CS2B-17: Copulas – Course Notes Page 1

17
Copulas

Course Notes
Syllabus objectives

1.3 Introduction to copulas

1.3.1 Describe how a copula can be characterised as a multivariate distribution

function which is a function of the marginal distribution functions of its
variates and explain how this allows the marginal distributions to be
investigated separately from the dependency between them.
1.3.2 Explain the meaning of the terms dependence or concordance, upper and
lower tail dependence; and state in general terms how tail dependence can
be used to help select a copula suitable for modelling particular types of
risk.
1.3.3 Describe the form and characteristics of the Gaussian copula and the
Archimedean family of copulas.

The Actuarial Education Company © IFE: 2019 Examinations

 Page 2 CS2B-17: Copulas – Course Notes

0 Introduction

0.1 Contents
There is no material in the Core Reading for this chapter that relates specifically to R.

Therefore we do not provide detailed Course Notes for this topic. Instead, we have provided an
exam-style worked example below that covers some areas that we think could be examined in the
practical exam.

0.2 Data requirements

These exercises do not require you to download any data files.

© IFE: 2019 Examinations The Actuarial Education Company

 CS2B-17: Copulas – Course Notes Page 3

Copulas – Exercise
Parts (i) and (ii) of this question involve simulating the values from a bivariate copula. You are not
expected to know how to do this for Subject CS2. However, the method we have used here
provides good practice with using some of the features of R. We thought this was more useful
than just giving you a table of pre-calculated values to read in and use for the graphs and
correlation calculations in parts (iii) to (v).

Ex 1 One way to simulate pairs of values from a bivariate copula is to use an acceptance / rejection
approach. Here we generate a large number of pairs of independent uniform random numbers,
then we use another random number to ‘accept’ (ie retain) or ‘reject’ each pair with the correct
probability. The probability is based on the probability density of the copula. However, this
method is sometimes not very efficient in terms of the number of pairs that are rejected in the
process.

The Clayton copula is defined by the function:

 
1 
C (u,v)  u  v   1 where   0 .

(i) Show that the probability density for this copula is:

 
2C 1  2
 (  1)u 1v  1 u  v   1 [6]
uv

(ii) (a) Generate 1,000 pairs of pseudo-random values (u,v) where u and v are
independent values from the U(0,1) distribution, using a ‘seed’ value of 17.

 2C
(b) Create a vector (also with 1,000 elements) of the values of with   2 for
uv
each of the pairs (u,v) .

(c) Rescale your vector in (ii)(b) so that all the values lie in the range [0,1] by dividing
the vector by the maximum value.

(d) Create a subset of the pairs in (ii)(a) where each pair is included or excluded with
the rescaled probabilities that you calculated in (ii)(c).

Hint: Create another vector of 1,000 pseudo-random U(0,1) values and ‘accept’
each pair if the random number is less than the rescaled value in (ii)(c).

(e) Count the number of pairs that have been ‘accepted’ using this algorithm and
comment on the efficiency of this method. [20]

(iii) (a) Create a scatterplot of the values of the subset you obtained in (ii)(e).

(b) Comment on the distribution of the points in your scatterplot. [12]

The Actuarial Education Company © IFE: 2019 Examinations

Page 4 CS2B-17: Copulas – Course Notes

(iv) (a) Calculate the Pearson, Spearman and Kendall coefficients of correlation for the
points on your scatterplot, quoting your answers to two decimal places.

(b) Comment on your answers to (iv)(a).

Hint: You are given that, with the Clayton copula, the theoretical value of Kendall’s

coefficient of correlation is Kendall  and the theoretical value of
 2
Spearman’s coefficient of correlation when   2 is  Spearman  0.68 . [12]

(v) (a) Use your subset of points from (ii)(b) to generate a set of pseudo-random pairs of
values (x , y) where x and y are values from a N(0,1) distribution connected by
the Clayton copula with   2 .

(b) Plot the values you obtained in (v)(a). [6]

[Total 56]

© IFE: 2019 Examinations The Actuarial Education Company

 CS2B-17: Copulas – Course Notes Page 5

Copulas – Solution
Ex 1 (i) Probability density

Differentiating first with respect to u gives:

 
C   1 
 u  v   1
u u

 u  v   1 u 1 
1 1  1



 
1  1
 u 1 u  v   1

Now differentiating with respect to v gives:

 
 2C    1  1  1 
 u u  v   1 
uv v  

   
 1  1  2
    1  u 1 u  v   1  v  1
  

 
1  2
 (  1)u 1v  1 u  v   1

(ii)(a) Generate pseudo-random values (U ,V )

We start by setting the seed value specified in the question:

set.seed(17)

We can now generate 1,000 pairs of independent pseudo-random numbers from a U(0,1)
distribution:

n=1000;u=runif(n);v=runif(n)

We can then put the values of our pairs in a table (matrix):

data=matrix(c(u,v),ncol=2)

head(data)

[,1] [,2]
[1,] 0.1550508 0.97418733
[2,] 0.9683788 0.71681979
[3,] 0.4682631 0.40560335
[4,] 0.7768197 0.10374040
[5,] 0.4078857 0.41535676
[6,] 0.5387971 0.08862091

The Actuarial Education Company © IFE: 2019 Examinations

Page 6 CS2B-17: Copulas – Course Notes

(ii)(b) Create a vector of values of the probability density

When   2 , the copula density we found in part (i) is:

2C
 
2.5
 3u3v 3 u2  v 2  1
uv

We can define a function density that operates on a pair of values x  (u,v) and returns the
value of the copula density:

density=function(x)
3*x[1]^(-3)*x[2]^(-3)*(x[1]^(-2)+x[2]^(-2)-1)^(-2.5)

The x[1] and x[2] in this definition refer to the two components of the vector x.

We can then apply our function to each row of the matrix:

clayton=apply(data,1,density)

Here we are applying the function to each row of a matrix. This is done using the apply function
and the argument 1, indicating that the function is to be applied across each row. (To apply a
function down each column, we would use the argument 2.)

head(clayton)

[1] 0.07775734 1.56096779 1.51794846 0.06767137 1.60684156 0.14363807

(ii)(c) Rescale the values

The maximum value is:

maxval=max(clayton)
maxval

[1] 25.9113

We can now divide the whole vector by this value:

probs=clayton/maxval

head(probs)

[1] 0.003000904 0.060242742 0.058582488 0.002611655 0.062013157

0.005543453

(ii)(d) Create a subset of the pairs with the Clayton probabilities

Following the hint, we can generate another vector of 1,000 uniform random numbers, which we
will use to make the decision whether to include each row:

random=runif(n)

Our outputs will look neater if we add the new columns we have calculated to our table:

fulldata=cbind(data,probs,random)

© IFE: 2019 Examinations The Actuarial Education Company

CS2B-17: Copulas – Course Notes Page 7

Just typing the cbind( ) command by itself would achieve the same objective of adding the
extra columns. However, R then automatically prints out the whole table (1,000 rows)! So we’ve
included the fulldata= to assign the answer to a variable and suppress this output.

Let’s label the columns in our table and see how it’s looking:

colnames(fulldata)=c("u","v","probs","random")

head(fulldata)

u v probs random
[1,] 0.1550508 0.97418733 0.003000904 0.07467454
[2,] 0.9683788 0.71681979 0.060242742 0.67884647
[3,] 0.4682631 0.40560335 0.058582488 0.32032500
[4,] 0.7768197 0.10374040 0.002611655 0.37374839
[5,] 0.4078857 0.41535676 0.062013157 0.10154273
[6,] 0.5387971 0.08862091 0.005543453 0.88611921

We can now select all the rows where the rescaled value of the copula density is less than the
random number. This will include (or exclude) them with the correct probability.

subset=fulldata[random<=probs,]

head(subset)

u v probs random
[1,] 0.5754613 0.4629139 0.05298045 0.014864425
[2,] 0.1188054 0.2178978 0.08469126 0.068626715
[3,] 0.3267991 0.6934728 0.02822582 0.019385549
[4,] 0.7242487 0.3391443 0.02735654 0.007623463
[5,] 0.3472285 0.3843515 0.06567128 0.008908249
[6,] 0.8366818 0.5159717 0.04017116 0.016946312

(ii)(e) Efficiency

We can check the size of the subset:

dim(subset)

[1] 41 4

So only 41 of the original 1,000 records have been included in the subset.

So this method is not very efficient as we have rejected almost 96% of the pairs that we
generated initially.

These values can differ quite a lot depending on the particular random values obtained using the
seed value chosen.

The Actuarial Education Company © IFE: 2019 Examinations

Page 8 CS2B-17: Copulas – Course Notes

(iii)(a) Scatterplot

We can now plot the pairs of values of (u,v) remaining in our subset as a scatterplot:

plot(subset[,1],subset[,2],xlab="u",ylab="v",las=1,
main="Simulation of Clayton copula with alpha=2")

We’ve included the las=1 option (label axis style) to display the numbers on the vertical axis
‘the right way up’, ie without rotation, but this is purely a matter of personal preference and you
wouldn’t have to do this in the exam.

Simulation of Clayton copula with alpha=2

1.0

0.8

0.6
v

0.4

0.2

0.0

0.0 0.2 0.4 0.6 0.8 1.0

© IFE: 2019 Examinations The Actuarial Education Company

CS2B-17: Copulas – Course Notes Page 9

(iii)(b) Comment on the distribution of the points

We can see the following features of the distribution of points:

 The points are definitely not distributed uniformly throughout the square (as they would be
with the independence / product copula).
 There are very few points in the top-left and bottom-right corners.
 The bulk of the points lie along the upward diagonal, which indicates a positive association
(correlation) between the values of u and v .
 The points in the bottom-left corner form quite a sharp angle, indicating that there is high
lower tail dependence.
 The points in the top-right corner are scattered widely, forming a blunt angle, indicating that
there is little or no upper tail dependence.

(iv)(a) Coefficients of correlation

To calculate the coefficients of correlation, we can use the cor function, specifying which
measure we wish to calculate. The default is Pearson. Note that, despite being names, the
options have to be entered with a small letter.

cor(subset[,1:2],method="pearson")

u v
u 1.0000000 0.6188123
v 0.6188123 1.0000000

This function actually gives the complete correlation matrix. The figure we want here for the
Pearson correlation is the 0.6188, or 0.62 to 2 decimal places.

Similarly:

cor(subset[,1:2],method="spearman")

u v
u 1.0000000 0.5714286
v 0.5714286 1.0000000

cor(subset[,1:2],method="kendall")

u v
u 1.0000000 0.4268293
v 0.4268293 1.0000000

So the Spearman correlation is 0.57 and the Kendall correlation is 0.43.

The Actuarial Education Company © IFE: 2019 Examinations

Page 10 CS2B-17: Copulas – Course Notes

(iv)(b) Comment on your answers

As expected, all three correlation coefficients are positive.

We can see that the Pearson and Spearman values (0.62 and 0.57) are quite similar, while the
Kendall value (0.43) is noticeably lower.

The Spearman value (0.57) is reasonably close to the theoretical value of 0.68, but there appears
to be quite a lot of random variation in the results. This is not surprising, given that the subset it
is based on contains only 41 pairs of values.

The Kendall value (0.43) is also reasonably close to the theoretical value of 0.5 (using the formula
given), but again there appears to be quite a lot of random variation in the results.

(v)(a) Pseudo-random pairs from a normal distribution

The values we have generated in our subset for u and v correspond to the percentile values
(ie the distribution function) for the two variables. We can now map these onto the correct
marginal distributions for x and y , which in this case is N(0,1) for both.

Recall that, to calculate the percentile values (quantiles) of a distribution in R, we prefix the
abbreviation for the distribution’s name with the letter ‘q’. So here we need:

qnorm(0.975)

[1] 1.959964

To apply this function to each of values in the vector of x values in our subset, we can use the
sapply function:

x=sapply(subset[,1],qnorm)
head(x)

Similarly for the y values:

[1] 0.1902958 -1.1809798 -0.4487693 0.5955101 -0.3928139 0.9809115

y=sapply(subset[,2],qnorm)
head(y)

[1] -0.09309544 -0.77931244 0.50571822 -0.41479962 -0.29407180

0.04004570

© IFE: 2019 Examinations The Actuarial Education Company

CS2B-17: Copulas – Course Notes Page 11

(v)(b) Plot

To see what they look like, we can plot the (x , y) pairs in a scatterplot.

plot(x,y,las=1,
main="N(0,1) values connected by Clayton copula with alpha=2")

Note that we don’t need to specify labels for the axes in this case, because R will by default use
the variables referred to in the arguments of the function, ie the “x” and “y” in this case.

N(0,1) values connected by Clayton copula with alpha=2

1
y

-1

-2

-2 -1 0 1 2

These 41 pairs of values (x , y) now each have a N(0,1) distribution but are connected via the
Clayton copula with   2 .

The Actuarial Education Company © IFE: 2019 Examinations