Peter Jäckela

Christian Kahlb
Lee Wildc
Ioannis Chryssochoosd

A default copula in a lattice-based credit model

First version: 11th September 2009
First published version: 21st June 2013
This version: 22nd October 2013

Abstract

We introduce a default event copula into the framework of a multi-factor credit model with correlated
stochastic hazard rates. We find that the codependence structure of default events in the limit of
infinitesimal time steps in a finite-differencing framework is dominated by the lower tail dependence
coefficient of the employed copula.

1 Introduction
It is a well known fact that mere correlation of hazard rates of underlyings, but otherwise independent
default events, in a multi-dimensional Cox process setting, fails to generate sufficient codependence
between default events to match market prices for first-to-default swaps, or n-th to default products in
general. In order to attain market realistic levels, one needs to consider explicit default event correlation.
In the following, we analyze how this can be done in a fully dynamic model by the aid of a generic event
copula in a finite-differencing setting, and what the implications are for the net codependence of defaults
over a multi-step horizon. Our focus is thus mainly on the small time step limit, as is often of importance
in a finite-differencing framework, since small time steps may be necessary for reasons of numerical
stability. Since we shall see that, in the limit of infinitesimal time steps, net codependence over a long
time horizon may significantly differ from the codependence structure applied over a single time step,
we are particularly interested in criteria that enable us to judge whether any one copula can, in principle,
at least to some degree, preserve correlation in the multi-step limit. Whilst our focus is thus mainly
on finite-differencing implementations, we emphasize that the considerations are generically applicable,
both to backward and to forward induction methods, i.e., Monte Carlo simulations. The key comparison
is between short time steps and single, long, time steps. First, however, we introduce the fundamental
modelling assumptions, specify the employed Cox processes and how a copula function can be used to
generate the full default codependence structure over a single time step.
Key words and phrases. Finite-differencing credit model, multi-name codependent default model.
a
OTC Analytics and VTB Capital
b
Commerzbank
c
Swiss Re
d
VTB Capital

1
2 Partial (integro) differential equations
The multi-dimensional Cox process we are going to work with throughout the following is defined by
choosing displaced diffusion Black-Karasinski dynamics [Rub83, BK91] for each individual hazard rate1
 
λ̂i (t) − βi2 ·V[yi (t)]+βi ·yi (t) 1
λi (t) = e 2 + 1− λ̂i (t) , (2.1)
βi βi

with driving Ornstein-Uhlenbeck processes

dyi (t) = −ai · yi (t) dt + σi (t) dWi (t) , (2.2)

and correlated Brownian motions dhWi (t), Wj (t)i = ρi,j dt. The displacement parameter βi allows us
to capture the implied volatility skew on credit default swaptions, whilst σi (·) is typically calibrated to
match the term-structure of at-the-money volatilities. In order to avoid negative hazard rates, which
in turn would imply negative default probabilities, we typically set β ≥ 1. We further use the time
dependent parameter λ̂i (·) to calibrate the model against credit-default swaps. Whilst this framework
can easily be extended to include other asset classes, we solely focus on credit specific aspects here.
Modelling credit events using a hazard rate approach essentially introduces an instantaneous jump
from a non-defaulted into a defaulted state. Thus, in order to price credit dependent options using the
fundamental theorem of asset pricing

V (t, x) = N (t) · EN V (T ) · N −1 (T ) Ft ,
 
(2.3)

with risk-neutral measure N associated to numeraire N , we have to apply a generalised version of the
theorem of Feynman-Kac [CT04, Øks03], leading to a partial (integro) differential equation,

∂t V + AV + EQ [V (t, y + ∆y) − V (t, y)| y] − rV = 0 , (2.4)

where we choose N to be the money market measure Q induced by the domestic short rate r and A to
denote the generator of the diffusion of the underlying Markovian drivers.
Given this basic modelling framework, we can now use a Monte Carlo simulation approach for nu-
merical implementation. Since many first-to-default baskets and similar products involve only a small
number m of underlyings (m . 5 or so), we can also use finite-differencing techniques for such con-
tracts. The lattice based approach makes it comparatively easy to include early exercise features which
are so commonplace in fixed income derivatives.
The standard implementation of a multi-factor stochastic hazard rate model defaultable underlyings
via partial (integro) differential equations is to use an extension of the state-space to monitor the default
of each underlying. Thus, not counting the stochasticity of the, say, m hazard rates, we are left with a
grid consisting of 2m states and each node ω ∈ {0, 1}m holds information about whether the underlying
is defaulted, with 0 representing default, and 1 representing non-default. To give an example, the node
ω = (0, 1) ∈ {0, 1}2 indicates that underlying #1 is defaulted, and underlying #2 is not in default.
Throughout the following we are going to denote the set of defaulted underlyings by ID and the set of
non defaulted ones by IS , with IS ∩ ID = ∅. Next, in order to discretise the integro differential part in
equation (2.4), we have to specify the instantaneous default probabilites to connect the nodes during the
rollback on the lattice. Typically, one chooses the jump measure according to

ν(∆yi , ∆) = 1{yi =0} · (λi ∆ · δ(∆yi − 1) + (1 − λi ∆) · δ(∆yi )) , (2.5)

1
Note that whilst (displaced) lognormal short rates are known to be ha problem, (displaced)
i lognormal hazard rates are
RT
λ (u) du
perfectly legitimate, since we are not computing quantities of the form E e t i
.

2
where δ refers to the Dirac-delta function, and ∆ is the time step. This effectively means that from
a non-defaulted state we default with probability λi ∆, whilst we stay non-defaulted with probability
1 − λi ∆. Thus, the discretised default probabilities, using the Fokker-Planck equivalent of equation (2.4)
and ignoring the generator A as well as the source term, read

P(ωi (T ) = 1) = (1 − λi ∆) · P(ωi (T − ∆) = 1) , (2.6)

P(ωi (T ) = 0) = λi ∆ · P(ωi (T − ∆) = 1) + P(ωi (T − ∆) = 0) , (2.7)

where λi is the hazard rate on this given node for underlying #i. This coupling implicitly defines the
probability for multiple defaults happening over the same time step ∆ to be given by the product of
individual default probabilities Y
λi ∆ . (2.8)
i ∈ ID (T )∩IS (T −∆)

Evidently, this means that, to order O(∆), this excludes joint defaults. This corresponds to the well
known result that independent multi-variate Poisson processes, almost surely, have no simultaneous
events.
Another way of viewing expression (2.8) is to say that, conditional on a specific hazard rate realiza-
tion, defaults over the next discrete time step ∆, are independent. For a practical credit model, however,
it would be desirable to allow for a certain degree of codependence of defaults over any one time step,
especially if the time step discretisation is coarse (e.g., ∆ ≈ 1/12) for reasons of numerical performance.
For this purpose, we extend the concept using a copula to link the nodes over a given time step. Using
the set of defaulted and non-defaulted underlyings IS and ID and introducing the following notation

P(ID , IS ) = P ({ωi = 0, i ∈ ID , ωj = 1, j ∈ IS , ∀i, j}) (2.9)

we link the defaults via

P(ID , ∅) = C(u(ID )) (2.10)
where C is a Copula and the components of the vector u ∈ [0, 1]m are given by
(
λi ∆ if i ∈ ID
ui = (2.11)
1 else .

We refer the reader to [Jäc02, ELM01, Nel98, CLV04] for more details about copulae and their applica-
tions in finance.
In figure 1, we illustrate the coupling induced by the use of the default copula for the three-dimensional
case. Note that, if node ω has k non-defaulted states, it is connected to 2k possible default states. Since
the copula (2.10) gives us the cumulative probability of default, we have to work out the default densities
via the recursive relation

P (ID , IS ) = P (ID ∪ {j}, IS ) + P (ID , IS ∪ {j}) , (2.12)

with {j} ∈
/ ID ∪ IS . We also know the event probability of all underlyings defaulting

P (I, ∅) = C(λ1 ∆, . . . , λm ∆) . (2.13)

Using the recursive relation (2.12) as well as the initial condition (2.13), one can further show that
X
P(ID , IS ) = (−1)(|ID |−|ID0 |) · P(ID0 , ∅) (2.14)
IS 0 ⊂IS
ID0 =ID ∪IS 0

3
 C(1,q,r) C(p,q,r)
node(1,0,0) node(0,0,0)

node(0,1,0)

C(1,1,r) C(p,1,r)
node(1,1,0)

C(1,q,1) C(p,q,1)
node(1,0,1) node(0,0,1)
y

C(1,1,1) C(p,1,1)
node(1,1,1) node(0,1,1)
x
Figure 1: Default transitions. The directions on the arrows indicate the one way relation between defaulted and non-defaulted
states.

where ID ∪ IS = I [GLN+ 01, theorem 2, page 7].

Another important aspect with regard to using a copula in (2.10) is the restriction that

λi ∆ < 1 , (2.15)

which in particular for stochastic hazard rates needs further consideration. Here, we use the approxima-
tion
1
pi := 1 − = λi ∆ + O(∆3 ) , (2.16)
1 + λi ∆(1 + λi ∆)
inspired by a Padé approximation of 1 − e−λi ∆ for the default probability over the time step. It is
straightforward to show that (2.16) is guaranteed to satisfy pi ∈ [0, 1] for all models with positive hazard
rates. Since we allow greater flexibility for the dynamics of the underlying hazard rate in (2.1) using a
displaced diffusion, we have to take extra care in the case of β less than 1, by adding the restriction

λi > 0 . (2.17)

This means we are effectively working with a displaced Black-Karasinski model with a truncating bound-
ary at λi = 0 when βi < 1. Since the Black-Karasinski model is always calibrated numerically, this poses
no extra operational difficulty.

3 Computing the default transition probabilities

We have previously explained that, in principle, any m-dimensional copula can be used to generate the
transition probabilities πω:ω0 between base state ω and transition state ω 0 conditional on a single re-
alization of all m stochastic hazard rates. It is in general conceivable, and possible, that one uses a
different copula for each base state. This would make it possible, for instance, to design a model that

4
shows accelerated default codependence by choosing the transition copula such that its codependence
parameter, whichever way it may be defined, is an increasing function of |ID | (i.e., the number of ele-
ments of the base state vector ω that indicate default). In the following, however, we simplify matters
by describing a procedure that generates all 2m × 2m /2 codependent default transition probabilities from
one copula function that represents the cumulative probability that, if initially all reference entities are
non-defaulted, at the end of the time step some names have defaulted with certainty, and all remaining
names may or may not have defaulted. We will expand on this below.
For a practical implementation, it is helpful to think of the elements of the default state vector ω as
the binary digits of an integer ω. Since the default state is absorbing, transitions between ω and ω 0 are
not possible if any binary digits of ω are in a defaulted state that are non-defaulted in ω 0 . In other words,

πω:ω0 = 0 if (ω ∨ ω 0 ) > ω (3.1)

where (· ∨ ·) represents the bitwise OR operator. Let us now define

cω0 = C(u(ω 0 )) (3.2)

with
(
pi if (ω 0 ∧ 2i ) = 0
ui = (3.3)
1 else ,

using the bitwise AND operator defined as (· ∧ ·), and pi given by (2.16). The function cω0 denotes the
probability that, given initially no default throughout, at the end of the time step, all those names have
defaulted whose associated bit in the integer ω 0 is in state 0, irrespective of what may have happened
to all the remaining names, i.e., they may have defaulted [(ω 0 ∧ 2i ) = 0] or not [(ω 0 ∧ 2i ) 6= 0]. In
other words, cω0 represents the cumulative probability over both defaulted and non-defaulted states for
all names #i for which (ω 0 ∧ 2i ) 6= 0 and certainty of default for all others. With this in mind, and with
the definition ω0 = (1, . . . , 1) ∈ {0, 1}m we can expand the recursive relationship (2.12) into the linear
rule
00 <ω 0
ωX
πω0 :ω0 = cω0 − πω0 :ω00 · 1{(ω0 ∨ω00 )≤ω0 } . (3.4)
ω 00 =0

This establishes the transition probabilities out of state ω0 .

For the remaining transition probabilities out of all other states ω 6= ω0 , as mentioned above, we
reuse the same copula information that was already computed for πω0 :ω0 . We condition on the base state
and, taking into account (3.1), compute
00 <ω 0
ωX
!,
πω:ω0 = 1{(ω∨ω0 )≤ω} · cω0 − πω0 :ω00 · 1{(ω0 ∨ω00 )≤ω0 } cω . (3.5)
ω 00 =0

This simplifies to

πω:ω0 = 1{(ω∨ω0 )≤ω} · πω0 :ω0 /cω . (3.6)

With this procedure, we can populate the full set of ∼ 2m × 2m /2 transition probabilities by evaluating
the copula function 2m − 1 times (cω0 is always 1), and of the order of 22m further multiplications and
additions. Since a copula function may involve exponential or power functions, avoiding its evaluation
for each element of the matrix can provide a significant computational saving.
As an example, consider the case m = 3. In this situation, the coefficients cω are

5
 ω 7 = 1112 6 = 1102 5 = 1012 4 = 1002 3 = 0112 2 = 0102 1 = 0012 0 = 0002

cω C(1, 1, 1) = 1 C(p1 , 1, 1) C(1, p2 , 1) C(p1 , p2 , 1) C(1, 1, p3 ) C(p1 , 1, p3 ) C(1, p2 , p3 ) C(p1 , p2 , p3 )

and the full matrix πω:ω0 is:

ω0
7 = 1112 6 = 1102 5 = 1012 4 = 1002 3 = 0112 2 = 0102 1 = 0012 0 = 0002
ω
1 − π7:6 − π7:5 − π7:4 c6 − π7:4 c5 − π7:4 c3 − π7:2
7 = 1112 c4 − π7:0 c2 − π7:0 c1 − π7:0 c0
−π7:3 − π7:2 − π7:1 − π7:0 −π7:2 − π7:0 −π7:1 − π7:0 −π7:1 − π7:0

6 = 1102 0 π7:6 /c6 0 π7:4 /c6 0 π7:2 /c6 0 π7:0 /c6

5 = 1012 0 0 π7:5 /c5 π7:4 /c5 0 0 π7:1 /c5 π7:0 /c5

4 = 1002 0 0 0 π7:4 /c4 0 0 0 π7:0 /c4

3 = 0112 0 0 0 0 π7:3 /c3 π7:2 /c3 π7:1 /c3 π7:0 /c3

2 = 0102 0 0 0 0 0 π7:2 /c2 0 π7:0 /c2

1 = 0012 0 0 0 0 0 0 π7:1 /c1 π7:0 /c1

0 = 0002 0 0 0 0 0 0 0 1

4 Short time step limit comparison

A practical aspect of the use of a default copula in a finite-differencing application is whether the code-
pendence behaviour of joint default, or survival, changes structurally when the number of steps that
amount to a total time horizon, say T , is taken to the limit, assuming constant hazard rates. Since the
default state is absorbing, we use as a criterion for the codependence structure the joint probability of
survival over the T -horizon,
P (ω1 (T ) = 1, . . . , ωn (T ) = 1) = P (τ1 > T, . . . , τn > T ) , (4.1)
where τi is the default time of underlying #i. Considering the two-dimensional case, and using equa-
tion (2.14), we have
P (ω1 (T ) = 1, ω2 (T ) = 1) = P∆ (∅, {1, 2}) · P (ω1 (T − ∆) = 1, ω2 (T − ∆) = 1) (4.2)
= P∆ (∅, {1, 2}) n · P (ω1 (0) = 1, ω2 (0) = 1) , (4.3)
where
P∆ (∅, {1, 2}) = 1 − C(λ1 ∆, 1) − C(1, λ2 ∆) + C(λ1 ∆, λ2 ∆) (4.4)
= 1 − λ1 ∆ − λ2 ∆ + C(λ1 ∆, λ2 ∆) (4.5)
and ∆ = T /n for n steps over a horizon from t = 0 to t = T . Note that in this section we set the
short-term default probabilities pi ≈ λi ∆ instead of equation (2.16) since we are interested in the limit
∆ → 0, and in this limit (2.16) becomes indistinguishable from λi ∆.
In the following, we discuss a small set of key copulae, and analyze how the joint probability of
survival is rendered for each one of them as we take the limit n → ∞ which is equivalent to ∆ → 0.
Before we do this, we mention that, due to the fact that the individual marginal default, as well as
survival, probabilities, i.e.,
P({i}, ∅) = 1 − P(∅, {i}) ∀ i, (4.6)
are not affected by the choice of codependence copula, in the two-dimensional case, our understand-
ing of the joint survival probability behaviour immediately gives us knowledge about the joint default
probability behaviour because of
P({1, 2}, ∅) = 1 − P(∅, {1}) − P(∅, {2}) + P(∅, {1, 2}) . (4.7)

6
4.1 Independence Copula
The independence copula is given by

CIndependence (u, v) = u · v . (4.8)

Here, we have
P∆ (∅, {1, 2}) = 1 − λ1 ∆ − λ2 ∆ + λ1 λ2 ∆2 . (4.9)
Using ∆ = T /n, we obtain for the limit

n
T 2
lim P Tn (∅, {1, 2}) n = lim 1 − λ1 Tn − λ2 Tn + λ1 λ2 n
= e−(λ1 +λ2 )T , (4.10)
n→∞ n→∞

which is the same as if we evaluate the same copula over a single time step with the marginal survival
probabilities e−λ1 T and e−λ2 T , i.e.,

P (τ1 > T, τ2 > T ) = 1 − CIndependence 1 − e−λ1 T , 1 − CIndependence 1, 1 − e−λ2 T

+ CIndependence 1 − e−λ1 T , 1 − e−λ2 T

(4.11)
= e−λ1 T · e−λ2 T . (4.12)

We note that the independence copula readily extends to arbitrary dimensions via
m
Y
CIndependence (u) = ui . (4.13)
i=1

4.2 Perfect (co-)dependence

A copula that represents perfect codependence, also known as comonotonicity, can be formulated as

CComonotonic (u) = min(u1 , . . . , um ) . (4.14)

The limit of the multi-step survival probability is

n
lim P Tn (∅, {1, 2}) n = lim 1 − λ1 Tn − λ2 Tn + min(λ1 Tn , λ2 Tn )
n→∞ n→∞
= min(e−λ1 T , e−λ2 T ) . (4.15)

This is the same as if we evaluate the same copula over a single time step with the marginal survival
probabilities e−λ1 T and e−λ2 T , i.e.,

P (τ1 > T, τ2 > T ) = 1 − CComonotonic 1 − e−λ1 T , 1 − CComonotonic 1, 1 − e−λ2 T

+ CComonotonic 1 − e−λ1 T , 1 − e−λ2 T

(4.16)
= min(e−λ1 T , e−λ2 T ) . (4.17)

4.3 Negative dependence

A copula that represents perfect negative codependence, also known as countermonotonicity, can, in two
dimensions, be formulated as
CCountermonotonic (u, v) = (u + v − 1)+ . (4.18)

7
It is worth noting that its generalization to more than two dimensions,
!
X
ui + 1 − m , (4.19)
i +

is not a copula. To prove this, one may compute the probability assigned to the uppermost sub-half-
hypercube [ 1/2, 1]m using (2.14). In three dimensions, we obtain

P([ 1/2, 1]3 ) = (1 + 1 + 1 + 1 − 3)+ (4.20)

− (1 + 1 + 1/2 + 1 − 3)+ − (1 + 1/2 + 1 + 1 − 3)+ − ( 1/2 + 1 + 1 + 1 − 3)+
+ (1 + 1/2 + 1/2 + 1 − 3)+ + ( 1/2 + 1/2 + 1 + 1 − 3)+ + ( 1/2 + 1 + 1/2 + 1 − 3)+
− ( 1/2 + 1/2 + 1/2 + 1 − 3)+
= 1 − 1/2 − 1/2 − 1/2 + 0 + 0 + 0 − 0
= − 1/2 (4.21)

It is straightforward to show that the probability assigned to the uppermost sub-half-hypercube, if com-
puted assuming (4.19) as the copula, turns out to be 1 − m2 and thus is negative for all m > 2, which
invalidates (4.19) as a copula in more than two dimensions.
It is intuitively easy to see that mutually perfect negative dependence is unattainable for more than
two entities. After all, if A is exactly the opposite of B, and B is exactly the opposite of C, in some
sense, then, invariably, A must be the same as C, which means that A and C are not opposite. In
terms of linear correlation, this relates to the well known result that the lower limit for the correlation
1
coefficient in a homogeneous correlation matrix is − m−1 . Whilst mutual antidependence is a tricky
concept, and whilst (4.19) does not comprise a valid copula in more than two dimensions, there are,
however, ways to generalize the concept of a certain degree of antidependence to higher dimensions such
as suggested by Kettler in 2008 [Ket08]. The copula suggested there is, however, difficult to visualize
as a density in more than two dimensions since it involves derivatives of the Dirac function and can,
possibly, only be understood in a measure theoretical sense. Here, we mention another possibility for a
negative dependence copula, namely
m
!m−1
X 1
m−1
Cnegative dependence (u) = ui +1−m . (4.22)
i=1 +

This particular copula concentrates all of the joint probability density

m
! m
∂m X 1
m−1
(m − 1)! Y m−1 1
−1
Cnegative dependence (u) = δ ui +1−m · m
· u i (4.23)
∂u1 · · · ∂um i=1
(m − 1) i=1

on the hypersurface defined by

m
X 1
uim−1 = m − 1 . (4.24)
i=1

Evidently, this means that, conditional on m−2 of the uniform marginal values, one of the two remaining
uniform variates is a convex function of the other. In three dimensions, for instance, we have
√ √ √
u2 = 2 − u1 − u3 . (4.25)

We show the surface on which all the probability density is concentrated for the three-dimensional ver-
sion of (4.22) in figure 2.

8
 1

0.8

0.6

u3
0.4

0.2

00
0.2
0.4
u1 0.6 1
0.8 0.8
u2 0.6
0.4
1 0 0.2

Figure 2: The Dirac density surface defined by copula (4.22) for m = 3. The colour temperature is an indication for the
relative density on the surface, with the highest density in the centre.

Negative dependence is for most credit modelling of no particular importance since, usually, strong
codependence is the more market relevant scenario. We included a discussion of negative dependence for
the sake of completeness, and to show that the multi-underlying credit modelling framework presented
in this article allows for this case, too, unlike conventional common factor copula models. Also, it is not
entirely inconceivable that negative dependence may perhaps be desirable in a model if credit baskets
such as Asda, Lidl, Aldi, and Argos (in the UK) are considered.
The limit of the multi-step survival probability for negative dependence is

n
lim P Tn (∅, {1, 2}) n = lim 1 − λ1 Tn − λ2 Tn + (λ1 Tn + λ2 Tn − 1)+
n→∞ n→∞
−(λ1 +λ2 )T
=e (4.26)

which is the same as for the independence copula. This highlights once more how fragile negative
dependence is, and serves as a warning that, if any negative dependence is to be evaluated in a financial
context as a limiting case, it inevitably has to be done with a view on long time steps.

4.4 Copula bounds

Any copula function must lie between (4.19) and (4.14), i.e.,
!
X
ui + 1 − m ≤ C(u) ≤ min(u1 , . . . , um ) . (4.27)
i +

This was firstPshown by Hoeffding [Hoe40], and independently, later, by Fréchet [Fré51]. It is for this
reason that ( i ui + 1 − m)+ is sometimes referred to as the lower Fréchet-Hoeffding bound, and the
comonotonic copula (4.14) as the upper Fréchet-Hoeffding bound.
Since we have already computed the multi-step limit for the upper and lower bound copulae in
sections 4.2 and 4.3, respectively, we now proceed to some specific parametric examples.

9
4.5 Ali-Mikhail-Haq Copula
The Ali-Mikhail-Haq copula [AMH78] belongs to the class of Archimedian copulae and is generated by
 
1 − θ(1 − u)
φAli-Mikhail-Haq (u) = ln . (4.28)
u

The copula is given as

u·v
CAli-Mikhail-Haq (u, v) = , (4.29)
1 − θ(1 − u)(1 − v)
with θ ∈ [−1, 1]. Interesting special cases for the Ali-Mikhail-Haq copula are the independence copula
for θ = 0 and the limiting case for θ → 1
u·v
CAli-Mikhail-Haq (u, v) = . (4.30)
u+v−u·v
If we consider the multi-step limit of the joint survival transition probabilities, we obtain

lim P Tn (∅, {1, 2}) n = e−(λ1 +λ2 )T , (4.31)

n→∞

for θ 6= 1 and
(λ2 2
1 +λ1 ·λ2 +λ2 )T
−
lim P Tn (∅, {1, 2}) n = e λ1 +λ2
, (4.32)
n→∞

for θ = 1. Thus, in the limit of infinitesimal discretisation, the Ali-Mikhail-Haq copula (4.29) effectively
degenerates to an independence copula (4.8).
This makes it clear that the Ali-Mikhail-Haq copula is only suitable in situations when sizeable time
steps are taken. This may be the case when a significant number of underlyings are involved with a
rather coarse discretisation in the spatial directions of the stochastic hazard rate drivers. In this case,
the high-dimensionality of the numerical calculation means that the fastest possible evaluation of all
involved analytical components is of paramount importance. Under such circumstances, and possibly
only under such circumstances, the Ali-Mikhail-Haq copula may be the copula of choice.

4.6 Clayton Copula

The Clayton Copula [Cla78] is another member of the class of Archimedian copulae and its generator is

φClayton (u) = u−θ − 1 . (4.33)

The copula is given by

− 1
CClayton (u, v) = u−θ + v −θ − 1 + θ (4.34)
for θ ∈ [−1, ∞). We consider three special cases:

lim CClayton (u, v) = CCountermonotonic (u, v) (4.35)

θ→−1

lim CClayton (u, v) = CIndependence (u, v) (4.36)

θ→0
lim CClayton (u, v) = CComonotonic (u, v) . (4.37)
θ→∞

This wide range of possible configurations makes the Clayton copula a good workhorse in many appli-
cations.

10
The limit of the multi-step survival probability for the Clayton copula is

−1/θ e−(λ1 +λ2 )T if θ ≤ 0
n −λ1 T −λ2 T +((λ1 T )−θ +(λ2 T )−θ )
lim P Tn (∅, {1, 2}) = e =
 
−θ −θ −1/θ
n→∞ e − λ 1 +λ 2 − (λ 1 +λ 2 ) T
else .
(4.38)
This compares to the evaluation of the same copula over a single time step with the marginal survival
probabilities e−λ1 T and e−λ2 T , i.e.,

P (τ1 > T, τ2 > T ) = 1 − 1 − e−λ1 T − 1 − e−λ2 T − CClayton 1 − e−λ1 T , 1 − e−λ2 T

(4.39)

−θ
−θ − θ1
= e−λ1 T + e−λ2 T − 1 + 1 − e−λ1 T + 1 − e−λ2 T −1 . (4.40)

We show in figure 3 how equations (4.38) and (4.40) compare for λ1 = 3%, λ2 = 5%, and T = 5. As

0.78
formula (4.38)
formula (4.40)
0.76

0.74

0.72

0.7

0.68

0.66
-2 -1.5 -1 -0.5 0 0.5 1
log10 θ
Figure 3: The multi-step limit (4.38) and the single step result (4.40) of the joint survival probability to T for the Clayton
copula with λ1 = 3%, λ2 = 5%, and T = 5.

we can see, the curves are sufficiently similar that the multi-step use of a Clayton default copula could
be adjusted, by means of calibration of θ, to perform like a single step Clayton copula.

4.6.1 Gaussian correlation coefficient mapping

Since the Clayton copula attains the same co- and counter-monotonic limits as the Gaussian copula, as
well as independence, it may be helpful in practical applications to have a homotopy transformation
from a Gaussian correlation coefficient ρ to the Clayton parameter θ. Intuitively, one may wish to use a
transformation that establishes equivalence in a certain sense of a codependence measure. For this, we
could equate a rank correlation measure for the Gaussian copula on one side, and the Clayton copula on
the other. As for Spearman’s rho, which is known for the Gaussian copula to be (6/π) · arcsin (ρ/2), it
turns out that its calculation for the Clayton copula involves an integral over the hypergeometric function
2 F1 , which doesn’t make for convenient evaluation. Kendall’s tau, however, is

Kendall
2
τGaussian = arcsin ρ (4.41)
π
11
for the Gaussian copula with correlation coefficient ρ, and

Kendall
θ
τClayton = (4.42)
θ+2
for the Clayton copula. It therefore seems natural to equate the respective expressions for Kendall’s τ ,
which is a measure for concordance, and thus to use the Clayton-Gaussian concordance correspondence
mapping
2
θClayton-Gaussian Concordance Correspondence =
ˆ π (4.43)
2 arcsin ρ
−1
when it is desirable to keep the Clayton copula’s co-dependence strength measured on a quasi-Gaussian
scale, i.e., parametrised by an approximately equivalent (in some sense) Gaussian correlation coeffi-
cient ρ. We show an example for this in figure 4. Note that, for the chosen parameters, the continuous

0.78
The multi-step limit, formula (4.38)
The single step case, formula (4.40)
0.76
The independence limit e−(λ1 +λ2 )·T
The comonotonic limit e− max(λ1 +λ2 )·T
0.74

0.72

0.7

0.68

0.66
0 0.2 0.4 0.6 0.8 1
ρ
Figure 4: The joint survival probability to T for the same data as in figure 3 via the (abscissa) transformation (4.43).

time-step limit of the joint survival probability (the red line) only starts to rise from the independence
limit case as of ρ & 0.1.

4.6.2 The multi-dimensional case

In more than one dimension [Kim74], like for any other Archimedean copula, we have for u ∈ Rd
P 
d
C (u) = φ−1 i=1 φ(ui ) , (4.44)

i.e.,
d
!− θ1
X
CClayton (u) = u−θ
i −d+1 . (4.45)
i=1

It is important to note that, when d > 2, the Clayton parameter θ is limited [MN09] to
1
θ≥− (4.46)
d−1

12
which, incidentally, is the same as the lower bound for the Gaussian copula correlation coefficient:
1
ρ≥− . (4.47)
d−1
This unfortunately implies that the lowest value for ρ that can be permitted in (4.43) for the resulting
value for θ to still meet condition (4.46), is given by
 
π
ρMinimum Clayton-Gaussian Concordance Correspondence = sin 2(3−2d) . (4.48)

Formula (4.48) is equal to (4.47) for d = 2 and d = 3, but slightly smaller in absolute value for larger
d, as is shown in figure 5. We emphasize that, whilst the restriction in ρ must be honoured if we want to

Figure 5: Lower bound for the Gaussian copula coefficient ρ as in (4.47) and the Minimum permissible Clayton-Gaussian
Concordance Correspondence ρ given by equation (4.48).

use the Clayton-Gaussian concordance correspondence parametrisation (4.43), the negative correlation
case is in practice rarely of significant importance, and if so, then only for benchmarking purposes, for
which the restriction is of no great consequence.

4.6.3 Homotopy parametrisation

A further point of note in practical applications is that of a smooth transition from the Gaussian copula,
which always comprises a reference case, to any other copula, such as the Clayton copula. For this, it is
helpful to have a parametrisation that in the sense of a homotopy transformation at one end comprises the
Gaussian copula, and at the other end, say, the Clayton copula. Conveniently, for this purpose, copulæ
have the feature that any convex combination of two copulæ also forms a valid copula. We can therefore
simply mix the Gaussian copula with the Clayton copula according to

CHomotopic mix (· · · ) = (1 − η) · CGaussian (· · · ) + η · CClayton (· · · ) (4.49)

for η ∈ [0, 1], which makes for a very convenient parametrisation with the homotopy parameter η repre-
senting the percentage of ab-normality of the blended copula.

13
4.7 Tail dependence
We have seen that the limit of the multi-step survival probability for any copula is given by

n
lim P Tn (∅, {1, 2}) n = lim 1 − λ1 Tn − λ2 Tn + C(λ1 Tn , λ2 Tn ) .
n→∞ n→∞

If we assume for a moment, without loss of generality, that λ1 ≤ λ2 , we find

C(λ1 Tn , λ1 Tn ) T T C(λ2 Tn , λ2 Tn )
λ1 Tn · ≤ C(λ1 , λ2 ) ≤ λ 2
T
· . (4.50)
λ1 Tn n n n
λ2 Tn

This enables us to derive the generic bounds

e−(λ1 +λ2 −min(λ1 ,λ2 )·L )T ≤ lim P Tn (∅, {1, 2}) n ≤ e−(λ1 +λ2 −max(λ1 ,λ2 )·L )T (4.51)
n→∞

with
C(u, u)
L = lim (4.52)
u u→0

being known as the lower tail dependence coefficient of the copula [CLV04]. Inequality (4.51) makes
it clear that only copulae with non-vanishing lower tail dependence will preserve any codependence
between default events in the multi-step limit n → ∞.
As it is known, for instance, that the Gaussian copula for ρ 6= 1 has no lower tail dependence, i.e.,
L = 0, we can immediately state that the multi-step limit of the joint survival probability for the Gaus-
sian copula is equal to that of the independence copula. This underlines that it is in general probably
not a good idea to deploy a Gaussian copula for default codependence in a multi-step finite-differencing
framework, unless it is intended to take sizeable time steps, possibly for reasons of numerical perfor-
mance. In that case, however, one may be better served to use a copula that gives the additional benefit
of being fast to evaluate (which the multi-dimensional Gaussian copula is not) such as the Ali-Mikhail-
Haq copula mentioned in section 4.5.
For completeness, we mention that for any Archimedean copula with generator φ(u), and copula
function C(u, v) = φ−1 (φ(u) + φ(v)), the lower tail dependence coefficient is

φ−1 (2φ(u)) φ−1 (2f )

Archimedean
L = lim = lim −1 . (4.53)
u→0 u f →∞ φ (f )

For the Clayton copula, for example, this gives

(
0 if θ ≤ 0
Clayton
L = 1 (4.54)
2− /θ else ,

which is consistent with (4.38) and (4.51).

For the Student-t copula [ELM01], in comparison, with ν degrees of freedom and homogeneous
correlation matrix where all off-diagonal elements are ρ, we have
Ct √ p p
Lν,ρ = 2tν+1 (− ν + 1 1 − ρ/ 1 + ρ) (4.55)

with tν (·) being the univariate Student-t distribution function with ν degrees of freedom.

14
5 Conclusion
In this article we have explained how a generic copula can be used to control the codependence structure
of default events in a multi-variate Cox process model for default. We presented a method to compute ef-
ficiently the discrete transition probabilities required in a numerical finite-differencing implementation.
We discussed a number of candidate copulae that can be used in the given context, specifically includ-
ing negative dependence, perfect codependence, and two particular Archimedean copulae, namely the
Ali-Mikhail-Haq and the Clayton copula, and compared their respective merits and disadvantages. We
further analyzed the limiting behaviour of the joint survival, and thus default, probabilities in a multi-step
finite differencing implementation. Finally, we established lower and upper bounds for the multi-step
limit of the joint survival, and thus default, probabilities for two reference entities in the given setting
that are determined entirely by the chosen copula’s lower tail dependence coefficient.

References
[AMH78] M.M. Ali, N.N. Mikhail, and M.S. Haq. A class of bivariate distributions including the bivariate
logistic. Journal of Multivariate Analysis, 8:405–412, 1978.
[BK91] F. Black and P. Karasinski. Bond and option pricing when short rates are lognormal. Financial
Analysts Journal, pages 52–59, July/August 1991.
[Cla78] D. Clayton. A Model for Association in Bivariate Life Tables and its Application in Epidemiological
Stuides of Fanilial Tendency in Chronic Disease Incidence. Biometrica, 84:567–577, 1978.
[CLV04] U. Cherubini, E. Luciano, and W. Vecchiato. Copula methods in finance. John Wiley and Sons, 2004.
[CT04] R. Cont and P. Tankov. Financial modelling with jump processes. Chapman & Hall / CRC, 2004.
[ELM01] P. Embrechts, F. Lindskog, and A. McNeil. Modelling Dependence with Copulas and Applications to
Risk Management. Working paper, Department of Mathematics, ETHZ CH-8092, Zürich, Switzer-
land, April 6 2001. www.risklab.ch/ftp/papers/DependenceWithCopulas.pdf.
[Fré51] M. Fréchet. Sur les tableaux de corrélation dont les marges sont données. Annales de l’Université de
Lyon, Sciences Mathematiques et Astronomie, 9:53–77, 1951.
[GLN+ 01] P. Georges, A.G. Lamy, E. Nicolas, G. Quibel, and T. Roncalli. Multivariate Survival Modelling: A
Unified Approach with Copulas. SSRN eLibrary, 2001. ssrn.com/paper=1032559.
[Hoe40] W. Hoeffding. Massstabinvariante Korrelationstheorie. Schriften des Mathematischen Seminars und
des Instituts für Angewandte Mathematik der Universität Berlin, 5:181–233, 1940.
[Jäc02] P. Jäckel. Monte Carlo methods in finance. John Wiley and Sons, February 2002.
[Ket08] P.C. Kettler. Fréchet–Hoeffding lower limit copulas in higher dimensions. Technical report, University
of Oslo, August 2008. folk.uio.no/paulck/FHcop.pdf.
[Kim74] C.H. Kimberling. A probabilistic interpretation of complete monotonicity. Aequationes Math., 10(2–
3):152–164, 1974.
[MN09] A.J. McNeil and J. Nešlehová. Multivariate Archimedean copulas, d-monotone functions and `1-norm
symmetric distributions. The Annals of Statistics, 37(5B):3059–3097, 2009.
[Nel98] R. B. Nelsen. An Introduction to Copulas. Number 139 in Lecture Notes in Statistics. Springer, 1998.
ISBN 0-387-98623-5.
[Øks03] B. Øksendal. Stochastic Differential Equations. Springer, 6th edition, 2003. ISBN 3540047581.
[Rub83] M. Rubinstein. Displaced diffusion option pricing. Journal of Finance, 38:213–217, March 1983.

15