Multivariate Dependence Modeling

Using Pair-Copulas

Doris Schirmacher
Ernesto Schirmacher 1

Copyright 2008 by the Society of Actuaries.

All rights reserved by the Society of Actuaries. Permission is granted to make brief excerpts for a published
review. Permission is also granted to make limited numbers of copies of items in this monograph for
personal, internal, classroom or other instructional use, on condition that the foregoing copyright notice is
used so as to give reasonable notice of the Society’s copyright. This consent for free limited copying without
prior consent of the Society does not extend to making copies for general distribution, for advertising or
promotional purposes, for inclusion in new collective works or for resale.

1 Corresponding author: Liberty Mutual Group, 175 Berkeley St., Boston, Mass. 02116.
[email protected].

1
Abstract

In the copula literature there are many bivariate distribution families but very few
higher dimensional ones. Moreover, most of these are difficult to work with. Some of the
bivariate families can be extended to more dimensions but in general the construction of
distribution functions with more than two variables is a difficult problem. We introduce
a construction method that is straightforward to implement and can produce multivari-
ate distribution functions of any dimension. In essence the method takes an arbitrary
multivariate density function and decomposes it into a product of bivariate copulas and
marginal density functions. Each of these bivariate copulas can be from any of the avail-
able families.

We also highlight the power of a graphical display known as a chi-plot to help

us understand the dependence between pairs of variables. One illustration, based on
changes in the exchange rate of three currencies, shows how we can specify the pair-
copulas and estimate their parameters. In another illustration we simulate data that ex-
hibits complex dependencies as would be found, for example, in enterprise risk manage-
ment or dynamic financial analysis.

2
1. Introduction

Actuaries are routinely called upon to analyze complex financial security systems.
The outcomes of these systems depend on many variables that have complex dependen-
cies. Until recently, actuaries have had a rather limited set of tools to analyze, extract and
make use of the information embedded in multivariate distributions. The best known
tools have been the linear correlation coefficient and the scatterplot. Linear correlation or
Pearson correlation is a global measure that attempts to summarize in a single number
the dependence between two variables. We cannot expect the correlation coefficient to
be able to adequately summarize complex dependencies into a single number, and it is
well known [2, 12] that two datasets with very different dependence patterns can have
the same correlation coefficient. To alleviate this situation the scatterplot has been used
very effectively to display the entire dataset. Here one can fully appreciate any patterns,
if they are strong enough. Unfortunately, our eyes would rather see a pattern where none
exists. We will introduce a graphical display designed to alleviate this problem.

In recent years there has been increased attention in the combined management
of all risk sources. It is no longer best practice to understand each risk source in isola-
tion. Today we not only need to consider each risk source but more importantly how all
risk sources relate to each other and their potential synergy to create catastrophic losses
when several factors align properly. Thus we need to understand the joint distribution
of all risk sources. Unfortunately, the number of tractable multivariate distributions of
dimension three or higher, such as the multivariate normal and t-distributions, is rather
limited. Moreover, for these two families the marginal distributions are also normal or
t-distributed, respectively. This restriction has limited their useful application in practical
situations. What is needed is a construction that would allow us to specify the marginal
distributions independently from the dependence structure. This we can do with the the-
ory of copulas. Early contributions to the application of copulas include [13, 15, 19, 27]
and in recent years we have seen more activity closely linked with applications in finance
[7, 26], insurance [5, 9, 25], enterprise risk management [7, 3] and other areas. Readers
new to copula methods should consult [14].

In a recent contribution [26] the authors state:

While a variety of bivariate copulas is available, when more than two vari-
ables are involved the practical choice comes down to normal vs. t-copula.
The normal copula is essentially the t-copula with high degrees of freedom
(df), so the choice is basically what df to use in that copula.

3
 They go on to introduce two new copula families; the IT and MCC. While these
two families are defined for any dimension their dependence properties are somewhat
limited and they are not very tractable even in three dimensions.

In this paper we introduce the work of Aas et al. [1] to construct multivariate
distributions of any dimension. Their method relies in using bivariate copulas and past-
ing them together appropriately to arrive at a distribution function. The construction is
straightforward and easy to implement. In Appendix F we have reproduced the algo-
rithms in [1] that carry out simulation and maximum likelihood estimation.

In Section 2, we start the study of dependence by showing that the marginal dis-
tributions affect our perception of the association between two variables. Therefore, if
we want to understand how two variables are interrelated we must remove the marginal
distributions and look at the rank transformed data [14, p. 349]. We also show that the
widely used Pearson correlation coefficient is a poor global measure of dependence be-
cause it is not invariant under all strictly increasing transformations of the underlying
variables. But the next two best known global measures of association; namely, Kendall’s
τ and Spearman’s ρs , are invariant under all such transformations.

Section 3 introduces some elementary properties of copulas and Sklar’s Theorem.

This theorem is a crucial result. Basically it states that any multivariate distribution can
be specified via two independent components:

1. marginal distribution functions, and

2. a copula function that provides the dependence structure.

We also state how both Kendall’s τ and Spearman’s ρs can be expressed in terms
of the underlying copula.

Section 4 provides all the details on how to decompose a multivariate density func-
tion into pair-copulas. For a given density function there are many possible pair-copula
decompositions. To help us organize them there is a graph construction known as a reg-
ular vine [4]. Regular vines are a rather large class of decompositions and so we will only
work with two subsets known as canonical and D-vines.

In Section 5, we show how to obtain the maximum likelihood parameters for a

canonical or D-vine. We also introduce a powerful graphical display, known as a chi-plot
or χ-plot [10, 11], to help us assess the dependence between two variables.

4
 Finally, Section 6 is devoted to a numerical example based, as in [26], on currency
rate changes and in Section 7 we show some of the flexibility of the pair-copula construc-
tion by simulating a D-vine structure with various parameters. Appendices A–E provided
some basic information on the following one-parameter families of copulas

1. Clayton 3. Galambos 5. Normal.

2. Frank 4. Gumbel

These copulas are just a small sample from all bivariate copulas. For more infor-
mation on these and other copulas refer to [18, 20, 22].

Appendix F reproduces four algorithms taken from [1]. These algorithms perform
simulation and maximum likelihood calculations for canonical and D-vine structures.

2. Understanding Dependence

The word “correlation” has been frequently used (or misused) as an over-arching
term to describe all sorts of dependence between two random variables. We will use the
word only in its technical sense of linear correlation or Pearson’s correlation, denote it by
ρ, and define it as

Definition 1. Let X and Y be two random variables with non-zero finite variances. The
linear correlation coefficient for (X, Y) is

Cov(X, Y)
ρ(X, Y) = p p , (1)
Var(X) Var(Y)

where Cov and Var are the covariance and variance operators, respectively.

This quantity, ρ, is a measure of linear dependence; that is, if Y depends on X

linearly (namely, Y = aX + b with a, b ∈ R and a 6= 0), then the absolute value of ρ(X, Y) is
equal to 1. We also know that linear correlation is invariant only under strictly increasing
linear transformations. In addition, linear correlation is the correct dependence measure
to use for multivariate normal distributions; but for other joint distributions it can give
very misleading impressions.

Suppose we have a set of observations (x1 , y1 ), (x2 , y2 ), . . . , (xn , yn ) from an unknown

bivariate distribution H(x, y). We would like to identify the distribution H that character-
izes their joint behavior. We could start our investigations by looking at the scatterplot
of the data and try to discover some pattern that would point us to the correct choice of
bivariate distribution. This approach has long been used with some successes and some

5
failures. The difficulty of effectively using the scatterplot stems from the fact that this dis-
play not only gives us information about the dependency between X and Y but also about
their marginal distributions. While marginal distributions are vital for other analyses it
distorts the information about their dependence.

For example, Figure 2.1 shows a conventional scatterplot on the left-hand side and
a monotone transformation on the right-hand panel. The Pearson correlation coefficient
on the left-hand side is approximately ρ = 0.07, but on the right-hand side is decidedly
different at about ρ = −0.15.

Figure 2.1
Effect of Monotone Transformations
1.0

7
● ●
●
● ●
●

6
● ●
0.5

●
● ●
●
5 ●
●
0.0

●
●
●
4

●
exp(2Y)

●
−0.5

●
Y

●
●
3

● ●
●
−1.0

●
● ●
2

●
−1.5

●
1

●
●
● ● ●
●
●
● ●
0

−2.0 −1.5 −1.0 −0.5 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5

X exp(X)

The left-hand panel shows 20 points (xi , yi ) sampled at random from a bivariate standard normal dis-
tribution. The points on the right-hand panel are given by the monotone transformation (xi , yi ) 7→
(exp(xi ), exp(2yi )).

Both scatterplots in Figure 2.1 definitely have a different qualitative feel to them. Nonethe-
less, in both displays the underlying bivariate dependence between the two variables has
not changed. Only their marginal distributions are different. This shows that the Pear-
son correlation coefficient is a poor measure of the association between two variables. In
particular, it is not invariant under strictly increasing transformations, and this is a major
objection to its use as a global measure of dependence.

The fact that the dependence between two variables is invariant under increasing
monotone transformations is based on two key results. The first one is a representation
theorem due to Sklar [24, 23] that states that the joint distribution function H(x, y) of any
pair of continuous random variables (X, Y) may be written in the form
 
H(x, y) = C F(x), G(y) , x, y ∈ R (2)

6
where F(x) and G(y) are the marginal distributions of X and Y, and C is a function mapping
[0, 1] × [0, 1] → [0, 1] known as a copula.

The second result says that if the pair (W, Z) is a monotone increasing transform
of the pair (X, Y), then the copula that characterizes the joint behavior of (W, Z) is exactly
the same copula as for the pair (X, Y). That is, copulas are invariant under monotone
increasing transformations [14, p. 348].

Since the copula that characterizes dependence is invariant under strictly monot-
one transformations, then a better global measure of dependence would also be invariant
under such transformations. Both Kendall’s τ and Spearman’s ρs are invariant under
strictly increasing transformations, and, as we will see in the next section, they can be
expressed in terms of the associated copula.

Kendall’s τ measures the amount of concordance present in a bivariate distribu-

tion. Suppose that (X, Y) and (X,
e Y)
e are two pairs of random variables from a joint distrib-
ution function. We say that these pairs are concordant if “large” values of one tend to be
associated with “large” values of the other, and “small” values of one tend to be associ-
ated with “small” values of the other. The pairs are called discordant if “large” goes with
“small” or vice versa. Algebraically we have concordant pairs if (X − X)(Y
e e > 0 and
− Y)
discordant pairs if we reverse the inequality. The formal definition is

Definition 2. Kendall’s τ for the random variables X and Y is defined as

τ(X, Y) = Prob((X − X)(Y

e e > 0) − Prob((X − X)(Y
− Y) e e < 0),
− Y) (3)

where (X,
e Y)
e is an independent copy of (X, Y).

Definition 3. Spearman’s ρs for the random variables X and Y is equal to the linear
correlation coefficient on the variables F1 (X) and F2 (Y) where F1 and F2 are the marginal
distributions of X and Y, respectively.

3. Copulas

Sklar’s Theorem is probably the most important theorem in the study of depen-
dence and as in (2) it allows us to write any joint distribution in terms of a copula function
and the marginal distributions. The theorem is valid not only in the bivariate case (n = 2),
but also in all higher dimensions (n > 2). Thus copula functions play a crucial role in our
understanding of dependence. Since the underlying copula is invariant under increasing
monotone transformations, the study of dependence should be free of marginal effects.

7
 That is, for dependence purposes, we should look at our data based on their ranks [14,
p. 349].

Copulas are fundamental for understanding the dependence between random vari-
ables. With them we can separate the underlying dependence from the marginal distrib-
utions. This decomposition is analogous to the way the multivariate normal distribution
is specified; namely, we need two components:

1. a vector of means and

2. a covariance matrix.

The vector of means provides the location for each of the marginal distributions
and the covariance matrix tells us about the dependence between the variables. Moreover,
these two components are independent of each other.

When we consider more general multivariate distributions we can have a simi-

lar decomposition as above. Basically, any multivariate distribution F(x1 , . . . , xn ) can be
specified by providing two components:

1. the marginal distributions F1 , F2 , . . . , Fn , and

2. a copula C: [0, 1]n → [0, 1] that provides the dependence structure.

For example, suppose that F1 , F2 , . . . , Fn are given marginal distributions. 2 Let the
copula function C(x1 , . . . , xn ) be equal to the product of its arguments; that is,

C(x1 , . . . , xn ) = x1 x2 · · · xn .

Then the resulting distribution F(x1 , . . . , xn ) defined via

F(x1 , . . . , xn ) = C(F1 (x1 ), F2 (x2 ), . . . , Fn (xn ))

gives us a multivariate distribution F where the n random variables are independent of

each other. The copula C(x1 , . . . , xn ) = x1 · · · xn is called the independence copula and it is
usually denoted by the symbol Π;

Π(x1 , . . . , xn ) = x1 x2 · · · xn . (4)

2 There are no restrictions on the marginal distributions besides being continuous and even this can be relaxed.
In particular, we don’t have to choose the margins to be all from the same family of distributions as would
happen if we used a multivariate normal distribution or a multivariate t-distribution.

8
So whenever we are dealing with a multivariate distribution F we should separate the
marginal distributions Fi from the dependence structure that is given by some copula
function C. Otherwise, we risk getting tangled up between whatever association there
might be between the variables and their marginal distributions.

The fact that we can always decompose a multivariate distribution F in its marginal
distributions Fi and a copula function C is known as Sklar’s Theorem [23, p. 83].

Theorem 1. Let F be an n-dimensional distribution function with marginal functions

F1 , F2 , . . . , Fn . Then there exists an n-dimensional copula C such that for all (x1 , . . . , xn ) ∈
Rn ,

F(x1 , . . . , xn ) = C(F1 (x1 ), . . . , Fn (xn )). (5)

If the functions F1 , . . . , Fn are all continuous, then C is unique; otherwise, C is uniquely

determined on Ran F1 × · · · × Ran Fn .

Conversely, if C is an n-copula and F1 , . . . , Fn are distribution functions, then the function

F defined via (5) is an n-dimensional distribution function with margins F1 , . . . , Fn .

Another key result for understanding the dependence among variables is that cop-
ulas are invariant under strictly increasing transformations [23, p. 91].

Theorem 2. For n ≥ 2 let X1 , X2 , . . . , Xn be random variables with continuous distribu-

tion functions F1 , F2 , . . . , Fn , joint distribution function F, and copula C. Let f1 , f2 , . . . , fn
be strictly increasing function from R to R. Then f1 (X1 ), f2 (X2 ), . . . , fn (Xn ) are random
variables with continuous distribution functions and copula C. Thus C is invariant under
strictly increasing transformation of X1 , X2 , . . . , Xn .

The implication of this result is that any property of the joint distribution function
that is invariant under strictly increasing transformations is also a property of their copula
and it is independent of the marginal distributions. Thus the study of dependence among
variables is really about the study of copulas.

We know from Figure 2.1 that the linear correlation coefficient is not invariant un-
der strictly increasing transformations. But as stated informally in the previous section,
both Kendall’s τ and Spearman’s ρs are invariant under such transformations. In fact,
these two measures of dependence can be expressed in terms of the underlying copula as
stated in the next two theorems. For the proofs see [22, p. 127, 135].

9
 Theorem 3. Let X and Y be continuous random variables with copula C. Then Kendall’s
τ is given by
ZZ
τ(X, Y) = 4 C(u, v) dC(u, v) − 1. (6)
[0,1]2

Theorem 4. Let X and Y be continuous random variables with copula C. Then Spear-
man’s ρs is given by
ZZ
ρs (X, Y) = 12 uv dC(u, v) − 3. (7)
[0,1]2

4. The Pair-Copula Construction

The construction of distribution functions with n > 2 dimensions is recognized as

a difficult problem. In [26] the authors say:

The MMC 3 copula densities get increasingly difficult to calculate as the di-
mension increases. For this reason, some alternatives to MLE 4 were explored.
One alternative is to maximize the product of the bivariate likelihood func-
tions, which just requires the bivariate densities.

And later they add:

It should be noted, however, that there are many local maxima for both like-
lihood functions, so we cannot be absolutely sure that these are the global
maxima.

In this section we present the results of Aas et al. [1] and follow their exposition
very closely. The basic idea behind the pair-copula construction is to decompose an arbi-
trary distribution function into simple bivariate building blocks and stitch them together
appropriately. These bivariate blocks are two-dimensional copulas and we have a large
selection to choose from [18, 22].

Before we present the general case let us illustrate the construction for two, three,
and four dimensions. The method is recursive in nature. For the base case in two dimen-
sions we can easily see that the density function f (x1 , x2 ) is given by
 
f (x1 , x2 ) = c12 F1 (x1 ), F2 (x2 ) · f1 (x1 ) · f2 (x2 ). (8)

3 These are multivariate copulas with general dependence introduced in [18, p. 163].
4 Maximum likelihood estimation.

10
This follows immediately by taking partial derivatives with respect to both arguments in
F(x1 , x2 ) = C(F1 (x1 ), F2 (x2 )), where C is the copula associated with F via Sklar’s Theorem.

Before we move on to the next case, note that from (8) we can determine what the
conditional density of X2 given X1 is; that is,

f (x1 , x2 )  
f2|1 (x2 |x1 ) = = c12 F1 (x1 ), F2 (x2 ) · f2 (x2 ). (9)
f1 (x1 )

This formula in its general form

f (xi , x j )  
f j|i (x j |xi ) = = cij Fi (xi ), F j (x j ) · f j (x j ). (10)
fi (xi )

will come in very handy as we move up into higher dimensions.

Next let us build a three-dimensional density function. Any such function can
always be written in the form

f (x1 , x2 , x3 ) = f1 (x1 ) · f2|1 (x2 |x1 ) · f3|1,2 (x3 |x1 , x2 ), (11)

and this factorization is unique up to a relabeling of the variables. Note that the second
term on the right-hand side f2|1 (x2 |x1 ) can be written in terms of a pair-copula and a mar-
ginal distribution using (9). As for the last term f3|1,2 (x3 |x1 , x2 ) we can pick one of the
conditioning variables, say x2 , and use a form similar to (10) to arrive at

f3|12 (x3 |x1 , x2 ) = c13|2 (F1|2 (x1 |x2 ), F3|2 (x3 |x2 )) · f3|2 (x3 |x2 ). (12)

This decomposition involves a pair-copula and the last term can then be further decom-
posed into another pair-copula, using (10) again, and a marginal distribution. This yields,
for a three-dimensional density, the full decomposition

f (x1 , x2 , x3 ) = f1 (x1 ) ·
 
c12 F1 (x1 ), F2 (x2 ) · f2 (x2 ) ·
 
c31|2 (F3|2 (x3 |x2 ), F1|2 (x1 |x2 )) · c23 F2 (x2 ), F3 (x3 ) · f3 (x3 ). (13)

For a four-dimensional density we start with

f (x1 , x2 , x3 , x4 ) = f1 (x1 ) · f2|1 (x2 |x1 ) · f3|1,2 (x3 |x1 , x2 ) · f4|1,2,3 (x4 |x1 , x2 , x3 ) (14)

11
and use (10) repeatedly together with the previous results to rewrite it in terms of six
pair-copulas and the four marginal densities fi (xi ) for i = 1, 2, 3, 4:

f (x1 , x2 , x3 , x4 ) = f1 (x1 ) ·
 
c12 F1 (x1 ), F2 (x2 ) · f2 (x2 ) ·
 
c23|1 (F2|1 (x2 |x1 ), F3|1 (x3 |x1 )) · c13 F1 (x1 ), F3 (x3 ) · f3 (x3 ) ·
 
c34|12 F3|12 (x3 |x1 , x2 ), F4|12 (x4 |x1 , x2 ) ·
 
c24|1 F2|1 (x2 |x1 ), F4|1 (x4 |x1 ) ·
 
c14 F1 (x1 ), F4 (x4 ) · f4 (x4 ) (15)

Notice that in the construction many of the pair-copulas need to be evaluated at a con-
ditional distribution of the form F(x|v) where v denotes a vector of variables. The calcu-
lation of these conditional distributions is also recursive. Let v−j denote the vector v but
excluding the jth component v j . For every j, Joe [17] has shown that
 
∂Cx,v j |v− j F(x|v−j ), F(v j |v−j )
F(x|v) = , (16)
∂F(v j |v− j )

where Cx,v j |v− j is a bivariate copula function. For the special case where v has only one
component we have
 
∂Cxv Fx (x), Fv (v)
F(x|v) = . (17)
∂Fv (v)

As we decompose a joint density function f (x1 , . . . , xn ) into a product of pair-copulas and

the marginal densities f1 , . . . , fn we need to make many choices in the conditioning vari-
ables. This leads to a large number of possible pair-copulas constructions. Bedford and
Cooke [4] have introduced a graphical model, called a regular vine, to organize all possi-
ble decompositions. But regular vine decompositions are very general; therefore, we will
only concentrate on two subsets called D-vines and canonical vines. Both models give
us a specific way of decomposing a density function. These models can be specified as a
nested set of trees. Figure 4.1 shows a canonical vine decomposition for a 4-dimensional
density function and Figure 4.2 shows a D-vine.

Similar constructions are possible for any number of variables. The intuition be-
hind canonical vines is that one variable plays a key role in the dependency structure and
so everyone is linked to it. For a D-vine things are more symmetric.

12
 Figure 4.1
Canonical Vine Representation

1 31
31 21|3
32 14|23
3 2 32 21|3 24|3
34 24|3

4 34
Three trees representing the decomposition of a four-dimensional joint density function. The circled nodes
on the left-most tree represent the four marginal density functions f1 , f2 , f3 , f4 . The remaining nodes on the
other trees are not used in the representation. Each edge corresponds to a pair-copula function.

Figure 4.2
D-Vine Representation

1 2 3 4
12 23 34
13|2 24|3

14|23
Three trees representing the decomposition of a four-dimensional joint density function into pair-copulas
and marginal densities. The circled nodes represent the four marginal density functions f1 , f2 , f3 , f4 . Each
edge is labeled with the pair-copula of the variables that it represents. The edges in level i become nodes for
level i + 1. The edges for tree 1 are labeled as 12, 23 and 34. Tree 2 has edges labeled 13|2 and 24|3. Finally,
tree 3 has one edge labeled 14|23.

In general, a canonical vine or D-vine decomposition of a joint density function

with n variables involves n2 pair-copulas. For the first tree in the representation we have

n − 1 edges. The second tree has n − 1 nodes (corresponding to the edges in the previous
tree) and n − 2 edges. Continuing in this manner we see that the total number of edges
across all trees in the representation is equal to
!
(n − 1)n n
(n − 1) + (n − 2) + · · · + 2 + 1 = = .
2 2

A four-dimensional joint density function requires the specification of six pair-copulas.

For a five-dimensional joint density we have to specify ten pair-copulas and most of these
need to be evaluated at conditional distribution functions. One way of reducing this com-
plexity is by assuming conditional independence. Suppose we have a three-dimensional

13
 problem where variable x1 is linked to both x2 and x3 and so we would use a canoni-
cal vine representation. 5 If we assume that conditional on x1 the variables x2 and x3 are
independent, then the construction simplifies to

f (x1 , x2 , x3 ) = f1 (x1 ) f2 (x2 ) f3 (x3 ) c12 (F1 (x1 ), F2 (x2 )) c13 (F1 (x1 ), F3 (x3 )) (18)

because c23|1 (F2|1 (x2 |x1 ), F3|1 (x3 |x1 )) = 1. Figure 4.3 shows the simulation of 150 points
from the joint distribution in (18) (assuming uniform margins) where the copula c12 is
from the Frank family with parameter 3, the c13 copula comes from the Galambos family
with parameter 2, and c23|1 is the independence copula. 6

Figure 4.3
Simulated Canonical Vine
●
● ●
●●● ● ●● ● ●●
●●●●●
●
●● ● ●●
●● ● ● ●●
● ●●●
● ●●
● ● ●
● ● ● ●●
●● ●
● ● ●● ●● ●● ● ● ● ●
● ●
● ● ● ● ● ●●
● ● ●●
● ● ● ●●●
● ● ●●● ●●● ● ● ● ●●● ● ●●
● ●● ● ● ●
● ● ●● ● ●● ●
● ●● ●● ● ● ● ●
●
● ● ●
●
●●
● ●●● ●
● ●● ●● ● ●●
●
●
●●● ● ● ●
●
●● ●● ● ●
● ●● ●●● ●
● ●
● Z
●● ● ● ●● ● ● ●● ●●● ●
●●●●● ●● ● ● ● ● ● ●●● ●●
● ● ● ●
● ● ●●● ● ● ●●● ●●●● ●
●● ● ● ●● ● ●● ●
●● ● ● ● ● ● ● ● ● ●● ●●
●
● ●
●●● ●● ●● ● ●●● ●●
● ●
●● ● ●● ●
●● ● ●● ● ●
● ●●● ● ● ●
●
●● ● ●● ●
●● ● ●● ●

●● ●●
● ● ●● ●● ● ● ● ●●
● ●● ●●●●●●
● ● ●● ●●●●●●●● ●●
●● ● ● ●
● ●● ● ●● ● ●● ●●●
● ● ●● ● ● ●● ● ●
● ● ● ● ●● ●●
● ●● ●● ● ● ●
● ●
●● ● ● ●● ● ● ● ● ●
●● ● ●●●●
● ● ●
● ● ● ● ● ● ●●● ● ●
●●●●●● ●● ●
● ● ● ●● ●
● ● ●
● ● ● ●
●
● ●
● ● ● ● ● ● ● ●
● ● ● ● Y ●● ● ●●
●●● ● ●● ● ● ● ● ●●● ●
● ● ●● ●● ● ●● ● ●● ● ●
● ●● ●●●
● ● ●
● ● ● ●●
●●● ● ● ● ●
● ● ● ●
● ● ●● ● ● ●●
●
● ● ●● ● ● ● ●● ● ● ● ● ●
●● ●
●● ●
●●●
●●●● ●●● ● ●● ● ●●● ● ● ● ● ●
● ● ● ● ● ● ● ●●● ●● ● ●
● ● ●● ●
● ● ● ●● ● ● ● ● ●
● ● ● ●

● ●
●● ●● ●● ●●● ●●● ●
●●
●
●●
●●
●●
●●
●● ●
● ●●
● ●●● ●● ●
● ●●
●● ●● ● ●● ● ●● ● ●● ●●●●●●
●●
●●●
●●●
● ● ●
●●●
● ● ● ●
● ● ● ● ●● ● ● ●●●●●●
●● ●● ●● ●● ●
●
● ●
● ● ●●● ● ● ●● ● ●●●
● ● ● ●
● ●● ●● ● ●●●●●
X ●
●
●
● ● ●
● ●
● ● ●
●●●●
●
●● ●
●
●
●● ●
● ● ● ●● ●●●● ● ●●
●● ●● ● ●● ●
●● ● ●
●●
● ●● ●
●
● ●●● ● ●● ●
●●●
●
●
● ●
●● ● ● ● ●
●● ●● ● ●● ●
●● ● ● ●● ● ●●
●●
● ● ● ● ● ●● ●● ● ● ●●
●● ●●
●
● ● ●● ● ● ●●● ● ●● ●
●● ● ● ● ● ●● ● ●
●
●
● ●● ● ● ● ●● ● ●
●● ● ● ● ●●●● ●

This display shows the pairwise scatterplots of 150 random points from a three-dimensional canonical
vine with uniform marginals where the c12 copula is from the Frank family, the c13 copula comes from
the Galambos family, and the conditional pair-copula c23|1 is the independence copula. The parameters of
the copulas c12 and c13 have been chosen so that the Kendall τ coefficient is approximately 0.3 and 0.65,
respectively.

Simulation from canonical and D-vines is relatively straightforward and it is based on

the following general sampling algorithm for n dependent uniform [0, 1] variables. First,
sample n independent uniform random numbers wi ∈ [0, 1] and now compute

5 In the case of three variables every canonical vine is a D-vine and vice versa. In higher dimensions this is
no longer the case.
6 The parameters for the Frank and Galambos copulas have been chosen so that their Kendall τ coefficients
are approximately equal to 0.3 and 0.65. For further parameter values see [18, Table 5.1].

14
 x1 = w1
x2 = F−1
2|1 (w2 |x1 )
−1
x3 = F3|1,2 (w3 |x1 , x2 )

x4 = F−1
4|1,2,3 (w4 |x1 , x2 , x3 )
..
.
−1
xn = Fn|1,2,3,...,n−1 (wn |x1 , . . . , xn−1 ).

To implement this algorithm we will need to calculate conditional distributions of the

form Fx|v (x|v) and their inverses where as before v is a vector of variables and v−j de-
notes the same vector but excluding the jth component. The case where v has only one
component this reduces to

∂Cxv (Fx (x), Fv (v))

F(x|v) = . (19)
∂Fv (v)

If we further assume that the marginal distributions Fx (x) and Fv (v) are uniform, then (19)
reduces further to

∂Cxv (x, v)
F(x|v) = . (20)
∂v

This last construction will occur often in the simulation algorithm. So define the function
h(·) via

∂Cxv (x, v; Θ)
h(x, v; Θ) = F(x|v) = , (21)
∂v

where the third argument Θ denotes the set of parameters for the copula associated with
the joint distribution of x and v. Also the second argument of h(·) always denotes the
conditioning variable. Moreover, let h−1 denote the inverse of h with respect to the first
variable; that is, the inverse of the conditional distribution function.

Note that equation (16) can be used recursively, and at each stage the number of
conditioning variables decreases by one. Eventually, we would arrive at the special case
shown in (19). This will allow us to recursively compute the conditional distributions and
so create a sample from the joint distribution function.

A appropriate choice of variable v j to use in (16) will give us either the canonical
vine or the D-vine. For the canonical vine we always choose v j to be the last conditioning
variable available

15
 ∂C j,j−1|1,2,...,j−2 (F(x j |x1 , . . . , x j−2 ), F(x j−1 |x1 , . . . , x j−2 ))
F(x j |x1 , . . . , x j−1 ) = , (22)
∂F(x j−1 |x1 , . . . , x j−2 )

and for the D-vine we always choose the first conditioning variable

∂C j,1|2,...,j−1 (F(x j |x2 , . . . , x j−1 ), F(x1 |x2 , . . . , x j−1 ))

F(x j |x1 , . . . , x j−1 ) = . (23)
∂F(x1 |x2 , . . . , x j−1 )

Algorithm 1, taken from [1], gives the pseudo-code for sampling from a canonical vine
with uniform marginals. The use of uniform marginals is for simplicity only and the
algorithm can easily be extended to other marginal distributions. We make heavy use of
the h function defined in (21) and in the algorithm we set

vi,j = F(xi |x1 , . . . , x j−1 ).

The symbol Θ j,i represents the set of parameters of the corresponding pair-copula density
c j,j+1|1,...,j−1 . Algorithm 2, also taken from [1], gives the pseudo-code to sample from a
D-vine.

5. Estimating the Pair-Copula Decomposition

The last section showed how the canonical or D-vine constructions decompose an
n-dimensional multivariate density function into two main components. The first one
is the product of each of the marginal density functions. The second component is the
product of the density functions of n(n−1)/2 bivariate copulas. To estimate the parameters
of either construction we need to

1. decide which family to use for each pair-copula and

2. estimate all necessary parameters simultaneously.

5.1 Chi-plots to Determine Appropriate Pair-Copulas

To specify which bivariate copulas we want to use in the canonical or D-vine de-
compositions we will pursue a graphical method based on a construction of Fischer and
Switzer [10, 11] known as a chi-plot or χ-plot. There are other more formal selection
techniques [6, 16, 18] that should be used in conjunction with this method.

The χ-plot is a powerful graphical representation to help us extract information

about the dependence between two random variables. Traditionally the scatterplot has
been used to detect patterns (or lack of patterns) of association between two variables.

16
We know that if two random variables are independent, then the scatterplot should show
a random arrangement of points. Unfortunately, the human eye is not very good at iden-
tifying randomness. We are all too eager to find some sort of pattern in the data. The
χ-plot was designed as an auxiliary display in which independence is itself manifested in
a characteristic way.

The essence of the χ-plot is to compare the empirical bivariate distribution against
the null hypothesis of independence at each point in the scatterplot. To construct this plot
from a set of points (x1 , y1 ), (x2 , y2 ), . . . , (xn , yn ) we calculate three empirical distribution
functions: the bivariate distribution H and the two marginal distributions F and G. For
each point (xi , yi ) let Hi be the proportion of points below and to the left of (xi , yi ). Also let
Fi and Gi be the proportion of points to the left and below of the point (xi , yi ), respectively.
Figure 5.1 shows graphically how to calculate Hi , Fi and Gi .

Each point (χi , λi ) of the χ-plot is then defined by

Hi − Fi Gi
χi = p (24)
Fi (1 − Fi )Gi (1 − Gi )

and

1 2 1 2
(    )
λi = 4Si max Fi − , Gi − , (25)
2 2

where

1 1
  
Si = sign Fi − Gi − . (26)
2 2

The formal definitions for H, F and G are

1 X
Hi = I(x j ≤ xi , y j ≤ yi ), (27)
n−1
j6=i

1 X
Fi = I(x j ≤ xi ), (28)
n−1
j6=i

1 X
Gi = I(y j ≤ yi ), (29)
n−1
j6=i

where I(A) is the indicator function of the event A. To avoid some erratic behavior at the
edges of the dataset only the points that satisfy |λi | < 4{1/(n − 1) − 1/2}2 are included in
the display. This restriction will eliminate at most eight points [10, p. 256].

17
 Figure 5.1
Chi-plot Construction I

8
(xi , yi ) Hi = 20

(a) (b)

13 12
Fi = 20 Gi = 20

(c) (d)
Panel (a) shows the scatterplot of n = 21 points with point (xi , yi ) highlighted. Panel (b) shows the region
where we need to count the number of points to calculate Hi . Panels (c) and (d) show the regions for the
calculation of Fi and Gi , respectively. In all cases we do not count the point (xi , yi ) as being inside the region.

The value of λi is a measure of the distance of the point (xi , yi ) from the center
of the dataset and the value of χi is a measure of the distance of the distribution H to
the distribution of independent pairs of random variables (X, Y). Note that the functions
Hi , Fi , Gi are the empirical distribution functions of the joint distribution and the marginal
distributions of (X, Y) and depend only on the ranks of the observations. Figures 5.1
and 5.2 graphically show how to calculate the points necessary to construct a χ-plot.

If X and Y are independent, then the numerator of χ, which equals H − F · G, is

equal to zero. In practice we have a sample from the bivariate distribution H and if X and
Y are independent we expect the χ-plot to show most points close to the line χ = 0. In
our χ-plots we have included a 95 percent confidence band around the null hypothesis of
independence. If most points lie within these control limits, then we have strong evidence
that the variables are indeed independent.

To understand what a χ-plot has to offer let’s consider three examples where we
know what the dependence between X and Y is and see how it is manifested in the χ-plot.

18
 Figure 5.2
Chi-plot Construction II

(xi , yi )
− +

+ −

(a) (b)

(xi , yi )

(c)
Panel (a) shows the scatterplot of n = 21 points with point (xi , yi ) highlighted. The center of the dataset is
defined as the point with coordinates equal to the medians of the marginal distributions and is shown as
the intersection of the vertical and horizontal lines. Panel (b) shows the sign that λi would have depending
on which quadrant the point (xi , yi ) is located. In panel (c) we want to calculate the distance from the point
(xi , yi ) to the center of the distribution. This distance is not the usual Euclidean distance but rather the
maximum of the squares of the distance from the marginal distributions Fi and Gi .

In Figure 5.3 we have 200 random points (xi , yi ) taken from a bivariate normal
distribution with mean µ = (0, 0), variance σ2 = (1, 1) and correlation coefficient equal
to 0. In this case, we know that X and Y are independent and so the χ-plot (see Figure 5.3)
should show that most of the points fall within the control bands.

Now consider again 200 points sampled at random from a standard bivariate nor-
mal distribution with correlation coefficient equal to 0.5 as shown in Figure 5.4. These
points now have a monotone positive association and the χ-plot shows that as a pattern
that increases from the point (−1, 0) towards the point (0, 0.5) and then decreases as λ
reaches the value +1. The majority of the points are now above the line χ = 0 signaling
positive association.

19
 Figure 5.3
Bivariate Normal with Zero Correlation

1.0

1.0
● ●
●● ● ● ●
● ● ●●
● ●
●
● ● ● ● ●
●● ● ●
● ● ● ● ● ● ●
● ●
● ● ●
0.8

●● ● ● ●
● ●
● ●

0.5
● ● ● ● ●
● ● ● ●
● ● ● ● ●
● ● ● ●
● ●
● ● ● ●
● ● ● ●● ●
● ● ●
0.6

● ● ● ●
●● ● ●
● ● ● ● ● ●●●●
● ● ●● ●● ●●● ●
●● ● ● ● ● ●●●●●●
● ●● ●●● ●
● ●●●● ● ●● ● ●●● ●
●●● ●●
● ●●
●

0.0
● ● ●●● ● ● ●●● ● ● ●●● ● ●● ●
● ● ● ● ●●●●● ● ●
●●●● ●●●●
● ● ●●●●● ●●●●
●●●
●
Y

●●

χ
● ● ● ●●● ●●●●●● ●●● ● ● ● ● ● ●●● ●● ●●●● ●●●● ● ● ●● ●● ●●● ●●● ● ●
● ● ● ●
● ●● ● ●
● ●●●●●●
● ●● ●
● ●●●
●● ● ● ●● ●● ●
●●
● ● ●● ● ●
● ● ●● ●
● ●●●●
● ●
0.4

●● ● ● ●
● ● ● ●
● ● ●
● ● ●
● ●● ●
●
● ●

−0.5
● ● ● ●
●
● ● ● ● ●
● ●
● ● ● ● ●● ●●
0.2

●● ● ●
●
● ●
● ● ● ●
●● ● ●
● ● ●●● ●
● ● ● ● ●
● ● ● ●
●
● ● ●●
● ● ●

−1.0
● ●
0.0

● ● ● ● ●

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of 200 random points from a bivariate normal distribution with mean
µ = (0, 0), variance σ2 = (1, 1) and correlation coefficient equal to 0. The right-hand panel shows the
corresponding χ-plot. Note that the majority of the points fall within the 95 percent control limits.

Figure 5.4
Bivariate Normal with 0.5-Correlation
1.0

1.0

● ● ●
● ● ●
● ●
● ● ● ● ●
● ● ● ● ●
● ●●
● ●
● ● ● ● ● ●
●
0.8

●●
● ● ●●
0.5

● ● ●
● ● ● ●
● ● ●● ●● ● ●● ●● ● ● ● ●●●●
● ●
● ● ● ● ● ● ●● ● ● ● ● ● ●● ●
● ●● ●
●
● ● ●● ●● ● ●●● ●● ●●●
● ●●●● ● ●● ●●● ●●●
● ●● ●●
● ●● ●●
● ●
● ● ● ●● ● ● ● ●●● ●●● ● ● ●● ●● ● ●
● ●●
●● ●●
● ●● ●●
● ● ● ●
● ● ●● ● ●● ●●●● ● ●●
● ●● ● ● ●●
● ●
0.6

● ● ● ●●●● ● ● ● ●
● ● ● ●● ● ●
● ● ● ●● ● ● ● ● ● ● ●● ● ● ●●●
●● ● ● ●
● ●●●●●● ●●●
● ●●● ●
● ● ●● ● ●● ● ● ●●
● ●●●●● ●
● ● ●
● ● ●
● ●● ● ● ● ● ●●●●●●●
0.0

● ●
Y

● ● ●
χ

● ●● ● ● ●● ● ●
● ●
● ● ● ●
●●
0.4

● ●●
● ●● ● ● ● ●
● ● ●
●● ● ● ● ●
● ● ●
●
● ●
−0.5

● ● ● ● ● ● ●● ● ● ●
● ●
●
0.2

● ●● ● ● ● ● ●
●
● ●●
● ●
● ● ●● ●●
● ● ●
● ●
● ● ●
● ●●
● ●
● ●
●● ● ●
−1.0

●
0.0

●●●● ● ●

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of 200 random points from a bivariate normal distribution with mean
µ = (0, 0), variance σ2 = (1, 1) and correlation coefficient equal to 0.5. The right-hand panel shows the
corresponding χ-plot. Most of the points are outside the control lines and above the line χ = 0 which
indicates positive dependence. Note how the peak is near the point (0, 0.5).

For the next example, taken from [10, example 4], consider a set of points where
there is no monotone association present. 7 The data consists of 200 points taken from the

7 We have slightly modified the example by plotting twice as many points as in the original.

20
standard bivariate normal distribution with zero correlation and satisfying the restric-
tions Y ≥ 0 and |X2 + Y2 − 1| ≤ 1/2. The scatterplot and the χ-plot are shown in Figure 5.5.

Figure 5.5
Non-Monotone Association

1.0
●
● ● ●
●●
1.4

● ● ●●
●
●
● ● ● ●
● ● ● ●
● ● ●●
● ●
1.2

● ● ●
● ●
●●

0.5
●
● ● ●
● ● ●●● ●
1.0

● ● ●
●● ●
● ● ● ● ● ● ● ● ●●
●●● ● ● ●●● ●● ●● ●●● ●
● ●● ● ●
● ● ●●● ● ● ● ●● ●● ●
●●●● ● ● ● ● ● ●
● ●
● ●●● ●● ● ● ●● ● ● ●● ●●● ●● ●
● ● ● ● ● ● ● ● ●●●● ●
● ● ●● ●●● ●●●
0.8

● ●● ● ●
● ● ● ●● ● ●● ● ●
● ● ● ● ●

0.0
● ● ● ● ● ●● ● ●● ● ● ●
Y

● ● ●●

χ
● ●
● ● ●● ●●
●●●●● ●● ●● ●
● ● ●
● ●●
● ● ● ●● ● ● ● ● ● ● ● ● ● ●
● ● ●● ● ● ● ●● ● ●
● ● ● ● ● ● ● ● ● ● ●● ●●
0.6

● ● ● ● ● ● ● ●
● ● ●● ● ● ●● ●● ●●● ●● ● ●
●●●●● ●
● ●
●
●
● ●● ● ● ● ● ●● ● ●●
●●●●● ● ●● ● ●●
● ● ●
●
●
● ● ● ●● ● ● ●
●
● ●
● ●
● ● ●●●
0.4

−0.5
● ● ● ●
● ●● ● ● ●
● ●
● ●● ● ● ●
●
● ● ●
● ● ● ●● ● ●
0.2

● ● ● ●●
● ● ●
● ● ●●
● ●●
● ●
● ●●

−1.0
● ● ● ●
● ●
●●
0.0

● ● ●

−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of 200 random points from a bivariate normal distribution with mean
µ = (0, 0), variance σ2 = (1, 1) and zero correlation satisfying the constraints Y ≥ 0 and |X2 + Y2 − 1| ≤ 1/2.
The right-hand panel shows the corresponding χ-plot.

The χ-plot shows that there are many points outside the control limits. This indicates
that the original variables are not independent. It also shows a more complex pattern
compared to the previous examples. In particular, there are four distinct regions in the
χ-plot roughly corresponding to the four quadrants defined by the median point of the X
and Y marginal distributions. In Figure 5.6 we have added a vertical and horizontal line
passing through the median of the marginal distributions and used different symbols to
plot in each quadrant. Note that in the lower-left and upper-left quadrants the association
between X and Y is positive and these points appear predominantly above the line χ = 0
in the χ-plot. Similarly, the points in the lower-right and upper-right quadrants of the
scatterplot appear mainly below the line χ = 0 in the χ-plot.

Appendices A–E show scatter- and χ-plots for various copulas under different pa-
rameters. This catalog serves as a good starting point to compare against an empirical
dataset. In general, to determine the best copula for a given application we would use
both formal and informal selection methods. For some of the formal techniques consult
[18, 6, 20]. One of the informal techniques is to look at the χ-plot of the empirical data and
compare it against the simulated χ-plots for various copula families to find an appropriate
match.

21
 Figure 5.6
Non-Monotone Association

1.0
1.4
1.2

0.5
1.0
0.8

0.0
● ●
Y

χ
● ●
●●●●● ● ●
●● ● ● ●
0.6

● ●
● ● ●
●● ● ● ●● ●● ●●● ●● ●
● ● ● ● ●● ● ●●
●●●●● ●● ●
●
● ● ●
●
● ●
● ● ●
● ● ●●
0.4

−0.5
● ●
●
●
● ● ●
● ●● ●
0.2

●● ●
●
● ●●
●●
●●

−1.0
● ●
●●
0.0

−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 −1.0 −0.5 0.0 0.5 1.0

X λ

This is the same as Figure 5.5 but we have added horizontal and vertical lines at the medians of the marginal
distributions. In each of the four quadrants we have used different plotting symbols. These symbols have
been carried over to the χ-plot.

5.2 Parameter Estimation via Maximum Likelihood

Once we have selected the appropriate pair-copula families we can proceed with
the estimation of the parameters via maximum likelihood. Suppose we have an n-dim
distribution function along with T observations. Let xs denote the vector of observations
for the s-th point with s = 1, 2, . . . , T. The likelihood function for a canonical vine decom-
position is
T Y
Y n
L(x; Θ) = f (xs,k ) ·
s=1 k=1
n− j
n−1 Y
Y
c j,j+i|1,..., j−1 (F(xs,j |xs,1 , . . . , xs,j−1 ), F(xs,j+i |xs,1 , . . . , xs,j−1 )). (30)
j=1 i=1

By taking logarithms and assuming that each of the n marginals is uniform on the unit
interval 8; that is, f (xs,k ) = 1 for all s and k, the log-likelihood function is
T X
X n− j
n−1 X  
`(x; Θ) = log c j,j+i|1,..., j−1 (F(xs,j |xs,1 , . . . , xs,j−1 ), F(xs,j+i |xs,1 , . . . , xs,j−1 )) (31)
s=1 j=1 i=1

Similarly, the likelihood function for a D-vine is

8 For simplicity we are using uniform margins. Extending our discussion to non-uniform margins is straight-
forward.

22
 T Y
Y n
L(x; Θ) = f (xs,k ) ·
s=1 k=1
n− j
n−1 Y
Y
ci, j+i|i+1,...,i+j−1 (F(xs,i |xs,i+1 , . . . , xs,i+ j−1 ), F(xs, j+i |xs,i+1 , . . . , xs,i+ j−1 )), (32)
j=1 i=1

and the log-likelihood (assuming uniform marginals) is

T X
X n− j
n−1 X  
`(x; Θ) = log ci,i+ j|i+1,...,i+j−1 (F(xs,i |xs,i+1 , . . . , xs,i+ j−1 ), F(xs,i+ j |xs,i+1 , . . . , xs,i+ j−1 )) . (33)
s=1 j=1 i=1

For each copula term in the log-likelihood (31) or (33) we have at least one parameter to
estimate. 9

The conditional distributions for a canonical vine

F(xs,j |xs,1 , . . . , xs,j−1 ) and F(xs,j+i |xs,1 , . . . , xs,j−1 )

or the conditional distributions for a D-vine
F(xs,i |xs,i+1 , . . . , xs,i+j−1 ) and F(xs,i+j |xs,i+1 , . . . , xs,i+j−1 )

are again determined by using the recursive relation (16) and the appropriate h(·) func-
tion (21). Then we can use numerical optimization techniques to maximize the log-
likelihood over all parameters simultaneously.

Algorithm 3 (shown in Appendix F), from [1], computes the log-likelihood func-
tion of a canonical vine for a given set of observations. The numerical maximization can
be carried out via the Nelder-Mead algorithm [21] or another optimization technique.

Starting values for the numerical maximization of the log-likelihood can be deter-
mined as follows:

1. Estimate the parameters of the copulas in tree 1 from the original data. That is, fit
each bivariate copula to the observations. This estimation is easy to do since we are
only looking at two dimensions at a time.
2. Compute the implied observations for tree 2 using the copula parameters from tree 1
and the appropriate h(·) functions.
3. Estimate the parameters of the copulas in tree 2 from the observations in step 2.

9 The number of parameters depends on the bivariate copula chosen. Many bivariate copula families have one
parameter, but others have two or more, such as the t-copula.

23
 4. Compute the implied observations for tree 3 using the copula parameters from
step 3 and the appropriate h(·) functions.
5. Continue along the sequence of trees in the pair-copula decomposition.

These starting values can then be passed to the full numerical maximization rou-
tine.

6. Illustration Based on Currency Rate Changes

In [26] various multivariate copula models were fitted to currency rate changes.
We will use the same currencies, but for a longer observation period, to illustrate the
pair-copula construction. The raw data consist of monthly rates of exchange between
the Canadian, Japanese and Swedish currencies and the U.S. dollar. This raw data spans
from January 1971 to July 2007, and the source is the Fred database of the Federal Reserve
Bank of St. Louis [8]. As in [26] we will apply the pair-copula construction to the monthly
changes in the rate of exchange. The left-hand panel of Figure 6.1 shows the scatterplot
matrix of this dataset.

Figure 6.1
Changes in Currency Exchange Rates
● ● ●● ● ●●● ● ●● ● ●● ●●●●● ●
●●
● ●●
●
● ●●● ●● ●● ●●● ●● ●●
●● ●● ● ●● ●● ● ● ●
●● ●●●● ●●
●● ● ●●●● ●
●
● ●●●
●
● ●● ●
●●
●
●
●
● ●● ●● ●●● ●●●●●●● ●● ● ● ●● ● ● ●●●● ●●
● ● ●●
● ● ●● ● ● ●●● ● ●●●●● ● ● ● ● ● ● ●● ●● ●● ●●● ●●●● ●●●
● ● ●● ● ● ● ●
●● ● ●● ●● ● ●●
●●
●● ● ● ● ●●●
●● ●● ●
● ●
●● ●●●● ●● ● ●● ●●
● ● ● ●● ●●●●
●
●●●●
●● ● ●● ● ● ● ● ● ●● ●
●●●
● ●●
●●
●●●●
● ● ●
●
● ● ● ●● ●● ● ●● ● ●● ● ● ●●●●●
● ●
●
● ●
●●●● ● ● ●●● ●● ● ● ● ●
● ●●● ●● ●● ●●● ●●
● ● ● ● ●●● ●●●● ●● ●● ● ●
●
● ●● ●●●● ●●●
●●●● ●
●●● ●
● ● ●● ●● ●●● ● ●● ● ● ●● ●●●●●●●●● ●
● ●● ●● ●●●● ●● ●●●●●●
●●● ● ●● ● ● ●●
● ● ● ● ●●● ● ●● ●
●● ● ●● ●● ●●● ● ●● ●
●
●●● ● ●●● ●● ● ●● ●
●●●●● ●● ●●
●●● ● ● ● ●●●●● ● ● Sweden ● ●●● ●● ●●●● ●● ●● ●
●●● ● ●●
●●●●● ● ● ●
Sweden
● ● ●● ●●● ●
●
●● ● ●● ●
●●●●● ● ●
● ● ●● ●●● ● ●● ● ● ● ● ● ● ●● ●●● ●●● ●● ● ●● ● ●● ●●
●● ●●
● ●●●
● ●
●● ●
●
● ●
●●●●
●●
●● ● ●●●● ●●●●●●●
●●●●
●●●
●●●●
● ● ●
● ●●●● ● ● ●●● ●
● ●
●● ●● ● ●●● ● ● ● ●● ● ●● ●
●
●
● ●●● ●
● ●● ●
●
●
●●
●
●● ●
● ●
● ●●
●●● ● ●● ●
●●
●●
● ●
●
●●●●
●●●● ●●● ● ● ●●● ● ●●●● ●●● ● ●
● ●●
● ●● ●● ● ●
● ●●●
● ●●●●●●
● ●●
●
●
●●
●
●
●●
●●
●●●●● ●● ●●● ●●
●●●
●●
●●
●●●●
●
●●
●
●●
● ●
●● ●●●●●
●
● ●● ●●● ● ● ●
● ● ●●●●
● ●●●●● ●● ●
●● ●● ● ●
●
●●
●●●●● ●●
●●
● ●
●●●
●
● ●
●
●●
●●●
●
●
● ●
●
●●●●●●● ●
● ● ●●●●●
●●●
●●●
●
●●●
●●
●
●
●
●
●
●●
●●
●●
●
●
●●
●● ● ● ●●●● ●●●
● ● ●●●●
● ● ● ●●●●●
●● ●●
● ●●● ● ● ●●●● ●
● ● ●● ●
●
●●●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●
●
●●
●●●
●●●●●● ● ●●●● ●
●● ● ●
●
●●●●
●
●●
●
●
●●
●
●
●
●●
●
●●
●●●
●
●●
●●
●
●●●●●● ● ●●●● ●●●●● ● ● ●● ● ● ●● ● ●● ● ●
●
●
●
●● ●● ● ● ●● ● ●●●
●
●●●●
●●●
●
●●●
● ●
●
●●●
● ●
●
●●●
●
●
● ●
●●●●●
● ●●●●
● ●●
●
● ●●
● ●●
●●
●
●●●
●
●
●●
●●
● ●
●●
●
●●
●●● ● ● ● ●●● ●● ●● ● ● ●● ●●
●● ●● ●●● ●
●● ● ●●●
●●●
● ●
●●● ●
●●
●● ●●
● ●
●
●●
●●●●
●
●
●●
●
●
●
●●●●
●● ● ● ● ●●●●●●●●●
●● ●●
●●
●●●●●
●●
●
●
●
●●
●
●●●
●
●● ●●● ●●●● ●● ●
● ●● ● ●
● ● ●● ● ●● ●● ●
● ●● ● ● ●● ● ●● ●
●
●● ●●● ●
●
●
● ●●
●
● ● ●
● ●● ● ●
● ● ●●
●
●●
●●●
●●● ●
●
●●
●● ●●●●●● ●● ●●●● ● ● ●
●●●
● ●● ● ●●
●●● ●●●●
●●●●●
●
● ●● ● ●● ●●
●
●●●
●●●●
●●●
●●●● ●
●
● ● ● ● ●
●● ●● ●●● ●
●● ● ● ●● ●● ● ● ●
● ●● ●
●
●●● ●●● ●● ●●
● ●●●
●● ●● ●● ●● ● ●●
●● ●● ● ● ●●● ●●●●● ● ●●● ●● ● ●●
● ●
● ● ●
●●
●
●● ● ●●●
●
● ●●
●●●
●● ● ●● ●●● ● ● ●● ● ●●
●● ●●●
● ●
●
●●● ●●
● ● ● ● ● ●●
●● ●● ●
● ● ● ●
●
● ● ●● ●●
● ● ●●● ●● ● ● ●● ● ●

● ● ● ● ● ● ●● ● ●●●●● ● ●●●●● ● ● ●●●● ●●● ● ●●

●● ●
●● ●● ● ●
● ●● ● ● ● ●●● ●●● ● ● ●●●● ● ●●
●● ● ● ●
●●
●●
● ●●● ● ●
● ●
●● ● ● ●
●
● ● ●●●● ● ● ● ●
●●●●●●●●●●
●●
● ● ● ● ● ● ● ● ●● ●● ●● ● ●● ● ●
●●● ● ● ●● ●●●● ● ● ●● ● ●● ●● ● ●
● ● ● ● ●●●●● ●
● ●
● ● ●●
●
● ●●
●●
●●
● ●●
●●●●● ●● ● ●●● ●●● ● ●● ● ● ● ● ●●●● ● ●
●●
●●●
●●
●●●
● ●●●
●
●● ● ● ●●
● ●●● ● ●●
●●●● ●●●● ●●● ● ● ●●● ● ● ●
●
● ● ●●● ● ●●●● ●●
●
●●
●
●●●●
●●
●
●●●●●● ● ● ●●● ●
●
●●●
●
● ● ● ●●● ● ● ●●●●●● ●
●●●●
●
● ●● ●● ●
●
● ●● ● ●
●
●●● ● ●●● ● ●
●●●●
●● ●●●
●●●
●●●●●●●
● ●●●● ● ●● ● ●●
● ●
●●●●
●● ●
●●
●●
●●●●●● ● ● ●● ●
●
●● ● ●
● ● ● ● ● ●● ●● ● ●
●● ●
●
●
●● ● ● ●● ● ● ●
●●● ● ●●●●
●●
● ●●●● ●
● ●●● ●●
●●
● ●● ●● ●● ● ●
●●●● ●●●●
●●
●●●
●
●
●●●●
●●● ● ●
●● ●● ●● ● ● ● ●●● ● ●● ●●●
● ● ●
●● ● ● ● ●
●
●
●●●
●●
●●
●●●
●
●
● ●●●
●●
●● ● ●● ●
●●●
●●●
●
●
●●●
●
●
●●●
● ●
●● ●● ● ●●●
● ● ●● ●●● ● ●●●●●
● ●●● ●● ●●● ●● ●● ● ●●
● ●
●
●●●
●●●●●
●●
●●
●●
●
●●
●
●
●
●●●●
● ●
●●●
●
● ●●●
●●●
● ● ● ● ●●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●
●
●●●●
●
●
● ●●●●● ● ● ● ●
●
● ●● ●●●●●● ●●● ● ● ● ● ● ● ●●
●●●●●●●●
●●●●
●●
●
●●●
●●
●●
●
●
●
●
●●
●●
●
●●●●
●●
●
●●
●● ●●●●●
●●
●●
●●
●
●●
●
●●
●●●
●
●●
●
●●
●
●
●
●
●●●
●●
●
● ●
● ● ● ●● ● ●
●●●● ●
●
●● ●●● ●● ● ●●●●●● ●● ● ●
●
● ●
● ● ●
●● ●
● ●
●●
●●●●
●● ●
●●●●
● ●●
●●● Japan ●●● ●●
●
● ●●
●
●●●
●
●●
●
●●●●
● ● ● ●●●● ● ●● ● ● Japan ●● ● ●●● ● ● ●
● ●● ●●
●
● ●●●
●●
● ●
●●
●● ●
●●● ● ●● ●●● ●
● ●
●
●●
● ●
●
●●●
●●
● ●● ● ● ●● ●● ●● ●●●● ● ●●●●
● ●● ● ●●●●
●●● ● ●● ●
● ● ●● ●●
●●
● ● ●
●●●●●●
● ●● ●●●
●
● ●● ●
●●●● ●●
●●
●
●●
●●
●
● ●●
●●●
●●●
●●
● ● ●
●●
●● ●●●●● ●● ● ● ● ● ●
● ● ●●●●● ●
● ●●
●
● ●● ●
● ●●●●●
●●
●● ●●●● ●●●●●● ● ●
● ●●
● ●● ●●
●●
●●● ● ● ● ●
● ● ●●● ●● ●●●● ●● ● ● ●
●● ●●●● ●●
●
●●●●●●● ●●
●●● ● ●● ●●●
●
●
●
●
●●● ● ●● ●● ●● ●
● ●● ● ● ● ● ●●●● ●● ● ●● ●●● ●●●
● ●●
● ● ●● ●●● ● ●
●●●● ● ●●● ●● ● ● ●● ● ●●●●● ● ● ● ● ●● ● ● ● ●●
● ●●● ● ● ●●●●●●● ● ● ●● ●
● ●●
● ●● ● ●● ●●● ●● ●● ●●●● ● ●●●● ●● ●●● ●●● ● ●
●● ● ●● ● ●●
●●●
●●● ●●●●
●
●●● ● ●●● ●● ● ● ● ●● ● ●● ●● ●
●
● ●
●
● ●
● ●●
● ● ● ●● ● ●●●
●● ●● ● ●● ● ● ●● ● ● ●● ●●
●●●
●
●●● ●
● ●●●● ● ● ●●●
●
● ●● ● ● ●● ● ●● ●●● ●●● ●●
●● ●
●●● ●
● ●●● ●●
● ● ● ●● ● ●
● ● ●● ● ● ●●● ●●● ● ● ●
● ● ● ●●●●●● ● ●●● ● ● ●●
● ●
● ● ● ●●●●● ●●● ● ● ●●●●
●
●●●● ●●●
●●● ●
●●●●●●● ● ●● ● ●●
●
●
● ● ●●●● ●● ● ●● ●●
● ● ●●
●
● ● ●
● ●● ●● ● ●● ●●
●
● ● ● ●● ●●● ● ● ●
● ● ● ● ● ● ●
●●●●●●●●●
● ● ● ●●● ●

● ● ● ●●● ● ●● ● ● ● ● ● ●●● ● ● ● ●
● ● ● ● ●
●●
●
●● ●●●●● ● ● ● ●●●●
●● ●●● ● ●● ● ●● ●●●●
● ● ● ●●●●
● ●● ● ●●
● ● ● ●● ●
●● ●● ●●●● ●●● ●●
●●●●
●●
●
● ●
● ●● ● ●●●●● ●● ●● ● ●●●●● ● ● ●● ●● ●
● ●● ●● ●●●
●● ● ● ●● ● ●● ● ●●●●●● ● ● ●● ●
● ● ● ● ● ●●●● ●●●● ●
●● ● ●●● ●●
●●●● ●●●●●
●●● ●●●●
●
●● ● ● ●●
● ●●● ●●●●
●●
●●●
● ● ● ●●
●
● ●● ●● ●
●● ● ●● ● ●●●● ●●●● ● ● ●● ●● ●●●●●●●
●
●●● ●
● ● ● ● ● ●●●●
● ●
● ● ●●●●●● ●
●
● ●●● ●
●●● ●●● ● ● ●● ● ●●●
● ● ● ●
● ●
● ● ●
●
●
●●●●●● ●●●
●●
●●● ●
● ●
●●● ● ●●●●● ●
●● ●
● ● ●
●● ● ● ●● ● ● ●
● ●●● ●● ● ●●●● ●● ●● ●●●● ●●
● ● ● ●
●●● ●●●
●
●●
●
●●
●●
●
●
●●
●●
●●
●
● ●●
● ●
● ●●●
● ●
●
●
●
●●●
●●
●●
●●
●●●
● ●●● ● ●●● ●
● ● ●● ● ●● ●● ●● ●
●●● ● ● ●●●●●●●● ●● ●
● ● ●
●●
●
● ●
●●
●●
●
●
●
● ●●
●●
●●●●●●● ● ● ●●●●●
● ●
●●
●
● ●●●
●
●
●
●●
●●
●
●
●●
●●●●●●●● ● ●● ● ●●● ● ● ●●
●● ●●●●
●● ●
● ● ● ●● ● ● ● ●
●● ● ●●●
●●
●●●● ●● ●●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●●●
●●●
●●● ●●
●● ● ●●● ●
●
●
●●
●●●
●●
●
●
●
●●●
●●
●●
●
●●
●
●●
●●
●●
●●● ●●● ● ● ● ● ●●● ●●●
● ●●●● ● ●●●● ●●●
●
●●● ● ●●
● ●● ●●● ●●●●●●●
●
●
● ●
● ●
●● ● ●●
● ●● ●●
●●●
●●●
●●●●●●●
●●●● ●● ● ● ●
● ●● ●● ●● ● ●● ● ●
● ● ●
● ●●●
●●●
●
●●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●●●●●
●
●●●● ● ●
●●● ●
● ●
●●●●
●
●
●
● ●
●●●●
●
●
●●●
●
●
●●
●●
●
●●
● ●
●● ●●
●●
● ● ●
● ● ●●● ●●● ●● ●●●● ● ● ● ● ●● ● ● ● ● ●● ●●●
●●●
● ●● ●● ●
●
●
●
●
●●●
●
●
●●
●
●●●
●
●●●
●●●●●
● ● ●
●●
●
●
●
●●●●
●
●
●
●
●
●●●
●
●
●
●●
●● ●●● ●●●●●
●●● ●●●
●● ●●●●● ● ● ● ● ●●
●● ●● ●●●●●● ●● ●●● ●
●
Canada ● ● ●●●●●●
●●
●●
●
●●
●●
●
●
●●
●
●
●
●●●
●
●●
●●
●●
●
●
●
● ●
●●●●●●●
● ●
●
●
●● ● ●●●
●●●
●●
●●●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
● ●
●●●●●●
●●●●●● ●
● Canada ●●●● ● ●●● ●● ●●●●●
●●
● ●●● ●
●
●
● ● ● ●●
●● ●●
●●
●●● ● ●●●●●● ● ● ●●
●
● ● ●
●●
●●● ●● ●●●●
● ●
● ● ●
●● ● ●●● ● ● ●
● ● ●● ●
● ●
● ●● ●● ●●
● ●●●●●●● ● ● ● ●●●● ● ●● ● ●
●●● ●● ●●●●● ●
● ● ●●●●●●● ●
●●
● ● ●● ● ● ● ●●● ● ● ● ● ●
●● ●● ●● ● ● ● ●●●●
● ● ●
●●● ● ● ●●● ●● ● ●
●●●●●
● ● ●●●● ●● ● ● ●● ●● ●
●●●● ●
●
●● ● ●● ●
●
●
● ● ●● ●● ●●
●● ● ● ● ●●●●
● ●●●● ● ● ●● ● ●● ●● ●
● ● ●● ● ● ●● ● ●● ●
● ●● ●●● ● ● ● ●● ●●●●●●●● ●●
● ● ●● ●
● ● ●● ●
●● ●● ●●● ●● ●●●●● ● ●●● ● ●●● ●● ●
● ● ●● ●
● ● ●
●● ● ●● ●●●●●● ●●●
● ●●● ●● ● ●● ●● ●●●●
● ●
●
● ● ●●●
●
●●●
●
● ●●● ●●● ●● ●●
●●● ● ●● ● ●●
●● ● ● ● ●● ● ●● ● ●● ●●● ●● ●● ● ●
● ●
●● ● ●
●● ●● ● ● ● ● ● ● ●● ●
●
● ●●● ●
●●
●●●● ● ●● ● ●
●●●●● ● ● ●●● ● ●● ●●●● ● ●●
●● ● ●●● ●●
●●●
● ●● ●●●●●● ●●
● ● ●

The left-hand panel shows the monthly changes in the foreign exchange rates between the Canadian, Japan-
ese and Swedish currencies against the U.S. dollar for the period January 1971 to July 2007. The right-hand
panel shows the ranked transformed data.

We are interested in understanding the dependence between these variables and so before
we proceed any further we will remove the marginal distributions from our analysis; that
is, we are only interested in the ranks of our dataset. Thus in the right-hand panel of
Figure 6.1 we have transformed our data via its empirical distribution function. If (xi , yi )
is a point in one of the scatterplots on the left-hand panel, then the corresponding point

24
  
on the right-hand panel is bFx (xi ), b
F y (yi ) ; where b
Fx and b
F y are the empirical marginal
distribution functions.

Figure 6.2 shows the three pairwise χ-plots based on the currency data shown in
Figure 6.1. Note that all three plots are very different. Most of the points on the Canada–
Japan χ-plot (left-most panel) are within the control bands implying that these two vari-
ables are slightly positively dependent but not too far from being independent. But notice
that within the control bands the points are not randomly scattered. Rather they seem to
steadily increase from about λ = −0.5 to λ = 0.15 and then they decrease as λ continues
towards the right. Also note that as we reach the right edge the points seem to scatter
more than before.

Figure 6.2
χ-plots for Changes in Currency Exchange Rates
1.0

1.0

1.0
0.5

0.5

0.5
●●
●
● ●●
● ●
●
●●●● ●
●●
●●
● ●
●
●● ●●
●●
●●
●
●●●●
●
● ●●●●●
●
● ●●
●● ●●●●●
● ●● ●●●
●
●●●●
● ●
●●●●●
●
●
●●
●
●
●
●
●●● ●
● ●●●●●●●
●
●
●●
●
●●●●●
●
●
●●●
●
● ●
●●
●
●
●●
●● ●
●●●● ●
● ●
●● ●● ●
●●●●●● ● ●
●● ●
●●
● ●●●●●
●● ●●
●
●●●●● ●
●
●●●●
●
● ●●●●●
●
●● ●
●●●●
●●●
● ●●●●●● ●●
● ● ●●●● ●● ●●
● ●● ● ●● ●●●● ●
● ●●● ●●
●
●●●●●●
●●
●●● ● ●●
●●●●
● ●●●●●●● ●
●●●●
● ● ●●
● ●●
●●● ●
●●
●●
● ●
● ●●
●● ●●●
●
● ●●
●●● ●●
● ●
●● ●●●● ●●
●●●●
●
● ●● ●● ● ●● ●
●● ●●
●●
●
●●
●●
●●●●
● ● ●
●●●●●●
● ●
●●
●●
●●
●●
●●
● ● ●●●●●
●●●
●
●
● ●●
●
● ●
●●
●
●●●
●
●●●
●●●
●●
●●●● ● ●
● ● ●● ●● ●● ●●
● ●● ●
●●● ●●● ●●
●●●●●●
●●●●●●● ●● ● ●● ● ● ●● ●●●●
●
● ●●●●●●●●●
●
●● ● ●● ●●
●
●●●●
●●●●
●●●
● ●●
●●●
● ●
●●●● ●
●●●●●●
●●
●●
● ●
●●●● ● ●● ●
●● ●
●
● ●●● ●● ●● ●●
●●
● ●
●
●●●●
●●●●
●
●●
●●● ●
● ●
●●●
●
●●●
●
●● ●
●●●●
● ●
● ●●●● ●●●● ● ● ●● ●●
●●●● ●●●
● ●●●
●
●●
●●
●
● ●
●●●
●●
●● ● ●
●● ●●●●●
●● ●●●●●
●●
●
●● ●
●●●●●
●
●
●
●●●●●●●
●●● ●●●●●● ●
● ●● ● ● ●●
● ●
●
●●●
●
●●●
● ●● ●● ● ●●● ● ● ● ●●
●●●
●●●●
●●
●●
●●●●
●
●●
●
●●
●●
●
●● ●
●●●
● ●●●
●●
●●● ●●
●
●●●
●● ●●
●●●●
●
●●
●
●●
●
●●
●
●●
●
●
●●●
●
●●
●
●
●
●
●●●●●●
●
● ●
●
●●●
●●
● ●●
●● ●●
● ●●● ●
●
●●●
●
●
●●●
● ●
●●●●● ●
●●
● ●
● ●
●●
●●●● ●●●
● ● ●● ● ● ●● ● ●●
●●
●●
●
● ●
● ●
● ● ●●● ● ●
●● ●
●●
●
●●
●
●●
●
●
●●● ●●● ● ●●●●
●●●●● ●●● ●
● ●
●●●
● ●
●
●●●●
●●
●●●
●●
●
● ●●● ●●
●
● ●●●
● ●●
● ●●●●
● ●●● ●
●
●●●
●● ●
●●
●
●●
●●
●●●●●●●
●●
●●●● ●● ●
●●
● ●●●●
●● ●● ●● ● ●●●●●●● ● ●●● ● ●●
●
●●● ●●●● ●●●
● ● ● ● ● ● ● ●
● ●● ● ●●
0.0

●
0.0

0.0
●●
● ●●
●●● ●● ●● ● ●
●●● ●● ●●●
● ●
● ●● ● ● ●●●
●● ●
●● ●
●
●●
●●●●●
●
●
●●
●●●●● ●●
●● ●●●● ●
● ● ●
●● ●● ●●●● ●
●● ● ● ●●
●●● ● ● ● ● ●
χ

●
χ

χ
● ● ●
● ● ● ●● ●● ● ● ●
● ● ●
●
● ●● ●●
● ● ●● ● ●●
●●●● ●
●●●
● ● ●●●●●
●
−0.5

−0.5

−0.5
−1.0

−1.0

−1.0

−1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0

λ λ λ

Canada–Japan Canada–Sweden Japan–Sweden

The left-most panel shows that there is a mild association between the Canadian and Japanese data. Note
that there is a slight up and down pattern with increasing spread as λ approaches 1 along the horizontal
axis. The middle panel clearly shows that the Canadian and Swedish data are positively associated and note
the increasing spread as we approach λ = 1 on the horizontal axis. In the right most panel the dependence
between the Japanese and Swedish data is strong. Note the peak around λ = 0.1 and how as λ approaches
1 on the horizontal axis there is no appreciable spread. This behavior is opposite from the other two panels.

On the middle panel of Figure 6.2 we clearly see that the monthly changes in the U.S.
exchange rate between the Canadian and Swedish currencies are positively associated as
most points are outside the control band and above the line χ = 0. Again we have an
increasing trend from λ = −1 to λ = 0 and then a decreasing trend as λ reaches the right-
hand side of the plot. Note also the increasing spread as λ moves from 0 to 1. The highest
point is approximately at (λ, χ) = (0, 0.15), and so we expect Kendall’s τ between these
two variables to be approximately equal to 0.15.

25
 The last panel in Figure 6.2 shows that the strongest positive association is between
the Japanese and Swedish currencies. Note that the highest point is approximately at
χ = 0.35, and so we suspect that Kendall’s measure of concordance would be about that
value. As with the previous panel we see the steady increase as we approach λ = 0
and then the decrease as we continue towards the right. But notice that we do not see
the increasing spread as λ approaches 1. This indicates that for values far away from
the center of distribution (λ near 1) there seems to be no dependence among them. This
feature of the χ-plot rules out copulas such as the Gumbel or Galambos families.

Figure 6.3
Implied Data for Tree 2 Construction
1.0

1.0
● ● ●
● ●● ● ● ● ● ● ●
● ● ● ● ●●
● ● ● ● ● ● ●
● ● ●● ●●
● ●● ● ●● ●● ●●● ● ● ● ●
● ● ●● ●
● ● ●●
●
● ● ● ●
● ●●
● ● ● ● ●
● ● ● ●● ●● ● ● ●● ●
0.8

●● ● ● ● ● ● ● ●
● ● ● ●● ●
● ●
●● ●● ●

0.5
● ● ●
●●● ● ● ● ●● ●
● ● ●
●
●● ● ● ●
● ● ● ●
● ● ● ● ● ● ● ●●
●
●●● ●● ●
●● ● ●● ● ●●
● ● ● ●
Japan−Sweden

● ● ● ●
● ● ● ●●
0.6

●● ● ● ● ● ● ●● ● ●● ●● ●
● ●● ● ●● ● ●
● ● ● ●
● ●● ● ● ● ●
● ●
● ● ● ● ●
● ●● ● ●● ●● ●● ● ●● ●●●●●
●●●● ●
●●●
●
●●●● ●●●●● ●●● ● ●
0.0

● ● ● ● ●●●●● ●● ● ●
● ● ●● ●●
●●● ●● ●● ●●
● ●● ● ●
● ● ●●● ●
●●●
●●●●●●
●●●
●
●●
● ●●●●●
●●●
● ●●
●● ●
●●
●●●
●
● ●
●
●● ●●● ● ●●●●
●●
● ●●●●● ●●
●●●●● ●●
● ●
●●●●●●● ● ●
χ

● ● ●●● ● ●● ●● ● ● ●
● ●●● ● ●●● ●●
●●● ●●● ●● ● ●
● ●●● ● ●● ● ●●●●●
● ●●
● ●●●● ● ●●● ●●
●●●●
● ●● ●●
●●●●
●
●
●
●
●●
●
●● ●●● ●
●
● ●●
●
●
●●
●●
●
●●●● ●● ●
●●●● ●
●●●●●●●
●●
●
●●●●
●●
●
●
●
●●
● ●
●
●
●
●●●●● ●
●
●●
●●●●
●●●● ●●●● ●● ●
● ●● ● ● ● ●
●●● ●●
● ●●●●●●
● ● ●
●●●
●●●
● ●●
● ●● ● ● ●● ● ● ●●●●
● ● ●● ●
●
●●● ● ●● ●
●●● ● ●●●
●● ● ● ●● ●● ● ● ● ●●●● ● ● ●●●●● ● ●● ● ● ● ●●
●
●● ●
● ● ● ●
● ● ● ● ●●●●●
0.4

● ● ●● ●
●●● ● ●● ● ● ●
● ● ●● ●● ●● ●● ●
●● ● ● ●
● ●
● ● ● ●
●
● ● ● ● ● ● ●
● ● ● ● ● ● ● ●
● ● ●
−0.5

● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ●● ●
● ● ●
● ●
●● ● ● ●
0.2

● ● ● ● ●
● ● ● ● ● ● ●
● ● ● ●● ● ● ●
● ● ● ●
● ● ● ●
● ●● ●● ● ●
● ●● ● ● ● ● ● ●●
● ●● ●● ● ● ● ● ●
●● ● ● ●●
●
● ● ● ● ●
● ●●● ●● ● ● ●
●
●● ●
● ● ●●
−1.0

● ●●
● ● ●
0.0

●
●

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

Canada−Sweden λ

The left-hand panel shows the scatterplot for the implied Canada–Sweden, Japan–Sweden data. The right-
hand panel shows the corresponding χ-plot.

The currency pairs Canada–Sweden and Sweden–Japan have strong positive dependence
and so we will use them to build our pair-copula construction. We need to select a copula
family that adequately represents the dependence between these pairs. Comparing the
empirical χ-plot for the pair Canada–Sweden with the copula families in the appendices
we can see that the Gumbel or Galambos families might be good candidates as their χ-
plots resemble the empirical one. We further explored this via simulation (and other more
formal techniques [6, 13]) and concluded that the Galambos family has features that are
not consistent with the empirical data. The Gumbel simulations provided evidence that
this family adequately describes the empirical data. The maximum likelihood estimate for
the parameter for the Gumbel family is 1.17. This estimate is based only on the bivariate
data Canada–Sweden and not on the full dataset. We will only use it as a starting point
for a full maximum likelihood estimation.

26
 A similar investigation between the Japanese and Swedish data led us to model
their dependence with a Frank copula. The maximum likelihood estimate of the parame-
ter is 3.45. Again this estimate will only be used as a starting point for the full maximum
likelihood estimation. We now have initial parameter estimates for the first tree in our
canonical vine decomposition. The second tree consists of two nodes and one edge. The
conditional pair-copula that we need to specify here is for the Canadian–Japanese given
the Swedish data. To this end we compute the implied observations for tree 2 from the
Gumbel and Frank copulas used in tree 1. Figure 6.3 shows the scatterplot and χ-plot for
the Canada–Sweden, Japan–Sweden implied data.

From the χ-plot in Figure 6.3 it is clear that the relationship Canada–Japan given
Swedish data should be modeled by the independence copula.

The canonical vine structure for the currency rate changes dataset is shown in Fig-
ure 6.4. These parameter estimates are only starting values to be used in a full maximum
likelihood estimation. Using Algorithm 3 with our dataset and our chosen pair-copulas
in an optimization routine to maximize the log-likelihood we arrive at the maximum like-
lihood parameter estimates in Table 6.1.

Table 6.1
Maximum Likelihood Parameter Estimates
Pair-copula Family ML estimate
Canada–Sweden Gumbel 1.11
Japan–Sweden Frank 1.62
Canada–Japan given Sweden Independent
Full model maximum likelihood parameter estimates for the canonical vine used to model the currency rate
changes dataset.

Figure 6.4
Initial Parameter Estimates for Maximum Likelihood
Gumbel Frank
ca sd jp
1.17 3.45

Independent
ca–sd sd–jp

The canonical vine used to model the Canadian (ca), Japanese (jp), and Swedish (sd) currency rate changes
along with the chosen pair-copula family and the individually estimated maximum likelihood parameters.
These parameters are only used as the starting values for the full model maximum likelihood estimation
procedure.

27
7. Illustration Based on Simulated Insurance Assets and Liabilities

In the context of enterprise risk management (ERM) one would like to either un-
derstand the dependencies among given risk factors or be able to simulate data that ex-
hibits specific dependence traits. In this section we will illustrate how one can use the
simulation algorithms of Appendix F to generate complex dependencies.

Suppose we would like to generate data for two asset classes (bonds and stocks)
and a liability portfolio (say, losses and expenses). For simplicity and because we want
to highlight the dependency structure we will work with uniform margins but extending
our discussion to other marginal distributions is straightforward. Let us use a simple D-
vine structure as our starting point. Figure 7.1 shows the copula families chosen and their
parameters.

Figure 7.1
Initial D-Vine Structure
Clayton Gumbel Frank
E L B S
0.50 1.43 −1.86

Independent Independent
EL LB BS

Independent
EB|L LS|B

This is a four-dimensional D-vine structure coupling one liability portfolio and two asset classes. The
liabilities are losses (L) and expenses (E) and the assets are bonds (B) and stocks (S). We have deliberately
kept this structure simple by choosing the independent copula for trees two and three. We have chosen to
model the dependence between the loss and expense with a Clayton copula. The asset classes are linked
via a Frank copula, and the losses are coupled with a Gumbel copula to the bonds. The parameters of these
copulas appear beneath the names. The independence copula does not have any parameters.

The copula used to link the expenses to the losses is Clayton with a parameter chosen
so that the Kendall’s τ between these two variables is equal to 0.2. Losses are linked to
bonds via a Gumbel copula with parameter chosen so that Kendall’s τ is equal to 0.3, and
finally bonds and stocks use a Frank copula with parameter −1.86; that is, the Kendall τ
coefficient is approximately equal to −0.2.

The left-hand panel of Figure 7.2 displays the D-vine structure, and the right-hand
panel shows scatterplots and corresponding χ-plots for a random sample of 200 points.
Notice how the χ-plots next to the main diagonal have the expected behavior given the D-
vine structure. The remaining three χ-plots also follow the assumptions of independence.

28
 Figure 7.2
Simulated Insurance Assets and Liabilities I
●●
●●● ● ● ●● ●●●● ● ● ● ● ●● ●●
●● ●● ●
●● ● ●●
● ● ●● ●●● ●● ●●
●
● ● ●
●●●
●●● ●
●●
●● ● ● ●● ●
●●
●
●●●
● ● ● ●● ●● ●● ● ●● ●
●●
●
● ● ●●
●● ●●● ●
● ●● ●
● ●●●●●
●● ●●
● ● ●● ●● ● ●●● ● ●●●
● ● ●● ● ● ●●● ●
●●●
●
●
●● ● ● ●● ●● ● ●● ● ●● ●● ●● ● ● ● ●● ●
●●● ● ● ●● ● ●● ● ● ●● ● ●● ● ● ● ●● ●●●● ●
● ● ●●●● ●
●● ● ● ●● ● ●●● ● ●●
● ● ●● ● ● ●●●
●●●● ●●
● ● ●● ●●
● ●
●●
● ●
● ● ● ●● ●●
●
●● ●●● ● ● ● ●●● ● ● ●●●● ●●●● ● ● ●● ●
● ●
● ●● ● ●●● ● ●
●
● ● ●●● ●
●
●●●
●
● ●● ● ●
●● ●
●●●
● ●
●
● ●
● ●●●● ●● ● ●
●● ●●
● ●
● ●● ●● ●●
● ●●
● ●● ● ●●
●
●
stocks
● ● ● ●
● ●● ● ● ●●● ●
●● ● ● ●
●● ●●
● ● ●● ● ● ● ●● ● ● ● ●●
●● ● ●●● ●●●●●●
●●● ● ●● ●
● ●● ● ● ●● ●●● ● ●● ●● ● ● ●● ●● ● ●● ●● ●
●●●●
● ●●
● ● ● ● ● ● ● ●● ●● ● ● ●● ●●●● ●●●
●● ●● ● ●● ● ●●● ●● ● ●● ● ● ● ●● ●
● ● ●● ● ● ● ●● ●●●● ● ●● ● ●
● ●● ● ●●● ● ●●
● ●
● ●● ●
●
● ●● ●●●● ● ● ●● ●●● ● ● ●●● ●●
●
● ● ●●
●●●● ● ●● ●
● ● ● ● ● ● ●
●●● ● ●●●● ● ● ●● ● ●● ● ●● ● ● ●● ●● ● ● ● ● ●

●● ●●● ● ●● ●● ● ●● ●●● ●
●
●●
Clayton Gumbel Frank ●●
● ●
● ●● ●
●
● ●
● ●●
● ●● ●
● ●●
●
●●●● ● ●
●●●●
●●
● ● ●● ●●
● ●●●● ●●
●●●●
● ● ●●● ● ●
●●
●
●●●
● ● ●●
E L B S ●
●●
● ●●
● ●●
● ● ●● ●● ● ●●
● ●●● ● ●●●●
●
● ●
●●● ●● ● ● ●●
●● ●
●● ●
0.50 1.43 −1.86 ● ●
●
●
●●●● ●●
●
● ● ● ●
● ●●●
●
● ●●
●
●●
●●
● ● ●
●●
● ●●● ●● ●●
● ●●
● ●● ●● ● ●●
●
●
●
●●● ● ●●● ●● ● ●● ● ● ●●● ●●● bonds ●
●
●
●●
●●
●●
●
●
●
●●
●●●
●●●
●
●● ● ●
●● ● ● ●●● ● ●●●
●●●● ●● ●
●
●●
●
●●
●
●
●
●●
●●
●
●
●●●
●
●●
●
●●
●●●
● ● ●●
●●
●
●
●
●●
●●
●
● ●● ● ●
●●●●●
●●●●
● ●
●●
●
●●●●●
●● ●●●●
●●●
●
●●
●●●
●●
●●
●
●
●●●
●●●● ●● ●
● ● ● ●● ●● ● ●● ●●●●● ●● ● ●
● ●●●
●●
●
●●●●
●●
●
● ●
●●●●
● ●●●
● ●
●●●● ● ● ● ● ● ●● ●● ● ● ●
●● ● ● ● ●● ●● ●
● ●● ● ●
● ●●●●● ●● ● ●
●●● ●●
●●●●● ●●●● ● ●● ● ● ●
●●●●● ● ● ● ●● ●
●
●
●●● ●● ● ● ●● ● ●●
●● ●●●●● ●● ●
Independent Independent ●
●●● ●
●● ● ●
● ●●
● ● ● ●
●●●
●
●
●
●
● ●
● ● ● ● ●

EL LB BS ●●
●
●
●● ● ●● ● ●●●● ●
●●● ● ●
●● ● ● ● ●●● ●●●●●● ●●
●●● ●●● ●●●
● ● ●
● ●●
● ●
●● ●● ●●● ●● ● ●
● ●●
●● ● ● ● ● ●●● ● ● ●
●●●●●●●
●● ●●● ●
● ●● ● ● ● ●● ●● ●●
● ●●●●
●● ●
●●●
● ●
●●
●●●●●● ●● ● ● ● ●●
●
●
● ●
●
●
●●
●
● ●●
●●●
●●
● ●●
●
●●
●
●
●●●●
●
●●
●
●●
●●●
●
●●
●
●
●●
●● ●
●
●●●●●
●
●
●●
●●
●
●●
●●●
●
●●
● ● ●● ●●●●●●● ●● ●
●●● ● ●●● ● ●● ● ●
●
●
●●
● ●●
●●●●
●
●●●● ● ●● ●
●●
●●
●
●
●●
●●● ●
●
●
●●
●●
●
●●
●
●●●
● ●●
●● ●
● ●
●●●●
●●
●
●●●●
●●
●
● ●●
●●
● ●●
●
●●
●●
●
● ●●● ● ● ● ●●● ●
●
●●
loss ● ●
●
●●●
●●
●
●
●●
●●
●●●
●
●
● ● ●●
●●●●
●
●
●●
●
●
● ●
●●
●
●●
●
●
●●●
●
●●
●● ●
●
●●
●
●
●
●●
● ●
●●
●
●
●●
●
●●
●●
●
●●
●
●
●
●
●
●
●●
● ● ●●● ● ●● ●●●●
Independent ●
● ●
● ●
●
●
●●
●
●●● ●
●●● ●
● ● ●
●
●
●
●
● ●● ●
EB|L LS|B ● ● ●●
●
●
●
●
●
● ●●●
●
●
●
●
●
●
●●●●●
● ●● ● ● ●

● ●
●
●●●
●
●●●●
● ●●
● ●●
●● ●
●
●
●●●
●●
●●
●●
●
●●
●
●●●●
●●
●● ●
●●
● ●●●
●●●
●●
●
●●
● ●
●●●
●
●
●●
●
●● ●
●●
●●●●●
●●
●
● ● ●●●●
●
●●
●
● ●
●
●●
●●
●●
● ●
●
●
●
● ●● ●
●●●
●●●●
●
●● ●●
●● ●●
●●
●●
●●
●●
●●
●●●●
● ●
●●●
●●
●●●●
●
●
● ●●● ● ● ●
●●
●
●●●
●
●
expense ●
●
●●
●●
● ●
●
●●
● ●
●
●●●●●● ●●
●
●●●●
● ●
●
● ●●●
● ●
●
●
●
●●
●●
●●
●
●●●●●
●● ●●●
●●● ●
●
●●
●● ●
●
●●●
●●●●●
● ●
●
●●
●●
●
●●●●
● ●
●
●●
●●●●
●
●●
●●● ●●
●● ●
● ●
●●
●
●
●●
●●
●
●●●●● ●
● ●
●
●●●
●●●
● ●●
● ●
●
●●●●
●●●● ● ●
●
●
●●
●
●●●
●
●
●●
●
●●
●●
●
●●
●
●
●●●
●●● ●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
●● ●●
●●●
●●●●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●●
● ●●
●● ●●●●

The left-hand panel shows the D-vine structure along with the parameters for each copula and the right-
hand panel has a random sample of 200 points from this D-vine. The upper portion of the graph show the
pairwise scatterplots and the lower portion displays the corresponding χ-plot.

In the next figure we change one of the parameters of our D-vine structure. For the Frank
copula (linking bonds and stocks) let us increase the strength of the dependence from
about a Kendall’s τ value of −0.2 to a value of −0.3. Figure 7.3 shows the new D-vine
structure and the simulated data.

Figure 7.3
Simulated Insurance Assets and Liabilities II
●● ● ●●●●● ●●● ● ●● ● ●● ●●●● ●●● ●● ● ● ●● ● ●● ● ●● ● ●
●
●●● ● ●●
●●
●● ● ● ● ●● ● ●● ● ●●● ●●●● ● ●
●●
●
●● ● ● ●
●
●●●● ●●● ● ● ● ● ● ●●● ● ●● ● ● ●
●● ●
● ●●● ●● ● ●●● ●
● ●●● ● ● ●● ●
● ● ● ●●
●
●
● ● ●● ●● ● ●
●●●●● ● ● ● ●● ● ● ●●● ●● ● ● ●
● ●
●●● ● ●● ●● ●● ● ●● ●● ● ●● ● ●● ●● ● ●●●
● ●●●● ●●● ● ●● ● ● ● ●●● ● ●●● ●●● ●●● ● ●● ● ●
●● ● ● ● ●
●●● ● ● ● ● ●● ● ● ●●●
● ●
●● ●●
●
●● ●
● ●● ●●
●●● ● ● ● ● ● ●● ● ●● ● ● ● ●●● ● ● ● ●● ●● ●●
● ● ● ● ●●● ● ●● ● ●● ● ●● ●●
● ●●●● ●● ● ● ● ●● ●
● ● ● ● ● ● ●
● ● ● ● ● ● ●
●●● ● ● ● ● ● stocks
● ●
● ●● ● ●● ●
●
●
●
● ●● ●● ● ●
●● ●●●● ●● ● ● ● ●●●● ● ● ●●●● ●● ●
●●●
● ● ●● ●● ●
●●
●● ●● ● ● ● ●●● ●● ●●●●● ●
● ●
●●●
● ●● ● ● ●●
● ●
● ●
●● ●● ● ●● ●● ●
● ● ● ●●●
●
● ● ● ●● ● ● ●
●● ● ● ●● ●
● ● ● ●●●● ● ●● ● ●● ●● ● ● ● ●●● ●
● ●● ●●● ●●● ● ●● ●● ●● ● ●
●
●
● ●● ● ● ● ● ● ●●
●
●●
●●●●
●● ● ●●● ● ●● ● ● ●●
●●● ●● ●●● ●●●●●● ●● ●
● ● ● ● ● ●●●
●
● ●
●●●
●
● ● ●● ● ● ● ● ●● ● ●● ● ●●●

● ●● ● ● ●● ● ● ●● ●●● ●● ●
●● ●● ●
●●● ● ● ●●
Clayton Gumbel Frank ●
● ●●
●●
● ●●● ●
●
● ●● ●
●●
● ●●
●
●
● ●
●● ●● ●
●
● ● ●●● ● ●● ●
●●●
● ●
●
●●●●

E L B S ● ●●
● ●● ●
● ● ●
● ●
●●
● ●● ●
● ●●
0.50 1.43 −3.00 ●
●
●●
●●● ●● ●●
●
●●● ● ● ●● ●●
● ●● ●
●
● ●● ● ●●
● ●● ●●
● ●●● ●●
●●
●
● ●●
● ● ●●
●
● ●●● ●●● ●
● ●●●● ●
●
●
●●
●●
●●● ●
● ●● ● ● ● ● ●●● ●
● ● ●
●●●●● ● ● ●● ●
●●● ●● ● ●
●● ●● ●
●
●●
bonds ●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●
●
●
●●●
●●
●●
●●●● ●●
● ●● ●
●
●
●●●
●
●
● ●● ●●●
●
●
●● ● ● ●● ●
●● ●
●●● ●
●
●
● ●●● ●
●● ● ●●●● ● ● ● ●●
●● ●
●●●
● ● ●
●
●●
●●
●
●●
●
●●
●● ●
● ●●●●●●●●●●
●
●●● ●
●●● ● ● ●
●● ●●● ●● ● ●●●●● ●● ●
●●
●
●●●●●
● ●
●
●●
●
●●
●●●
●●
●
●●
●
●●●
●
●
●
●●●
●
●
●● ●●
● ●● ● ● ●● ●● ●
●
● ●●●●●
● ●● ● ● ●●● ●● ●●
●
●
●●●
●●
●●
●
●●
●●
●
●●
●●●
●
●●
●
●
●
● ●●● ● ●● ● ●● ● ● ● ●●
●●● ● ●
● ● ● ●
●
● ●● ●●
● ●● ● ● ●
Independent Independent ●●●● ● ●
● ●●● ● ●
● ● ●
●● ● ● ●●
● ● ●●●● ●● ●●
● ●● ● ●
●
●
●
●● ●
EL LB BS ● ● ●● ●
●● ●●
● ●●●●
●
●● ● ●●●
● ●●
● ● ●● ●●● ●●● ●
● ●●●● ● ●●●●●●●●
●● ● ●
●● ● ●●● ●
●●● ● ●● ●
●●
●● ● ● ●● ● ●●● ● ●●●
●●●
●●
●
●
●
●●
●●●
●
●●
●
●●
●●
●
●●●
●
●
●●
●●
●
●
●
●
●
●●●●
●
●
●● ● ●
●●●●
●●●
●●●● ●
● ● ●●● ●
● ●● ●
● ●● ●● ●
●●●●●●●
●●● ●● ●
●
● ●●
●
●
●
●●●
●
●●
●
●
●
●●●●●
●
●●
●●
●
●
●
●
●
●
●● ● ●● loss ●
●
● ●●●
●●
●●
●
●
●●
●●
●
●●
●
●●
●
●
●
● ●●
●●
●●
●
●● ●●
● ● ●● ●
●●●● ●
● ● ● ●
●● ●● ● ●●
● ●
●●
●
●●
●●
●●
●● ●
●●●
●●●
●
●●
●●
●
● ●
●●●●
●●●
●●●
●●
●●
●●●
●●●
● ●
●
●●
●●
●●
●
●●
●●● ● ●●● ●● ●●● ● ●● ●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
● ●●●
●●
●●
●●
● ●
●● ●●●●
●●●
●●●
●●●●●
●
●
Independent ●● ● ●
●
● ● ● ●●
● ●● ●● ●●●●●●●●● ●
●●●● ● ●● ● ● ●● ● ●
EB|L LS|B ●●●
●
●●● ● ●
● ●● ●●
● ● ● ●●● ●
●●● ● ●● ● ●●
●●●●● ●●
●

●●●●
●●●●●●
●●●
●● ●●●
●
●●●
● ●
●●●
●●●●
●●
●●●●● ●
●●
●
●●● ●● ●
●
●●●●
●●
●
●
●
●●
●
●●●
●●
●
●●●●●
●●●
●
●
●●● ● ●●●
●●
●
● ●●●●●●●
●●
●●
●●
●●●●
●●●●
●● ●●●
●●
●
●●
●
●●
● ●●
●
● ●
●
●●
● ●●●●●●
●
● ●●●
● ●●●● ●●
●
● ●
●●
●● ●●
●
● ●
●●
● ●
●●
●●●
●
●●
●●
●●●●●
● ●●●
●●●
●
●●
●●●●●
● ●
●●●
●
●
●
●●
●
●●
●●●● ● ● ●
expense ●●
●●
●● ●
●●●●
● ● ●●
●
●●●●
●●
●
●
●
●
●
●
●
●●●
● ●
●
●●
●
●●
●
●●●● ●●
●
●● ●●
●●
● ● ●
●
●●●
●●
●●
●●
●
●
●●● ●●●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●●
●●
●
●
●
● ● ● ●●●
●
●●●
●●● ●● ●●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●● ● ●
●●●●●
●
●●●●
●
●
●
●
●●
●
●●
●
●●
●●●
●
●
●
●●●●
●●
●● ● ● ●●

The left-hand panel shows the D-vine structure along with the parameters for each copula. The right-hand
panel displays a random sample of 200 points from this D-vine. The upper portion of the graph shows the
pairwise scatterplots and the lower portion displays the corresponding χ-plots.

29
First notice that, as expected, the bonds–stocks χ-plot shows a stronger negative depen-
dence. Also this change has influenced the other χ-plots; in particular, the loss–expense
pair the conditional expense–stocks pair given loss and bonds.

Now let us make a change to one of the conditional copulas in our structure; that
is, one of the copulas linking two variables given others. In Figure 7.4 we have changed
the copula linking expenses to bonds given losses from independent to normal with a
parameter equal to 0.5. Again comparing this new simulated data to the one in Figure 7.2
we see that the loss–bonds dependence has increased as well as the loss–stocks given
bonds.

Figure 7.4
Simulated Insurance Assets and Liabilities III
● ● ●● ● ● ●● ● ● ● ●● ●●● ● ●●● ● ● ●●●● ●●● ● ● ●●
● ●●
● ●●●● ● ●● ● ● ● ●● ● ● ●● ● ●● ●● ● ●
● ● ● ●● ● ●● ● ●● ●
● ●●
●
● ● ● ● ● ● ● ● ●● ●
●● ●●● ● ● ● ● ● ● ●● ●
● ● ● ● ●● ● ● ● ●●● ● ● ●●●
● ●● ●●● ●● ●●●
●●● ● ● ● ● ● ●
●●●● ● ● ● ● ●●●●●●● ● ●●
● ● ●
● ●●
●● ● ●● ●● ●
● ●
● ●
●●
● ●●
●●
●●● ●● ● ●
● ●●● ● ● ● ● ●● ● ●●●● ● ● ●● ●●●●
●● ●●●● ● ● ●●●● ● ●●● ●● ● ●● ● ●● ● ●●
●● ●
●● ●● ●● ● ●
● ● ● ● ● ● ●● ● ● ● ●● ●●
● ●● ●●●●
● ● ●●
●
●●● ●●
●
●
●● ● ●● ●
●●
●●
●
●
●
● ●●●
● ●
● ●● ● ●●●● ● ● ●
● ● ● ● ● ● ● ●
●●●● stocks
● ●● ●● ● ● ● ●● ● ● ● ● ●● ● ●● ●
●● ● ● ● ● ● ●● ●
● ● ● ●●● ●● ●●● ● ● ● ● ●● ● ● ●
● ●●● ●● ●● ●
● ● ● ● ● ●● ● ●● ● ● ● ● ●●● ●●● ● ● ●●● ● ● ● ● ●● ●
● ●
●
●● ●
●● ● ●● ● ● ●●
● ●●●●●
● ● ● ●● ●●● ● ●● ● ● ●
● ●●● ● ● ●●●● ● ●●●
● ●●●●●
● ●● ● ●
●
●● ●● ●●●●
● ●● ●● ●● ● ● ●●
●● ● ●●●
●
●● ● ●●●●● ●●● ●● ●●● ●●
● ● ●● ● ●●● ● ● ●●●● ● ● ●● ● ●● ●●● ●
● ● ●
●●●
● ● ●● ●● ● ●● ● ● ●● ●●●

●● ●● ●● ● ● ● ● ●●●● ●● ● ● ●
●●●
●●
● ● ●●●
Clayton Gumbel Frank ●● ●
●●● ●●
●●
●●●●●●
● ● ● ●
●●●●●
●
●
● ●●
●● ● ●●
●●
● ●●●
● ●●●●● ●
●●
● ●●
●
●●
●
●●
●●
E L B S ●
●
●● ●
● ●● ●
● ●● ●●●
●● ● ●●●●
●●
0.50 1.43 −1.86 ●●● ●● ●●
● ● ● ●
● ●
●● ● ●●
● ●●● ● ●
●
● ●●
● ● ●
●● ●● ●
●●
●
● ●●●
● ●
●●● ●
●
● ●
●● ●●● ●●●●●●
●●
● ● ●●●● ●
●●
● ● ●●● ●●●
●● ●●
●●●● ● ● ●
●
●
●
●
●● ● ● ●● ●● ●● ● bonds ●
●
●
●
●
●
●
●
●
●●●
●
●●● ●●
●●
●
●●
●
●●
●● ● ● ● ●● ●● ● ●
●●●●● ● ● ●● ● ●
●
●
●
●
●● ●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●●● ● ●●
●
●●
● ●
●●
●●●
●
●
●
●
●
●●●●
●
●●
● ●● ●●● ● ● ●●● ● ●● ●●● ● ●
●●
●●
●
●
●
●●
●
●●
●
●
●●●
●●
●●●
●
●●●
●
●●●
●●●● ●● ●●
● ●
●
● ● ● ● ● ●●●●● ●●● ●● ●● ●●● ● ●
●
●
● ●
●● ●●● ● ● ●
●● ● ● ●●● ●● ● ●
●● ●●
●● ● ● ● ● ●●● ● ● ● ●●
● ●● ●
● ●● ●● ●
●●●● ●
●●●● ●● ● ●● ●●
● ●●
Normal Independent ●
●
●●
●●●●
●
●
● ●● ● ●● ●● ● ●
EL LB BS ● ● ●● ●● ●●
0.50 ●
● ● ●
●●●●●
●
●● ●● ●
●
● ●
●●
●●
●●● ●●
●●
● ● ●●● ●● ● ●●●●●●
●
●●●
● ●●●●
● ●●● ●● ● ●
●● ● ●●
● ●
●●●●●
● ●●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●●
●●●●
● ●●●
●
●● ●
●
●
●●
●
●
●●
●
● ● ●● ●●
● ●● ● ●
●●●● ●●
● ●
●
● ●●● ●●●●●●● ● ● ● ●
●● ● ● ● ●●
●
● ●
●
●
●
●
●
●
●
●
● ●●●● ●
●● ●●
●
●●●●
● ●●
●●
●●●●
●●
●
●●
●
●
●
●
●
●
●
●
●● ●●
●
● ● ●●● ●●●● ●
●
● ●
●
●
●
●●
●●
●●
● ●
●
●
●● ●●
●
●●●
● ● ● ●●● ● loss ●●● ●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●●●
●●●●
●
●●
●
●
●
●●
●●● ●
●●●●
●●●
●
●
●
●●
●
●
●
●
●●
●●●
●●
●
●●
●●
●● ●
●
●
●
●●
●
●●● ●●● ●●
● ● ●
●● ● ●
● ●
●●●
●
●●
●
●●
●
●●
●●●
●
●
●●●
●●
●
●●
● ●
●●●●●●
●●●●● ●
●●
Independent ●
● ●● ● ●
●● ●● ●
●
● ● ●
●
● ●
●
●●
● ● ●
EB|L LS|B ●● ●
● ●
●
●
● ●
● ●
●●
●●●●●● ●
● ●
●●●
●●
● ● ● ● ●● ●

●●
●●
●●
●●
●
●
●●●
●
● ●●●
●●●
● ● ●●
●
● ● ● ●
● ●
●
●●●●
●●
●
●
●
●
●●
●
●
●
●●●
●●●
●● ●
●
●
●●●●
●●
●
●●
●
● ●
●
● ●●●●
●●●●
●●●●
●
●●
●
●●●
●●
●● ● ●
●
●
●
●
●
●●●●
● ●●
●
●
●
●
●
●●●●
●
● ●●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●●
●
●●
●●
●●
●●
●
●●
●
●●
●
●
●●
●●●
●●
●
● ●●
●
●
●●
●●●●●
● ●
●● ●
●●●
●● ●●
● ● ●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●●●
●●
● ●
●●●
●●●
● ●
●
●●●
●
●●
● ●●
●●●
●●●●
●
●●
●
●●●
●●● ●
●
● ● ●●
●●
●●
●●
●
●● ●● ● ●●●
●
●
●
●
●
expense ● ● ●●●
●●
●●
● ● ●●
●
●
●●
●
●●
●
●
●●●
●
●
●
●
●
●
●●
●●●●
●
●●●●
●
●
●●
●
●●●●●
●
●
●●
●
●●
●
●●●
●
●●
●
●
●●
●●
●
●
●●
●●
●●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
● ●
●
●
●●
●
●●
●
●
●
●●
●●
● ●
●● ●●
● ●

The left-hand panel shows the D-vine structure along with the parameters for each copula, and the right-
hand panel has a random sample of 200 points from this D-vine. The upper portion of the graph show the
pairwise scatterplots, and the lower portion displays the corresponding χ-plot.

Finally, the last change involves the conditional copula linking expenses and stocks given
loss and bonds. We remove the independent copula and put in its place a normal copula
with parameter equal to 0.5.

In this last case, the bottom-right χ-plot shows the expected behavior. But the adja-
cent χ-plots do not differ significantly from those in the original Figure 7.2. Surprisingly
the loss–expense χ-plot now has a different feel to it. For λ near 1 we see increasing
spread. Also the dependence between bonds and stocks seems to be weaker.

As we gain more experience with the canonical and D-vines, the selection of pair-
copulas, and how they interact with each other, we will be able to build high-dimensional
models that better capture the most relevant aspects of the problem in question.

30
 Figure 7.5
Simulated Insurance Assets and Liabilities IV
● ●● ● ●●
● ● ● ● ●● ●
●
● ● ● ●● ●● ● ●●
● ●●●● ●●● ● ●● ●●● ●●●
●● ●
● ● ● ● ● ● ●●● ●● ● ●● ●● ●●
● ●● ●
●●● ●●● ●●● ●
●● ● ●● ●●
● ● ● ●● ●●● ● ● ●● ● ●
● ● ● ●●● ●●●●
● ● ●● ● ● ● ●●● ● ● ●●●●● ●● ● ●
●● ● ●●●●●●● ●● ● ● ● ● ●● ●● ●
● ●●
●●●● ●●● ● ● ●● ●
●●●● ●●●● ●● ● ● ●● ●●● ● ●●
●●●●● ●● ●●●
● ●●● ●●● ●
●
●
●●
● ●
● ●
●● ● ●● ●
● ● ● ●●●
●
● ● ● ● ● ● ● ● ●
● ●●● ●
● ● ●● ●● ● ●●● ●● ● ● ●●
●● ● ●●● ● ● ●
●●● ●● ●●● ●●
●
● ●●●
● ● ● ● ●●●● ● stocks
●● ●●● ●● ● ● ●● ●● ● ● ●●● ● ●●● ● ● ●● ●● ●● ●● ●●●
● ●
● ● ●●●
●● ● ●● ● ● ● ●
● ● ● ● ●
● ● ● ● ● ●●● ●●●●
● ● ●●
● ●● ● ● ●
● ●●
● ●
●
●●●
●● ●●
●● ● ● ● ●●●●●●● ●● ●● ●● ●
● ●●●● ●●●● ●●●● ●● ● ●● ●●●●
●●●●●● ●●
●●●
●
● ●●
●● ●●
● ●● ●●● ● ●● ●
●● ● ●●● ●
● ●
●●
●● ●
●● ●
● ●● ●●●● ● ●●
●
●
● ●● ● ● ● ●● ●●●● ●● ●
●●
●●
●● ● ●
● ●●● ●● ● ●
● ● ●● ●● ●
● ●● ●
●●●● ● ● ●● ● ● ● ●●●● ● ● ●
●● ● ●
●● ● ● ● ●
●● ● ●● ● ● ●● ● ●● ● ● ● ●● ●●● ●● ●
● ●● ●● ●● ●● ●● ●●●●● ● ● ●●●●
●● ● ●●●●●●●
Clayton Gumbel Frank ● ● ●
●
●
●●●
●●
●●
●●●● ●● ●●●● ●●
●
●
●
●
● ●
●
● ●●
● ● ●●
● ●●●
● ● ●●
●● ●●● ● ●●●
●
● ●●
E L B S ●● ● ●● ●
● ● ● ●● ●●●
●● ●
●●●
●● ● ● ● ● ●
●
●●●● ●
●
● ●● ● ●●
0.50 1.43 −1.86 ●●
●
●●
● ●
●●● ●
● ●●●
●●
●
●
●
●
●
●
● ●
●
●● ●●●
● ● ●●
●
● ●● ●
●●
●
●
●
● ● ●●
● ●
● ●
●
● ●● ●
● ●● ●●
●●●
●
● ● ●●
●
● ●●
● ●● ●
●● ●● ● ●
●
●● ● ● ● ● ●●●●●● ● ● ●● ●
● bonds ●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●●● ●●
● ●
●
●●
●
●●●
●
● ●
●●●
●●●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●●
● ●●●●● ● ● ●● ●● ● ● ●● ●
●●● ●●●
●●
●●●
● ● ●●
● ●●●
●●
●●
●●
●●●
●
● ●● ●
●
● ●● ● ●● ●● ●
●●
●●● ●
● ●●● ●● ●● ●●●
● ● ●
●
●●
●●
●●
●●
●●●
●●
●●
●●
●
● ●
●●●
● ●● ● ● ●● ● ●● ● ●●●● ● ●
●
●●
● ● ● ●● ● ● ● ●
●
● ● ●● ● ● ●●●
●
●
● ●●● ●● ● ●● ● ● ● ●●●
●●
● ● ● ● ● ● ● ●●●● ● ● ●●●●●
● ● ●● ● ●● ●
● ●
Independent Independent ●● ● ●● ●
● ●
● ●● ●●
●●●●● ●
● ●
EL LB BS ●●● ●
●
●
● ●
●
●●
● ●
● ●●
●● ●● ● ●
● ● ●
● ● ● ●●
● ●●●●●●
●●
● ●●●● ●
● ●● ● ●
●● ● ● ●●● ●●● ● ● ● ●● ●
●●● ●
●
● ●● ● ● ●● ●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●●●●●●
●
●
●●
●●● ●
●
●●
●
●●
●
●
●
●●
●●●
●
●●●
●
●
●●●●● ● ●● ●
●●●● ● ●
●
●●●●●●
●● ●●
●
●●
● ●●
●●●
●
●●
●●
●●
●
●●
●●
●●●
●●
●●
●
●
●●
●
●● ● ●●● ●●● ●●
●●●● ●
●
●●●●
●
●
●●
● ●
● ●●●●●●
● ●
● ●●●●
●
● ●● ●●
●●●●
●● ●●
● ●●
●●●●
●
●●
●● ● ● ● ●●● ●● ●
● loss ● ●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●●
●●
●
●●
●●●
●
●
●
●●
●
●
● ●● ● ● ● ● ●● ● ●
Normal ●●
●●●●● ●
●●●●●●● ●● ● ●● ●
● ●

EB|L LS|B ●●
●●
●●
●
●●
●
●●
●
● ●●●
●
● ● ●
●
●● ●
●● ●●
● ●
0.50 ●● ●● ●
●
●
●● ● ● ● ● ●

●
●
●
●
● ●●●
●
●●●●●
●●
●●●●●
●●
●
●●
●
●
● ●
●●● ●
●●●
●
●
●●
●●
●
●●
●●●
●
●●●●
●● ●
●
●● ● ●
●● ●●
●● ● ●● ●
●●●●●
●●●
●● ● ● ● ●●●●● ●
●
●●● ●
● ● ●
●●
●
●●
●●●●
●●●
●● ●
●● ● ●●●
●●
●●
●●
●●●
●●
●
●
●●
●
●
● ●
●●
●● ●●
●●
●
●
● ●●
● ● ●●
●●●
●● ●
●
●●
● ●
●●●●
● ●
●●
●●●
●●
●●
●
●●●
●
●●
●●
●●
●●
●
●●
●
● ● ●
●●●●
●● ●
●●●●
●● ● ●
●●●
●●
● ●
●●
● ●
●
●●
●●
●
●●
●
●
●●
●●
● ●
●
●
●●
●
●●
●●
●
●
●●●
●
●●
●●
●●
●●
●●● ●●
●
●●●●●
● ●
●●
●
● ●
●●●●
●●
●●●
●●
●●●
●● ●●●
● ●
●
●●
●
●
●
●●
●
●
●●
●●
●●
●
●●
●●
● ●
●
●
●●
●●
●
●●● ●
●
●
●
●
●
●
●●
●
●●
●
●
● ●
●●
●
● ●
●
●●● ●
● ●●
● ● ●
●●
●●●
●
●●
●
●
●●
●
●
●
●
●●●
●
●
●●
●
expense ●
● ● ●● ●
●
●●
●
● ● ● ●●
●● ●●●
●
●●●
●
●●● ● ● ●● ●●
●● ●●
●
●
● ● ●●●
●

The left-hand panel shows the D-vine structure along with the parameters for each copula, and the right-
hand panel has a random sample of 200 points from this D-vine. The upper portion of the graph show the
pairwise scatterplots, and the lower portion displays the corresponding χ-plot.

31
References
[1] Aas, K., Czado, C., Frigessi, A., and Bakken, H. 2006. “Pair-copula constructions
of multiple dependence” Tech. Rep. SAMBA/24/06, Oslo, Norway: Norwegian
Computing Center.
http://www.nr.no/files/samba/bff/SAMBA2406.pdf
[2] Anscombe, F. J. 1973. “Graphs in statistical analysis” The American Statistician 27(1):
17–21.
[3] Brehm, P. J., et al. 2007. Enterprise Risk Analysis for Property & Liability Insurance
Companies. Guy Carpenter & Company.
[4] Bedford, T., and Cooke, R. M. 2002. “Vines: A new graphical model for dependent
random variables” The Annals of Statistics 30(4): 1031–1068.
[5] Charpentier, A., and Segers, J. 2006. “Lower tail dependence for archimedean copu-
las: Characterizations and pitfalls” Discussion Paper 29, Tilburg University, Center
for Economic Research.
http://ideas.repec.org/p/dgr/kubcen/200629.html
[6] Durrleman, V., Nikeghbali, A., and Roncally, T. 2000. “Which copula is the right
one?” Tech. Rep., Crédit Lyonnais.
http://ssrn.com/abstract=103245
[7] Embrechts, P., Lindskog, F., and McNeil, A. 2003. “Modelling dependence with cop-
ulas and applications to risk management” in Handbook of Heavy Tailed Distributions
in Finance, Rachev, S. T., ed. North Holland: Elsevier.
[8] Federal Reserve Bank of St. Louis.
http://research.stlouisfed.org/fred2
[9] Faivre, F. 2003. “Copula: A new vision for economic capital and applications to a
four line of business company” Tech. Rep., ASTIN Colloquium.
http://www.actuaries.org/ASTIN/Colloquia/Berlin/Faivre.pdf
[10] Fisher, N. I., and Switzer, P. 1985. “Chi-plots for assessing dependence” Biometrika
72(2): 253–265.
[11] Fisher, N. I., and Switzer, P. 2001. “Graphical assessment of dependence: Is a picture
worth 100 tests?” The American Statistician 55(3): 233–239.
[12] Fox, J. 1991. “Regression diagnostics: An introduction” Vol. 07–079 of Sage Uni-
versity Paper Series on Quantitative Applications in the Social Sciences. Newbury Park,
Calif.: Sage Publications.
[13] Frees, E. W., and Valdez, E. A. 1998. “Understanding relationships using copulas”
North American Actuarial Journal 2(1): 1–25.
http://www.soa.org/library/journals/north-american-actuarial-journal/1998/january/naaj9801_1.pdf
[14] Genest, C., and Favre, A.-C. 2007. “Everything you always wanted to know about
copula modeling but were afraid to ask” Journal of Hydrologic Engineering 12(4):
347–368.
http://archimede.mat.ulaval.ca/pages/genest/publi/JHE-2007.pdf

32
[15] Genest, C., and MacKay, J. 1986. “The joy of copulas: Bivariate distributions with
uniform marginals” The American Statistician 40(4): 280–283.
http://archimede.mat.ulaval.ca/pages/genest/publi/TAS-1986.pdf
[16] Genest, C., and Rivest, L.-P. 1993. “Statistical inference procedures for bivariate
archimedean copulas” Journal of the American Statistical Association 88: 1034–1043.
[17] Joe, H. 1997a. “Families of m-variate distributions with given margins and m(m−1)/2
bivariate dependence parameters,” in Distributions with Fixed Marginals and Related
Topics, Rüschendorf, L., Schweizer, B., and Taylor, M. D., eds.
[18] Joe, H. 1997b. Multivariate Models and Dependence Concepts, Vol. 73 of Monographs on
Statistics and Applied Probability. Boca Raton, Fla.: Chapman & Hall/CRC.
[19] Klugman, S. A., and Parsa, R. 1999. “Fitting bivariate loss distributions with copu-
las” Insurance: Mathematics and Economics 24(1–2): 139–148.
http://ideas.repec.org/a/eee/insuma/v24y1999i1-2p139-148.html
[20] Mari, D. D., and Kotz, S. 2004. Correlation and Dependence. London: Imperial College
Press.
[21] Nelder, J. A., and Mead, R. 1965. “A simplex algorithm for function minimization”
Computer Journal 7: 308–313.
[22] Nelsen, R. B. 1999. An Introduction to Copulas, Vol. 139 of Lecture Notes in Statistics.
New York: Springer-Verlag.
[23] Schweizer, B., and Sklar, A. 2005. Probabilistic Metric Spaces. Mineola, N.Y.: Dover
Publications.
[24] Sklar, A. 1959. “Fonctions dé repartition á n dimensions et leurs marges” Publ. Inst.
Stat. Univ. Paris 8: 229–231.
[25] Valdez, E. A., and Tang, A. 2005. “Economic capital and the aggregation of risks
using copulas” Tech. Rep., University of New South Wales.
http://www.gloriamundi.org/detailpopup.asp?ID=453058136
[26] Venter, G., Barnett, J., Kreps, R., and Major, J. 2007. “Multivariate copulas for finan-
cial modeling” Variance 1(1): 103–119.
[27] Venter, G. G. 2002. “Tails of copulas” Proceedings of the Casualty Actuarial Society
89: 68–113.
http://www.casact.org/pubs/proceed/proceed02/02068.pdf
[28] Venter, G. G. 2003. “Quantifying correlated reinsurance exposures with copulas”
Casualty Actuarial Society Forum: 215–229.
http://www.casact.org/pubs/forum/03spforum/03spf215.pdf

33
Appendix A. Bivariate Clayton Copulas

The bivariate Clayton copulas are given for 0 ≤ δ < ∞ by

C(u, v; δ) = (u−δ + v−δ − 1)−1/δ . (34)

The density function is

c(u, v; δ) = (1 + δ)[uv]−δ−1 (u−δ + v−δ − 1)−2−1/δ . (35)

The h and h−1 functions are

 −1−1/δ
h(u, v; δ) = v−δ−1 u−δ + v−δ − 1
 −1/δ
δ+1 −δ/(1+δ)

−1 −δ
h (u, v; δ) = uv +1−v .

The derivation of these formulas is in [1].

Figure A.1
Clayton Copula 0-Correlation
1.0

1.0

●● ● ●●
● ● ●
●
● ● ●
●
● ● ●
● ●
● ● ●
● ● ●
0.8

● ●● ●
●
●
0.5

● ●
●
● ● ●
● ● ●
●
0.6

● ● ●
●
● ●
● ● ● ●
●
● ● ●● ● ● ●● ● ●●
● ●●
●● ●●
● ●● ●
0.0

● ● ● ● ● ●
Y

● ●● ● ●●● ●
● ● ●● ● ●● ● ● ●●●● ● ●●● ● ● ●
● ● ● ● ● ● ● ●
● ● ●● ●● ●● ●● ● ●● ●
● ● ● ●
●●●
0.4

● ● ● ● ● ●● ●
● ● ● ● ●●● ● ●
●
● ●
●
● ● ●
−0.5

●
● ● ● ● ●
● ●
0.2

● ●
● ● ●
● ● ●
● ● ● ●
● ●
● ● ●
●
●
−1.0
0.0

● ● ●

0.0 0.2 0.4 0.6 0.8 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Clayton copula with a correlation coefficient of zero. The right-
hand panel shows the corresponding χ-plot. Notice that most of the points are within the 95 percent control
bands.

34
 Figure A.2
Clayton Copula 0.2-Correlation

1.0

1.0
● ●
● ●
●
● ● ● ●
● ●
● ●● ● ●
● ●
0.8

● ● ● ● ●
● ●
●

0.5
● ●

● ● ● ●
● ● ● ●
● ● ● ● ● ● ● ●
● ● ●●● ● ●● ● ● ●
0.6

●● ● ● ● ● ● ● ●
●
● ● ● ●● ● ● ●●● ● ●● ●
● ● ●● ● ● ● ● ●●●
● ●
● ● ●● ● ● ●●
● ● ● ●●
● ●● ● ● ● ●● ● ● ● ●●●●● ●
● ● ● ●●
●● ● ●

0.0
● ● ● ● ● ●
Y

χ
● ●
● ●
● ●
● ●
0.4

● ●
● ●
●
● ●
● ●● ●●
● ●● ●
● ● ●

−0.5
● ● ●●
0.2

● ● ●
● ●
●
●
●
● ●
●● ● ● ●
●
● ● ●
●

−1.0
●
0.0

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Clayton copula with a correlation coefficient equal to 0.2. The
right-hand panel shows the corresponding χ-plot.

Figure A.3
Clayton Copula 0.5-Correlation
1.0

1.0

● ●
● ●
● ● ●
●
● ●
● ● ● ●
● ● ●
●
0.8

●● ●
●●●●● ●●●● ● ●●
● ● ●●
●
● ● ●●● ●
0.5

● ● ●
●● ●● ●●●
●
● ●● ● ●● ●
● ● ● ●●● ● ●
● ● ● ● ● ●
● ● ● ● ●● ●● ● ● ●
● ●
● ● ● ●● ● ●
● ● ● ●
● ● ●
● ● ● ●
0.6

● ● ● ● ● ● ● ● ●
●
● ● ● ● ●●
● ●● ●●
● ● ●●
● ●
0.0

● ●
Y

● ● ● ● ● ●●
●
●
●●
0.4

● ●
● ● ●
● ●
● ●
●
●
−0.5

● ● ●
● ● ●
●
0.2

● ●
●● ●
● ●
● ● ●
●
● ● ● ● ●
● ● ●
●
●● ● ●
●
−1.0

●
0.0

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Clayton copula with a correlation coefficient equal to 0.5. The
right-hand panel shows the corresponding χ-plot.

35
 Figure A.4
Clayton Copula 0.9-Correlation

1.0

1.0
● ● ●●●● ●
● ●
●●
●
●
●●
●●●● ●● ● ●● ● ●
● ● ●●●
●●●
● ●● ●●
● ●● ● ● ●●●●●●
●● ●● ● ● ●● ●
● ● ●● ●●
● ● ● ● ● ●●● ●
● ●
● ● ●
● ● ●● ●
● ● ●● ● ●●●
0.8

● ●
●
●● ●●●

0.5
● ● ●
●
●
● ● ●●
● ●
● ● ●
● ●● ●
0.6

● ●
● ●
● ● ●
●

0.0
● ●
Y

χ
●
●● ●
● ●●
●● ●
●
0.4

● ●
●
●● ●
●● ● ●
●

−0.5
●
0.2

● ●●
● ●
●●●
●
●●●
●●
●●
●●
●
●●
●●

−1.0
0.0

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Clayton copula with a correlation coefficient equal to 0.9. The
right-hand panel shows the corresponding χ-plot.

36
Appendix B. Bivariate Frank Copulas

The bivariate Frank copula has parameter space 0 ≤ δ < ∞ and distribution func-
tion

C(u, v; δ) = −δ−1 log([η − (1 − eδu )(1 − eδv )]/η), (36)

where η = 1 − e−δ . The density is

δηe−δ(u+v)
c(u, v; δ) = , (37)
[η − (1 − eδu )(1 − eδv )]2

and the h and h−1 functions are

e−δv
h(u, v; δ) =
1−e−δ
+ e−δv − 1
1−e−δu
1 − e−δ
( ),
−1
h (u, v; δ) = − log 1 − −1 δ.
(u − 1)e−δv + 1

The derivation of these formulas is in [1].

Figure B.1
Frank Copula 0-Correlation
1.0

1.0

● ● ● ●
● ● ● ●
● ●
● ●● ●
●
●
●
●
0.8

●● ● ●
● ●
● ●
0.5

● ● ●
● ● ● ●
● ● ●
● ●
●
● ● ●
● ● ● ●
0.6

● ●
● ●
● ● ● ● ● ●
● ● ● ●
● ● ● ●● ● ●
● ● ● ● ● ● ●● ● ●● ●
● ● ●● ● ●● ●
0.0

● ● ● ● ● ●
● ●● ●● ● ● ●
Y

● ●
● ● ● ● ● ● ●● ● ● ●●
● ● ●● ● ● ●
● ● ● ●● ●
● ● ● ●● ● ● ●●● ●● ●● ●●● ● ●●
● ● ● ● ●
●
0.4

●
● ● ● ●
● ●
● ● ●
−0.5

● ● ●
●●
●
0.2

● ● ●
● ●
● ●
●●
●● ●
● ●
−1.0

● ●
0.0

●● ●

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Frank copula with a correlation coefficient of zero. The right-
hand panel shows the corresponding χ-plot. Notice that most of the points are within the 95 percent control
bands.

37
 Figure B.2
Frank Copula 0.2-Correlation

1.0

1.0
●
● ● ●
● ●
● ●
● ●
● ● ●
0.8

● ●●
● ●

0.5
●
● ●
● ●
●
● ● ●
● ● ●
● ● ● ● ●●
●
● ● ●● ● ● ●●● ● ●●
● ● ●
● ● ● ●●● ●●●● ●
● ● ● ● ● ●● ●●
0.6

● ● ●● ●
● ● ● ● ● ●● ● ● ●●
● ● ●● ●● ● ● ●●● ●●● ●●
●
● ●
● ●●
● ● ● ●
●● ● ● ●● ● ●
● ● ●
● ● ●●
● ●

0.0
●
Y

χ
● ●
● ● ● ●
●
●
0.4

●
●
● ●
● ● ●● ● ●
●●
● ●● ●● ● ●

−0.5
●
●● ●
● ● ● ●
0.2

● ●● ●● ●
● ●
●●
●
● ●
● ● ● ●
● ● ●
● ●

−1.0
●
0.0

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Frank copula with a correlation coefficient equal to 0.2. The
right-hand panel shows the corresponding χ-plot.

Figure B.3
Frank Copula 0.5-Correlation
1.0

1.0

● ● ●
● ● ● ●
● ●
● ●● ●
● ● ●
●
● ● ● ●
●
0.8

● ●●●
● ● ● ●● ● ● ●
● ●
0.5

● ● ● ●●
● ● ● ●●●
● ● ●●●●
●
●● ●
● ● ●● ● ●
● ●
● ● ●● ● ● ● ●
● ● ● ●● ●● ●
● ●● ● ● ● ● ● ●
● ● ● ●
● ●●● ●● ●●●●●●
● ● ● ●● ●
●● ● ●
0.6

● ● ● ● ●●●● ● ●
● ● ● ● ● ●
●
●
●●
● ● ●
●●
0.0

● ●
Y

● ● ●
● ● ●●
● ● ●
●
0.4

● ● ● ●
● ●
●
● ● ●
●
●
−0.5

● ●
● ●
0.2

●
●
●● ● ●
● ●
●
● ● ●
●
● ●
−1.0

●
●
0.0

0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Frank copula with a correlation coefficient equal to 0.5. The
right-hand panel shows the corresponding χ-plot.

38
 Figure B.4
Frank Copula 0.9-Correlation

1.0

1.0
● ●
●● ●
● ● ●●● ●
● ● ●● ●
●
●
● ●
● ●●●●●●●●
●
● ●
●●● ● ●● ●
●●● ●
● ● ● ● ●● ●● ●●● ●● ●
● ● ● ●● ●●●
● ● ●
● ● ●● ●● ●
●● ● ● ●
●●● ●
●●●
0.8

● ● ●
● ● ● ● ●
● ●●

0.5
●
● ●● ● ●
● ● ●
●
●
●● ● ● ●●●●●
● ● ●
●
●
● ● ● ●
● ●
0.6

● ●
●
●

0.0
● ●
Y

χ
● ● ●●
●
●
● ● ●●
●
●
0.4

● ●
●
● ● ●
●
● ● ●
●
● ●

−0.5
● ●
●
0.2

●
● ● ●●
●
●●
● ●
● ●● ● ●●
●● ●
●

−1.0
● ●
0.0

● ●

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Frank copula with a correlation coefficient equal to 0.9. The
right-hand panel shows the corresponding χ-plot.

39
Appendix C. Bivariate Galambos Copulas

Let ũ = − log u and ṽ = − log v and for 0 ≤ δ < ∞ the distribution function is
n o
C(u, v; δ) = uv exp (ũ−δ + ṽ−δ )−1/δ , (38)

and the density function is

  −1−1/δ  
c(u, v; δ) = C(u, v; δ)(uv)−1 1 − ũ−δ + ṽ−δ ũ−δ−1 + ṽ−δ−1
 
−δ −2−1/δ −δ −1/δ
   
−δ −δ−1 −δ
+ ũ + ṽ (ũṽ) 1 + δ + ũ + ṽ .

The conditional density function h(u, v; δ) is given by

C(u, v; δ)
!
  δ −1−1/δ
h(u, v; δ) = 1 − 1 + ṽ/ũ . (39)
v

The inverse function, h−1 , cannot be written down in closed form but we can find its
inverse numerically. The derivation of these formulas is in [1].

Figure C.1
Galambos Copula 0-Correlation
1.0

1.0

● ●
●●

● ● ● ● ●
●
●
●
●
0.8

●
● ●
● ●
● ●
● ● ●
0.5

● ● ●
●
● ●
● ●
●
● ● ● ●
● ●
●
0.6

●●
●● ● ● ● ●●
●
● ● ● ●
●
●
● ● ● ● ●
●● ●
●● ● ●
● ● ● ● ●● ●● ● ●● ● ● ●● ●●
●● ●
●● ● ● ● ● ●●●● ●
0.0

● ●●●● ● ●●● ● ● ● ●●●●●●

Y

●●
χ

● ● ● ●●
●●●● ●● ●
● ●
● ● ●
●● ●● ● ●●● ●
● ●● ●
●●
0.4

●
● ●

● ● ● ●
●
● ●●
−0.5

● ●
● ●
● ●
●
0.2

● ●
● ● ● ●
●
● ● ●●
● ● ●
● ● ●
●
−1.0

● ●
● ●●
0.0

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Galambos copula with a correlation coefficient of zero. The right-
hand panel shows the corresponding χ-plot. Notice that most of the points are within the 95 percent control
bands.

40
 Figure C.2
Galambos Copula 0.2-Correlation

1.0

1.0
●
● ● ●●
● ●
● ●
● ● ●
●
●
●● ●
● ●
0.8

● ●
● ●● ●
● ●●

0.5
●
● ● ●
●
● ● ● ● ●
● ●
● ● ● ●●● ●●● ● ●●● ●
●
●
●
●
● ●● ● ●● ●
● ● ● ●●●● ● ●●
0.6

● ●● ● ●● ●●●● ●● ●● ● ●●● ●● ●
● ● ●● ● ● ● ● ● ●●
●
●● ●
● ●
● ● ● ●● ● ●
● ● ● ● ● ● ●●●●
● ● ● ●
●

0.0
● ● ● ●
Y

χ
● ● ●
●
● ●
● ●
●
0.4

● ●● ●
● ●
● ● ● ●
●
● ●
●
● ● ●

−0.5
● ●
● ●
● ●
0.2

● ●
● ●
●

● ● ●
● ● ●● ●
●

−1.0
0.0

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Galambos copula with a correlation coefficient equal to 0.2. The
right-hand panel shows the corresponding χ-plot.

Figure C.3
Galambos Copula 0.5-Correlation
1.0

1.0

●
● ●
● ● ●●
●
● ●
● ● ●
● ● ● ● ●
●
0.8

● ●●● ●
● ● ● ● ●●
● ●
● ● ● ● ● ●● ● ●●
● ●●● ● ●
0.5

●
● ● ● ●● ● ● ●●●●●● ●● ●● ●●
●●●● ● ● ● ● ● ● ●
● ●● ●● ●
● ● ●●●
● ●●● ●●●●●
●
● ●● ●●●●
●
●● ● ● ● ● ●● ●● ●
● ● ●
0.6

● ● ●
● ● ● ●
● ●
● ●
● ● ● ●
0.0

● ●●
Y

● ● ●
● ● ●
●
● ● ●
0.4

● ●●●●
● ●
●
● ●
●
● ●
−0.5

● ● ●
● ● ● ●
0.2

● ●
● ●
●
● ●
●
● ●●
●● ●
● ●
● ● ●
−1.0
0.0

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Galambos copula with a correlation coefficient equal to 0.5. The
right-hand panel shows the corresponding χ-plot.

41
 Figure C.4
Galambos Copula 0.9-Correlation

1.0

1.0
● ●●● ●●●
●●
● ●● ●
●● ● ●● ●●●● ● ●●● ●
●●
●● ●●
●
●
●●●
● ●●
●● ●●● ● ●●●● ●● ●
● ●● ●●●●● ● ● ●●●
●● ● ● ● ● ● ●
●
●● ● ● ●● ● ●●
●● ●
●● ● ● ●
● ●
0.8

●●
●● ● ●
●● ●

0.5
● ●
● ●● ●
●
● ●
●●
●
● ●
● ●
0.6

● ●●●
●
●
●

0.0
● ●
Y

χ
●
● ● ●
● ●
0.4

●● ●
●● ● ●
● ●
●
● ● ●
●
●●

−0.5
● ●●
● ●●●
● ●
0.2

● ●
●●●● ●
● ●●●●
●
●

−1.0
●●
0.0

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Galambos copula with a correlation coefficient equal to 0.9. The
right-hand panel shows the corresponding χ-plot.

42
Appendix D. Bivariate Gumbel Copulas

Let ũ = − log u and ṽ = − log v. For 1 ≤ δ < ∞ the Gumbel copula distribution is

C(u, v; δ) = exp −(ũδ + ṽδ )1/δ

n o
(40)

and its density function is

(ũṽ)δ−1 h
δ δ 1/δ
i
c(u, v; δ) = C(u, v; δ)(uv)−1 (ũ + ṽ ) + δ − 1 . (41)
(ũδ + ṽδ )2−1/δ

The h function is
 δ !−1+1/δ
−1

δ δ 1/δ
 ũ
h(u, v; δ) = v exp −(ṽ + ũ ) · 1+ (42)
ṽ

and the function h−1 (u, v; δ) cannot be written in closed form; therefore, we need to use a
numerical routine to invert it. The derivation of these formulas is in [1].

Figure D.1
Gumbel Copula 0-Correlation
1.0

1.0

● ● ●
●
● ● ●
●
● ● ●
● ● ● ● ●
0.8

●
● ●
● ●
0.5

● ●
●
●
● ●
●
● ●
●● ●
0.6

● ● ●
● ● ● ● ●
● ● ● ●
● ● ● ●●
●
● ● ● ●● ●● ●●●●● ●●● ● ●
●
●● ● ●● ● ●
0.0

● ● ●●● ●
●● ● ●● ● ● ●
● ● ●● ● ● ●●
Y

● ● ●●
●●
● ●● ● ● ● ● ● ● ● ●●
● ●● ● ● ●●
●● ● ● ● ●● ● ●●●
● ● ● ● ● ● ●●
● ●● ●
● ● ● ●
0.4

● ●
● ●
●
●
● ● ●
−0.5

● ● ●

● ●
●
0.2

● ●
●● ● ● ●
● ●
●
● ●
● ● ●
● ●
● ●
−1.0

● ● ● ●
●
0.0

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Gumbel copula with a correlation coefficient of zero. The right-
hand panel shows the corresponding χ-plot. Notice that most of the points are within the 95 percent control
bands.

43
 Figure D.2
Gumbel Copula 0.2-Correlation

1.0

1.0
● ● ●
● ● ● ● ●
● ●
● ● ● ●
● ●
● ●
● ●
0.8

● ●
●

0.5
● ● ● ● ●
● ● ● ● ●
● ● ● ● ●●● ●
● ●● ● ● ● ●● ● ● ●
● ● ● ● ● ● ●● ●●●● ●● ● ● ●
● ● ●●
● ●● ● ● ●
● ● ●● ●
● ● ●● ●
0.6

● ● ● ● ●●
● ●● ● ● ● ● ●
●● ●● ● ● ● ● ●
● ●●
● ● ●
●● ● ● ●
●● ● ●
●●
●
●● ●
●
●●

0.0
●● ●
Y

χ
● ● ● ●
●
● ●● ●
●
0.4

● ●
● ●
●
●
● ●● ● ●
●● ● ● ● ● ●

−0.5
●
● ●
● ●
0.2

● ●
●
● ●
●● ●
●
● ● ●
● ● ●

−1.0
●
●● ●
0.0

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Gumbel copula with a correlation coefficient equal to 0.2. The
right-hand panel shows the corresponding χ-plot.

Figure D.3
Gumbel Copula 0.5-Correlation
1.0

1.0

●●
●
●
● ● ●
●
●●
●
● ● ●
● ● ●
● ● ● ●
●● ● ● ● ● ●
0.8

●● ● ●● ●
● ●●● ●● ●●● ●● ● ●●
● ● ●●
● ● ●●● ●●● ●
●
0.5

● ● ● ●
● ● ● ●●●●● ● ● ● ●● ● ●
●
● ● ● ● ●
● ●
● ● ● ● ●● ● ●
● ●● ●●●
● ● ● ● ●● ●
● ● ●
● ● ●
0.6

● ● ● ●
●
●
● ●● ● ● ● ●
●
● ● ●
0.0

●
Y

●
χ

●
● ●
0.4

● ● ●
●
● ●
● ●
● ●
●
● ●
●
−0.5

● ● ●
● ● ●●
●
0.2

●●
● ●
●
● ● ●
● ●
●●
●
● ● ● ● ● ● ●
● ●
−1.0
0.0

● ●● ●

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Gumbel copula with a correlation coefficient equal to 0.5. The
right-hand panel shows the corresponding χ-plot.

44
 Figure D.4
Gumbel Copula 0.9-Correlation

1.0

1.0
● ● ●● ●●
●● ●
●●●
● ●●●
●●●●● ● ●● ● ● ●
●●
●● ●●●●●
●● ● ● ●● ●●●
● ● ●
● ●
●●
●
● ●
●
●
●●
● ●●
●●●
●
●●●
●● ●● ●
● ● ● ●
●
●
●
● ● ●●● ●
● ●●● ● ● ● ●
●
●
0.8

●
●● ● ● ●●
●

0.5
● ● ●
●●●
●●
● ●●●
●
● ●
0.6

●● ●● ●
●
●

0.0
●
●
Y

χ
●●
●
●
●● ●
0.4

●
● ● ●
● ● ●●
● ● ●
● ●●
● ● ●

−0.5
● ●
●● ●
● ●
0.2

● ●
●● ●●
● ●
●
● ● ●●
●● ●
● ●
● ●
●

−1.0
0.0

●
●

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a Gumbel copula with a correlation coefficient equal to 0.9. The
right-hand panel shows the corresponding χ-plot.

45
Appendix E. Bivariate Normal Copulas

The bivariate normal copula is given by

C(u, v; δ) = Φδ (Φ−1 (u), Φ−1 (v)), (43)

where Φ is the standard normal distribution N(0, 1) with mean zero and unit variance,
Φ−1 is its inverse, and Φδ is the bivariate standard normal distribution with correlation δ.

The density is given by

δ2 (u2 + v2 ) − 2δuv
!
1
c(u, v; δ) = √ exp − , (44)
1 − δ2 2(1 − δ2 )

and the h and h−1 functions are

Φ−1 (u) − δΦ−1 (v)

!
h(u, v; δ) = Φ √
1 − δ2
 √ 
h−1 (u, v; δ) = Φ Φ−1 (u) 1 − δ2 + δΦ−1 (v) .

The derivation of these formulas is in [1].

Figure E.1
Normal Copula 0-Correlation
1.0

1.0

● ●
●
● ● ● ●
●
●
● ●
● ● ●
● ● ●
0.8

●
● ● ● ● ●
● ●
● ●
0.5

● ●● ●
● ●●
●
● ● ● ● ●
● ●● ●
● ● ●
0.6

●
● ● ● ●
●● ●● ●● ●●
● ● ● ● ● ●● ● ● ●
● ● ●
●
● ●
● ● ●● ●●●●●● ● ●●
● ● ● ●● ● ● ● ●●
●
●● ● ● ●●●● ●● ● ● ●
●●
● ● ● ●● ●
0.0

● ●● ●●● ●
Y

● ● ● ● ●
● ● ●● ● ●● ●●
● ●
● ● ●
0.4

● ● ● ●
●● ● ●
● ●
● ● ●
● ● ●
●● ● ● ●
−0.5

● ● ●
● ●
●
0.2

●●
● ● ● ●●
●
● ●● ●
●
−1.0

● ●
0.0

●●

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a normal copula with a correlation coefficient of zero. The right-
hand panel shows the corresponding χ-plot. Notice that most of the points are within the 95 percent control
bands.

46
 Figure E.2
Normal Copula 0.2-Correlation

1.0
● ● ● ●
● ●
● ●
●
●
● ● ●
● ●
0.8

● ●
●
● ● ● ●

0.5
● ●
●● ●
●
● ●
●● ●
● ● ● ● ●
0.6

● ●
● ● ●●
● ● ●● ● ●
● ● ●
● ● ●● ● ●●● ●●● ● ●●
●●● ●●● ● ● ●
● ● ● ●
●● ●● ● ● ● ●● ● ● ●●● ● ● ●
●
● ● ●● ●● ●●●●●

0.0
● ● ● ● ●
● ●●●●● ● ● ● ● ●
Y

● ●●

χ
● ● ● ● ● ● ● ●
●● ● ●●●● ●
0.4

● ●
●
● ● ● ●
● ● ● ●
● ● ●
●●
● ● ● ●
●

−0.5
● ● ●
●
●
0.2

●● ● ● ●
● ● ●
● ●
●
● ● ●
●
● ●
● ●

−1.0
0.0

● ●

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a normal copula with a correlation coefficient equal to 0.2. The
right-hand panel shows the corresponding χ-plot.

Figure E.3
Normal Copula 0.5-Correlation
1.0

● ● ●

● ● ●
● ● ●●
● ●
●
0.8

● ● ● ●
●
● ● ● ●
● ●
● ●●● ● ●
0.5

● ● ● ● ● ● ● ●●● ● ●● ●● ● ● ● ● ●
● ● ● ● ● ●●● ● ●
● ● ● ●●● ●● ●● ● ●
● ● ● ● ● ●● ● ●
● ● ●●●● ●●
● ●● ● ●● ●● ●●●
●●
● ● ● ● ●● ●
●
●● ● ●● ●●
0.6

● ● ●
● ●
● ●●
● ●● ●●
●
●
0.0
Y

●
● ● ●
●
● ●
0.4

● ● ● ●
● ● ● ●
●
●● ● ● ●
●
● ●

● ●
−0.5

● ●
● ●
● ● ●
0.2

●● ● ●
● ●
● ●
● ● ● ●
● ● ●
●
● ●
● ●
●
−1.0
0.0

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a normal copula with a correlation coefficient equal to 0.5. The
right-hand panel shows the corresponding χ-plot.

47
 Figure E.4
Normal Copula 0.9-Correlation

1.0

1.0
● ●●
● ●
● ●
●
● ●
● ●●●●● ● ●
● ● ●
● ●●
● ●●● ●
●● ● ● ● ● ●
● ● ● ● ●●●●●●
● ● ● ●●●●●
0.8

● ● ●● ● ●●● ●●●●
●
● ● ● ●
●
● ● ● ● ● ●●● ●
● ● ●● ●●
● ●●●

0.5
● ●
● ●● ● ●● ● ●●
●● ●
● ● ● ● ● ●●
● ●● ● ●
●
● ● ● ●
● ●
● ● ●
●
● ●● ●
0.6

●
● ● ●●
●

0.0
● ●
● ●
Y

χ
● ● ●
● ●
● ●
● ●
0.4

● ●
● ● ●

● ● ●
●

−0.5
●
● ● ● ●
● ●●
● ●
● ●●
0.2

●
● ●
●
● ●
●
●● ●

−1.0
●●● ●● ●
●
●
0.0

0.0 0.2 0.4 0.6 0.8 1.0 −1.0 −0.5 0.0 0.5 1.0

X λ

The left panel displays the scatterplot of a normal copula with a correlation coefficient equal to 0.9. The
right-hand panel shows the corresponding χ-plot.

48
Appendix F. Canonical and D-Vine Algorithms

Algorithm 1 Simulation for a Canonical Vine.

This algorithm will generate a sample x1 , x2 , . . . , xn from a canonical vine. The parameter Θ j,i represents
the necessary parameters for the copula c j,j+1|1,...,j−1 used in the construction.

For i = 1, 2, . . . , n, let wi be independent uniform random numbers on [0, 1].

x1 ← v1,1 ← w1
3 for i ← 2, 3, . . . , n
vi,1 ← wi
for k ← i − 1, i − 2, . . . , 1
6 vi,1 ← h−1 (vi,1 , vk,k ; Θk,i−k )
end for
xi ← vi,1
9 if i = n then
Stop
end if
12 for j ← 1, 2, . . . , i − 1
vi,j+1 ← h(vi,j , v j,j ; Θ j,i− j )
end for
15 end for

49
Algorithm 2 Simulation for a D-Vine.
This algorithm will generate a sample x1 , x2 , . . . , xn from a D-vine. The parameter Θi,j represents the
necessary parameters for the i, j-th copula in the construction.

For i = 1, 2, . . . , n, let wi be independent uniform random numbers on [0, 1].

x1 ← v1,1 ← w1
3 x2 ← v2,1 ← h−1 (w2 , v1,1 , Θ1,1 )
v2,2 ← h(v1,1 , v2,1 , Θ1,1 )
for i ← 3, 4, . . . , n
6 vi,1 ← wi
for k ← i − 1, i − 2, . . . , 2
vi,1 ← h−1 (vi,1 , vi−1,2k−2 , Θk,i−k )
9 end for
vi,1 ← h−1 (vi,1 , vi−1,1 , Θ1,i−1 )
xi ← vi,1
12 if i = n then
Stop
end if
15 vi,2 ← h(vi−1,1 , vi,1 , Θ1,i−1 )
vi,3 ← h(vi,1 , vi−1,1 , Θ1,i−1 )
if i > 3 then
18 for j ← 2, 3, . . . , i − 2
vi,2 j ← h(vi−1,2 j−2 , vi,2 j−1 , Θ j,i− j )
vi,2 j+1 ← h(vi,2 j−1 , vi−1,2 j−2 , Θ j,i− j )
21 end for
end if
vi,2i−2 ← h(vi−1,2i−4 , vi,2i−3 , Θi−1,1 )
24 end for

50
Algorithm 3 Log-likelihood for Canonical Vine.
Evaluation of the log-likelihood function for a canonical vine.

log-likelihood ← 0
for i ← 1, 2, . . . , n
3 v0,i ← xi
end for
for j ← 1, 2, . . . , n − 1
6 for i ← 1, 2, . . . , n − j
log-likelihood ← log-likelihood + `(v j−1,1 , v j−1,i+1 ; Θ j,i )
end for
9 if j = n − 1 then
Stop
end if
12 for i ← 1, 2, . . . , n − j
v j,i = h(v j−1,i+1 , v j−1,1 ; Θ j,i )
end for
15 end for

51
Algorithm 4 Log-likelihood for D-Vine.
Evaluation of the log-likelihood function for a D-vine decomposition.

log-likelihood ← 0
for i ← 1, 2, . . . , n
3 v0,i ← xi
end for
for i ← 1, 2, . . . , n − 1
6 log-likelihood ← log-likelihood + `(v0,i , v0,i+1 ; Θ1,i )
end for
v1,1 ← h(v0,1 , v0,2 ; Θ1,1 )
9 for k ←1, 2, . . ., n-3
v1,2k ← h(v0,k+2 , v0,k+1 ; Θ1,k+1 )
v1,2k+1 ← h(v0,k+1 , v0,k+2 ; Θ1,k+1 )
12 end for
v1,2n−4 ← h(v0,n , v0,n−1 ; Θ1,n−1 )
for j ← 2, 3, . . . , n − 1
15 for i ← 1, 2, . . . , n − j
log-likelihood ← log-likelihood + `(v j−1,2i−1 , v j−1,2i ; Θ j,i )
end for
18 if j = n − 1 then
Stop
end if
21 v j,1 ← h(v j−1,1 , v j−1,2 ; Θ j,1 )
if n > 4 then
for i ← 1, 2, . . . , n − j − 2
24 v j,2i ← h(v j−1,2i , v j−1,2i+1 ; Θ j,i+1 )
v j,2i+1 ← h(v j−1,2i+1 , v j−1,2i+2 ; Θ j,i+1 )
end for
27 end if
v j,2n−2 j−2 ← h(v j−1,2n−2 j , v j−1,2n−2 j−1 ; Θ j,n−j )
end for

52