Tails of Copulas

Abstract: Actuaries who want to model correlated joint distributions have a choice of
quite a few copulas, but little basis for choosing one over another. Methods are provided
here to describe the features of different copulas, so that more informed choices can be
made.

Copulas differ not so much in the degree of association they provide, but rather in which
part of the distributions the association is strongest. Often needed for property and casu-
alty applications are copulas that emphasize correlation among large losses, i.e., in the
right tails of the distributions. Several copulas that do this are discussed.

To describe aspects of the copulas, univariate functions of copulas are introduced, for ex-
ample tail concentration functions. These descriptive functions can be thought of as an
intermediate step between correlation coefficients, such as Kendall, Spearman, Gini, etc.,
which are zero-dimensional measures of association, and the multi-dimensional copula
function itself.

The descriptive functions can be used to select copulas having desired characteristics,
such as tail concentration, and they can also be used in the fitting process to judge how
well the fitted copulas match those aspects of the data.
 Tails of Copulas

What are copulas?

Copulas provide a convenient way to express joint distributions of two or more random
variables. With a copula you can separate the joint distribution into two contributions: the
marginal distributions of each variable by itself, and the copula that combines these into a
joint distribution. One basic result is that any joint distribution can be expressed in this
manner. Another convenience is that the conditional distributions can be readily ex-
pressed using the copula.

Some measures of association depend only on the copula and not on the marginal distri-
butions. Both Kendall’s tau and Spearman’s rank correlation are examples, but the usual
Pearson linear product-moment correlation depends on the marginal distributions. Corre-
lation coefficients measure the overall strength of the association, but give no information
about how that varies across the distribution. Through the choice of copula, a good deal of
control can be exercised over what parts of the distributions the variables are more
strongly associated. One aspect emphasized below is controlling the strength of the rela-
tionship in the tails of the distributions. For instance, workers compensation and property
losses might be correlated in the extreme tails, but not elsewhere in the distributions, and
there are copulas with this kind of behavior.

A previous PCAS example of the use of copulas was provided in Wang [8], who provided
details of calculation methods for aggregate distributions, with some examples using
copulas.

Technically, copulas are joint distributions of unit uniform variates. In application, the
unit uniform variates are viewed as probabilities from some other variates. Then the joint
distribution of those variates is produced from those probabilities using their individual
inverse distribution functions. Copulas thus provide a ready method for describing joint
distributions and simulating correlated variables. Quite a few copulas are available, and

2
they have differing characteristics that lead to different relationships among the variables
generated.

This paper reviews several popular copulas, introduces some others, and also introduces
methods for selecting which copulas may be most appropriate for a given application. In
particular, the behavior of the copulas in the right and left tails can be used to distinguish
among joint distributions that produce the same overall correlation.

The organization of the paper is first to review copula methods in general, then to exam-
ine several specific copulas, and finally to look at measures that can be used to identify
key characteristics of copulas. An example is provided to illustrate how these measures
are applied to some correlated loss data.

1. General Considerations

Copulas – Formal Definition

It would be convenient to be able to express a joint distribution function F(x,y) as a func-
tion of FX(x) and FY(y), the individual (or marginal) distribution functions for X and Y, i.e.,
as F(x,y) = C(FX(x),FY(y)). To do this, C can be defined by C(u,v) = F(FX-1(u),FY-1(v)). This
gives C(FX(x),FY(y)) = F(FX-1(FX(x)),FY-1(FY(y))) = F(x,y). The function C(u,v) is called a cop-
ula. For many bivariate distributions, the copula form is the easiest way to express and
generate the joint probabilities. It allows a separate description of the individual distribu-
tions and their association. Copulas work in the multi-variate context also, but this paper
will primarily look at bivariate copulas, especially those defined by a single parameter.

In this context, a copula is a joint distribution of two unit uniform random variates U and
V with C(u,v) = Pr( U≤ u, V≤v). Also, c(u,v) will be used to denote the corresponding
probability density, which is the mixed second partial derivative of C(u,v). The simplest
copula is the uniform density for independent draws, i.e., c(u,v) = 1, C(u,v) = uv. Two
other simple copulas are M(u,v) = min(u,v) and W(u,v) = (u+v–1)+, where the “+” means
“zero if negative.” A standard result, given for instance by Wang[8], is that for any copula

3
C, W(u,v) ≤ C(u,v) ≤ M(u,v). M and W are called the Fréchet upper and lower bounds, re-
spectively.

Conditioning with Copulas

The conditional distribution can be defined using copulas. Let C1(u,v) denote the deriva-
tive of C(u,v) with respect to the first argument. When the joint distribution of X and Y is
given by F(x,y) = C(FX(x),FY(y)), then the conditional distribution of Y|X=x is given by:

FY|X(y) = C1(FX(x),FY(y))

For example, in the independent case C(u,v) = uv, the conditional distribution of V given
U=u is C1(u,v) = v = Pr(V<v|U=u). This is of course independent of u.

If C1 is simple enough to invert algebraically, then the simulation of joint probabilities can
be done using the derived conditional distribution. That is, first simulate a value of U, say
u, then simulate a value of V from C1, the conditional distribution of V|U=u.

Correlation
The linear correlation coefficient based on the covariance of two variates is not preserved
by copulas. That is, two pairs of correlated variates with the same copula can have differ-
ent correlations. However, the Kendall correlation, usually denoted by τ, is a constant of
the copula. That is, any correlated variates with the same copula will have the τ of that
copula.

There are different ways of defining τ, but the simplest may be τ = 4E[C(u,v)] – 1. For in-
dependent variates with C(u,v) = uv, E[C(u,v)] = ¼ so τ = 0. Also, for perfectly correlated
variates U = V, E[C(u,v)] = ½, so τ will be 1. Thus the scaling makes τ look like a correla-
tion coefficient. The key measure though is E[C(u,v)], which is a basic constant of a copula
and generalizes to the case of several variates. The limiting values are obtained for the
Fréchet upper and lower bound copulas, with τ = -1 for W and τ = 1 for M. These copulas
thus express complete negative correlation and complete positive correlation, respec-
tively.

4
 2. Some Particular Copulas
Some well-known copulas and a few designed particularly for loss severity distributions
are reviewed here.

Frank’s Copula
Define gz = e-az – 1. Then Frank’s copula with parameter a ≠ 0 can be expressed as:
C(u,v) = -a-1ln[1 + gugv/g1], with conditional distribution
C1(u,v) = [gugv+gv]/[gugv+g1]
c(u,v) = -ag1(1+gu+v)/(gugv+g1)2 and Kendall’s τ of
a
τ(a) = 1 – 4/a + 4/a2 ∫0 t/(et-1) dt

For a<0 this will give negative values of τ.

C1 can be inverted, so correlated pairs u,v can be simulated using the conditional distribu-
tion. First simulate u and p by random draws on [0,1]. Here p is considered a draw from
the conditional distribution of V|u. Since this has distribution function C1, v can then be
found as v = C1-1(p|u). The formula for this, which can be found from the formula for C1,
is:
v = -a-1ln{1+pg1/[1+gu(1–p)]}

Once u and v have been simulated, the variables of interest X and Y can be simulated by
inverting the marginal distributions, i.e., x = FX-1(u) and y = FY-1(v).

Gumbel Copula
This copula has more probability concentrated in the tails than does Frank’s. It is also
asymmetric, with more weight in the right tail. It is given by:
C(u,v) = exp{- [(- ln u)a + (- ln v)a]1/a}, a ≥ 1.
C1(u,v) = C(u,v)[(- ln u)a + (- ln v)a]-1+1/a(-ln u)a-1/u
c(u,v) = C(u,v)u-1v-1[(-ln u)a +(-ln v)a]-2+2/a[(ln u)(ln v)]a-1{1+(a-1)[(-ln u)a +(-ln v)a]-1/a}
τ(a) = 1 – 1/a

Unfortunately, C1 is not invertible, so another method is needed to simulate variates.

5
Embrechts, etal.[1] discuss the Gumbel copula and give a procedure to simulate uniform
deviates from a general class of copulas to which it belongs. For the Gumbel this proce-
dure starts by simulating two independent uniform deviates u and v, and then solving
numerically for 1>s>0 with ln(s)s = a(s–u). Then the pair [exp(ln(s)v1/a), exp(ln(s)(1-v)1/a)]
will have the Gumbel copula distribution.

Heavy Right Tail Copula and Joint Burr

For some applications actuaries need a copula with less correlation in the left tail, but high
correlation in the right tail, i.e., for the large losses. Here is one:
-a
C(u,v) = u + v – 1 + [(1 – u)-1/a + (1 – v)-1/a – 1] a>0
-a-1
C1(u,v) = 1 – [(1 – u)-1/a + (1 – v)-1/a – 1] (1 – u)-1-1/a
-a-2
c(u,v) = (1+1/a)[(1 – u)-1/a + (1 – v)-1/a – 1] [(1 – u)(1 – v)]-1-1/a
τ(a) = 1/(2a + 1)

The conditional distribution given by the derivative C1(u,v) can be solved in closed form
for v, so simulation can be done by conditional distributions as in Frank’s copula.

Frees and Valdez [2] show how this copula can arise in the production of joint Pareto dis-
tributions through a common mixture process. Generalizing this slightly, a joint Burr
distribution is produced when the a parameter of both Burrs is the same as that of the
heavy right tail copula.

Given two Burr distributions, F(x) = 1 – (1 + (x/b)p)-a and G(y) = 1 – (1 + (y/d)q)-a, the joint
Burr distribution from the heavy right tail copula is:

F(x,y) = 1 – (1 + (x/b)p)-a – (1 + (y/d)q)-a + [1 + (x/b)p + (y/d)q]-a

The conditional distribution of y|X=x is also Burr:

FY|X(y|x) = 1 – [1 + (y/dx)q]-(a+1), where dx = d[1 + (x/b)p/q]

6
By analogy to the joint normal, this can be called the joint Burr because the marginal and
conditional distributions are all Burr. In practice, the degree of correlation can be set with
the a parameter, leaving the p and q parameters to fit the tails, and b and d to set the
scales of the two distributions.

The Normal Copula

Useful for its easy simulation method and generalized to multi-dimensions, the normal
copula is lighter in the right tail than the Gumbel or HRT, but heavier than the Frank cop-
ula. The left tail is similar to the Gumbel.

To define the copula functions, let N(x;m,v) denote the normal distribution function with
mean m and variance v, N(x) abbreviate N(x;0,1), and B(x,y;a) denote the bivariate stan-
dard normal distribution function with correlation = a. Also let p(u) be the percentile
function for the standard normal, so N(p(u)) = u. Then with parameter a, which is the
normal correlation coefficient:

C(u,v) = B(p(u),p(v);a)
C1(u,v) = N(p(v);ap(u),1-a2)
c(u,v) = 1/{(1-a2)0.5exp([a2p(u)2-2ap(u)p(v)+a2p(v)2]/[2(1-a2)])}
τ(a) = 2arcsin(a)/π

The Kendall tau is somewhat less than a. The following table shows a few values.
a 0.15643 0.38268 0.70711 0.92388 0.98769
τ 0.10000 0.25000 0.50000 0.75000 0.90000

Simulation uses the conditional distribution C1. Simulate p(u) from a standard normal
and then p(v) from the conditional normal C1. The standard normal distribution function
can then be applied to these percentiles to get u and v.

7
 Visualizing Copulas
The copula densities can be graphed as surface plots, and these are somewhat informa-
tive, but to get a better feeling for what the copulas will do in practice it is helpful to look
at the joint distributions they produce from a standard sample distribution. The unit log-
normal (where ln(x) is standard normal) is used for this in the contour plots of the joint
densities for the copulas defined so far, using τ = .35.
Figure 1
Frank Joint Unit Lognormal Density Tau = .35
10

8.9

7.8

0.187-0.204
0.17-0.187 6.7
0.153-0.17
0.136-0.153
0.119-0.136 5.6
0.102-0.119
0.085-0.102
0.068-0.085 4.5
0.051-0.068
0.034-0.051
0.017-0.034 3.4
0-0.017

2.3

1.2

0.1
0.1 1.2 2.3 3.4 4.5 5.6 6.7 7.8 8.9 10

8
 Figure 2
Normal Joint Unit Lognormal Density Tau = .35
10

8.9

7.8

0.153-0.17
6.7
0.136-0.153
0.119-0.136
0.102-0.119
5.6
0.085-0.102
0.068-0.085
0.051-0.068
4.5
0.034-0.051
0.017-0.034
0-0.017
3.4

2.3

1.2

0.1
0.1 1.2 2.3 3.4 4.5 5.6 6.7 7.8 8.9 10

The normal and Frank copulas graphed in Figures 1 and 2 do not produce a strong rela-
tionship between large losses, although the normal shows a slightly stronger relationship,
whereas the Frank is stronger around the mode.

9
 Figure 3
Gumbel Joint Unit Lognormal Density Tau = .35
10

8.9

7.8

0.187-0.204
0.17-0.187 6.7
0.153-0.17
0.136-0.153
0.119-0.136 5.6
0.102-0.119
0.085-0.102
0.068-0.085 4.5
0.051-0.068
0.034-0.051
0.017-0.034 3.4
0-0.017

2.3

1.2

0.1
0.1 1.2 2.3 3.4 4.5 5.6 6.7 7.8 8.9 10

In contrast, the Gumbel copula keeps a strong relationship even for the large losses, as
seen in the higher values of the density function in the upper right of Figure 3.

10
 The heavy right tail (HRT) copula is even stronger in right tail correlation than is the
Gumbel. While difficult to see in Figure 4, it is also weaker in the left tail. This will be
more clear with the tail concentration functions discussed below.
Figure 4
HRT Joint Unit Lognormal Density Tau = .35
10

8.9

7.8

0.187-0.204
0.17-0.187 6.7
0.153-0.17
0.136-0.153
0.119-0.136 5.6
0.102-0.119
0.085-0.102
0.068-0.085
4.5
0.051-0.068
0.034-0.051
0.017-0.034
3.4
0-0.017

2.3

1.2

0.1
0.1 1.2 2.3 3.4 4.5 5.6 6.7 7.8 8.9 10

11
 Kreps’ Partial Perfect Correlation Copula Generator
A family of copulas has been developed by Rodney Kreps [6]. This is based on a method
for generating copulas that are mixtures of perfectly correlated and totally independent
variates This is easier to describe as a simulation procedure, and then look at the copulas.

The basic idea is to draw two perfectly correlated deviates in some cases and two uncorre-
lated deviates otherwise. More specifically, let h(u,v) be a symmetric function of u and v
mapping the unit square to the unit interval. To implement the simulation, draw three
unit random deviates u, v, and w. If h(u,v) < w, simulate x and y as FX-1(u) and FY-1(v) re-
spectively. Otherwise take the same x
Pareto(1,4) with h=(uv)^.3
5 but let y = FY-1(u) = x. Thus some draws
4.5
are independent and some are perfectly
4
3.5 correlated. The choice of the h function
3 provides a lot of control over how often
2.5
2 pairs will be correlated and what parts
1.5 of the distributions are correlated.
1
0.5 Figure 5
0 For instance, h can be set to 0 or 1 in
0 1 2 3 4 5
some interval like j < u,v < k to provide
independence or perfect correlation in that
Pareto(1,4) with h=(uv)^.3
10
interval, or it could be set to a constant p
1
to provide correlation in 100p% of the 0. 00001 0.0001 0.001 0.01 0.1 1 10
0.1
cases in that interval. Another choice is
h(u,v) = (uv)a. This creates more 0.01
correlation for larger values of u and v, 0.001
with the parameter a controlling how
0.0001
much more. Figure 6
0.00001

12
Figures 5 and 6 illustrate simulations in the case where h(u,v) = (uv)0.3 and both X and Y
are distributed Pareto with F(x) = 1 – (1 + x)-4. The correlated and uncorrelated instances
clearly show up separately, in either the log or regular scale.

For larger values of a, h(u,v) is smaller, so it is less likely that h(u,v) exceeds the random
value w and thus less likely that the case u=v will be selected. For small values of a, on the
other hand, h(u,v) will be larger, approaching one as a goes to zero. Thus h(u,v)>w is
more likely, so u=v will also be more likely. The partial perfect correlation copula genera-
tor thus provides a good deal of flexibility and control over how much correlation is in-
corporated and where in the distribution it occurs.

To describe the copulas that result, it will be convenient to adopt the notation used in
spreadsheets where a logical expression in parentheses will evaluate to zero if the expres-
sion is false and to one if it is true. Thus (u=v) is one if u=v and zero otherwise, etc.

PP Max Data Pairs t = .5

Although Kreps considers more general
1
situations, a relatively simple copula results
0.9
in the case where h(u,v) breaks out as a 0.8

product of a univariate function evaluated 0.7

at u and v, i.e., h(u,v) = h(u)h(v). If we 0.6

x 0.5
define H(x) = ∫0 h(t)dt, the copula formulas
0.4
become: Figure 7
0.3

0.2

C(u,v) = uv – H(u)H(v) + H(1)H(min(u,v)) 0.1

C1(u,v) = v – h(u)H(v) + H(1)h(u)(v>u) 0

0 0.2 0.4 0.6 0.8 1
c(u,v) = 1 – h(u)h(v) + H(1)h(u)(u=v)

For a concrete example, pick an a between zero and one, and let h(u) = (u>a). Thus if both
u and v exceed a, the simulated values of u and v will be identical, and otherwise they
x
will be independent. If x>a, H(x) = ∫a dt = x – a, and if not, H(x) = 0. Thus H(u) =

13
(u – a)(u>a). Also, H(1) = 1 – a, and H(min(u,v)) = [min(u,v) – a](u>a)(v>a). The copula
formulas above can then be computed directly for this h. The Kendall correlation is τ(a) =
(1 – a)4. Sometimes this copula is called PP max, for partial perfect max function. The scat-
ter plot of a simulated sample is graphed in Figure 7 for the case τ = ½.

PP (uv)^a Data Pairs τ =.5

Another example is to take h(u) =
1
ua. Then H(u) = ua+1/(a+1), and 0.9

H(1) = 1/(a+1). Here, 0.8

0.7
τ(a) = 1/[3(a+1)4] +
0.6
8/[(a+1)(a+2)2(a+3)]. As a
0.5
increases, this approaches zero, 0.4
reflecting the fact that selecting 0.3

u=v becomes less likely, and at a 0.2

0.1 Figure 8
= 0, τ =1, as this gives the perfect
0
correlation case. 0 0.2 0.4 0.6 0.8 1

Figure 8 shows simulated pairs for the case τ = ½. More correlated pairs occur at higher
values of u and v, as can be seen from the growing paucity of independent pairs when
going to the upper right.

3. Distinguishing among copulas

A few functions are introduced here to help illustrate different properties that can distin-
guish the various copulas. These functions can also be approximated from data, and so
can be used to assess which copulas more closely capture features of the data.

Tail Concentration Functions

Given a copula, right and left tail concentration functions can be defined with reference to
how much probability is in regions near <1,1> and <0,0>. For any z in (0,1) define:
L(z) = Pr(U<z,V<z)/z and R(z) = Pr(U>z,V>z)/(1 – z). In terms of the copula functions,
L(z) is just C(z,z)/z. To calculate R(z), note that 1 - Pr(U>z,V>z) = Pr(U<z) + Pr(V<z) -
Pr(U<z,V<z) = z + z – C(z,z). Then R(z) can be calculated by R(z) = [1 – 2z +C(z,z)]/(1 – z).
Also, note that Pr(U<z,V<z) = Pr(U<z|V<z)Pr(V<z). But Pr(V<z) is just z, as copulas are

14
 defined with uniform unit marginals, so L(z) = Pr(U<z|V<z) = Pr(V<z|U<z), and simi-
larly R(z) = Pr(U>z|V>z). Joe [4] uses the term “upper tail dependence parameter” for R

= R(1) = lim(z→1)R(z), and “lower tail dependence parameter” for L = L(0) =

lim(z→0)L(z).

The left tail function approaches unity for z near 1, so does not distinguish much between
copulas there, and similarly for the R function near 0. Thus they can be combined into an
LR function which is L below ½ and R above ½.. This is graphed in Figure 9 for the copu-
las discussed above and for the Clayton copula, a heavy left tailed copula discussed later.

LR Functions for Tau = .35 Figure 9

0.8

0.7

0.6

0.5

Gum
HRT
Frank
0.4
Max
Power
Clay
Norm
0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

15
A basic feature that distinguishes copulas in the right tail is whether R=R(1) = 0 or some-
thing greater. The Gumbel, HRT, and partial perfect copulas all have R>0. The HRT is
heavier in the right tail than the Gumbel, but less so than the partial perfect copulas. The
Clayton is the only copula here showing positive left tail dependence. The HRT and PP
Max copulas are very lowly dependent in the left tail. In fact, for the PP Max L(z) function
the variates are independent in the left tail. Thus for low z, L(z) = 1. The normal and
Frank copulas do not show tail dependence in the limits, but away from the extremes the
normal shows greater tail concentration than the Frank on both sides.

For the four copulas with R > 0, R is shown below:

Gumbel HRT PP Power PP Max

R 2 – 21/a 2– a 1/(1+a) 1–a

Since R and τ are functions of the same parameter, they can be viewed as functions of
each other. Once one is determined, the other is fixed for single-parameter copulas. Figure
10 graphs this relationship.

R as a Function of Tau Figure 10

1

0.9

0.8

0.7

0.6
Gumbel
HRT
0.5
R

Power
Max
0.4

0.3

0.2

0.1

0
0 0.2 0.4 0.6 0.8 1
Tau

16
A good starting point for choosing a copula would be to look at the target pair <τ,R> and
find which copula is closest. But since for the copulas above R is usually greater than
τ, lower values of R would not be matched by any of them. R is somewhat tricky to de-
termine for empirical data, as the far tail values have increasingly less data. Some projec-
tion of the lower values of R(z) might be necessary. Also the fitting should look at the R(z)
function, not just R.

Cumulative Tau
Other descriptive functions can be defined that show different aspects of copulas. The
cumulative tau function decomposes the integral defining the Kendall tau. Recall that tau
1 1
is defined as –1+4∫0 ∫0 C(u,v)c(u,v)dvdu. The cumulative tau can be defined as J(z) =
z z
–1+4∫0 ∫0 C(u,v)c(u,v)dvdu/C(z,z)2.

The full double integral is a probability weighted average of C(u,v), i.e., EC(u,v). To com-
pare to this on the square from (0,0) to (z,z), the partial integral has to be divided by the
weights, hence the first power of C(z,z) in the denominator. This quotient will give the
average value of C(u,v) in the square from (0,0) to (z,z). This will increase as a function of
z for any copula. The second C(z,z) divisor expresses this average relative to C(z,z), i.e.,
shows how the average C compares to the maximal C in the square. This may or may not
increase as a function of z, which makes it a more interesting property of the copula.

The normalization to the range of a correlation with the –1 and 4 is a matter of conven-
ience and familiarity, and gives J(1) = τ. The integration can be done numerically, al-
though for some copulas, formulas are given in Appendix A. The shape of the J function
depends on the copula and the tau. It is graphed for several taus for each copula in Fig-
ures 11 – 16. All the graphs end up at τ for z=1, but can start off high or low, and can in-
crease or decrease at varying rates.

Some other descriptive functions are discussed in Appendix B.

17
 Frank Cumulative Tau τ = .1, .5, .9 Gumbel Cumulative Tau τ = .1, .5, .9
1
1
0.9
0.9
0.8
0.8
0.7
0.7
Figure 11 Figure 12
0.6
0.6
0.5
0.5
0.4
0.4
0.3 0.3
0.2
0.2
0.1 0.1
0 0

Normal Cumulative Tau, t = .1, .5, .9 HRT Cumulative Tau τ = .1, .5, .9
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6 Figure 14
0.5 Figure 13 0.5
0.4 0.4
0.3 0.3

0.2 0.2

0.1 0.1

0 0

PP Max Cumulative Tau  = .1, .5, .9 PP Power Cumulative Tau τ = .1, .5, .9
1
1
0.9
0.9
0.8
Figure 15 0.8
0.7 Figure 16
0.7
0.6
0.6
0.5 0.5
0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0

18
 4. Flipping a Copula

The notation S(x) = 1 – F(x) is often used to describe the survival function Pr(X>x). The
joint survival function S(x,y) = Pr(X>x, Y>y) is not 1 – F(x,y), however, as that would be
the probability that either X>x or Y>y, but not necessarily both. In fact, S(x,y) = 1 – FX(x) –
FY(y) + F(x,y), i.e., Pr(X>x, Y>y) = 1 – [Pr(X<x) + Pr(Y<y)] + Pr(X<x, Y<y).

Similarly for a copula C(u,v) = Pr(U<u, V<v) the survival function of the copula, i.e.,
CS(u,v) = Pr(U>u, V>v), is CS(u,v) = 1 – u – v + C(u,v). Since C(FX(x),FY(y)) = F(x,y), we
have CS(FX(x),FY(y)) = S(x,y).

For a copula C, define CF(u,v) = CS(1 – u, 1 – v) = u + v – 1 +C(1 – u, 1 – v). Then

CF(SX(x),SY(y)) = CS(FX(x),FY(y)) = S(x,y). Note that CS is not a copula as it is zero at (1,1),
but CF is a copula. Call CF the flipped copula of C. When the flipped copula is applied to
the survival functions it gives the joint survival function for the copula. However, the
flipped copula can be applied to distribution functions, and then it can have quite differ-
ent properties than the original copula has. The next copula is an example.

Clayton’s Copula
This copula has a heavy concentration of probability near (0,0) so it correlates small losses.
It is not intuitively interesting for property-liability claims, but it may have some applica-
tion.

-a
C(u,v) = [u-1/a + v-1/a – 1] a>0
-a-1
C1(u,v) = u-1-1/a[u-1/a + v-1/a – 1]
-a-2
c(u,v) = (1+1/a)[uv]-1-1/a[u-1/a + v-1/a – 1]
τ(a) = 1/(2a + 1)

19
What is interesting here is that the heavy right tail copula is actually the flipped Clayton
copula. The tau is the same for both copulas1, and the tail concentration functions are
swapped. This is actually how the HRT copula was defined, and suggests defining other
copulas by flipping known copulas. The copula would have to have some asymmetry to
make this worthwhile. One candidate would be Gumbel’s copula.

The Flipped Gumbel

Gumbel’s copula is heavier in the right tail than the left. Flipping it would produce a cop-
ula with the opposite property:

C(u,v) = u + v – 1 + exp{-([-ln(1–u)]a +[-ln(1–v)]a)1/a}, a ≥ 1.

C1(u,v)=1–exp{-([-ln(1–u)]a +[-ln(1–v)]a)1/a}{[-ln(1–u)]a +[-ln(1–v)]a}1/a –1[-ln(1–u)]a-1/[1–u]
c(u,v) =(1–u)-1(1–v)-1{[-ln(1–u)]a +[-ln(1–v)]a}-2+1/a[ln(1–u)ln(1–v)]a-1 x
[a+{[-ln(1–u)]a +[-ln(1–v)]a}1/a –1]exp{-([-ln(1–u)]a +[-ln(1–v)]a)1/a}
τ(a) = 1 – 1/a

5. Applications

Loss Adjustment Expense

Two recent actuarial papers fit parameters to the joint distribution of loss and loss ad-
justment expense for a liability line using 1500 claims supplied by Insurance Services Of-
fice, Inc. The two studies may or may not have used the same data, but they present scat-
ter plots that are similar. They both use copulas to describe the joint distribution.

There were a couple of methodological differences between the two papers. Frees and
Valdez [2] assume Pareto marginals for both distributions, but compare fits for several
copulas. Klugman and Parsa [5], on the other hand, compare fits for a number of severity
distributions, but select Frank’s copula arbitrarily. The papers may have taken different
approaches to the censoring of claims by policy limits as well. Klugman and Parsa say

1 Tau for a sample is the average value of sign[(u – x)(v – y)] among all distinct pairs (u,v), (x,y).
This value is the same for the flipped pairs (1–u, 1–v), (1–x, 1–y), so tau will be the same for the
original and the flipped sample for any copula.

20
they omit claims for which either loss or expense is zero, so they can get true severity dis-
tributions for both. Frees and Valdez probably do this as well.

Frees and Valdez used the K(z) function discussed in Appendix B to select among copu-
las. Plotting the empirical K(z) against the values from several copulas, they found the
Gumbel looked best. The Gumbel also gave the best value for the Akaike information cri-
terion (AIC). Optimizing the AIC is equivalent to finding the copula with the highest
maximum likelihood in this case, as all the copulas they tried had one parameter. The best
fit they found was produced by the Gumbel copula with a = 1.453. This gives τ = 0.31.
Klugman and Parsa estimate the Frank a = 3.07438, which also gives τ = 0.31.

A convenient way to compare heavy-tailed severity fits is to look at the median and the
heaviness of the tail, which can be quantified as the smallest positive moment that does
not converge. For the Pareto, for example, this moment is just the shape parameter.

If we express the Pareto as F(x) = 1 – (1+x/b)–a , then Frees and Valdez find: for loss, a =
1.122, b = 14,036, and for expense, a = 2.118 and b = 14,219. Klugman and Parsa find the
best severity fits with the inverse Burr, which can be expressed as F(x) = (1+(x/b)–c)–a.
They estimate2 for loss, a = 1.046 = c, b = 11,577.7, and for expense, a = 1.57658, b =
10,100.2, c = 0.573534. These parameters are converted to median and tail heaviness ( = c
for the inverse Burr) below. There is reasonably close agreement among these values ex-
cept for the tail heaviness for loss expense, for which the divergence is a little greater.
Loss Median Loss Tail Expense Median Expense Tail
Frees & Valdez 12,000 1.12 5500 2.12
Klugman & Parsa 12,275 1.05 5875 1.58

2 The inverse Burr with a = c they call the inverse paralogistic, which is actually a name I coined
some years ago. For the loglogistic, F(x) = 1–(1+(x/b)a)–1, whereas the Pareto has F(x) = 1–
(1+(x/b)1)–a, so the combined form F(x) = 1–(1+(x/b)a)–a could be called the paralogistic. The in-
verse of a distribution in this context is the distribution of 1/X from that distribution, which gener-
ates the inverse Burr, inverse paralogistic, etc.

21
Neither paper looked at the heavy right tail copula. For τ of 0.31, this is not too different
from the Gumbel. In fact it is similar to the Gumbel in the right tail and more like the
Frank in the left tail. This suggests that the joint Burr discussed above, which is built from
the HRT copula, may provide a reasonable approximation to the loss and expense distri-
bution, particularly in the right tail. This could be useful for excess-of-loss reinsurance es-
timates, especially when data is scarce. Recall that the joint Burr distribution is given by:

F(x,y) = 1–(1+(x/b)p)-a –(1+(y/d)q)-a +[1+(x/b)p +(y/d)q]-a

The a parameter comes from the HRT copula, with τ = 1/(1+2a). For τ = 0.31, the implied
a is 1.11. The tail heaviness factors are ap and aq, so p and q can be estimated from these
parameters for this value of a. The tail heaviness can be estimated from available data or
industry values could be used. A simple choice given the table above would be to take the
loss factor as 1.11, which would give p = 1. A reasonable choice for q might be 1.5. Finally,
b and d can be estimated from the respective medians. E.g., for b and p, the median is
b(21/a –1)1/p. For a = 1.11, then, b = (median)1.151/p. The medians from Klugman and Parsa
with p = 1 and q = 1.5 give (rounded):

F(x,y) = 1–[1+x/14150]-1.11–[1+(y/6450)1.5]-1.11+[1+x/14150 +(y/6450)1.5]-1.11

Given a loss of x, the conditional distribution of loss expense is also Burr:
FY|X(y|x) = 1–[1+(y/dx)1.5]–2.11 with dx = 6450 +11x 2/3

22
 Simulated Hurricane Losses

A simulation of n=727 losses from a hurricane loss generator for a sample data set of

MD & DE Joint Empirical Probabilities Maryland and Delaware

exposures will be used as an
example of some of the issues
that arise in copula estimation.
As the emphasis is on the
copula, not the marginal
severities, the simulated losses
were converted to probabilities
by dividing the loss ranks for
each state by n+1=728. The
probability pairs were grouped
into 20 intervals of 5%
probability in each state for the
graph here. The graph in Figure
17 shows there is a positive
relationship between the loss
Figure 17
probabilities for the two states,
with some degree of concentration near (0,0) and (1,1). This is given in table form in Ap-
pendix C. A scatter plot of the empirical probabilities is shown in Figure 18.

DE vs. MD copula Figure 18

1.000 The usual estimate for the Kendall tau is
0.900 to compute the average value over all
0.800
pairs of observations (ui,vi), (uj,vj), i<j of
0.700
sign[(ui – uj)(vi – vj)]. In this case the
0.600

0.500
estimate is τ = .4545.
0.400

0.300 An empirical copula can also be built at

0.200
each point by counting the other points
0.100
that are less in both states. As there are
-
- 0.200 0.400 0.600 0.800 1.000

23
 n–1 other pairs, the count divided by n–1 can be taken as an estimate of the copula at that
point. For this data, the maximum empirical copula value is 0.9821 and the average is
0.36363. Four times this less 1 is another estimate of tau, and this also is 0.4545.

Empirical L and R functions can be computed similarly. An estimate for L(z) can be ob-
DE and MD L(z) & R(z)
Figure 19 tained as C(z,z)/z where C(z,z) is
1.0
computed as the proportion of
0.9

0.8 pairs with u and v both less than z.

0.7 Then with this C, R is estimated by
0.6 by R(z) = [1 – 2z +C(z,z)]/(1 – z).
0.5
These functions are graphed in
0.4
Figure 19. The limiting values L
0.3
and R are problematic immedi-
0.2

0.1 ately, in that they appear to be

positive, yet much less than tau. All the copulas reviewed above have either R=0 or R>τ.
The tails are fairly symmetrical, which poses additional fitting difficulties for single-tailed
copulas like the HRT, PP Max, and Clayton. The Frank and normal copulas are thus likely
to fit best, even though they are too light in the extreme tails.

An empirical cumulative tau can also be calculated. For each z, the empirical C(u,v) can be
computed for each (u,v) pair with both u and v less than z. Then the average of these val-
ues estimates the average copula in the square from (0,0) to (z,z). This divided by C(z,z),
times four less one, is the estimate of J(z).
DE and MD J(z) Figure 20
0.60

0.40
Its graph in Figure 20 is not like

0.20
the J(z) for any of the copulas for

0.00
small values of z, but the empiri-

-0.20
cal calculation is based on few

-0.40
points when z is small. For larger

-0.60
z it is most similar to the almost

-0.80
linear J of the Frank copula.

-1.00

24
 M(z) for MD|DE<z Figure 21 The M(z) function discussed in
0.5
Appendix B can be calculated either
0.45

0.4
for DE|MD or MD|DE. Figure 21

0.35 graphs MD|DE. It is most like the M

0.3 function for the normal copula.
0.25

0.2
The descriptive functions thus suggest
0.15
that the normal and Frank copulas
0.1
should provide the best fits to this
0.05
data, but they will be light in the tails.
0
0 0.2 0.4 0.6 0.8 1

0.9 Maximum likelihood

LR Function for DE/MD and Fits Figure 22 estimation of the
0.8
parameter was
performed for several
0.7
of these copulas. The
0.6 parameter and the
maximal likelihood
0.5
are shown below. As

0.4 all the copulas here

have a single pa-
Data
0.3 rameter, the ordering
Frank
Normal of the likelihood
0.2
PP Power function is the same
0.1 as those from the
various information
0
criteria like AIC, etc.
HRT Gumbel Frank Normal Flipped Gumbel
Parameter 0.968 1.67 4.92 0.624 1.68
Ln Likelihood 124 157 183 176 161
Tau 0.34 0.40 0.45 0.43 0.40

25
 The partial perfect copulas are difficult to estimate by MLE, as it is rare to have observa-
tions with exactly equal marginal probabilities. Nonetheless these copulas may be reason-
able as scenario generators. An alternative is to estimate the parameter by matching tau.
For the PP Power copula this gives a = 0.314. However for this data some of the descrip-
tive functions seem to make this copula unlikely.

The likelihood function favors the Frank copula in this case. Some of the descriptive func-
tions are graphed for the fit and the data for this copula and in some cases some other
copulas in Figures 22 and 23. The L and R functions are combined in the graph above.
R(z) is shown for z>0.5, and L(z) for z < 0.5. The Frank copula looks like a close fit all
along except in the tails, where the normal is a little better. The PP Power appears to be
too heavy in the right tail for this data.
0.6

J(z) Data and Fits

The graph at left shows the
0.4

J(z) function for the data and

0.2
the normal and Frank copu-
0 las. The two copulas provide
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

quite different fits to this

-0.2 Data
data, but it is a subjective
Normal
-0.4
Frank matter as to which is better,
-0.6 with the Frank probably
having the edge for its close
-0.8

Figure 23 fit for z>0.5. The Frank

-1
copula has a lower sum of
squared errors, but this disappears if the first two points (at –1) are omitted.

Even though the Frank copula provides the best fit according to the likelihood function,
there are fitting problems in the tails. Somewhat heavier-tailed copulas with strength in
both tails would be useful here. See Appendix D for an example. Another alternative
would be to use the Frank copula but model the extreme events separately.

26
 6. Conclusion

Copulas provide a convenient way to model and simulate correlated variates. Several
copulas with varying shapes are available for modeling these relationships. Shape differ-
ences among copulas can be discerned with the descriptive functions. These can be used
both in fitting copulas to data and in applying informed judgment to select a copula for a
given application.

Statisticians have identified a fair number of copulas – e.g., see [7]. The use of the descrip-
tive functions provides an avenue for researching their properties. There may also be
more descriptive functions that can reveal other aspects of a copula. For instance, the J
and M functions looked at average probabilities between 0 and z. Mirror functions could
look at the same probabilities between z and 1, analogous to the way that R mirrors L. It
would also be possible to define more functions over non-rectangular parts of the unit
square, such as the region where C(u,v) is less than z, as in the K function, or sections like
u and v both less than z.

This paper focused on bivariate copulas but many of the concepts can be generalized to
the multi-variate case. The descriptive functions have multi-variate analogs except for
M(z) which would have to be done pairwise. Only the normal and partial perfect copulas
fully generalize to multi-variate forms that allow specification of all pairwise correlations,
but there are other multivariate copulas – e.g., see [4].

In summary, actuaries now have a number of copulas to chose among and a number of
techniques for refining that choice, yet more copulas and more techniques could still be
worth uncovering.

27
 Appendix A – J(z)
For a copula with distribution function C(u,v) define:
z z
I(z) = ∫0 ∫0 C(u,v)c(u,v)dvdu. Then J can be expressed as:
J(z) = 4I(z)/C(z,z)2 – 1.
For the following distributions the formula for 4I(z) is given.

Gumbel:
∞
(2–1/a)exp[21+1/aln(z)] – 4(-ln(z))a(1–1/a) ∫y e–2w w–a dw, where y = -21/aln(z)

Heavy Right Tail:

y
8z –8+ 4(2y–1)–a + [4a(1–z)2 +2(1+(2y–1)–2a)(a+1)]/[2a+1] + 8a∫1 (w+y –1)–a–1w–a dw,

where y=(1–z)–1/a.

Partial Perfect Max

z4 + (z>a)(a4 – 4a3 + 2(1 + 2z)a2 – 4az + 2z2 – z4)

Partial Perfect Power

z4 +4(a+1)–2[(y4–2y3/3+y2/2)(a+1)–2 +za+3(a2+3a+4)(a+2)–1(a+3)–1 –z2(a+2)(a2+2a+2)(a+2)–2],

where y=za+1.

Clayton
y–b(b+1–b/y)(b+2)/(b+1), where b=2a and y=2/z1/a–1.

BB1
y
2(1+(ac)– 1)y2 – 4(a+1)y2+a/[ac(a+2)] –4x(ac)-1∫0 (w– a–1)–c w1+a [(ac+1)w- a–a–1] dw, where
x=2(z– a–1)c, and y=(1+x1/c)–1/a.

28
 Appendix B – Other Descriptive Functions

Cumulative Conditional Mean

A function of interest is the conditional expected value of V|U=z. However this is often
difficult to estimate from data, as there are usually not too many values of V for any given
value of U. So a related function is chosen: the expected value of V given U<z. Let
z 1
M(z) = E(V|U<z) = ∫0 ∫0 vc(u,v)dvdu/z
Since E(V) = ½, every copula will have M(1) = ½ so the differences in M among copulas
will be for lower values of z and the shape of the curve approaching z = 1.
Often the integral has to be done numerically, but for a few copulas it is done explicitly at
the end of this appendix. Graphs of this function for several copulas are shown in Figures
24 – 29. For this function, the lower τ is, the closer the values stay to ½.

Frank M(z) for τ = .1, .5, .9 Figure 24 Gumbel M(z) for τ = .1, .5, .9 Figure 25
0.5 0.5

0.45 0.45

0.4 0.4

0.35 0.35

0.3 0.3

0.25 0.25

0.2 0.2

0.15 0.15

0.1 0.1

0.05 0.05

0 0

HRT M(z), τ = .1, .5, .9

Normal M(z) for τ = .1, .5, .9 Figure 26
0.5 0.6
Figure 27
0.45
0.5
0.4

0.35 0.4

0.3

0.25 0.3

0.2
0.2
0.15

0.1 0.1

0.05
0
0

29
 PP Power M(z), τ = .1, .5, .9 Figure 28 PP Max M(z), τ = .1, .5, .9 Figure 29
0.6
0.6

0.5
0.5

0.4
0.4

0.3
0.3

0.2
0.2

0.1
0.1

0
0

Copula Distribution Function

Genest and Rivest [3] define a function K(z) which is basically Pr(C(u,v)<z). It is the area
of the unit square in which Pr(C(u,v)<z). An empirical K(z) can be calculated for any z as
the proportion of empirical values of C(u,v) that are less than z. Although C(u,v) ap-
proaches one as u and v approach one, it is possible that C is low for most values of u and
v, which would make K(z) high for most z’s. Or C could grow fairly quickly through
lower values of u and v, which would tend to make K(z) smaller.

Genest and Rivest show how to calculate K for a number of copulas. In particular,
Copula K(z)
Gumbel z(1 – ln z1/a)
Frank z + a–1(1–eaz)ln[(1–e–az)/(1–e–a)]

Hurricane Application
M(z) for the hurricane data and the Frank and normal copulas is graphed in Figure 30.
The normal copula is the one with the better fit for small events, and the Frank fits better
in the middle of the range.

A scatter plot of the empirical K percentiles as a function of the Frank K percentiles (often
called the QQ plot) is shown in Figure 31, along with the line x=y. The values are very

30
close to the line. This supports the fit, but as K(0) = 0 and K(1) = 1 for any copula, empiri-
cal or parametric, fit problems in the tails are difficult to discern with this function.

0.5
M(z) Data and Fits
0.45
0.4
0.35
0.3
0.25 Data

0.2 Frank
Normal
0.15
0.1
0.05
0
Figure 30

K percentiles, Data vs. Frank

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 Figure 31
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

31
M(z) Formulas
Partial Perfect Maximum
M(z) = ½ – ½ (z > a)(1 – a)(1 – z)(z – a)/z
Partial Perfect Power
M(z) = ½ + (za+1 – za)/[(a+1)(a+2)]

32
 Appendix C – Delaware and Maryland Probabilities by Range
Range Upper Limits - Maryland
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
0.05 7 9 2 8 1 0 3 1 2 1 0 0 0 0 1 0 0 1 0 0
0.1 2 7 9 5 4 1 0 1 5 1 0 1 0 0 0 0 0 0 0 0
0.15 10 5 2 3 6 1 4 2 0 1 0 1 0 0 0 1 0 0 1 0
0.2 2 4 2 3 2 6 5 1 2 4 1 0 0 0 1 0 1 1 1 0
0.25 2 2 2 1 4 5 2 3 3 4 4 2 0 1 0 0 0 0 0 1
0.3 3 0 3 4 4 0 5 3 3 1 2 6 0 0 1 0 0 0 0 2
0.35 3 0 3 4 1 2 2 2 5 2 2 4 3 0 0 1 1 0 0 1
0.4 2 1 2 4 2 1 1 2 2 3 3 2 2 2 1 1 0 0 2 4
0.45 1 1 1 0 3 4 4 3 2 0 4 3 1 2 1 2 1 0 0 3
0.5 1 1 0 0 1 6 1 1 1 1 3 1 4 1 5 1 2 3 0 3
0.55 1 3 1 0 1 2 3 4 3 1 3 0 5 4 4 1 1 0 0 0
0.6 1 2 3 1 2 2 0 1 1 6 1 2 1 1 3 4 0 2 2 1
0.65 1 0 4 3 0 1 1 2 2 2 4 0 3 5 1 1 0 2 2 3
0.7 0 1 2 0 1 3 1 3 2 3 2 2 3 3 3 2 0 1 3 1
0.75 0 0 1 0 3 2 0 0 0 3 3 2 1 3 2 5 8 3 0 0
0.8 0 0 0 0 1 1 1 4 1 0 2 2 1 3 2 3 7 2 3 4
0.85 0 0 0 0 0 0 2 3 0 0 1 3 7 3 1 1 6 5 3 1
0.9 0 0 0 0 0 0 1 1 2 3 0 2 5 3 1 5 2 7 4 1
0.95 0 0 0 0 0 0 0 0 0 0 2 2 1 5 7 4 4 3 6 2
1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 2 5 3 7 9 9

33
 Appendix D – Joe’s BB1 Copula
Several examples of two-parameter bivariate copulas are provided by Joe [4]. One that has a
closed form for tau and can be heavy in both tails he labels BB1. It is a generalization of the
Gumbel and Clayton copulas.

C(u,v) = {1+[(u-a – 1)c + (v-a – 1)c]1/c}-1/a , a > 0, c ≥ 1

C1(u,v) = {1+[(u-a – 1)c + (v-a – 1)c]1/c}-1/a – 1[(u-a – 1)c + (v-a – 1)c]1/c – 1(u-a – 1)c – 1u–a – 1

c(u,v) = {1+[(u-a – 1)c + (v-a – 1)c]1/c}-1/a – 2[(u-a – 1)c + (v-a – 1)c]2/c – 2 ×

{ac + 1 + a(c– 1) [(u-a – 1)c + (v-a – 1)c]– 1/c} (u-a – 1)c– 1u–a– 1 (v-a – 1)c– 1v–a– 1

τ = 1 –2/[c(a+2)]
R(1) = 2 – 21/c , L(0) = 2– 1/(ac)

The Gumbel is the limiting case a → 0. The Clayton arises when c = 1, but here the a parameter
is the reciprocal of the Clayton a parameter in the text.

With R, L, and τ all closed form it is possible to find a and c to set two of them and then see
what the third is. Not all combinations are possible. Figure 32 graphs L as a function of R for
several values of τ. For each τ, there is an inverse relationship between R and L. Either can get as
low as needed, approaching zero, for any value of τ, but then the other becomes large. Each be-
comes somewhat higher than τ if the other one is low. Higher τ allows higher R and L. The left
tails appear to be somewhat heavier than the right tails, so flipping this copula could be useful
for some applications.

This copula was not all that useful for the hurricane data, in that it is so heavy-tailed. The MLE
log-likelihood was 170, which was not as good as some other copulas. The a and c were 0.386
and 1.434, which gave τ = 0.415, L = 0.286, and R = 0.379. So the τ was a little low and the tail
parameters higher than the data would suggest.

34
 BB1 L as a Function of R, tau = .1, .2, …, .9 Figure 32
1

0.9

0.8
tau = .7

0.7

0.6
tau = .5
0.5
L

0.4
tau = .3
0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
R

35
 References

1. Embrechts, Lindskog, and McNeil Modelling Dependence with Copulas and Applications to
Risk Management at http://www.risklab.ch/Papers.html#MTLindskog

2. Frees, EW and Valdez, EA, “Understanding Relationships Using Copulas,” North Ameri-
can Actuarial Journal, vol.2, #1, pp. 1-25

3. Genest, C and Rivest, L 1993 “Statistical Inference Procedures for Bivariate Archimedean
Copulas,” Journal of the American Statistical Association 88: 1034–1043

4. Joe, H, Multivariate Models and Dependence Concepts, Chapman and Hall, 1997

5. Klugman, SA and Parsa, R, “Fitting Bivariate Loss Distributions with Copulas,” Insur-
ance: Mathematics and Economics 24, pp. 139-148, 1999

6. Kreps, R “A Partially Comonotonic Algorithm For Loss Generation,” ASTIN Colloquium

Papers 2000

7. Nelson, R, An Introduction to Copulas, Springer Lecture Notes in Statistics, 1999

8. Wang, S, “Aggregation of Correlated Risk Portfolios: Models and Algorithms,” 1999

PCAS LXXXV

36