Valuation of a Homogeneous Collateralized Debt Obligation

by

Fabio Mibielli Peixoto

An essay presented to the University of Waterloo in partial fulllment of the requirements for the degree of Masters in Mathematical Finance

Waterloo, Ontario, Canada, 2004

c Fabio Mibielli Peixoto 2004

1 E-mail:

fabiop@aya.yale.edu

Abstract Monte Carlo simulation and a semi-analytical method are used to value a basket default swap and an homogeneous Collateralized Debt Obligation (CDO). The semianalytical technique is based on the one factor copula model proposed by J.P. Laurent and J. Gregory [1]. We study the properties of a CDO with Monte Carlo and compare the spread calculation with the one obtained by the factor model.

Contents

List of Figures List of Tables 1 Introduction 1.1 1.2 Denition and Classication . . . . . . . . . . . . . . . . . . . . . . How a CDO works . . . . . . . . . . . . . . . . . . . . . . . . . . .

iii v 1 2 4 6 6 7 8 10 11 14 15 21

2 Mathematical Background 2.1 2.2 2.3 2.4 2.5 Stopping Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hazard Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Point Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Distribution of Default Times . . . . . . . . . . . . . . . . . . . . . Copula Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3 Monte Carlo simulation of a CDO 3.1 3.2 Expected Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . CDO Spread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i

4 Single Factor Copula Model 4.1 4.2 Basket Default Swap . . . . . . . . . . . . . . . . . . . . . . . . . . CDO spread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

26 28 33 36 38

5 Conclusion Bibliography

ii

List of Figures
1.1 Estimated global number of CDOs from 1987 to 2000. Source: JP Morgan [3]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 1.3 1.4 3.1 General classication of CDOs. . . . . . . . . . . . . . . . . . . . . Schematics of how a CDO is setup. . . . . . . . . . . . . . . . . . . Capital structure of a simple CDO. . . . . . . . . . . . . . . . . . . Dependence of expected loss with correlation, for 50,000 iterations, h = 0.03 and R = 0.4. . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Variation of expected loss with recovery rate for 50,000 iterations, h = 0.03 and = 0.3. . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Portfolio loss distribution for correlation at 10%, recovery rate 25%, hazard rate 0.03 and 50,000 trials . . . . . . . . . . . . . . . . . . . 3.4 Portfolio loss distribution for correlation at 10%, recovery rate 50%, hazard rate 0.03 and 50,000 trials . . . . . . . . . . . . . . . . . . . 3.5 Equity spread versus correlation with R = 0.4, h = 0.03 and 20,000 iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Equity spread versus recovery rate with = 0.3, h = 0.03 and 20,000 iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii 23 23 20 20 19 18 3 4 5 5

3.7

Mezzanine spread versus correlation with R = 0.4, h = 0.03 and 20,000 iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.8

Mezzanine spread versus recovery rate with = 0.3, h = 0.03 and 20,000 iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.9

Senior spread versus correlation with R = 0.4, h = 0.03 and 20,000 iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.10 Senior spread versus recovery rate with = 0.3, h = 0.03 and 20,000 iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

iv

List of Tables
3.1 Expected loss for each CDO tranche calculated with 50,000 iterations, h = 0.03, = 0.3 and R = 0.4. . . . . . . . . . . . . . . . . . 3.2 Fair spread of a homogeneous CDO calculated with Monte Carlo with 30,000 iterations, constant hazard rate 0.03, correlation parameter 0.3, recovery rate 0.4 and quarterly premium payments ( = 1/4). . 4.1 Spreads for a k th to default basket swap on a homogeneous portfolio with 10 instruments, xed correlation factor a = 0.3 and recovery rate 40%. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Spreads for a k th to default basket swap on a homogeneous portfolio with 10 instruments with xed hazard rate 0.03 and recovery rate 40%. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Comparison of spreads calculated with Monte Carlo and a single factor model. CDO has 100 obligors with h = 0.03 each, a = 0.3 and R = 40%. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 33 32 22 17

Chapter 1 Introduction
In the last twenty years one of the main innovations in the eld of Finance has been securitization. This is the process of pooling together a portfolio and issuing liability and equity notes backed by this pool of assets. It started in the late 1970s with mortgages backed securities and it has expanded to other instruments like credit card debt, student and car loans, junk bonds, etc. Its main purpose for the originator is to transfer some of the risk to the investors and to free up regulatory capital. The main advantage for the investors is diversication: they are able to invest in products that they would not have access otherwise. In this report we will study one of the main examples of securitization: a Collateralized Debt Obligation (CDO). This is a security that has had a tremendous growth in the last couple of years linked to the surge of the credit derivatives market. Using Monte Carlo simulation and a semi-analytical method we analyze and nd the no-arbitrage price of a CDO. We explain both methods in details and compare their results. The semi-analytical method is based on the single factor copula model proposed by J.P. Laurent and J. Gregory [1] and recently used also by J. 1

CHAPTER 1. INTRODUCTION Hull and A. White [2].

In this introduction we dene and classify a CDO explaining how it works. In chapter 2 we make a review of the mathematical tools necessary to study a defaultable portfolio. In Chapter 3 we apply Monte Carlo simulation to study a CDO and in chapter 4 we explain and use the single factor copula model to price a CDO. In chapter 5 we present our conclusion.

1.1

Denition and Classication

A CDO is an asset backed security whose collateral consists mainly of a portfolio of defaultable instruments like loans, junk bonds, mortgages, etc. If the portfolio contains only credit default swaps (CDS), it is called a synthetic CDO. This has become one of the most popular CDOs in the last couple of years. The reason is that with a CDS the bank can create a portfolio with exposure to defaultable instruments without the requirement of owning them, facilitating their creation. The main economic reason for the existence of CDOs is to address market imperfections. It frees regulatory capital for nancial institutions and it improves liquidity for some bonds and loans. For investors it creates risk-return proles hard to achieve in other ways and it gives them opportunity to invest in less liquid assets. The issuance of CDOs has had an exponential increase for the last 10 years (see Fig. 1.1). This sudden growth is due to its better acceptance by investors, especially hedge funds, and credit risk hedgers [3]. Also the surge of the credit derivatives market in the last couple of years has contributed to the increase in the number of synthetic CDOs.

CHAPTER 1. INTRODUCTION

Figure 1.1: Estimated global number of CDOs from 1987 to 2000. Source: JP Morgan [3].

CDOs can be classied in two categories: arbitrage and balance sheet (see Fig. 1.2). An arbitrage CDO is setup to make money from the dierence between the cost of acquiring the collateral portfolio and the sale of the notes. A balance sheet CDO has the purpose of freeing up regulatory capital that is tied up to the collateral asset. In this case the portfolio consists mainly of loans [4]. We can further divide CDOs into cash ow and market value types. A cash ow CDO is not subjected to active trading by the manager and its main source of risk is due to credit default. A market value CDO has its collateral portfolio markedto-market frequently, therefore it has an additional risk linked to the performance of the CDO manager. In this report we will analyze only cash ow CDOs, since nine out of ten CDOs, both by number and by volume, are of this type [3].

CHAPTER 1. INTRODUCTION

CDO

Arbitrage

Balance sheet

Cash flow

Market value

Cash flow

Figure 1.2: General classication of CDOs.

1.2

How a CDO works

A Special Purpose Vehicle (SPV) is setup to hold the portfolio of assets. It issues notes with dierent subordination, so called tranches, that is sells to investors (see Fig. 1.3). The principal payment and interest income (LIBOR + spread) are allocated to the notes according to the following rule: senior notes are paid before mezzanine and lower rated notes. Any residual cash is paid to the equity note. The equity note is also called the rst-loss position because it is the rst to be aected by a default in the portfolio. Therefore it oers a larger spread than the more senior notes. In practice we can say that the CDO investors are selling protection, choosing a tranche according to their risk - return preference. A consequence of this payment subordination is that a senior note is less risky than the collateral. As we move down the CDOs notes, the risk level increases. Each tranche, except for the equity, is rated by credit rating agencies. The senior notes usually have Aaa/AAA rating. The originating bank has an incentive to keep

CHAPTER 1. INTRODUCTION

Assets s old

Interest & Principal

Senior Mezzanine

Ban k
Funds

SPV
Funds

Equity

Figure 1.3: Schematics of how a CDO is setup.

Sen ior
14%- 100%

Mezza nine 3% - 14%

Equ ity 0% - 3%

Figure 1.4: Capital structure of a simple CDO.

a signicant share of the equity tranche to demonstrate its commitment to the CDO performance. As an example consider a CDO with three tranches: equity, mezzanine and senior. Suppose that the total CDO notional is 100 million and the capital structure is the one shown in Fig. 1.4. During the CDO lifetime some instruments in the collateral portfolio might default. If at maturity the total default loss is less than 3 million, only the equity tranche is aected. If it is between 3 and 14 million, the equity tranche does not get the principal back and the mezzanine gets only part of it. If the loss is more than 14 million than the equity and mezzanine do not get anything back and the senior tranche gets whatever is left.

Chapter 2 Mathematical Background

In this chapter we will present the basic mathematical tools to study the default time of a credit portfolio. We assume a probability space (, F, {Ft }t0 , P), where is the set of all states, F is a -eld representing the collection of all events, {Ft }t0 is a ltration representing the increase in information with time and P is a probability measure. When pricing an instrument we always assume no-arbitrage and in this case P = Q is the risk-neutral measure.

2.1

Stopping Time

When studying a credit portfolio, the rst thing we have to model is the time of default for each instrument. It should be a random variable taking values in R+ {}, where we include {} to take into account a no-default event. Also, if we assume a market with free ow of information, the event of default should be known to everyone when it occurs. Therefore is not only a random time but also

CHAPTER 2. MATHEMATICAL BACKGROUND a stopping time with respect to {Ft }t0 : { < t} Ft , t 0,

(2.1)

where { < t} is the inverse image of () < t. Notice that the requirement that is a stopping time does not mean that the default event must come as a surprise. The concept of a predictable stopping time can be used for predictable defaults [5]. For modelling purposes we can represent the default event at time t as: N = I{ t} , where I is the indicator function (I{ t} = 1 if t, I{ t} = 0 if > t). (2.2)

2.2

Hazard Rate

The hazard rate function h plays an important role in intensity models for credit risk. Intuitively a large value of h means a higher probability of default. Its formal denition is the following [5]. Denition 2.2.1 Let be a stopping time and F (t) = P( t) its distribution function. Assuming that F (t) < 1 for all t and that f (t) is the density function, the hazard rate h of is given by: h(t) = dF (t) f (t) 1 = . 1 F (t) 1 F (t) dt (2.3)

Using F = P (t < t + t, > t) and 1 F (t) = P ( > t) in (2.3) we can write: h(t) = lim P (t < t + t| > t) . t0 t (2.4)

CHAPTER 2. MATHEMATICAL BACKGROUND

Equation (2.4) gives us an interpretation of h(t). It is the probability of default, per unit of time, in (t, t + t), given that no default happened before t. From Eq.(2.3) we can re-write F (t) as: F (t) = 1 e
Rt
0

h(s)ds
Rt
0

, ,

(2.5) (2.6)

S(t) = 1 F (t) = e

h(s)ds

where S(t) = P ( > t) is called the survival function.

2.3

Point Processes

A single default time is represented by a stopping time as seen in section 2.1. The generalization for multiple default events constitutes a point process. Denition 2.3.1 A point process is an increasing sequence of stopping times {1 , 2 , . . .} where 0 < i < i+1 i N. A point process is a good way to represent a sequence of defaults in a credit portfolio or multiple defaults of a single obligor. The total number of defaults is represented by a counting process:
n

N (t) =
i=1

I{i t} ,

(2.7) (2.8)

N (0) = 0.

CHAPTER 2. MATHEMATICAL BACKGROUND

This is a stochastic process with natural ltration Gt = {N (s), 0 s t}. More specically N (t) is a submartingale as shown below:
n

E[N (s)|Gt ] =
i=1

E[I{i s} |Gt ], s > t > 0,

(2.9) (2.10) (2.11) (2.12)

I{i s} = I{i t} + I{ti s} ,

n

E[N (s)|Gt ] = N (t) +

i=1

P (t i s|Gt ),

E[N (s)|Gt ] N (t).

We can then apply the Doob-Meyer decomposition theorem to the counting process N (t) [6]. Theorem 2.3.1 (Doob-Meyer decomposition) Let X(t) be a right-continuous, non-negative submartingale with respect to (, F, {Ft }t0 , P). Then there exists a right-continuous martingale M (t) and an increasing, right-continuous process A(t) such that E[A(t)] < and X(t) = M (t) + A(t) a.s. for any t 0. If we assume that A(t) is predictable, this decomposition is unique. The theorem above implies that there exists a unique, predictable process A(t) such that {N (t) A(t)}t0 is a right-continuous Ft -martingale. The process A(t) is called the compensator of (N (t), Ft ). Its importance is that it gives some information about the probability of defaults over the next period of time: E[N (t + t) N (t)|Ft ] = E[A(t + t) A(t)|Ft ], since {N (t) A(t)}t0 is an Ft -martingale. Denition 2.3.2 If A(t) is dierentiable, we can write A(t) =
t 0

(2.13)

(s)ds, where

(t) is a positive, predictable process called the stochastic intensity of N (t).

CHAPTER 2. MATHEMATICAL BACKGROUND

10

Credit models where A(t) is dierentiable are called intensity models. We can relate (t) to the probability of one default at time t, given the information at time t : E[dN (t)|Ft ] = dA(t), (2.14)

where we have used (2.13) and the fact that A(t) is predictable. Assuming that dN (t) = {0, 1} in a short period of time we have: P [dN (t) = 1|Ft ] = (t)dt. (2.15)

If we use the natural ltration Gt = {N (s), 0 s t} and some regularity conditions, we can show that the intensity process is equal to the hazard rate [5]: (t) = h(t). (2.16)

2.4

Distribution of Default Times

In this report we will consider only the simple case of a piece-wise constant hazard rate. This is a simplifying approximation that is actually used quite often in the nancial industry. We will also assume that N(t) follows a homogeneous Poisson process with intensity = h. Denition 2.4.1 The counting process {N (t)}t0 follows a homogeneous Poisson process with intensity if and only if: {N (T ) N (t)}T >t0 are stationary and independent increments. N (t) has a Poisson distribution with parameter . P [NT Nt = k] = 1 [(T t)]k e(T t) . k! (2.17)

CHAPTER 2. MATHEMATICAL BACKGROUND

11

Following Landos paper [7] we can simulate the rst time of default . Assuming N0 = 0, we have from (2.17) the probability of no-default: P (Nt = 0) = et . (2.18)

The survival function S (t) = P ( > t) and the distribution function F (t) = P ( t) of the rst time of default are: S (t) = et , F (t) = 1 et . Using the inverse transform method on (2.20) we can simulate : = ln U , (2.21) (2.19) (2.20)

where U is a uniform random variable in [0, 1]. We can apply Eq.(2.21) to simulate the time of default for each instrument in a credit portfolio. This will be useful in chapter 3.

2.5

Copula Functions

When analyzing a portfolio with n defaultable instruments, one of the most important parameters to know is the joint default distribution: F (t1 , . . . , tn ) = P (1 t1 , . . . , n tn ). (2.22)

Finding out the format of F is not an easy task due to the fact that dierent joint probability distributions can have the same marginal functions. The connection between the two is given by the copula function. It separates the structural dependence of the random variables from their marginal behavior [8], [9].

CHAPTER 2. MATHEMATICAL BACKGROUND

12

Denition 2.5.1 An n-dimensional copula function C : [0, 1]n [0, 1], n N is a joint distribution function with uniformly distributed marginal on [0, 1]. C(u1 , . . . , un ) = P (U1 u1 , . . . , Un un ), Ui U [0, 1], i = {1, . . . , n}. (2.23) (2.24)

The most important theorem in the theory of copulas is Sklars theorem [10], [11]. It proves that copulas are the link between the marginal functions and the joint distribution function. Theorem 2.5.1 (Sklar) Let F be a joint distribution function on Rn with marginal distribution Fi , i = {1, . . . , n}. Then there exists a copula function C such that (x1 , . . . , xn ) Rn , F (x1 , . . . , xn ) = C(F1 (x1 ), . . . , Fn (xn )). If the Fi s are continuous, the copula is unique. Corollary 2.5.1 Let F be an n-dimensional joint distribution function with marginal distribution Fi , i = {1, . . . , n}. An n-dimensional copula C can be built as:
1 1 C(u1 , . . . , un ) = F [F1 (u1 ), . . . , Fn (un )],

(2.25)

(2.26)

where Fi1 is the quasi-inverse function dened by Fi1 (ui ) = inf {x|Fi (x) ui }. The corollary above is the key to generate many dierent copulas. One of the rst to be used in credit risk was the Gaussian copula [12]. It is obtained by using the multivariate normal distribution ( is the correlation matrix) and normal marginal distribution in (2.26): CGauss (u1 , . . . , un ) = [1 (u1 ), . . . , 1 (un )]. (2.27)

CHAPTER 2. MATHEMATICAL BACKGROUND

13

If we assume as in [12] that the dependence structure in a credit portfolio is represented by a Gaussian copula, the joint probability distribution of default is obtained by using (2.27) in (2.25): P (1 t1 , . . . , n tn ) = [1 (F1 (t1 )), . . . , 1 (Fn (tn ))], (2.28)

where Fi (ti ) is given by (2.5). Notice that each default time i can have a dierent marginal distribution. There are many other copulas that can be used in credit risk besides the Gaussian, including the t-copula, the Clayton copula, etc [8], [9]. However, because we will not study the CDO price sensitivity to the copula choice, we will always assume a Gaussian copula in our analysis.

Chapter 3 Monte Carlo simulation of a CDO

Consider a portfolio with n instruments, each one with notional Ni , i = {1, . . . , n}. We associate a random variable i to each one of them representing their default time with marginal default distribution: Fi (t) = P (i t), t 0. The default event correlation between instruments i and j is given by: ij (t) = corr I{i <t} , I{j <t} , where I is the indicator function. The default loss for instrument i is a random variable represented by: Li (t) = (1 Ri )Ni I{i <t} , (3.3) (3.2) (3.1)

where Ri is the recovery rate, assumed deterministic in our case. Therefore the total portfolio loss is given by:
n

L(t) =
i=1

Li (t). 14

(3.4)

CHAPTER 3. MONTE CARLO SIMULATION OF A CDO

15

In the case of a CDO we are interested in knowing how the total losses (3.4) aect a tranche with a lower bound a% and an upper bound b%. From the payment subordination we know that the tranche [a,b] suers a loss at time t if: a%NT < L(t) b%NT , where NT =
n i=1

(3.5)

Ni is the total portfolio value. Due to condition (3.5) we can

represent the tranche loss LA,B (t) as: LA,B (t) = [L(t) A] I{L(t)[A,B]} + (B A)I{L(t)[B,NT ]} , where A = a%NT and B = b%NT . (3.6)

3.1

Expected Loss

The rst property of the CDO that we will analyze is the expected loss for each tranche at maturity. Using Monte Carlo simulation and Eq.(3.6) this becomes a straightforward task. For this purpose we are going to assume a simple structure as the one depicted in Fig. 1.4. Moreover we want to focus on the dependence of the expected loss on the correlation and recovery rate, hence we will simplify our model assuming a homogeneous CDO: same notional N , recovery rate R and hazard rate h for all obligors. We also use a constant correlation matrix with only one parameter: ij = , i = j, ij = 1, i = j, 0 < < 1. (3.7) (3.8) (3.9)

CHAPTER 3. MONTE CARLO SIMULATION OF A CDO

16

This is a major assumption, but it is a very common one in the industry. A one dimensional correlation matrix can be treated in a similar way as implied volatility in the Black-Scholes equation. We will study a portfolio with n = 100 instruments, maturity T = 5 years and we assume that the risk-free interest rate r is constant at 5% p.a. Using these assumptions we can simplify (3.5) and (3.6): a <L(t) b, L(t) = L(t)/N. The tranche loss is given by: La,b (t) = [L(t) a] I{L(t)[a,b]} + (b a)I{L[b,100]} . N (3.12) (3.10) (3.11)

We use a Gaussian copula to model the dependence structure in the portfolio. The procedure to estimate the expected loss at maturity with Monte Carlo simulation is the following [7], [13]: 1. Generate n multinormal random variates vi with mean zero and correlation matrix given by (3.7) and (3.8). 2. Assuming a constant hazard rate h for n obligors, the default time is given by (2.21): ln ui , i = 1, . . . , n , h where ui = (vi ) and is the standard normal distribution. i = 3. Using (3.12) and (3.13) we calculate the loss at maturity Lk (T ). a,b 4. Repeat procedure 1 to 3 m times. The estimator for the expected loss is: 1 E[La,b (T )] = m
m

(3.13)

Lk (T ). a,b
k=1

(3.14)

CHAPTER 3. MONTE CARLO SIMULATION OF A CDO

17

Table 3.1 shows the value and the standard error for the expected loss calculated with 50,000 iterations, h = 0.03, = 0.3 and R = 0.4. Notice that crude Monte Carlo is already good enough to give us a relative error less than 2%. Tranche 14% - 100% 3% - 14% 0% - 3% Exp. Loss (%) 1.83 39.23 82.59 Std. error (%) 0.02 0.18 0.14

Table 3.1: Expected loss for each CDO tranche calculated with 50,000 iterations, h = 0.03, = 0.3 and R = 0.4.

Figure 3.1 shows the dependence of the expected loss at maturity on correlation with xed recovery rate. The position of an investor that buys an equity or mezzanine tranche is long correlation: larger the correlation in the future, smaller will be his expected loss. An investor who buys a senior tranche is short correlation: he is betting that correlation will remain small or go down in the future. This terminology is commonly used in the industry [14]. Figure 3.2 shows the dependence of the expected loss with recovery rate for a xed correlation ( = 0.3). All tranche losses decrease with higher recovery rate as expected. Figures 3.3 and 3.4 show the loss distribution of a portfolio with dierent recovery rates and xed correlation. Notice that the one with larger recovery rate (Fig. 3.4) has a smaller right tail, meaning that the senior tranche gets less aected by the losses. This can also be observed in Fig. 3.2 where the senior expected loss gets close to zero at high recovery rates.

CHAPTER 3. MONTE CARLO SIMULATION OF A CDO

18

R = 40%
100 Equity Mezzanine Senior

90

80

70

Expected loss (%)

60

50

40

30

20

10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Correlation

Figure 3.1: Dependence of expected loss with correlation, for 50,000 iterations, h = 0.03 and R = 0.4.

CHAPTER 3. MONTE CARLO SIMULATION OF A CDO

19

= 30%
90

80

Equity Mezzanine Senior

70

60

Expected loss (%)

50

40

30

20

10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Recovery rate

Figure 3.2: Variation of expected loss with recovery rate for 50,000 iterations, h = 0.03 and = 0.3.

CHAPTER 3. MONTE CARLO SIMULATION OF A CDO

20

6 5 Probability (%) 4 3 2 1 0

10

20 30 Loss (%)

40

50

Figure 3.3: Portfolio loss distribution for correlation at 10%, recovery rate 25%, hazard rate 0.03 and 50,000 trials
6 5 Probability (%) 4 3 2 1 0

10

20 30 Loss (%)

40

50

Figure 3.4: Portfolio loss distribution for correlation at 10%, recovery rate 50%, hazard rate 0.03 and 50,000 trials

CHAPTER 3. MONTE CARLO SIMULATION OF A CDO

21

3.2

CDO Spread

Pricing a CDO tranche is very similar to pricing a basket of credit default swaps. First we have to estimate the present value of tranche losses triggered by credit events during the CDO lifetime. Second we calculate the present value of the premium payments weighted by the outstanding capital (original tranche amount minus accumulated losses). The former is called the default leg (DL) and the latter is the premium leg (PL). The fair spread s is dened such that the expected value of both legs is equal. The expectation is taken under the risk-neutral measure. s= E[DL] . E[P L] (3.15)

Let us now describe in more details how to calculate DL and PL for a homogeneous CDO tranche [a,b] with n = 100. Consider the k th iteration of a Monte Carlo simulation. The accumulated loss at time t, in percentage, is given by (3.11):
100

L (t) = (1 R)
i=1

I{ik <t} ,

(3.16)

k k where {1 , . . . , 100 } is a sequence of default times obtained as described in sec-

tion 3.1. Equation (3.12) gives us the tranche loss, therefore the k th default leg is: DLk =
i=1 100

eri La,b (ik ) ,

(3.17)

where r is the risk-free interest rate. The premium leg is paid over the outstanding capital in the tranche. If during the lifetime of the CDO the tranche is wiped out, there are no more premium payments. The formula is given by:
w

P Lk = N
j=1

j ertj min max b Lk (tj ), 0 , b a ,

(3.18)

CHAPTER 3. MONTE CARLO SIMULATION OF A CDO

22

where {t1 , . . . , tw } are the premium payment dates with frequency j = tj tj1 . Taking the average of m values of (3.17) and (3.18) and applying them in (3.15), we obtain the estimate of s for the CDO from Fig. 1.4: Tranche 14% - 100% 3% - 14% 0% - 3% s (bps/year) Std. error (bps/year) 35.4 961 4073 0.6 7 26

Table 3.2: Fair spread of a homogeneous CDO calculated with Monte Carlo with 30,000 iterations, constant hazard rate 0.03, correlation parameter 0.3, recovery rate 0.4 and quarterly premium payments ( = 1/4).

As expected the equity tranche demands the highest spread because it is the riskiest one. Crude Monte Carlo again gives us a precise estimation with a relative error of the order of 1%. Notice that when calculating the standard error s we have to take into account the covariance between DL and PL: s s = m
2 2 DL P L cov(DL, P L) + 2 , (E[DL])2 (E[P L])2 E[DL]E[P L]

(3.19)

2 2 where DL and P L are the variance estimators.

Figures (3.5) to (3.10) show the spread dependence on correlation and recovery rate. It is natural that they should behave similarly to the expected loss: riskier the tranche, higher the premium demanded by the protection seller. Notice also that the equity and the senior tranches are very sensitive to the recovery rate, therefore indicating that this is an important parameter to estimate when pricing a CDO [14].

CHAPTER 3. MONTE CARLO SIMULATION OF A CDO

Equity tranche (0% 3%) R = 40%
12000

23

10000

8000

Spread (bps/year)

6000

4000

2000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Correlation

Figure 3.5: Equity spread versus correlation with R = 0.4, h = 0.03 and 20,000 iterations.
Equity tranche (0 3%) = 30%
5500 5000 4500 4000 3500 3000 2500 2000 1500 1000 500 0

Spread (bps/year)

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Recovery rate

Figure 3.6: Equity spread versus recovery rate with = 0.3, h = 0.03 and 20,000 iterations.

CHAPTER 3. MONTE CARLO SIMULATION OF A CDO

Mezzanine tranche (3% 14%) R = 40%
1200

24

1100

1000

900

Spread (bps/year)

800

700

600

500

400

300

200

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Correlation

Figure 3.7: Mezzanine spread versus correlation with R = 0.4, h = 0.03 and 20,000 iterations.
Mezzanine tranche (3% 14%) = 30%
1800

1600

1400

1200

Spread (bps/year)

1000

800

600

400

200

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Recovery rate

Figure 3.8: Mezzanine spread versus recovery rate with = 0.3, h = 0.03 and 20,000 iterations.

CHAPTER 3. MONTE CARLO SIMULATION OF A CDO

Senior tranche (14% 100%) R = 40%
180

25

160

140

120

Spread (bps/year)

100

80

60

40

20

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Correlation

Figure 3.9: Senior spread versus correlation with R = 0.4, h = 0.03 and 20,000 iterations.
Senior tranche (14% 100%) = 30%
120

100

80

Spread (bps/year)

60

40

20

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Recovery rate

Figure 3.10: Senior spread versus recovery rate with = 0.3, h = 0.03 and 20,000 iterations.

Chapter 4 Single Factor Copula Model

Latent factor models are well known in Actuarial Science where they have been used to model loan loss distributions for some time (see references in [1]). Their main property is that the default events, when conditioned on some factor, are independent. This allows the use of semi-analytical methods to estimate the joint default probability and other relevant quantities. Especially in high-dimension problems this could be a good alternative approach to Monte carlo simulation. For the particular case of a CDO, a factor approach has been proposed by J.P. Laurent and J. Gregory [1]. The motivation in applying this technique is because CDOs usually contain a large number of obligors (100 or more) where a semi-analytical method could prove to be more time ecient than simulation. Notice however that the factor model is not purely analytical because a numerical integration is still necessary. In this report we will study the application of a one factor model with a Gaussian copula to a homogeneous CDO. We start by assuming a simplied version of a rms value model [5]. 26

CHAPTER 4. SINGLE FACTOR COPULA MODEL

27

Assumption 4.0.1 The default of each obligor is triggered by the change of value of its assets. This value is denoted by Vi (t) and we assume that it is normalized and it has a standard normal distribution N (0, 1). The assets of each rm is correlated with each other through a correlation matrix . Each obligor defaults when Vi Ki , where Ki is a stochastic barrier. Even with this simple model there are still N (N 1)/2 parameters to be estimated in for a portfolio with N instruments. This can be a daunting task as in the case with N = 100, where N (N 1)/2 = 4950. Therefore we need to make more assumptions. A common one is to assume that all the rm values Vi depend on M common factors. In this case the number of free parameters is M N . Let us assume a one factor model (M = 1) as proposed in [1]: Assumption 4.0.2 The rm values are driven by a common factor V and a noise term
i.

Vi (t) = ai V + where V and

i

1 a2 i , i

(4.1)

are independent, standard normal random variables.

From Eq.(4.1) we get: Cov(Vi , Vj ) = ai aj , V ar(Vi ) = 1. (4.2) (4.3)

Therefore in this model there are only N free parameters in . Later on we will simplify even further by assuming that ai = aj = a, i = j, just like we did in section 3.1. The factor V can be understood as an underlying economic variable, like the business cycle or the interest rate, that aect dierent industries at the same time. The noise
i

is related to rm specic factors as management, balance

CHAPTER 4. SINGLE FACTOR COPULA MODEL

28

sheet, etc. The important thing is that, conditioned on V , the rms values and default distribution are independent. Since a default is triggered by Vi crossing the barrier Ki , we can say that the distribution of default time is: Fi (t) = P (i t) = P (Vi Ki ), Ki = 1 [Fi (t)], where we used the assumption that Vi N (0, 1). The conditional probability of default is: pi (t|V ) = P (Vi Ki |V ) = 1 [Fi (t)] ai V 1 a2 i
i

(4.4) (4.5)

(4.6)

where we have used (4.1), (4.5) and the fact that

N (0, 1). Using the assumption

that the defaults are independent when conditioned on V , we obtain the joint probability of default:
n

F (t1 , . . . , tn ) =
i=1 n

i=1

1 [Fi (t)] ai v 1 a2 i ai v 1 [Fi (t)] 1 a2 i

g(v)dv, g(v)dv,

(4.7) (4.8)

S(t1 , . . . , tn ) = where g(v) = ev

2 /2

/ 2 is the density function of V . Equations (4.7) and (4.8)

can be integrated using numerical methods and they allow us to price any kind of credit portfolio.

4.1

Basket Default Swap

Because of their pricing similarities, we will rst apply the one factor model to a basket default swap before pricing a CDO. We consider a k th to default swap that

CHAPTER 4. SINGLE FACTOR COPULA MODEL

29

only pays when the k th default happens in a portfolio with n k obligors. In order to price this instrument, we need to know the risk-neutral probability of k defaults by time t. Equation (4.8) can be written as:
n

S(t1 , . . . , tn ) =
i=1

qi (t|v)g(v)dv,

(4.9)

where qi (t|V ) is the conditional survival function of instrument i. The number of defaults at time t is given by the counting process (2.7) and in a one factor model the probability of k defaults can be written as: P [N (t) = k] = P [N (t) = k|v]g(v)dv. (4.10)

Therefore we have to nd a formula for P [N (t) = k|V ], k = {0, . . . , n}. Let us start with the conditional probability of no-default:
n

P [N (t) = 0|V ] =
i=1

qi (t|V ).

(4.11)

The conditional probability of one default is:

n

P [N (t) = 1|V ] =
i=1 n

(1 qi )

qj , qj . qi

(4.12)

j=1 j=i n

P [N (t) = 1|V ] =
i=1

(1 qi )
j=1

(4.13)

Using (4.11) in (4.13):

n

P [N (t) = 1|V ] = P [N (t) = 0|V ]

i=1

wi (t|V ),

(4.14) (4.15)

1 qi (t|V ) . wi (t|V ) = qi (t|V )

CHAPTER 4. SINGLE FACTOR COPULA MODEL Repeating a similar procedure for k = {2, 3, . . .}, we can quickly verify that: P [N (t) = k|V ] = P [N (t) = 0|V ] wp(1) . . . wp(k) ,

30

(4.16)

where the sum is taken over all n!/(n k)!n! dierent combinations of {1, . . . , n}. Reference [2] gives an algorithm to calculate all the elements of this sum. Calling Ak = relation: A1 = B1 , 2A2 = B1 A1 B2 , 3A3 = B1 A2 B2 A1 + B3 , kAk = B1 Ak1 B2 Ak2 + B3 Ak3 . . . + (1)k Bk1 A1 + (1)k+1 Bk . (4.17) (4.18) (4.19) (4.20) wp(1) . . . wp(k) and Bj =
n j i=1 (wi ) ,

they prove the following recurrence

The algorithm (4.17) - (4.20) can be easily implemented to generate Ak for all k {0, . . . , n}. Knowing how to calculate (4.10) is enough to price any basket default swap, because the k th survival function can be written as: S k (t) = P ( k > t), S k (t) = P [N (t) < k],
k1

(4.21) (4.22) (4.23)

S (t) =
j=0

P [N (t) = j],

where k is the k th default time. When pricing a basket we assume that interest rates and recovery rates are deterministic and that they are independent of default times. Similarly to what we did in chapter 3, we have to calculate the premium leg (PL) and the default leg (DL) for an mth to default swap. The expected value of the premium leg is given

CHAPTER 4. SINGLE FACTOR COPULA MODEL by the expected value of the discounted premium cashows:
I

31

E[P Lm ] = E sm i (ti )
i=1

I{ m >t} ,

(4.24)

where sm is the mth to default premium, ti i {1, . . . , I} are the premium payment dates, i = ti ti1 are the payment frequencies and (t) is the discount factor. Using (4.23) in (4.24) we get:
I m1

E[P Lm ] = sm
i=1

i (ti )
k=0

P [N (ti ) = k] .

(4.25)

In order to calculate the default leg, we will assume a homogeneous basket where every obligor has the same notional value of 1 and recovery rate R. This simplies the calculation since now we only need to know (4.23) to price it. The expected value of the default leg is given by the expected value of any default before maturity: E[DLm ] = E[(1 R)( m )I{ m T } ] = (1 R)
0 T

(t) dS m (t),

(4.26)

where T is the maturity date. Integrating (4.26) by parts: E[DLm ] = (1 R) 1 (T )S m (T ) +

0 T

S m (t) d(t) .

(4.27)

Using our assumption of deterministic interest rates, under the risk-neutral measure we have: E[DLm ] = (1 R) 1 erT S m (T ) r
0 T

S m (t)ert dt .

(4.28)

The premium for the basket is given by the ratio of (4.28) and (4.26). sm = E[DLm ] . E[P Lm ] (4.29)

CHAPTER 4. SINGLE FACTOR COPULA MODEL

32

Table 4.1 shows the spread (4.29) in bps/year for the k th default from a basket with 10 obligors. We use a single parameter a = 0.3 for the correlation matrix and the recovery rate is constant at 40%. Notice that from (4.2) we have a = , where is the correlation parameter (3.7) used in the Monte Carlo simulation. Table 4.2 shows the spread in bps/year for dierent correlation factors, with a xed hazard rate 0.03. k 1 2 3 4 5 6 7 8 9 10 h = 0.01 445 140 53 21 8 3 1 0.3 0.1 0 h=0.03 1194 519 266 141 73 36 16 6 2 0.4

Table 4.1: Spreads for a k th to default basket swap on a homogeneous portfolio with 10 instruments, xed correlation factor a = 0.3 and recovery rate 40%.

These tables agree with what we would expect. In table 4.1 a higher hazard rate means a larger spread for all swaps, because the chance of default increases. Table 4.2 shows that increasing the correlation among the instruments in the portfolio causes early default swaps to get cheaper and late ones to get more expensive. This is similar to the behavior observed in the CDOs in Figs. 3.5 and 3.9. This is

CHAPTER 4. SINGLE FACTOR COPULA MODEL k 1 2 3 4 5 6 7 8 9 10 a=0 a= 1880 596 184 45 8 1 0 0 0 0 0.3 a= 0.6

33

1194 519 266 141 73 36 16 6 2 0.4

755 421 277 192 135 93 63 40 22 9

Table 4.2: Spreads for a k th to default basket swap on a homogeneous portfolio with 10 instruments with xed hazard rate 0.03 and recovery rate 40%.

an evidence that their pricing should be similar as we will see in the next section.

4.2

CDO spread

For an homogeneous portfolio with zero recovery rate, a CDO tranche is equivalent to a basket default swap. For the general case of recovery rate dierent than zero, the stochastic process that we have to observe is the total portfolio loss L(t) (3.4). The premium leg for a tranche [A, B] is:
I

PL =
i=1

i (ti ) (B A)I{L(ti )[0,A]} + [B L(ti )]I{L(ti )[A,B]} .

(4.30)

CHAPTER 4. SINGLE FACTOR COPULA MODEL Taking the expected value of (4.30) we have:
I B

34

E[P L] =
i=1

i (ti ) (B A)P [L(ti ) < A] +

A

(B x) dFL (x) ,

(4.31)

where FL is the distribution of L(t). If we consider a homogeneous CDO with each notional being 1, the set of all possible values for L(t) is {0, (1 R), . . . , n(1 R)}. Therefore we can make an association between P [L(t) = C] and P [N (t) = k] from (4.10): P [L(t) = C] = P N (t) = C +1 1R , (4.32)

where x = max{m Z|m x} is the oor function. Using (4.32) we can re-write (4.31) as: A I 1R E[P L] = i (ti ) (B A) P [N (ti ) = k] + i=1 k=0 k=

B 1R

A 1R

[B k(1 R)]P [N (ti ) = k] . +1 (4.33)

Now we have to calculate the default leg. The loss accumulated in a tranche is given by: M (t) = [L(t) A]I{L(t)[A,B]} + (B A)I{L(t)>B} . (4.34)

Since M (t) is an increasing jump process, we can dene a Riemann-Stieltjes integral to represent the default leg [1].
T

DL =
0

(t) dM (t).

(4.35)

Integrating (4.35) by parts and taking into account that (t) = ert we have: E[DL] = erT E[M (T )] + r
0 T

ert E[M (t)] dt.

(4.36)

CHAPTER 4. SINGLE FACTOR COPULA MODEL Tranche 14% - 100% 3% - 14% 0% - 3% MC 35.4 0.6 961 7 4073 26 Factor model 35.8 966 4096

35

Table 4.3: Comparison of spreads calculated with Monte Carlo and a single factor model. CDO has 100 obligors with h = 0.03 each, a = 0.3 and R = 40%.

The expected value of (4.34) is:

B

E[M (t)] = (B A)P [L(t) > B] +

A

(x A) dFL (x).

(4.37)

Applying (4.32) in (4.37):

n
B 1R

E[M (t)] = (B A)
k=
B 1R

P [N (t) = k] +
+1 k=
A 1R

[k(1 R) A] P [N (t) = k].

+1

(4.38) Using (4.10), (4.33), (4.36) and (4.38) we can calculate the tranche spread of a homogeneous CDO. Table 4.3 compares the spreads from table 3.2 with the ones obtained with a single factor model. The values obtained with the factor model agree with the Monte Carlo simulation within one standard deviation.

Chapter 5 Conclusion
In this report we have priced a homogeneous CDO using Monte Carlo simulation and a semi-analytical method. Both prices agreed within one standard deviation, indicating that the factor model is a valid alternative to price a CDO. According to references [1] and [2], the main advantage of the factor model is that it can be much faster than Monte Carlo when pricing a CDO with a large number of instruments (n 100). However some care must be taken in order to achieve this gain. First the factor model should be implemented in C/C++ . Second we should choose an ecient algorithm for the numerical integration, because this is one of the most time consuming parts of the program. The values in this report were obtained using Matlab. At the end we found out that this is not a good method to implementthe factor model. The reason is due to the fact that Matlab cannot numerically integrate functions with matrices in its denition. We had to implement the Simpson algorithm to do the integration. This is not the most ecient way and therefore the factor model in our case was as slow as Monte Carlo. Another point that we must be careful is during the 36

CHAPTER 5. CONCLUSION

37

calculation of P [N (t) = k|V ] (4.16). Notice that when qi (t|V ) goes to zero in (4.11), P [N (t) = 0|V ] goes to zero also and wi (t|V ) in (4.15) goes to innity. However their product in (4.16) should always be upper bounded by one by the denition of probability. In Matlab we have observed that for k 90 and qi (t|V ) 0.05 the value of P [N (t) = k|V ] sometimes diverged ( 1). After doing a revision of the algorithm, we are convinced that this was caused by a oating point error of the product of a very small P [N (t) = 0|V ] and a very large wp(1) . . . wp(k) in (4.16).

We have xed this problem by substituting any value above one in P [N (t) = k|V ] by zero. A last observation to be made is that the gain in speed reported in [1] and [2] was only for the special case of a homogeneous CDO. In the non-homogeneous situation things get more complicated and the gain will probably not be as good. However the factor model is still a valid way to double check results obtained with Monte Carlo and, due to its semi-analytical nature, it has always the potential to speed up the calculation if it is programmed in an ecient way.

Bibliography
[1] Jean-Paul Laurent and Jon Gregory. Basket default swaps, CDOs and factor copulas. Technical report, University of Lyon and BNP Paribas, September 2003. Available in www.defaultrisk.com/ps crderivatives.htm. [2] John Hull and Alan White. Valuation of a CDO and an nth to default CDS without Monte Carlo simulation. Technical report, University of Toronto, November 2003. Available in www.defaultrisk.com/ps crderivatives.htm. [3] Douglas Lucas. CDO Handbook. Technical report, J.P. Morgan, 2001. [4] Domenico nical Picone. Royal Collateralized Bank of Debt Obligations. 2003. Techin

report,

Scotland,

Available

www.defaultrisk.com/ps crderivatives.htm. [5] P.J. Schnbucher. Credit Derivatives Pricing Models. Wiley Finance, 2003. o [6] I. Karatzas and S. Shreve. Brownian Motion and Stochastic Calculus. Springer, second edition, 1991. [7] David Lando. On Cox processes and credit risky securities. Review of Derivatives Research, 2:99120, 1998. [8] R. Nelsen. An Introduction to Copulas. Springer, 1999. 38

BIBLIOGRAPHY

39

[9] P. Embrechts, F. Lindskog, and A. McNeil. Modelling dependence with copulas and applications to risk management. Technical report, September 2001. Available in http://www.math.ethz.ch/ baltes/ftp/papers.html. [10] A. Sklar. Fonctions de rpartition ` n dimensions et leurs marges. Publications e a de lInstitut de Statisque de lUniversit de Paris, 8:229231, 1959. e [11] A. Sklar. Random variables, joint distribution functions and copulas. Kybernetika, 9:449460, 1973. [12] D.X. Li. On default correlation: a copula approach. The Journal of Fixed Income, 9:4354, 2000. [13] Stefano Galiani. Copula functions and their application in pricMas-

ing and risk managing multiname credit derivative products. ters thesis, Kings College London, November 2003. www.defaultrisk.com/ps crderivatives.htm.

Available in

[14] Hans Boscher and Ian Ward. Long or short in CDOs. Risk Magazine, pages 125129, June 2002.