Decision and Risk

Lecture 2: Value-at-Risk Using Copulas

Last updated on January 20, 2025

1
Contents

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Sklar’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 Examples of copulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4.1 Fundamental copulas . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4.2 Implicit copulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 Copula densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.6 Explicit copulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.6.1 Empirical Application . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6.2 Bivariate Copula Simulation . . . . . . . . . . . . . . . . . . . . . . 17
2.6.3 Empirical application . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.1 Introduction
The “era of i.i.d.” is over: and when dependence is taken seriously, copulas naturally
come into play. It remains for the statistical community at large to recognize this
fact. And when every statistics text contains a section or chapter on copulas, the
subject will have come of age.
- Berthold Schweizer

In previous lecture we considered portfolio Value-at-Risk using the assumption that

the joint distribution of log-returns was bivariate normal. In that case the dependence
structure was completely described by the linear correlation coefficient ρ. If the joint dis-
tribution of log-returns is not bivariate normal, then the linear correlation coefficient may
provide an inaccurate measure of the dependence structure, and therefore the Value-at-
Risk estimates may be misleading. For example, a particular portfolio of assets may seem
less risky than it actually is. Hence, alternative methods for describing the dependency
structure should be considered. Copulas offer a popular framework for both defining
multivariate distributions and modelling multivariate data.
Before introducing the basic definitions and properties of copulas, it is worthwhile to
motivate the need for adequate and flexible models to describe the dependence structure

2
2.1. Introduction 3

Figure 2.1: Anscombe’s quartet, correlation ρ = 0.8

Figure 2.2: Correlation ρ = 0

y
2

0
−2 −1 0 1 2
x

between random variables. Understanding and quantifying dependence is at the core of

all modelling efforts in multivariate analysis. Since multivariate normal distribution is
one of the most widely used models, we will start by highlighting the limitations of the
linear correlation coefficient with the help of scatterplots.
First, let’s consider the Anscombe’s quartet in Figure 2.1. In 1973, Frank Anscombe
developed idealized data sets to demonstrate the importance of graphing data before
analyzing it. This figure presents four data sets that have very similar statistical properties
in the sense that the correlation coefficient is approximately 0.8 in all four plots. However,
it can be noticed that the relationships in all four plots are quite different. If we consider
Figure 2.1, the interpretation of a strong correlation is only appropriate in Plot 1, and to
some extent in Plot 3. However, Plot 2 and 4 exhibit a non-linear relationship between
the two variables, and therefore the correlation coefficient cannot adequately describe this
relationship.
Next, let’s consider Figure 2.2 which illustrates another inaccurate belief that zero
correlation implies independence. This figure depicts two variables, X ∼ N (0, 1), and
Y = X 2 . A linear correlation coefficient of zero does not necessarily mean that the two
variables are independent. In this plot it can be seen clearly that there exists a non-linear
U-shaped relationship between two variables X and X 2 despite the linear correlation
coefficient being equal to zero.
Let’s consider a third graphical example with simulated log-returns presented in Figure
2.3. In all these bivariate plots the linear correlation coefficient is the same, ρ = 0.7, and
all marginal distributions are standard normal. However, it can be clearly seen that the
dependence structure is not the same in these plots. For example, in Plot 2 we can observe
that extreme negative log-returns have a tendency to occur together, that is, under this
model it is much more likely to see large joint losses. We can easily verify this by counting
the number of observations in the highlighted green rectangle in all plots (note that the
sample size is the same in all plots). Therefore, this model can be considered to be the
4 Lecture 2. Copula Theory

Figure 2.3: Identical correlation ρ = 0.7, but different dependence structure

most dangerous out of these four models from the point of portfolio diversification and
risk management. However, if we use the linear correlation as a measure of dependence,
then we will not be able to distinguish these differences in the dependence structure.
Another interesting fact that these plots demonstrate is that the joint distribution of two
normally distributed random variables is not necessarily normal (remember that marginal
distributions are standard normals in all these plots). Here only Plot 1 contains draws
from the bivariate normal distribution.
The model presented in Plot 1 is called Gaussian copula 1 . It can be shown that the
tail dependence of the Gaussian copula approaches zero when one goes far enough into
the lower (or upper) tail. This is referred to as ”asymptotic independence”, and means
that regardless of how high the correlation between two assets is, if we just go far enough
into the lower (or upper) tails, extreme events appear to occur independently.
In the context of portfolio risk management, failure to take into account extreme
joint losses could have devastating consequences. Therefore, if the joint distribution of
log-returns is not normal, then using the linear correlation coefficient as a measure of
dependence may lead to misleading conclusions. Therefore, alternative methods for cap-
turing the dependence structure are required. Copulas offer a popular framework for
modelling multivariate data.

1
When the multivariate Gaussian copula is used with normal marginal distributions, the resulting
joint distribution is the multivariate normal
2.1. Introduction 5

Figure 2.4: In the mid-’80s, Wall Street turned to the quants – brainy financial engineers – to invent new
ways to boost profits. Their methods for minting money worked brilliantly...until one of them devastated
the global economy. Source: [Link]

To sum up, the correlation coefficient can adequately describe only a specific type
of dependence - a linear relationship, which indicates how close the points lie around a
straight line. If the dependence structure between two random variables is non-linear, then
the linear correlation coefficient is unable to accurately describe that type of dependence.

Recipe for Disaster: The Formula That Killed Wall Street

In February 2009, journalist Felix Salmon described Gaussian copula as the formula that
“devastated the global economy” and that it “will go down in history as instrumental in
causing the unfathomable losses that brought the world financial system to its knees.”

One of the main limitations of Gaussian copula (model in Plot 1 in Figure 2.3) in the
financial context is that it has the property known as “asymptotic independence” between
random variables (in simpler words, extreme joint negative or positive returns are very
unlikely). This particular property manifested itself during the financial crisis of 2007,
and as a result the Gaussian copula attracted a lot criticism. Models are not perfect, but
that doesn’t mean they’re not useful. The Gaussian copula has been misused and and
subsequently criticised, but that does not mean it is useless. It has certain properties
which are appropriate in certain circumstances.
6 Lecture 2. Copula Theory

2.2 Preliminaries
In this section we consider well-known results that play important role in all statistical
simulations, and also prepare the ground for the definition of copulas.
The first result is so special and important that it has its own name – The Probability
Integral Transform. In short, what this result is saying is that, we can convert any
continuous random variable X to a standard uniform random variable using its own CDF
FX .
The Probability Integral Transform

Let X be a continuous r.v. with CDF FX (x) which is strictly increasing over the
range of X. Then U ≡ FX (X) ∼ U(0, 1).

The simple proof of this result is as follows:

FU (u) = P(U ≤ u)
= P(FX (X) ≤ u)
= P(FX−1 (FX (X)) ≤ FX−1 (u))
= P(X ≤ FX−1 (u))
= FX (FX−1 (u))
=u

where u ∈ [0, 1]
Second result is known as The Inverse Probability Integral Transform, which is simply
the inverse of the previous result. What this result is saying is that, we can transform the
standard uniform random variable U into a random variable X with the desired target
distribution simply by choosing FX−1 . This result is extremely useful for generating random
samples from any desired distributions, using as input only a standard uniform number
generator.
The inverse Probability Integral Transform

Let U ∼ U(0,1), and FX be any CDF. Then, X ≡ FX−1 (U ) ∼ FX

Again, the proof of this result is quite simple:

P(X ≤ x) = P(FX−1 (U ) ≤ x)
= P(FX (FX−1 (U )) ≤ FX (x))
= P(U ≤ FX (x))
= FX (x)
2.3. Sklar’s Theorem 7

where x∈R
Now that we are fully equipped with these two results, we introduce the definition of
a copula. Let’s consider the following definition for the bivariate case, which can be easily
extended to a multivariate case.

Definition 1 Copula

A 2-dimensional copula C : [0, 1]2 → [0, 1] is a joint cumulative distribution

function (CDF) of a 2-dimensional random vector with standard uniform marginals
U (0, 1).

A 2-dimensional copula is a function with the following properties:

1. C(0, u2 ) = 0 and C(u1 , 0) = 0

2. C(1, u2 ) = u2 and C(u1 , 1) = u1

3. C is 2-increasing. That is, for all u1 , u01 , u2 , u02 ∈ [0, 1] such that u1 u01 and
u2 u02 :

C(u01 , u02 ) − C(u01 , u2 ) − C(u1 , u02 ) + C(u1 , u2 )0

From the above we can see that a copula can be characterised both from a probabilistic
and analytic point of view. A probabilistic interpretation is that a two dimensional
copula is a bivariate probability distribution function with standard uniform marginal
distribution functions. That is, we can describe the joint distribution of X1 , X2 using the
marginal distribution functions and the copula C. An analytic definition is that a copula
can be viewed as an analytic function which is a mapping from a unit square to a unit
interval which satisfies all three properties. Therefore, if you can find a function that
satisfies all three properties, then you may call this function a copula.

2.3 Sklar’s Theorem

Next we will consider Sklar’s Theorem, which is central to the theory of copulas and is the
foundation of most applications of copula theory to statistics. Sklar’s Theorem is a very
important result that enables the splitting of a multivariate distribution into marginal
distributions and an associated copula function.
Sklar’s Theorem (in the bivariate case) states that if we take a random vector with
elements X1 and X2 , which has joint distribution function F1,2 (x1 , x2 ) and marginal dis-
tribution functions F1 (x1 ) and F2 (x2 ), then there exists a 2-dimensional copula C for all
values of x. The word “copula” stems from the Latin verb “copulare” meaning to bind,
8 Lecture 2. Copula Theory

to connect, to link. Thus, the copula links the marginal distributions to form the joint
distribution. From a modelling perspective, Sklar’s Theorem allows us to separate the
modelling of the marginal distributions from the dependence structure, which is captured
by a copula function C.

Sklar’s Theorem
If (X1 , X2 )0 has joint distribution function F1,2 (x1 , x2 ) and marginal distribution
functions F1 (x1 ), F2 (x2 ), then there exists an appropriate 2-dimensional copula C
such that, for all x1 , x2 in R̄ = [−∞, ∞]:

F1,2 (x1 , x2 ) = C (F1 (x1 ), F2 (x2 ))

The joint probability density function f1,2 (x1 , x2 ) for F1,2 (x1 , x2 ) is:

f1,2 (x1 , x2 ) = c(F1 (x1 ), F2 (x2 )) · f1 (x1 ) · f2 (x2 )

where c is a 2-dimensional copula density.

2.4 Examples of copulas

There are broadly three widely known classes of copulas: fundamental copulas, implicit
copulas and explicit copulas. First, we will consider the fundamental copulas, which
consist of copulas that represent perfect positive dependence, independence and perfect
negative dependence.

2.4.1 Fundamental copulas

Independence Copula

The independence copula is based on the fact that, if random variables are independent,
then their joint distribution function equals the product of marginal distribution functions,
F (x1 , x2 ) = F1 (x1 ) · F2 (x2 ). Therefore, the independence copula Π(u1 , u1 ) of u1 and u2 is
the product of u1 and u2 , i.e. Π(u1 , u1 ) = C (u1 , u2 ) = u1 · u2 since C (F1 (x1 ) , F2 (x2 )) =
F (x1 , x2 ) = F1 (x1 )·F2 (x2 ). Therefore, X1 , X2 are independent, if and only if, their copula
is Π. The density of the independence copula equals 1, i.e. c(u1 , u1 ) = 1, (u1 , u2 )0 ∈ [0, 1]2 .
The importance of product copula is that it is often used as a benchmark because it
corresponds to independence.
2.4. Examples of copulas 9

Fréchet-Hoeffding bounds

Other members of fundamental copulas are also known as Fréchet-Hoeffding bounds.

Fréchet-Hoeffding bounds correspond to cases of extreme forms of dependency that rep-
resent perfect positive dependence and perfect negative dependence, comonotonicity and
countermonotonicity respectively. The comonotonicity copula M (u1 , u2 ) is the Fréchet-
Hoeffding upper bound, and the countermonotonicity copula W (u1 , u2 ) is Frećhet-Hoeffding
lower bound (for the lower bound this is true only in the two-dimensional case 2 ). That is,
they are distributions of perfectly positively dependent and perfectly negatively dependent
random variables respectively.
According to the Fréchet-Hoeffding bounds, every copula has to lie between these two
functions M (u1 , u2 ) and W (u1 , u2 ), which are the bivariate distribution functions. That
is, for any 2-dimensional copula C : [0, 1]2 → [0, 1], and any (u1 , u2 )0 ∈ [0, 1]2 , the following
inequalities hold:
W (u1 , u2 ) ≤ C (u1 , u2 ) ≤ M (u1 , u2 )

where:
W (u1 , u2 ) = max {u1 + u2 − 1, 0} and
M (u1 , u2 ) = min {u1 , u2 }
The concept of comonotonicity is important in financial risk management. In par-
ticular, the sum of the components X1 + X2 + · · · + Xn will be the riskiest if the joint
probability distribution of the random vector (X1 , X2 , . . . , Xn ) is comonotonic, i.e. the
copula of the random vector (X1 , X2 , . . . , Xn ) is the comonotonic copula.
Figure 2.5 illustrates the scatterplots for fundamental copulas, comonotonicity copula,
independence copula, and countermonotonicity copula. The first plot displays comono-
tonicity copula, which represents perfect positive dependence. It can be observed that the
data points lie on a (45 degree) straight line with a positive slope. In the second plot, we
have a sample of observations from the independence copula. Here we can observations
in the unit square which exhibit no relationship between the variables. The third plot
displays the countermonotonicity copula, which represent perfect negitive dependence. It
can be observed that in this case the data points lie on a perfect straight line with a
negative slope.

2.4.2 Implicit copulas

Next, let’s consider implicit copulas. Implicit copulas are called implicit because they
do not have a simple closed form. Implicit copulas are constructed using well-known
multivariate distributions. The two most known and widely used elliptical copulas are
the Gaussian and the Student’s t copulas, which are the dependence structures of the
2
In higher dimensions greater than 2, the Frećhet-Hoeffding lower bound is no longer a copula, i.e.
it does not satisfy the properties (1) - (3) in Definition 1
10 Lecture 2. Copula Theory

Figure 2.5: Comonotonicity copula, independence copula, and countermonotonicity copula.

Figure 2.6: Gaussian (Normal) Copula

Normal copula, ρ=−0.999 Normal copula, ρ=0

0.8

0.8
u2

u2
0.4

0.4
0.0

0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
u1 u1

Normal copula, ρ=0.9 Normal copula, ρ=0.999

0.8

0.8
u2

u2
0.4

0.4
0.0

0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
u1 u1

multivariate normal and t distributions. One of the prominent characteristics of elliptical

copulas is that they exhibit symmetry in the sense that the dependence is the same in
the lower and upper tails.

Gaussian (Normal) Copula

The Gaussian copula is perhaps the most widely used of all copulas. In the bivariate
case, normal copula has the linear correlation coefficient ρ as its dependence parameter.
Furthermore, the Gaussian copula is neither lower- nor upper-tail dependent3 , i.e. both
lower- and upper-tail dependence parameters are equal to 0. Hence, the Gaussian copula
3
Do not worry about the mathematical definition of tail dependence, it will not be covered in this
course. In simple terms, you can think of upper (or lower) tail dependence as the conditional probability
of one variable taking extremely large (or small) values given that the other variable has taken extremely
large (or small) values.
2.4. Examples of copulas 11

approach cannot model tail dependence. In the context of financial risk management, the
lack of lower tail dependence loosely means that, the probability of extreme loss in one
investment given an extreme loss experienced in another investment, is equal to 0. This
means that a “reckless” use of Gaussian copula may give a false sense of security when
financial assets are in fact tail-dependent. The Gaussian copula can be written as follows:

Φ−1 (u1 ) Φ−1 (u2 )

−(r2 − 2ρrs + s2 )
Z Z  
1
CG (u1 , u2 |ρ) = p exp drds
−∞ −∞ 2π 1 − ρ2 2(1 − ρ2 )

where Φ−1 (·) is the inverse cumulative distribution function of a standard normal, and
ρ ∈ (−1, 1).
Figure 2.6 illustrates four scatterplots of observations drawn from a Gaussian copula
with different values of the correlation coefficient. In the first plot ρ = −0.999, where the
points are accumulated along the main diagonal. The second plot represents the indepen-
dence case where ρ = 0, i.e. uniform values exhibit no relationship. As ρ approaches 1
the cloud of points start to accumulate around a (45 degree) straight line with a positive
slope. Plots 3 and 4 illustrate that as ρ approaches 1, the Gaussian copula approaches the
Fréchet-Hoeffding upper bound, i.e. the comonotonicity copula, which represents perfect
positive dependence.

Student (Student-t) copula

Another member of elliptical copulas is the Student-t copula, which also has the linear
correlation coefficient ρ as a measure of dependence. In contrast to the Gaussian copula,
the Student copula is both lower- and upper-tail dependent, although it imposes symmetry
in both tails. Furthermore, the Student’s t-dependence structure introduces an additional
parameter for the degrees of freedom ν. Increasing the value of ν decreases the tendency
to exhibit extreme co-movements. Hence, the ν parameter controls the heaviness of the
tails. The stronger the linear correlation ρ, and the lower the degrees of freedom ν,
the stronger is the tail dependence. Therefore, in contrast to the Gaussian copula, the
Student’s t-copula allows for joint fat tails and an increased probability of joint extreme
events. The Student’s copula can be written as follows:

t−1
ν (u1 ) t−1
ν (u2 )
− ν+2
r2 − 2ρrs + s2
Z Z 
1 2
Ct (u1 , u2 |ρ, ν) = p 1+ drds
−∞ −∞ 2π 1 − ρ2 ν(1 − ρ2 )

where ν is the degree-of-freedom parameter, t−1 ν (·) is the inverse of the standard
Student-t cumulative distribution function, and ρ ∈ (−1, 1).
Figure 2.7 illustrates four scatterplots of observations drawn from Student copula.
12 Lecture 2. Copula Theory

Figure 2.7: Student (Student-t) copula

Student’s t copula, ρ=−0.8, ν=3 Student’s t copula, ρ=−0.8, ν=30

0.8
0.8
u2

u2
0.4

0.4
0.0

0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
u1 u1

Student’s t copula, ρ=0.8, ν=3 Student’s t copula, ρ=0.8, ν=30

0.8

0.8
u2

u2
0.4

0.4
0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.0 0.2 0.4 0.6 0.8 1.0
u1 u1

Similar to the Gaussian copula, the correlation coefficient controls the strength of the
dependence. Furthermore, Student’s t copula has an additional parameter – degrees of
freedom ν. The lower the degrees of freedom ν, the stronger is the tail dependence. In
the top panel ρ = −0.8 and ν = 3, and at the top right corner ρ = −0.8 and ν = 30.
Note that the t copula has one peculiar feature when ν is low – observations are widely
scattered in the middle of the distribution and in the corners. The same behaviour can
be observed in the lower panel. It is worth noting that, as ν → ∞, the Student’s t copula
approaches the Gaussian copula.

2.5 Copula densities

If the bivariate copula C (u1 , u2 ) has a density c(u1 , u2 ), then it can be obtained by simply
taking the partial derivative with respect to both underlying variables:

∂ d C (u1 , u2 )
c(u1 , u2 ) =
∂u1 , ∂u2

Suppose that the bivariate copula is given in the following form where the bivariate
CDF F1,2 (·) is known:

C(u1 , u2 ) = F1,2 F1−1 (u1 ) , F2−1 (u2 )

Then the bivaraite copula is implicit, and, therefore, the copula density can be obtained
2.6. Explicit copulas 13

by dividing the bivariate (joint) density function f1,2 by the marginal densities f1 and f2 .

f1,2 F1−1 (u1 ) , F2−1 (u2 )

c(u1 , u2 ) =
f1 F1−1 (u1 ) · f2 F2−1 (u2 )

Using this technique, we can calculate the density of the Gaussian and the Student’s
t-copulas. For example, the density of the Gaussian copula can be derived as follows:
 2 2
ρ (x1 + x22 ) − 2ρx1 x2

1
c(u1 , u2 ) = p exp −
1 − ρ2 2(1 − ρ2 )
where x1 = Φ−1 (u1 ), x2 = Φ−1 (u2 ) are the standard normal inverse functions of u1
and u2 respectively, and ρ is the correlation coefficient.

2.6 Explicit copulas

In this section we will consider explicit copulas. They are called explicit because their
distribution functions have (explicit) closed forms, something not possible for the Gaussian
and Student t copulas. The well-known family of explicit copulas is the Archimedean
copula family. Archimedean copulas are popular (especially in finance, insurance, etc.)
and are widely used in applications due to their simple form and nice properties.

Frank copula

First, let’s consider Frank copula, which is a symmetric Archimedean copula, and has the
following form:

(
)
−θu1

−θu
e − 1 e 2
− 1
C(u1 , u2 |θ) = −θ−1 log 1 +
e−θ − 1

where θ ∈ (−∞, ∞). Frank copula has the following probability density function:

∂ 2 C(u1 , u2 )
c(u1 , u2 ) =
∂u1 ∂u2

−2
= θ 1 − e−θ e−θ(u1 +u2 ) 1 − e−θ − 1 − e−θu1 1 − e−θu2



The dependence parameter θ of a Frank copula may assume any real value, θ ∈
(−∞, ∞). As θ approaches −∞, the Frank copula approaches the Fréchet-Hoeffding
lower bound; the value of 0 corresponds to the independence case; and θ approaches
∞, the Frank copula approaches the Fréchet-Hoeffding upper bound. The Frank copula
has some useful properties. First, the Frank copula allows negative dependence amongst
random variables. Second, similar to the Gaussian and Student-t copulas, Frank copula
14 Lecture 2. Copula Theory

Figure 2.8: Frank copula

Frank copula, θ=−20 Frank copula, θ=0

0.8

0.8
u2

u2
0.4
0.4

0.0
0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
u1 u1

Frank copula, θ=10 Frank copula, θ=20

0.8

0.8
u2

u2
0.4

0.4
0.0
0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
u1 u1

exhibits symmetry in both tails. Third, Frank copula is considered to be ”comprehensive”

in the sense that both Fréchet-Hoeffding bounds are included in the range of permissible
dependence, i.e. as θ → −∞ and θ → ∞ the Frank copula approaches the Fréchet-
Hoeffding lower and upper bound respectively. What this means is that, the Frank copula
can be used to model the dependence structure between random variables with strong
positive or negative dependence.
Figure 2.8 illustrates four scatterplots of observations drawn from Frank copula with
different parameter values θ. These simulations illustrate that the strongest dependence
is centered in the middle of the distribution.

Gumbel copula

Next, let’s consider Gumbel copula which is an asymmetric Archimedean copula. The
distribution function of Gumbel copula has the following form:

  1 
C(u1 , u2 |θ) = exp − (− log u1 )θ + (− log u2 )θ θ

where θ ∈ [1, ∞). Gumbel copula has the following density:

∂ 2 C(u1 , u2 )
c(u1 , u1 ) =
∂u1 ∂u2

−2+2/θ
= C(u1 , u2 )(u1 u2 )−1 (− log u1 )θ + (− log u2 )θ (log u1 log u2 )θ−1
n
−1/θ o
× 1 + (θ − 1) (− log u1 )θ + (− log u2 )θ
2.6. Explicit copulas 15

Figure 2.9: Gumbel copula

Gumbel copula, θ=1 Gumbel copula, θ=2

0.8

0.8
u2

u2
0.4

0.4
0.0

0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
u1 u1

Gumbel copula, θ=10 Gumbel copula, θ=20

0.8
0.8
u2

u2
0.4

0.4
0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.0 0.2 0.4 0.6 0.8 1.0
u1 u1

The Gumbel copula is upper-tail dependent, but not lower tail dependent. For this
reason, Gumbel copula is known as an asymmetric Archimedean copula. The dependence
parameter θ is restricted to the interval [1, ∞). Value of 1 corresponds to the independence
case, and as θ approaches ∞, the Gumbel copula approaches the Fréchet-Hoeffding upper
bound. Nevertheless, the Gumbel copula does not attain the Fréchet-Hoeffding lower
bound for any value of θ. Furthermore, the Gumbel copula does not allow negative
dependence, but it does exhibit strong upper tail dependence.
Figure 2.9 illustrates four scatterplots of observations drawn from the Gumbel copula
for different parameter values of θ. For θ = 1 we have the independence case, then as θ
increases and approaches ∞, the cloud of points starts to concentrate around a straight
line with a positive (45 degree) slope. It can also be observed that values are more
concentrated in the upper tail than in the lower tail. Therefore, if outcomes are known to
be strongly dependent at high values but less dependent at low values, then the Gumbel
copula may be a good candidate for modelling the dependence structure.

Clayton copula

Another asymmetric member of the Archimedean family is the Clayton copula, which is
an asymmetric Archimedean copula. The distribution of Clayton copula has the following
form:

− θ1
C(u1 , u2 |θ) = u−θ −θ
1 + u2 − 1
16 Lecture 2. Copula Theory

Figure 2.10: Clayton copula

Clayton copula, θ=0.0001 Clayton copula, θ=4

0.8

0.8
u2

u2
0.4
0.4

0.0
0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
u1 u1

Clayton copula, θ=10 Clayton copula, θ=20

0.8
0.8
u2

u2
0.4
0.4

0.0
0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
u1 u1

where θ ∈ (0, ∞). Clayton copula density:

∂ 2 C(u1 , u2 )
−1/θ−2
c(u1 , u2 ) = = (1 + θ)(u1 u2 )−1−θ u−θ
1 + u−θ
2 − 1
∂u1 ∂u2
In contrast to Gumbel copula, Clayton copula has lower-tail dependence, but no upper
tail dependence. The dependence parameter θ is restricted to (0, ∞) interval. As θ
approaches zero, Clayton copula approaches the independence copula. As θ approaches
infinity, Clayton copula approaches the Fréchet-Hoeffding upper bound, but for no value
of θ does it attain the Fréchet-Hoeffding lower bound. Hence, similar to Gumbel copula,
the Clayton copula cannot account for negative dependence.

Figure 2.10 illustrates four scatterplots of observations drawn from Clayton copula for
different parameter values of θ. When θ aproaches 0, we obtain the independence case;
then as θ increases and approaches ∞, the cloud of points starts to concentrate around
a (45 degree) straight line with a positive slope. It can also be observed that values are
more concentrated in the lower tail than in the upper tail. Clayton copula has been widely
used to study events that exhibits strong lower tail dependence. Because Clayton copula
allows for the occurrence of extreme downside events, it has been successfully applied in
portfolio risk management.
2.6. Explicit copulas 17

2.6.1 Empirical Application

Simulations play an important role in statistics. They are necessary to understand the
properties of an underlying multivariate distribution. Very often one may need to perform
simulations in order to get an idea of the shape of the distribution. In particular, risk
management applications require simulation-based experiments. In finance, you may
require to simulate possible future scenarios in order to evaluate, for example, Value-at-
Risk. If the simulation issue can not be easily solved in higher dimensions for a given
copula, then copula is not useful (tractable) to compute the Value-at-Risk. This is because
in practice, the analysis will very often take place in the higher dimensional space (i.e.
we need to analyse more than two variables).

2.6.2 Bivariate Copula Simulation

Sklar’s theorem allows modelling the dependency structure separately from the marginals.
To simulate bivariate data with a particular copula dependence structure, the following
steps are performed:

1. Simulate uniform random variables {U1 , U2 } with the given copula dependency
structure.

2. Generate sample {X1 , X2 } = F1−1 (U1 ) , F2−1 (U2 ) .



Note that a random variable Xi can have any desired distribution by choosing arbitrary
marginal quantile function FX−1i . It can be noted here that Step 2 is straightforward, since
all we need is to generate random sample {X1 , X2 } by transforming each Ui using the
corresponding marginal inverse distribution function. However, it is Step 1 that might
require some extra work 4 .

Implicit copulas: Bivariate Gaussian copula

It turns out that is relatively simple to simulate {U1 , U2 } from a 2-dimensional Gaus-
sian copula (which is constructed from a multivariate standard normal distribution over
R2 ). First, we simulate random variables {X1 , X2 }, and then use the Probability Inte-
gral Transform (PIT) to obtain random variables {U1 , U2 } with the dependence structure
described by the Gaussian copula:

{U1 , U2 } = {Φ (X1 ) , Φ (X2 )}

where Φ is the cumulative distribution function of a standard normal.

4
Do not worry about the details in Step 1, we will not consider them in this course. Simulations
will be performed using R. However, I should mention that the difficulty arises because of the need to
induce the dependence structure implied by a copula. We cannot simply simulate a pair of values from
the univariate standard uniform distribution.
18 Lecture 2. Copula Theory

Figure 2.11: Weekly log-returns

FTSE100

0.15
0.05
log−returns

−0.05
−0.15

2001 2004 2007 2010 2013 2016 2019

S&P500
0.15
0.05
log−returns

−0.05
−0.15

2001 2004 2007 2010 2013 2016 2019

Archimedean (explicit) copulas

Previous approach can only be implemented when the functional form of the joint dis-
tribution is known. This is main drawback of simulating from implicit copulas because
it requires the knowledge of the joint distribution. In practice, the joint distribution is
usually unknown, and the number of available multivariate distributions is considerably
limited. Therefore, this approach limits the usefulness of the inversion method for appli-
cations where the researcher or practitioner does not know the true joint distribution.
However, we might have some information about the dependence structure, which can
be described using a particular copula. In this situation, we simulate {U1 , U2 } directly
from the copula. It is true that we also do not have the knowledge about the true
copula. However, the number of available explicit copula functions is relatively higher,
and therefore, one is able to draw samples from a variety of multivariate distributions with
different dependence structures without knowing the functional form of the multivariate
distribution.

2.6.3 Empirical application

In this section I present a brief application of the copula theory to the estimation of
Value-at-Risk using Monte Carlo simulation approach. We use weekly log-returns for the
period from 1999 to 2018. Figure 2.11 we have time-series plots of log-returns for the two
stock indices.
The fist step in copula modelling is to find models for marginal distributions. We will
not go into details about marginal modelling in this lecture, but we will come back to this
in Lecture 6 using simulated data.
Once we find the adequate marginal mode, we can apply Probability Integral Trans-
form to our data. According to the Probability Integral Transform result, if the marginal
2.6. Explicit copulas 19

Figure 2.12: Probability integral transform (PIT) histograms.

FTSE100 (PIT) S&P500 (PIT)

1.0

1.0
0.8

0.8
Density

Density
0.6

0.6
0.4

0.4
0.2

0.2
0.0

0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

models we have chosen are adequate, then transforming a random variable by its contin-
uous distribution function always leads to the standard uniform distribution. Hence, the
transformed observations should look like a draw from a standard uniform distribution.
Figure 2.12 presents Probability Integral Transform histograms. These histograms seem
essentially uniform. We can also employ formal procedures, such as Kolmogorov-Smirnov
test, to test for uniformity (more on this in Lecture 6). Without presenting the results, I
carried out the test, and the p-values were quite large for both stock indices, which means
that we do not reject the null hypothesis. In this context, the null hypothesis is that these
transformed observations come from a standard uniform distribution.
The next step is to fit various copula models to the transformed data. Using AIC we
choose the copula that best fits the data (AIC values are not reported here), and out of the
copulas provided in the table, the BB1 is the copula of best fit. Note that BB1 is another
bivariate copula which is a member of the Archimedean family of copulas (we haven’t
considered its functional form in this lecture as this is not essential for our purposes).
Figure 2.13 presents a table that contains the 99% and 95% VaR estimates using
Historical simulation approach, highlighted in dark orange, and Monte Carlo simulation
based approach using several copula models. The reason we used Monte Carlo approach
for computing VaR is because it is not easy to derive parametric formula for VaR based
on copulas. It is not surprising that the Gumbel and Frank copulas provide a smaller
estimates of VaR. Recall that these two copulas lack lower tail dependence. Also note that
different copula models result in different VaR estimates, hence getting the copula model
right is very important from the risk management perspective (e.g. if log-returns are lower
tail dependent, then using the Gaussian, Gumbel or Frank copulas will underestimate the
riskiness of a portfolio, which in this case is reflected in the relatively lower VaR estimates).
20 Lecture 2. Copula Theory

Figure 2.13: Value-at-Risk: Monte Carlo simulation approach