ST 810-006

Statistics and Financial Risk

Copulas
A copula is the joint distribution of random variables
U
1
, U
2
, . . . , U
p
, each of which is marginally uniformly distributed
as U(0, 1).
The term copula is also used for the joint cumulative distribution
function of such a distribution,
C (u
1
, u
2
, . . . , u
p
) = P (U
1
u
1
, U
2
u
2
, . . . , U
p
u
p
) .
1 / 39
ST 810-006
Statistics and Financial Risk
Examples
If U
1
, U
2
, . . . , U
p
are independent,
C (u
1
, u
2
, . . . , u
p
) = u
1
u
2
u
p
.
If they are completely dependent (U
1
= U
2
= = U
p
with
probability 1),
C (u
1
, u
2
, . . . , u
p
) = min (u
1
, u
2
, . . . , u
p
)
2 / 39
ST 810-006
Statistics and Financial Risk
Sklars Theorem
Copulas are important because of Sklars Theorem:
For any random variables X
1
, X
2
, . . . , X
p
with joint c.d.f.
F(x
1
, x
2
, . . . , x
p
) = P (X
1
x
1
, X
2
x
2
, . . . , X
p
x
p
)
and marginal c.d.f.s
F
j
(x) = P (X
j
x) , j = 1, 2, . . . , p,
there exists a copula such that
F (x
1
, x
2
, . . . , x
p
) = C [F
1
(x
1
) , F
2
(x
2
) , . . . , F
p
(x
p
)] .
If each F
j
(x) is continuous, C is unique.
3 / 39
ST 810-006
Statistics and Financial Risk
That is, we can describe the joint distribution of X
1
, X
2
, . . . , X
p
by the marginal distributions F
j
(x) and the copula C.
The copula (Latin: link) links the marginal distributions together
to form the joint distribution.
From a modeling perspective, Sklars Theorem allows us to
separate the modeling of the marginal distributions F
j
(x) from
the dependence structure, which is expressed in C.
4 / 39
ST 810-006
Statistics and Financial Risk
The proof is simple in the case that all F
j
(x) are continuous,
because in this case each has an inverse function F
1
j
() such
that
F
j
_
F
1
j
(u)

= u, for all 0 u 1.
If U
j
= F
j
(X
j
), then U
j
U(0, 1):
P (U
j
u) = P [F
j
(X
j
) u] = P
_
X
j
F
1
j
(u)

= F
j
_
F
1
j
(u)

= u.
Write C for the c.d.f. of this copula; then
F (x
1
, x
2
, . . . , x
p
) = P (X
1
x
1
, X
2
x
2
, . . . , X
p
x
p
)
= P [U
1
F
1
(x
1
) , U
2
F
2
(x
2
) , . . . , U
p
F
p
(x
p
)]
= C [F
1
(x
1
) , F
2
(x
2
) , . . . , F
p
(x
p
)] .
5 / 39
ST 810-006
Statistics and Financial Risk
Copula Density
When F() and C() are dierentiable, the equation
F (x
1
, x
2
, . . . , x
p
) = C [F
1
(x
1
) , F
2
(x
2
) , . . . , F
p
(x
p
)]
for the joint cumulative distribution function implies that the
joint probability density function (pdf) satises
f (x
1
, x
2
, . . . , x
p
)
f
1
(x
1
) f
2
(x
2
) . . . f
p
(x
p
)
= c [F
1
(x
1
) , F
2
(x
2
) , . . . , F
p
(x
p
)]
where c() is the pdf of the copula distribution:
c (u
1
, u
2
, . . . , u
p
) =

p
u
1
u
2
. . . u
p
C (u
1
, u
2
, . . . , u
p
) .
6 / 39
ST 810-006
Statistics and Financial Risk
That is, the copula pdf is the ratio of the joint pdf to what it
would have been under independence.
So we can also interpret the copula as the adjustment that we
need to make to convert the independence pdf into the joint pdf.
7 / 39
ST 810-006
Statistics and Financial Risk
Classes of Copulas
An Archimedean copula has the form
C(u
1
, u
2
, . . . , u
p
) =
_

1
(u
1
) +
1
(u
2
) + +
1
(u
p
)
_
for an appropriate generator ().
The Clayton, Frank, Gumbel, and Joe copulas are Archimedean.
Each has a single parameter that controls the degree of
dependence.
8 / 39
ST 810-006
Statistics and Financial Risk
A copula can be extracted from any joint distribution F():
C(u
1
, u
2
, . . . , u
p
) = F
_
F
1
1
(u
1
), F
1
2
(u
2
), . . . F
1
p
(u
p
)
_
Unlike the Pearson (or linear) correlation, the copula is invariant
under monotone increasing transformations of the random
variables.
If F() is a multivariate normal distribution N
p
(, ), then C()
is a Gaussian copula.
If the location or scale of the distribution is changed, the copula
does not change, so conventionally = 0 and = R, a
correlation matrix.
9 / 39
ST 810-006
Statistics and Financial Risk
Gaussian copula for p = 2, = 0.5
0.5
0.5

1


1

1.5
1.5

2


2


2
.
5


2
.
5


3


4

0.0 0.2 0.4 0.6 0.8 1.0
0
.
0
0
.
2
0
.
4
0
.
6
0
.
8
1
.
0
Gaussian copula density, rho = 0.5
10 / 39
ST 810-006
Statistics and Financial Risk
Gaussian copula for p = 2, = 0.5
u
u
c
o
p
Gaussian copula density, rho = 0.5
11 / 39
ST 810-006
Statistics and Financial Risk
If F() is a multivariate t-distribution t

(, ), then C() is a
t-copula.
Again, conventionally = 0 and = R, a correlation matrix.
But note that when X t

(, ):
if > 1 then
E(X) = ;
if > 2 then
Var(X) =
_

2
_
.
Y =
_
2

X has the standardized multivariate t-distribution.

12 / 39
ST 810-006
Statistics and Financial Risk
t-copula for p = 2, = 5, = 0.5

1

1


2


2

3

3


5


5

0.0 0.2 0.4 0.6 0.8 1.0
0
.
0
0
.
2
0
.
4
0
.
6
0
.
8
1
.
0
t copula density, rho = 0.5 df = 5
13 / 39
ST 810-006
Statistics and Financial Risk
t-copula for p = 2, = 5, = 0.5
u
u
c
o
p
t copula density, rho = 0.5 df = 5
14 / 39
ST 810-006
Statistics and Financial Risk
When R = I, the multivariate normal distribution is that of
independent standard normal variables, and the copula has a
constant density.
But the t-copula still shows dependence.
15 / 39
ST 810-006
Statistics and Financial Risk
t-copula for p = 2, = 5, = 0

0
.
8

0.8
0.8

0
.
8


1


1


1


1

1
.2

1
.2


1
.
2

1
.2

0.0 0.2 0.4 0.6 0.8 1.0
0
.
0
0
.
2
0
.
4
0
.
6
0
.
8
1
.
0
t copula density, rho = 0 df = 5
16 / 39
ST 810-006
Statistics and Financial Risk
t-copula for p = 2, = 5, = 0
u
u
c
o
p
t copula density, rho = 0 df = 5
17 / 39
ST 810-006
Statistics and Financial Risk
Application to Portfolio Risk
Consider a portfolio of p corporate bonds.
Ignore interest rate risk, and write the value of bond i at time t
when its rating is as B
i
(t, ).
If the rating of bond i at time t is
i
(t), the portfolio value is
V(t) =

i
B
i
[t,
i
(t)] .
The distribution of the change in value from t to t + t can be
calculated from the joint conditional distribution of

1
(t + t),
2
(t + t), . . . ,
p
(t + t).
18 / 39
ST 810-006
Statistics and Financial Risk
Historical data give plausible values for the marginal transition
probabilities
P
,
= P [
i
(t + t) =

|
i
(t) = ] .
We can write
i
(t + t) as a function of a uniform variable U
i
:

i
(t + t) =

, where

satises
P [
i
(t + t) <

|
i
(t)] < U
i
P [
i
(t + t)

|
i
(t)] .
That is, we break the interval (0, 1) into subintervals with
endpoints at the cumulative probabilities
P [
i
(t + t)

|
i
(t)], and nd the subinterval containing U
i
.
19 / 39
ST 810-006
Statistics and Financial Risk
To model the joint transition probabilities, we need a joint
distribution for U
1
, U
2
, . . . , U
p
.
Independence is not plausible: if the bonds were all at the same
initial rating, independence implies a multinomial distribution for
the counts in the various ratings at a later time, and historical
data show considerable over-dispersion relative to the
multinomial.
We therefore need a copula; the question is, which one?
20 / 39
ST 810-006
Statistics and Financial Risk
The Gaussian Copula
Modeling and simulation of dependent transitions is usually
described in terms of Gaussian random variables instead of
uniform random variables.
Write Z
i
=
1
(U
i
), so Z
i
N(0, 1), and break the whole real
line (, ) into subintervals at the Gaussian quantiles
corresponding to the cumulative probabilities,

1
{P [
i
(t + t)

|
i
(t)]}.
Then Z
i
falls into one of the new subintervals i U
i
falls into the
corresponding subinterval of (0, 1), and we can determine the
new rating as a function of Z
i
by nding the subinterval
containing Z
i
.
21 / 39
ST 810-006
Statistics and Financial Risk
The question of what copula to use now becomes: how to model
the joint distribution of Z
1
, Z
2
, . . . , Z
p
, given that each of them
is marginally N(0, 1).
The obvious choice is the multivariate normal distribution
N
p
(0, R) with zero mean vector and with dispersion matrix equal
to some correlation matrix R.
This is equivalent to using the Gaussian copula dened by
U
i
= (Z
i
) , i = 1, 2, . . . , p,
where the Zs have this multivariate normal distribution.
22 / 39
ST 810-006
Statistics and Financial Risk
Choice of R
The Gaussian copula for p = 2 has only the single parameter .
When p > 2, the
1
2
p(p 1) o-diagonal correlations must be
specied, constrained by the non-negative deniteness of R.
When the variables correspond to entities with appropriate
structure, a factor model may be used, such as in the Moodys
capital model:
Z
i
=
1

G
+
2

R
i
+
3

I
i
+
4

F,i
where
2
1
+
2
2
+
2
3
+
2
4
= 1, and the s are independent
standard normal variables.
So each Z
i
is N(0, 1), and the correlation of Z
i
and Z
j
depends
on which s they share.
23 / 39
ST 810-006
Statistics and Financial Risk
Specically, for i = j ,
Corr(Z
i
, Z
j
) =
2
1
+
2
2
1
{R
i
=R
j
}
+
2
3
1
{I
i
=I
j
}
,
where the rst term (global) is present for all i and j , and the
second and third are present for entities in the same region R or
industry I , respectively.
The resulting correlation matrix R is necessarily non-negative
denite.
24 / 39
ST 810-006
Statistics and Financial Risk
In the absence of such structure, general parametric models may
be used.
An exchangeable correlation matrix has
Corr(Z
i
, Z
j
) = , i = j .
The resulting correlation matrix R is non-negative denite for
1/p 1.
For 0, it is the same as the Moodys model with

2
=
3
= 0.
An exchangeable correlation matrix may be a reasonable choice
when the dependence of all pairs is expected to be the same.
25 / 39
ST 810-006
Statistics and Financial Risk
Sometimes the variables have a natural ordering, such as by
time, and nearby variables are expected to be more dependent
than widely separated variables.
The simplest correlation structure of this form is that of the rst
order autoregressive model, AR(1):
Corr(Z
i
, Z
j
) =
|i j |
for some [1, 1].
Like the exchangeable correlation matrix, the AR(1) correlation
matrix is characterized by a single parameter, .
26 / 39
ST 810-006
Statistics and Financial Risk
Tail Dependence
In risk management we are concerned about the tail of the
distribution of losses, and often large losses in a portfolio are
caused by simultaneous large moves in several components, so
one interesting aspect of any copula is the probability it gives to
simultaneous extremes in several dimensions.
The lower tail dependence of X
i
and X
j
is dened as

l
= lim
u0
P
_
X
i
F
1
i
(u)|X
j
F
1
j
(u)

.
It depends only on the copula, and equals
lim
u0
1
u
C
i ,j
(u, u).
27 / 39
ST 810-006
Statistics and Financial Risk
Tail dependence is symmetric: the tail dependence of X
i
and X
j
is the same as that of X
j
and X
i
.
Upper tail dependence
u
is dened similarly.
The Gaussian copula has the curious property that its tail
dependence is 0, regardless of the correlation matrix.
The t-copula is similarly derived from the multivariate
t-distribution, and has positive tail dependence even when X
i
and X
j
are uncorrelated:

l
=
u
= 2t
+1
_

( + 1)
_
1
i ,j
1 +
i ,j
_
_
where t
+1
() is the cumulative distribution function of the
univariate t-distribution with + 1 degrees of freedom.
28 / 39
ST 810-006
Statistics and Financial Risk
Table : Tail dependence in t-distributions

-0.5 0 0.5 0.9 1

= 2 0.058 0.182 0.391 0.718 1
4 0.012 0.076 0.253 0.630 1
10 0.000 0.007 0.082 0.463 1
0 0 0 0 1
29 / 39
ST 810-006
Statistics and Financial Risk
Tail Correlation
In some cases, the quantity of interest is the correlation of
extreme tail events.
Write
I
i
= 1
{X
i
F
1
i
(u)}
.
For instance, I
i
could be the indicator of default for the i
th
bond
in a portfolio.
Then
E(I
i
I
j
) = P(I
i
= 1 I
j
= 1)
= P
_
X
i
F
1
i
(u) X
j
F
1
j
(u)

= C
i ,j
(u, u).
30 / 39
ST 810-006
Statistics and Financial Risk
Then
Corr(I
i
, I
j
) =
C
i ,j
(u, u) u
2
u(1 u)
.
So the lower tail correlation of X
i
and X
j
,
lim
u0
Corr(I
i
, I
j
) = lim
u0
1
u
C
i ,j
(u, u),
is the same as the lower tail dependence.
Similarly for the upper tail correlation.
31 / 39
ST 810-006
Statistics and Financial Risk
Application to CDO Pricing
In a CDO (Collateralized Debt Obligation), cash ows from a
pool of credit-sensitive instruments are allocated to investors in
various tranches, in order of seniority.
In a cash CDO, the instruments are bonds, and the investors
make initial investments to fund the purchase of the bonds.
32 / 39
ST 810-006
Statistics and Financial Risk
In a synthetic CDO, the instruments are CDSs (Credit Default
Swaps).
A CDS duplicates the cash ows of a loan + bond purchase,
but without money or bonds changing hands.
Since no assets are actually purchased, initial investment is not
required (unfunded synthetic CDO).
However, in most cases, investors in the lower tranches are
required to invest their notional amounts in the vehicle, since
they are at most risk of making payments under the pool of
CDSs (partially funded synthetic CDO).
33 / 39
ST 810-006
Statistics and Financial Risk
In either type of CDO, an investor in a particular tranche faces a
stream of cash ows, say C
j
at times t
j
.
Those cash ows have maximum value when no name in the
pool defaults, but may be reduced if defaults have occurred.
Let
i
be the time of default of name i . Then C
j
depends on
which names have defaulted by time t
j
:
C
j
= C
j
_
1
{
1
t
j
}
, 1
{
2
t
j
}
, . . . , 1
{
N
t
j
}
_
.
34 / 39
ST 810-006
Statistics and Financial Risk
The value of the tranche is therefore
E
_

j
e
rt
j
C
j
_
,
where the expected value is with respect to the risk-neutral
measure, and r is the risk-free interest rate, assumed to be
constant.
The marginal risk-neutral distribution of
i
can be inferred from
the bond market, if this name has issued bonds with the relevant
maturities.
To construct the full joint risk-neutral distribution, we therefore
need a copula.
35 / 39
ST 810-006
Statistics and Financial Risk
The market convention is to use the Gaussian copula with the
correlation matrix
R =
_
_
_
_
_
_
_
1 . . .
1 . . .
1 . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. . . 1
_
_
_
_
_
_
_
.
This is known as a single-factor copula, since the underlying Zs
can be constructed as
Z
i
=
G

+
name
i
_
1 ,
where
G
N(0, 1) is common to all Zs,
name
i
N(0, 1) is
specic to name i , and all s on the right hand side are
independent of each other.
36 / 39
ST 810-006
Statistics and Financial Risk
Since the copula involves only a single parameter, , the
resulting risk-neutral distribution can be calibrated to the market
price of a single tranche.
A rst-loss tranche is the rst to take losses, up to a detachment
point, expressed as a percentage of the pool. A correlation
calibrated to a rst-loss tranche is called a base correlation.
The more senior tranches are characterized by lower and upper
detachment (or attachment) points. A correlation calibrated to
a non-rst-loss tranche is called a compound correlation.
37 / 39
ST 810-006
Statistics and Financial Risk
Unfortunately:
Base correlations are found to depend on the detachment point:
no one correlation gives prices that match the market price for
all detachment points. Instead, they almost always increase with
the detachment point.
Compound correlations may not be unique: for dierent values
of , the expected value matches the market price.
Compound correlations may not even exist: for no value
1 < < 1 does the expected value match the market price.
38 / 39
ST 810-006
Statistics and Financial Risk
Current market practice is to view a non-rst-loss tranche as the
combination of a long position in the rst-loss tranche for the
upper detachment point with a short position in the rst-loss
tranche for the lower detachment (or attachment) point.
Its value is then the dierence of the values for the two rst-loss
tranches.
That is, two dierent risk-neutral distributions are used to nd
the value of a single tranche!
39 / 39