not tractable to simulate from. In addition, another advantage over importance

sampling is that IPS is dynamically consistent in time. Often, when using

importance sampling and employing Girsanov’s theorem, there is a dependence

on the maturity time T in the new dynamics. This dependence means that each

maturity date needs its own simulation. IPS doesn’t suffer from this problem

and we will show how to use the same simulation to compute the tranche spreads

for all maturities.

4.2 Problem Formulation

We follow both [26, 3] and assume that the value of the firm, St , follows

geometric Brownian motion. We also assume that interest rates are constant.

Under the risk-neutral probability measure P we have,

dS(t) = rS(t)dt + σS(t)dW (t) (4.1)

where r is the risk-free interest rate, and σ is constant volatility. At any time

t ≤ T , the price of the nondefaultable bond is Γ(t, T ) = e−r(T −t) .

67
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

We also assume that at time 0, the firm issued non-coupon corporate bonds

expiring at time T . The price of the defaultable bond at time t ≤ T is denoted

Γ̄(t, T ).

In [26] default was assumed to only occur at the expiration date T . Fur-

thermore, default at time T is triggered if the value of the firm was below some

default threshold B; that is if ST ≤ B. Therefore, assuming zero recovery, the

price of the defaultable bond satisfies,


Γ̄(t, T ) = E e−r(T −t) 1ST >B |St

= Γ(t, T )P(ST > B|St )

= Γ(t, T )N(d2 )

where N(·) is the standard cumulative normal distribution function and,



S
ln t
+ (r − 12 σ 2 )(T − t)
d2 = B

σ T −t

The next model, the Black-Cox model, is also known as the first passage ap-

proach. Developed in [3], default can occur anytime before the expiration of

the bond and the barrier level, B(t), is some deterministic function of time. In

[3], they assume that the default barrier is given by the function B(t) = Keηt

(exponentially increasing barrier) with K > 0 and η ≥ 0. The default time τ is

68
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

defined by

τ = inf {t : St ≤ B(t)}

That is, default happens the first time the value of the firm passes below the

default barrier. This has the very appealing economical interpretation that

default occurs when the value of the firm falls below its debt level, thereby not

allowing the firm to pay off its debt.

Assuming zero recovery, we can price the defaultable bond by pricing a

barrier option. Therefore,


Γ̄(t, T ) = 1τ >t E e−r(T −t) 1τ >T |St

= 1τ >t Γ(t, T )P(τ > T |St )


  p 
St −
= 1τ >t Γ(t, T ) N(d2 ) −
+
N(d2 )
B(t)

where
 
± ln + (r − η − 12 σ 2 )(T − t)
St
B(t)

2 = √
σ T −t
2(r − η)
p = 1−
σ2

In addition, for a constant barrier level B, we denote the probability of default

for the firm between time t and time T by P (t, T ). Hence,


  1− 2r2 
St σ
P B (t, T ) = 1 − 2)−
N(d+ N(d−
2) (4.2)
B

69
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

4.2.1 Multi-Name Model

For our purposes, we consider a CDO written on N firms under the first

passage model. That is, we assume that the firm values for the N names have

the following dynamics,

dSi(t) = rSi (t)dt + σi Si (t)dWi (t) i = 1, . . . , N (4.3)

where r is the risk-free interest rate, σi is constant volatility, and the driving

innovations dWi(t) are infinitesimal increments of Wiener processes Wi with

correlation

d Wi , Wj t = ρij dt.

Each firm i is also assumed to have a deterministic boundary process Bi (t) and

default for firm i is given by

τi = inf {t : Si (t) ≤ Bi (t)} (4.4)

We define the loss function as


N
L(T ) = 1{τi ≤T } (4.5)
i=1

That is, L(T ) counts the number of firms among the N firms that have de-

faulted before time T . We remark that in the independent homogeneous port-

folio case, with constant barrier, the distribution of L(T ) is Bin(N, P B (0, T ))

where P B (0, T ) is defined in Equation (4.2).

70
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

Recall that for a homogeneous portfolio, it is well known that the spread

on a single tranche can be computed from the knowledge of expectations of the

form,

E{(L(Ti ) − Kj )+ }

where Ti is any of the coupon payment dates, and Kj is proportional to the

attachment points of the tranche.

The most interesting and challenging computational problem is when all of

the names in the portfolio are correlated. In [29], the distribution of losses for

N = 2 is found by finding the distribution of the hitting times of a pair of

correlated Brownian motions. However, the distribution is given in terms of

modified Bessel functions and a general result for N > 2 has yet to be found.

Since the distribution of L(T ) isn’t known in the dependent case, for N > 2, MC

methods are generally used to calculate the spread on the tranches. Since N

is typically very large (125 names is a standard contract), PDE-based methods

are ruled out and one has to use MC.

Instead of computing the spread on the tranches numerically, our goal is to

calculate the probability mass function for L(T ), that is calculate

P(L(T ) = i) = pi (T ) i = 0, . . . , N (4.6)

71
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

In this manner, we will then be able to calculate all expectations that are a

function of L(T ), not just the spreads. In addition, as the reader will see, our

method is dynamically consistent in time so that we can actually calculate, for

all coupon dates Tj ≤ T ,

P(L(Tj ) = i) = pi (Tj ) i = 0, . . . , N (4.7)

with one MC run. This is in contrast to a lot of importance sampling techniques

where the change of measure has a dependence on the final time through the

Girsanov transformation thereby requiring a different MC run for each coupon

date.

4.3 Feynman-Kac Path Measures and Interact-

ing Particle Systems

Feynman-Kac path measures, and their subsequent interacting particle sys-

tem interpretation, are closely related to the stochastic filtering techniques used

in mathematical finance. In this chapter, we adapt an original interacting par-

ticle system developed in [12] to the computation of equation (4.7). In [12],

the authors develop a general interacting particle system method for calculat-

ing the probabilities of rare events. As discussed in Chapter 3, the method

can be rapidly described as formulating a Markov process and conducting a

72
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

mutation-selection algorithm so that the chain is “forced” into the rare event

regime.

For the sake of the reader, we review the most important components of

Feynman-Kac and interacting particle systems before describing our model in

more detail.

We suppose that we have a continuous time non-homogenuous Markov chain,

(X̃t )t∈[0,T ] . However, for the variance analysis and to foreshadow the method

ahead, we only consider the chain (Xp )0≤p≤n = (X̃pT /n )0≤p≤n , where n is fixed.

The chain Xn takes values in some measurable state space (En , En ) with Markov

transitions Kn (xn−1 , dxn ). We denote by Yn the historical process of Xn , that

is
def. def.
Yn = (X0 , . . . , Xn ) ∈ Fn = (E0 × · · · × En )

Let Mn (yn−1, dyn ) be the Markov transitions associated with the chain Yn . Let

Bb (E) denote the space of bounded, measurable functions with the uniform

norm on some measurable space (E, E). Then, given any fn ∈ Bb (Fn ), and the

pair of potentials/transitions (Gn , Mn ), we have the following Feynman-Kac

measure defined by
 

γn (fn ) = E f (Yn ) Gk (Yk ) (4.8)
1≤k<n

73
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

We denote by ηn (·) the normalized measure defined as



E f (Yn ) 1≤k<n Gk (Yk )
ηn (fn ) =  = γn (fn )/γn (1) (4.9)
E 1≤k<n Gk (Yk )

In addition, in [12] they assume that the potential functions are chosen such

that

sup Gn (yn )/Gn (ȳn ) < ∞


(yn ,ȳn )∈Fn2

However, the authors note that this condition can be relaxed by considering

traditional cut-off techniques, among other techniques (see [12, 11] for more

details).

A very important observation is that


n
γn+1(1) = γn (Gn ) = ηn (Gn )γn (1) = ηp (Gp )
p=1

Therefore, given any bounded measurable function fn , we have


γn (fn ) = ηn (fn ) ηp (Gp )
1≤p<n

The above relationship is crucial because it enables us to relate the un-normalized

measure in terms of only the normalized “twisted” measures. In our study, we

will also make use of the distribution flow (γn− , ηn− ) defined exactly the same

way as (γn , ηn ) except we replace Gp by its inverse

G−
p = 1/Gp

74
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

Then, using the definition for γn and ηn it is easy to see that E[fn (Yn )] admits

the following representation,


 
 
E [fn (Yn )] = E fn (Yn ) G−
p (Yp ) × Gp (Yp )
1≤p<n 1≤p<n
 

= γn fn G−
p
1≤p<n
 
 
= ηn fn G−
p ηp (Gp )
1≤p<n 1≤p<n

Finally, it can be checked that the measures (ηn )n≥1 satisfy the nonlinear recur-

sive equation

def.
ηn = Φn (ηn−1 ) = ηn−1 (dyn−1 )Gn−1 (yn−1)Mn (yn−1 , ·)/ηn−1(Gn−1 )
Fn−1

starting from η1 = M1 (x0 , ·).

4.3.1 Interacting Particle System Interpretation and Gen-

eral Algorithm

The above definitions and results lend themselves to a very natural inter-

acting path-particle interpretation. We denote the Markov chain taking values

in the product space FnM with transformation Φn by ξn = (ξni )1≤i≤M , for each

time n ≥ 1. In [12], the authors constructed a numerical algorithm so that each

path-particle

ξni = (ξ0,n
i i
, ξ1,n i
, . . . , ξn,n ) ∈ Fn = (E0 × · · · × En )

75
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

is sampled almost according to the twisted measure ηn .

We start with an initial configuration ξ1 = (ξ1i )1≤i≤M that consists of M

independent and identically distributed random variables with distribution,

η1 (d(y0 , y1 )) = M1 (x0 , d(y0 , y1)) = δx0 (dy0 )K1 (y0 , dy1)

def. i
i.e., ξ1i = (ξ0,1 i
, ξ1,1 i
) = (x0 , ξ1,1 ) ∈ F1 = (E0 × E1 ). Then, the elementary

transitions ξn−1 → ξn from Fn−1


M
into FnM are defined by


M
P(ξn ∈ d(yn1 , . . . , ynM )|ξn−1) = Φn (m(ξn−1 ))(dyni ), (4.10)
i=1

def. 1 M
where m(ξn−1 ) = M i=1 δξn−1
i , and d(yn1 , . . . , ynM ) is an infinitesimal neighbor-

hood of the point (yn1 , . . . , ynM ) ∈ Fnm . From the definition of Φn , one can see

that (4.10) is the overlapping of a simple selection and mutation transition,

M selection ˆ M mutation
ξn−1 ∈ Fn−1 −→ ξn−1 ∈ Fn−1 −→ ξn ∈ FnM

The selection stage is performed by choosing randomly and independently M

path-particles

ξˆn−1
i
= (ξˆ0,n−1
i
, ξˆ1,n−1
i
, . . . , ξˆn−1,n−1
i
) ∈ Fn−1

according to the Boltzmann-Gibbs particle measure


M j
Gn−1 (ξ0,n−1 j
, . . . , ξn−1,n−1 )
M i i
δ(ξj j
0,n−1 ,...,ξn−1,n−1 )
j=1 i=1 Gn−1 ((ξ0,n−1 , . . . , ξn−1,n−1 )

76
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

Then, for the mutation stage, each selected path-particle ξˆn−1


i
is extended by

ξni = ((ξ0,n
i i
, . . . , ξn−1,n i
), ξn,n )

= ((ξˆ0,n
i
, . . . , ξˆn−1,n
i i
), ξn,n ) ∈ Fn = Fn−1 × En

i
where ξn,n is a random variable with distribution Kn (ξˆn−1,n−1
i
, ·). In other words,

the transition is made by applying the original kernel Kn . All of the mutations

are performed independently. We just quote the results from [11, 12] in stating

the weak convergence result

1 
M
def. N →∞
ηnM = i ,ξ i ,...,ξ i ) −→ ηn
δ(ξ0,n n,n
M i=1 1,n

Furthermore, there are several propagation of chaos estimates that ensure that

i i i
(ξ0,n , ξ1,n , . . . , ξn,n ) are asymptotically independent and identically distributed

with distribution ηn [11]. Therefore, we can form the particle approximation

γnM defined as

γnM (fn ) = ηnM (fn ) ηpM (Gp )
1≤p<n

Recall the following lemma.

Lemma 4.3.1 ([12]). γnM is an unbiased estimator for γn , in the sense that for

any p ≥ 1 and fn ∈ Bb (Fn ) with ||fn || ≤ 1, we have


E(γnM (fn )) = γn (fn ) and sup M E[|γnM (fn ) − γn (fn )|p ]1/p ≤ cp (n)
M ≥1

for some constant cp (n) < ∞ whose value does not depend on the function fn .

77
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

Proof. Refer to [12]

4.4 Pricing Using Interacting Particle Systems

In this section, we present our adaptation of the interacting particle system

approach to computing rare event probabilities in credit risk with a structural-

based approach by applying it to the following model.

Our Markov process is the 3 × N dimensional process (X̃t )t∈[0,T ] defined as:

X̃t = (S1 (t), min S1 (u), 1τ1 ≤t , S2 (t), min S2 (u), 1τ2 ≤t , · · ·
u≤t u≤t

· · · , SN (t), min SN (t), 1τN ≤t )


u≤t

where the dynamics of Si (t) are given in Equation (4.3). We assume a constant

barrier Bi for each firm 1 ≤ i ≤ N. While it is redundant to also include

1τi ≤t in the above expression since we also know minu≤t Si (u), we keep track of

it because it will tell us the default time of the firm when we implement the

algorithm numerically. We divide the time interval [0, T ] into n equal intervals

[ti−1 , ti ], i = 1, 2, . . . , n. These are the times we stop and do the selection and

mutation. We introduce the chain (Xp )0≤p≤n = (X̃pT /n )0≤p≤n and the whole

history of the chain denoted by Yp = (X0 , . . . , Xp ).

Since it isn’t possible to sample directly from the distribution of (Xp )0≤p≤n

for N > 2, we will have to apply an Euler scheme during the mutation stage;

78
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

we let t denote the sufficiently small time step used. In general t will be

chosen so that t << T /n.

Our general strategy is to find a potential function that increases the likeli-

hood of default among the firms. In the IPS algorithm, given a particular choice

of weight function G(·), particles with low scores are replaced by particles with

high scores. Therefore, we would like to select a potential function G(·) that

progressively only selects the particles whose firms are defaulting. Since the rare

event in this case is that the minimum of the firm value falls below a certain

level, we would like to put more emphasis on particles whose firm minimums

are decreasing during a mutation step.

Therefore, we fix some parameter α < 0 and define the potential function,

Gα (Yp ) = exp[α(V (Xp ) − V (Xp−1 ))] (4.11)

where

N
V (Xp ) = log(min Si (u))
u≤tp
i=1

The choice of α < 0 may seem peculiar initially, but it is chosen to be

negative because the potential function Gα (Yp ) can be written in the form,

Gα (Yp ) = exp[α(V (Xp ) − V (Xp−1 ))]


 N  

= exp α log min(Si (u)/ min Si (u)
u≤tp u≤tp−1
i=1

79
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

where,
 
log min(Si (u))/ min Si (u) ≤0
u≤tp u≤tp−1

Therefore, to place more weight on the firms whose minimum has decreased, we

must multiply by α < 0.

Hence, we will be putting more weight onto path-particles whose minimum

has decreased the most between two mutation times.

4.4.1 Note on Choice of Weight Function

There are several comments about this weight function that should be ad-

dressed.

For the moment, we only consider one firm name (N = 1) and drop the

subscript i. As written above, we have

 
log min(S(u))/ min S(u) ≤ 0
u≤tp u≤tp−1

However,

 
log min S(u)/ min S(u) = 0 ⇐⇒ min S(u) = min S(u)
u≤tp u≤tp−1 u≤tp u≤tp−1

and

80
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

 
log min S(u)/ min S(u) < 0 ⇐⇒ min S(u) < min S(u)
u≤tp u≤tp−1 u≤tp u≤tp−1

Therefore with α < 0,

  
α
G (Yp ) = exp α log min S(u)/ min S(u)
u≤tp u≤tp−1




⎨1, if minu≤tp S(u) = minu≤tp−1 S(u),
=



⎩> 1, if minu≤tp S(u) < minu≤tp−1 S(u),

Therefore, the particles whose minimums have not decreased in the last time

step will have a weight function value of 1. The particles whose minimums have

decreased greatly in the last time step will have a weight value greater than 1.

In addition, the more negative α is and the greater the decrease, the higher the

weight function value will be. Therefore, since we are sampling more frequently

from the particles with a large weight function value, we will be sampling from

the particles whose minimum has decreased the most. Thereby, we will be

putting more weight on the trajectories that are approaching the rare event

region.

In addition, there are several computational advantages for choosing the

weight function above. Chiefly among them are:

81
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

1. Our choice of weight function, while not unique in this regard, will only

require us to keep track of (Xp−1 , Xp ) instead of the full history Yp =

(X0 , X1 , · · · , Xp ) thereby minimizing the increased dimensionality of using

an IPS scheme.

2. In addition, our weight function has the added advantage of having the

property that 1≤k<p G(Yk ) = exp[α(V (Xp−1 )−V (X0 ))] thereby ensuring

that the Feynman-Kac measures defined in equations (6.6) and (6.7) are

more simple to analyze.

4.4.2 Algorithm on the Structural Model

Our algorithm is built with the weight function defined in Equation (4.11).
(i)
Initialization. We start with M identical copies, X̂0 , 1 ≤ i ≤ M, of the

initial condition X0 . That is,

(i)
X̂0 = (S1 (0), S1 (0), 0, S2(0), S2 (0), 0, · · · , SN (0), SN (0), 0), 1 ≤ i ≤ M

(i) (i) (i)


We also have a set of “parents”, Ŵ0 , defined by Ŵ0 = X̂0 . We denote
def. (i) (i) (i)
V0 = V (Ŵ0 ). This forms a set of M particles (Ŵ0 , X̂0 ), 1 ≤ i ≤ M.
(i) (i)
Now suppose that at time p, we have the set of M particles (Ŵp , X̂p ),

1 ≤ i ≤ M.

Selection Stage

82
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

We first compute the normalizing constant,

1 
M #  $
η̂pM = exp α V (X̂p(i) ) − V (Ŵp(i) ) (4.12)
M i=1

Then, we choose independently M particles according to the empirical dis-

tribution,

1 
M #  $
ηpM (dW̌ , dX̌) = exp α V (X̂p ) − V (Ŵp ) × δ(Ŵp(i) ,X̂p(i) ) (dW̌ , dX̌)
(i) (i)
M η̂pM i=1
(4.13)
(i) (i)
The particles that are selected are denoted (W̌p , X̌p ).

Mutation Stage
(i) (i)
For each of the selected particles, (W̌p , X̌p ), we apply an Euler scheme
(i) (i)
from time tp to time tp+1 with step size t for each X̌p so that X̌p becomes
(i) (i) (i)
X̂p+1 . We then set Ŵp+1 = X̌p . It should be noted, that each of the particles

are evolved independently and that the true dynamics (given in equation (4.3))

of Xp are applied rather than some other measure. It is this fact that separates

IPS from IS.

Then let,

N
f (X̂n(i) ) = 1{min (i)
Sj (u)≤Bj }
u≤T
j=1

83
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

denote the number of firms that have defaulted by time T for the ith particle.

Then, the estimator for P (L(T ) = j) = pj (T ), 1 ≤ j ≤ N, is given by


  n−1 
1 M 
PjM (T ) = 1f (X̂n(i) )=j exp(−α(V (Ŵn(i) ) − V0 )) × η̂pM (4.14)
M i=1 p=1

This estimator is unbiased in the sense that E[PjM (T )] = pj (T ). The unbiased-

ness follows directly from Lemma 4.3.1.

4.4.3 Note on Algorithm for Pricing Collateralized Debt

Obligations

From the above formulation, we can immediately see how the pricing of the

CDO occurs with IPS. Suppose that {T1 , . . . , TK } is the set of coupon dates.

Recall from Chapter 2 that the spread for each tranche is given by functions

of different expectations of functionals of the loss function at the coupon dates.

Therefore, it suffices to calculate the probability density function of losses at

each coupon date. That is, we need to calculate, for each 1 ≤ k ≤ K

P (L(Tk ) = j), 1≤j≤N

Therefore, our solution is to replace P (L(Tk ) = j) with the IPS approxima-

tion PjM (Tk ) formed above.

84
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

At first glance, this may not seem to be that novel. However, as mentioned

before, IPS is dynamically consistent in time. This means that we can use the

same simulation to compute PjM (T1 ) as we use to compute PjM (TK ). This is in

direct contrast to the most used form of variance reduction; importance sam-

pling. In most applications of importance sampling, there is a dependence on

the maturity time T through the Girsanov change of measure. Therefore, given

a collection of dates {T1 , . . . , TK }, a separate simulation is required for each

maturity. However, since IPS simulates under the original measure, there is no

inconsistency in using the same simulation to compute all maturities. There-

fore, we run the IPS algorithm until time T1 and form the vector (PjM (T1 ))1≤j≤N

and then continue the simulation and IPS algorithm until time T2 . We continue

this way until time TK .

4.5 Variance Analysis: Single-Name Case

While the unbiasedness of PjM (T ) is given by Lemma 4.3.1, it says nothing

of the variance of the estimator. Since there are many such unbiased estima-

tors, the key is analyzing the variance of the estimator. To analyze the variance

of equation (4.14), the reader would need to know the multidimensional dis-

tribution of the stock prices and their running minimum. As mentioned, this

85
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

distribution is not known. Therefore, we analyze the more simple case where

we only have one name.

Therefore, we take N = 1, with constant barrier B and are interested in

computing, using IPS, the probability of default before maturity T . That is, we

compute

P B (0, T ) = P(τ ≤ T ) = P(min S(u) ≤ B)


u≤T

Of course, we have an explicit formula for P B (0, T ) defined in equation (4.2).

However, we are interested in comparing IPS and MC in this case precisely

because we have an explicit formula for which to compare our results. The

more general case, where N is large and the names are correlated, will just be

analyzed numerically as there aren’t even formulas for traditional MC. It should

be noted that we are only interested in values of B that make the above event

rare.

We remark that it is a standard result that the variance associated with

the traditional MC method for computing P B (0, T ) is P B (0, T )(1 − P B (0, T )).

That is, the variance for traditional MC is of order P B (0, T ).

We also remark that the Markov chain (Xp )0≤p≤n defined in Section 4.4

simplifies to

Xp = (S(tp ), min S(u), 1τ ≤tp )


u≤tp

86
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

Then, following the setup described in Section 4.3, we see that the rare event

probability defined above has the following Feynman-Kac representation:

P B (0, T ) = γn (Ln(B) (1))

(B)
Where Ln (1) is given by the weighted indicator function defined for any path

yn = (x0 , . . . , xn ) ∈ Fn by


L(B) (B)
n (1)(yn ) = Ln (1)(x0 , . . . , xn ) = 1{minu≤tn S(u)≤B} G−
p (x0 , . . . , xp )
1≤p<n

= 1{minu≤tn S(u)≤B} e−α(V (xn−1 )−V (x0 ))

= 1{minu≤tn S(u)≤B} e−α(log(minu≤tn−1 S(u)/S0 )) (4.15)

and recalling that tn = T by the initial assumptions. Also, notice that


(B)
||Ln (1)(yn )|| ≤ 1 since log(minu≤tn−1 S(u)/S0) ≤ 0 and −α > 0 by assumption.

Therefore, if we let

B
PM (0, T ) = γnM (L(B) M (B)
n (1)) = ηn (Ln (1)) ηpM (Gp ) (4.16)
1≤p<n
B
then we see that by Lemma 4.3.1, PM (0, T ) is an unbiased estimator of P B (0, T )

such that
M →∞
B
PM (0, T ) −→ P B (0, T ) a.s.

While many estimators are unbiased, the key to determining the efficiency of

our estimator is to look at it’s variance. As such, we have the following central

limit theorem for our estimator.

87
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

Theorem 4.6. The estimator PBM (0, T ) given in equation (4.16) is unbiased,

and it satisfies the central limit theorem

√ M →∞
M E[PM
B
(0, T ) − P B (0, T )] −→ N(0, σnB (α)2 )

with the asymptotic variance

σnB (α)2
n # %
 & % & $
α log(minu≤tp−1 S(u)) −α log(minu≤tp−1 S(u))
= E e × E PB,p,n e
2
− P (0, T )
B 2

p=1

(4.17)

where PB,p,n is the collection of functions defined by

' (
PB,p,n (x) = E 1minu≤T S(u)≤B |Xp = x

and P B (0, T ) is defined in equation (4.2)

Proof. The proof follows directly by applying Theorem 2.3 in [12] with the

weight function that we have defined in equation (4.11). However, we present

the proof for our specific case for completeness.

First, notice that we have the following representation

√ √
B
M [PM (0, T ) − P B (0, T )] = M [γnM (Ln(B) (1)) − P B (0, T )]

= Wnγ,M (Ln(B) (1))

88
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

(B)
where Ln (1) is the weighted indicator function defined in Equation 4.15.

Recall that Wnγ,M (fn ) is defined in Equation 3.11 to be


n
Wnγ,M (fn ) = γpM (1)WpM (Qp,n fn )
p=1
(B)
Therefore, taking fn = Ln (1) in Equation 3.11 and Equation 3.12 we have
(B)
that Wnγ,M (Ln (1)) converges in law, as N goes to infinity, to a centered Gaus-
(B)
sian random variable Wnγ (Ln (1)) with variance

σnB (α)2 = E(Wnγ (L(B) 2


n (1)) )


n

n (1)) − ηp Qp,n (Ln (1))] )


γp (1)2 ηp ([Qp,n (L(B) (B) 2
=
p=1

Now, recalling that

 

Qp,n (fn )(yp ) = E fn (Yn ) Gk (Yk )|Yp = yp
p≤k<n

(B)
and substituting fn = Ln (1) and

 
Gk (Yk ) = eα log(minu≤tk S(u)/ minu≤tk−1 S(u)) (4.18)
p≤k<n p≤k<n

= eα log(minu≤tn−1 S(u)/ minu≤tp−1 S(u)) (4.19)

we find that

89
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

 
−α log(minu≤tp−1 S(u)/S0 )
Qp,n (L(B)
n (1))(y p ) = E 1{minu≤tn S(u)≤B} e |Y p = y p



= e−α log(minu≤tp−1 S(u)/S0 ) E 1{minu≤tn S(u)≤B} |Yp = yp


= e−α log(minu≤tp−1 S(u)/S0 ) E 1{minu≤tn S(u)≤B} |Xp = xp

= e−α log(minu≤tp−1 S(u)/S0 ) P(min S(u) ≤ B|Xp = xp )


u≤tn

where we used the Markov property on the second to last line. Recall that

Xp = (S(tp ), minu≤tp S(u), 1τ ≤tp ).

Now, using the definition of ηp , we find that

ηp (Qp,n (Ln(B) (1))) = P(min S(u) ≤ B)/γp (1)


u≤tn

= P B (0, T )/γp(1)

In addition, we have that


γp (1) = E( Gk (Yk ))
1≤k<p
 
 α log(minu≤tk S(u)/ minu≤tk−1 S(u))
= E e
1≤k<p
 
= E eα log(minu≤tp−1 S(u)/S0 )

Therefore, combining the above results and plugging them into the formula

for σnB (α)2 we have

90
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

σnB (α)2

n

n (1)) − ηp Qp,n (Ln (1))] )


γp (1)2 ηp ([Qp,n (L(B) (B) 2
=
p=1
n
= γp (1)2 ηp ([e−α log(minu≤tp−1 S(u)/S0 ) P(min S(u) ≤ B|Xp = xp ) − P B (0, T )/γp (1)]2 )
u≤tn
p=1
n  
−2α log(minu≤tp−1 S(u)/S0 )
= γp (1) 2
ηp (e P(min S(u) ≤ B|Xp = xp ) ) − P (0, T ) /γ (1)
2 B 2 2
u≤tn
p=1
n    
−α log(minu≤tp−1 S(u)/S0 )
= γp (1)E e P(min S(u) ≤ B|Xp = xp ) − P (0, T )
2 B 2
u≤tn
p=1
n #    
α log(minu≤tp−1 S(u)/S0 ) −α log(minu≤tp−1 S(u)/S0 )
= E e E e P(min S(u) ≤ B|Xp = xp ) 2
u≤tn
p=1
$
−P B (0, T )2
n # %
 & % & $
α log(minu≤tp−1 S(u)) −α log(minu≤tp−1 S(u))
= E e × E PB,p,n e
2
− P (0, T )
B 2

p=1

where PB,p,n is the collection of functions defined by

' (
PB,p,n (x) = E 1minu≤T S(u)≤B |Xp = x

and P B (0, T ) is defined in Equation (4.2).

The above theorem gives the asymptotic variance of the single-name estima-

tor in terms of expectations. The following corollary provides a more analytical

expression for the variance that will be useful to compare with traditional MC.

91
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

Corollary 4.6.1. Given the choice of weight function in Equation (4.11) and

any α < 0, and constant barrier B, we have that

σnB (α)2 (4.20)


   

n ( σ1 ) ln(B/S0 ) 0 ∞
− 12 θ 2 tp
= f (α, tp−1)e e−ασx+θz Ψ(tp−1 ,tp ) (x, y, z)dzdxdy
p=1 −∞ y y
 ( σ1 ) ln(B/S0 )  ∞
+ e−ασx+θz Υ(tp−1 ,tp ) (x, z)dzdx
−∞ x
 0  0 ∞
+ h(tp , z)2 e−ασx+θz Ψ(tp−1 ,tp ) (x, y, z)dzdxdy
( σ1 ) ln(B/S0 ) y y
  
0 ∞
+ h(tp , z)2 e−ασx+θz Υ(tp−1 ,tp ) (x, z)dzdx
( σ1 ) ln(B/S0 ) x
  1− 2r2 2 
S0 σ
− 1 − N(d+
2 (0, 0)) + N(d−
2 (0, 0))
B
(4.21)

where,

f (α, tp−1)
 √   √ 
ασ + θ ασ(ασ+2θ)tp−1 /2 (ασ + θ) tp−1 θ −θ tp−1
= e Erfc √ + Erfc √
ασ + 2θ 2 ασ + 2θ 2
 ∞
2 2
Erfc(x) = √ e−v dv
π x
r − 12 σ 2
θ =
σ

92
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

Ψ(tp−1 ,tp ) (x, y, z)


√ 2
2 2e−(2x−2y+z) /2tp
= ) ×
πtp t2p−1 (tp − tp−1 )2
 (1)
σ(tp−1 ,tp ) (μ(tp−1 ,tp ) (x, y, z) − x + z − 2y) −(x−μ(1) (x,y,y))2 /2σ(t
2
√ e (tp−1 ,tp ) p−1 ,tp )



(1)
+ σ(t2 p−1 ,tp ) + (μ(tp−1 ,tp ) (x, y, z))2
   
 x − μ
(1)
(x, y, z)
(1) (t ,t
p−1 p )
+μ(tp−1 ,tp ) (x, y, z)(z − 2y − 2x) − 2x(z − 2y) 1−Φ
σ(tp−1 ,tp )

Υ(tp−1 ,tp ) (x, z)


*  
−(2x−z)2 /2tp σ(tp−1 ,tp ) −(x−μ(tp−1 ,tp ) (x,z)) /2σ(tp−1 ,tp )
(2)
2 2 2
= × e √ e
πtp tp−1 2 2π

  x − μ(2) (x, z) 
(2) (t ,t
p−1 p )
+(μ(tp−1 ,tp ) (x, z) − 2x) 1 − Φ
σ(tp−1 ,tp )

2 σ(tp−1 ,tp ) −(x−μ(3) (x,z))2 /2σ(t
2
−e−z /2tp √ e (tp−1 ,tp ) p−1 ,tp )



  x − μ(3) (x, z) 
(3) (tp−1 ,tp )
+(μ(tp−1 ,tp ) (x, z) − 2x) 1 − Φ
σ(tp−1 ,tp )

93
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

 x
1 2
Φ(x) = √ e−y /2 dy
−∞ 2π
(1) 2x(tp − tp−1 ) + (2y − z)tp−1
μ(tp−1 ,tp ) (x, y, z) =
tp
(2) 2x(tp − tp−1 ) + ztp−1
μ(tp−1 ,tp ) (x, z) =
tp
(3) 2xtp − ztp−1
μ(tp−1 ,tp ) (x, z) =
tp
tp−1
σ(t2 p−1 ,tp ) = (tp − tp−1 )
tp

and,
  1− 2r2 
S0 eσz σ
h(tp , z) = 1 − N(d+
2 (tp , z)) + N(d−
2 (tp , z))
B
± (ln(S0 ) + σz − ln(B)) + (r − 12 σ 2 )(T − tp )

2 (tp , z) = 
σ T − tp
1 2 )u+σW
Proof. First, recall that since S(u) = S0 e(r− 2 σ u
, σ > 0 by assumption,

and log is an increasing function (so we can interchange the order of log and

min) we have

 „ «
% & 1 2
α log minu≤tp−1 S0 e(r− 2 σ )u+σWu )
α log(minu≤tp−1 S(u))
E e = E e
% &
α(minu≤tp−1 (r− 12 )u+σWu )
= S0α E e
% cu )
&
ασ(minu≤tp−1 W
= S0α E e

94
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

where Wu = θu + Wu is a Brownian motion with deterministic drift θ =


 1 2
r− 2 σ
σ
. Therefore, the above computation simplifies to computing the moment

generating function of the running minimum of Brownian motion with drift.

This formula is well known (see for instance [4]) and so we have,

% &
α log(minu≤tp−1 S(u))
E e
  √   √ 
ασ + θ ασ(ασ+2θ)tp−1 /2 (ασ + θ) tp−1 θ −θ tp−1
= S0 α
e Erfc √ + Erfc √
ασ + 2θ 2 ασ + 2θ 2
:= S0α f (α, tp−1) (4.22)

where,
 ∞
2 2
Erfc(x) = √ e−v dv
π x
% &
Now, we will compute E PB,p,n
2
e−α log(minu≤tp−1 S(u)) . The general goal will

u . First, we
be to write everything in terms of expectations of functionals of W

note that

' (
PB,p,n (x) = E 1minu≤T S(u)≤B |Xp = x
+ ,
= E 1minu≤T S(u)≤B |Xp = (min S(u), S(tp ), 1τ ≤tp )
u≤tp

= 1minu≤tp S(u)≤B + 1minu≤tp S(u)>B P( min S(u) ≤ B|S(tp ))


tp ≤u≤T
  1− 2r2 
S(t ) σ
N(d−
p
= 1minu≤tp S(u)≤B + 1minu≤tp S(u)>B 1 − N(d+ 2)+ 2)
B

95
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

where,
 
S(tp )
± ln B
+ (r − 12 σ 2 )(T − tp )

2 = 
σ T − tp

In the formula for d±


2 we will find it useful to substitute the formula S(tp ) =

c tp explicitly as,


S0 eσWtp and write the dependence on tp and W

 c

σWtp
± ln + (r − 12 σ 2 )(T − tp )
S0 e
B
d±  
2 (tp , Wtp ) =
σ T − tp
 

± ln(S0 ) + σ Wtp − ln(B) + (r − 12 σ 2 )(T − tp )
= 
σ T − tp

c
In addition, we also substitute S(u) = S0 eσWu into the expression for PB,p,n

and rearrange to get,

PB,p,n (x) = 1minu≤t cu ≤( 1 ) ln(B/S0 )


W
p σ

+1minu≤t 1 − N(d+ 
p
cu >( 1 ) ln(B/S0 )
W 2 (tp , Wtp ))
σ

 ctp
1− 2r2 
σW σ
S0 e 
+ N(d−
2 (tp , Wtp ))
B

Hence,

PB,p,n (x)2 = 1minu≤t cu ≤( 1 ) ln(B/S0 ) + 1minu≤t 


cu >( 1 ) ln(B/S0 ) h(tp , Wtp )
2
p W σ p W σ

96
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

where

⎛  1− 2r2 ⎞
ctp
σW σ
tp ) = ⎝1 − N(d+ (tp , W
tp )) + S0 e  ⎠
h(tp , W 2 N(d−
2 (tp , Wtp ))
B
% &
−α log(minu≤tp−1 S(u))
Hence, plugging in the expression for 2
PB,p,n into E PB,p,n e
2

we have,

% &
E PB,p,n
2
e−α log(minu≤tp−1 S(u))
% cu
&
−ασ minu≤tp−1 W
= S0−α E 1minu≤t Wcu ≤( ) ln(B/S0 )
1 e
 p σ

c
+S0−α E 1minu≤t W  2 −ασ minu≤tp−1 Wu
cu >( 1 ) ln(B/S0 ) h(tp , Wtp ) e (4.23)
p σ

where the expectation above is taken with respect to the measure P for which

u is a Brownian motion with drift. Recall that under P,


W

t = θdt + dWt
dW
 
r− 12 σ2
where Wt is a P standard Brownian motion and θ = σ
.

t is a standard Brownian mo-


Using Girsanov’s theorem (see [24] or [28]), W

 and the Radon-Nikodym derivative, Z(t), is given by


tion under P

97
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

+  t  ,
1 t 2
Z(t) = exp − θdWu − θ du
0 2 0
+  t ,
 1 2
= exp − θ(dWu − θu) − θ t
2
+ 0
,
t + 1 θ2 t
= exp −θW
2

 to get
Therefore, we rewrite Equation (4.23) as an expectation under P

% &
E 2
PB,p,n e−α log(minu≤tp−1 S(u))
% cu
&
= 
S0−α E 1minu≤t Wcu ≤( 1 ) ln(B/S0 ) e
−ασ minu≤tp−1 W
Z(tp ) −1

 p σ

cu
 1
+S0−α E 
cu >( 1 ) ln(B/S0 ) h(tp , Wtp )
2 −ασ minu≤tp−1 W
e Z(tp )−1
minu≤t p W σ
% cu θ W
&
 1 −ασ minu≤tp−1 W ctp − 1 θ 2 tp
= S0−α E cu ≤( 1 ) ln(B/S0 ) e
minu≤tp W e 2

 σ

cu θ W
c
+S −α E 1 c 1
tp )2 e−ασ minu≤tp−1 W
h(tp , W
1 2
e tp − 2 θ tp
0 minu≤tp Wu >( σ ) ln(B/S0 )
  
− 12 θ 2 tp
= S0−α e 1y≤( 1 ) ln(B/S0 ) e−ασx+θz Γ(tp−1 ,tp ) (dx, dy, dz)
σ

   
+ 1y> 1 ln(B/S0 ) h(tp , z)2 e−ασx+θz Γ(tp−1 ,tp ) (dx, dy, dz) (4.24)
σ

where,

 min W
Γ(tp−1 ,tp ) (dx, dy, dz) = P( u ∈ dx, min W tp ∈ dz)
u ∈ dy, W
u≤tp−1 u≤tp

98
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

 The
u is a standard Brownian motion under P.
and as stated before, W

formula for Γ(tp−1 ,tp ) (dx, dy, dz) is given by Lemma A.0.3 in the Appendix and

is,

Γ(tp−1 ,tp ) (dx, dy, dz) = Ψ(tp−1 ,tp ) (x, y, z)1y<x,y≤z,x≤0dzdydx

+Υ(tp−1 ,tp ) (x, z)1x≤z,x≤0 δx (dy)dzdx

where,

Ψ(tp−1 ,tp ) (x, y, z)


√ 2
2 2e−(2x−2y+z) /2tp
= ) ×
πtp t2p−1 (tp − tp−1 )2
 (1)
σ(tp−1 ,tp ) (μ(tp−1 ,tp ) (x, y, z) − x + z − 2y) −(x−μ(1) (x,y,y))2 /2σ(t
2
√ e (tp−1 ,tp ) p−1 ,tp )



(1)
+ σ(t2 p−1 ,tp ) + (μ(tp−1 ,tp ) (x, y, z))2
   
 (1)
x − μ(tp−1 ,tp ) (x, y, z)
(1)
+μ(tp−1 ,tp ) (x, y, z)(z − 2y − 2x) − 2x(z − 2y) 1−Φ
σ(tp−1 ,tp )

99
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

Υ(tp−1 ,tp ) (x, z)


*  
−(2x−z) /2tp σ(tp−1 ,tp ) −(x−μ(tp−1 ,tp ) (x,z)) /2σ(tp−1 ,tp )
(2)
2 2 2 2
= × e √ e
πtp tp−1 2 2π

  x − μ(2) (x, z) 
(2) (tp−1 ,tp )
+(μ(tp−1 ,tp ) (x, z) − 2x) 1 − Φ
σ(tp−1 ,tp )

2 σ(tp−1 ,tp ) −(x−μ(3) (x,z))2 /2σ(t
2
−e−z /2tp √ e (tp−1 ,tp ) p−1 ,tp )



  x − μ(3) 
(3) (tp−1 ,tp ) (x, z)
+(μ(tp−1 ,tp ) (x, z) − 2x) 1 − Φ
σ(tp−1 ,tp )

and,

(1) 2x(tp − tp−1 ) + (2y − z)tp−1


μ(tp−1 ,tp ) (x, y, z) =
tp
(2) 2x(tp − tp−1 ) + ztp−1
μ(tp−1 ,tp ) (x, z) =
tp
(3) 2xtp − ztp−1
μ(tp−1 ,tp ) (x, z) =
tp
tp−1
σ(t2 p−1 ,tp ) = (tp − tp−1 )
tp

Then, Equation 4.24 can be written as,

100
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

% &
E PB,p,n
2
e−α log(minu≤tp−1 S(u))
  
1 2
= S0−α e− 2 θ tp 1y≤( 1 ) ln(B/S0 ) e−ασx+θz Γ(tp−1 ,tp ) (dx, dy, dz)
σ

   
+ 1y> 1 ln(B/S0 ) h(tp , z)2 e−ασx+θz Γ(tp−1 ,tp ) (dx, dy, dz)
σ

1 2
= S0−α e− 2 θ tp
  
× 1y≤( 1 ) ln(B/S0 ) e−ασx+θz Ψ(tp−1 ,tp ) (x, y, z)1y<x,y≤z,x≤0dzdydz
σ

  
+ 1y≤( 1 ) ln(B/S0 ) e−ασx+θz Υ(tp−1 ,tp ) (x, z)δx (dy)1y≤z,x≤0dzdx
σ
  
+ 1y> 1 ln(B/S0 ) h(tp , z)2 e−ασx+θz Ψ(tp−1 ,tp ) (x, y, z)1y<x,y≤z,x≤0dzdydz
σ

   
+ 1y> 1 ln(B/S0 ) h(tp , z)2 e−ασx+θz Υ(tp−1 ,tp ) (x, z)δx (dy)1y≤z,x≤0dzdx
σ

1 2
= S0−α e− 2 θ tp
 1
( σ ) ln(B/S0 )  0  ∞
× e−ασx+θz Ψ(tp−1 ,tp ) (x, y, z)dzdxdy
−∞ y y
 ( σ1 ) ln(B/S0 )  ∞
+ e−ασx+θz Υ(tp−1 ,tp ) (x, z)dzdx
−∞ x
 0  0 ∞
+ h(tp , z)2 e−ασx+θz Ψ(tp−1 ,tp ) (x, y, z)dzdxdy
( σ1 ) ln(B/S0 ) y y
  
0 ∞
+ h(tp , z)2 e−ασx+θz Υ(tp−1 ,tp ) (x, z)dzdx
( σ1 ) ln(B/S0 ) x

(4.25)

Therefore, substituting Equations 4.22 and 4.25 and our equation for P B (0, T )

into Equation 4.17 of Theorem 4.6 we have,

101
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

σnB (α)
   

n ( σ1 ) ln(B/S0 ) 0 ∞
− 12 θ 2 tp
= f (α, tp−1)e e−ασx+θz Ψ(tp−1 ,tp ) (x, y, z)dzdxdy
p=1 −∞ y y
 ( σ1 ) ln(B/S0 )  ∞
+ e−ασx+θz Υ(tp−1 ,tp ) (x, z)dzdx
−∞ x
 0  0 ∞
+ h(tp , z)2 e−ασx+θz Ψ(tp−1 ,tp ) (x, y, z)dzdxdy
( σ1 ) ln(B/S0 ) y y
  
0 ∞
+ h(tp , z)2 e−ασx+θz Υ(tp−1 ,tp ) (x, z)dzdx
( σ1 ) ln(B/S0 ) x
  1− 2r2 2 
S0 σ
− 1 − N(d+
2 (0, 0)) + N(d−
2 (0, 0))
B

While it is possible to construct many different unbiased estimators, the

critical quantity to consider is the variance of each estimator. We derived an

explicit (up to a multiple integrals) expression (Equation 4.21) for the vari-

ance of the IPS estimator with our particular choice of weight function. Our

weight function, Equation 4.11, puts more emphasis on the path-particles whose

minimums have decreased the most between two time steps.

Since we have an explicit formula for the variance of our IPS estimator, we

can compare the variance of performing the selection and mutation algorithm

with performing traditional MC. As stated in previous sections, the variance of

the MC estimate of the single-name default probabilities is given by P B (0, T ) ×

102
Chapter 4. Credit Derivatives Pricing with Interacting Particle Systems

(1 − P B (0, T )) which is of order (P B (0, T )). The formula for P B (0, T ) is given

in Equation 4.2.

To compute the integrals appearing in Equation 4.21 and hence the vari-

ance of IPS, we relied on numerical integration techniques. Specifically, we used

the functions triplequad.m and dblquad.m in Matlab which uses adaptive

Simpson quadrature to evaluate numerical integrals over a rectangular region.

However, the integrals appearing in Equation 4.21, as written, don’t have rect-

angular regions of integration. Therefore, prior to using these functions, we

transformed the integrals appearing in Equation 4.21 into a sequence of in-

tegrals over rectangular regions before using triplequad.m and dblquad.m.

The numerical results of comparing the asymptotic variance of IPS and pure

MC are presented in Section 5.2.

103
Chapter 5

Numerical Implementation and


Results

5.1 Introduction

In this chapter, we investigate numerically the results of implementing the

IPS procedure for estimating the probability mass function of the loss function

for single names and multinames. In addition, we demonstrate numerically that

IPS performs better than traditional Monte Carlo for computing rare event

probabilities in a credit derivatives framework with a substantial reduction in

variance.

5.2 Single Name

For the single name case, we compute the probability of default for different

values of the barrier using IPS and traditional MC. In addition, for each method,

104
Chapter 5. Numerical Implementation and Results

we implemented the continuity correction for the barrier level described in [7]

to account for the fact that we using a discrete approximation to the continuous

barrier for both IPS and MC. For the different values of the barrier we use, we

can calculate the exact probability of default from Equation (4.2).

The following are the parameters we used for both IPS and MC.

α r σ S0 t T n M
-18.5 .06 .25 80 .001 1 20 20000

In addition, we also used 20000 simulations in the traditional MC. The

results are shown in Figure 5.1.

One can see that IPS is capturing the rare event probabilities for the single

name case whereas traditional MC isn’t able to capture these values.

Indeed probabilities of order 10−14 may be irrelevant in the context of de-

fault probabilities but the user can see that IPS is capturing the rare events

probabilities for the single name case whereas traditional Monte Carlo is not

able to capture these values below 10−4 .

In Figure 5.2, we show how the variance decreases with the barrier level,

and therefore with the default probability for MC and IPS. In the IPS case, the

variance is obtained empirically and using the integral formulas derived in the

Appendix. We deduce that the variance for IPS decreases as p2 (p is the default

probability), as opposed to p in the case of MC simulation.

105
Chapter 5. Numerical Implementation and Results

MC vs IPS
0
10

2
10

IPS
4 MC
10
True Value
Probabilitiy of Default

6
10

8
10

10
10

12
10

14
10
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Barrier/S0

Figure 5.1: Default probabilities for different barrier levels for IPS and MC

Each MC and IPS simulation gives an estimate of the probability of default

(whose theoretical value doesn’t depend on the method) as well as an estimate

of the standard deviation of the estimator (whose theoretical value does depend

on the method). Therefore, it is instructive from a practical point of view to

compare the two methods by comparing the empirical ratios of their standard

deviation to the probability of default for each method. If P B (0, T ) is the

106
Chapter 5. Numerical Implementation and Results

Comparison of MC with IPS


5
10

0
10

5
10

10
10
Variance

IPS Empirical Variance


IPS Theoretical Variance
15
10 MC Theoretical Variance
MC prob. default squared

20
10

25
10

30
10
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
B/S0

Figure 5.2: Variances for different barrier levels for IPS and MC

probability of default for a certain barrier level B, then the standard deviation,

P2 (B), for traditional MC is given by,

 
P2MC (B) = P B (0, T ) × (1 − P B (0, T ))

and the theoretical ratio for MC is given by



P2MC (B) (1 − P B (0, T ))
= 
P B (0, T ) P B (0, T )

107
Chapter 5. Numerical Implementation and Results

We have formulas given in equation (4.2) to calculate the above exactly.

For IPS, the corresponding ratio is

P2IPS (B) σnB (α)


=
P B (0, T ) P B (0, T )

where σnB (α) is given in Corollary 4.6.1.

For both methods, we will use the theoretical formulas as well as empirically

derived quantities.

In Figure 5.3, the reader sees that there are specific regimes where it is

more efficient to use IPS as opposed to traditional MC for certain values of the

barrier level. This is to be expected because IPS is well suited to rare event

probabilities whereas MC isn’t.

108
Chapter 5. Numerical Implementation and Results

Standard Deviation to Probability Ratio for MC and IPS


7
10
IPS Theoretical
IPS Empirical
6
10 MC Theoretical
MC Emprical

5
10

4
10
p2(x)/p(x)

3
10

2
10

1
10

0
10
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
B/S0

Figure 5.3: Standard Deviation to Probability ratio for MC and IPS

5.3 Multi-Name

For the multi-name case, we tested using 25 firms (N = 25). In addition,

we took all of the firms to be homogenous, meaning that they have the same

parameters, starting value, and default barrier in equation (4.3). The following

109
Chapter 5. Numerical Implementation and Results

are the parameters that we used,

α r σi Si (0) Bi t T n M
-18.5/25 .06 .3 90 36 .001 1 20 10000

In addition, we took the correlation between the driving Brownian motions

to be ρij = .4 for i = j. The above parameters give, in the independent case,

a probability of default of .0018, a realistic default probability for highly-rated

firms. For the MC simulation we also used 10, 000 simulations.

The following two pictures illustrate the difference between using MC and

IPS to estimate the loss probability mass function. That is, we calculate nu-

merically,

P (L(T ) = k)

where L(T ) is the number of firms that have defaulted before time T defined

in equation (6.3). Figure 5.4 shows the shape of the pmf. Here, one sees that

on the linear scale, it is hard to distinguish IPS from MC. However, Figure 5.5

is much more interesting because the pmf is on a log scale (in y) and one can

see that the IPS method is picking up more of the tail events for the pmf of

the loss function and the distinction between IPS and MC becomes clear. Also,

one can see that for this regime, MC is only good for estimating the pmf for

K = 0, 1, 2, 3, 4, but that IPS is good for estimating the pmf for K = 0, 1, . . . , 10.

Considering the fact that most contracts are written only on the first 40% of

110
Chapter 5. Numerical Implementation and Results

losses and 10 = .4 × 25, we see that IPS does a good job of describing the rare,

but economically (in the sense of contracts) significant events.

Density function for Loss given by different numerical procedures

0.8 MC
IPS
0.7

0.6

0.5
P(Loss = K)

0.4

0.3

0.2

0.1

0 5 10 15 20
Number of Defaults (K)

Figure 5.4: Probability mass function of the loss function described in equation
(6.3) with N = 25

111
Chapter 5. Numerical Implementation and Results

Density function for Loss given by different numerical procedures


0
10
MC
IPS
2
10

4
10
P(Loss = K)

6
10

8
10

10
10
0 2 4 6 8 10
Number of Defaults (K)

Figure 5.5: Probability mass function of the loss function described in equation
(6.3) with N = 25 in log scale

112
Chapter 6

Other Applications of
Interacting Particle Systems

6.1 Introduction

In this chapter, we present some applications of interacting particle systems

in the first passage model context. We focus on the first passage model to

continue the framework developed in earlier chapters. Specifically, we will look

at two applications of interacting particle systems to risk management. These

two risk management metrics that we will consider are portfolio Value-at-Risk

(VaR) and expected shortfall calculations.

6.2 Portfolio Value at Risk

Suppose that we have a portfolio consisting of N names. For our purposes,

we consider firms behave under the first passage model. That is, we assume

113
Chapter 6. Other Applications of Interacting Particle Systems

that the firm values for the N names have the following dynamics,

dSi(t) = rSi (t)dt + σi Si (t)dWi (t) i = 1, . . . , N (6.1)

where r is the risk-free interest rate, σi is constant volatility, and the driving

innovations dWi(t) are infinitesimal increments of Wiener processes Wi with

correlation

d Wi , Wj t = ρij dt.

Each firm i is also assumed to have a deterministic boundary process Bi (t) and

default for firm i is given by

τi = inf {t : Si (t) ≤ Bi (t)} (6.2)

In addition, for each firm i there corresponds the amount of notional ni that’s

written on that firm in the portfolio. This is the amount of capital that is at risk

if the ith firm defaults. The notional is assumed to be constant throughout the

life of the portfolio. We assume a positive recovery rate ri (typically assumed

to be 40% in industry). Therefore, the portfolio loss at time T can be written

as


N
L(T ) = ci 1{τi ≤T } (6.3)
i=1

where

ci = (1 − ri )ni

114
Chapter 6. Other Applications of Interacting Particle Systems

We will consider the problem of estimating the 99% level Value-at-Risk

(VaR) for the above portfolio. That is, we would like to find x.99 such that

P (L(T ) ≥ x.99 ) = .01

The interpretation of x.99 is that it is the amount of money one can be sure

that the portfolio loss will not exceed with a probability of 99%.

The above problem is an inverse problem because we have to determine the

rare-level set [x.99 , ∞) for the fixed probability .01. Without explicit formulas

for the distribution of L(T ), we have to rely on numerical results to estimate

x.99 .

6.3 Review of Monte Carlo for VaR Calcula-

tion

The use of MC for the calculation of VaR levels is a very well developed

field and is straightforward to implement. We briefly describe the basics of

using traditional MC for VaR and refer the reader to [18] for a more complete

account of the various methods.

115
Chapter 6. Other Applications of Interacting Particle Systems

We assume that our portfolio is as described in the previous section. Let M

be the number of MC trajectories. We denote by X j (T ), 1 ≤ j ≤ M, the 1 × N

vector of simulated default times for the j th trajectory. That is,

X j (T ) = (τ1j , τ2j , . . . , τij , . . . , τNj )

where the default times above were simulated by using the dynamics in

Equation (4.3) and keeping track of the default time on a path by path basis.

Across each j, the simulated trajectories are independent so we that we have M

independent realizations coming from the joint distribution of (τ1 , τ2 , . . . , τN ).

For each X j (T ) we form the corresponding portfolio loss for that trajectory,


N
j
L (T ) = ci 1τ j ≤T
i
i=1

Notice that Lj (T ) above is a deterministic function of X j (T ).

Denote the reverse order statistics of Lj (T ) by L̄j (T ) so that

L̄1 (T ) ≥ L̄2 (T ) ≥ . . . ≥ L̄M (T )

Recall, for any given x,


1 
M
1 i
M i=1 L̄ (T )≥x

is an unbiased estimator of P(L(T ) ≥ x).

116
Chapter 6. Other Applications of Interacting Particle Systems

We can reverse the idea above to find x.99 . We set j  = .01 × M and by

setting x.99 = L̄j (T ) we have that

1 
M

P(L(T ) ≥ L̄j (T )) ∼ 1 i j
M i=1 L̄ (T )≥L̄ (T )
1
= .01 × M −→ .01
M

Therefore, the user sees that determining the VaR level consists in gener-

ating the independent losses, ordering them, and then picking the correct one

depending on the VaR level desired.

6.4 Interacting Particle System Interpretation

of Value at Risk

Following on the material written in previous chapters, the IPS interpreta-

tion for estimating

P(L(T ) ≥ x)

is to replace the standard MC procedure by the selection and mutation proce-

dure with a particular choice of weight function. Again, we would use the weight

function that’s given by the running minimum of the firm value. Therefore, we

117
Chapter 6. Other Applications of Interacting Particle Systems

fix some parameter α < 0 and define the potential function,

Gα (Yp ) = exp[α(V (Xp ) − V (Xp−1 ))] (6.4)

where

N
V (Xp ) = log(min Si (u))
u≤tp
i=1

Then, we run the mutation and selection algorithm with the weight function

above until time T . Then, we know that the unbiased estimator for P(L(T ) ≥ x)

is given by

  n−1 
1  
M
1Li (T )≥x exp(−α(V (Ŵn(i) ) − V0 )) × η̂pM (6.5)
M i=1 p=1

Let,
n−1 

ki = exp(−α(V (Ŵn(i) ) − V0 )) × η̂pM
p=1

Therefore, the IPS estimator can be written in the more compact version,

1 
M
ki 1Li (T )≥x
M i=1

Now, for each Li (T ) we have the corresponding weight ki. Given the set

(Li (T ), ki )1≤i≤M , we denote the reverse order statistics of Li (T ) and its corre-

sponding weight by (L̄i (T ), k̄i).

As before, given any x, the unbiased estimator of P(L(T ) ≥ x) is

1 
M
k̄i 1L̄i (T )≥x
M i=1

118
Chapter 6. Other Applications of Interacting Particle Systems

Now, to determine the VaR level we do the following. We form the vector

(j1 , j2 , . . . , jM ) where ji is defined as



i
ji = k̄p
p=1

and set

j  = min{i : ji ≥ .01 × M}


Then, our IPS estimate of VaR is x.99 = L̄j (T ).

We can see that

1  1

M
k̄i 1L̄i (T )≥L̄j (T ) = k̄1 + k̄2 + · · · + k̄j  −→ .01
M i=1 M

6.5 Expected Shortfall Calculation

Suppose we wish to compute the expected shortfall contribution,

E(L(T )|L(T ) ≥ a)

The a above may come from a VaR type analysis (i.e. a = x.99 ) or different

stress test scenarios.

To estimate the conditional expectation above using MC, we simulate inde-

pendently the loss events Li (T ), 1 ≤ i ≤ M. Again, we order the Li (T )’s into

119
Chapter 6. Other Applications of Interacting Particle Systems

M
L̄i (T ) and let j  = i=1 1L̄i (T )≥a . Then the unbiased estimator for the above

conditional expectation is

1
M
Li (T )1Li (T )≥a
E(L(T )|L(T ) ≥ a) ∼ M
1
i=1
M
M i=1 1Li (T )≥a
1
M i
M i=1 L̄ (T )1L̄i (T )≥a
= 1
M
M i=1 1L̄i (T )≥a

1  i
M
= L̄ (T )1L̄i (T )≥a
j  i=1
j
1  i
= L̄ (T )
j  i=1

We can immediately see a problem. The above routine is highly inefficient.

Suppose a = x.99 . Then if we use 10,000 MC trajectories, j  = .01×10000 = 100

by the definition of j  and the way we calculate VaR levels. Therefore, only 100

trajectories are used for our estimate, which results in a very large variance.

This problem is only exacerbated for even higher level of a’s. This problem is

due to the nature of expected shortfall. To calculate the expected shortfall, we

must investigate the behavior of the density in the tail. However, in using MC

simulations, reaching the tail of the distribution is itself very difficult, let alone

getting enough resolution once we do hit the tail.

120
Chapter 6. Other Applications of Interacting Particle Systems

One of the main culprits of the above shortcomings with using MC to es-

timate expected shortfall contributions is the direct relationship between the

probability of the rare event and the number of MC trajectories that exceed the

rare event. However, IPS doesn’t suffer from this problem because it twists the

empirical distribution of the particles so that more trajectories are in the tail.

6.5.1 Interacting Particle Systems’ Answer to Expected

Shortfall Contributions

IPS is well developed for answering questions about the conditional proba-

bilities of rare event regions. Del Moral and Garnier included the formulation

for calculating conditional probabilities in their paper [12].

Looking at the way L(T ) is defined, we can see that we want to put more

emphasis on the trajectories that see more defaults. In this way, we will be

looking at the tail events of L(T ), i.e. the regions where the portfolio loss is the

greatest. Therefore, we will use the same weight function that we have always

used.

Recall the Feynman-Kac measure,


 

γn (fn ) = E f (Yn ) Gk (Yk ) (6.6)
1≤k<n

121
Chapter 6. Other Applications of Interacting Particle Systems

We denote by ηn (·) the normalized measure defined as



E f (Yn ) 1≤k<n Gk (Yk )
ηn (fn ) =  = γn (fn )/γn (1) (6.7)
E 1≤k<n Gk (Yk )

The Markov process for our set up is the one we introduced earlier. That is,

our Markov process is the 3 × N dimensional process (X̃t )t∈[0,T ] defined as:

X̃t = (S1 (t), min S1 (u), 1τ1 ≤t , S2 (t), min S2 (u), 1τ2 ≤t , · · ·
u≤t u≤t

· · · , SN (t), min SN (t), 1τN ≤t )


u≤t

We introduce the chain (Xp )0≤p≤n = (X̃pT /n )0≤p≤n and the whole history of the

chain denoted by Yp = (X0 , . . . , Xp ).


N
In addition, we can easily see that L(T ) is bounded by i=1 ci which we

assume is finite. Then the rare event {L(T ) ≥ a} has the following Feynman-

Kac representation

P(L(T ) ≥ a) = γn (Ina (1))

where,


Ina (1)(y) = Ina (x0 , x1 , . . . , xn ) = 1L(T )≥a G−
p (x0 , x1 , . . . , xp )
1≤p<n

In fact, we can say something a little more general as well. Given any

bounded-measurable function φn , then

E(φn (X0 , . . . , Xn ); L(T ) ≥ a) = γn (Ina (φn ))

122
Chapter 6. Other Applications of Interacting Particle Systems

where,


Ina (φn )(x0 , x1 , . . . , xn ) = φn (x0 , . . . , xn )1L(T )≥a G−
p (x0 , x1 , . . . , xp )
1≤p<n

Recall that


n
γn+1(1) = γn (Gn ) = ηn (Gn )γn (1) = ηp (Gp )
p=1

Therefore, given any bounded measurable function fn , we have


γn (fn ) = ηn (fn ) ηp (Gp )
1≤p<n

Hence, we have the following representations


P(L(T ) ≥ a) = ηn (Ina (1)) ηp (Gp )
1≤p<n

E(φn (X0 , . . . , Xn ); L(T ) ≥ a) = ηn (Ina (φn )) ηp (Gp )
1≤p<n

E(φn (X0 , . . . , Xn )|L(T ) ≥ a) = ηn (Ina (φn ))/ηn (Ina (1))

Therefore, setting φn = L(T ), we see that the expected shortfall contribution

admits the following Feynman-Kac interpretation

E(L(T )|L(T ) ≥ a) = ηn (Ina (L(T )))/ηn (Ina (1))

where

123
Chapter 6. Other Applications of Interacting Particle Systems


Ina (L(T ))(x0 , . . . , xn ) = L(T )1L(T )≥a G−
p (x0 , . . . , xp )
1≤p<n

We can estimate this conditional expectation by replacing the correspond-

ing normalized Feynman-Kac measures (ηn ) with their corresponding particle-

approximation measures ηnM , where M denotes the number of particles. Del

Moral and Garnier state that

ηnM (Ina (φn ))/ηnM (Ina (1)) −→ ηn (Ina (φn ))/ηn (Ina (1)) = E(φn (X0 , . . . , Xn )|L(T ) ≥ a)

In fact, using Theorem 2.2 in Del Moral and Garnier, we have the following

theorem,

N
Theorem 6.6. Given that L(T ) = i=1 ci 1τi ≤T , the estimator ηnM (Ina (L(T )))/ηnM (Ina (1))

satisfies the central limit theorem


M ηnM (Ina (L(T )))/ηnM (Ina (1)) − E(L(T )|L(T ) ≥ a) −→ N(0, σn (a, L(T ))2 )

with the asymptotic variance


n
−2
σn (a, L(T )) = P(L(T ) ≥ a)
2
γp (1)γp− (1)ηp− (Pp,n (a, L(T ))2 )
p=1

and the collection of functions Pp,n (a, L(T )) defined by

Pp,n (a, L(T ))(x0 , . . . , xp )

= E[(L(T ) − E(L(T )|L(T ) ≥ a))1L(T )≥a |(X0 , . . . , Xp ) = (x0 , . . . , xp )]

124
Chapter 6. Other Applications of Interacting Particle Systems

6.6.1 Further Analysis of Interacting Particle Systems

Assume that we have followed the algorithm for IPS on the Markov chain

(Xp )0≤p≤n = (X̃pT /n )0≤p≤n

where

X̃t = (S1 (t), min S1 (u), 1τ1 ≤t , S2 (t), min S2 (u), 1τ2 ≤t , · · ·
u≤t u≤t

· · · , SN (t), min SN (t), 1τN ≤t )


u≤t

Assume that we are at time T. Then we have the M simulated values for

the chain XT denoted by XTi , 1 ≤ i ≤ M. Again, we form the portfolio loss for

each simulated particle,


N
i
L (T ) = cj 1τji ≤T
j=1

Therefore, our IPS estimate of E[L(T )|L(T ) ≥ a] is

ηnM (Ina (L(T )))/ηnM (Ina (1))


M (i)
1
M i=1 Li (T )1Li (T )≥a exp(−α(V (Ŵn ) − V0 ))
= M (i)
1
M
1Li (T )≥a exp(−α(V (Ŵn ) − V0 ))
i=1
M i ˆ (i)
i=1 L̄ (T )1L̄i (T )≥a exp(−α(V (W̄n ) − V0 ))
1
M
=
1
M ˆ n(i) ) − V ))
M
1 i
i=1 exp(−α(V (W̄
L̄ (T )≥a 0

125
Chapter 6. Other Applications of Interacting Particle Systems

However, we also have another way of interpreting the above expression. We

have that

1
M ˆ n(i) ) − V ))
M
L̄i (T )1L̄i (T )≥a exp(−α(V (W̄
i=1 0
M ˆ (i)
i=1 1L̄i (T )≥a exp(−α(V (W̄n ) − V0 ))
1
M
j  i ˆ (i)
i=1 L̄ (T )1L̄i (T )≥a exp(−α(V (W̄n ) − V0 ))
= j  ˆ n(i) ) − V ))
1 i
i=1 exp(−α(V (W̄
L̄ (T )≥a 0
j

= L̄i (T )pi1L̄i (T )≥a
i=1

Since for i ≤ j  we have L̄i (T ) ≥ a by definition of j  we have




j
ηnM (Ina (L(T )))/ηnM (Ina (1)) = L̄i (T )pi (6.8)
i=1

where
ˆ n(i) ) − V ))
exp(−α(V (W̄ 0
pi = j  , 1 ≤ i ≤ j
exp(−α(V ( ˆ n ) − V ))

(k)
k=1 0

and


j
pi = 1
i=1

or we can see the above equations as being


1  i
j
ηnM (Ina (L(T )))/ηnM (Ina (1)) =  L̄ (T )bi (6.9)
j i=1

where

bi = pi j 

126
Chapter 6. Other Applications of Interacting Particle Systems

6.6.2 Remarks

Equation (6.8) is interesting because it admits the representation of the exact

definition of an expectation. In addition, each of the L̄i (T )’s are simulated as

are the pi ’s.

Equation (6.9) is more like the equations you would see with traditional MC.

However, the important items to be noted here are the normalizing constants

bi that don’t appear in traditional MC, and the fact that j  is determined in

quite a different way under IPS than in MC.

In MC, we know that j  is given by P(L(T ) ≥ a) × M. That is, among

M i.i.d samples, only P(L(T ) ≥ a) × M will be in the rare level set [a, ∞).

This is precisely because all of the samples are i.i.d. and come from the original

distribution, for which it is difficult to simulate the tail. However, IPS is quite

different.

Recall that even though IPS is performed under the original measure, the

empirical distribution of the particles converge weakly to the distribution under

the twisted change of measure. In addition, recall that j  is defined as

j  = min{ji : ji ≥ P(L(T ) ≥ a) ∗ M}

127
Chapter 6. Other Applications of Interacting Particle Systems

where,

i
ji = k̄p
p=1
n−1 

ˆ (i) ) − V )) ×
k̄i = exp(−α(V (W̄ η̂pM
n 0
p=1

are the weights associated with the reverse ordered statistics Li (T ).

Suppose that P(L(T ) ≥ a) = .01. As we know, this is not a small probability

for IPS to handle and we would expect that the number of particles that exceed

a (i.e. Li (T ) ≥ a) would be very close to M. This would mean that j  is very

close to M. One can see that this could potentially have a profound impact.

It means that we will be using closer to the number of particles we started

with, instead of a substantially reduced number, to compute the conditional

expectations empirically. This will generate a very significant savings.

6.7 Numerical Example

In this section we consider a homogenous portfolio consisting of N = 10

names. We use the following parameters,

r σ S0 t T n M
.06 .3 90 .001 1 10 10000

128
Chapter 6. Other Applications of Interacting Particle Systems

In addition, we take ρij = .4. We set ci = 1000, 1 ≤ i ≤ 1000. Therefore,

the possible values for L(T ) are

L(T ) ∈ {0, 1000, 2000, . . . , 10000}

The barrier we use is Bi = B = 44.5263, which gives P(τi ≤ T ) = .01604

and P(L(T ) ≥ 1000) ∼ .1103.

Then we compute the estimates for E(L(T )|L(T ) ≥ a) for a ∈ {0, 1000, 2000, . . . , 10000}.

In the graph below we display the results. The estimates for MC go to zero (af-

ter a certain a) because none of the simulated trajectories entered that rare

event region.

The most interesting thing to look at is the j  value for MC and IPS for all

of the different a’s.


a Monte Carlo j  IPS j 
0 10000 10000
1000 1154 9598
2000 307 9034
3000 93 8392
4000 34 7548
5000 14 6612
6000 3 5644
7000 2 4634
8000 1 3548
9000 0 2371
10000 0 1241

The above table shows the substantial improvement one can expect in using

IPS for expected shortfall versus MC. Note that the number of trajectories for

129
Chapter 6. Other Applications of Interacting Particle Systems

Estimated Expected Shortfall Contributions  P(L(T) 1000) = .11


12000
IPS
MC

10000

8000
E(L(T) | L(T) a)

6000

4000

2000

0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
a

Figure 6.1: Estimated Expected Shortfall Contributions

IPS that yield a loss of 10,000 (the maximum) is 1241, where as the number

of trajectories for MC that exceed a loss of 1,000 is only 1154. While these

numbers will change from simulation run to simulation run, they do give one a

good idea of what to expect.

Next, we look at a particular case were defaults are more rare. Here we take

B = Bi = 35.05 so that P(τi ≤ T ) = .001 and P(L(T ) ≥ 1000) ∼ .0129 (which

130
Chapter 6. Other Applications of Interacting Particle Systems

is close to a .99% VaR level).

The graph and chart that follow show the results.

Estimated Expected Shortfall Contributions  P(L(T) 1000) = .0129


12000

10000

8000
E(L(T) | L(T) a)

6000

4000

2000

0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
a

Figure 6.2: Estimated Expected Shortfall Contributions with Rare


Probabilities

131
Chapter 6. Other Applications of Interacting Particle Systems

a Monte Carlo j  IPS j 


0 10000 10000
1000 129 7904
2000 11 6417
3000 2 5076
4000 1 3993
5000 0 2974
6000 0 2205
7000 0 1509
8000 0 946
9000 0 514
10000 0 193

The above example is indicative of the behavior of MC in the rare event

region for estimating expected shortfall. That is, there are very few trajectories

that exceed the rare event for the calculation of the expected shortfall. However,

one can see that IPS pushes more of the trajectories into the tail so that there

are more simulated paths in the rare event region to compute the expected

shortfall with more accuracy. Though these j  numbers will change with each

simulation, it gives one a good idea of their magnitudes.

132
Chapter 7

Conclusion

In this dissertation, we adapted an original IPS approach to the computation

of rare event probabilities to the field of credit risk under the first passage model.

We showed that with our choice of weight function, IPS is more efficient than

traditional MC methods for computing the probability of tail events in credit

risk. We derived an explicit formula for the asymptotic variance of applying IPS

using our particular choice of weight function. We showed that applying IPS to

multi-name credit derivatives is more advantageous than importance sampling

for the following reasons. First, IPS doesn’t require the user to compute a change

of measure explicitly as is needed to apply importance sampling. Second, IPS is

dynamically consistent in time which allows the user to use the same simulation

to calculate many different maturities in contrast to importance sampling. This

is a distinct advantage when dealing with portfolio credit derivatives such as

CDOs with many different coupon dates and maturities.

133
Chapter 7. Conclusion

In addition, we showed that there are specific regimes where IPS is much

better suited to credit risk than traditional MC methods. We also started the

development of applying IPS to risk management by showing how portfolio VaR

and expected shortfall contributions are calculated in our model.

Our future research will be to adapt our algorithm to more complicated mod-

els that incorporate more realistic market structures such as stochastic volatility

or regime switching models. In addition, we would like to continue the devel-

opment of IPS towards risk management. Finally, we would like to investigate

the calculation of the Greeks by using IPS.

134
Appendix A

Proofs

Recall that by using the reflection principle,

1 −(|z−y|−y)2 /2t
P( min Ws ≤ y, Wt ∈ dz) = √ e 1y≤0 dz (A.1)
0≤s≤t 2πt

The above admits the following density,

φt (y, z)dzdy = P( min Ws ∈ dy, Wt ∈ dz)


0≤s≤t
1
2 −(z−2y)2 /2t
= (z − 2y)e 1y≤z,y≤0 dzdy (A.2)
πt3

Lemma A.0.1. Let Wt be a standard Brownian motion and tp−1 < tp . Let

g(a,c,tp−1 ) (db, de, tp ) denote the following conditional distribution,

g(a,c,tp−1 ) (db, de, tp ) = P(min Wu ∈ db, Wtp ∈ de| min Wu ∈ da, Wtp−1 ∈ dc)1a≤c,a≤0
u≤tp u≤tp−1

Then,

135
Appendix A. Proofs

g(a,c,tp−1 ) (db, de, tp )


*
2 2
= (e + c − 2b)e−(e+c−2b) /2(tp −tp−1 ) dbde1b<a,b≤e,b≤0,a≤c,a≤0
π(tp − tp−1 ) 3

1  
−(c−e)2 /2(tp −tp−1 ) −(e+c−2a)2 /2(tp −tp−1 )
+  e −e de1a≤e,a≤c,a≤0 δa (db)
2π(tp − tp−1 )

136
Appendix A. Proofs

Proof.

g(a,c,tp−1 ) (db, de, tp )

= P(min Wu ∈ db, Wtp ∈ de| min Wu ∈ da, Wtp−1 ∈ dc)1a≤c,a≤0


u≤tp u≤tp−1

= P(min( min Wu , min Wu ) ∈ db, Wtp ∈ de| min Wu ∈ da, Wtp−1 ∈ dc)1a≤c,a≤0
u≤tp−1 tp−1 <u≤tp u≤tp−1

= P( min Wu ∈ db, Wtp ∈ de| min Wu ∈ da, Wtp−1 ∈ dc)1b<a,a≤c,a≤0


tp−1 <u≤tp u≤tp−1

+P( min Wu ≥ a, Wtp ∈ de| min Wu ∈ da, Wtp−1 ∈ dc)1a≤c,a≤0 δa (db)


tp−1 <u≤tp u≤tp−1

= P( min Wu ∈ db, Wtp ∈ de|Wtp−1 ∈ dc)1b<a,a≤c,a≤0


tp−1 <u≤tp
 ∞
+ P( min Wu ∈ dx, Wtp ∈ de|Wtp−1 ∈ dc)1a≤c,a≤0 δa (db)
a tp−1 <u≤tp

= P((c + min Wu ) ∈ db, (c + Wtp −tp−1 ) ∈ de)1b<a,a≤c,a≤0


u≤tp −tp−1
 ∞
+ P((c + min Wu ) ∈ dx, (c + Wtp −tp−1 ) ∈ de)1a≤c,a≤0 δa (db)
u≤tp −tp−1
a
 ∞ 
= φtp −tp−1 (b − c, e − c)dbde1b<a + φtp −tp−1 (x − c, e − c)dx de1a≤c,a≤0 δa (db)
a

= φtp −tp−1 (b − c, e − c)dbde1b<a,a≤c,a≤0


 * 

2 2
+ (e + c − 2x)e−(e+c−2x) /2(tp −tp−1 ) 1x≤e,x≤cdx de1a≤c,a≤0 δa (db)
a π(t p − tp−1 ) 3

= φtp −tp−1 (b − c, e − c)dbde1b<a,a≤c,a≤0


 * 
min(e,c)
2 2
+ (e + c − 2x)e−(e+c−2x) /2(tp −tp−1 ) dx de1a≤c,a≤0 δa (db)
a π(tp − tp−1 ) 3

= φtp −tp−1 (b − c, e − c)dbde1b<a,a≤c,a≤0


1  
−(c−e)2 /2(tp −tp−1 ) −(e+c−2a)2 /2(tp −tp−1 )
+ e −e de1e≤c,a≤c,a≤0δa (db)
2π(tp − tp−1 )
1  2 2

+ e−(e−c) /2(tp −tp−1 ) − e−(e+c−2a) /2(tp −tp−1 ) de1c≤e,a≤c,a≤0δa (db)
2π(tp − tp−1 )
*
2 2
= (e + c − 2b)e−(e+c−2b) /2(tp −tp−1 ) dbde1b<a,b≤e,a≤c,a≤0
π(tp − tp−1 ) 3
137
1  
−(c−e)2 /2(tp −tp−1 ) −(e+c−2a)2 /2(tp −tp−1 )
+  e −e de1a≤e,a≤c,a≤0 δa (db)
2π(tp − tp−1 )
Appendix A. Proofs

Lemma A.0.2. Let Wt be a standard Brownian motion and tp−1 < tp . Let

Φ(tp−1 ,tp ) (da, db, dc, de) denote the following density,

 
Φ(tp−1 ,tp ) (da, db, dc, de) = P min Wu ∈ da, min Wu ∈ db, Wtp−1 ∈ dc, Wtp ∈ de
u≤tp−1 u≤tp

Then,

Φ(tp−1 ,tp ) (da, db, dc, de)


*
2 2
= (e + c − 2b)e−(e+c−2b) /2(tp −tp−1 ) dbde1b<a,b≤e
π(tp − tp−1 ) 3

1  
−(c−e)2 /2(tp −tp−1 ) −(e+c−2a)2 /2(tp −tp−1 )
+ e −e de1a≤e δa (db)
2π(tp − tp−1 )

*
2 2 /2t
× 3 (c − 2a)e−(c−2a) p−1
1a≤c,a≤0 dadc
πtp−1

Proof.

Φ(tp−1 ,tp ) (da, db, dc, de)


 
= P min Wu ∈ da, min Wu ∈ db, Wtp−1 ∈ dc, Wtp ∈ de
u≤tp−1 u≤tp

= P(min Wu ∈ db, Wtp ∈ de| min Wu ∈ da, Wtp−1 ∈ dc) × P( min Wu ∈ da, Wtp−1 ∈ dc)
u≤tp u≤tp−1 u≤tp−1

= g(a,c,tp−1 ) (db, de, tp ) × φtp−1 (a, c)dadc

Substitute the equations for g(a,c,tp−1 ) (db, de, tp ) (Lemma A.0.1) and φtp−1 (a, c)dadc

(Equation A.2) and rearrange.

138
Appendix A. Proofs

Lemma A.0.3. Let Wt be a standard Brownian motion and tp−1 < tp . Let

Γ(tp−1 ,tp ) (a, b, e) denote the following distribution

Γ(tp−1 ,tp ) (da, db, de) = P( min Wt ∈ da, min Wt ∈ db, Wtp ∈ de)
t≤tp−1 t≤tp

Then,

Γ(tp−1 ,tp ) (da, db, de) = Ψ(tp−1 ,tp ) (a, b, e)1b<a,b≤e,a≤0 dedbda

+Υ(tp−1 ,tp ) (a, e)1a≤e,a≤0 δa (db)deda

(A.3)

where,

Ψ(tp−1 ,tp ) (a, b, e)


√ 2

2 2e−(2a−2b+e) /2tp σ(μ(1) − a + e − 2b) −(a−μ(1) )2 /2σ2
= ) × √ e
πtp t2p−1 (tp − tp−1 )2 2π
 

2 a − μ (1)
+ σ + (μ(1) )2 + μ(1) (e − 2b − 2a) − 2a(e − 2b) 1 − Φ( )
σ

139
Appendix A. Proofs

Υ(tp−1 ,tp ) (a, e)


*   
2 −(2a−e)2 /2tp σ −(a−μ(2) )2 /2σ2 a − μ(2)
= e √ e + (μ − 2a)(1 − Φ(
(2)
)
πtp tp−1 2 2π σ
 
−e2 /2tp σ −(a−μ(3) )2 /2σ2 a − μ(3)
−e √ e + (μ − 2a)(1 − Φ(
(3)
)
2π σ

and,

2a(tp − tp−1 ) + (2b − e)tp−1


μ(1) =
tp
2a(tp − tp−1 ) + etp−1
μ(2) =
tp
2atp − etp−1
μ(3) =
tp
tp−1
σ2 = (tp − tp−1 )
tp

Proof. Our strategy is to integrate the four-dimensional distribution Φ(tp−1 ,tp ) (da, db, dc, de)

that we found in Lemma A.0.2 with respect to c. That is, we calculate explicitly,

 ∞
Γ(tp−1 ,tp ) (da, db, de) = Φ(tp−1 ,tp ) (da, db, dc, de)
−∞

where the integral above is taken with respect to c. Recall that the formula

for Φ(tp−1 ,tp ) (da, db, dc, de) is separated into two distinct regions; either b < a

140
Appendix A. Proofs

or b = a. Therefore, to integrate Φ(tp−1 ,tp ) (da, db, dc, de) we will deal with each

region separately.

For notational convenience we drop the functional dependence on (tp−1 , tp )

and write Φ(tp−1 ,tp ) (da, db, dc, de) as

Φ(tp−1 ,tp ) (da, db, dc, de) = f (a, b, c, e)dedcdbda + χ(a, c, e)δa (db)dedcda

where,

f (a, b, c, e)
2 2 /2t 2
= ) (c − 2a)(e − 2b + c)e−(c−2a) p−1 −(e−2b+c) /2(tp −tp−1 )
1b<a,b≤e,a≤c,a≤0
π 2 t3p−1 (tp − tp−1 )3

and,

χ(a, c, e)
(c − 2a)e−(c−2a) /2tp−1  −(c−e)2 /2(tp −tp−1 ) 
2
−(e+c−2a)2 /2(tp −tp−1 )
=  e −e 1a≤e,a≤c,a≤0
π 2 tp−1 3 (tp − tp−1 )

Therefore,

 ∞   ∞ 
Γ(tp−1 ,tp ) (da, db, de) = f (a, b, c, e)dc dedbda+ χ(a, b, c, e)δa (db)dc deda
−∞ −∞

141
Appendix A. Proofs

∞
We first compute −∞
f (a, b, c, e)dc.

By combining the terms in order of c and completing the square in c, we can

write f as,

2
2e−(2a−2b+e) /2tp (1) 2 2
f (a, b, c, e) = ) (c2 +c(e−2b−2a)−2a(e−2b))e−(c−μ ) /2σ 1b<a,b≤e,a≤c,a≤0
π 2 t3p−1 (tp − tp−1 )3

where,

2a(tp − tp−1 ) + (2b − e)tp−1


μ(1) =
tp
tp−1
σ2 = (tp − tp−1 )
tp

We must calculate

Ψ(tp−1 ,tp ) (a, b, e)1b<a,b≤e,a≤0


 ∞
:= f (a, b, c, e)dc
−∞
 ∞ 2
2e−(2a−2b+e) /2tp (1) 2 2
= ) (c2 + c(e − 2b − 2a) − 2a(e − 2b))e−(c−μ ) /2σ 1b<a,b≤e,a≤0 dc
a π 2 t3p−1 (tp − tp−1 )3
−(2a−2b+e)2 /2tp
Let A = √2e 2 3 . We drop the indicator function 1b<a,b≤e,a≤0 for
π tp−1 (tp −tp−1 )3

notational convenience. Therefore,

142
Appendix A. Proofs

Ψ(tp−1 ,tp ) (a, b, e)


√  ∞
1 (1) )2 /2σ 2
= A 2πσ 2 √ (c2 + c(e − 2b − 2a) − 2a(e − 2b))e−(c−μ dc
a 2πσ 2
 (1)

Make the change of variable x = c−μσ
√  ∞
1
−x2 /2
= A 2πσ 2 √ (σx + μ (1) 2
) + (σx + μ (1)
)(e − 2b − 2a) − 2a(e − 2b) e dx
a−μ(1) 2π
σ

Let

h(a, b, e)
 ∞
1
−x2 /2
= √ (σx + μ (1) 2
) + (σx + μ (1)
)(e − 2b − 2a) − 2a(e − 2b) e dx
a−μ(1) 2π
σ

so that


Ψ(tp−1 ,tp ) (a, b, e) = A 2πσ 2 h(a, b, e)

Therefore, we have

143
Appendix A. Proofs

h(a, b, e)
 ∞
1
2 2 2
= (1)
√ σ x + xσ(e − 2b − 2a + 2μ(1) ) e−x /2 dx
a−μ 2π
σ ∞
1 2
+ √ ((μ(1) )2 + μ(e − 2b − 2a) − 2a(e − 2b))e−x /2 dx
a−μ(1) 2π
σ

(Using integration by parts for x2 )


  ∞ 
σ2 2 2
= √ −xe−x /2 |∞ a−μ(1) + (1)
e−x /2 dx
2π x= σ
a−μ

 ∞ σ
1 −x2 /2
+σ(e − 2b − 2a + 2μ(1) ) √ xe dx
a−μ(1) 2π
σ
 ∞
1 2
+((μ ) + μ (e − 2b − 2a) + 2a(e − 2b))
(1) 2 (1)
(1)
√ e−x /2 dx
a−μ
σ

σ(a − μ(1) ) −(a−μ(1) )2 /2σ2 a − μ(1)
= √ e + σ 2 (1 − Φ( ))
2π σ
σ(e − 2b − 2a + 2μ(1) ) −(a−μ(1) )2 /2σ2
+ √ e

a − μ(1)
+((μ(1) )2 + μ(1) (e − 2b − 2a) − 2a(e − 2b))(1 − Φ( ))
σ
σ(μ(1) − a + e − 2b) −(a−μ(1) )2 /2σ2
= √ e

a − μ(1)
+(σ 2 + (μ(1) )2 + μ(1) (e − 2b − 2a) − 2a(e − 2b))(1 − Φ( ))
σ

Since

√ −(2a−2b+e)2 /2t
√ 2 2e p
A 2πσ 2 = )
πtp t2p−1 (tp − tp−1 )2

144
Appendix A. Proofs

then Ψ(tp−1 ,tp ) (a, b, e) is given by

Ψ(tp−1 ,tp ) (a, b, e)1b<a,b≤e,a≤0


 ∞
= f (a, b, c, e)dc
−∞

= A 2πσ 2 h(a, b, e)
√ 2

2 2e−(2a−2b+e) /2tp σ(μ(1) − a + e − 2b) −(a−μ(1) )2 /2σ2
= ) × √ e
πtp t2p−1 (tp − tp−1 )2 2π

a−μ (1)
+(σ 2 + (μ(1) )2 + μ(1) (e − 2b − 2a) − 2a(e − 2b))(1 − Φ( )) 1b<a,b≤e,a≤0
σ
(A.4)

where for the last line above we recall the indicator function 1b<a,b≤e,a≤0 that

we had been suppressing.


∞
We now calculate Υ(tp−1 ,tp ) (a, e)1a≤e,a≤0 δa (db) := −∞
χ(a, c, e)δa (db)dc. Re-

call that,

χ(a, c, e)
(c − 2a)e−(c−2a) /2tp−1  −(c−e)2 /2(tp −tp−1 ) 
2
−(e+c−2a)2 /2(tp −tp−1 )
=  e −e 1a≤e,a≤c,a≤0
π 2 tp−1 3 (tp − tp−1 )

We drop the indicator function 1a≤e,a≤c,a≤0 for notational simplicity. We

perform the multiplication above and separate the two terms. Completing the

square in c, we can write

145
Appendix A. Proofs

2
(c − 2a)e−(c−2a) /2tp−1 −(c−e)2 /2(tp −tp−1 ) (c − 2a) 2 (2) 2 2
 e =  e−(2a−e) /2tp e−(c−μ ) /2σ
π tp−1 (tp − tp−1 )
2 3
π (tp − tp−1 )tp−1
2 3

and,

2
(c − 2a)e−(c−2a) /2tp−1 −(e+c−2a)2 /2(tp −tp−1 ) (c − 2a) −e2 /2tp −(c−μ(3) )2 /2σ2
 e =  e e
π 2 (tp − tp−1 )tp−1 3 π 2 (tp − tp−1 )tp−1 3

where,

2a(tp − tp−1 ) + etp−1


μ(2) =
tp
2atp − etp−1
μ(3) =
tp

and σ 2 is as before.

We also have that,

 ∞
(c − 2a) −(2a−e)2 /2tp −(c−μ(2) )2 /2σ2
 e e dc
a π 2 (tp − tp−1 )tp−1 3
*  
2 −(2a−e)2 /2tp σ −(a−μ(2) )2 /2σ2 a − μ(2)
= e √ e + (μ − 2a)(1 − Φ(
(2)
)
πtp tp−1 2 2π σ

c−μ(2)
after making the change of variables x = σ

146
Appendix A. Proofs

In addition, we also have,

 ∞
(c − 2a) 2 /2t (3) )2 /2σ 2
 e−e p
e−(c−μ
dc
a π 2 (tp − tp−1 )tp−1 3
*  
2 −e2 /2tp σ −(a−μ(3) )2 /2σ2 a − μ(3)
= e √ e + (μ − 2a)(1 − Φ(
(3)
)
πtp tp−1 2 2π σ

c−μ(3)
after making the change of variables x = σ

Therefore, combining the above results, we have that

Υ(tp−1 ,tp ) (a, e)1a≤e,a≤0 δa (db)


 ∞
= χ(a, c, e)δa (db)dc
−∞
 ∞
(c − 2a) 2 (2) 2 2
=  e−(2a−e) /2tp e−(c−μ ) /2σ dcδa (b)1b≤e,a≤0
a π 2 (tp − tp−1 )tp−1 3
 ∞
(c − 2a) 2 (3) 2 2
−  e−e /2tp e−(c−μ ) /2σ dcδa (b)1b≤e,a≤0
a π (tp − tp−1 )tp−1
2 3
*   
2 −(2a−e)2 /2tp σ −(a−μ(2) )2 /2σ2 a − μ(2)
= e √ e + (μ − 2a)(1 − Φ(
(2)
)
πtp tp−1 2 2π σ
 
2 σ (3) 2 2 a − μ (3)
−e−e /2tp √ e−(a−μ ) /2σ + (μ(3) − 2a)(1 − Φ( ) 1b≤e,a≤0 δa (db)
2π σ
(A.5)

Therefore, by combining Equations A.4 and A.5, we arrive at Equation A.3

and the proof is complete.

147
Bibliography

[1] E. Altman, A. Resti, and A. Sironi, Default recovery rates in credit

risk modelling: A review of the literature and empirical evidence, Economic

Notes, 33 (2004), pp. 183–208.

[2] A. Bassamboo and S. Jain, Efficient importance sampling for reduced

form models in credit risk, in WSC ’06: Proceedings of the 38th conference

on Winter simulation, Winter Simulation Conference, 2006, pp. 741–748.

[3] F. Black and J. C. Cox, Valuing corporate securities: Some effects of

bond indenture provisions, Journal of Finance, 31 (1976), pp. 351–67.

[4] A. Borodin and P. Salminen, Handbook of Brownian motion: Facts

and formulae, Basel: Birkauser, 2002.

[5] D. Brigo and M. Tarenghi, Credit default swap calibration and equity

swap valuation under counterparty risk with a tractable structural model,

www.defaultrisk.com, (2005).

148
Bibliography

[6] M. Britten-Jones and S. Schaefer, Non-linear value-at-risk, Euro-

pean Finance Review, 2 (1999), pp. 161–187.

[7] M. Broadie, P. Glasserman, and S. Kou, A continuity correction for

discrete barrier options, Mathematical Finance, 7 (1997), pp. 325–349.

[8] J.-M. Courtault, Y. Kabanov, B. Bru, P. Crépel, I. Lebon,

and A. L. Marchand, Louis bachelier on the centenary of théorie de la

spéculation, Mathematical Finance, 10 (2000), pp. 341–353.

[9] J. C. Cox, J. Ingersoll, Jonathan E, and S. A. Ross, A theory of

the term structure of interest rates, Econometrica, 53 (1985), pp. 385–407.

[10] L. David, On default correlation: A copula function approach, 2001.

[11] P. Del Moral, Feynman-Kac Formulae: Genealogical and Interacting

Particle Systems with Applications, Springer-Verlag, 2004.

[12] P. Del Moral and J. Garnier, Genealogical particle analysis of rare

events, Annals of Applied Probability, 15 (2005), pp. 2496–2534.

[13] D. Duffie, Credit swap valuation, Financial Analyst Journal, 55 (1999).

[14] D. Duffie and D. Lando, Term structures of credit spreads with incom-

plete accounting information, Econometrica, 69 (2001), pp. 633–64. avail-

able at http://ideas.repec.org/a/ecm/emetrp/v69y2001i3p633-64.html.

149
Bibliography

[15] A. Elizalde, Credit risk models iv: Understanding and pricing cdos, work-

ing papers, CEMFI, Apr. 2006.

[16] Y. Eom, J. Helwege, and J. Z. Huang, Structural models of corporate

bond pricing: An empirical analysis, Review of Financial Studies, 17 (2004),

pp. 499–544.

[17] J.-P. Fouque, R. Sircar, and K. Solna, Stochastic volatility effects on

defaultable bonds, Applied Mathematical Finance, 13 (2006), pp. 215–244.

[18] P. Glasserman, Monte Carlo Methods in Financial Engineering,

Springer, 2003.

[19] P. Glasserman, P. Heidelberger, and P. Shahabuddin, Efficient

monte carlo methods for value-at-risk, Mastering Risk: Vol 2, (2001).

[20] P. Glasserman and J. Li, Importance sampling for portfolio credit risk,

Management Science, 51 (2005), pp. 1643–1656.

[21] J. Hull and A. White, Valuing credit default swaps i: No counterparty

default risk, Journal of Derivatives, 8 (2000).

[22] , Valuing credit default swaps ii: Modeling default correlation, Journal

of Derivatives, 8 (2001).

150
Bibliography

[23] E. Jones, S. Mason, and E. Rosenfeld, Contingent claims analysis of

corporate capital structures: An empirical investigation, Journal of Finance,

39 (1984), pp. 611–625.

[24] I. Karatas and S. E. Shreve, Brownian Motion and Stochastic Calcu-

lus, Springer, 1991.

[25] D. Lando, Credit Risk Modeling, Princeton University Press, 2004.

[26] R. C. Merton, On the pricing of corporate debt: the risk structure of

interest rates, Journal of Finance, 29 (1974), pp. 449–470.

[27] P. J. Schonbucher, Credit Derivative Pricing Models: Models, Pricing

and Implementation, Wiley Finance Series, 2003.

[28] S. E. Shreve, Stochastic Calculus for Finance II: Continuous-Time Mod-

els, Springer, 2003.

[29] C. Zhou, An analysis of default correlations and multiple defaults, Review

of Financial Studies, 14 (2001), pp. 555–76.

151