Correlation Smile Matching with -Stable Distributions and

Fitted Archimedan Copula Models

Dirk Prange

and Wolfgang Scherer

Dresdner Kleinwort Wasserstein

Risk Methodology Trading
Model Validation Credit Derivatives
14 March 2006
Abstract
As an extension of the standard Gaussian copula model we present a generalization
based on stable distributions. For special parameter values these distributions coincide
with Gaussian or Cauchy distributions, but changing the parameters allows a continuous
deformation away from the Gaussian copula to others which provide fatter tails. All these
factor copulas are embedded into a framework of stochastic correlations.
We furthermore generalize the linear dependency in the usual factor approach to a
more general copula dependency between the individual trigger variable and the common
latent factor.
Our analysis is carried out on a non homogeneous correlation structure of the under-
lying portfolio. Market premia, even through the correlation crisis, can be reproduced by
certain models. From a numerical perspective all these models are simple since calcula-
tions can be reduced to one dimensional numerical integrals.
1 Introduction
Over the recent years, dependency modeling for basket type credit derivatives has evolved
along with the increase in trading volume of such instruments.
After a rst approach via binomial expansion techniques and other modeling attempts
the copula based models have become more widely discussed and used in the last two years.
Perhaps the historical starting point was the by now standard Gaussian copula of Li [1]. Since
then, various modications and extension thereof have been proposed. In order to replicate
the market observed correlation smile Andersen and Sidenius [2] investigated extending the
Gaussian one factor copula model to include random recovery and/or making the factor
loading state dependent as well, thus introducing what has become known as local correlation,
while still keeping the Gaussian distribution of the latent and idiosynchratic random factors.
While both random recovery and local correaltion does induce some correlation smile this
does not appear to be suciently pronounced to replicate the full market. In a similar spirit

Email: Dirk.Prange@drkw.com

Email: Wolfgang.Scherer@drkw.com
1
Burtschell et al. [3] have extended the Gaussian copula with state dependent correlation to
produce a market skew.
Rather than generalizing the factor loadings Kalemanova et al. [4] replace the Gaussian
copula by the copula generated from normal inverse Gaussian distributed factors. Their,
analysis, however, was done using the large homogenous portfolio approximation, rather than
being based on the true individual characteristics of the underlying portfolio. As such their
approach is not suitable to provide a basis for consistently pricing sub-portfolios or delivering
sensitivities w.r.t. individual names. A recent comparison of dierent copula models such as
Student, Double Student, Clayton, and Marshall-Olkin copulas is given in [5].
An approach not based on factor copulas, but on generalized Archimedean copulas has
been presented by Rogge and Schonbucher [6], but while this is an interesting idea, which
can be generalized even further, it is not clear at present which, if any, of the generalizations
actually reproduces the observed market tranche prices. Recent work by Hull and White [7]
also takes the direction of generalizing the underlying copula by implying out the distribution
of the hazard rate path from observed market prices.
In this paper, we present several approaches based on a combination and extension of
various of the ideas mentioned above and investigate their suitability to reproduce the tranche
market for most of 2005, with particular emphasis on the period of the correlation crisis in
May:
We rst consider a class of one-factor models based on -stable distributions for the
factors. This gives rise to -stable copulas, which include the Gaussian copula as a
special case, but allow for continuous deformation away from the Gaussian case to
distributions with increasingly fatter tails. The parameters of the -stable distribution
can be determined by calibration to the tranche market and provide a good t, in
particular for mezzanine to super senior tranches.
Calibration to market can be further enhanced, if these copulas are combined with the
above copula mixing via state dependent correlations as outlined in [3]. This results
in a near perfect t under normal market conditions and a very good t for markets
during the correlation crisis.
In what can be viewed as a further generalization of local correlations, we generalize the
factor dependence from merely linear coupling to a two-dimensional coupling between
latent variable and common driver based on a copula mixture, including a number of
Archimedean copulas.
Lastly, we introduce a piecewise linear generator function for an Archimedean copula
such that the interpolation points for that generator are determined from observed
tranche market data. This is somewhat akin to Hull and Whites search for the Perfect
Copula.
The paper is organized as follows: In section 2 we summarize the basic assumptions
underlying one factor models based on conditional independence. We introduce the extension
to stochastic correlation in section 2.2 and the -stable copula in section 2.3. In section 2.4
we generalize the linear dependency of the latent factor to a two dimensional copula which
can be either a mixture of copulas or an Archimedan copula with a generalized piecewise
linear generator. In section 3 we present the calibration procedure and the results for the
2
iTraXX tranche market from 26 April 2005 until 14 June 2005 which arguably comprises the
most dicult market for tranches so far. The paper then ends with a concluding section.
2 Default Dependencies
2.1 Basic Assumptions and Resulting Dependency Structures
We consider a portfolio of u underlyings i U := {1, . . . , u}. The following assumptions form
the basis of all conditional independence loss models:
1. A credit event of the ith underlying taking place on or before t is equivalent to a random
variable X
i
falling below a barrier b
i
(t), i.e.
{
i
t} = {X
i
b
i
(t)} (1)
which has as a consequence
P
i
(t) = P{
i
t} = P{X
i
b
i
(t)} = F
X
i
(b
i
(t)) (2)
i.e.
b
i
(t) = F
1
X
i
(P
i
(t)) (3)
where F
A
(z) := P{A z} denotes the the cumulative distribution function of a random
variable A.
2. The variables X
i
are given by a common risk driver (the latent variable) X and idiosyn-
cratic risks X
i
such that
X
i
= c
i
X +c
i
X
i
i U (4)
where X, X
1
, . . . , X
u
are independent random variables and all X
i
have the same dis-
tribution.
The common risk driver X can be one- or d > 1-dimensional, resulting in a so-called
one- or d-Factor Model. Here we concentrate on the one factor case, but this can easily be
generalised
1
.
Note that it is the joint distribution of the X
i
which gives rise to the joint distribution
of defaults via Eqn. (1) (see also Eqn. (11) below). The introduction of X and the

X
i
and
the linear relation Eqn. (4) is only one (albeit comfortable) way to induce a dependency
structure. In case that second moments of X
i
exist on nds from Eqn. (4) that
Cov(X
i
, X
j
) = c
i
c
j
Var(X) +
ij
c
2
i
Var(X
i
) (5)
such that the values of the c
i
are often interpreted as determining a correlation in the
portfolio. For the -stable distributions considered below, these second (or higher) moments
( i.e. the right hand side of Eqn. (5)) no longer exist, but nevertheless these models give rise
to strong correlation features, which now, however, are not solely determined by the c
i
alone.
Moreover, we can replace the linear relation in Eqn. (4) by something more general, hence,
in Section 2.4 below we will generalize the linear dependency structure encoded in Eqn. (4)
to a dependence structure between X and X
i
originating from a two-dimensional copula.
1
Although it is far from clear, whether more factors will always improve the model.
3
It follows from Eqn. (4) that
P
i
(t|x) := P{
i
t|X = x} = F
X
i
_
F
1
X
i
(P
i
(t)) c
i
x
c
i
_
(6)
and that conditional on X the defaults of the underlyings are independent. Eqn. (6) also
allows to calculate the default density
p
i
(t|x) := P{
i
[t, t +dt]|X = x} =
dP
i
(t|x)
dt
= f
X
i
_
F
1
X
i
(P
i
(t))
c
i
_
p
i
(t)
c
i
f
X
i
(F
1
X
i
(P
i
(t)))
(7)
where f
A
(z) =
d
dz
F
A
(z) denotes the probability distribution function of a random variable A.
The conditional independence can then be used to calculate the full default (and thus
loss) distribution as
P{
1
t
1
, . . . ,
u
t
u
} = E [P{X
1
b
1
(t
1
), . . . , X
u
b
u
(t
u
)|X = x}] = E
_

iU
P
i
(t
i
|X)
_
=
_

iU
P
i
(t
i
|x)dF
X
(x) (8)
and similarly for the default density. This distribution is required to calculate payos which
depend on identity of the defaulted obligor, like NTDs with dierent recoveries.
To obtain the portfolio loss, it is however, simpler to use by now well known recursive
methods as e.g. outlined in [2]. Suppose then, that P{L(t) = Z|X = x} has been obtained
by such a recursive procedure. This yields the full (unconditional) loss distribution at time t
as
P{L(t) = Z} =
_
P{L(t) = Z|X = x}dF
X
(x) (9)
and this distribution can be used to calculate expected payos for derivatives depending on
the losses in a portfolio such as e.g. NTDs with homogeneous recoveries or CDO Portfolio
Tranche Swaps.
Let
F
X
U
(a
1
, . . . , a
u
) := P{X
1
a
1
, . . . , X
u
a
u
)}
denote the joint distribution of the X
U
:= (X
1
, . . . , X
u
) and
C
X
U
(a
1
, . . . , a
u
) := F
X
U
(F
1
X
1
(a
1
), . . . , F
1
X
u
(a
u
)) (10)
their copula. It follows from their respective denitions and from Eqn. (1) and Eqn. (3) that
this is equal to the copula of default times
C

U
(a
1
, . . . , a
u
) := F

U
(F
1

1
(a
1
), . . . , F
1

u
(a
u
))
= P{
1
P
1
1
(a
1
), . . . ,
u
P
1
u
(a
u
)}
= P{X
1
b
1
[P
1
1
(a
1
)], . . . , X
u
b
u
[P
1
u
(a
u
)])}
= P{X
1
F
1
X
1
(a
1
), . . . , X
u
F
1
X
u
(a
u
)}
= C
X
U
(a
1
, . . . , a
u
) (11)
In the rst generation copula model [1] the joint distribution F
U
on the right hand side of
Eqn. (10) was chosen to be a multidimensional Gaussian thus leading in Eqn. (11) to the
Gaussian copula. Since then other choices for F
U
(see e.g. [5]) have been investigated as
well.
4
2.2 Incorporating stochastic correlation
Since the conditional independent Gaussian model does not provide any further parameters,
its ability to t to observed market tranche premia for all tranches is insucient, such that
the observed premia can only be replicated with dierent correlation input for each tranche,
giving rise to what is now referred to as the correlation smile. Burtschell et al. [8] there-
fore considered to take stochastic correlations into account. We adopt their approach here,
allowing the dependence to be a mixture of co-monotonicity, independence and the original
conditional independent model. Therefore we introduce two Bernoulli variables B
a
, B
b
, which
are independent and independent of the factors X
i
, X. Their probabilities are denoted by
q
l
= P{B
l
= 1}, l = a, b, and we set

X
i
= (1 B
a
)(1 B
b
)X
i
+B
b
X + (1 B
b
)B
a
X
i
(12)
= [c
i
(1 B
a
)(1 B
b
) +B
b
]X + [ c
i
(1 B
a
) +B
a
](1 B
b
)

X
i
such that the default event for name i is now triggered by

X
i
falling below a barrier, i.e.
{
i
t} = {

X
i
b
i
(t)}. (13)
The conditional probabilities are given by
P{

X
i
< z|X = x} =
1

k,l=0
P{

X
i
< z|X = x B
a
= k B
b
= l}P{B
a
= k}P{B
b
= l}
= I
{xz}
q
b
+F
X
i
(z)q
a
(1 q
b
) + (1 q
a
)(1 q
b
)F
X
i
_
z c
i
x
c
i
_
(14)
where
I
{A}
:=
_
1, if event A is true
0, else
denotes the indicator function. Integrating out the latent variable X in Eqn. (14) leaves us
with
P{

X
i
z} = q
b
F
X
(z) +q
a
(1 q
b
)F
X
(z) + (1 q
a
)(1 q
b
)F
X
i
(z) (15)
In what follows, we choose the distributions for X, X
i
and thus X
i
such that F
X
= F
X
i
= F
X
i
.
Note that then Eqn. (15) implies F
e
X
i
= F
X
i
= F
X
.
2.3 -Stable Copulas
To obtain a versatile copula structure which includes the Gaussian copula as a special case,
but allows for tail dependency and fatter tails in the portfolio loss distribution, we choose the
latent and idiosynchratic factors from the following -stable distributions (using the notation
of Nolan in [9]) X, X
i
S(, , , ; 1). Except for certain values (see below) these
distributions cannot be expressed with known functions, but it is only their characteristic
function which can be given explicitly as follows

S(,,,;1)
(x) =
_
exp
_

|x|

[1 i(tan

2
) sgn(x)] +ix
_
, if = 1
exp
_
|x|[1 +i
2

sgn(x) log |x|] +ix

_
, if = 1.
(16)
The parameter takes values in ]0, 2] and its decrease leads to fatter tails. The parameter
[1, 1] denes the skewness of the distribution, while R
+
is a scaling and R
5
-6 -4 -2 2 4 6
0.05
0.1
0.15
0.2
0.25
0.3
alpha 1.3, beta 0
alpha 1.7, beta 0
alpha 2.0, beta 0
Figure 1: -stable distributions with dierent values for . The value = 2 is the Gaussian
distribution. Fatter tails for < 2 are clearly visible.
is a location parameter. For our purposes the scale and location parameters and can
be set to some convenient values other scale and location only change the intermediate b
i
functions. The term -stable refers to the fact that the sum of two -stable random variables
is again -stable random variable (stability under summation), albeit with possibly dierent
(skewness, scale, or location) parameters.
Note that the parameter allows us to change the distribution away from the Gaussian
case
GAUSS
= 2 in a smooth fashion to other distributions which have ever increasing fatter
tails as decreases away from
GAUSS
= 2. A value for ]0, 2] may thus be obtained by
calibration. Figures 1 and 2 show the distribution functions for dierent values of and .
Choosing = 1, = 0, the coecients as c
i
, c
i
= (1 c

i
)
1

and
X S(, , 1, 0; 1) (17)
X
i
S(, , 1, 0; 1) (18)
implies
X
i
= c
i
X + (1 c

i
)
1

X
i
S(, , 1, 0; 1) (19)
6
-6 -4 -2 2 4 6
0.05
0.1
0.15
0.2
0.25
alpha 1.3, beta 0.8
alpha 1.7, beta 0.2
alpha 1.3, beta 0.8
Figure 2: -stable distributions with dierent values for and . Changing away from
= 0 introduces skew in the distribution.
is again an -stable distributed random variable with the same , and = 1, = 0.
Denoting the -stable cumulative probability distribution function by F

, the conditional
default probabilities are then given by
P
i
(t|x) = F

_
F
1

(P
i
(t)) c
i
X
(1 c

i
)
1

_
. (20)
In the following we will denote the models tested from this group as follows
STABLE A stable model without stochastic correlation. The free model parameters are
(c
i
, , ).
STABLEmix A stable model with stochastic correlation. The free model parameters are
(q
a
, q
b
, c
i
, , ).
The Gaussian copula is obtained as a special subcase, if in Eqn. (17) and in Eqn. (18)
the values = 2 and = 0 are used The conditional default probabilities are then given by
P
i
(t|x) =
_
_

1
(P
i
(t)) c
i
X
_
1 c
2
i
_1
2
_
_
. (21)
where denotes the standard cumulative normal distribution function.
In order to compare we include the Gaussian case as a separate model denoted as
7
GAUSSmix A Gaussian model with stochastic correlation. The free model parameters are
(q
a
, q
b
, c
i
).
Choosing = 1 and = 0 in Eqn. (17) and in Eqn. (18) gives a special case of the
Cauchy distribution
2
and which we denote by F
C
F
C
(x) =
1

arctan(x) +
1
2
, (22)
The inverse of F
C
is explicitely given by: F
1
C
(x) = tan
_
x
1
2
_
, and the conditional default
probability becomes:
P
i
(t|x) =
1

arctan
_
tan
_
x
1
2
_
c
i
x
1 c
i
_
+
1
2
. (23)
In order to compare we also include the Cauchy case as a separate model denoted as
CAUCHYmix A Cauchy model with stochastic correlation. The free model parameters are
(q
a
, q
b
, c
i
).
2.4 Generalizing Latent Variable Dependence to Copula
In the usual conditional loss setup a linear relation between (asset value) X
i
and the latent
variable X (global factor) as given in Eqn. (4) is assumed. This has already been extended
in the random factor loading approach [2] and the stochastic correlation approach [3]. Here
we generalize this further in that, instead of using a linear coupling in Eqn. (4) a general
dependence of X and X
i
given through a 2-dimensional copula function C : [0, 1]
2
[0, 1]
can be assumed. The auxiliary random variables X
i
are not needed in this case.
Hence, we specify a 2-dimensional copula C, which gives the joint distribution of the
variables
X, X
i
U[0, 1], (24)
such that
F
X
i
(x) = x (25)
F
X
(y) = y (26)
P{X
i
x, X y} = C(x, y) (27)
and the corresponding default barriers b
i
(t) are given by
b
i
(t) = P{X
i
b
i
(t)} = P{
i
t} = P
i
(t). (28)
Like in the previous section 2.1, X
i
and X
j
, conditionally on X, are independent. Their
distribution conditional on X is given by:
P{X
i
< z|X = x} =

x
C(z, x) =:
2
C(z, x) (29)
2
The full parametrized class of Cauchy distributions can be obtained as special cases of -stable distribu-
tions, but is not needed for our copula modeling here.
8
such that
P
i
(t|x) := P{ t|X = x} = P{X
i
< b
i
(t)|X = x} =
2
C(P
i
(t), x). (30)
The joint default time distribution of the
i
is still given by (8) with dF
X
(x) = I
{0<x<1}
dx
being the uniform measure.
In this copula setup we investigated two approaches:
Mixture Copula: Any convex combination of copula functions will again be a copula.
We choose as base functions the FrechetHoeding bounds, the independent copula and
an Archimedean copula C

with lower generator and form the combination:

C(x, y) = w
1
C

(x, y) +w
2
(x, y) +w
3
M(x, y) +w
4
W(x, y), with
4

i=1
w
i
= 1
and where
C

(x, y) =
1
((x) +(y)) is an Archimedean copula with generator ,
(x, y) = xy the independent copula,
M(x, y) = min(x, y) the comonoton copula,
W(x, y) = max(x +y 1, 0) the countermonoton copula.
There is a lot of freedom for the choice of the generator function in this mixture copula.
From historical and from implied data several authors have found a trend for increasing
correlation in bad market environment, which in this setup corresponds to a small value
of X. We therefore look for an associated copula with positive lower tail dependence.
Following the notation of Nelsen [10][table 4.1] we used copulas 1, 16, 19 and 20 of
which the last three gave the most convincing results.
Here we report the gures for the Nelsen copula no. 20 which has generator function

(t) = exp(t

) e, (0, ). For this model we investigate two possibilities:

COPmixHOM A copula mixture model where the Archimedean copula has gener-
ator

(x) = exp(x

) exp(1) and at correlation structure. The free model

parameters are (w
1
, w
2
, w
3
, w
4
, ).
COPmixINH An inhomogenous copula mixture model where the Archimedean copula
has the issuer specic generator

i
(x) = exp(x

i
) exp(1) leading to a non-at
correlation. The free model parameters are thus (w
1
, w
2
, w
3
, w
4
,
i
).
Archimedean copula with nonstrict piecewise linear generator: On a set of
nodes 0 = x
0
< x
1
< < x
n
= 1 we specify y values 1 = y
0
> y
1
> > y
n
= 0, and
the generator is dened to be a piecewise linear convex function with (x
i
) = y
i
. The
points y
i
, i = 1, . . . , n 1 must be chosen to satisfy the convexity of . This model is
denoted as
ARCHIMpl A homogenous mixture copula with piecewise linear generator. The free
model parameters are abscissa points x
i
and corresponding function values y
i
for
the linearly interpolated grid: (x
0
, . . . , x
n
; y
0
, . . . , y
n1
), since y
n
= (x
n
) = (1) =
0 by construction.
9
3 Calibration
3.1 Reducing the number of parameters
For the application of the models to real world prices a choice of parameters and functions
needs to be made. We have chosen to have the parameters
q
a
, q
b
in all mixture models GAUSSmix, CAUCHYmix, STABLEmix
, in all -stable models STABLE, STABLEmix
, w
1
, w
2
, w
3
, w
4
in the copula mixture models COPmixHOM, COPmixINH
y
1
, y
2
, y
3
, y
4
in the generalized Archimedan copula model ARCHIMpl
determined by multi-dimensional calibration.
What is left is the determination of the c
i
for GAUSSmix, CAUCHYmix, STABLEmix,
STABLE, which we consider next below and the
i
for COPmixINH, which we consider
thereafter. Both will add one more parameter to be calibrated.
In many applications of the conditional default models the c
i
from equation (4) are set to
the same number, which will be determined by a calibration procedure ( e.g. base correlation).
We prefer, however, to retain more specic information as to how each issuer is coupled to
the economy, hence we do not set all c
i
equal. Rather than determining the c
i
directly as
input ( e.g. from KMV data), we apply a one parameter approach that allows to retain the
relative issuer specic character, but at the same time is amenable to a calibration procedure.
Hence, we take couplings
i
from a database ( e.g. KMV) which should already express
the dependency of obligor i on a joint global factor. In the spirit of transforming historical
market parameters into implicit ones we then apply a global scaling of these by setting
c
i
=

i
, i = 1, . . . , u and [0, ), s. th. for > 1 it follows that c
i
<
i
and < 1 implies
c
i
>
i
. Moreover, lim

c
i
= 0 and lim
0
c
i
= 1,
i
[0, 1). In such an approach we take
the variability of the couplings into account while keeping their ordering. The boundaries

i
= 0 meaning independence and
i
= 1 corresponding to comonotonicity remain preserved.
The single parameter is then obtained within the multi-dimensional calibration.
To determine
i
for COPmixINH in order to take the inhomogenity of the dependence
into account we chose
i
= log (1
i
) with the
i
obtained from some some database
( e.g. KMV) such that it reects the dependency of obligor i on a common economic wide
factor and with [0, ). Then
i
= 0 means independence and
i
= 1 corresponds to
comonotonicity. Again, the single parameter is then obtained within the multi-dimensional
calibration.
3.2 Calibration procedure
In applying the model to the pricing of CDO tranche swaps we focused on the iTraxx series.
During May 2005 the so called correlation crises lead to high equity tranche spreads while the
lowest mezzanine kept relatively stable. We investigate the capability to calibrate the models
to weekly quotes starting from 26 April until 14 June.
The calibration is performed by rst bootstrapping all 125 individual default probability
curves and using these as inputs for the models. All swaps (single name and tranche) are
quarterly and use Act/360 daycount for the premium leg. Recovery rates for the single name
10
swaps are set to 40% for the single names as well as for the iTraxx loss payments. Parameters
for the various models are found by using a multidimensional minimization routine.
3
3.3 Calibration results
Below we will thus consider the following models for their suitability to match a tranche
market
GAUSSmix A Gaussian model with stochastic correlation. The model parameters to be
calibrated are (q
a
, q
b
, ).
CAUCHYmix A Cauchy model with stochastic correlation. The model parameters to be
calibrated are (q
a
, q
b
, ).
STABLEmix A stable model with stochastic correlation. The model parameters to be
calibrated are (q
a
, q
b
, , , ).
STABLE A stable model without stochastic correlation. The model parameters to be cali-
brated are (, , ).
COPmixHOM A copula mixture model where the Archimedean copula has generator

(x) =
exp(x

) exp(1) and at correlation. Instead of nding the convex coecients w

i
di-
rectly we parametrized them in form of spherical coordinates , , according to
w
1
= (cos
a
)
2
,
w
2
= (sin
a
cos
b
)
2
,
w
3
= (sin
a
sin
b
cos
c
)
2
,
w
4
= (sin
a
sin
b
sin
c
)
2
The model parameters to be calibrated are (
a
,
b
,
c
, ). In the result table given
below the parameters are found in the following columns:
a
= q
a
,
b
= q
b
.
COPmixINH An inhomogenous copula mixture model where the Archimedean copula has
the issuer specic generator

i
(x) = exp(x

i
) exp(1) where
i
=
_
1
1
i
1
_
. The
model parameters to be calibrated are (
a
,
b
,
c
, ) and the w
i
are parametrized as for
COPmixHOM.
ARCHIMpl A homogenous mixture copula with piecewise linear generator. The abscissa
points are chosen to be
x
0
= 0.000, x
1
= 0.003, x
2
= 0.010, x
3
= 0.030, x
4
= 0.050, x
5
= 0.0150, x
6
= 1.000.
Since an overall scaling of the generator function , R
+
, leads to the same
copula we set y
0
= 1. Then the model parameters to be calibrated are the corresponding
function values y
i
for the linearly interpolated grid: (y
1
, . . . , y
5
).
3
We use a Levenberg Marquardt algorithm on weighted square deviations.
In case of convergence, the solver usually nds a minimum in less than 50 steps, which on an average pc
takes between 30s for a Gaussian Model with 2 parameters and 2 min for a stable model with 5 parameters.
11
26-Apr-05 Tranche Parameters
Model 0-3% 3-6% 6-9% 9-12% 12-22% q
a
q
b

GAUSSmix 28.4% 177 97 52 19 0.540 0.010 0.114
CAUCHYmix 27.7% 109 67 45 24 0.003 0.213 1.137
STABLE 27.0% 156 59 33 22 0.000 0.000 0.416 1.768 -0.354
STABLEmix 27.0% 159 60 33 22 0.005 0.002 0.417 1.766 -0.256
COPmixINH 27.7% 177 76 45 22 0.203 0.020 0.758
COPmixHOM 27.7% 175 78 46 21 0.593 0.002 0.223
INDEP 44.6% 75 0 0 0
y
1
y
2
y
3
y
4
y
5
ARCHIMpl 25.8% 158 50 22 15 0.745 0.325 0.229 0.143 0.113
Market 25.9% 156 50 24 14
Table 1: Calibration results for 5 year iTraXX Series 3 for 26 April 2005
03-May-05 Tranche Parameters
Model 0-3% 3-6% 6-9% 9-12% 12-22% q
a
q
b

GAUSSmix 31.6% 194 97 50 17 0.545 0.002 0.123
CAUCHYmix 33.1% 132 79 64 27 1.000 0.300 0.885
STABLE 30.3% 185 58 31 20 0.000 0.000 0.520 1.500 0.631
STABLEmix 30.2% 185 55 30 19 0.004 0.006 0.586 1.313 0.771
COPmixINH 31.0% 192 77 44 20 0.001 0.732 0.689
COPmixHOM 31.0% 189 79 45 20 0.574 0.003 0.207
INDEP 47.3% 92 0 0 0
y
1
y
2
y
3
y
4
y
5
ARCHIMpl 30.0% 181 60 33 17 0.780 0.433 0.303 0.198 0.097
Market 30.0% 183 57 29 17
Table 2: Calibration results for 5 year iTraXX Series 3 for 03 May 2005
INDEP The independence model. Serves as a comparison for pricing all tranches as if the
issuers were independent.
The calibration results for the time period from 26 April 2005 to 14 June 2005 are shown
in in Tables 1 to 8
We included the results from an independent model INDEP in order to show how close the
equity tranches priced to complete independence during the correlation crisis. The calibration
results show that the stable model with stochastic correlation STABLEmix performs best
throughout this time of market turmoil in the spring of 2005. But even the pure stable
model STABLE (without stochastic correlation), which uses only two parameters ( and )
to describe the distribution performs calibrates reasonably well to the market. In particular,
these two models show mezzanine tranche and super senior spreads much closer to the market
than the other models. Given the market turbulence in may 2005, the deviations for the stable
models from the market ( i.e. the calibration error) may well be within the bid/oer spread
at that time, but we were unable to conrm this. During normal markets the stable model
STABLE alone may be sucient to calibrate to the quoted tranche spreads. The piecewise
linear Archimedean copula ARCHIMpl also performed reasonably well in that it could deliver
higher spreads for the upper tranches, but tends to overprice the rst mezzanine tranche.
12
10-May-05 Tranche Parameters
Model 0-3% 3-6% 6-9% 9-12% 12-22% q
a
q
b

GAUSSmix 39.8% 171 82 38 11 0.721 0.001 0.049
CAUCHYmix 39.4% 139 60 36 15 0.000 0.167 1.582
STABLE 38.9% 157 41 24 16 0.000 0.000 0.607 1.682 -0.288
STABLEmix 38.9% 158 41 24 16 0.012 0.030 0.788 1.320 0.629
COPmixINH 39.6% 173 60 38 16 0.748 0.011 1.000
COPmixHOM 39.6% 167 64 39 16 0.924 0.003 0.311
INDEP 50.6% 116 0 0 0
y
1
y
2
y
3
y
4
y
5
ARCHIMpl 38.3% 166 39 19 15 0.590 0.282 0.249 0.188 0.082
Market 39.0% 160 42 26 17
Table 3: Calibration results for 5 year iTraXX Series 3 for 10 May 2005
17-May-05 Tranche Parameters
Model 0-3% 3-6% 6-9% 9-12% 12-22% q
a
q
b

GAUSSmix 47.6% 210 107 73 17 0.233 0.261 10.281
CAUCHYmix 48.0% 178 53 38 27 0.377 0.070 0.795
STABLE 47.7% 170 43 28 24 0.000 0.000 0.752 1.291 -0.161
STABLEmix 47.6% 172 45 29 25 0.011 0.011 0.757 1.292 -0.187
COPmixINH 47.6% 224 89 55 21 1.064 0.453 0.989
COPmixHOM 47.8% 217 85 57 23 1.012 0.095 0.479
INDEP 59.3% 215 1 0 0
y
1
y
2
y
3
y
4
y
5
ARCHIMpl 45.3% 232 61 30 15 0.699 0.322 0.274 0.211 0.090
Market 48.0% 178 54 36 27
Table 4: Calibration results for 5 year iTraXX Series 3 for 17 May 2005
24-May-05 Tranche Parameters
Model 0-3% 3-6% 6-9% 9-12% 12-22% q
a
q
b

GAUSSmix 40.8% 155 101 61 14 0.010 0.254 12.213
CAUCHYmix 40.1% 125 45 37 20 0.000 0.077 1.015
STABLE 39.9% 127 43 30 20 0.000 0.000 0.627 1.412 -0.406
STABLEmix 39.9% 127 43 30 20 0.000 0.000 0.624 1.417 -0.433
COPmixINH 41.0% 166 81 57 20 1.384 0.526 1.000
COPmixHOM 41.2% 168 79 54 21 1.172 0.399 0.452
INDEP 53.4% 142 1 0 0
y
1
y
2
y
3
y
4
y
5
ARCHIMpl 38.3% 158 41 21 16 0.606 0.228 0.196 0.149 0.064
Market 39.8% 124 43 27 19
Table 5: Calibration results for 5 year iTraXX Series 3 for 24 May 2005
13
31-May-05 Tranche Parameters
Model 0-3% 3-6% 6-9% 9-12% 12-22% q
a
q
b

GAUSSmix 35.7% 138 97 54 14 0.698 0.257 8.554
CAUCHYmix 34.3% 111 51 36 17 0.000 0.131 1.102
STABLE 34.2% 110 39 32 18 0.000 0.000 0.541 1.585 -0.946
STABLEmix 34.2% 114 45 33 18 0.011 0.048 0.629 1.483 -0.359
COPmixINH 35.6% 153 69 43 20 0.849 0.368 0.999
COPmixHOM 35.6% 152 68 43 21 0.915 0.015 0.341
INDEP 48.2% 97 0 0 0
y
1
y
2
y
3
y
4
y
5
ARCHIMpl 33.5% 127 43 22 15 0.666 0.268 0.232 0.179 0.068
Market 33.8% 120 42 29 15
Table 6: Calibration results for 5 year iTraXX Series 3 for 31 May 2005
07-Jun-05 Tranche Parameters
Model 0-3% 3-6% 6-9% 9-12% 12-22% q
a
q
b

GAUSSmix 33.4% 132 97 54 15 0.188 0.268 8.474
CAUCHYmix 31.7% 99 47 36 17 0.000 0.117 1.010
STABLE 31.5% 97 39 33 18 0.000 0.000 0.518 1.572 -0.990
STABLEmix 31.5% 101 43 33 19 0.010 0.044 0.619 1.414 -0.119
COPmixINH 33.1% 149 66 42 21 0.744 0.277 0.999
COPmixHOM 33.1% 144 67 43 22 0.898 0.007 0.338
INDEP 46.2% 84 0 0 0
y
1
y
2
y
3
y
4
y
5
ARCHIMpl 30.9% 114 29 18 16 0.672 0.277 0.202 0.137 0.080
Market 31.0% 106 34 24 15
Table 7: Calibration results for 5 year iTraXX Series 3 for 07 June 2005
14-Jun-05 Tranche Parameters
Model 0-3% 3-6% 6-9% 9-12% 12-22% q
a
q
b

GAUSSmix 32.1% 130 98 54 16 0.222 0.274 8.537
CAUCHYmix 30.4% 94 46 36 18 0.000 0.112 0.973
STABLE 30.1% 103 40 31 16 0.000 0.000 0.769 1.127 0.719
STABLEmix 30.1% 98 43 34 20 0.030 0.046 0.566 1.480 -0.460
COPmixINH 31.8% 144 67 43 22 0.741 0.298 1.000
COPmixHOM 31.7% 142 67 43 22 0.878 0.004 0.328
INDEP 45.2% 78 0 0 0
y
1
y
2
y
3
y
4
y
5
ARCHIMpl 29.3% 102 23 17 16 0.660 0.241 0.176 0.118 0.068
Market 29.3% 96 33 20 14
Table 8: Calibration results for 5 year iTraXX Series 3 for 14 June 2005
14
4 Conclusion
We have introduced a two parameter (, ) class of distributions closed under addition, so
called -stable distributions, which include the standard Gaussian distribution as a special
case ( = 2, = 0), but allow a continuous deformation of the latter to distributions with
ever fatter tails as decreases away from 2 and more skew as is changed away from 0.
Using these -stable distributions for the idiosynchratic and latent variables in conditional
independence loss models results in signicantly higher tranche spreads for the upper tranches
than those which can be achieved with the standard Gaussian models. Combining the -stable
distribution with the framework of stochastic correlation further enhances the ability to match
the market yields quite good calibration properties for these models, as evidenced by applying
them to the markets during the correlation crisis of 2005.
Furthermore, we have carried out the search for a perfect Archimedean copula with a
generator which is allowed to be piecewise linear and whose interpolation points are deter-
mined by calibration. The resulting class of models also allow for higher spreads in the upper
tranches and a correlation smile, but at the same time tend to deliver too high spreads for
the rst mezzanine tranches.
Overall we believe that the -stable mixture models with stochastic correlation comprise
a class of models capable of capturing the tranche market.
References
[1] David X. Li. On default correlation: A copula function approach. Technical report, The
RiskMetrics Group, 2000. working paper 99-07.
[2] Leif Andersen and Jakob Sidenius. Extension to the gaussian copula: Random recovery
and random factor loadings. http://www.happycool.com/papers/GaussExtensions6.pdf,
2004.
[3] X. Burtschell, J. Gregory, and J-P. Laurant. Beyond the gaussian copula: Stochastic and
local correlation. http://laurent.jeanpaul.free.fr/stochastic local correlation.pdf, 2005.
[4] A. Kalemanova, B. Schmidt, and R. Werner. The nor-
mal inverse gaussian distribution for synthetic cdo pricing.
http://www.mathnance.ma.tum.de/personen/KalemanovaSchmidWerner.pdf, 2005.
[5] X. Burtschell, J. Gregory, and J-P. Laurant. A comparative analysis of cdo pricing mod-
els. http://www.mathematik.uni-ulm.de/nmath/ss 05/fe/jplaurentcomparative.pdf,
2005.
[6] Philipp J. Schonbucher and Ebbe Rogge. Modeling dynamic portfolio credit risk.
www.nasto.uni-bonn.de/schonbuc/papers/schonbucher rogge dynamiccreditrisk.pdf,
2003.
[7] J. Hull and A. White. The perfect copula.
http://www.rotman.utoronto.ca/hull/DownloadablePublications/PerfectCopula.pdf,
2005.
[8] X. Burtschell, J. Gregory, and J-P. Laurant. Beyond the gaussian copula: Stochastic and
local correlation. http://laurent.jeanpaul.free.fr/stochastic local correlation.pdf, 2005.
15
[9] John P. Nolan. Stable Distributions. in preparation, 2005.
[10] Roger B. Nelsen. An Introduction to Copulas. Number 139 in Lecture Notes in Statistics.
Springer, 1999.
16