Valuation of Multi-Name

Credit Derivatives

Bernd Will
Exeter College
University of Oxford

A thesis submitted in partial fulfilment of the requirements for the

MSc in
Mathematical Finance

September 29, 2003

 Acknowledgements

I owe my gratitude to my employer, d-fine GmbH, who made my parti-

cipation in the course of Mathematical Finance possible by financing the
course and allowing me to take the necessary time off.
Special thanks to Christian Hille who gave the initial impuls to me to chose
this topic and who supported my work with a lot of useful information.
 “Valuation of Multi-Name Credit Derivatives”
Bernd Will
MSc in Mathematical Finance
Trinity 2003

Abstract

The subject of this thesis is modelling and valuation of multi-name credit

derivatives such as basket default swaps and basket CDS tranches. After
a description of the relevant products, models for valuation are presented.
The key idea of modelling correlated default is the usage of copulae. In this
thesis the valuation models are set up with Gaussian normal-, Student t-
and Clayton copulae. Two different methods for valuation are described:
The first is the standard Monte Carlo method for simulating the default
times, with which multi-name credit derivatives can be priced. In the
second approach the correlation structure is simplified by a factor copula
model, in which semi-explicit fomulae for valuation of the multi-name
credit products can be formulated.
In this thesis the Monte Carlo valuation approach is implemented for the
Gaussian normal-, the Student t- and the Clayton copula; the semi-explicit
approach is implemented for the Gaussian normal copula.
The implemented valuation methods are used to price a simple three credit
basket and a real world basket CDS tranche. It is investigated what
influence parameters like correlation and recovery rates have on the spread
of credit products. For the Student t- and Clayton copula correlation
structure the dependence of the spread on the copula parameters ν and α
is examined.
Contents

1 Introduction 1
1.1 Credit Risk Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Merton’s Structural Approach to Credit Risk . . . . . . . . . . 2
1.1.2 Reduced Form or Hazard Rate Approach . . . . . . . . . . . . 4
1.2 Single-Name Credit Derivatives . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Credit Default Swap . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.2 Pricing Credit Default Swaps . . . . . . . . . . . . . . . . . . 7
1.3 Multi-Name Credit Derivatives . . . . . . . . . . . . . . . . . . . . . 8
1.3.1 Credit Default Swap . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.2 Basket Default Swap . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.3 Basket CDS Tranche or CDO . . . . . . . . . . . . . . . . . . 9

2 Pricing Multi-Name Credit Derivatives 12

2.1 Copulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.1 Sklar’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.2 Gaussian Normal Copula . . . . . . . . . . . . . . . . . . . . . 13
2.1.3 Student t Copula . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.4 Archimedean Copulae . . . . . . . . . . . . . . . . . . . . . . 15
2.1.5 Measure of dependence . . . . . . . . . . . . . . . . . . . . . . 18
2.2 Pricing Basket Default Swaps . . . . . . . . . . . . . . . . . . . . . . 19
2.2.1 Monte Carlo Method . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.2 Semi-Explicit Approach . . . . . . . . . . . . . . . . . . . . . 22
2.3 Pricing Basket CDS Tranches . . . . . . . . . . . . . . . . . . . . . . 25
2.3.1 Monte Carlo Method . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.2 Semi-Explicit Approach . . . . . . . . . . . . . . . . . . . . . 26

i
3 Implementation 28
3.1 Benchmark Results - Reference Case . . . . . . . . . . . . . . . . . . 28
3.2 Monte Carlo Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2.2 Comparison with Reference Case . . . . . . . . . . . . . . . . 31
3.2.3 Convergence of Monte Carlo Method . . . . . . . . . . . . . . 33
3.3 Semi-Explicit Approach . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.1 Semi-Explicit Approach: Basket CDS . . . . . . . . . . . . . . 34
3.3.2 Semi-Explicit Approach: Basket CDS Tranche . . . . . . . . . 35
3.3.3 Comparison with Reference Case . . . . . . . . . . . . . . . . 35
3.3.4 Bimodal and binomial distribution . . . . . . . . . . . . . . . 37
3.3.5 Runtime Comparison . . . . . . . . . . . . . . . . . . . . . . . 38

4 Results 39
4.1 Results for the Reference Case . . . . . . . . . . . . . . . . . . . . . . 39
4.1.1 Gaussian Normal Copula - Basket Spread as Function of the
Correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.1.2 Student t Copula - Influence of ν on the Basket Spread . . . . 40
4.1.3 Clayton Copula - Influence of α on the Basket Spread . . . . . 40
4.2 Valuation of a real world Basket CDS Tranche . . . . . . . . . . . . . 42
4.2.1 Tranche Spread as Function of the Recovery Rates . . . . . . 43
4.2.2 Tranche Spread as Function of the Correlations . . . . . . . . 44
4.2.3 Tranche Spread as Function of the Default Probabilities . . . . 44
4.2.4 Tranche Spread as Function of the regarded Tranche . . . . . 45
4.2.5 Student t Copula - Tranche Spread as Function of ν . . . . . . 46
4.2.6 Clayton Copula - Tranche Spread as Function of α . . . . . . 47

5 Conclusion 48

A Derivation of the default probability Zki (t) 50

B Limitation of Factor Models 52

C Evaluation Data for the real world Basket CDS Tranche 59

References 66

ii
Chapter 1

Introduction

In recent years the credit derivatives market has grown rapidly in volume and com-
plexity of the products offered. While the biggest part of outstanding notionals is still
found in simple products like credit default swaps, complex products having payoff
profiles depending on a whole credit portfolio are becoming more popular. Examples
of such products are basket default swaps (BDS) and basket CDS tranches. A basket
default swap extends the credit protection, which a simple credit default swap (CDS)
grants for a single underlying, to a portfolio of underlyings with the restriction that
the default of only one underlying is compensated. Depending on the ranking of the
protected default, the product is called 1st -to-default basket, 2nd -to-default basket
or, generally, k th -to-default basket. A basket CDS tranche is a generalization of the
basket default swap in the sense that a certain percentile of the portfolio notional
is protected. The price of such products depends on the joint default probability of
the underlyings in the credit portfolio. Modelling joint default is difficult, as observa-
tions of joint defaults are even scarcer than the already rare event of a single default.
Thus, modelling the joint distribution is the crucial point in pricing multi-name credit
derivatives.
The thesis is structured as follows: The subsequent sections of this chapter contain
an outline of the relevant credit models as well as of the products of interest; these are
single-name and multi-name credit derivatives. Chapter 2 contains a description of
the methods used for valuation of multi-name credit derivatives. The implementation
of the valuation methods is given in Chapter 3. The implemented programs are used
to investigate the influence of different parameters on a simple basket default swap
with an underlying portfolio of three credits and a real world basket CDS tranche.
The results are presented in Chapter 4.

1
1.1 Credit Risk Models
Starting in 1974, modelling credit risk has been of central interest to theorists as
well as to practitioners, and different credit risk models have been developed. In the
following sections a short overview over the fundamental ideas of the different models
is given.

1.1.1 Merton’s Structural Approach to Credit Risk

In the structural approach, which was developed by Merton in 1974 [16] and which
is based on the principles of Black and Scholes’ option pricing theory [4], the asset
value V of a company is modelled as Brownian motion where the change in the asset
value dV is given by:

dV = µV dt + V dW (1.1)

The parameters µ and σ denote the drift and the volatility of the Brownian motion.
The liabilities of a company are summarized in a zero coupon bond with face value F
and maturity T . If the value of the assets V at maturity T is smaller than the value
of the liabilities F , the company cannot meet its obligation to pay the liabilities and
the company defaults. If the value of the assets V is greater than the liabilities F ,
then the company can fulfil its obligation to pay its liabilities at maturity T and no
default occurs;

V >F : No Default
V ≤F : Default

In this model the firm’s equity can be regarded as an European call option on its
assets with a strike of F and maturity T . In Merton’s model default can occur only
at maturity T of the zero coupon bond. An extension of Merton’s model - that default
can occur before maturity if the asset value V falls below a certain threshold D - was
provided by Black and Cox [3]. A commercial product which follows this approach is
CreditGrades [20]: The value of the assets V (t) is modelled as random walk (1.1) with
zero drift and default occurs if the value of the assets hits a certain lower threshold
RD for the first time (with recovery R and debt per share D). RD follows a lognormal
random walk with mean R̄D and standard deviation Stdev(ln(RD)) = λ.

V (t) > RD(t) : No Default at time t

V (t) ≤ RD(t) : Default at time t

2
The default time τ is given by the first time the condition V (t) ≤ RD(t) is satisfied.
Summarizing the description, the firm is modelled as a “down-and-out” barrier option
with a time dependent knock-out barrier RD. Using the known distributions for the
first stopping time of a Brownian motion, the survival probability S(t) = 1 − F (t) =
P(τ > t) is
   
ln(d) At ln(d) At
S(t) = Φ − − dΦ − − (1.2)
At 2 At 2

with Φ denoting the cumulative standard normal distribution function, the parame-
(0) λ2
ters A2t = σt + λ2 and d = VRD e (σ denotes the asset volatility and λ the barrier
volatility).

Latent Variable Models Models where the default event depends on the evolu-
tion of a fundamental property of the considered company, are called latent variable
models. The models developed by KMV [11] and CreditMetrics [17] are such la-
tent variable models. In these models the relative change in the latent variables
in a portfolio consisting of m debtors is given by a m-dimensional random vector
X = (X1 , ..., Xm )T where the random variables Xi follow the related marginal distri-
butions and the correlations between the latent variables. For each debtor i a lower
threshold Di (of the relative change of the latent variables) is given and when hit this
leads to default of the debtor. As an indicator whether a debtor has defaulted or not,
a default indicator Yi is used which has a value of 1 in case of default and 0 in case
of no default.

Yi = 1 ⇐⇒ Xi ≤ Di (1.3)

The cumulative distribution P(Xi ≤ Di ) gives the probability that the random
variable Xi is less or equal to the threshold Di . Given historic time dependent default
probabilities1 the threshold Di can be used to calibrate the cumulative distribution
to the given historic default probabilities; for example, if Xi follows a normal dis-
tribution with a time dependent default probability of Fi (t), the threshold equals
Φ−1 (Fi (t)). That is, default occurs if Xi ≤ Φ−1 (Fi (t)),2 where Φ−1 is the inverse
normal cumulative distribution.
1
The default probability F (t) is defined as the probability that the default time τ is smaller than
or equal to time t: F (t) = P(τ ≤ t). The condition Xi ≤ Di at time t is equivalent to the condition
τi ≤ t.
2
As Xi is normally distributed, the probability for condition Xi ≤ Di is P(Xi ≤ Di ) = Φ(Di ).
This probability is equal to P(τi ≤ t) = Fi (t) so that Φ(Di ) = Fi (t) which provides the sought
threshold Di = Φ−1 (Fi (t)).

3
 This last step of linking the thresholds to default probabilities is not part of
standard Merton’s model. Some authors, like Mashal and Naldi [14], name this
approach “hybrid approach”. The name derives from the fact that this approach
uses typical features of Merton’s approach (default happens if V falls below a certain
threshold D) as well as typical features of the hazard rate model (calibration to
market implied default probabilities).
If the evolution of the asset values of several companies is considered, the asset
values of the companies cannot change independently as each company acts in the
same macroeconomic environment which has influence on the evolution of the asset
values. Thus the Wiener processes W , which determine the random walk of the
asset value V , cannot be independent but have to be correlated. As the value of the
latent variable relative to the threshold determines whether a firm defaults or not,
the dependence structure of default is given by the dependence structure of the latent
variables.

Correlation structure in Merton’s Approach In Merton’s approach it is not

only assumed that the marginal distribution of the latent variables is a normal distri-
bution but also that the probability of several debtors defaulting can be modelled by
a multi-variate Gaussian distribution. For a two debtor universe the probability that
debtor 1 and debtor 2 default is given by the two dimensional Gaussian distribution:

Zx1 Zx2
s2 − 2ρ12 st + t2
 
1
P(X1 ≤ x1 , X2 ≤ x2 ) = exp − ds dt (1.4)
2(1 − ρ212 )
p
2π 1 − ρ212
−∞ −∞

In the equation above ρ12 is the correlation between default of credit 1 and default of
credit 2.
In order to determine in Merton’s model which credits have defaulted at time T ,
correlated random numbers following a standard normal distribution have to be gen-
erated for the change in the latent variables Xi and the relative changes in the latent
variables have to be compared with the lower thresholds.

1.1.2 Reduced Form or Hazard Rate Approach

In contrast to the structural approach, where the default probability is conditioned
explicitly on the value of the regarded firm, reduced form models use actual credit
prices to extract the required default probabilities. In the model developed by Duffie
and Singleton [6] default events are determined by the hazard rate function h(t). The

4
hazard rate function gives the instantaneous default probability for a certain time t
conditional on no default before time t,

F (t + δt) − F (t)
h(t) = P[t < τ ≤ t + δt|τ > t] = (1.5)
1 − F (t)
f (t)δt
= (1.6)
1 − F (t)
S 0 (t)
= − (1.7)
S(t)

The distribution function F (t) gives the probability that default happens before or
at time t. The survival function S(t) gives the probability that default happens after
time t. With the equations above the distribution function F (t) = P(τ ≤ t) of τ
(=default time) and the survival function S(t) = P(τ > t) = 1 − F (t) can be related
to the hazard rate function h(t) as follows:
Rt
− h(s)ds
S(t) = e 0 (1.8)
Rt
− h(s)ds
F (t) = 1 − e 0 (1.9)

F (t) is the marginal distribution of the default time for a certain credit; this marginal
distribution can be computed from market-quoted credit default swap spreads or
defaultable bonds [9]. With F (t) given for all credits, the joint distribution function
can be derived from the marginal distributions via a copula approach.

1.2 Single-Name Credit Derivatives

In this chapter single-name derivatives - in the sense that there is only one underlying
product - are introduced. The name is somewhat misleading as even in a single-name
derivative normally there are at least two counterparties involved (e.g. in a credit
default swap there is the reference entity of the underlying and the CDS issuing
party).

1.2.1 Credit Default Swap

A credit default swap (CDS) is a product which gives the protection buyer the right to
demand compensation payments from the protection seller if the underlying defaults.
This protection is guaranteed until the maturity T of the credit default swap. In
return for the guaranteed protection, the protection buyer has to pay a premium to

5
 Premium
Protection Seller Protection Buyer
Compensation

Underlying Credit

Figure 1.1: Schematic picture of a credit default swap (CDS).

the protection buyer until maturity or until the underlying defaults. This premium
payment is called spread payment. The periodic payments of the protection buyer
to the protection seller are the fixed side of the swap and denoted as premium leg.
The compensation payment of the protection seller to the protection buyer in case
of default of the underlying credit is the floating side of the swap and is also called
default leg.
To price a credit default swap the following steps have to be accomplished:

ˆ Determine the default probability until maturity.

ˆ Determine the present value of the fixed leg of the CDS (these are the payments
the protection buyer pays to the protection seller).

ˆ Determine the present value of the float leg of the CDS (this is the payment the
protection buyer receives from the protection seller in case of default).

An elementary way to get the required default probabilities as function of time is to

discount the cashflows of market-quoted bonds of the underlying party to today (with
the risk free rate). The present value calculated in this way does not enclose the credit
risk and is too big compared to the market price. The factors with which the dis-
counted cashflows must be multiplied in order to make their sum equal to the market
price are the default probabilities. To get the term structure of credit default proba-
bilities, a procedure which is comparable to bootstrapping coupon bearing bonds to
get the zero yield curve (for further information see [9]) has to be accomplished:

ˆ Sort bonds with respect to their maturities from small to big maturities.

6
 ˆ Compute the present value of the first bond by discounting with the risk free
rate.

ˆ Find the factor with which the present value has to be multiplied to make the
product equal to the market price of the bond.

ˆ Compute the present value of the next bond. Coupons are discounted by multi-
plying with the risk free rate and the corresponding already determined default
probability.

ˆ Find the factor with which the nominal has to be multiplied to make the sum
of all coupons and the nominal equal to the market price.

1.2.2 Pricing Credit Default Swaps

A credit default swap guarantees that the protection buyer receives a compensation
payment from the protection seller if the underlying security defaults before maturity
T . The price of the default protection is the premium s (spread) which has to be
paid from the protection buyer to the protection seller periodically.
The following properties are required to compute the present value of the fixed
and float leg of credit default swaps:

T Maturity of the credit default swap

N Nominal of the credit default swap
tj Time of the jth premium payment
∆j−1,j Time between payment j − 1 and payment j in years
τ Default time of the underlying credit
P(τ ≤ tj ) Risk neutral probability that underlying defaults before or at time tj
R Recovery Rate
B(0, tj ) Discount factor
s Spread

With the properties given above, the present value of the fixed leg (=premium leg)
can be computed by:
X
P VP remium = sN B(0, tj )∆j−1,j (1 − P(τ ≤ tj )) (1.10)
j

This is the present value of the payments the protection seller receives from the
protection buyer. The present value of the payments the protection buyer receives in

7
case of default is:
ZT
P VDef ault = N B(0, t)(1 − R)P(τ = t)dt (1.11)
0

As the present values of the two legs have to be equal at initiation of a fair CDS, the
spread to be paid on the nominal from the protection buyer to the protection seller
is given by:

RT
B(0, t)(1 − R)P(τ = t)dt
0
s= P (1.12)
B(0, tj )∆j−1,j (1 − P(τ ≤ tj ))
j

1.3 Multi-Name Credit Derivatives

In this section multi-name credit products such as

ˆ Credit default swap3

ˆ Basket default swap

ˆ Basket CDS tranche or Collateralized Debt Obligation (CDO)

are introduced.

1.3.1 Credit Default Swap

As already mentioned in the section about single-name credit derivatives, in reality
there is nothing like a single-name credit derivative. Even for a credit default swap
there are two parties to be taken into account: The first is the party of the referenced
underlying, and the second is the issuer of the credit default swap. If the issuer
defaults before maturity, the credit default swap is worthless to the protection buyer
as he will receive no compensation payments in case of default of the referenced
underlying. Thus, to be exact, one has to regard the credit default swap as a credit
product with two underlying parties and therefore one cannot neglect the correlation
between the party of the referenced underlying and the issuer of the credit default
swap. For an exact treatment refer to [10].
3
If one pays attention to the fact that the counterparty in the credit default swap is also prone
to default, the CDS also has to be regarded as multi-name credit derivative.

8
1.3.2 Basket Default Swap
A basket default swap (BDS) is a product that, like a plain vanilla credit default swap,
guarantees protection against loss in case of default. In contrast to a credit default
swap that gives protection against losses in one underlying, a basket credit default
swap guarantees protection against losses in all underlying credits that are contained
in the basket. There are different kinds of basket default swaps - the most popular
one is the 1st -to-default BDS, which gives the protection buyer the right to claim
compensation for the losses in the first credit defaulted. Similarly, a 2nd -to-default
BDS covers the losses of the second underlying defaulted. In general these products
are called k th -to-default basket default swap.

Premium
Protection Seller Protection Buyer
Compensation

Pool of Underlying
Credits

Figure 1.2: Schematic picture of a basket default swap (BDS).

1.3.3 Basket CDS Tranche or CDO

A basket CDS tranche is constructed as slice of a credit portfolio containing a large
number n of credits. The basket CDS tranche is defined by its lower boundary a and
its upper boundary b. The tranche notional at time t is not affected by a default if
the total portfolio loss before default is bigger than b or if the total portfolio after
loss is smaller than a. The tranche notional of a basket CDS tranche has the same
profile as a put spread (Figure 1.3).
In general there are at least three tranches:

ˆ Equity tranche: This tranche starts from a = 0 and covers the very first losses
in the portfolio.

9
 Tranche notional

a b Portfolio Loss
Figure 1.3: Tranche notional of a basket CDS tranche with lower boundary a and
upper boundary b.

ˆ Mezzanine tranche: If the portfolio losses have already consumed the complete
Equity tranche notional, subsequent defaults are compensated with the Mezza-
nine tranche notional. Expressed in fractions of the total nominal this tranche
is characterized by a > 0 and b < 100%.

ˆ Senior tranche: The senior tranche notional is affected by credit defaults only
if the notional of all other tranches has already be consumed to cover losses.

The buyer of the tranche can also be called protection seller as he compensates the
seller of the tranche for all losses in the tranche (the tranche notional decreases).
In return the buyer of the tranche receives periodic payments from the seller of the
tranche on the remaining tranche notional.
Like a credit default swap the basket CDS tranche consists of two sides: The
regular payments of the protection buyer (tranche seller) are the fixed side. The
payment of the protection seller (tranche buyer) in case of default is the float side.
For each side there are two ways a default can influence the notional:

ˆ Binary: In case of default the notional decreases with the full amount of the
credit. Recovered parts of the credit are not taken into account.

ˆ Recovery: In case of default the decrease in the notional is given by the part of
the credit that cannot be recovered.

It is not necessary that the way a default affects the notional is the same for both
sides. For example, the effect of a default on the fixed side (premium payments) can
be of binary-type (the amount, the fixed side notional decreases, is the notional of

10
the defaulted credit no matter how much of the credit can be recovered) whereas the
effect on the float side (default payment) can be recovery-type. Such a setup would
be necessary, for example, if a basket CDS has to be modelled as basket CDS tranche.
If such a tranche must be priced, two tranche notionals must be distinguished: The
fixed side tranche notional and the float side tranche notional.

11
Chapter 2

Pricing Multi-Name Credit

Derivatives

With the increasing popularity of credit products having a payoff profile depending
on a whole credit portfolio modelling correlated defaults has also gained importance.
Examples of such products are basket default swaps (k th -to-default swap) and basket
CDS tranches or CDOs (Collateralized Debt Obligations). To price such multi-name
credit products, it is not suffcient to know the default probabilities for all credits,
but it is also necessary to determine the joint distribution of all credits in order to
evaluate correlated default effects.
The next sections contain an outline of the theoretical background for pricing multi-
name credit derivatives. The key idea of modelling correlated default, the so-called
copulae, are described. After the copula concept has been introduced, pricing basket
default swaps and basket CDS tranches are discussed in detail. These concepts are
used to implement different programs to price k th -to-default basket default swaps and
basket CDS tranches. The implementation is described in Chapter 3.

2.1 Copulae
A useful concept for computing joint distribution functions are copulae (a detailed
description can be found in [18]). Copulae can be used to generate the joint distribu-
tion function of a credit portfolio containing n credits if the marginal distributions of
all credits are known. Mathematically, a copula is a function [0, 1]n → [0, 1] with the
following properties:

C(u1 , u2 , ..., 0, ..., un ) = 0 (2.1)

C(1, . . . , 1, uk , 1, . . . , 1) = uk (2.2)

12
Furthermore, for every ā and b̄ in [0, 1]n such that ai ≤ bi for all i, the volume of the
n-box [ā, b̄] is V ([ā, b̄]) ≥ 0.

2.1.1 Sklar’s Theorem

The probably most important property of copulae is summarized in Sklar’s theorem:
Sklar’s theorem states that for any n-dimensional distribution function H(x1 , ..., xn )
with marginal distribution functions F1 (x1 ), ..., Fn (xn ) there exists a n-dimensional
copula, such that

H(x1 , ..., xn ) = C(F1 (x1 ), ..., Fn (xn )) (2.3)

Using different copulae C1 , C2 , .. different joint distribution functions can be gener-

ated having the same marginal distribution functions F1 (x1 ), ..., Fn (xn ) of the random
variables X1 , . . . , Xn .

2.1.2 Gaussian Normal Copula

The most popular copula is the Gaussian normal copula, which models joint distribu-
tion as n-dimensional Gaussian normal distribution. For the general case the copula
of a n-dimensional distribution is given by
Φ−1
Z (u1 ) Φ−1
Z (un )  
1 1 −1
CΣ (u1 , . . . , un ) = n√ ... exp − x̄Σ x̄ dx1 . . . dxn (2.4)
(2π) 2 det Σ 2
−∞ −∞

where Σ is the linear correlation matrix.

Although often not explicitly mentioned, the Gaussian normal copula is used in
many models. For example, in the models of KMV [11] and CreditMetrics [17] the
joint distribution is also a multi-variate normal, and if one is familiar with market
Value-at-Risk calculations, one recognizes the algorithm given above to be exactly
the same as for modelling the evolution of market risk factors.
Equation (2.4) describes the n-dimensional joint distribution for a Gaussian nor-
mal copula, but this is not sufficient to create correlated random numbers following
the joint distribution of a Gaussian normal copula. These numbers are required for
Monte Carlo simulations. An algorithm generating the required random numbers Xi
can be formulated as follows:

ˆ Generate the Cholesky decomposition of the correlation matrix Σ = AAT ;

13
 ˆ Generate vector z̄ as m independent standard normally distributed random
numbers z1 , . . . , zm ∈ N (0, 1);

ˆ Generate vector X̄ as X̄ = Az̄.

Having performed all steps above, correlated random numbers fitting the joint normal
distribution can be found in Xi .

2.1.3 Student t Copula

Besides the Gaussian normal copula there are other copulae which can be used to gen-
erate joint distribution functions. The Student t distribution is a generalization of the
normal distribution in the sense that the Student t distribution contains the Gaus-
sian normal distribution as limit. The Student t distribution contains the additional
parameter ν, named degrees of freedom. With ν → ∞ the Student t distribution con-
verges to the normal distribution. Compared to the normal distribution the Student
t distribution shows higher probabilities at the margins1 .
The probability density function of a Student t distribution is given by
− ν+1
Γ( ν+1 ) x2
 2
fν (x) = √ 2 ν 1+ (2.5)
νπΓ( 2 ) ν

where Γ(x) is the Gamma function

Z∞
Γ(x) = tx−1 e−t dt (2.6)
0

If the Γ function is used with even values of ν, then only the following properties
of the Γ function are required:

Γ(x + 1) = xΓ(x) (2.7)

Γ(1) = 1 (2.8)
√
Γ(0.5) = π (2.9)

The Student t distribution is

Zx
tν (x) = fν (y)dy (2.10)
−∞

1
The higher probability mass at the margins is often called “fat tails”.

14
In the case n = 2 the Student t copula is given by
t−1 −1
Z (u) tνZ (v)
ν − ν+2
s2 − 2ρ12 st + t2

1 2
Cν,ρ12 (u, v) = 1+ ds dt (2.11)
ν(1 − ρ212 )
p
2π 1 − ρ212
−∞ −∞

where ρ12 is the correlation between random variable 1 and random variable 2.
The procedure to generate random numbers fitting a Student t copula is similiar
to the generation of standard normally distributed random numbers and comprises
the following steps:

ˆ Generate the Cholesky decomposition of the correlation matrix Σ = AAT ;

ˆ Generate vector z̄ as m independent standard normally distributed random

numbers z1 , ..., zm ∈ N (0, 1);

ˆ Generate one χ2ν distributed random number s;

ˆ Generate vector ȳ as ȳ = Az̄;

√
ˆ Generate vector x̄ as x̄ = ν
√ ȳ.
s

2.1.4 Archimedean Copulae

In order to introduce the Archimedean copula family, a continuous, strictly decreasing
and convex function ϕ(u) : [0, 1] → [0, ∞) for all u ∈ [0, 1] is considered. The function
ϕ is called generator of the copula. The pseudo-inverse of ϕ defined as
n ϕ−1 (z) if 0 < z ≤ ϕ(0)
−1
ϕ (z) =
0 if ϕ(0) ≤ z < ∞

is required if ϕ(0) = ∞. If ϕ(0) 6= ∞, then ϕ−1 is the ordinary inverse function.

With ϕ and ϕ−1 the function C : [0, 1]n → [0, 1]

C(u1 , u2 , . . . , un ) = ϕ−1 (ϕ(u1 ) + ϕ(u2 ) + . . . + ϕ(un )) (2.12)

is an Archimedean copula if ϕ−1 is monotone. Examples of the Archimedean copula

family are

Gumbel Copula The Gumbel copula has the generator ϕ(u) = (− ln(u))α with
α ∈ (1, ∞) and is defined as
n 1
o
α α α
C(u1 , u2 ) = exp −[(− ln(u1 )) + (− ln(u2 )) ] (2.13)

15
Clayton Copula The Clayton copula is defined via the generator ϕ(u) = u−α − 1.
For α ∈ (0, ∞) the Clayton copula can be written as
1
C(u1 , u2 ) = (u−α −α
1 + u2 − 1)
−α
(2.14)
 
Frank Copula For the Frank copula the generator is ϕ(u) = ln exp(−αu)−1
exp(−α)−1
with
α ∈ R\{0}. The Frank copula is
 
1 (exp(−αu1 ) − 1)(exp(−αu2 ) − 1)
C(u1 , u2 ) = − ln 1 + (2.15)
α exp(−α) − 1

General algorithm to create random numbers for a given copula Embrechts

et al. [7] propose an algorithm with which random numbers can be created for any
copula. They point out that this algorithm is efficient only under certain conditions (it
must be possible to write Ck−1 (·|u1 , . . . , uk−1 ) in closed form). First the k-dimensional
marginal distribution of C is defined as

Ck (u1 , u2 , . . . , un ) = C(u1 , . . . , uk , 1, . . . , 1) (2.16)

The random numbers U1 , . . . , Un follow the joint distribution C. The conditional

distribution of Uk for given U1 , . . . , Uk−1 is

Ck (uk |u1 , . . . , uk−1 ) = P(Uk ≤ uk |U1 = u1 , . . . , Uk−1 = uk−1 ) (2.17)

∂ k−1 Ck (u1 ,...,uk ) (k−1)
∂u1 ...∂uk−1 Ck (u1 , . . . , uk )
= ∂ k−1 Ck−1 (u1 ,...,uk−1 )
= (k−1)
(2.18)
∂u1 ...∂uk−1
Ck−1 (u1 , . . . , uk−1 )

In order to generate random numbers, the following steps have to be accomplished:

ˆ Create n independent uniform random numbers v1 , . . . , vn ∈ U (0, 1);

ˆ First random number: u1 = v2 ;

ˆ Second random number u2 ∈ C2 (u2 |u1 )

with v2 = C2 (u2 |u1 ) follows u2 = C2−1 (v2 |u1 )

ˆ Third random number u3 ∈ . . .

ˆ nth random number un ∈ Cn (un |u1 , . . . , un−1 )

with vn = Cn (un |u1 , . . . , un−1 ) follows un = Cn−1 (vn |u1 , . . . , un−1 )

16
In case of an Archimedean copula Ck−1 (uk |u1 , . . . , uk−1 ) can be written in closed form
and thus the algorithm described above can be used to generate random numbers.
For an Archimedean copula (2.12) the conditional distribution Ck (uk |u1 , . . . , uk−1 )
can be formulated as
ϕ−1(k−1) (ϕ(u1 ) + ϕ(u2 ) + . . . + ϕ(uk ))
Ck (uk |u1 , . . . , uk−1 ) = (2.19)
ϕ−1(k−1) (ϕ(u1 ) + ϕ(u2 ) + . . . + ϕ(uk−1 ))

For the Clayton copula with the generator ϕ(u) = u−α − 1 and its inverse ϕ−1 (v) =
1 1
(v+1)− α the first derivative of ϕ−1 is ϕ−1(1) (v) = − α1 (v+1)− α −1 and the k th derivative
of ϕ−1 is

(α + 1)(α + 2) · . . . · (α + k − 1) 1
ϕ−1(k) (v) = (−1)k (v + 1) −α −k
(2.20)
αk
According to the above algorithm the following steps have to be performed:

ˆ Generate n independent uniform random variables (v1 , v2 , . . . , vn ) from U (0, 1)

ˆ The first random variable is u1 = v1

ϕ−1(1) (c2 )
ˆ Set v2 = C2 (u2 |u1 ) = ϕ−1(1) (c1 )
. With c1 = ϕ(u1 ) = u−α
1 − 1 and c2 = ϕ(u1 ) +
ϕ(u2 ) = u−α −α
1 + u2 − 2 this is

− α1 −1
u−α −α

1 + u2 − 1
v2 = (2.21)
u−α
1

This can be solved in u2 giving

  − α  − α1
u2 = v1−α v2 α+1 − 1 + 1 (2.22)

ˆ ...
ϕ−1(n−1) (cn )
ˆ Set vn = Cn (un |u1 , . . . , un−1 ) = ϕ−1(n−1) (cn−1 )
.

− α1 −n+1
u−α −α −α −α
1 + u2 + . . . + un−1 + un − n + 1

vn = (2.23)
u−α −α −α
1 + u2 + . . . + un−1 − n + 2

This can be solved in un giving

n α

 α(1−n)−1  o− α1
un = u−α
1 + u−α
2 + ... + u−α
n−1 − n + 2 vn −1 +1 (2.24)

17
2.1.5 Measure of dependence
Given two random variables X1 and X2 , there are different measures characterizing the
dependence structure between X1 and X2 . For some distributions, like for example the
elliptical distributions (e.g. the Gaussian normal or Student t distribution), the linear
correlation coefficient provides a good method to describe dependence. As pointed out
by Embrechts et al. [7] for other distributions like the Archimedean distributions the
linear coefficient is inappropriate or can even be misleading2 . For such distributions
Kendall’s tau and Spearman’s rho are alternatives to describe dependence.

Pearson’s correlation coefficient ρ The most popular correlation measure is

Pearson’s linear correlation coefficient. For two random variables X1 and X2 with
finite variances the linear correlation coefficient is
Cov(X1 , X2 ) E(X1 , X2 ) − E(X1 )E(X2 )
ρ(X1 , X2 ) = p = p (2.25)
Var(X1 )Var(X2 ) Var(X1 )Var(X2 )

Kendall’s tau Two observations (x1 , x2 ) and (x01 , x02 ) of the random vector (X1 , X2 )
are concordant if (x1 − x01 )(x2 − x02 ) > 0 and discordant if (x1 − x01 )(x2 − x02 ) < 0.
Kendall’s tau is defined for the random variables X1 and X2 as the probability of
concordance minus the probability of discordance:

τ (X1 , X2 ) = P[(X1 − X10 )(X2 − X20 ) > 0] − P[(X1 − X10 )(X2 − X20 ) < 0] (2.26)

where (X10 , X20 ) is a independent copy of (X1 , X2 ).

With the copula C(u1 , u2 ) modelling the correlation structure of the random vari-
ables X1 and X2 this can be written as
ZZ
τ (X1 , X2 ) = 4 C(u1 , u2 )du1 du2 − 1 (2.27)
[0,1]2

Kendall’s tau is equal to −1 if two random variables are countermonotonic, +1 if two

random variables are comonotonic and 0 if two random variables are independent.
α
For the Clayton copula Kendall’s tau is equal to τ = α+2
. Thus the parameter
α describes the dependence in a Clayton copula correlation structure; increasing α
has the same qualitative effect as increasing correlation within the Clayton copula
correlation structure.
2
An example, how the usage of the linear correlation coefficient for Archimedean copulae results
in misleading implications, can be found in [7].

18
Spearman’s rho Spearman’s rho is defined as

ρS (X1 , X2 )
= P[(X1 − X10 )(X2 − X̄2 ) > 0] − P[(X1 − X10 )(X2 − X̄2 ) < 0] (2.28)
3
where (X10 , X20 ) and (X̄1 , X̄2 ) are independent copies of (X1 , X2 ).

2.2 Pricing Basket Default Swaps

A k th -to-default basket default swap gives protection against the k th default in the
underlying pool of credits. Because of this, the whole joint distribution of all under-
lyings must be taken into account. There are two ways of computing the value of a
basket default swap:

ˆ Simulate a large number of random scenarios with random numbers following

the joint distribution.

ˆ Use a factor model to make semi-explicit computations tractable.

2.2.1 Monte Carlo Method

The present value of a k th -to-default basket depends on the time the k th credit de-
faults. In order to determine the default time of the k th credit, the default times of
all credits in the underlying basket must be known. Pricing a k th -to-default basket
default swap thus can be split up into the following steps:

ˆ Generate the correlated default times τi for all underlyings in the basket

ˆ Sort the credits with respect to their default time τi

ˆ Determine the k th default time τ k

ˆ Determine the present value of the premium leg

ˆ Determine the present value of the default leg

ˆ Repeat all steps above until the required number of scenarios has been simulated

The subsequent sections give a detailed description of these steps.

19
Generating the correlated default times At this point the marginal distribution
of the underlying credits as well as the chosen model for the joint distribution become
important. In general this step consists of creating random numbers, which are
transformed to follow the joint distribution, and computing the default times τi out
of the random numbers. In order to make things a bit more specific, a Gaussian
normal copula is considered. As already described in Section 2.1.2 about Gaussian
normal copulae the computation consists of

ˆ generating correlated normal random numbers as in Section 2.1.2

ˆ generating [0, 1] distributed numbers ui = Φ(xi ) where Φ(x) is the normal

distribution function

ˆ computing the default times τi = Fi−1 (ui ) where Fi (t) = P(τi ≤ t) is the
marginal distribution of credit i

With the last two steps the realization xi of the latent variable Xi is linked to the
default probabilities of the underlying credit and the default time τ is computed.3
The vector τ̄ contains the default times τi of all n underlyings in the basket. For
a k th -to-default basket default swap the default times have to be brought into an
ascending order, and the k th credit defaulting at time τ k has to be found.

Computing the present value of the premium leg The payments of the pre-
mium leg are the compensation the protection seller receives for taking over the credit
risk of the underlying. The premium is paid as long as the underlying credit has not
defaulted but not longer than to the maturity of the contract. Whether accrued
premium payments between payment dates are taken into account, depends on the
contractual agreement. With the k th default time τ k given, the present value of the
premium leg can be computed by
X
P VP remium = sN B(0, tj )∆j−1,j (2.29)
j

where N is the nominal of the BDS, tj are the payment dates of the premium leg
(tj ≤ T ), s is the percentile amount of the nominal to be paid at payment dates and
3
In the section about Merton’s model the equation Di = Φ−1 (Fi (t)) is given: Expressed in terms
of the latent variable default occurs if Xi ≤ Di = Φ−1 (Fi (t)). Let τi be the time the threshold Di is
hit and default occurs. Then the above equation can be reformulated as τi = Fi−1 (Φ(Di )): In this
context default occurs if t ≥ τi = Fi−1 (Φ(Di )).

20
∆j−1,j is the time between two premium payments:
n (tj − tj−1 ) 1{τ k >tj }
∆j−1,j =
(τ − tj−1 ) 1{τ k ≤tj ∧τ k >tj−1 } (accrued premium payment)

If τ k ≤ tj and accrued premium payments are agreed, the year fraction ∆j−1,j is
τ − tj−1 ; if accrued premium payments are not agreed, ∆j−1,j is 0.

Computing the present value of the default leg With the k th default time
τ k given, the present value of the default leg can be computed with the subsequent
equation

P VDef ault = B(0, τ k )N (1 − Rk )1{τ k ≤T } (2.30)

where T is the maturity of the BDS, Rk is the recovery rate of the k th defaulted credit
and B(0, t) is the discount factor, which gives the present value of one unit paid at
time t.

Pricing the Basket Default Swap The term of pricing a basket default swap
can be understood in two ways:

ˆ For a given BDS with a fixed spread rate the value of the contract has to be
determined. This can be done by calculating the difference between the present
values of the default and premium leg. From the protection seller’s point of
view this can be formulated as
N N
!
1 X X
PV = P VP remium (i) − P VDef ault (i) (2.31)
N i=1 i=1

where i is the index of the scenario.

ˆ For a given BDS the basket spread rate has to be determined, which makes the
BDS worthless for both sides. This can be done by dividing the present value
of the default leg through the present value of the premium leg (calculated with
s = 1 in Equation 2.29).
P
P VDef ault (i)
s= Pi (2.32)
i P VP remium (i)

21
2.2.2 Semi-Explicit Approach
While the Monte Carlo method prices credit products by simulating the default times
and calculating the price of the product with respect to the simulated default times
in a large number of scenarios, in the semi-explicit approach developed by Laurent
and Gregory [12] the probabilities required to compute the present value of the k th -to-
default basket CDS, namely the probability P(N (t) = m) that a certain number of
credits has defaulted at time t and the probability Zki (t) that at time t credit i has
defaulted as k th credit in the basket, are determined. Given these probabilities, the
present value of the premium and default leg can be calculated analytically.

Factor Copulae In order to determine the required default probabilities, Laurent

and Gregory use a factor copula approach to model Xi , which is the default deter-
mining property of a credit. In the factor model Xi consists of two parts: The first
part is a for all credits identical Gaussian random variable V and the second part is
a for each credit specific Gaussian random variable Vi :
q
Xi = βi V + Vi 1 − βi2 (2.33)

The factor βi determines how strong Xi is linked to the evolution of the global random
variable V . V and Vi are independent standard normally distributed random numbers.
Because of this, the correlation between two credits is Cov(Xi , Xj ) = βi βj . The link
between the Xi and the default time τi is given via the marginal distribution of the
default time P(τi ≤ t) = Fi (t):
q
Xi = βi V + Vi 1 − βi2 ≤ Φ−1 (Fi (t)) (2.34)

With the equation above the condition that credit i defaults can be expressed in terms
of Vi as

Φ−1 (Fi (t)) − βi V

Vi ≤ p ⇐⇒ τi ≤ t (2.35)
1 − βi2
i|V
The conditional default probability pt that credit i defaults at time t conditional
on V is
!
i|V Φ−1 (Fi (t)) − βi V
pt = P(τi ≤ t|V ) = Φ p (2.36)
1 − βi2

22
Because of the independence of V and Vi the Gaussian factor copula can be written
as:
n
!!
Φ−1 (Fi (t)) − βi v
Z Y
C(u1 , . . . , un ) = Φ p ϕ(v)dv (2.37)
i=1 1 − βi2

where ϕ(v) is the normal probability density function of v.

Probability P(N (t) = k) of having k credits defaulted at time t Let N (t)

P
be the number of defaulted credits at t: N (t) = i Ni (t) with Ni (t) = 1{τi ≤t} . The
probability P(N (t) = k) is required to compute the present value of the premium leg
of a basket default swap. Furthermore, P(N (t) = k) is required for a homogeneous
basket default swap to compute the present value of the default leg.
Using a probability generating function approach, the probability P(N (t) = k)
that k credits have defaulted at time t can be determined:
n
X
N (t)
ψN (t) (u) = E[u ]= P(N (t) = k)uk (2.38)
k=0

This can be written as

" " ##
Y
N (t) N (t) Ni (t)
E[u ] = E[E[u |V ]] = E E u |V (2.39)
i

i|V i|V
and if it is taken into account that E[uN i(t) |V ] = 1 − pt + u pt , the probability
generating function looks like
n
" n #
X Y i|V i|V

ψN (t) (u) = E[uN (t) ] = P(N (t) = k)uk = E 1 − pt + u pt (2.40)
k=0 i=1

The required probability P(N (t) = k) can be determined by calculating the coefficient
of the uk term.

Present value of the premium leg If the probability P(N (t) = k) is known,
the present value of the premium leg can be determined. The basket default swaps,
regarded by Laurent and Gregory, have the following modalities:

ˆ The premium s is paid on the protected nominal N ;

ˆ The premium is paid at certain payment dates tj ;

ˆ For simplification reasons Laurent and Gregory neglect accrued premium pay-
ments;

23
 ˆ The time between the payment dates tj−1 and tj is given as year fraction ∆j−1,j ;

ˆ B(0, t) is the discount factor which discounts a payment at time t to time 0.

With these assumptions the present value of the premium leg is

X
P V = sN ∆j−1,j B(0, tj )P(N (tj ) < k) (2.41)
j

X k−1
X
= sN ∆j−1,j B(0, tj ) P(N (tj ) = i) (2.42)
j i=0

Determining the default probability Zki (t). In order to evaluate the default leg
for an inhomogeneous (k + 1)th -to-default basket default swaps, the probability that
credit i defaults at time t having k credits already defaulted before time t is required
for all credits in the basket. This probability is denoted as Zki (t) and can be written
with Ni (t) = 1{τi ≤t} and N (−i) (t) = j6=i Nj (t) as
P

1
Zki (t) = lim P(Ni (t0 ) − Ni (t) = 1, N (−i) (t) = k) (2.43)
t →t t0
0 −t
P(Ni (t0 ) − Ni (t) = 1, N (−i) (t) = k) can be computed using the joint probability
generating function of (Ni (t0 ) − Ni (t), N (−i) (t)) defined by:
h 0 (−i)
i
ψ(u, v) = E uNi (t )−Ni (t) v N (t) (2.44)

After some computations (refer to Appendix A), the subsequent equation can be
determined, in which the required probability Zki (t) emerges:
n−1
" #
i|V Y 
X dp t j|V j|V

Zki (t)v k = E 1 − pt + pt v (2.45)
k=1
dt j6=i

The probability that credit i defaults as (k + 1)th is given by the coefficient of the v k
term.

Default Leg With the default probability Zki (t) given for all credits in the basket,
the present value of the default leg can be determined. The fraction of the nominal
which is paid in case of default is given by the recovery rate Ri . As consequence the
protection seller has to pay Mi = Ni (1−Ri ) in case of default to the protection buyer.
With the properties defined above, the value of the default leg is
ZT X
n
i
PV = B(0, t)Mi Zk−1 (t)dt (2.46)
0 i=1

where k is the order of the basket (k th -to-default).

24
2.3 Pricing Basket CDS Tranches
A basket CDS tranche, as described in detail in Section 1.3.3, gives protection to
the protection buyer against losses between the lower boundary a and the upper
boundary b. In this section a basket CDS tranche is analyzed from the perspective of
the tranche buyer (that is the protection seller). The tranche buyer receives regular
payments at times tj on the notional remaining in the tranche from the tranche seller.
Pricing basket CDS tranches can be accomplished in two ways:

ˆ In Monte Carlo simulations the loss as well as the tranche notional are computed
for every scenario individually. With equation (2.51), (2.52) and (2.53) the
present value of the tranche is computed.

ˆ In the semi-explicit approach the loss distribution is computed via the char-
acteristic function of the loss. A Fourier Transformation of the characteristic
function provides the loss distribution. With the loss distribution the expected
tranche notionals of the float and fixed leg can be calculated. If the expected
tranche notionals are known, the present value of the basket tranche can be
determined.

In the subsequent sections the two techniques to evaluate basket CDS tranches are
described in detail.

2.3.1 Monte Carlo Method

In order to get the loss for a credit portfolio containing n credits, the default time τi
must be determined for all n credits. As the manner a default influences the tranche
notional can be different for fixed and float side, two portfolio losses are distinguished.
Let Lf ixed (t) be the portfolio loss for the fixed side at time t and Lf loat (t) the portfolio
loss for the float side.
n
X
Lf ixed (t) = Ni (1 − Ri 1{f ixed type=recovery} ) 1{τi ≤t} (2.47)
i=1
n
X
Lf loat (t) = Ni (1 − Ri 1{f loat type=recovery} ) 1{τi ≤t} (2.48)
i=1

In the equations above Ni is the nominal of credit i, Ri is the recovery rate of credit i
and τi is the default time of credit i. With a lower boundary a and an upper boundary

25
b the tranche notional is

Nf ixed side (t) = max(b − Lf ixed (t), 0) − max(a − Lf ixed (t), 0) (2.49)
Nf loat side (t) = max(b − Lf loat (t), 0) − max(a − Lf loat (t), 0) (2.50)

The present value of the fixed side payments can be written as

m
X
P Vf ixed side = s ∆j−1,j B(0, tj )Nf ixed (tj ) (2.51)
j=1

where tm = T is the maturity of the basket CDS tranche, ∆j−1,j is the year fraction
between two payments, B(0, tj ) is the discount factor and s is the percentile premium
the protection seller receives.
If it is assumed that compensation payments for default are paid in certain inter-
vals only, the present value of the float side payments is
l
X
P Vf loat side = (Nf loat (tj−1 ) − Nf loat (tj ))B(0, tj ) (2.52)
j=1

The present value of the tranche (from the protection seller’s point of view) is the
difference between the two present values computed above:

P VT ranche = P Vf ixed − P Vf loat (2.53)

Performing all described steps N times, adding up all present values and dividing the
sum by N gives the expected value of the tranche.

2.3.2 Semi-Explicit Approach

While in the Monte Carlo method the portfolio loss (and thus also the present value of
the tranche) is computed for each scenario with respect to the simulated default times
τi individually, and the expected value of the basket CDS tranche is the average of the
present values, in the semi-explicit approach the whole loss distribution P(L(t) = k)
is used to determine the tranche spread.
The first step in the semi-explicit approach is to determine the loss distribution
P(L(t) = k). This is accomplished by the characteristic function of L(t):

ΨL(t) (u) = E[eiL(t)u ] (2.54)

" n #
Y  j|V

j|V

= E 1 − pt + pt · eiMj u
j=1
Z +∞ n 
Y  
j|V j|V
= ϕ(V ) 1 − pt + pt · eiMj u dV
−∞ j=1

26
with the loss Mj on credit j in case of default (this is Nj (1 − Rj ) for recovery type or
Nj for binary type) and ϕ(V ) the normal probability density function for the global
latent variable V . The characteristic function (2.54) can be expressed in terms of the
Fourier Transform of φ(k):
Z ∞
ΨL(t) (u) = eiku φ(k)dk (2.55)
−∞

The interesting point is that the probability P (L(t) = k) is equal to the Fourier
Transform P (L(t) = k) = φ(k). The required loss distribution for a certain time t
can be determined by Fourier Transformation of the characteristic function E[eiL(t)u ]:
Z ∞
1
φ(k) = e−iku ΨL(t) (u)du (2.56)
2π −∞

In practice the first step to determine the probabilities P(L(t) = k) is to calculate the
characteristic function with equation (2.54) then to make a Fourier Transformation
of the characteristic function as described in equation (2.55) to get φ(k).
The second step is to determine what impact a certain portfolio loss k has on the
tranche notional. The connection between portfolio loss and tranche notional N (k)
is displayed in Figure 1.3 and can mathematically be stated as

N (k) = (b − a) + (a − k)1{k≥a} + (k − b)1{k≥b} (2.57)

This function maps a certain portfolio loss to the corresponding tranche notional. If
the leg types of fixed and float leg differ, two portfolio losses have to be determined
in the first step and in the second step the two portfolio losses have to be mapped to
two tranche notionals with equation (2.57).
With the loss distribution determined and the mapping function N (k) between
portfolio loss and tranche notional the expected tranche notional is
+∞
X
N = E[N ] = N (k)P(L(t) = k) (2.58)
k=0

If the expected tranche notional has been computed, equation (2.51) and (2.52) pro-
vide the present value of the basket tranche.

27
Chapter 3

Implementation

In this chapter the implementation of the different programs is described in detail.

The chapter starts with the definition of reference results which are used to check
the implemented programs. Then the programs using Monte Carlo simulations are
described. The chapter ends with the description of the programs using the semi-
explicit approach to price multi-name credit derivatives.

3.1 Benchmark Results - Reference Case

In order to verify whether the results in the implemented programs are reasonable,
benchmark results are required.
Schmidt and Ward published an article [21] in which they investigate a k th -to-
default basket with three underlying credits. They assume that the spread of credit
default swaps on the three underlyings are 0.9%, 1.0% and 1.1% for maturities 1
to 5 years. The recovery rate and the linear correlation of the underlyings is 20%
and 50%. This credit basket is called reference case throughout this thesis. The
results of the reference case are shown in Table 3.1. The first step to check the

Maturity (yrs) 1st 2nd 3rd

1 2.63 0.34 0.04
2 2.56 0.42 0.06
3 2.51 0.47 0.08
4 2.47 0.51 0.09
5 2.44 0.55 0.10

Table 3.1: Benchmark basket spreads from Schmidt and Ward [21].

implementation with Schmidt and Ward is to determine default probabilities for the

28
underlyings which give the spreads mentioned above for the CDS. The spread of a
credit default swap is given by equation (1.12). As the CDS used by Schmidt and
Ward have the same spread for all maturities, determining the default probabilities
is done by a procedure similar to bootstrapping. Firstly, the default probability for
one year is determined by changing it as long as the resulting spread does not agree
with the required spread. Then with the one year default probability determined,
the two year default probability is changed until the resulting spread agrees with
the required one. This procedure is done for all maturities and for all three CDS in
an auxiliary Excel spreadsheet. The resulting default probabilities are shown in the
following table:

Credit 1 Credit 2 Credit 3

Spread 0.9% 1.0% 1.1%

Maturity
0 0.00000 0.00000 0.00000
1 0.01097 0.01217 0.01337
2 0.02172 0.02417 0.02653
3 0.03249 0.03600 0.03950
4 0.04304 0.04768 0.05229
5 0.05346 0.06519 0.06487
6 0.06380 0.07056 0.07730

Table 3.2: Default probabilities for the reference case.

3.2 Monte Carlo Method

General Computations Starting point for the Monte Carlo evaluation of multi-
name credit products are the correlated default times τi .
For a Gaussian normal copula correlation structure correlated N (0, 1)-distributed
random numbers are required. These numbers are generated by two algorithms.
N (0, 1) distributed random numbers are created according to Marsaglia [5] and the
uncorrelated random numbers are transformed into correlated ones (with the corre-
lation matrix Σ) by multiplying them with the Cholesky decomposed matrix A (the
relationship between Σ and A is Σ = AAT ).
For a Student t copula correlation structure the correlated N (0, 1)-distributed
p
random numbers are multiplied additionally with ν/s where ν is the degree of
freedom of the t distribution and s is a χ2ν -distributed random number.

29
 The correlated random numbers are transformed into [0, 1]-distributed numbers
(these are the default probabilities):

ˆ ui = Φ(xi ) for Gaussian normal copula

ˆ ui = tν (yi ) for Student t copula (yi =

p
ν/s xi )

For the Clayton copula the generation of random numbers is accomplished with the
algorithm described in Section 2.1.4.
With the [0, 1]-distributed numbers, the default times τi are determined by τi =
Fi−1 (ui )
where Fi (t) is the marginal default probability of underlying i.

3.2.1 Implementation
An overview for which products and copulae a Monte Carlo pricing program has been
implemented is given in Table 3.3:

Copula Gaussian normal Student t Clayton

Product
k th -to-default Program 1 Program 3 Program 5
Basket CDS Tranche Program 2 Program 4 Program 6

Table 3.3: Matrix of products and copulae for which Monte Carlo pricing programs
are implemented.

Progam 1 Monte Carlo simulation of an inhomogeneous k th -to-default basket de-

fault swap. Joint default is modelled as Gaussian normal copula. After the
correlated default times τi have been simulated, the value of the premium and the
default leg are determined with equations (2.29) and (2.30). With the present value
of the default and premium leg for all scenarios the spread of the BDS is determined
as quotient of the sum of the present values of the default legs divided by the sum of
the present values of the premium legs (equation (2.32)).

Program 2 Monte Carlo simulation of an inhomogeneous basket CDS tranche.

Joint default is modelled as Gaussian normal copula. With the correlated default
times τi given, the loss of the float (Nf loat as in (2.48)) and the fixed side (Nf ixed as
in equation (2.47)) of the tranche can be determined at any time. Equation (2.51)
and (2.52) provide the present value of the fixed and float side.

30
Program 3 Monte Carlo simulation of an inhomogeneous k th -to-default basket de-
fault swap. Joint default is modelled as Student t copula. The Student t dis-
tribution is given as integral over the Student t probability function from −∞ to
x:
Zx
tν (x) = fν (y)dy (3.1)
−∞

Calculating the integral using Simpson’s rule would be far too slow therefore inte-
gration is accomplished with the Gaussian quadrature method.1 After the correlated
default times are determined, the spread of the BDS is computed as in Program 1.

Program 4 Monte Carlo simulation of an inhomogeneous basket CDS tranche. In

contrast to Program 2 joint default is modelled as Student t copula; all other steps
are equal to Program 2.

Program 5 Monte Carlo simulation of an inhomogeneous k th -to-default basket de-

fault swap. Joint default is modelled as Clayton copula. With exception of the
copula used this program is identical to Program 1.

Program 6 Monte Carlo simulation of an inhomogeneous basket CDS tranche.

Joint default is modelled as Clayton copula. With exception of the copula used
this program is identical to Program 2.

3.2.2 Comparison with Reference Case

In order to check the implementation of the programs, every program is parameterized
to compute spreads for the reference case (see Section 3.1):
1
Rb
In the Gaussian quadrature framework the integral a f (x)dx is computed as sum of N function
values f (xi ) at fixed abscissae xi ∈ (a, b), where each function value is multiplied with a for the
abscissa xi specific weight w(xi ):

Zb N
X
f (x)dx = w(xi )f (xi ) (3.2)
a i=1

The abscissae xi as well as the weights w(xi ) do not depend on the function f (x) to be integrated.
An algorithm providing the abscissae xi and the weights w(xi ) is given in [19].

31
Program 1 Normal copula, k th -to-default basket: A basket CDS consisting of three
credits with marginal default probabilities that give CDS spreads of 0.9%, 1.0% and
1.1% is examined. Recovery is set to 20% and linear correlation coefficients to 50%.
In each simulation run N = 5·106 scenarios are created to compute the basket spread.
Table 3.4 shows the results of the calculations.

Program 2 Normal copula, basket CDS tranche: In order to map the reference case
as basket CDS tranche, a portfolio consisting of the three credits mentioned above
is set up. The premium leg type is set to binary while the default leg type is set to
recovery. Lower and upper boundary of the tranche are chosen automatically in such
a way that the relevant tranche notional is equivalent to a k th -to-default (e.g. for the
2nd -to-default case with a nominal of 300 and a recovery rate of 20% the lower and
upper boundary for the premium tranche notional are set to 100 and 200 whereas the
boundaries for the default tranche notional are set to 80 and 160).

Schmidt & Ward k th -basket Basket Tranche

Maturity 1st 2nd 3rd 1 st
2nd 3rd 1st 2nd 3rd
1 2.63 0.34 0.04 2.647 0.331 0.039 2.635 0.330 0.039
2 2.56 0.42 0.06 2.558 0.412 0.060 2.546 0.407 0.061
3 2.51 0.47 0.08 2.508 0.468 0.077 2.500 0.460 0.076
4 2.47 0.51 0.09 2.463 0.509 0.092 2.451 0.501 0.090
5 2.44 0.55 0.10 2.437 0.542 0.104 2.424 0.534 0.102

Table 3.4: Comparison of k th -to-default basket (Program 1) and basket CDS tranche
(Program 2) with Schmidt & Ward [21]. Correlation structure is modelled with
Gaussian normal copula. The relative errors of the basket spreads (defined as standard
deviation of the basket spread divided by the basket spread) are found to be of equal
size for Program 1 and Program 2: The relative errors are smaller than 0.7% for
1st -to-default, smaller than 1.8% for 2nd -to-default and smaller than 5.3% for 3rd -to-
default.

Program 3 Student t copula, k th -to-default basket: In the limit of ν → ∞ the

Student t distribution is equal to the Gaussian normal distribution. To Compare
values with the reference case, very large values of ν have to be regarded. The results
displayed in Table 3.5 have been determined with ν = 1000.

Program 4 Student t copula, basket CDS tranche: The reference case is mapped as
basket tranche as already described. In order to compare the results of the Student t
copula with those computed with a normal copula, the parameter ν is set to ν = 1000.

32
 As can be seen in Table 3.4 and 3.5, there are only little differences between the
results of the implemented programs and the reference results provided by Schmidt
and Ward.
Schmidt & Ward k th -basket Basket Tranche
Maturity 1st 2nd 3rd 1st 2nd 3rd 1st 2nd 3rd
1 2.63 0.34 0.04 2.63 0.34 0.04 2.63 0.33 0.04
2 2.56 0.42 0.06 2.56 0.41 0.06 2.58 0.42 0.06
3 2.51 0.47 0.08 2.50 0.47 0.08 2.51 0.47 0.08
4 2.47 0.51 0.09 2.46 0.51 0.09 2.48 0.51 0.09
5 2.44 0.55 0.10 2.43 0.55 0.10 2.46 0.54 0.10

Table 3.5: Comparison of k th -to-default basket (Program 3) and basket CDS tranche
(Program 4) with Schmidt & Ward [21]. Correlation structure is modelled with
Student t copula. The parameter degree of freedom is set to ν = 1000. The relative
errors of the basket spreads (defined as standard deviation divided by the basket
spread) are found to be of equal size for Program 3 and Program 4: The relative
errors are smaller than 0.7% for 1st -to-default, smaller than 1.9% for 2nd -to-default
and smaller than 5.4% for 3rd -to-default.

3.2.3 Convergence of Monte Carlo Method

In order to get an impression how fast the Monte Carlo method converges, a Monte
Carlo simulation2 with 30 Mio. scenarios is performed. The error in scenario i is:
Spread(imax ) − Spread(i)
Errori = ∗ 100 (3.3)
Spread(imax )
Figure 3.1 shows the relative error as function of the scenario. As can be seen in
the figure, the biggest relative error is about 0.35%, and after 4 Mio. simulations the
relative error stays smaller than 0.1%.

3.3 Semi-Explicit Approach

In the semi-explicit approach a factor copula model is used to reduce the complexity
in correlation structure. This approach is implemented for a Gaussian normal copula
only. The programs implemented in this approach are:

ˆ Program 7: This program prices inhomogeneous k th -to-default basket CDS with

a Gaussian normal copula correlation structure as described in Section 2.2.2.
2
The simulation computes the spread of the 1st -to-default for Schmidt & Ward’s reference basket
where the contract has a 5 year maturity.

33
 0.4
0.3

Relative Error (%)

0.2
0.1
0.0
-0.1 0 5 10 15 20 25 30
-0.2
-0.3
-0.4
Scenario (Mio.)

Figure 3.1: This figure shows the relative error in the tranche spread as function of
the number of simulation runs.

ˆ Program 8: This program prices inhomogeneous basket CDS tranches with a

Gaussian normal copula correlation structure via the characteristic function of
the loss (Section 2.3.2).

3.3.1 Semi-Explicit Approach: Basket CDS

The first step is to implement a function which computes the conditional default
i|V
probability as given in equation (2.36). With pt the probability P(N (t) = k), that
the number of defaulted credits N (t) at time t is equal to k, can be determined.
This can be achieved by directly calculating the coefficient of the uk term in equation
(2.40). The expectation has to be taken with respect to V . The integral is computed
with Simpson’s-rule, that is the interval b − a is divided into n parts. The integral is
given by
b n    
xi+1 − xi
Z X xi + xi+1
f (x)dx = f (xi ) + 4 f + f (xi+1 ) (3.4)
a i=1
2 6
b−a
where x1 = 1, xi+1 = xi + n
and xn = b.
With the number of defaulted credits known, the present value of the premium leg
can already be computed with (2.41). In order to compute the value of the default
leg, the probability Zki (t), that credit i defaults at time t after k credits already
have defaulted, is required. Equation (2.45) provides Zki (t) as coefficient of the vk -
term. Expanding the product in (2.45) and determining the coefficient of the v k -term
delivers the probability Zki (t).3 If the probabilities Zki (t) are calculated for the relevant
3
The implementation showed that Zki (t) is extremely sensitive to errors in the inverse normal
distribution. A first approximation of the inverse normal distribution, which had a relative error
less that 10−6 , led to errors in the spread bigger than 10%. The final version used to calculate the
inverse normal distribution accomplishes a rational approximation with a relative error less than
10−8 and uses this result for the Halley method [1] to get an accurate result.

34
k for all credits i and all relevant times t, the present value of the default leg can be
computed with equation (2.46).
This method is used for the calculation of the spread of k th -to-default baskets only.
For basket CDS tranches expanding the probability generating function ΨL(t) = E[uL(t) ],
like it is done for a k th -to-default in (2.45), is not very useful, as the complete loss
distribution is needed and not only one coefficient as for a k th -to-default basket.

3.3.2 Semi-Explicit Approach: Basket CDS Tranche

In contrast to a k th -to-default basket, where simply the probability that credit i
defaults as k th credit is needed to price the default leg, for a basket CDS tranche the
complete loss distribution of the portfolio is required for valuation. Thus the main
focus is to determine the loss distribution of the portfolio, that is the probability that
the portfolio is hit by a certain loss as function of the loss P(L(t) = x). The sought
function P(L(t) = x) can be determined as Fourier Transform4 of the characteristic
function of the portfolio loss ΨL(t) (u) = E[eiL(t)u ]. The integration is done with
Gaussian quadrature method.

3.3.3 Comparison with Reference Case

k th -to-default Basket In the following table the results of the Semi-Explicit ap-
proach are compared with the Schmidt & Ward case:

Schmidt & Ward Semi-Explicit

Maturity (yrs) 1st 2nd 3rd 1st 2nd 3rd
1 2.63 0.34 0.04 2.68 0.33 0.04
2 2.56 0.42 0.06 2.59 0.41 0.06
3 2.51 0.47 0.08 2.54 0.46 0.08
4 2.47 0.51 0.09 2.50 0.51 0.09
5 2.44 0.55 0.10 2.46 0.54 0.10

Table 3.6: Comparison of the semi-explicit approach with Schmidt & Ward.

Comparing the numbers above, one has to keep in mind that Schmidt and Ward
take the accrued premium payments into account whereas in Laurent and Gregory’s
semi-explicit approach these payments are neglected. Because of this, the difference
in the case of a 1st -to-default basket with maturity T = 1 between 2.63 and 2.68 is
not that surprising. A more useful indicator of the correctness of the implementation
4
The Fourier Transformation is done as FFT (Fast Fourier Transformation) with the algorithm
given in [19].

35
 Monte Carlo Semi-Explicit
st
Maturity (yrs) 1 2nd 3rd 1st
2nd 3rd
1 2.6797 0.3312 0.0395 2.6779 0.3317 0.0403
2 2.5862 0.4123 0.0601 2.5928 0.4139 0.0608
3 2.5362 0.4685 0.0765 2.5375 0.4688 0.0772
4 2.4925 0.5101 0.0916 2.4960 0.5103 0.0914
5 2.4660 0.5435 0.1037 2.4631 0.5437 0.1039

Table 3.7: Comparison between the results of the semi-explicit approach and the
Monte Carlo results. In this special implementation accrued premium payments are
not considered. The relative errors of the basket spreads (standard deviation of the
basket spread divided by the basket spread) are smaller than 0.7% for 1st -to-default,
smaller than 1.8% for 2nd -to-default and smaller than 5.3% for 3rd -to-default.

is the difference between the numbers computed in the semi-explicit approach and
those that are computed via a special Monte Carlo implementation, where the accrued
premium payments are also not considered. As can be seen in Table 3.7, there is quite
a good congruence between the results in the Monte Carlo approach and in the semi-
explicit approach.

Basket CDS tranche To check the implementation of the pricing of basket CDS
tranches, Program 8 is parameterized such that the reference case can be described
as a basket CDS tranche. In order to do so, the fixed leg type is set to binary and
the float leg type is set to recovery. The results of this computation can be seen in
the subsequent Table 3.8.

Schmidt & Ward Semi-Explicit

Tranche
Maturity (yrs) 1st 2nd 3rd 1st 2nd 3rd
1 2.63 0.34 0.04 2.639 0.328 0.040
2 2.56 0.42 0.06 2.550 0.407 0.060
3 2.51 0.47 0.08 2.497 0.461 0.076
4 2.47 0.51 0.09 2.455 0.501 0.090
5 2.44 0.55 0.10 2.421 0.534 0.102

Table 3.8: Comparison between the results of the semi-explicit approach and Schmidt
and Ward’s results if the basket CDS is modelled as homogeneous basket CDS tranche.

36
3.3.4 Bimodal and binomial distribution
For certain credit portfolios the loss distribution can be computed analytically in a
simple manner. This is the case if:

ˆ all credits have the same nominal N ;

ˆ all credits have the same default probability p;

ˆ all credits have the same recovery rate of R = 0;

ˆ the correlation between all credits is the same;

ˆ the correlation between the credits is ρ = 0 or ρ = 1.

In the ρ = 1 case the loss distribution is bimodal; that is, there are only two values
populated: Either all credits have defaulted, so that the portfolio loss is 100%, or no
credit has defaulted, so that the portfolio loss is 0%. The probability of all credits
having defaulted at time t is p and the probability of no credit having defaulted is
1 − p. This situation is shown in Figure 3.2 for p = 10%. In the ρ = 0 case the loss
100%
Bimodal
90%
Binomial
80%
70%
Probability of Loss (%)

60%
50%
40%
30%
20%
10%
0%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Loss (%)

Figure 3.2: This figure shows the relative error in the tranche spread as function of
the number of simulation runs.

distribution is binomial:
 
n k
P(Loss(t) = N k) = N p (1 − p)n−k (3.5)
k

Figure 3.2 also shows the binomial distribution for a default probability of p = 10%
and n = 20. This special setup of a credit portfolio can be used to check whether the

37
implementation of Program 8 delivers the expected loss distributions. Hence, Program
8 is parameterized as described above, and the loss distribution is determined and
compared with the theoretical expectation (as shown in Figure 3.2). There are no
differences to the theoretical values.

3.3.5 Runtime Comparison

The big advantage of the semi-explicit approach is its ability to provide results much
faster than Monte Carlo computations. In order to give an impression how big the
differences in the required runtimes implied by the chosen method are, the basket
spreads are calculated with maturities T = 1, ..., 5 for the 1st -to-default basket refer-
ence case and the observed runtimes are shown in Table 3.9.
The basket spreads are calculated via:

ˆ Program 1 - Direct Monte Carlo: 5 · 106 scenarios;

ˆ Program 2 - Monte Carlo Tranche: 5 · 106 scenarios, 1st -to-default is modelled

as basket CDS tranche;

ˆ Program 8 - Semi-Explicit Approach: 1st -to-default is modelled as basket CDS

tranche, integration in equation (2.54) is done with Simpson’s rule;

ˆ Program 8 - Semi-Explicit Approach: 1st -to-default is modelled as basket CDS

tranche, integration in equation (2.54) is done with a 20-point Gaussian Legen-
dre quadrature method5 ;

Table 3.9 shows the observed runtimes6 .

Monte Carlo Monte Carlo Semi-Explicit Semi-Explicit

Direct Tranche Simpson’s rule Gaussian Quadrature
Runtime(s) 210 321 5 1

Table 3.9: Observed runtime to compute five 1st -to-default basket spreads for matu-
rities T = 1, 2, 3, 4, 5.

It is obvious that the semi-explicit approach is able to provide results much faster
than Monte Carlo methods.

5
An implementation of a 10-point Gaussian integration algorithm can be found in [19] (see Chap-
ter 4.5).
6
The calculations are performed on a 600 MHz Intel Pentium III PC with 64 MByte memory.

38
Chapter 4

Results

In this chapter the results computed with the implemented programs are presented.
Section 4.1 gives the results for the reference credit basket, and in Section 4.2 the
results for a real world basket CDS tranche are shown.

4.1 Results for the Reference Case

In this section Schmidt and Ward’s reference basket [21] is subject of examination.

4.1.1 Gaussian Normal Copula - Basket Spread as Function

of the Correlations
For Gaussian normal copulae linear correlation is an important parameter. In this
section the influence of ρ on the basket spread is investigated. Changing correlation
3.0
1st
2.5 2nd
3rd
Basket Spread (%)

2.0

1.5

1.0

0.5

0.0
0.0 0.2 0.4 0.6 0.8 1.0
ρ

Figure 4.1: Basket spread as function of the linear correlation coefficient ρ. The
number of Monte Carlo simulation runs is N = 5 · 106 .

39
between the credits has a big influence on the basket spread of a k th -to-default basket.
Figure 4.1 displays the spread1 of a 1st -, a 2nd - and a 3rd -to-default basket as function
of the correlation for the reference case [21]. With increasing correlation the spread
of the 1st -to-default decreases whereas the spread of the 2nd - and 3rd -to-default in-
creases with increasing correlation. The explanation for this dependence is that with
increasing correlation the probability for a second or third default increases whereas
the probability for a first default decreases. Therefore the spread of the 1st -to-default
decreases while 2nd - and 3rd -to-default protection gets more expensive.

4.1.2 Student t Copula - Influence of ν on the Basket Spread

If the correlation structure is modelled as Student t copula, the additional parameter ν
influences the spread of a k th -to-default basket. The dependence of the 1st -to-default2
for the reference case is displayed in the upper left graph of Figure 4.2. With increasing
ν the spread increases also and converges to the normal copula spread. As can be seen
in the figure the difference between the spread computed with the Student t copula
and the spread computed with the Gaussian normal copula of 2.63% is already for
ν = 64 very small (about 1% difference). For small values of ν the probability of
correlated defaults increases leading to a decrease in the spread. The right upper
graph of Figure 4.2 displays the influence of ν on the spread of the 2nd -to-default
for the reference case and the lower left graph the dependence of the 3rd -to-default
basket spread on the parameter ν.
What can be seen clearly in Figure 4.2 is that the parameter ν of the Student t
copula has a dramatic influence on the basket spread.
The results are summarized in the lower right graph of figure 4.2, which shows
the relative differences of the Student t basket spread to the Gaussian normal copula
basket spread as function of ν. The maximal change is −23% for the 1st -to-default,
115% for the 2nd -to-default and 448% for the 3rd -to-default. This shows clearly what
impact the copula used for pricing has on the determined prices.

4.1.3 Clayton Copula - Influence of α on the Basket Spread

The last copula investigated for the reference case is an example of the Archimedean
copula family - the Clayton copula. The results are computed with a Monte Carlo
simulation as described for the Clayton copula in Section 2.1.4. Figure 4.3 shows the
1
These results have been computed with Program 1.
2
The results are computed with Program 4 for T = 1. The number of Monte Carlo simulation
runs is set to N = 5 · 106 .

40
 2.7 0.8

2.6 0.7

Basket Spread (%)

0.6
Basket Spread (%)

2.5
0.5
2.4
0.4
2.3
0.3
2.2 0.2
2.1 0.1
2.0 0.0
1 10 100 1000 1 10 100 1000

ν ν

0.3 500
1st

Change in Basket Spread (%)

0.2 400 2nd
Basket Spread (%)

3rd
0.2 300

0.1 200

100
0.1

0
0.0
1 10 100 1000
1 10 100 1000 -100

ν ν

Figure 4.2: Basket Spread as function of the parameter ν; the left upper graph shows
the 1st -to-default basket spread - the relative errors (standard deviation divided by
basket spread) are smaller than 1.0%; the 2nd -to-default basket spread is displayed in
the right upper graph - the relative errors are smaller than 2.9%; the right lower graph
shows the 3rd -to-default - the relative errors are smaller than 5.5%; the right lower
graph relates the Student t basket spreads to the Gaussian normal basket spreads.

basket spread for maturity T = 1 for the 1st -, 2nd - and 3rd -to-default as function of
the parameter α. The parameter α is related to the “correlation” within the Clayton
α
copula correlation structure. Kendall’s tau as measure of dependence is τ = α+2 .
Thus increasing α is analogous to increasing correlation. For the 1st -to-default basket
increasing α results in a decrease of the basket spread whereas for the 2nd - and 3rd -to-
default a decrease leads to an increase of the basket spread. The influence α has on
the basket spread thus can be compared to the influence the parameter ρ has on the
basket spread for the Gaussian normal copula. The form of the curves in Figure 4.1
and 4.3 is different but the qualitative influence of increasing correlation is in both
figures the same: An increase in correlation leads to a decrease in the 1st -to-default
basket spread and increases in the 2nd and 3rd -to-default basket spreads.

41
 3.0
1st
2nd
2.5
3rd

2.0
Basket Spread (%)

1.5

1.0

0.5

0.0
0 1 2 3 4 5
α

Figure 4.3: Basket spread of the reference case as function of α if joint default is
modelled as Clayton copula. Maturity is set to T = 1 and the number of Monte
Carlo simulation runs is N = 5 · 106 . Standard deviation: For the 1st -to-default the
relative errors (standard deviation divided by the basket spread) are for all values of α
smaller than 2.1%, for the 2nd -to-default smaller than 1.2% and for the 3rd -to-default
smaller than 5.2%.

For a specific value of α the Clayton copula produces nearly equal results as the
Gaussian normal copula. This value is α = 0.27 for the reference case basket. The
basket spreads are 2.626 for the 1st -to-default, 0.338 for the 2nd -to-default and 0.038
for the 3rd -to-default.
Again, the results for this copula show that the choice of the copula has an enor-
mous influence on the basket spread.

4.2 Valuation of a real world Basket CDS Tranche

Given the contractual characteristics of a real basket tranche, the implemented rou-
tines can be used to find a price for this product. The basket tranche to be priced
consists of 31 underlyings with individual marginal default probabilities and recovery
rates. The notional per underlying is 10, 000, 000 Euro and the lower limit of the
tranche is 20, 000, 000 Euro whereas the upper limit is 40, 000, 000 Euro. The fixed
and float leg type of the tranche are both recovery style. As the basket tranche to be
priced is a real world product, the data has been anonymized; that is, the underly-
ing names are called Credit 1 to Credit 31. The data can be found in Appendix C.
Some key information about the tranche is listed in the subsequent Table 4.1. Using
Program 2 for basket CDS tranche pricing with the data above, a present value of
1, 520, 203 is computed. The difference to the price provided by the protection buyer

42
 Property Value
Start 02-May-2003
Maturity 02-May-2008
Coupon Frequency 3 monthly
Lower Boundary 20,000,000
Upper Boundary 40,000,000
Currency Euro
Premium-Leg-Type Recovery
Default-Leg-Type Recovery
Spread 120 bps
Evaluation time (=Today) 02-May-2003
Present Value 1,529,295

Table 4.1: Basic Data of the real world basket CDS tranche.

of 9, 092 corresponds to a percentile deviation less than 1%. In the next sections this
real world basket is examined3 regarding the influence recovery rates, correlations,
default probabilities and the lower and upper boundary have on the tranche spread4 .

4.2.1 Tranche Spread as Function of the Recovery Rates

In this section the influence of the recovery rate on the tranche spread is investigated.
Starting with the initial recovery rates given in Appendix C the recovery rates are

4.5

4.0
Tranche Spread (%)

3.5

3.0

2.5

2.0

1.5
-20 -10 0 10 20
Change in Recovery Rate (%)

Figure 4.4: Tranche spread of the real world basket CDS tranche as function of the
shift in the recovery rates. The shift is given as percentage.
3
Computations are done with Program 2.
4
With the given parameters the fair tranche spread is 3.1027% for valuation date 02-May-2003.

43
changed5 and the tranche spread with the changed recovery rates is computed. Fig-
ure 4.4 shows that with increasing recovery rates the tranche spread decreases. This
is exactly what one would expect as with increasing recovery rates the severity of a
default decreases and thus also the present value of the default leg decreases.

4.2.2 Tranche Spread as Function of the Correlations

A similar analysis as with recovery rates can be done to determine the influence the
correlation has on the tranche spread. Commencing with the correlation coefficients
given in Appendix C every coefficient is shifted by the same amount. Figure 4.5
shows how the spread changes if the correlations between the underlying credits are
changed. For the regarded tranche an increase in correlation leads to an increase in
the tranche spread.
4.0

3.8
Tranche Spread (%)

3.6

3.4

3.2

3.0
0 5 10 15 20 25
Change in Correlation (%)

Figure 4.5: Tranche spread of the real world basket CDS tranche as function of the
shift in the correlations. The shift is given as percentage.

4.2.3 Tranche Spread as Function of the Default Probabili-

ties
The influence of the default probabilities can be investigated by changing the default
probabilities for all underlyings by the same amount, calculating the new spread and
comparing this changed spread with the original. Figure 4.6 shows how the spread
5
The recovery rates are changed on a global scale, that is, the recovery rate of all credits is shifted
by the same amount. A value of 5 in figure 4.4 means that the recovery rate of all credits of the
basket is increased by 5%. For example, credit 1 having an original recovery rate of 45% has a
changed recovery rate of 50%.

44
changes if the default probabilities of all underlyings for all maturities6 are increased
by the same amount.
4.0

3.8

Tranche Spread (%)

3.6

3.4

3.2

3.0
0.0 0.5 1.0 1.5
Change in Default Prob (%)

Figure 4.6: Tranche spread of the real world basket CDS tranche as function of the
shift in the default probabilities. The shift is given as percentage.

As expected, an increase in the default probabilities causes an increase in the

tranche spread.

4.2.4 Tranche Spread as Function of the regarded Tranche

In this section it is investigated how the tranche spread depends on the lower and
upper boundary of the basket CDS tranche. Figure 4.7 shows the spread as function

Boundary
lower upper Tranche Spread
0 20 10.1251
20 40 3.1027
40 60 1.4895
60 80 0.8012
80 100 0.4340
100 120 0.2132

Table 4.2: This table shows the tranche spread for different tranches.

of the lower boundary. With increasing lower boundary the tranche spread decreases.
This result is quite intuitive as with an increasing lower boundary the probability
that the tranche is affected by defaults decreases.
6
This corresponds to a parallel shift of the complete default curve.

45
 12.0

10.0

Tranche Spread (%)

8.0

6.0

4.0

2.0

0.0
0 10 20 30 40 50 60 70 80 90 100 110
Lower Boundary (Mio €)

Figure 4.7: Tranche spread of the real world basket CDS tranche as function of the
shift in the lower boundary of the tranche. Upper boundary is always lower boundary
plus 20 Mio. Euro.

4.2.5 Student t Copula - Tranche Spread as Function of ν

In this section it is investigated what influence the chosen correlation structure has on
the tranche spread7 . In the previous sections correlated defaults are modelled via the
Gaussian normal copula whereas in this section the tranche spread is determined for a
Student t copula correlation structure. Figure 4.8 shows the tranche spread computed
1.6

1.5
0-20
1.4 20-40
Student t / Normal

1.3 30-50
35-55
1.2
40-60
1.1 60-80
1.0 80-100

0.9

0.8
1 10 100 1000
ν

Figure 4.8: Student t copula tranche spread relative to Gaussian normal copula
tranche spread as function of ν.

if correlated default is modelled via a Student t copula divided by the tranche spread
computed if correlated default is modelled via Gaussian normal copula for different
7
Computations are done with Program 4.

46
tranches. As can be seen in Figure 4.8, there are two different regimes with respect
to the influence increasing ν has on the tranche spread. If the lower boundary is
smaller than 35 · 106 , then increasing ν leads to an increase in the tranche spread.
This behaviour is similar to a 1st -to-default. If the lower boundary is bigger than
35 · 106 , then increasing ν results in decreasing tranche spreads. This behaviour can
be compared to 2nd -to-default-basket.

4.2.6 Clayton Copula - Tranche Spread as Function of α

In this section the influence of the parameter α of the Clayton copula on the tranche
spread is examined. As already mentioned the parameter α describes dependence
within the Clayton copula. Using Program 8 provides the results presented in Figure
4.9. Depending on the location of the tranche within the portfolio the influence of
increasing α on the tranche spread is different. For the tranche from 0 to 20 · 106 Euro
the tranche spread decreases with increasing α, whereas for the tranches with a lower
boundary bigger than 20·106 Euro the tranche spread increases with increasing α. As

20 3.5 0-20
18 20-40
3.0
Spread(alpha)/Spread(alpha=4.8)

16 40-60
2.5
Tranche Spread (%)

14 60-80
12 2.0 80-100
10
1.5
8
6 1.0
4
0.5
2
0 0.0
0 1 2 3 4 5 0 1 2 3 4 5
α α

Figure 4.9: Left side: Clayton copula tranche spread as function of the parameter α.
Right side: Clayton copula tranche spread at given α divided by the tranche spread
at α = 4.8.

it is difficult to see the curve shapes in the left side of the figure for higher tranches,
the right side of the figure compares the curve shapes by dividing the tranche spread
for a certain α with the tranche spread for α = 4.8.

47
Chapter 5

Conclusion

In this thesis basket default swaps and basket CDS tranches (CDOs) are priced with
two different methods. The first method is standard Monte Carlo, the second is a
factor copula based approach providing semi-explicit formulae for pricing. The semi-
explicit approach is implemented for a Gaussian normal copula correlation structure
whereas the Monte Carlo method is implemented for a Gaussian normal-, a Student
t- and a Clayton copula.
If the correlation structure of the regarded BDS is modelled with a Student t
copula, higher probability mass compared to a Gaussian normal copula is found at
the margins of the distribution leading to lower 1st -to-default and higher 2nd - and
3rd -to-default basket spreads. This effect is found to be especially large for the 3rd -
to-default basket, where the basket spread for ν = 2 is nearly 5 times the size of the
basket spread if the correlation structure is modelled via a Gaussian normal copula. A
qualitatively similar effect can be found in basket CDS tranches. Tranches covering
the first losses also show lower tranche spreads if joint default is modelled with a
Student t copula. The explanation for this behaviour is that for a Student t copula
the probability of joint defaults is higher than for a Gaussian normal copula.
In order to use the semi-explicit approach, the correlation matrix has to be mapped
to the correlation vector β of a factor copula model. If it is not possible to map the
correlation matrix to the correlation vector for the semi-explicit approach accurately,
then there is a nearly linear dependence between the error in the correlation vec-
tor determined with the least square method and the error in the calculated basket
spreads.
In certain cases an accurate mapping is possible; then both methods provide equal
results for basket and tranche spreads. The advantage of the semi-explicit method is
that the complete loss distribution can be determined via a FFT of the characteristic

48
function of the loss. Compared to Monte Carlo methods this method is much faster
and thus even for very large portfolios computations are possible in a reasonable time.
In this thesis a real world basket CDS tranche is priced. The determined price
agrees with the price provided by the protection buyer. For the real world basket
CDS tranche it is examined what influence changes in the correlation matrix, in the
recovery rates, in the default probabilities and in the position of the tranche within
the credit portfolio have on the tranche spread. The results agree with the expectation
that

ˆ increasing recovery rates lead to a decrease in the tranche spread

ˆ increasing default probabilities result in an increase in the tranche spread

ˆ a shift of the lower boundary of the tranche to higher values leads to lower
tranche spreads

The influence of increasing correlation on the tranche depends on the position of the
tranche within the credit portfolio. For the tranche given in the regarded real world
basket CDS tranche increasing correlation leads to an increase in the basket spread.
If a Clayton copula is used to model the correlation structure, an increase in the
parameter α describes an increase in the correlation. For a basket default swap 1st -
to-default protection gets cheaper with increasing α whereas 2nd - and 3rd -to-default
protection gets more expensive with increasing α. This observation is in good con-
cordance with the influence the linear correlation coefficient has on the basket spread
for a Gaussian normal copula. For this copula increasing correlation leads also to
higher costs for 1st -to-default protection and lower costs for 2nd - and 3rd -to-default
protection.
If the results gained with different copulae are compared, it is getting obvious that
the choice of the used copula has an enormous influence on the determined basket or
tranche spreads.

49
Appendix A

Derivation of the default

probability Zki (t)

In order to price an inhomogeneous BDS, the probability Zki (t) defined as

1
Zki (t) = lim P(Ni (t0 ) − Ni (t) = 1, N (−i) (t) = k) (A.1)
t →t t0 − t
0

is required. Ni (t0 ) − Ni (t) gives the number of defaulted credits between t and t0 and
N (−i) (t) = j6=i Ni (t) the number of defaulted credits before or at t where credit i is
P

not taken into account. The probability P(Ni (t0 ) − Ni (t) = 1, N (−i) (t) = k) can be
computed via the joint probability generating function of (Ni (t0 ) − Ni (t), N (−i) (t)):
h 0 (−i)
i
ψ(u, v) = E uNi (t )−Ni (t) v N (t) (A.2)

This equation can be written as

n
X
ψ(u, v) = P(Ni (t0 ) − Ni (t) = 0, N (−i) (t) = k)v k
k=1
n−1
X
+ P(Ni (t0 ) − Ni (t) = 1, N (−i) (t) = k)u v k (A.3)
k=1

ψ(u, v) can also be stated as

h h 0 (−i)
ii
ψ(u, v) = E E uNi (t )−Ni (t) v N (t) |V (A.4)

By conditional independence this is

h h 0
i h (−i) ii
ψ(u, v) = E E uNi (t )−Ni (t) |V × E v N (t) |V (A.5)

With
i|V i|V
P (Ni (t0 ) − Ni (t) = 1) = pt0 − pt (A.6)
 
i|V i|V
P (Ni (t0 ) − Ni (t) = 0) = 1 − pt0 − pt (A.7)

50
 i|V
this can be written in terms of the conditional default probabilities pt
" #
    Y 
i|V i|V i|V i|V j|V j|V
ψ(u, v) = E 1 − pt0 + pt + pt0 − pt u × 1 − pt + pt v (A.8)
j6=i

Comparing equation (A.3) with (A.8), the subsequent identity is found

n−1
" #
X i|V i|V
Y j|V j|V
P(Ni (t0 ) − Ni (t) = 1, N (t) = k)v k = E (pt0 − pt ) (1 − pt + pt v) (A.9)
k=1 j6=i

i|V
For smooth conditional default probabilities pt the limit t0 → t is defined for both
sides:
n−1
" #
i|V Y 
X dp t j|V j|V

Zki (t)v k = E 1 − pt + pt v (A.10)
k=1
dt j6=i

In order to determine Zki (t) the product on the right side of (A.10) has to be expanded.
Then the expectation of the v k term with respect to V has to be computed.

51
Appendix B

Limitation of Factor Models

The idea behind factor models is to reduce the complexity of the correlation structure.
In certain cases this can be done without loss of accuracy, but in general it is not
possible to map any correlation matrix to a correlation vector for a factor model. In
this section such cases are studied, and it is investigated what impact the error in the
correlation vector for the semi-explicit approach has on the calculated spread.

General Remarks The investigation is performed for a portfolio consisting of three

credits with properties already described in the reference case (Schmidt and Ward).
Starting from a correlation structure which can be reproduced perfectly in the factor
model, the correlation matrix is changed so that the correlation structure cannot be
reproduced by the correlation β vector any longer. The initial values are chosen as
ρ12 = 0.4, ρ13 = 0.5 and ρ23 = 0.8 and the best correlation vector is determined with
the least square method. That is, 2 = (ρ12 − β1 β2 )2 + (ρ13 − β1 β3 )2 + (ρ23 − β2 β3 )2 is
minimized under the additional restriction β3 ≤ 0.9999.1 The results of this procedure
can be found in Table B.1.
In order to determine the impact of an error in the β-factors on the calculated
spread in the semi-explicit approach, the difference between the spread calculated
with the correct correlation matrix with a Monte Carlo simulation and the spread
calculated with erroneous β-factors is computed. This difference is called basket
spread error and is defined as

∆Spread = SpreadM onte Carlo − SpreadSemi−Explicit Approach (B.1)

This procedure is performed for the 1st -, 2nd - and 3rd -to-default basket for maturity
T = 1, ..., 5. The data obtained in this way is displayed in Table B.2. Scenario 1
1
This condition is required as otherwise there are problems in computing the conditional default
i|V
probability pt .

52
 Monte Carlo Semi-Explicit Approach Error
Scenario ρ12 ρ13 ρ23 β1 β2 β3 2
1 0.4 0.5 0.8 0.500016043 0.800058688 0.9999 3.39e-9
2 0.3 0.5 0.8 0.457648610 0.775042609 0.9999 0.005416045
3 0.2 0.5 0.8 0.415130051 0.753282217 0.9999 0.022103115
4 0.1 0.5 0.8 0.372251102 0.735388483 0.9999 0.050702225
5 0.0 0.5 0.8 0.328658468 0.722068881 0.9999 0.091771662

Table B.1: This table contains the correlations used for Monte Carlo simulations and
the least square minimized β-factors for the semi-explicit approach.

shows a basket spread error not equal to zero. This basket spread error is regarded as
constant offset and is subtracted from all basket spread errors; the corrected basket
spread error values are displayed in Table B.3. It is clear that the basket spread error
decreases with increasing maturity. This effect can be seen also in Figure B.1, which
shows the quotient of the basket spread error at maturity T divided by the basket
spread error at maturity T = 1 for the 1st -to-default. A similar effect is found also
1.02
Spread Error (T) / Spread Error (T=1)

1.00
0.98
0.96 Scenario 2
0.94 Scenario 3
0.92 Scenario 4
0.90 Scenario 5
0.88
0.86
0.84
0.82
0 1 2 3 4 5 6

Maturity (yrs)

Figure B.1: This figure shows the quotient of the basket spread error at maturity T
divided by the basket spread error at maturity T = 1.

for the 2nd - and 3rd -to-default.

It is of particular interest how the size of the β-error affects the basket spread
error. The basket spread errors in Table B.3 are averaged and shown in Figure B.2 as
function of the β-error. This figure implies that there is a linear dependence between
the β-error and the basket spread error.
The absolute error in the basket spread is not that meaningful as one has no
impression whether the very small basket spread errors of 3rd -to-default baskets re-
lated to the basket spread of the 3rd -to-default baskets are relatively big, small or

53
 0.20
0.15
1st
0.10
2nd
Spread Error (%) 0.05
3rd
0.00
-0.05 0.0 0.1 0.2 0.3 0.4

-0.10
-0.15
-0.20
∆β

Figure B.2: Average of the basket spread errors for maturities T = 1, ..., 5 as function
of the error in β

of comparable size. Therefore the basket spread error is translated into a property
which can be understood better. The starting point is the difference between the
correct basket spread sM C (Monte Carlo) and the erroneous basket spread sSE (Semi-
explicit approach). It is possible to find a special flat correlation ρ̄M C such that a
Monte Carlo simulation with the flat correlation ρ̄M C results in the correct basket
spread sM C . The same holds for the erroneous basket spread ρ̄SE : A Monte Carlo
simulation with the flat correlation ρ̄SE results in the erroneous basket spread sSE :
s(M C : ρ = ρ̄SE ) = sSE . Thus the basket spread error can be transformed into a flat
correlation error: ∆ρ̄ = ρ̄M C − ρ̄SE .
Table B.4 presents the flat correlation errors for the scenarios 1 to 5 and for the
maturities T = 1, ..., 5. Again, scenario 1 shows values not equal to zero which are
regarded as constant offset and which are subtracted from the scenarios. Table B.5
contains the corrected data. With increasing maturity the amount of the flat correla-
tion errors decreases again. Figure B.3 displays the flat correlation errors as function
of the error in β. 1st - and 2nd -to-default baskets show a linear dependence whereas
the 3rd -to-default basket seems to decrease stronger than linearly as function of the
β-error.

54
 20

15
Flat Correlation (%)
10
1st
5 2nd
3rd
0
0.0 0.1 0.2 0.3 0.4
-5

-10

∆β

Figure B.3: Average of the errors in the flat correlation for the maturities T = 1, ..., 5
as function of the error in β.

Scenario 1 2 3 4 5
 0 0.0736 0.1487 0.2252 0.3029
Maturity 1st -to-default
1 -0.0020 -0.0437 -0.0856 -0.1190 -0.1545
2 -0.0224 -0.0606 -0.0989 -0.1340 -0.1698
3 0.0067 -0.0304 -0.0686 -0.1066 -0.1405
4 0.0088 -0.0269 -0.0636 -0.1006 -0.1359
5 0.0170 -0.0179 -0.0546 -0.0909 -0.1273
2nd -to-default
1 0.0038 0.0438 0.0847 0.1217 0.1558
2 -0.0066 0.0319 0.0721 0.1119 0.1488
3 -0.0014 0.0346 0.0742 0.1136 0.1517
4 0.0079 0.0419 0.0788 0.1184 0.1575
5 0.0071 0.0405 0.0773 0.1157 0.1556
3rd -to-default
1 0.0000 -0.0010 -0.0044 -0.0077 -0.0111
2 0.0004 -0.0012 -0.0051 -0.0104 -0.0152
3 -0.0007 -0.0015 -0.0052 -0.0109 -0.0178
4 0.0016 0.0001 -0.0041 -0.0107 -0.0183
5 0.0004 -0.0013 -0.0056 -0.0121 -0.0203

Table B.2: This table shows the error in the basket spread for maturity T = 1, ..., 5
and order 1,2 and 3.

55
 Scenario 1 2 3 4 5
 0 0.0736 0.1487 0.2252 0.3029
st
Maturity 1 -to-default
1 0.0000 -0.0417 -0.0836 -0.1170 -0.1524
2 0.0000 -0.0382 -0.0765 -0.1116 -0.1474
3 0.0000 -0.0371 -0.0753 -0.1133 -0.1472
4 0.0000 -0.0358 -0.0724 -0.1094 -0.1447
5 0.0000 -0.0349 -0.0716 -0.1079 -0.1443
Average 0.0000 -0.0375 -0.0759 -0.1118 -0.1472
2nd -to-default
1 0.0000 0.0400 0.0809 0.1179 0.1521
2 0.0000 0.0385 0.0787 0.1185 0.1554
3 0.0000 0.0360 0.0756 0.1150 0.1531
4 0.0000 0.0340 0.0709 0.1105 0.1496
5 0.0000 0.0334 0.0702 0.1087 0.1485
Average 0.0000 0.0364 0.0753 0.1141 0.1517
3rd -to-default
1 0.0000 -0.0010 -0.0044 -0.0077 -0.0111
2 0.0000 -0.0016 -0.0055 -0.0108 -0.0156
3 0.0000 -0.0008 -0.0045 -0.0102 -0.0170
4 0.0000 -0.0015 -0.0058 -0.0123 -0.0199
5 0.0000 -0.0017 -0.0060 -0.0125 -0.0207
Average 0.0000 -0.0013 -0.0052 -0.0107 -0.0169

Table B.3: This table shows the error in the corrected basket spread: The basket
spread of the first scenario (see Table B.2) is subtracted from all basket spreads.

56
 Scenario 1 2 3 4 5
 0 0.0736 0.1487 0.2252 0.3029
Maturity 1st -to-default
1 0.0012 0.0250 0.0519 0.0740 0.0990
2 0.0126 0.0355 0.0598 0.0840 0.1082
3 -0.0038 0.0174 0.0412 0.0656 0.0881
4 -0.0050 0.0154 0.0381 0.0615 0.0844
5 -0.0096 0.0104 0.0328 0.0556 0.0789
2nd -to-default
1 0.0031 0.0369 0.0721 0.1056 0.1361
2 -0.0059 0.0294 0.0668 0.1047 0.1398
3 -0.0013 0.0345 0.0742 0.1141 0.1529
4 0.0083 0.0445 0.0840 0.1265 0.1687
5 0.0079 0.0456 0.0873 0.1309 0.1762
3rd -to-default
1 0.0000 -0.0032 -0.0188 -0.0437 -0.0885
2 0.0008 -0.0032 -0.0166 -0.0431 -0.0793
3 -0.0014 -0.0036 -0.0148 -0.0380 -0.0756
4 0.0030 0.0003 -0.0106 -0.0332 -0.0664
5 0.0008 -0.0026 -0.0134 -0.0346 -0.0669

Table B.4: This table shows the error in the flat correlations corresponding to the
basket spread for maturity T = 1, ..., 5 and order 1,2 and 3.

57
 Scenario 1 2 3 4 5
 0 0.0736 0.1487 0.2252 0.3029
st
Maturity 1 -to-default
1 0.0000 0.0238 0.0507 0.0728 0.0978
2 0.0000 0.0228 0.0472 0.0714 0.0955
3 0.0000 0.0212 0.0450 0.0693 0.0919
4 0.0000 0.0204 0.0431 0.0665 0.0893
5 0.0000 0.0200 0.0424 0.0651 0.0885
Average 0.0000 0.0216 0.0457 0.0690 0.0926
2nd -to-default
1 0.0000 0.0339 0.0691 0.1025 0.1331
2 0.0000 0.0353 0.0727 0.1107 0.1457
3 0.0000 0.0358 0.0755 0.1155 0.1542
4 0.0000 0.0362 0.0757 0.1182 0.1604
5 0.0000 0.0377 0.0794 0.1230 0.1683
Average 0.0000 0.0358 0.0745 0.1140 0.1523
3rd -to-default
1 0.0000 -0.0032 -0.0188 -0.0437 -0.0884
2 0.0000 -0.0041 -0.0174 -0.0440 -0.0802
3 0.0000 -0.0022 -0.0134 -0.0366 -0.0742
4 0.0000 -0.0027 -0.0136 -0.0362 -0.0694
5 0.0000 -0.0034 -0.0142 -0.0354 -0.0676
Average 0.0000 -0.0031 -0.0155 -0.0392 -0.0760

Table B.5: This table shows the error in the corrected flat correlations: The flat
correlation of the first scenario (see Table B.4) is subtracted from all flat correlations.

58
Appendix C

Evaluation Data for the real world

Basket CDS Tranche

The subsequent tables show the recovery rates, the survival probabilities of the un-
derlying credits and the correlations between the credits as of the evaluation date
02-May-2003.

Table Content
C.1 Recovery Rates and Survival Probabilities from 3m to 18m
C.2 Survival Probabilities from 21m to 39m
C.3 Survival Probabilities from 42m to 60m
C.4 Correlations
C.5 Correlations
C.6 Correlations

59
 Survival Probability
Underlying Recovery 3m 6m 9m 12m 15m 18m
Credit 1 0.45 0.9996 0.9992 0.9986 0.9980 0.9973 0.9966
Credit 2 0.45 0.9984 0.9968 0.9952 0.9938 0.9923 0.9909
Credit 3 0.45 0.9984 0.9968 0.9952 0.9938 0.9923 0.9909
Credit 4 0.45 0.9950 0.9897 0.9840 0.9779 0.9717 0.9654
Credit 5 0.23 0.9989 0.9978 0.9963 0.9946 0.9926 0.9905
Credit 6 0.23 0.9994 0.9987 0.9978 0.9968 0.9956 0.9943
Credit 7 0.23 0.9994 0.9988 0.9979 0.9968 0.9954 0.9938
Credit 8 0.23 0.9973 0.9944 0.9915 0.9886 0.9857 0.9827
Credit 9 0.23 0.9992 0.9983 0.9972 0.9959 0.9944 0.9927
Credit 10 0.45 0.9979 0.9956 0.9932 0.9908 0.9882 0.9855
Credit 11 0.23 0.9981 0.9960 0.9936 0.9910 0.9881 0.9851
Credit 12 0.45 0.9993 0.9986 0.9975 0.9962 0.9947 0.9931
Credit 13 0.45 0.9996 0.9993 0.9988 0.9982 0.9976 0.9969
Credit 14 0.45 0.9944 0.9884 0.9829 0.9775 0.9723 0.9673
Credit 15 0.45 0.9956 0.9909 0.9860 0.9807 0.9752 0.9695
Credit 16 0.23 0.9989 0.9977 0.9965 0.9951 0.9936 0.9920
Credit 17 0.45 0.9988 0.9975 0.9957 0.9935 0.9911 0.9885
Credit 18 0.45 0.9984 0.9968 0.9952 0.9938 0.9924 0.9910
Credit 19 0.45 0.9996 0.9993 0.9985 0.9975 0.9963 0.9948
Credit 20 0.45 0.9961 0.9919 0.9875 0.9829 0.9781 0.9733
Credit 21 0.45 0.9828 0.9648 0.9474 0.9303 0.9135 0.8970
Credit 22 0.45 0.9993 0.9985 0.9977 0.9969 0.9961 0.9952
Credit 23 0.45 0.9966 0.9931 0.9893 0.9854 0.9813 0.9772
Credit 24 0.23 0.9992 0.9985 0.9976 0.9966 0.9956 0.9945
Credit 25 0.45 0.9862 0.9717 0.9576 0.9437 0.9301 0.9166
Credit 26 0.45 0.9996 0.9992 0.9987 0.9982 0.9976 0.9970
Credit 27 0.45 0.9957 0.9912 0.9869 0.9827 0.9786 0.9746
Credit 28 0.45 0.9968 0.9933 0.9896 0.9856 0.9813 0.9769
Credit 29 0.45 0.9970 0.9938 0.9906 0.9873 0.9841 0.9809
Credit 30 0.23 0.9993 0.9986 0.9978 0.9969 0.9959 0.9949
Credit 31 0.45 0.9826 0.9646 0.9472 0.9303 0.9138 0.8978

Table C.1: Recovery Rate and survival probability for 3m to 18m.

60
 Survival Probability
Underlying 21m 24m 27m 30m 33m 36m 39m
Credit 1 0.9958 0.9949 0.9939 0.9929 0.9918 0.9907 0.9897
Credit 2 0.9894 0.9881 0.9865 0.9848 0.9831 0.9815 0.9795
Credit 3 0.9894 0.9881 0.9865 0.9848 0.9831 0.9815 0.9795
Credit 4 0.9585 0.9519 0.9452 0.9381 0.9312 0.9243 0.9174
Credit 5 0.9880 0.9855 0.9828 0.9798 0.9767 0.9735 0.9701
Credit 6 0.9928 0.9912 0.9896 0.9878 0.9859 0.9840 0.9819
Credit 7 0.9920 0.9900 0.9880 0.9857 0.9833 0.9808 0.9783
Credit 8 0.9796 0.9767 0.9734 0.9700 0.9667 0.9634 0.9596
Credit 9 0.9908 0.9888 0.9868 0.9845 0.9821 0.9797 0.9770
Credit 10 0.9826 0.9797 0.9767 0.9735 0.9702 0.9670 0.9634
Credit 11 0.9818 0.9784 0.9749 0.9711 0.9672 0.9632 0.9591
Credit 12 0.9912 0.9892 0.9872 0.9849 0.9827 0.9803 0.9777
Credit 13 0.9962 0.9954 0.9946 0.9937 0.9927 0.9918 0.9907
Credit 14 0.9624 0.9579 0.9535 0.9492 0.9453 0.9416 0.9378
Credit 15 0.9632 0.9571 0.9507 0.9439 0.9371 0.9304 0.9231
Credit 16 0.9902 0.9884 0.9865 0.9844 0.9823 0.9801 0.9777
Credit 17 0.9854 0.9822 0.9791 0.9756 0.9721 0.9684 0.9647
Credit 18 0.9896 0.9884 0.9870 0.9857 0.9844 0.9831 0.9817
Credit 19 0.9931 0.9912 0.9894 0.9874 0.9852 0.9830 0.9807
Credit 20 0.9680 0.9630 0.9576 0.9518 0.9460 0.9402 0.9340
Credit 21 0.8803 0.8649 0.8493 0.8335 0.8185 0.8042 0.7892
Credit 22 0.9943 0.9934 0.9925 0.9915 0.9906 0.9896 0.9887
Credit 23 0.9727 0.9683 0.9637 0.9588 0.9539 0.9490 0.9436
Credit 24 0.9933 0.9920 0.9907 0.9893 0.9878 0.9863 0.9847
Credit 25 0.9029 0.8902 0.8773 0.8642 0.8517 0.8397 0.8272
Credit 26 0.9964 0.9957 0.9950 0.9941 0.9932 0.9923 0.9913
Credit 27 0.9706 0.9670 0.9633 0.9596 0.9562 0.9529 0.9496
Credit 28 0.9720 0.9672 0.9623 0.9570 0.9517 0.9464 0.9406
Credit 29 0.9775 0.9743 0.9709 0.9674 0.9640 0.9607 0.9572
Credit 30 0.9938 0.9926 0.9913 0.9900 0.9885 0.9871 0.9855
Credit 31 0.8817 0.8670 0.8521 0.8372 0.8230 0.8097 0.7961

Table C.2: Survival probability for 21m to 39m.

61
 Survival Probability
Underlying 42m 45m 48m 51m 54m 57m 60m
Credit 1 0.9887 0.9877 0.9867 0.9855 0.9842 0.9829 0.9815
Credit 2 0.9774 0.9752 0.9731 0.9706 0.9680 0.9653 0.9626
Credit 3 0.9774 0.9752 0.9731 0.9706 0.9680 0.9653 0.9626
Credit 4 0.9107 0.9040 0.8973 0.8906 0.8839 0.8772 0.8705
Credit 5 0.9667 0.9631 0.9594 0.9557 0.9519 0.9480 0.9440
Credit 6 0.9798 0.9776 0.9753 0.9731 0.9707 0.9683 0.9658
Credit 7 0.9757 0.9729 0.9701 0.9672 0.9643 0.9612 0.9579
Credit 8 0.9559 0.9520 0.9481 0.9442 0.9402 0.9361 0.9319
Credit 9 0.9744 0.9715 0.9686 0.9658 0.9629 0.9599 0.9568
Credit 10 0.9598 0.9561 0.9523 0.9486 0.9448 0.9409 0.9370
Credit 11 0.9550 0.9507 0.9464 0.9421 0.9377 0.9332 0.9286
Credit 12 0.9750 0.9721 0.9692 0.9665 0.9638 0.9609 0.9580
Credit 13 0.9897 0.9886 0.9875 0.9865 0.9855 0.9844 0.9834
Credit 14 0.9343 0.9309 0.9278 0.9251 0.9224 0.9200 0.9179
Credit 15 0.9158 0.9083 0.9008 0.8932 0.8854 0.8775 0.8694
Credit 16 0.9753 0.9728 0.9702 0.9676 0.9647 0.9618 0.9588
Credit 17 0.9609 0.9569 0.9528 0.9489 0.9449 0.9408 0.9366
Credit 18 0.9803 0.9789 0.9775 0.9759 0.9743 0.9726 0.9710
Credit 19 0.9784 0.9760 0.9735 0.9710 0.9685 0.9659 0.9631
Credit 20 0.9278 0.9214 0.9150 0.9082 0.9012 0.8940 0.8867
Credit 21 0.7750 0.7608 0.7472 0.7337 0.7204 0.7072 0.6945
Credit 22 0.9878 0.9870 0.9861 0.9851 0.9840 0.9828 0.9817
Credit 23 0.9381 0.9324 0.9266 0.9208 0.9149 0.9088 0.9026
Credit 24 0.9830 0.9813 0.9796 0.9777 0.9758 0.9738 0.9717
Credit 25 0.8152 0.8032 0.7917 0.7803 0.7688 0.7576 0.7466
Credit 26 0.9903 0.9893 0.9882 0.9870 0.9858 0.9846 0.9833
Credit 27 0.9465 0.9434 0.9405 0.9377 0.9350 0.9323 0.9299
Credit 28 0.9350 0.9291 0.9231 0.9173 0.9114 0.9054 0.8993
Credit 29 0.9538 0.9504 0.9470 0.9436 0.9402 0.9367 0.9332
Credit 30 0.9839 0.9823 0.9806 0.9789 0.9772 0.9755 0.9737
Credit 31 0.7832 0.7706 0.7586 0.7466 0.7348 0.7233 0.7122

Table C.3: Survival probability for 42m to 60m.

62
 Credit
1 2 3 4 5 6 7 8 9 10
Credit 1 1 0.651 0.607 0.225 0.751 0.710 0.595 0.647 0.641 0.432
Credit 2 0 1.000 0.848 0.309 0.796 0.808 0.701 0.686 0.529 0.575
Credit 3 0 0 1.000 0.319 0.702 0.680 0.615 0.611 0.401 0.498
Credit 4 0 0 0 1.000 0.301 0.274 0.301 0.269 0.233 0.332
Credit 5 0 0 0 0 1.000 0.909 0.808 0.878 0.848 0.661
Credit 6 0 0 0 0 0 1.000 0.730 0.714 0.693 0.582
Credit 7 0 0 0 0 0 0 1.000 0.690 0.699 0.596
Credit 8 0 0 0 0 0 0 0 1.000 0.667 0.644
Credit 9 0 0 0 0 0 0 0 0 1.000 0.498
Credit 10 0 0 0 0 0 0 0 0 0 1.000
Credit 11 0 0 0 0 0 0 0 0 0 0
Credit 12 0 0 0 0 0 0 0 0 0 0
Credit 13 0 0 0 0 0 0 0 0 0 0
Credit 14 0 0 0 0 0 0 0 0 0 0
Credit 15 0 0 0 0 0 0 0 0 0 0
Credit 16 0 0 0 0 0 0 0 0 0 0
Credit 17 0 0 0 0 0 0 0 0 0 0
Credit 18 0 0 0 0 0 0 0 0 0 0
Credit 19 0 0 0 0 0 0 0 0 0 0
Credit 20 0 0 0 0 0 0 0 0 0 0
Credit 21 0 0 0 0 0 0 0 0 0 0
Credit 22 0 0 0 0 0 0 0 0 0 0
Credit 23 0 0 0 0 0 0 0 0 0 0
Credit 24 0 0 0 0 0 0 0 0 0 0
Credit 25 0 0 0 0 0 0 0 0 0 0
Credit 26 0 0 0 0 0 0 0 0 0 0
Credit 27 0 0 0 0 0 0 0 0 0 0
Credit 28 0 0 0 0 0 0 0 0 0 0
Credit 29 0 0 0 0 0 0 0 0 0 0
Credit 30 0 0 0 0 0 0 0 0 0 0
Credit 31 0 0 0 0 0 0 0 0 0 0

Table C.4: Table of correlations.

63
 Credit
11 12 13 14 15 16 17 18 19 20
1 0.582 0.650 0.560 0.397 0.660 0.584 0.576 0.586 0.654 0.444
2 0.692 0.542 0.543 0.544 0.747 0.627 0.672 0.739 0.769 0.506
3 0.640 0.461 0.566 0.604 0.612 0.525 0.599 0.780 0.634 0.389
4 0.221 0.364 0.405 0.380 0.281 0.350 0.241 0.258 0.278 0.235
5 0.776 0.706 0.521 0.463 0.854 0.673 0.778 0.815 0.822 0.569
6 0.747 0.643 0.468 0.480 0.835 0.686 0.791 0.695 0.911 0.610
7 0.683 0.611 0.472 0.490 0.700 0.505 0.676 0.666 0.672 0.441
8 0.623 0.593 0.460 0.430 0.714 0.597 0.618 0.681 0.641 0.411
9 0.648 0.700 0.390 0.253 0.702 0.511 0.640 0.518 0.621 0.484
10 0.484 0.443 0.424 0.426 0.523 0.565 0.508 0.570 0.523 0.369
11 1.000 0.591 0.350 0.381 0.730 0.533 0.726 0.650 0.664 0.516
12 0 1.000 0.422 0.324 0.680 0.452 0.614 0.473 0.572 0.456
13 0 0 1.000 0.632 0.398 0.529 0.439 0.493 0.436 0.348
14 0 0 0 1.000 0.380 0.484 0.458 0.445 0.500 0.188
15 0 0 0 0 1.000 0.620 0.788 0.685 0.750 0.640
16 0 0 0 0 0 1.000 0.614 0.514 0.659 0.524
17 0 0 0 0 0 0 1.000 0.617 0.740 0.584
18 0 0 0 0 0 0 0 1.000 0.633 0.432
19 0 0 0 0 0 0 0 0 1.000 0.631
20 0 0 0 0 0 0 0 0 0 1.000
21 0 0 0 0 0 0 0 0 0 0
22 0 0 0 0 0 0 0 0 0 0
23 0 0 0 0 0 0 0 0 0 0
24 0 0 0 0 0 0 0 0 0 0
25 0 0 0 0 0 0 0 0 0 0
26 0 0 0 0 0 0 0 0 0 0
27 0 0 0 0 0 0 0 0 0 0
28 0 0 0 0 0 0 0 0 0 0
29 0 0 0 0 0 0 0 0 0 0
30 0 0 0 0 0 0 0 0 0 0
31 0 0 0 0 0 0 0 0 0 0

Table C.5: Table of correlations.

64
 Credit
21 22 23 24 25 26 27 28 29 30 31
1 0.411 0.580 0.571 0.403 0.660 0.506 0.129 0.602 0.539 0.509 0.614
2 0.333 0.492 0.750 0.503 0.832 0.488 0.175 0.714 0.546 0.697 0.739
3 0.159 0.439 0.652 0.461 0.723 0.410 0.228 0.608 0.442 0.569 0.659
4 0.153 0.239 0.265 0.386 0.353 0.123 0.285 0.189 0.254 0.218 0.316
5 0.375 0.615 0.642 0.461 0.859 0.576 0.113 0.786 0.562 0.616 0.776
6 0.376 0.582 0.657 0.465 0.864 0.613 0.117 0.792 0.640 0.678 0.760
7 0.302 0.461 0.561 0.433 0.805 0.457 0.151 0.648 0.455 0.465 0.714
8 0.357 0.488 0.552 0.367 0.724 0.421 0.146 0.679 0.408 0.494 0.684
9 0.372 0.585 0.452 0.385 0.662 0.501 0.084 0.582 0.472 0.428 0.585
10 0.347 0.389 0.542 0.375 0.621 0.328 0.106 0.461 0.335 0.405 0.626
11 0.228 0.504 0.505 0.367 0.714 0.552 0.045 0.700 0.494 0.554 0.685
12 0.438 0.493 0.458 0.412 0.625 0.461 0.088 0.564 0.472 0.383 0.562
13 0.324 0.421 0.554 0.509 0.556 0.321 0.259 0.355 0.416 0.444 0.492
14 0.261 0.263 0.588 0.678 0.528 0.202 0.381 0.312 0.415 0.449 0.547
15 0.336 0.592 0.596 0.470 0.802 0.572 0.066 0.937 0.560 0.652 0.721
16 0.220 0.599 0.570 0.430 0.706 0.482 0.201 0.592 0.498 0.614 0.568
17 0.203 0.458 0.599 0.493 0.765 0.593 0.104 0.783 0.603 0.649 0.711
18 0.146 0.447 0.551 0.357 0.704 0.425 0.027 0.658 0.372 0.508 0.646
19 0.341 0.546 0.622 0.445 0.797 0.593 0.086 0.710 0.676 0.705 0.708
20 0.227 0.626 0.380 0.277 0.581 0.529 -0.019 0.619 0.733 0.653 0.510
21 1.000 0.269 0.450 0.326 0.307 0.103 0.113 0.249 0.430 0.243 0.326
22 0 1.000 0.387 0.276 0.540 0.513 0.076 0.528 0.516 0.514 0.428
23 0 0 1.000 0.617 0.677 0.438 0.277 0.540 0.523 0.580 0.709
24 0 0 0 1.000 0.484 0.233 0.333 0.375 0.368 0.407 0.510
25 0 0 0 0 1.000 0.566 0.153 0.763 0.568 0.674 0.795
26 0 0 0 0 0 1.000 0.006 0.551 0.534 0.528 0.461
27 0 0 0 0 0 0 1.000 0.047 0.120 0.146 0.177
28 0 0 0 0 0 0 0 1.000 0.504 0.639 0.665
29 0 0 0 0 0 0 0 0 1.000 0.724 0.565
30 0 0 0 0 0 0 0 0 0 1.000 0.605
31 0 0 0 0 0 0 0 0 0 0 1.000

Table C.6: Table of correlations.

65
References

[1] Acklam P.J.: An Algorithm for computing the inverse normal cumulative distri-
bution. Available at http://home.online.no/∼ pjaklam/notes/invnorm

[2] Boscher H. and Ward I.: Long or short in CDOs. Risk, June 2002

[3] Black F. and Cox J: Valuing Corporate Securities: Some Effect of Bond Indenture
Provisions. Journal of Finance, 31, p. 351-367

[4] Black F. and Scholes: The Pricing of Options and Corporate Liabilities. Journal
of Political Economy, 81, p. 637-654

[5] Brent R.P.: Fast Normal Random Numbers Generators for Vector Proces-
sors. March 1993, available at http://cs.anu.edu.au/techreports/1993/TR-CS-
93-04.pdf

[6] Duffie D. and Singleton K. J.: Modeling Term Structures of Defaultable Bonds.
Review of Financial Studies, 12, p. 687-720

[7] Embrechts P., Lindskog F. and McNeil A.: Modelling Dependence with Copulas
and Applications to Risk Management. Available at http://www.defaultrisk.com

[8] Frey R., McNeil A. and Nyfeler M.: Copulas and credit models. Risk, October
2001

[9] Hull J. and White A.: Valuing Credit Default Swaps I: No Counterparty Default
Risk. Working paper, April 2000

[10] Hull J. and White A.: Valuing Credit Default Swaps II: Modeling Default Cor-
relations. Working paper, April 2000

[11] KMV Corporation: Modeling Default Risk. Technical Document, 1997, available
at http://www.kmv.com

66
[12] Laurent J.-P. and Gregory J.: Basket Default Swaps, CDOs and
Factor Copulas. October 2002, available at http://www.maths.univ-
evry.fr/mathfi/JPLaurent.pdf

[13] Li D.: On default correlation: a copula function approach. Journal of Fixed

Income, March 2000

[14] Mashal R. and Naldi M.: Pricing Multiname Credit Derivatives: Heavy Tailed
Hybrid Approach. Available at http://www.defaultrisk.com

[15] Mashal R. and Naldi M.: Extreme events and default baskets. Risk, June 2002

[16] Merton R.: On the pricing of corporate debt: the risk structure of interest rates.
Journal of Finance, 29, p. 449-470

[17] J.P. Morgan & Incorporated: CreditMetrics - Technical Document. Technical

Document, 1999, available at http://www.riskmetrics.com

[18] Nelson R.: An Introduction to Copulas. Springer, New York, 1999

[19] Press W.H., Teukolsky S.A. and Vettling W.T.: Numerical Recipes in C. Cam-
bridge University Press, 1993

[20] RiskMetrics Group: CreditGrades Technical Document. Available at

http://www.creditgrades.com/resources/resources

[21] Schmidt W. and Ward I.: Pricing default baskets. Risk, June 2002

67