Munich Personal RePEc Archive

Copula based simulation procedures for

pricing basket Credit Derivatives

Fathi, Abid and Nader, Naifar

Faculty of Business and Economics, University of Sfax, UR:

MODESFI„ Institute of the Higher Business Studies, University of
Sfax, UR: MODESFI,

March 2007

Online at https://mpra.ub.uni-muenchen.de/6014/
MPRA Paper No. 6014, posted 29 Nov 2007 13:41 UTC
 Copula ba se d sim ula t ion proc e dure s for pric ing

ba sk e t Cre dit De riva t ive s

Fa t hi Abid*
Faculty of Business and Economics, University of Sfax, UR: MODESFI,
Sfax, Tunisia
N a de r N a ifa r ¤
Institute of the Higher Business Studies, University of Sfax, UR: MODESFI,
Sfax, Tunisia

First ve rsion: M a rc h 2 0 0 7

Abst ra c t

This paper deals with the impact of structure of dependency and the choice of procedures for rare-
event simulation on the pricing of multi-name credit derivatives such as nth to default swap and
Collateralized Debt Obligations (CDO). The correlation between names defaulting has an effect on the
value of the basket credit derivatives. We present a copula based simulation procedure for pricing
basket default swaps and CDO under different structure of dependency and assessing the influence of
different price drivers (correlation, hazard rates and recovery rates) on modelling portfolio losses.
Gaussian copulas and Monte Carlo simulation is widely used to measure the default risk in
basket credit derivatives. Default risk is often considered as a rare-event and then, many studies
have shown that many distributions have fatter tails than those captured by the normal distribution.
Subsequently, the choice of copula and the choice of procedures for rare-event simulation govern the
pricing of basket credit derivatives. An alternative to the Gaussian copula is Clayton copula and t-
student copula under importance sampling procedures for simulation which captures the dependence
structure between the underlying variables at extreme values and certain values of the input random
variables in a simulation have more impact on the parameter being estimated than others .

K e yw ords: Collateralized Debt Obligations, Basket Default Swaps, Monte Carlo method,
One factor Gaussian copula, Clayton copula, t-student copula, importance sampling.

* Professor of finance. Fathi Abid can be contacted at : fathi.abid@fsegs.rnu.tn (Corresponding author).

¤
Associate assistant Professor of finance. Nader Naifar can be contacted at: doctoratnader@yahoo.fr

1
1 . I nt roduc t ion

A credit derivative is an over-the-counter derivative designed to transfer credit risk from one party to
another. By synthetically creating or eliminating credit exposures, they allow institutions to more
effectively manage credit risks. Four of the most common credit derivatives are credit default swap,
credit linked notes, total return swap and credit spread options. The dominant product in the credit
derivatives market is the credit default swap. However, the last ten years or so has seen the growth of
‘portfolio credit derivatives’ such as basket default swap and Collateralised Debt Obligation (CDO)1.
These financial instruments have been used successfully by large financial institutions to diversify and
reduce credit risk. Many empirical works has been done on single name credit derivative products.
Hull, Predescu & White (2004) analyze the impact of credit rating announcements in the pricing of
credit default swap. Norden & Weber (2004) analyze the empirical relationship between credit default
swap, bond and stock markets. Ericsson, Jacobs & Oviedo (2004) investigate the relationship between
theoretical determinants of default risk (firm leverage, volatility and the riskless interest rate) and
actual market spread of credit default swap using linear regression. Abid & Naifar (2006 (a)) explain
empirically the determinants of credit default swap rates using a linear regression. They find that credit
rating, maturity, riskless interest rate, slope of the yield curve and volatility of equities explain more
than 60% of the total level of credit default swap.

Actually, more substantial empirical studies are devoted on structured credit derivatives instruments,
in particular basket default swap and CDO. The main problem in the pricing of such instruments is
modelling the structure of dependency of the default times. Defaults are rarely observed. Copulas can
be introduced to model these correlations by using the correlations of corresponding default time. We
know that Kendall’s tau remains invariant under monotone transformations. This is the foundation of
modelling the correlation of credit events by using the correlation of underlying default time via
copulas. Li (2000) present a Gaussian copula method for the pricing of first to default swap. Other
studies of elliptical copulas with higher tail dependence, such as the t-copula, can be found in Mashal
and Naldi (2002). The Marshall-Olkin copula is yet another class of copula functions, which stems
from the multivariate compound Poisson process. In this model, individual defaults are constructed
from a series of independent common shock. Previous work on the use of the Marshall-Olkin copula
in the context of credit risk modelling includes Duffie and Pan (2001), Wong (2000), Lindskog and
McNeil (2003).Hull & White (2004) develop two procedures to pricing tranches of CDO and nth to
default swap. The first procedure involves calculating the probability distribution of the number of
defaults by a time T and suited to the situation where companies have equal weight in the portfolio and
recovery rates are assumed to be constant. The second involves calculating the probability distribution
of the total loss from defaults by time T. Jobst (2002) propose a pricing model that draws expected
loan loss of CDO based on parametric bootstrapping through extreme value theory under the impact of
asymmetric information. Tavares et al. (2004) present a basket model to deal with the Gaussian copula
smile. They combine the copula model (to model the default risk that is driven by the economy) with
independent Poisson processes (to model the default risk that is driven by a particular sector and by
the company in question).Hull and White (2005) introduce the technique of perfect copulas. Their
copula model can be regarded as ‘perfect’ in that it hits the tranche quotes exactly. The hazard-rate-
path probability distribution is the only input about the underlying copula in order to value a CDO.
Burtschell et al. (2005) employ the technique of the double Student-t copula model for the calibration
of CDO. They find that this copula model fit better the features to the CDO market in comparison to
other models like Gaussian, t-Student, stochastic correlation, Clayton and Marshall-Olkin copulas.
Madan (2004) provide details for the pricing of nth to default contracts using the one factor Gaussian
and the Clayton copulas. He model the marginal default time densities using Weibull and Frechet

1
According to SIFMATM, Global CDO issuance through the third quarter of 2006, at $322 billion, has exceeded
full year 2005 issuance by 20%. Issuance in the third quarter of 2006, at $117.8 billion, also exceeded issuance
in the third quarter of last year by 30%. European Securitisation Forum (ESF) Forecasts Issuance to Grow to a
New Record of €531 Billion in 2007, Led by residential mortgage-backed securities (RMBS), commercial
mortgage-backed securities (CMBS) and CDO. Then, a 16.4 percent growth rate from the €456 billion issued in
2006.

2
families and the joint densities are obtained using the method of copula. Verschuere (2006) present a
factor approach combined with copula functions to price tranches of synthetic Collateralized Debt
Obligation (CDO) having totally inhomogeneous collateral (the obligors in the CDO pool have
different spreads and different notional). Sircar and Zariphopoulou (2006) study the impact of risk
aversion on the valuation of basket credit derivatives. They use the technology of utility-indifference
pricing in intensity based models of default risk.

In our paper, we analyse the impact of structure of dependency and the nature of simulation
procedures on the pricing of multi-name credit derivatives such as nth to default swap and
Collateralized Debt Obligations (CDO). To express dependencies between times of default, Gaussian,
student and Clayton copulas have been considered. The copula function links the univariate margins
with their full multivariate distribution. It presents a useful tool when modelling non Gaussian data
since the Pearson’s correlation coefficient is adapted for linear dependence and normal distribution.
One appealing feature of a copula function is that the margins do not depend on the choice of the
dependency structure and then, we can model and estimate the structure of dependency and the
margins separately.

The remainder of this paper is organised as follows: section two describes some mathematical
background about the concept of copula and its properties. In section three, we present some tools for
modelling joint default times. Section four and five present a copula based simulation procedures for
pricing basket default swaps and CDO, assessing the influence of different price drivers (correlation,
hazard rates and recovery rates) on modelling portfolio losses. Section six summarizes the findings
and concludes.

2 . Stylised facts about Copula functions

The most important problem in the pricing of basket credit derivatives and CDO tranches is the
modelling of the joint default times. In this section, we will introduce the concept of a copula function.
Copula was first used in survival analysis and actuarial sciences. Copula functions are getting more
and more popular credit correlation modelling due to its simplicity and fast computation. Embrechts,
et al (1999) clarified many issues concerning dependence and its relationship to correlation, especially
in financial data such as market crashes, credit crises. According to Gennheimer (2002) there are
several reasons why copulas are such an attractive tool for modelling dependence:
1. They provide us with a powerful tool for building a large number of multivariate models and are
extremely useful in the Monte Carlo simulation of dependent risk factors.
2. They allow us to overcome the fallacies and dangers of approaches to dependence that focus only
on correlation.
3. They provide a way of studying scale-free measures of dependence.
4. They express dependence on a quantile scale, which we will find is useful for describing the
dependence of extreme outcomes.

Frey & McNeil (2002) analyse default correlation and the pricing and hedging of credit sensitive
instruments with copulas functions. Jouanin et al (2001) address the problem of incorporating default
dependency in intensity-based credit risk models by using copulas functions to model the joint
distribution of the default times. Abid & Naifar (2005, 2006(b)) study the impact of Stock returns
volatility of reference entities on credit default swap rates using Archimedean copula.
For n uniform random variables u1 , u 2, ..., u n , the joint distribution function C is defined as:

C(u1 , u 2 ,..., u n , θ ) = Pr [U1 ≤ u1, U 2 ≤ u 2 ,..., U n ≤ u n ] (2.1)

With θ is the dependence parameter.
We present the following definition for the bivariate case: A copula function is the restriction to
[0,1]2 of a continuous bivariate distribution function whose margins are uniform on [0,1]. . A (bivariate)
copula is a function C : [0,1]2 → [0,1] which satisfies the boundary conditions:

3
 C (t , 0 ) = C (0 , t ) = 0 and C (t ,1) = C (1, t ) = t for t ∈ [0 ,1]. (2.2)
Similarly, copula satisfies the 2-increasing property:
C (u 2 , v 2 ) − C (u 2 , v1 ) − C (u1 , v 2 ) + C (u1 , v1 ) ≥ 0 (2.3)
For all u 1 , u 2 , v 1 , v 2 in [0,1] and u 1 ≤ u 2 and v 1 ≤ v 2 .
A copula is symmetric if:
C(u , v ) = C(v, u ) for all (u , v ) in [0,1] and is asymmetric otherwise.
2
(2.4)

Sklar (1959) shows the importance of copulas as a universal tool for studying multivariate
distributions2. By definition, applying the cumulative distribution function (CDF) to a random variable
(r.v.) results in a r.v. that is uniform on the interval [0, 1]. Let X a random variable with continuous
distribution function FX, FX(X) is uniformly distributed on the interval [0,1]. This result is known as
the probability integral transformation theorem and present many statistical procedures. With this
result in hand, we may introduce the copula using basic statistical theory. In particular, the copula C
for (X,Y) is just the joint distribution function for the random couple FX(X), FY(Y) provided FX and FY
are continuous.
The previous representation is called canonical representation of the distribution. Thus, copulas link
joint distribution functions to their margins. Then, in continuous distribution, the problem of obtaining
the joint distribution has reduced to selecting the appropriate copula. We can build multidimensional
distributions with different marginals.
Copula functions allow us to separate the structure of dependency between default times into two
parts: the first part is the specification of the marginal distribution function (the distribution of default
time of each obligor. The second part is the choice of the appropriate copula which describes the
structure of dependency between default times.
Numerous copulas can be found in the literature (see Nelson (1999) and Joe (1997)). The most
commonly applied copula function (especially in finance modelling) is the Gaussian copula3. This
could be justified by the fact that the multivariate normal distribution has two appealing
characteristics: first, their marginal distributions are normal and second, it can be fully described by
their marginal distribution and a variance-covariance matrix. For univariate margins F1,…,Fn which
are Gaussians, the dependence structure among the margins is described by a unique normal copula
function. Let X 1 ,..., X n be random variables which are standard normal distributed with
means µ 1 ,..., µ n , standard deviations σ1 ,..., σ n and correlation matrix Σ . Then, the distribution
⎛ X − µi ⎞
function C Σ (u 1 ,..., u n ) of the random variables U i = Φ⎜⎜ i ⎟⎟, i ∈ {1,..., n} is a Gaussian copula
⎝ σi ⎠
with correlation matrix Σ . Φ (.) denotes the cumulative univariate standard normal distribution
function.

The Gaussian copula can be written as:

Φ −1 ( u1 )Φ −1 ( u n )
1 ⎛ 1 r r T r r ⎞
C Σ (u 1 ,..., u n ) = ∫ ... ∫ exp⎜ − (ν − µ ) Σ −1 (ν − µ )⎟dν 1 ...dν n
⎝ 2 ⎠
n
(2π) 2 det Σ −∞ −∞

(2.5)
With Φ −1 is the inverse of the standard univariate Gaussian distribution function.
By differentiating the precedent equation with respect to u 1 ,..., u n , we obtain the density of the
Gaussian copula:

2
The original definition of copula is given by Sklar (1959) and the Sklar’s theorem is considered as the most
important theorem about copula functions. The problem of obtaining a joint distribution is reduced to selecting
the appropriate copula.
3
Credit Metrics TM and KMV model implicitly incorporate copula functions based on the multivariate Gaussian
distribution of asset value process.

4
 ⎛ ⎛ ν1 ⎞ ⎞
⎜ ⎜. ⎟ ⎟
C Σ (u 1 ,..., u n ) =
1 1 −1
⎜. ⎟
( )
exp⎜ − (ν 1 ,..., ν n ) ∑ − 1 ⎜ . ⎟ ⎟⎟ ; With ν i = Φ −1 (u i ), i ∈ {1,..., n}.
⎜ (2.6)
det Σ 2
⎜ ⎜ν ⎟⎟
⎝ ⎝ n ⎠⎠
The following algorithm generate random variates (u 1 ,..., u n ) which are determination of correlated
uniform variates on [0,1] from the Gaussian copula with the correlation matrix Σ :
- Find the Cholesky decomposition4 A of the correlation matrix Σ , such that Σ = A ⋅ A T ;
- Simulate n independent standard normal random variates Z = (z 1 , z 2 ,..., z n )T ;
- Set x = A ⋅ Z ;
- Set x back to an n-dimensional vector u of uniform variates on [0,1] by computing u = Φ (x ) .
The vector u is a random variate from the n-dimensional Gaussian copula C Σ .

Figure 1:1000 simulated standard uniform random variables under Gaussian copula

If we use a Gaussian copula, we preserve the underlying distribution of the individual random
variables but the joint distribution is like a multidimension Gaussian. This naturally assigns very little
weight to the tails. In reality, we find that within the financial markets, tail events occur much more
frequently. So we would like a joint distribution which has fatter tails but preserves the same (bell
shaped, non-skewed) characteristics of the Gaussian, hence we use the t-Student copula.
⎛ ν + 1⎞ ν +1
Γ⎜ ⎟ −
⎝ 2 ⎠ x ⎛ 2 ⎞ 2
t ν (x ) = ∫ ⎜1 + s ⎟ ds . (2.7)
−∞ ⎛ ν ⎞
⎜ ν ⎟⎠
Γ⎜ ⎟ ν Π ⎝
⎝2⎠
The t-student copula with the correlation matrix Σ and ν degrees of freedom is presented as follow:

ν+n
−
t ν−1 ( u1 ) t ν−1 ( u n ) ⎛ ⎛ → →⎞⎞ 2
⎜ ⎟⎟
T
1 ⎛→ →⎞ −1 ⎜
C ν ,Σ (u 1 ,..., u n ) = ∫ ... ∫ k ν , Σ ⎜1 + ⎜ v − u ⎟ ∑ ⎜ v − u ⎟ ⎟ dv1 ...dv 2 (2.8)
−∞ −∞ ⎜ ν⎝ ⎠ ⎜ ⎟⎟
⎝ ⎝ ⎠⎠
With:

4
A symmetric and positive definite matrix can be efficiently decomposed into a lower and upper triangular
matrix. For a given matrix, this is achieved by the LU decomposition which factorizes A=LU. If A satisfies the
above criteria, one can decompose more efficiently into A = LLT , where L (which can be seen as the ``matrix
square root'' of A) is a lower triangular matrix with positive diagonal elements. L is called the Cholesky triangle.

5
 ⎛ν+n⎞
Γ⎜ ⎟
k ν ,Σ = ⎝ 2 ⎠
. (2.9)
⎛ν⎞
Γ⎜ ⎟ (Πν ) det (Σ )
n

⎝2⎠

To simulate random variates from the t-Student copula C ν,Σ with the correlation matrix Σ and ν
degrees of freedom, we can use the following algorithm:
- Find the Cholesky decomposition A of the correlation matrix Σ , such that Σ = A ⋅ A T ;
- Simulate n independent standard normal random variates Z = (z 1 , z 2 ,..., z n )T ;
- simulate a random variate, s, from χ ν2 distribution, independent of Z;
- Set y = A ⋅ Z ;
ν
- Set x = y ;
s
- Set x back to an n-dimensional vector u of uniform variates on [0,1] by computing u = t ν (x ) .
The vector u is a random variate from the n-dimensional t-Student copula C ν,Σ .

Figure 2: 1000 samples from t-Student copula with degree of freedom equal to 7.

In terms of the appropriate choice for the number of degrees of freedom, it is often necessary to carry
out some statistical tests with historical data to ascertain how fat we require the tails to be, Galiani
(2003) use an Exact Maximum Likelihood Method (EML). Other woks explain how to calibrate t-
student copula to real market data (Mashal and Zeevi (2003), Romano (2002), Meneguzzo and
Vecchiato (2002)…). The difference between Gaussian copulas and the t-Student copulas can be
described with the concept of tail dependence. If a bivariate copula C(u , v ) such as:
1 + C(u , u ) − 2u
lim = λU > 0 , (2.10)
u →1 1− u
Then C has upper tail dependence with parameter λ U . If:
C(u , u )
lim = λL > 0 (2.11)
u →0 u
Then C has lower tail dependence with parameter λ L .
Numerous copulas can be found in the literature (see Nelson (1999) and Joe (1997)). The
Archimedean copula has many families that are capable to present different structure of dependency
and different methods are developed to estimates its parameters. We only need to find functions which
will serve as generators and define the corresponding copula. Clayton copula is an example of
Archimedean copula. This family proposed by Clayton (1978) is the following:

6
Let Φ (t ) =
(t −θ
−1 )
with θ ∈ [− 1, ∞ ) /{0} , then:
θ
⎡ −θ ⎤
( )
1
C clayton
θ (u , v ) = max ⎢ u + v −θ
− 1
−
θ ,0
⎥ (2.12)
⎣ ⎦
If θ > 0 , then φ(0 ) = ∞ and we can simplify the above expression:

( )
1
C clayton
θ (u, v ) = u −θ + v −θ − 1 θ−
(2.13)
With θ expresses the degree of dependence among the marginal components.
To illustrate the range of bivariate behaviour that can be represented by Clayton copula, consider the
following figures:

Figure 1: 1000 simulated standard uniform random variables under Clayton copula.

The Clayton copula has lower tail dependence but not upper tail dependence. The contour generated
by the Clayton copula implies fat tailed distribution. The contour or the level curves of a copula C are
{ }
given by (u, v ) ∈ I 2 / C(u, v ) = t .

3 . M ode lling joint de fa ult t im e s

In this section, we present some mathematical background for modelling joint default times. We
assume a probability space (Ω, ℑ, {ℑ t }t ≥0 , Ρ ) , where Ω is the underlying probability space containing all
possible events over a finite time horizon. ℑ is a σ -field representing the collection of all events.
{ℑ t }t ≥0 is a filtration that carries information with times and Ρ is a probability measure5. The pricing
is assumed under no arbitrage and then, Ρ is risk-neutral measure.
Default time τ i for each obligor i = {1,..., n} should be a random variable and the event of default should
be known for everybody at any times because we assume a perfect market with a free flow of
information. Default is a stopping time τ with respect to the filtration: {τ < t} ⊂ ℑ t , ∀t ≥ 0 . Let
F (t ) = Ρ(τ ≤ t ) be the distribution function and f (t ) is the density function of stopping time, the hazard
rate h of τ or intensity process is defined such that we have for the probability of default until time (t
+ ∆t ) given survival till t is given by:

5
According to Bielecki et al (2005), the probability space is endowed with a filtration ℑ = H ∨ F , where the
filtration H carries information about evolutions of credit events, such as changes in credit ratings of perspective
credit names and F is some reference filtration.

7
 f (t ) dF(t ) P(t < τ ≤ t + ∆t / τ > t )
h (t ) =
1
= × = lim (3.1)
1 − F(t ) 1 − F(t ) dt ∆ →0 ∆t

With dF = (t < τ ≤ t + dt , τ > t ) , 1 − F (t ) = P(τ > t ) = S (t ) is called the survival function that gives the
probability that a security will attain age t .the hazard rate function gives the instantaneous default
probability for a security that has attained age t .The marginal survival distributions S i (t i ) is assumed
to be smooth and strictly decreasing, this can be written as:
t
− ∫ h i (s )ds
S i (t i ) = 1 − Fi (t i ) = e 0
(3.2)

t
− hi (s )ds
∫
Then Fi (t i ) = 1 − e 0
(3.3)
hi (.) is the default intensity process for entity i. The default times τ i are defined:
⎧⎪ t ⎫⎪
τ i := inf ⎨t ≥ 0 : hi (s )ds ≥ θ i ⎬
∫ (3.4)
⎪⎩ 0 ⎪⎭
Where θ i has an exponential distribution with unit intensity6.

For a long time, finance and risk management treated dependence and correlation as basically
equivalent. Based upon law of large number arguments, it was assumed that risk factors were normally
distributed. This implies that the joint distribution function is determined by the vector of means and
the covariance matrix. Effects regarding correlation are particularly strong for some of the most recent
innovations in credit markets, namely single-tranche CDOs and basket default swap. Consequently, to
value such a contract the joint distribution of default times τ i :
F (t1 ,..., t n ) = P(τ 1 ≤ t1 ,..., τ n ≤ t n ) (3.5)
The joint survival time distribution is given by:
S (t1 ,..., t n ) = P(τ 1 > t1 ,..., τ n > t n ) (3.6)
⎛ ⎛⎜ t ⎞ ⎛ t ⎞
⎜ − h (s )ds ⎟ ⎞
⎛ t ⎞
⎜ ⎜ − ∫ h1 (s )ds ⎟⎟ ⎜⎜ − ∫ h2 (s )ds ⎟⎟ ⎜ ∫ n ⎟⎟
= C⎜ e ⎝ 0 ⎠ , e⎝ 0 ⎠ ,..., e ⎝ 0 ⎠⎟ (3.7)
⎜ ⎟
⎜ ⎟
⎝ ⎠
For deterministic intensities, this framework converges to Li (200) model .The times of default τ i are
defined as the first time the default countdown processes:
⎛ t ⎞
⎜ − h ( s ) ds ⎟
⎜ ∫i ⎟
λ i (t ) := e⎝ 0 ⎠ reach the level of the trigger variables U i :
τ i := inf {t ≥ 0 : λ i (t ) ≤ U i } (3.8)

The choice of a dependence structure between default times drives the prices of basket default swaps
and CDO tranches. Copulas functions allow us to separate the problem of modelling the default times
into two parts: first, the specification of the marginal distribution functions and second, the choice of a
suitable copula which describes the dependence structure between the default times. Then, the
marginal distributions together with the choice of a suitable copula are sufficient to specify the full
joint distribution of the default times.
The benefits for using copulas to modelling joint default times:
- Maintains input correlation matrix reasonably well.
- Distribution-free approach.

6
The exponential distribution is used to model Poisson processes, which are situations in which an object
initially in state A can change to state B with constant probability per unit time λ. The time at which the state
actually changes is described by an exponential random variable with parameter λ

8
- Can be employed in simulation procedures.
- Allows for various dependence structures (including tail dependence).
- Generates an exact joint distribution.

4 . Ba sk e t de fa ult Sw a ps spre a d

The most common type of basket default swaps is the first-to-default swap (FTDS), where the seller
compensates the buyer any loss of the principal and also, possibly, the accrued interest of the asset in
the reference basket which defaults first. The main difference between (FTDS) and a credit default
swap (CDS) is the event causing payout for the contract (in one case, it is the first default of any of a
list of names and in the other is default of a single name). A nth to default basket default swap gives
protection against the nth default in the underlyings pool of credits. We are going to present different
pricing methodologies for this product.

Regular Periodic premium till contract

Protection expiration or nth credit event occurs Protection
Buyer th
Payment in case of n Default seller

Reference entity
(Basket of names)
Figure 1: nth to Default Basket

4.1. Pricing under Gaussian copula and t-student copulas using Monte Carlo simulations

The pricing of nth to default basket default swap depends on the time the nth credit defaults. The default
times of different obligors are connected to each other by a Gaussian or t-student copula. The marginal
default times of all credits in the basket must be known.
Suppose a basket of credit default swap with the following characteristics:

Basket default swap

A is the notional amount of the contract. The
N
Initial Par Value total value of the basket is V = ∑ A i
i =1
Number of obligors N
δ =1 for annual payment frequency, δ i i
Payment frequency
δ =0.25 for quarterly payment frequency…
Maturity date T
Basket seniority n=1 for first to default basket, n=2 for second
to default basket
Fair price s
Table 1: Basket default swap notations

According to Galiani (2003), the risk neutral price of the nth to default basket swap is computed by
equating the expected value of the discounted premium payment leg (fixed cash flow to be paid till
contract expiration T or nth credit event occurs) with the expected value of the discounted default leg
(contingent payment in case of default), under the equivalent martingale measure Ρ * . Under this
measure, the price processes of any tradeable security, discounted by the money market account, are
Ρ * -martingales with respect to some filtration.

9
The premium legs are paid as long as the underlying credit has not defaulted until the maturity of the
contract. The present value of the premium leg of the nth to default basket default swap can be
computed as follows:
⎡ K ⎤
E (PL n ) = E ⎢ pδβ (t i )
⎢⎣ i =1
∑
I {τ n >t } ⎥
⎥⎦
(4.1.1)

Where ( p = f .A ) the premium leg as function as the fair spread of the contract (f) as a fraction of is
notional amount (A) in basis point, t i , i ∈ {1,..., k} are the payment dates either until T or until τ < T in
case of default, δ i is the frequency of payment and β(t ) is the discount factor.
( )
Let Let F(n ) (t ) = Ρ * τ n ≤ t be the distribution function of τ , then we can rewrite the premium leg as:
n

∑ [1 − F (t i )]
K
E (PL n ) = pδβ (t ) n
(4.1.2)
i =1

The second part for pricing nth to default swap is the default leg E[DL n ] . The default leg ban be
expressed as the difference between the expected discounted default payment E [DPn ] and the
expected discounted accrued premium E [APn ] . Then, E[DL n ] = E [DPn ] - E [APn ] .
With:
⎡ ⎤
( ) ( )∑ Ι {τ
N
E [DPn ] = E ⎢ A. 1 − R n β τ n n
≤T }
⎥
⎢⎣ j =1 ⎥⎦

∑ (1 − R )∫ β (τ ) F( )
N T
=A n n nth = j
n (dt ) (4.1.3)
j =1 0

F(nn ) = j (t ) is the distribution function of n basket default relative to the jth defaulter
th
th
We notice that
for allowing different recovery rates for the obligors.
⎡ K ⎛τ n − t ⎞ ⎤
E [APn ] = E ⎢ p ⎜ ∑ i −1 ⎟
( )
δ β τ n Ι {t <τ n ≤t } ⎥
⎢⎣ i =1 ⎜⎝ t i − t i −1 ⎟⎠ i −1 i
⎥⎦

ti
K
v − t i −1
=p ∑ ∫t
i =1 ti −1 i − t i −1
δβ (v )F( n ) (dv ) (4.1.4)

The fair spread of the basket default swap is given as:

∑ (1 − R )∫ β (τ )F( )
N T
n n n th = j
(dt )
E [DL n ]
n
j =1
s= = 0
E [PL n ] + E [APn ]
[ ] ∫ tv −−tt
K ti (4.1.5)
∑i =1
δβ (t )1 − F n (t i ) + i −1

i −1
δβ (v )F(n ) (dv )
ti −1 i

Pricing a nth to default basket default swap under Gaussian and t-student copula using Monte Carlo
simulations can be presented as the following steps:
Step 1: simulate N-dimensional vector of correlated uniform random variates from a copula
( C Σ or C ν,Σ ) as described in § 2.
Step 2: Translate the corresponding uniform variates into default time for each obligors.
Step 4: Sort the credits with respect to their default time τ i and determine the nth default time τ n . For
first to default swap, we find the first default time τ * = min τ i .
i

10
Step 5: Based on specific realization of τ n Determine the present value of the premium leg E [PLn ] .
Step 6: Determine the present value of the default leg E [DLn ] .
Step 7: Repeat all steps above until the required number of scenarios has been simulated and the
sample average fair spread of the nth to default basket swap as described in the equation (4.1.5).

To determine the impact of the structure of dependency via Gaussian and t-student copulas on the nth
to default basket swap spread, a simulation study was performed for different baskets of different A
50000 Monte Carlo simulations is performed.

Basket default swap (BDS 1): Homogeneous baskets of five names, maturity date T=5, constant
default intensities h=0.01; constant interest rate 5%, a deterministic recovery rate of 40%, δ =0.25 for
quarterly payment frequency and correlation between each pair of entities rho=0.3

Basket default swap (BDS 2): Homogeneous baskets of five names, maturity date T=5, constant
default intensities h=0.01; constant interest rate 5%, a deterministic recovery rate of 40%, δ =0.25 for
quarterly payment frequency and correlation between each pair of entities rho=0.6

Basket default swap (BDS 3): Homogeneous baskets of 10 names, maturity date T=5, constant default
intensities h= 0.05; constant interest rate 5%, a deterministic recovery rate of 40%, δ =0.25 for
quarterly payment frequency and correlation between each pair of entities rho=0.5

Fair price
Fair price
(Gaussian copula)
(t-student copula
Second
First to Second to Third to First to Third to
to
default default default default default
default
(BDS 1) 446.7645 140.4528 52.8958 385.7386 153.3905 40.3990
(BDS 2) 294.6252 138.3890 78.8875 302.1418 122.9784 93.5541
(BDS 3) 1.1597e+003 690.5959 480.8674 1.7200e+003 810.5678 447.4202
Table 2: Basket default swap premiums(on basis points)

We notice that the premium computed under Gaussian and t-student copulas are different. Burtschell
et al (2005) find that Gaussian and t-student copulas lead to quite similar premium for first to default
swap when they change the number of names in the basket. Similarly, the differences are minor when
they change the rank of default

Figure 2: Gaussian copula spread of first default swap as function of correlation

using Monte Carlo simulation.

11
4.2. Pricing under one factor Gaussian copula model

In this section, we implement a one factor model approach first presented by Laurent and Gregory
(2003), and Andersen, Sidenius and Basu (2003). The one factor Gaussian copula is the copula
associated with multivariate normal random variables that display the correlation structure induced by
linear dependence on a single common normally distributed factor. This model results in semi-
analytical expressions for the price of basket credit derivatives, avoiding the time consuming
simulation step needed with the Monte Carlo method. In the semi-explicit approach, we must
determine the probability P(N(t ) = m ) that a certain number of credits has defaulted at time t and the
probability that at time t, credit i has defaulted as nth credit in the basket. Given these probabilities, we
can compute analytically the present value of the premium and default leg and then, the price of the nth
basket default swap.
Suppose that V1 ,..., Vm are marginally standard normal variates. In a one factor model we write:
(4.2.1)
Vi (t ) = a i C + 1 − a i2 ε i

With C is a common factor for all firms and ε i is the error terms. C and ε i are independent standard
normally distributed random variables and then, cov(Vi , V j ) = a i a j and Var(Vi ) = 1. The term ai
determines how strong Vi is linked to the evolution of the common factor C.
The default of an obligor is triggered when Vi ≤ K i with K i is the default barrier.
The link between the Vi and the default time τ i is given via the marginal distribution of the default
time: Fi (t ) = P(τ i ≤ t ) = P(Vi ≤ K i ) (4.2.2)
K i = Φ −1 [Fi (t )] , with Φ −1 is the inverse cumulative distribution function.
The conditional default probability that credit i defaults at time t conditional on C is :
⎡ −1 ⎤
Φ [Fi (t ) − a i C]⎥
p i (t \ C ) = P(Vi ≤ K i \ C ) = Φ ⎢ (4.2.3)
⎢ 1 − a i2 ⎥
⎣ ⎦
The times of default conditionally on C are independent. The probability of default joint is written as:
⎡ ⎤
Φ −1 [Fi (t )] − a i c ⎥
+∞ n
F(t 1 ,...., t n ) = ∫ ∏ Φ ⎢⎢ g(c )dc, (4.2.4)
− 2 ⎥
− ∞ i =1 ⎣ 1 a i ⎦
⎡ ⎤
a i c − Φ −1 [Fi (t )]⎥
+∞ n
S(t 1 ,...., t n ) = ∫ ∏ Φ ⎢⎢ g(c )dc. (4.2.5)
1 − a i2 ⎥
− ∞ i =1 ⎣ ⎦
e −c
2
/2
With g(c ) = represent the density function of C.
2π
The pricing of nth to default basket default swap involve the knowledge of risk neutral probability of
+∞ n
nth default according to time: S(t 1 , t 2 ,..., t n ) = ∫ ∏ q i (t \ c )g(c )dc, (4.2.6)
− ∞ i =1
With q i (t \ C ) is the survival function conditional to credit i. The number of default until time t is
N(t ) = ∑ I {τ i ≤ t} .
n
presented via the following process: (4.2.7)
i =1
The probability of nth to default in the basket following the one common factor can be written as:
P[N (t ) = n ] = ∫ P[N (t ) = n \ c]g (c )dc (4.2.8)
We must determine P[N(t ) = n \ C] , n = {0,1,..., N} with N is the total number of credits in the basket
n ≤ N . For n = 0 , then no default in the basket. The conditional probability of not default can be
written as :
n
P[N(t ) = 0 \ C] = ∏qi (t \ C). (4.2.9)
i =1

12
The conditional probability of one default n = 1, can be written as:
⎡ ⎤
P[N( t ) = 1 \ C] = ∑ ⎢⎢(1 − q i )∏ q j ⎥⎥
n n

i =1 j=1
⎢⎣ j≠ i ⎥⎦
⎡ ⎤
n qj
P[N( t ) = 1 \ C] = ∑ ⎢⎢(1 − q i )∏ ⎥⎥
n
(4.2.10)
i =1 j=1 q i
⎢⎣ ⎥⎦
We use (4.2.9) and (4.2.1), then:
P[N( t ) = 1 \ C] = P[N(t ) = 0 \ C]∑ w i (t \ C ) ,
n
(4.2.11)
i =1
1 − q i (t \ C )
w i (t \ C ) = .
q i (t \ C )
We repeat the same procedure to nth default in the basket, n = {2,3,..., N} we can easily verify:
P[N( t ) = 1 \ C] = P[N(t ) = 0 \ C]∑ w p (1),..., w p (k )
The nth survival function can be presented as follows: S n (t ) = P τ n > t , ( )

S n (t ) = P[N(t ) < n ],
n −1 (4.2.12)
Sn (t ) = ∑ P[N(t ) = j].
j=0
th
With τ is the n default time.
n

The premium legs are paid as long as the underlying credit has not defaulted until the maturity of the
contract. The accrued premium payments are taken into account. We suppose that free risk interest rate
and recovery rate are constant. The present value of the premium leg of the nth to default basket default
swap can be computed as:
⎡ ⎤
E(PL n ) = E ⎢p n δ i β (t i )∑ I {τ n > t }⎥
I
(4.2.13)
⎣ i =1 ⎦
Where p n is the nth primium leg, t i , i ∈ {1,..., I}are the payment dates of the premium leg, δ i is the
frequency of payment and β(t ) is the discount factor. We combine (4.2.12) and (4.2.13), we can
rewrite the present value of the premium leg of the nth to default basket default swap as:
n −1
E[PL n ] = p n ∑ δ i β (t i ) ∑ P[N(t i ) = n ]
I
(4.2.14)
i =1 n =0

The second part for pricing nth to default swap is the default leg E[DL n ] . We suppose that the basket is
composed by homogeneous credit with the same notional and then, the same recovery rate. The loss in
case of default can be written as:

[ ( ) ] T
E[DL n ] = E (1 − R )β τ n I {τ n ≤T} = −(1 − R ) ∫ β(t )dS n (t ),
0
(4.2.15)

With T is the maturity date. The integration of the precedent equation can be written as:
[
E[DL n ] = (1 − R )1 − β (T )S n (T ) + ∫0T S n (t )dβ (t ) ] (4.2.16)

Following the hypothesis of risk neutral probability and constant free risk interest (r) The loss in case
of default can be written as:
⎡ T ⎤
E[DL n ] = (1 − R )⎢1 − e − rT S n (T ) − r ∫ S n (t )e − rt dt ⎥ , (4.2.17)
⎣ 0 ⎦

13
The term of pricing a basket default swap (spread: S) can be computed by dividing the present value of
the default leg E[DL n ] through the present value of the premium leg E[PL n ] :
E[DL n ]
s=
E[PL n ]

Suppose a basket default swap with 20 names with constant nominal amount and assuming correlation
matrix with one factor a = 0.6 ( a = ρ ) and a recovery rate of R= 40%. We can find the ongoing
premiums and see the effect of hazard rate.

nth to default h=0.01 h=0.04 h=0.06 h=0.08

1 419.11 1313.2 1860.6 2410.9
2 229.91 852.5 1219.5 1607.3
3 159.88 648.5 943.3 1238.1
4 118.38 511.1 779.5 1019
5 87.49 406.6 659.8 872

Table3: Spread of nth to default swap with n=20, correlation factor a = 0.6
and recovery rate R=40%.

Note that the effect of hazard rate is important for the pricing of the nth to default swap. For h=0.01,
the spread of first to default swap is 419.11 basis point. However, for h=0.08, the spread of first to
default swap is2410.9. Indeed, the highest spread always defaults first since the value of hazard rate
reflect the intensity of default probability.

Figure 3: Spread of nth to default swap for different hazard rates.

Figure 4 shows that spread of basket default swap decrease according to the number of defaults with
different hazard rate. From the point of view of the protection seller, the higher the nth to default, the
less likely the spread is to pay. Similarly, we notice that the level of spread increase when the hazard
rate increase.
Now, assuming hazard rate h=0.06 and a recovery rate of R= 40%. We can find the ongoing premiums
and see the effect of correlation factor.
nth to default a =0 a = 0.1 a = 0.3 a = 0.6

1 8449.1 6130.5 3635.6 1860.6

2 3695.2 2925.1 2004.2 1219.5
3 2205.8 1817.3 1360.5 943.3
4 1421.9 1219.9 993.8 779.5
5 903.9 834.8 751.6 659.8

Table 4: Spread of nth to default swap with n=20, h=0.06 and recovery rate R=40%.

14
 Figure 4: Spread of nth to default swap for different correlation factors.

Figure 5: one factor Gaussian copula spread of first default swap as function of correlation

Similarly, Figure 5 shows that spread of basket default swap decrease according to the number of
defaults with different correlation coefficient. However, we notice that the level of spread increase
when the correlation coefficient increase.

4.3. Pricing under one factor Clayton copula

The structure of dependency via Clayton copula in credit risk analysis has been considered by many
authors like Laurent and Gregory (2003), Madan et al (2004), Burtschell et al (2005). Let C any
positive random variable with f (c ) be the marginal density. C is called the common factor. The
conditional distribution function as conditionally independent given C is some distribution function
Gic (x ) :
p i ( X i < x \ C = c ) = Gic (x ) . (4.3.1)
The joint density of variables X i is given by:
M
∞
M
∞ c ∑ Ln (Gi ( xi ))
F (x1 ,..., x M ) = ∏Gic (x i ) f (c )d (c ) == e
∫ ∫ i =1
f (c )d (c ) (4.3.2)
i =1
0 0

⎛ M ⎞
= ψ⎜ − ∑ Ln(G i (x i ))⎟ (4.3.3)
⎝ i =1 ⎠
∞
where ψ(s ) is the Laplace transform of C and ψ(s ) = ∫ e −sc f (c )d(c ) . From (4.3.2), we can present the
0
implied marginal densities:
∞ ∞
Fi (x i ) = ∫ G ic (x i )f (c )d(c ) == ∫ e cLn (G i (x i )) f (c )d(c )
0 0

15
 = ψ(− Ln(G i (x i ))) (4.3.4)
− ψ −1 (Fi ( x i ))
Then: G i (x i ) = e
Substituting (4.3.4) into (4.3.3),, we obtain:
⎛M ⎞
F(x 1 ,..., x M ) = ψ⎜ ∑ ψ −1 (Fi (x i ))⎟
⎝ i =1 ⎠
A typical example is the Clayton copula where the variable C has a Gamma distribution with
1
parameter , where θ > 0 . Then, we obtain:
θ

1
f (v ) = e − v C (1−θ ) / θ , with ψ(s ) = (1 + s )− θ ,
1
(4.3.5)
⎛1⎞
Γ⎜ ⎟
⎝θ⎠
ψ −1 (u ) = u − θ − 1 .
1
−
⎛M ⎞
C(u 1 ,..., u M ) = ⎜ ∑ u i−θ − M + 1⎟
θ
Then, (4.3.6)
⎝ i =1 ⎠

The conditional default probability that credit i defaults at time t conditional on C is given by
equation(4.2.3) in case of one factor Gaussian copula. In the one factor Clayton copula:

((
pi (t \ C ) = exp C 1 − Fi (t )−θ )) (4.3.7)

The joint survival function on one factor Gaussian copula is given by (4.2.5). In the one factor
Clayton copula:
+∞ n 1−θ
S (t1 ,...., t n ) = ∫ ∏ (1 − pi (t \ C))
1 ( )
e −C C θ .dc (4.3.8)
⎛1⎞
− ∞ i =1 Γ⎜ ⎟
⎝θ ⎠

Clayton
Clayton Gaussian Gaussian Clayton Gaussian
θ = 0.36 Rho=0.5
θ = 0.18 Rho=0.3 θ = 0.66 Rho=0.7
First to default 462.6 454.9 293.8241 287.8186 110.6611 92.6339

Table 5: Spread of nth to default swap under one factor Clayton copula

Figure 6: Spread of first default swap as function of theta (one factor Clayton copula)

16
4.4. Pricing under Optimised Monte Carlo simulations via Importance sampling

Variance reduction has always been a central issue in Monte Carlo experiments. Importance sampling
is a variance reduction technique that can be used in the Monte Carlo simulation. The idea behind
Importance sampling is that certain values of the input random variables in a simulation have more
impact on the parameter being estimated than others. If these "important" values are emphasized by
sampling more frequently, then the estimator variance can be reduced.

Joshi & Kainth (2004) apply importance sampling to the pricing of nth to default credit swaps within
the Li model and obtain stable and sizeable speed ups. They show that Monte Carlo simulations in the
Li model can be slow to converge and present procedures for accelerating the computation of prices
and sensitivities to hazard rates. Glasserman &Li (2005) develop importance sampling (IS) procedures
for rare-event simulation for credit risk measurement. They focus on the normal copula model
originally associated with J.P. Morgan’s CreditMetrics system. Dependence between obligors is
captured through a multivariate normal vector of latent variables; a particular obligor defaults if its
associated latent variable crosses some threshold. Glassserman &Juneja (2006) have considered the
problem of simultaneous estimation of the probabilities of multiple rare events. Successful
applications of importance sampling for rare event simulation typically focus on the probability of a
single rare event. As a way of demonstrating the effectiveness of an importance sampling technique,
the probability of interest is often embedded in a sequence of probabilities decreasing to zero. They
shows that Importance sampling based on exponential twisting produces aymptotically efficient
estimates of rare event probabilities in a wide range of problems.

According to Milicia (2006), Importance Sampling is a change of probability measure which takes as
from the original probability measure to a new, biased, one which better suits our purposes
Assume we want to estimate by simulation the probability p of an event (X ≥ t ) where X is a random
variable with probability density function f (x ) . Importance sampling is concerned with the
determination of an alternate density function f Θ (x ) , usually referred to as a biasing density, for the
simulation experiment. This density allows to the above event to occur more frequently, so the
sequence lengths δ gets smaller for a given estimator variance:

f (x )
p = E(X ≥ t ) = ∫ (x ≥ t ) f Θ (x )dx = E Θ ((X ≥ t )L(X ))
f Θ (x ) (4.4.1)

f (x )
L(x ) = (4.4.2)
f Θ (x )
L(x ) is the the Radon-Nikodym derivative or likelihood ratio of the change of measure. The
Importance sampling estimator is then:

∑ (X i ≥ t )L(X i )
1 K
p IS = (4.4.3)
K i =1
Where X i ~ f Θ

σ MC
2
In order to identify successful IS techniques, the variance ratio , and this can be interpreted as the
σ IS
2

speed-up factor by which the Importance sampling estimator achieves the same precision as the Monte
Carlo estimator.
Joshi and Kainth (2004) explain the implementation difficulties of Li (2000) model. For a short team
deal, many Monte Carlo paths result in a zero or constant pay-off. Then, they sample more intensively
the areas where the pay-off is rapidly changing. For basket credit derivative, this means ensuring that
every path has enough defaults to result in a non-trivial pay-off and achieving a reasonable distribution

17
within that area. Subsequently, they achieve this by redefining the probability measure, and weighting
the pay-off’s value to compensate for the adjustment.
Assume a basket default swap of size N, The Joshi and Kainth method first determines whether a
particular asset defaults within the life of the swap, and, if a default is to occur, it then determines the
time of the default in the interval [0; T]. Only the default probabilities are changed7.

Let default indicator variables Yi = I(τ i ≤ T ) , (i=1,…,N) be the indicator variables for defaults which
occur within the lifetime of the contract. Conditional default probabilities are then given by P1 = F1 (T )
and Pi = P(Yi = 1 / τ 1 ,..., τ i −1 ) = Fi (T / τ 1 ,..., τ i −1 ) , i=2,…N.
The importance sampling probabilities are defined as:
⎡ i −1 ⎤
⎢ n − j∑=1
Yj
i −1 ⎥
⎢ ∑ Yj < n ⎥
n ⎢ N − i +1 j=1 ⎥
p1 = , ~
pi ⎢ ⎥ (4.4.4)
N
⎢ i −1
⎥
⎢ pi ∑ Yj ≥ n ⎥
⎢⎣ j=1 ⎥⎦

i −1
With ∑ Y j the number of assets which have defaulted by time T.
j=1

Joshi and Kainth (2004) algorithm can be presented as:

Step1: draw a N-dimentional vector U uniform variates

⎡ pi Ui ⎤
⎢ ~ if U i ≤ ~
pi ⎥
⎢ pi ⎥
Step2: Let Vi ⎢ ⎥ (4.4.5)
⎢ 1- p ⎥
⎢ pi + ~ i (U i − ~
pi ) otherwise ⎥
⎣⎢ 1- pi ⎦⎥

And τ i = Fi−1 (Vi / τ 1 ,..., τ i −1 ) (4.4.6)

Step 3: Calculate the weight (likelihood ratio):

⎡ pi ⎤
⎢~ if Yi = 1 ⎥ (4.4.7)
p
Li = ⎢ i ⎥
⎢ 1 - pi if Yi = 0 ⎥
⎢⎣ 1 - ~
pi ⎥⎦

Once τ 1 ,..., τ N have been generated, we evaluate V(τ 1 ,..., τ N ) and return the weighted
estimate V(τ 1 ,..., τ N ) L with L = L1 L 2 ...L N the weight for the path.
To implement this algorithm, we must sample from the conditional default time distribution in
(4.4.6). Let ∑ = A ⋅ A ' be the positive definite matrix, let W = A ⋅ Z , then Wi = Φ −1 (Fi (τ i )) .Then
conditioning on τ 1 ,..., τ i is equivalent to conditioning on Z 1 ,..., Z i

7
Chen and Glasserman (2006) show that the importance sampling probabilities used by Joshi and Kainth (2004)
may be interpreted as follows: Consider an urn initially containing n black balls and N-n white balls. Balls are
drawn from the urn at random, without replacement. If j < n of the first i-1 balls are black, then the probability
that the ith draw produces a black ball is (n-j)/(N-i+1). All n black balls will eventually be drawn.

18
 ⎛ ⎛ i =1 ⎞⎞
⎜ Φ −1 ⎜ F (t ) − ⎟⎟
⎜ ⎜ i ∑ A Z
ij j ⎟ ⎟
⎝ ⎠⎟
P(τ i ≤ t / Z 1 ,..., Z i −1 ) = Φ⎜⎜
j =1
⎟ (4.4.8)
Aii
⎜ ⎟
⎜⎜ ⎟⎟
⎝ ⎠
Milicia (2006) adopt the Importance Sampling method8 to other elliptical copula like t-student
ν
copula. Let W = ⋅ A ⋅ Z with s ~ χν2 and τ i = Fi −1 (Tν (Wi )) , then :
s

⎛ Tν−1 (Fi (t )) i −1 ⎞
⎜
⎜ ν
− ∑A Z
j =1
ij j
⎟
⎟
⎜ ⎟
P(τ i ≤ t / Z 1 ,..., Z i −1 ) = Φ⎜ ⎟
s
(4.4.9)
⎜ Aii ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎝ ⎠

Fair price Fair price Fair price Fair price

(Gaussian (t-student Gaussian Gaussian
copula) copula) copula) JK copula)
MC MC IS JKM IS

446.7645 385.7386 438.11 439.36

(BDS 1) (0.0084788) (0.008478) (0.08478) (0.000612)

Table 6: Spread of First to default swap under Monte Carlo and Importance sampling

We notice that the spread of first to default swap change with the structure of dependency and the
simulation techniques. Then, the choice of copula and the choice of procedures for rare-event
simulation govern, also, the pricing of basket credit derivatives.

5 . Colla t e ra lize d de bt obliga t ion spre a d

Collateralized debt obligation (CDO) refers to securitization9 of pools of assets. A CDO cashflow
structure allocates interest income and principal repayments from a collateral pool of different debt
instruments to a prioritized collection (tranches) of CDO securities. Following the classification of
Tvakoli (2003), a CDO is backed by portfolios of assets that may include a combination of bonds,
loans, securitised receivables, asset-backed securities, tranches of other CDO’s, or credit derivatives
referencing any of the former. Some market practitioners define a CDO as being backed by a portfolio
including only bonds. A Collateralized loan obligation (CLO) is a type of CDO that is backed by a
portfolio of loans. A Collateralized bond obligation (CBO) is a type of CDO that is backed by a
portfolio of bonds issued by a variety of corporate or sovereign obligors. The development of
structured credit derivatives leads to the emergence of synthetic Collaterized Debt Obligations which
transfer the risk of a pool of single-name Credit Default Swaps. This realizes an exposure to a variety
of names.

8
We use the Matlab code for Importance Sampling as implemented in the paper of Milicia (2006). An
electronic version of the article and the corresponding Matlab code are available online at
http://www.brics.dk/~milicia/libraryGM.htm
9
According to Tvakoli (2003), securitization has been a means for banks to reduce the size of their balance
sheets and to reduce the risk on their balance sheets. This allowed banks to do more business and allowed
investors accsess to diversified pools of assets to which they otherwise not have had access.

19
 Ca Cash Cash
Loan Bond proceeds proceeds
Special Senior Class
collateral Purpose
Note 30%-100%
Vehicle
(Bank) issuance
Loan (SPV)
Assets Mezzanine Class
10%-30%

Equity Class
0%-10%
Figure 7: CDO

Figure 6 illustrate an example of a simple CDO structure with three tranches: equity, mezzanine and
senior. Suppose that the total CDO notional is 100 millions and during the lifetime of CDO some debts
in the collateral portfolio might default. At maturity, if the total default loss is less than 10 millions,
only the equity tranche is affected. If the total losse is between 10 and 30 millions, the equity tranche
does not get the principal back and the mezzanine gets only part of it. If the loss is more than 30
millions than the equity and mezzanine do not get anything back and senior tranche gets is left.
From an economic perspective, CDO structures thus create custom exposures that investor’s desire
and cannot achieve in any other way. However, CDO address some important market imperfections.
First, banks and certain other financial institutions have regulatory capital requirements that make it
valuable for them to securitize and sell some portion of their assets, reducing the amount of
(expensive) regulatory capital that they must hold. Second, individual bonds or loans my be illiquid,
leading to a reduction in their market values. Securitization may improve liquidity, and thereby raise
the total valuation to the issuer of the CDO structure. Third, adverse selection10 can be mitigated by
securitization of assets in a CDO. The seller achieves a higher total valuation (for what is sold and
what is retained) by designing the CDO structure so as to concentrate the “lemon’s premium”11 into
small subordinate tranches, leaving the large senior tranch relatively immune to the effects of adverse
selection.According to Duffie & Gârleanu (2001), the effect of adverse selection can be discovered in
the transfer of bank loans or junk bonds. There is an informational asymmetry between the potentially
better-informed seller of such assets and the potentially uninformed buyer, so there is a price reduction
sometimes called a lemon’s premium.

There are a number of ingredients involved in pricing individual tranches: default probability,
recovery rate and default correlation. The first two are self-explanatory. Correlation, however, deals
with the distribution of defaults throughout a portfolio and the likelihood of a single default causing a
succession of defaults. Mahadevan and Schwartz (2001) identify three broad types of CDO pricing
methodologies: rating methodologies that infer a credit rating for the CDO from the ratings of its
constituent parts and the relationships between them12which is then used to price the CDO off
similarly rated bonds and CDOs. Market value methodologies that essentially equate the CDO price to
the sum of the market values of the constituent parts (Duffie & Gârleanu (2001), Mashal (2002)…).
Finally, Cash flow methodologies that involve discounting back simulated future cash flows.

10
There may be a significant amount of private information regarding the bank loan. An investor may be
concerned about being “picked off” when trading such instruments.
11
The reduction in price due to adverse selection is sometimes called a “lemon’s premium”.
12
For example the process used by Moody’s, they calculate “diversity scores” by which the analysis of a
portfolio of correlated assets is effectively simplified into an analysis of a portfolio of uncorrelated assets.

20
5.1. Pricing under Gaussian and t-student copula using Monte Carlo simulations
Consider an homogeneous CDO with n obligors with nominal amount A i and recovery rate Ri with i=
1,2,…,n, (assumed deterministic), maturity T years and we assume constant risk free interest rate. The
total value of the portfolio is VT = ∑ A i and L i = (1 − R i ) A i will denote the loss given default for the
n

i =1

i credit. Let τ i be the default time of the i th name and N i (t ) = ∑ I {τ ≤ t } be the counting process which
th n
i
i =1
jumps from 0 to 1 at default time of name i. let L(t ) will denote the cumulative loss on the collateral
portfolio at time t:

L(t ) = ∑ L i N i (t ) .
n
(5.1.1)
i =1

The tranche [a,b] suffers a loss at time t if a %VT < L(t ) ≤ b%VT , where a % and b% are respectively
lower and upper bound. Suppose that a %Vt = a ' and b%Vt = b ' , then, the tranche loss :

[ ] ( )
L a i ,b ' (t ) = L(t ) − a ' Ι {L(t )∈[a ' , b' ]} + b ' − a ' Ι {L(t )∈[b ' ,V ]}
t (5.1.2)

Using Monte Carlo simulation, the estimation of tranche loss becomes a straightforward task.
According to Peixoto (2004), Pricing a CDO using Monte Carlo simulation involves creating sample
paths of correlated default times. These default times are used to calculate the payments on two legs
and value each leg. The first is the present value of tranche losses triggered by credit events during the
CDO lifetime and is called default leg [DL] and the second is the present value of the premium
payments weighted by the outstanding capital (original tranche amount minus accumulated losses) and
is called premium leg [PL]. The fair spread of CDO can be computed by dividing the present value of
the default leg E[DL] through the present value of the premium leg E[PL] :
E[DL]
S= (5.1.3)
E[PL]
The Kth default leg can be computed as:
DL k
i =1
n
= ∑ e − rτi L
k

a ,b (τ )
k
i (5.1.4)

Where r is the free risk interest rate and τ1k , τ k2 ,..., τ kN { } the sequence of default times with K th

iteration of a Monte Carlo simulation. The accumulated loss is given by:

n
ℑ k (t ) = (1 − R )∑ I {τ k < t } (5.1.5)
i
i =1
The premium leg is paid over the outstanding capital in the tranche. If during the lifetime of the CDO
the tranche is wiped out, there are no more premium payments:
PL k = N ∑ δ j e
w

j=1
− rt j
{ [
min max b − ℑ k t j , 0 , b − a ( ) ] } (5.1.6)

Where {t 1 ,..., t w }are the premium payment dates with frequency δ j .

Table 3 presents fair spread of an homogeneous CDO with Monte Carlo simulation. Standard errors of
estimates are less than 1 basis point.

Tranche Spread (basis point)

0% à 10% (Equity) 2952,4
10% à 30% (Mezzanine) 779.3024
30% à 100 % (Senior) 43.4713
Table 8: Fair spread of an homogeneous CDO with h = 0.06, recovery rate R=0.4,
correlation coefficient rho=0.4, 50,000 iterations and quarterly spread payment.

21
Similarly to the basket default swap, we use t-student copulas to take into account the occurrence of
joint extreme events among obligors.

Tranche Spread (basis point)

0% à 10% (Equity) 3172.895
10% à 30% (Mezzanine) 762.065
30% à 100 % (Senior) 30.210
Table 9: Fair spread of an homogeneous CDO with h = 0.06, recovery rate R=0.4,
correlation coefficient rho=0.4, 50,000 iterations and Degree of freedom(DoF= 10)

Equity tranche prices appear to be higher when computed through the t-copulas, with the opposite
verified for the Senior tranche. Mezzanine tranches seem not to be particularly affected by the copula
function.

The principal payment and interest income are allocated to the notes according to the following rule:
Senior CDO notes are paid before mezzanine and lower subordinated notes are paid, with any residual
cash flow, to an equity tranche. Therefore, equity tranche offers a larger spread than the more senior
notes because is the first to be affected by a default in the portfolio. The price sensitivity with respect
to price driving factors such as correlation, recovery rate, and the credit-worthiness of the underlying
portfolio is examined. The following table present spread of CDO with different borns (a% and b%)

Tranche Spread (basis point)

0% à 5% (Equity) 4438.5
5% à 15% (Mezzanine) 1705.5
15% à 100 % (Senior) 130.4230

Table10: Fair spread of an homogeneous CDO with h = 0.06, recovery rate R=0.4,
correlation coefficient rho=0.4, 50,000 iterations and quarterly spread payment.

When we change the upper and the lower boundary of each trench, the CDO’s spread change also. In
fact, when we change the boundary of Equity tranch of (0%-10%) to (0%-5%), we note that the
CDO’s spread increase (2952,4 to 4438.5) because the expected loss in case of default incre for all the
equity tranche.. The same argument for Mezzanine and Senior tranche.
Now, we study the sensitivity of CDO’s spread to recovery rate. We assume that all other variables are
constant.

Tranche Spread (basis point), R=0.1 Spread (basis point), R=0.7

0% à 10% (Equity) 3860.6 1830.4
10% à 30% (Mezzanine) 1237 191.5912
30% à 100 % (Senior) 161.5789 9.5762e-004

Table 11: Fair spread of an homogeneous CDO with h = 0.06, recovery rate,
correlation coefficient rho=0.4, 50,000 iterations and quarterly spread payment.

22
 Figure 8: Fair spread of an homogeneous CDO with correlation coefficient rho=0.4,
50,000 iterations and quarterly spread payment.

Currently, we study the sensitivity of CDO’s spread to correlation coefficient13. We assume that all
other variables are constant.

Tranche Spread (basis point), rho=0.3 Spread (basis point), rho=0.7

0% à 10% (Equity) 3.4904e+003 1.7873e+003
10% à 30% (Mezzanine) 760.2878 776.2190
30% à 100 % (Senior) 27.1457 104.2831

Table 12: Fair spread of an homogeneous CDO with h = 0.06, recovery rate R=0.4
and 50,000 iterations and quarterly spread payment.

Figure 9: Fair spread of an homogeneous CDO with recovery rate R=0.4,

50,000 iterations and quarterly spread payment.

Unfortunately when correlations are implied from all of the tranche prices observed in the market,
each tranche has its own individual correlation which is different from the others. The value of the
lowest tranches of a CDO, say an equity tranche, increases as the correlation between defaults falls,
and decreases as default correlation rises. According to Gibson (2004), the effect of correlation on
CDO tranches is intuitive. The more the defaults within a portfolio become correlated, the more the
portfolio behaves like a single credit. A higher correlation of defaults implies a greater likelihood that
losses will wipe out the equity and mezzanine tranches and inflict losses on the senior tranche.
According to Rachev (2006), the value of the equity tranche rises as correlation rises. Equity tranche

13
CDO tranches are quoted in the market, they incorporate a correlation calculation. For example, a junior
mezzanine tranche (6%–9% of loss) might be quoted at a bid offer of 75/95bp, with a 16.5% correlation. An
equity tranche, however, might be seen as 1,400/1,600bp with a 25% correlation on the bid and a 20%
correlation on the offer.

23
investors gain more in a scenario with very few defaults than they lose from a scenario with many
defaults (they are only exposed to the first few defaults). Mezzanine tranches are subject to both
effects, which can broadly cancel each other out and make mezzanine tranches less sensitive to
correlation. The relationship is the other way round for the most senior tranches: as correlation
decreases, value decreases because the probability of a large number of defaults decreases.

As option premiums translate into a volatility smile through the Black & Scholes formula, CDO
spreads generate a correlation smile. Turc & Very (2004) use the beta model to imply a correlation
from the market spreads of CDO tranches. They present a bootstraping technique as a means for
incorporating the smile into CDO prices and hedges.

Tavares et al (2004) explain the difference on the behaviour of the correlation for each tranches firstly,
by the supply and demand argument that investors are vervous of the risk inherent in equity tranches
while Mezzanine tranches are extremely popular with investors. Sellers of protection on senior
tranches seem to require a minimum coupon irresepective of subordination. Secondly, the Gaussian
copula loss distribution underestimates the perceived chances of a very low or very high number of
defaults whilst significantly over-estimating the chances of observing a few defaults. They proposed a
new method named “Composed basket model” while it attemts to rationalise how market participants
allocate the risk.

Finally, we present the relationship between expected loss and recovery rate. In the Monte Carlo
simulations, the losses as well as the tranche notional are computed for every scenario individually and
the value of CDO tranches is the average of the present values. We use a Gaussian copula to model the
dependence structure in the portfolio. Following Galiani (2003), the pricing of CDO tranches with
Monte Carlo method can be presented as the following steps:
Step 1: simulate vi correlated N(0,1) random variables, i =1,…n.
Step 2: Find Ui’s such that U i = Φ (v i ) and Φ is the cumulative normal distribution function.
− ln U i
Step 3: Find the default times τ ' s by τ = , i =1,…,n.
λ
Step 4: Given the simulated default times, we can compute the value of losses at maturity for [a,b]
tranche L T [a , b] .
Step 5: Repeat all steps above until the required number of scenarios has been simulated (m=50000
iterations in our example). The estimator of expected loss can be computed as the average for every
scenario individually: E[L a ,b (T )] = ∑ L a ,b (T ).
1 m k
m K =1

Expected loss given Expected loss given Expected loss given

Tranche
default (%) with R=0.4 default (%) with R=0.1 default (%) with R=0.7
0% à 10% (Equity) 81.3169 87.0294 70.3784
10% à 30% (Mezzanine) 32.6471 52.4393 17.8161
30% à 100 % (Senior) 1.3022 4.7677 0.1222

Table 13 : Expected loss tranche with h=0.06, correlation coefficient rho=0.3 and 50.000 iterations.

According to Tables 9, 10 and 11 we notice that the expected losses of equity tranches is higher than
the mezzanine and senior tranche because the risk is higher. The equity note is also called the first loss
position because it is the first to be affected by a default in the portfolio. Also, we notice the
dependence of the expected loss given default and recovery rates for a fixed correlation. Table 14
shows that for recovery rate R=0.1, we obtain the expected losses of equity tranches is 87.0294%.The
expected losses of equity tranches of the same CDO but with recovery rate R=0.7

Figure 9 shows the relationship between expected loss tranche and recovery rate for different
correlation coefficient. Notice that the expected loss for equity tranche is larger meaning that

24
Mezzanine and senior tranches get less affected by the losses. This can be observed where the senior
expected loss gets close to zero at high recovery rates.

Figure 10: The relationship between Expected loss tranche and recovery rate.

Others extensions can be done with stochastic recovery rates. Garcia et al (2004) analyzes the effect of
stochastic recoveries on individual ratings. They develop a Monte Carlo simulation model that
generates credit events with both constant and stochastic recovery rates. They shown that non-constant
recovery rates can have a significant impact on binomial expansion technique (BET) ratings,
particularly for mezzanine tranches of CDOs with noninvestment- grade underlying collateral pools.

5.3. Pricing under One factor Gaussian copula method

The current standard for pricing CDO is the one-factor Gaussian copula. The market-standard version
of this copula is characterized by a single parameter to summarize all correlations among default
times. These individual default variables of the homogeneous portfolio are dependent on one
systematic risk factor which results both in a complexity reduction due to the factor-model and also in
conditionally independent defaults, whose properties simplify computations. For modelling purpose,
the conditional default probability that credit i defaults at time t conditional on C (equation (4.2.3)), is
the only inputs which include all model specification. The one-factor structure of this model implies
that, conditional on the realization of the common factor, the m individual credits are independent.
When pricing a standard CDO, this conditional independence greatly facilitates the calculation of the
conditional loss distribution of the tranche.

Tranche Spread on basis point

0% à 10% (Equity) 3628.1
10% à 30% (Mezzanine) 757.6
30% à 100 % (Senior) 46.91

Table 14: Fair spread of homogenous CDO with h=0.06, correlation factor of Gaussian copula a = 0.6
recovery rate R=0.4.

For the pricing of CDO, a one-factor Gaussian copula model with constant and equal pairwise
correlations, default intensities and recovery rates for all assets in the reference portfolio has become
the standard market model. Some drawbaks in the applicability of the one factor Gaussian model are
enumerated by Gregory and Laurent (2004). They present an extension to the Gaussian one-factor
copula model that allows a clustered correlation structure by specifying inter and intra sector
correlation. Additionally they introduce dependence between recovery rates and default events which
leads to an improved modeling of the smile. Hull and White (2005) argue that the market‘s focus on
implied correlations is misplaced. Andersen and Sidenius (2005) extend the one factor Gaussian
copula in order to match the “correlation smiles” in the CDO market. Instead of equation (4.2.1), the
latent variables are given by: Vi (t ) = B i ⎛⎜ a i C + 1 − a i2 ε i ⎞⎟ + (1 − B i )⎛⎜ β i C + 1 − β i2 ε i ⎞⎟ ; Where Bi are
⎝ ⎠ ⎝ ⎠

25
Bernoulli random variables, ai and β i determine how strong Vi is linked to the evolution of the
common factor C ( (0 ≤ β i ≤ a i ≤ 1) . Subsequently, there two states: one corresponding to a high
correlation and the other to a low correlation: Vi (t ) = (B i a i + (1 − B i )β)C + 1 − (B i a i + (1 − B i )β i )2 ε i .
They focus on a stochastic correlation Gaussian model which ai is linked to the evolution of the
common factor C with probability p and β i with probability (1- p).
Others extension of the one factor Gaussian copula is the one factor t-student copula and double t-
copula. In the t-Student approach, the random vector V follows a t-student distribution with ν degrees
of freedom. In the Double t-copula, C and ε i are independent random variables following t-student
distribution with respectively ν and ν ' degrees of freedom (Hull and White (2004), Burtschell et al
(2005)).

6 . Conc lusion
The aims of this paper is is to assess the importance of the dependence structure and the choice of
simulations procedures regarding the valuation on the of multi-name credit derivatives such as basket
default swaps and CDO tranches. The key idea of modelling correlated default is the usage of copulas
functions. The valuation models are set up with Gaussian, Clayton and Student t- copulas. 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. The advantage of
Monte-Carlo is its simplicity and generality. Its main drawbacks, however, are the quality of the
convergence, especially when one computes sensitivities. A good convergence is particularly hard to
achieve for credit products since default events are usually rare, and probabilities in the tail of the
distribution are difficult to estimate. On the other hand, the direct implementation in closed form is
very accurate but is less trivial to implement; it is also very expensive computationally. Indeed, this
method is based on enumerating the 2n default configurations of the basket and computing the
probability of each configuration. This algorithm explodes exponentially as the number of credits
increases.
In the second approach the correlation structure is simplified by a factor copula approach. This
approach enables us to provide semi-explicit expressions that reduce the computational times
compared with Monte Carlo simulations techniques. We investigate the influence of different price
drivers (correlation, hazard rates and recovery rates) on basket credit derivatives modelling portfolio
losses and basket credit derivatives spreads.
Furthermore, the Gaussian distribution has thin tails compared to other distributions. As we are
concerned of default events that are by nature tail events, we use distributions with fat tails such as the
t-student and Clayton distribution. Similarly, the choice of procedures for rare-event simulation
governs the pricing of basket credit derivatives. Joshi and Kainth (2004) shows that For baskets of
high-quality credits or short-maturity swaps, Monte Carlo simulation produces few paths with n or
more defaults; nearly all paths simply return a value of 0. Then, they introduce Importance Sampling
methods that forces all paths to produce at least n defaults. An alternative to the Monte Carlo
simulation is Clayton copula and t-student copula under importance sampling procedures for
simulation which captures the dependence structure between the underlying variables at extreme
values and certain values of the input random variables in a simulation have more impact on the
parameter being estimated than others.

26
References
Abid, F and N. Naifar (2005), “The impact of stock returns volatility on credit default swap rates: a copula
study”, International Journal of Theoretical and Applied Finance, Vol.8, No.8 pp: 1135-1155.

Abid, F and N. Naifar (2006 (a)), “The determinants of credit default swap rates: An explanatory study”,
International Journal of Theoretical and Applied Finance, Vol. 9, No. 1, pp. 23-42.

Abid, F and N. Naifar (2006 (b)), “Credit default swap rates and equity volatility: a nonlinear relationship”,
The Journal of Risk Finance, Volume 7, Number 4, August 2006, p: 348-371.

Andessen, L. and J. Sidenius, (2005), “Extention to the Gaussian copula: Random recovery and Random
factor loadings”, Journal of Credit Risk.

Bielecki, T. R., S. Crepey, M. Jeanblanc, and M. Rutkowski (2005),“Valuation of basket credit derivatives
inthe credit migrations environment”. Working paper, March 2005.

Bielecki, T. R., A. Vidozzi, and L. Vidozzi. (2006), “An efficent approach to valuation of credit basket
productsand rating triggered step-up bonds” Working paper, May 2006.

Burtschell, X.; Gregory, J.; Laurent, J.-P. (2005), “A Comparative Analysis of CDO pricing models”,
Working Paper, BNP Paribas.

Duffie, D.; Gârleanu, N. (2001), “Risk and Valuation of Collateralized Debt Obligations”, Working Paper,
Graduate School of Business, Stanford University.

Embrechts, P., McNeil, A., and D. Staumann, (1999), “Correlation and dependence in risk management:
Properties and pitfalls”, RISK (1999), 69–71.

Ericsson, J., K. Jacobs and R.A. Oviedo (2004), “The determinants of Credit Default Swap Premia”,
CIRANO, Montréal, 2004s-55.

Frey, R., and A. McNeil (2002), “Modeling dependent defaults”, working Paper, Department of Mathematics,
ETHZ.

Garcia, T., Maghakian, A., AND S. Sharma, (2004), “Implications of Stochastic Recovery Rates in
Evaluating CDO Tranches”, The Journal of Fixed Income, p: 64-71.

Gennheimer, H. (2002), “Model Risk in Copula Based Default Pricing Models”, Working paper n. 19, National
Centre of Competence in Research Financial Valuation and Risk Management.

Gibson, M.S. (2004), “Understanding the Risk of Synthetic CDOs”, Finance and Economics Discussion
Series 2004-36, Board of Governors of the Federal Reserve System (U.S.).

Gregory, J. and J.P. Laurent, (2004), “In the core of correlation”, Risk, October, pp: 87-91.

Hull, J and A. White, (2004) “Valuation of CDO and nth to default CDS without Monte Carlo simulation”,
Journal of Derivatives, Vol. 12(2), p. 8-23.

Hull, J., M. Predescu and A. White (2004), “The relationship between credit default swap spreads, bond
yields and credit rating announcements”, Working Paper, University of Toronto.

Jobst, A., (2002), “The pricing puzzle: the default term structure of collateralized loan obligation”, (CFS
working paper n°. 2002/14, 2002).

Joe, H., (1997), “ Multivariate models and dependence concepts”, (Chapman & Hall, London, 1997).

Jouanin, J-F., G. Riboulet and T. Roncalli (2001), “Beyond conditionnally independent defaults”, Credit
Lyonnais.

27
Li, D., (2000), “On default correlation: a copula function approach”, Journal of Fixed Income 9, (2000) 43-
54.

Madan, D. B., Konikov, M. and M. Marinescu, (2004), “Credit and basket default swaps”, (Working paper,
Robert H. Smith School of Business and Bloomberg L.P, 2004).

Mashal, R., and A. Zeevi (2002), “Beyond correlation: extreme co-movements between financial assets”,
Working paper, Columbia Graduate School of Business.

Nelsen, R., (1999), An Introduction to Copulas, (Springer Verlag, New York, 1999).

Norden, L. and M. weber (2004), “The comovement of credit default swap, bond and stock markets: an
empirical analysis”, Center for Financial Studies, Working Paper, No. 2004/20.

Papenbrock, J. (2006), “Price Calibration and Hedging of Correlation Dependent Credit Derivatives using
a Structural Model with alpha-Stable Distributions”,Diploma Thesis,Institute for Statistics and Mathematical
Economic Theory.

Sicar, R., and T. Zariphopoulou (2006), “Utility valuation of credit derivatives and application to CDOs”,
Working paper, Princeton University.

Sklar, A., (1959), “Fonctions de répartition à n dimensions et leurs marges”, Vol. 8 (Publications de l’Institut
de Statistique de l’Université de Paris, Paris, 1959).

Tavakoli, J.M (2003), “Collateralised debt obligations and structured finance”, John Wiley & Sons, Inc.,
Hoboken, New Jersey.

Tavares, P.A.C., Nguyen, T.U., Chapovsky, A., Vaysburd, I. (2004), “Composite basket model”, working
Paper, Merrill Lynch Credit Derivatives.

Turc, J., and P. Very., (2004), “Pricing and hedging correlation products”, Quantitative strategy, Bloomberg:
SGCT.

28
 Appendix
Some Matlab codes used in the paper

function y1 = Gaussrnd(rho,N);
% This function generates random numbers from Gaussian copula as presented in Figure 1.
% rho = the Gaussian copula correlation matrix
% N = number of random numbers to be generated
% y1 = random numbers from Gaussian copula, N by 2 matrix
if N == 0
y1=[];
else
A = chol(rho);
z = randn(N,2)*A;
u = cdf('Normal', z(:,1),0,1);
v = cdf('Normal', z(:,2),0,1);
y1 = [u v];
end

function y2 = Claytonrnd(Theta, N);

% This function generates random numbers from Clayton copula as presented in Figure 1.
% input:
% Theta: the Clayton copula parameter with Theta ∈ [−1, ∞ ) /{0}
% N = number of random numbers to be generated
% output:
% y = random numbers from Clayton copula, n by 2 matrix
if N== 0
y2=[];
else
y2 = NaN*ones(N,2);
q = rand(N,1);
u = rand(N,1);
v = ((q.^(-Theta/(1+Theta))-1).*u.^(-Theta)+1).^(-1/Theta);
y2 = [u v];
end

function y3 = studrnd(rho, Dof, N);

% This function generates random numbers from Student’s t copula as presented in Figure 2.
% rho = the correlation matrix
% Dof: degrees of freedom
% N = number of random numbers to be generated
% y3 = random numbers from Student’s t copula, N by 2 matrix
if N == 0
y3=[];
else
A = chol(rho);
z=randn(N,1);
s=chi2rnd(Dof);
y = z' ⋅ A
x = (sqrt (Dof ) / sqrt (s )) ⋅ y;
y3 = tcdf(x,Dof);
end

function[Fairspread]=Gaussn2dMc[N,n2d,rho,T,r,R,h]
% inputs:
% h: Default intensity for all obligors (assumed deterministic)
% N= number of obligors on the basket default swap (basket size)
% k: number to default in the basket default swap eg k=1 for first to default swap
% runs: number of Monte Carlo simulations
% rho: correlation matrix between obligors
% T: maturity of credit default swap, typical maturity is three, five and ten years.
% r: risk free interest rate (assumed deterministic)
% R: recovery rate

29
% delta: day count fraction on year eg delta=0.25 quarterly period
% ntimesubsteps=10; %no. of sub timesubsteps within each delta for integration
% outpout:
% Fairspread: return fair spread of nth to default swap as described in § 4.1

n=T/delta %time steps for indexing preminum payments eg. T=5, delta=0.25, n=5/0.25=20
dt=delta;
dt2=dt/ntimesubsteps;

%%%%%%%%%%%%%%%Simulation of default times%%%%%%%%%%%%%%%%%%%%

mu=zeros(N,1);
sig=ones(N,N)*rho + (1-rho)*eye(N);
%%%%%%%%%%%%%%% Simulation of default times via Gaussian copula %%%%%%%%%%%%%%
matrix=mvnrnd(mu,sig,runs);
defTime=-log(cdf('Normal',matrix,0,1))/h;
%%%%%%%%%%%%%%% Simulation of default times via t-student copula %%%%%%%%%%%%%%
%CorMat=(eye(N)-eye(N)*rho) + ones(N,N)*rho;
%A=chol(CorMat); %CorMat=A'*A here
%z=randn(scenarios,N);
%s = chi2rnd(DoF,1,N);
%y=z*A;
%x=repmat((sqrt(DoF)./sqrt(s)),scenarios,1).*y;
%u=tcdf(x,DoF);
%defTime = -log(u) ./ repmat(h,scenarios,N);
%%%%%%% %%%%%%%%% The value of premium leg %%%%%%%%%%%%%%%%%%%
t=dt;
PL=0; %Expected value of average Premium leg payments
for i=1:n,
%find probability of survival of kth default until each indexing period
B=exp(-r*t); %B is the discount factor at time t
SurvivalProb=1-sum(sum(defTime<t,2)>=k,1)/size(defTime,1);
PL=PL+delta*B*SurvivalProb;
t=t+(T/n);
end
PL
%%%%%%%%%%%%% The value of default leg %%%%%%%%%%%%%%%%%%%%%%%
DP=0; %Expected value of average default leg payments
kdefaultTime=sort(defTime,2);
kdefaultTime=kdefaultTime(:,k); %default time for each scenario
DPvec=(1-Recovery_rate)*exp(-r*kdefaultTime).*(kdefaultTime<T);
DP=sum(DPvec)/scenarios

%%%%%%%%%%%% The value of accrued premium leg %%%%%%%%%%%%%%%

AP=0; %Expected value of average Accrued premium
t=0;
for i=1:n,
curt=t;
for ti=1:ntimesubsteps,
%find probability that k th default happened
%between ti and ti+1
DefaultsVec=(kdefaultTime>(curt)) & (kdefaultTime<=(curt+dt2));
ProbOfDefault=sum(DefaultsVec)/scenarios;
curt=curt+dt2/2;
B=exp(-r*curt);
AP=AP+(curt-t)*B*ProbOfDefault;
curt=curt+dt2/2;
end
t=t+dt;
end
AP
spread=10000*DP/(PL+AP)

30