You are on page 1of 13

Insurance: Mathematics and Economics 76 (2017) 113

Contents lists available at ScienceDirect

Insurance: Mathematics and Economics


journal homepage: www.elsevier.com/locate/ime

Hierarchical Archimedean copulas through multivariate compound


distributions
Hlne Cossette, Simon-Pierre Gadoury, tienne Marceau *, Itre Mtalai
cole dactuariat, Universit Laval, Quebec (Qc), Canada

article info a b s t r a c t
Article history: In this paper, we propose a new hierarchical Archimedean copula construction based on multivariate
Received November 2016 compound distributions. This new imbrication technique is derived via the construction of a multivariate
Received in revised form June 2017 exponential mixture distribution through compounding. The absence of nesting and marginal conditions,
Accepted 6 June 2017
contrarily to the nested Archimedean copulas approach, leads to major advantages, such as a flexible
Available online 19 June 2017
range of possible combinations in the choice of distributions, the existence of explicit formulas for the
distribution of the sum, and computational ease in high dimensions. A balance between flexibility and
Keywords:
Archimedean copulas parsimony is targeted. After presenting the construction technique, properties of the proposed copulas
Mixing random variables are investigated and illustrative examples are given. A detailed comparison with other construction
Compounding methodologies of hierarchical Archimedean copulas is provided. Risk aggregation under this newly
MarshallOlkin proposed dependence structure is also examined.
Hierarchical structure 2017 Elsevier B.V. All rights reserved.

1. Introduction The use of multivariate Archimedean copulas in high dimension


can be restrictive due to their exchangeability property. One can
Copulas are now well known tools used for dependence model- resort to another interesting class of copulas, namely vine copulas
ing purposes in many research topics. A d-dimensional copula is a (see e.g. Bedford and Cooke (2002) and Joe (1997)). They are
d-variate probability distribution function for which the marginals pair copula constructions allowing a cascade decomposition of a
are uniformly distributed on (0, 1), with d 2. One important multivariate distribution into the product of bivariate copulas. Vine
class of copulas is the Archimedean copula family, popular for its d(d1)
copulas force the use of 2 bivariate copulas which requires a
simple construction procedure and multivariate generalization. A
high number of parameters when d, the dimension of the copula,
d-dimensional copula C is said to be an Archimedean copula if
increases.
C (u1 , . . . , ud ) = 1 (u1 ) + + 1 (ud ) , Hierarchical Archimedean copulas provide an interesting al-
( )
ternative to allow asymmetries. The first approach to construct
for (u1 , . . . , ud ) [0, 1]d . (1) hierarchical Archimedean copulas was proposed by Joe (1997) who
The continuous and strictly decreasing function is called the introduced the so-called nested Archimedean copulas in three and
generator of the copula, where : [0, ) [0, 1], (0) = 1 and four dimensions. They are obtained by nesting into each other
limt (t) = 0. In the same manner, 1 : [0, 1] [0, ), for Archimedean copulas. They are able to capture different depen-
which 1 (0) = inf{t : (t) = 0}, where 1 is the inverse of the dence relations between and within different groups of risks with
generator . The set of all such functions is denoted by . In fact, a relatively small number of parameters (see e.g. Grecki et al.,
from Kimberlings and Bernsteins theorems, see e.g. Kimberling 2016a). They were further studied by e.g. McNeil (2008), Hofert
(1974), Feller (1971), and Hofert (2010), representation (1) leads to (2012), and Hofert (2011), in a general setting. For an Archimedean
a proper copula for all d 2 if and only if is the LaplaceStieltjes hierarchical structure to be a proper copula, a given nesting condi-
Transform (LST) of a strictly positive random variable (rv) with
tion must be verified. For example, a 3-dimensional fully nested
cumulative distribution function (cdf) F , where the LST of the rv
Archimedean copula can be written as
is given by
C (u1 , u2 , u3 ) = C (u1 , C (u2 , u3 ))

et x dF (x) = E et .
[ ]
L (t) = (2) ( ))
= 0 01 (u1 ) + 01 1 11 (u2 ) + 11 (u3 ) .
(
0

For the hierarchical structure to be a proper copula, 01 1


* Corresponding author.
E-mail address: etienne.marceau@act.ulaval.ca (. Marceau). must have completely monotone derivatives, where 0 and 1 are

http://dx.doi.org/10.1016/j.insmatheco.2017.06.001
0167-6687/ 2017 Elsevier B.V. All rights reserved.
2 H. Cossette et al. / Insurance: Mathematics and Economics 76 (2017) 113

generators of the parent and the child copulas respectively. If gen- and marginal pgfs given by
erators 0 and 1 belong to the same family, the verification of this [
Mi
]
sufficient nesting condition can be done without much problem, PMi (t) = E ti , (i = 1, . . . , d), (4)
see Hofert (2010) for restrictions on the parameters. However,
if generators 0 and 1 of different Archimedean families are where |t1 |, . . . , |td | 1.
A vector of rvs = (1 , . . . , d ) is said to follow a multivariate
combined, the sufficient condition may not hold for any choice of
parameters. Also, sampling from these copulas can be difficult.
compound distribution if each Mi component i can be represented as
To circumvent these constraints, other approaches were pro- a random sum i.e. i = j=1 Bi, j , where i = 1, . . . , d for d 2.
posed to build hierarchical Archimedean copulas. Hering et al. For each i, the elements of the sequence Bi = {Bi,j , j = 1, 2, . . .}
(2010) suggest to build hierarchical Archimedean copulas based are assumed independent and identically distributed (iid) strictly
on Lvy subordinators. Also, Brechmann (2014) used the Kendall positive rvs, while the sequences are independent from each other
and from the vector of rvs M. By convention, for each i = 1, . . . , d,
distribution function, the multivariate analog to the probability
Bi,j Bi with cdf FBi and LST LBi (j = 1, 2, . . .).
integral transform for univariate random variables, as a tool to
Since the rv i is defined as a random sum, its LST is given by
construct hierarchical Archimedean copulas.
[ [ ]M ]
In this paper, our objective is to propose an alternative ap-
Li (t ) = E et i = EMi E et i |Mi = EMi E etBi i
[ ] [ [ ]]
proach to build hierarchical Archimedean copulas via multivari-
= PMi LBi (t ) ,
( )
ate compound distributions. The absence of nesting and marginal (5)
conditions leads to major advantages, such as a flexible range of
possible combinations in the choice of distributions, the existence for i = 1, . . . , d.
Similarly, the multivariate LST of the vector of mixing rvs is
of explicit formulas for the distribution of the sum, and compu-
given by
tational ease in high dimensions. The approach given here has
similarities with the one of Hering et al. (2010). Both approaches
L (t1 , . . . , td ) = E et1 1 . . . etd d
[ ]
derive hierarchical Archimedean copulas from the joint survival
= EM E et1 1 . . . etd d M
[ [ ]]
function of a multivariate mixed exponential distribution. Also,
sampling algorithms are obtained in a similar fashion. The proba- = EM [(E [etB1 ]M1 ) . . . (E [etBd ]Md )]. (6)
bilistic arguments at the basis of these two construction strategies
however differ. Here, random sums are key elements compared to With (6) and (3), the expression for the multivariate LST of is
Lvy processes in Hering et al. (2010). In our opinion, the proposed given by
approach uses simpler mathematical tools to obtain hierarchical
L (t1 , . . . , td ) = PM LB1 (t1 ) . . . LBd (td ) .
( )
Archimedean copulas. We provide in Section 4.2 a detailed com- (7)
parison of both constructions. In this paper, we consider the specific case of multivariate
The following sections are organized as follows: in Section 2, compound distribution for assuming that Mi = M (i = 1, . . . , d),
we provide the steps for the construction of a multivariate copula where M is a strictly positive discrete rv with
with the use of multivariate compound distributions, followed
PM (t) = E t M .
[ ]
by sampling algorithms, accompanied by relevant notations and (8)
properties. Section 3 proposes a representation of the copula by
In other words, the components of M are comonotonic and
common mixtures. In Section 4, we compare the proposed ap-
identically distributed (with Mi M, i = 1, . . . , d). Then, it means
proach with nested Archimedean copulas and Hering et al. (2010)s
that the vector of mixing rvs follows a multivariate compound
construction method and provide a brief discussion on estimation
distribution and its multivariate LST in (7) becomes
procedures. Finally, Section 5 is devoted to risk aggregation.
L (t1 , . . . , td ) = E et1 1 . . . etd d
[ ]
2. Construction of hierarchical Archimedean copulas with mul-
= EM E et1 1 . . . etd d |M
[ [ ]]
tivariate compound distributions
= EM E et1 1 |M . . . E etd d |M
[ [ ] [ ]]
As explained in Marshall and Olkin (1988), multivariate distri- [( [ ])M ]
= EM E et1 B1 . . . E etd Bd .
] [
butions can easily be constructed through the use of exponential (9)
mixtures. In this section, we propose an alternative method based
Combining (9) and (8) leads to the following expression for the
on a probabilistic argument framework using multivariate com-
multivariate LST of :
pound distributions to build hierarchical Archimedean copulas.
Note that the copula obtained under the proposed approach is
L (t1 , . . . , td ) = PM LB1 (t1 ) . . . LBd (td ) .
( )
(10)
represented as a specific function of LSTs of rvs and their inverse.
To obtain this representation, we have recourse to rvs with mul- Also, for each i = 1, . . . , d, (5) becomes
tivariate compound distributions. For this reason and given the
Li (t ) = PM LBi (t ) .
( )
context, we will refer either directly to LSTs or to the rvs associated (11)
to them. The key element of our construction is that the latter are
defined with a vector of mixing rvs, denoted by , which follows 2.2. Multivariate mixed exponential distributions defined with multi-
a multivariate compound distribution. We examine the properties variate compound distributions
of this new copula and provide simulation algorithms.
Let Y = (Y1,1 , . . . , Y1,n1 , . . . , Yd,1 , . . . , Yd,nd ) be a vector of
2.1. Multivariate compound distributions n1 + ( + nd rvs which can be more conveniently represented
as Y = Y1 , . . . , Yd , where Yi = Yi,1 , . . . , Y1,ni is the vector of
) ( )
Let M = (M1 , . . . , Md ) be a vector of discrete strictly positive ni rvs for the subgroup i (i = 1, 2, . . . , d). Given = , where
rvs with joint multivariate probability generating function (pgf) = (1 , . . . , d ), it is assumed that
given by
[ ] (Y1,1 | = ), . . . , (Y1,n1 | = ), . . . ,
M Md
PM (t1 , . . . , td ) = E t1 1 . . . td , (3) (Yd,1 | = ), . . . , (Yd,nd | = )
H. Cossette et al. / Insurance: Mathematics and Economics 76 (2017) 113 3

1 ( )
are conditionally independent. The conditional distributions of the e.g. Sklar, 1959 or Nelsen, 2007) and letting yi,j = F Yi,j ui,j in (13),
components of Yi are only influenced by the component i of , a hierarchical Archimedean copula C defined with a multivariate
i.e. (Yi,j | = ) is identically distributed as (Yi,j | i = i ), compound distribution is obtained from the multivariate survival
for i = 1, . . . , d and j = 1, . . . , ni . We assume that (Yi,1 | function given in (19) of the multivariate mixed exponential dis-
i = i ), . . . , (Yi,ni | i = i ) are exponentially distributed with tribution as follows :
parameter i for i = 1, . . . , d. The univariate distribution of Yi, j is ( 1 1
C u = F Y F Y1,1 u1,1 , . . . , F Y1,n u1,n1 , . . . ,
( ) ( ) ( )
therefore a mixed exponential distribution with survival function 1
given by 1 1 ))
F Yd,1 ud,1 , . . . , F Yd,n
( ) (

ud,nd
d
F Yi,j |i =i (yi,j ) dFi (i )
( )
F Yi,j yi,j =
n1 nd

0
1
u1,j , . . . , 1
ud,j ,
( ) ( )
= L L1 Ld (20)
eyi,j i dFi (i ) = Li yi,j ,
( )
= (12) j=1 j=1
0

for i = 1, . . . , d and j = 1, . . . , ni . The inverse of the survival where u = (u1 , . . . , ud ) with ui = (ui,1 , . . . , ui,ni ) for i = 1, . . . , d.
Let us examine more deeply the structure of C . The expression
function F Yi,j in (12) is given by
for the multivariate Archimedean copula in (20) is written in terms
1 of the multivariate LST of the multivariate mixing random vector
i ui,j ,
1
F Yi,j ui,j = L
( ) ( )
(13)
with the sum of the inverse of the corresponding marginal LSTs
where of 1 , . . . , d evaluated at each element of a subgroup as compo-
nents. Hierarchical Archimedean copulas built with our proposed
1 1 1
,
( ) ( ( ))
Li ui,j = LB PM ui,j (14) approach can be seen as multivariate extensions to the classical
i
univariate Archimedean copulas which are defined as a univariate
for ui,j [0, 1], i = 1, . . . , d and j = 1, . . . , ni . Since
LST evaluated at the sum of the inverse of the LST evaluated at each
element of a multivariate uniform random vector as follows:
PM (s) = LM ( ln(s)) (15) ( )
d

and C (u) = L 1
L (ui ) .
i=1
M (u) .
(u) = exp L
1 1
( )
PM (16)
Given the connection with the mixing random vector and the
1 1
We can rewrite L
i in terms of LM as follows: random variables M, Bi (i = 1, . . . , d), an alternative representa-
tion of the copula C in (20) can be established in terms of LM and LBi
1 1
exp(L1
ui,j ) , (i = 1, . . . , d), and Li (i = 1, . . . , d). Indeed, letting (10) and (15)
( ) ( ( ))
Li ui,j = LB M (17)
i
in (20), the hierarchical Archimedean copula C can also be written
for ui,j [0, 1], i = 1, . . . , d and j = 1, . . . , ni . as follows:
The vector of rvs Yi follows a multivariate mixed exponential
ni
( d ( ) )
distribution, which is defined only in terms of LST of the mixing rv ( ))
1
.
( ) (
C u = LM ln LBi Li ui,j (21)
i given in (11) and for which the multivariate survival function
i=1 j=1
corresponds to
Finally, inserting (17) in (21), the hierarchical Archimedean
F Yi (yi ) = F Yi |i =i (yi ) dFi (i ) copula C in (21) can be defined solely in terms of LM and LBi
0 (i = 1, . . . , d) as
ni

F Yi,j |i =i (yi,j ) dFi (i ) C u; LM , LB1 , . . . , LBd
( )
=
0 j=1 ni
( d ( ))
( ( ))))
ni 1
exp L 1
.
( (
= LM ln LBi LB ui , j (22)
eyi,j i dFi (i ) i M
= i=1 j=1
0 j=1
Note that in (22) the terms LM , LB1 , . . . , LBd are explicitly inserted
= Li (yi,1 + + yi,ni ), (18)
as arguments of the function C . We will add these arguments only
where yi = (yi,1 , . . . , yi,ni ), for i = 1, . . . , d. when we find relevant to insist on their implication in the copula
Finally, the vector of rvs Y follows a multivariate mixed ex- representation.
ponential distribution. Most importantly for us, the latter has the
interesting feature of being defined by the vector of mixing rvs , 2.4. Dependence structure of hierarchical Archimedean copulas de-
which follows a multivariate compound distribution. The multi- fined with multivariate compound distributions
variate survival function of Y is represented in terms of the LST
(10) of i.e. Either from (21) or (22), it is clear that the proposed hierarchical
Archimedean copulas have explicit forms and offer many possibil-
F Y y = L y1,1 + + y1,n1 , . . . , yd,1 + + yd,nd ,
( ) ( )
(19) ities of dependence structures through the choices of distributions
for the strictly positive discrete rv M and the strictly positive rvs
where y = y1 , . . . , yd = y1,1 , . . . , y1,n1 , . . . , yd,1 , . . . , yd,nd .
( ) ( )
B1 , . . . , Bd . These copula constructions allow flexibility without
requiring a large number of parameters. As stated in Brechmann
2.3. Construction of hierarchical Archimedean copulas defined with (2014), a major issue of any copula model is to find a good balance
multivariate compound distributions between parsimony and flexibility.
The dependence structure associated to the hierarchical
We use the inversion method to identify the proposed hi- Archimedean copula defined in either (21) or (22) can be illustrated
erarchical Archimedean copula C defined with a multivariate with a tree representation as shown in Fig. 1 in which U =
compound distribution. Indeed, applying Sklars Theorem (see (U1 , . . . , Ud ) with Ui = (Ui,1 , . . . , Ui,ni ) for i = 1, . . . , d, is a vector
4 H. Cossette et al. / Insurance: Mathematics and Economics 76 (2017) 113

2.5. Dependence properties

In light of the previous results, we aim to measure and compare


the strength of the dependence relations between the rvs of a given
subgroup and also between rvs from different subgroups. In other
words, we want to compare whether the copula generated from L
is more dependent than the one with a generator LM . To do so, we
Fig. 1. One level tree structure. need to introduce the concept of concordance ordering as defined
in Joe (1997) page 37.

of rvs with multivariate cdf C . This shows a one level tree structure Definition 1. Let C1 and C2 be two d-dimensional copulas. C2 is
composed of d subgroups joined together by their common link more concordant that C1 , written C1 c C2 , if
with the rv M. The dependence among the components of Ui is
and C1 u C2 u ,
( ) ( ) ( ) ( )
C1 u C2 u
captured by Archimedean copulas generated by the mixing rvs i ,
for i = 1, 2, . . . , d. Then, the dependence relation between the for u [0, 1]d .
vectors U1 , . . . , Un is defined through the rv M.
Let us examine the dependence structure for specific subsets A very useful consequence of the concordance ordering is its
of U. First, the multivariate cdf of (Ui1 , . . . , Uim ) is given by the relation with dependence measures. For instance, if C1 and C2 are
(1)
copulas with respective Kendall taus 1 , 2 , Spearman rhos S ,
hierarchical Archimedean copula (2)
S , tail dependence parameters 1 , 2 , and C1 c C2 , then 1 2 ,
C (ui1 , . . . , uim ; LM , LBi , . . . , LBim ),
1
(23) S(1) S(2) , and 1 2 . (See Joe, 1997.)
Suppose now that C1 and C2 are Archimedean copulas with
for i1 , . . . , im {1, 2, . . . , d} and for m = 2, . . . , d. Also, if we respective generators L1 and L2 . Theorems 4.1 and 4.7 in Joe (1997)
choose at least two subgroups i1 and i2 , the multivariate cdf of any provide the condition that
vector formed with at least two rvs from each one of them is a ( 1guarantees
) that C1 c C2 is verified. Such
a condition states that L 1 L2 is a completely monotone func-
hierarchical Archimedean copula of the similar form. tion.
Now we consider the dependence structure within a subgroup i We recall that our objective is to compare the copula C , gener-
(i = 1, 2, . . . , d). From (20), the multivariate cdf of (Ui,j1 , . . . , Ui,jl ) ated from L , with the one generated from LM (written CM ). Since
(for j1 , . . . , jl {1, 2, . . . , ni } and for l = 2, . . . , ni ) is given by L is related to LM , this condition is easily verified as follows:
( l
)
LM L = LM LM ( ln (LB )) = ln (LB ) .
1 1
C ui,j1 , . . . , ui,jl ; Li = Li 1
( ) ( )
Li ui,k (24)
k=1 Therefore, if ( ln (LB )) is completely monotone, then the de-
pendence within a subgroup is stronger than the outer dependence
which corresponds to an l-variate Archimedean copula with ( gen-) between subgroups, i.e. CM c C and hence M .
erator Li . As a special case, the multivariate cdf of Ui is C ui ; Li
It is well known that ( ln(LBi )) is completely monotone if and
for i = 1, . . . , d. Due to (11) and (17), copulas of the form as given
only if Lm Bi is the LST of a positive rv i.e. LBi for all m N
m
in (24) are Archimedean copulas based on two generators LM and
(see Joe, 1997). Table 1 provides a list of distributions, used to
LBi . Such copulas allow for more flexibility by using at least two
generate an Archimedean copula, for which it is easy to directly
parameters (see e.g. Nelsen, 2007, Section 4.5, for other examples
of two parameter Archimedean copulas).
verify that Lm Bi is the LST of a positive rv. We denote by SNB(m; r , q)
a shifted negative binomial distribution i.e. Bi,1 + + Bi,m admits
Finally, we turn our attention to the dependence structure
a stochastic representation N + m, where N NB(r , q), r R+ and
for vectors of rvs containing at most one component from each m
q (0, 1). For the logarithmic distribution, j=1 Bi,j is an mfold
subgroup of U. For j1 {1, 2, . . . , n1 }, . . . , jd {1, 2, . . . , nd }, the
sum
m of iid logarithmic rvs. If Bi follows a Sibuya distribution,
multivariate cdf of (U1,j1 , . . . , Ud,jd ) is obtained from (10) and (20)
and is given by j=1 Bi,j is distributed as an unknown rv N with probability mass
( m ) ( i )
(1) , k = 1, 2, . . .. Finally,)for a
i+k
function pk = i=1 i k
(
C u1,j1 , . . . , ud,jd = L1 ,...,d L1 u1,j1 , . . . , Ld ud,jd
1 1
( ) ( ( ) ( ))
1
stable distribution i.e. Bi S , 1, cos(/2) , 1{=1} ; 1 , Lm
B i
PM LB1 L1 u1,j1 . . . LBd Ld ud,jd .
( ( 1 ( )) ( 1 ( )))
= is(the LST of a positive rv exponentially tilted
1
) stable distributed
(25) S , 1, (cos(/2)m) , m 1{=1} , h 1{=1} ; 1 with h = 0. For
Using (14), (25) becomes sampling procedures, we refer to Hofert (2011) and Devroye
(2009).
C u1,j1 , . . . , ud,jd
( )
( ( ))) ( ))))
1
. . . 2.6. Examples of hierarchical Archimedean copulas
( 1 ( 1
( 1 (
= PM LB1 LB1 PM u 1 , j1
L B d
L Bd PM u d, j d

= PM PM u1,j1 . . . PM ud,jd
( 1 ( ) 1
( ))
In the following two examples, we give a glimpse of the
flexibility of our approach based on multivariate compound dis-
= LM ln PM1 u1,j1 ln PM1 ud,jd .
( ( ( )) ( ( )))
(26)
tributions in the construction of a large variety of hierarchical
Then, from (16) and (26), we conclude that the multivariate cdf Archimedean copulas C . One must first choose the distribution
of (U1,j1 , . . . , Ud,jd ) is for the strictly positive discrete rv M. Here, we limit ourselves to
( ) shifted-geometric and logarithmic distributions for M. Of course,
d
other candidates would have been suitable, such as the Sibuya and
C u1,j1 , . . . , ud,jd ; LM = LM 1
,
( ) ( )
LM ui,ji (27)
shifted-negative binomial distributions (see e.g. Joe, 2014). Next,
i=1
distributions for the rvs Bi are selected. Either shifted-geometric,
which corresponds to a single-generator Archimedean copula with logarithmic, or gamma distributions are picked. Among the pos-
LM . Given (27), it means that the dependence relation between the sible combinations of distributions, we choose to depict 6 hier-
components of (U1,j1 , . . . , Ud,jd ) is only defined through the rv M. archical Archimedean copulas constructed under the proposed
H. Cossette et al. / Insurance: Mathematics and Economics 76 (2017) 113 5

Table 1
Distributions for the sum of common laws.
Distribution of Bi Lm
B (t) Distribution of Bi,1 + + Bi,m
i
( )m
qet
ShiftedGeometric (q) 1(1q)et
SNB(m; m, q)
( )m r
qet
ShiftedNegativeBinomial (r ; r , q) 1(1q)et
SNB(rm; rm, q)
( )m
ln(1 et ) m
Logarithmic ( ) ln(1 ) i=1 Bi , Bi Logarithmic( )
m
( m ) ( i )
Sibuya ( )
( )
1 (1 et ) N pk = i=1 i k
(1)i+k
)m
Gamma (, 1) 1
Gamma(m, 1)
(
1+t
( 1
) ( 1
)

Stable , 1, cos(/2) , 1{=1} ; 1 (exp(t )) m
S , 1, (cos(/2)m) , m 1{=1} , 0; 1

approach. In the first example, we present 3 copulas with an un- Example 2. We assume that M Logarithmic ( ) with LM (t ) =
ln(1 et ) (1(1 )t )
derlying shifted-geometric distribution for M, while in the second
ln(1 ) M (t ) = ln(
and L1

). Letting = ln(1 )
example a logarithmic distribution describes the behavior of the the multivariate cdf of (U1,j1 , . . . , Ud,jd ) corresponds to the Frank
rv M. By convention, when the distributions for the rvs Bi (i = copula with generator LM , i.e.,
1, . . . , d) are from the same family of distributions (with different ( n ( ui
))
parameters), we refer to the resulting copula as distribution for 1 i=1 1 e
C u1,j1 , . . . , ud,jd ; LM .
( )
= ln 1
M-distribution for Bi hierarchical Archimedean copula. (1 e )n1
qet 1. LogarithmicGeometric hierarchical Archimedean copula:
Example 1. Let M ShifGeo (q) with LM (t ) = 1(1q)et
and
qi et
LM (t ) = ln(
1
q
1
). Letting = 1 q, the multivariate Bi ShifGeo (qi ) with LBi (t ) = 1(1qi )et
and
t +1q
cdf of (U1,j1 , . . . , Ud,jd ) corresponds to the AliMikhailHaq (AMH) LB (t ) = ln( qi 1
1
);
i +1qi
copula with generator LM , i.e., t

n parameters: = (1 e ), qi , . . . , qd (0, 1);


i=1 ui copula C :
C u1,j1 , . . . , ud,jd ; LM = n .
( )
( + (1 ) ui ) (1 ) ni=1 ui

i=1
( )
C u
)1

d nj ( qi (1e )
1 (1 e ) qi + 1 qi
1. GeometricGeometric hierarchical Archimedean copula: i=1 j=1 1e
ui,j
)1 ) .

= ln 1

(
d ni ( qi (1e )
qi et 1 (1 qi )
Bi ShifGeo (qi ) with LBi (t ) = 1(1qi )et
and i=1 j=1 1e
ui,j + 1 qi

LB (t ) = ln( qi 1
1
);
i
t +1qi 2. LogarithmicLogarithmic hierarchical Archimedean copula:
parameters: q, qi , . . . , qd (0, 1); t
copula C (see equation given in Box I). Bi Logarithmic (i ) with LBi (t ) = ln(1 i e )
ln(1i )
and
( t
)
(1(1i ) )
LB (t ) = ln
1
2. GeometricLogarithmic hierarchical Archimedean copula: i
;
i
ln(1i et ) parameters: = (1 e ), 1 = (1 e1 ), . . . , d =
Bi Logarithmic (i ) with LBi (t ) = and
ln(1i ) (1 ed ) (0, 1);
(1(1i )t )
LB (t ) = ln(
1
i
); copula C :
i
parameters: q, 1 = (1 e1 ), . . . , d = (1 ed ) ( d
(0, 1); ( ) 1 1
C u = ln 1 (1 e )
copula C : i
i=1
ui,j
( ) ( )
C u i 1e
ni 1e
( ) 1 1 e
ln 1 (1 ei ) .
q
i ui,j +1q
d ni

q i=1 ln
1 (1 ei ) 1 e
1e i
j=1 (1ei )
j=1
= ( )1 .
q
i ui,j +1q
d
i + (1 q)
d
ln 1 (1 ei )
ni
1 e 3. Logarithmic-Gamma hierarchical Archimedean copula:
i=1 i=1 j=1 (1ei )
i
Bi Gamma (i , 1) with LBi (t) = 1
and LB1 (t ) =
( )
1+t i
1
3. Geometric-Gamma hierarchical Archimedean copula: t 1;
i

i parameters: = (1 e ) (0, 1), 1 , . . . , d R+ ;


Bi Gamma (i , 1) with LBi (t ) = 1
and L1
( )
1+t B (t) = i copula C :
1
t 1;i

parameters: q (0, 1), i , . . . , d R+ ; ( )
C u =
1
ln 1 (1 e )
copula C :
( )
C u i
d ni ( ) 1
( ) 1 )i 1 e ui,j i
(ni 1) .
d ni ( q
q i=1 j=1 +1q i
(ni 1)
ui,j
1 e
= ( ) 1 )i . i=1 j=1
d ni (
1 (1 q)
q i
i=1 j=1 ui,j
+1q (ni 1) Obviously, hierarchical Archimedean copulas can be built as-
suming that the distributions of the rvs Bi belong to different
6 H. Cossette et al. / Insurance: Mathematics and Economics 76 (2017) 113

d ni ( ui,j
)
q qi
i=1 j=1 (
ui,j qqi ui,j 1 )
.
( )
C u = d ui,j
ni
( )
(1q) i =1 qi
d ( ni ( j=1
( ) ))
))
ui,j u qqi ui,j 1
1 (1 qi ) 1 ( i,j
i=1 j=1 ( )
(
ui,j qqi ui,j 1 d ni ui,j
i=1 1(1qi ) j=1 u qq u 1
i,j i i,j ( )

Box I.

( ) 1 ) ( ( ( q
)))
n2 ( n1 1 u1,i +1q 1
ln 1 exp
q
q j=1 u2,j
+1q 1 i=1 ln
C (u1 , u2 ) = ))) .
)1 ( ( ( q
1 u1,i +1q 1
(
n2
q n1
j=1 u2,j +1q 1 ln 1 exp i=1
ln
ln (1 ) 1

ln(1 )(1q)1

Box II.

families. Since we cannot list all possible hierarchical Archimedean It is clear that the sampled values of (U1,1 , U1,2 ) provide scat-
copulas that can obtained this way, we limit ourselves to pro- ter plots that are similar to those of the Clayton copula with a
viding, in the following example, an illustration of a hierarchical dependence subdued by the shifted geometric distribution. The
Archimedean copula assuming d = 2 subgroups with different same phenomenon happens for the sampled values of (U2,1 , U2,2 ).
families of distributions B1 and B2 . Since the dependence structures of the four other pairs (U1,1 , U2,1 ),
(U1,1 , U2,2 ), (U1,2 , U2,1 ), (U1,2 , U2,2 ) depend on M, their scatter plots
Example 3. Let d = 2 and M ShiftedGeo(q) with LM (t ) = are those of an AMH copula.
qet ln(1 et )
1(1q)et
, B1 Log( ) with LB1 (t) = ln(1 ) and finally, B2
)
Gamma(, 1) with LB2 (t) = t +1 1 . Then, the expression of the
(
2.8. Multi-level hierarchical Archimedean copulas
3-parameters copula is given by the equation given in Box II.
Until now, we have only considered one-level hierarchical
Archimedean copulas, for which the associated typical one-level
2.7. Sampling hierarchical Archimedean copulas tree representation is provided in Fig. 1. Naturally, it is possible
to design multi-level hierarchical Archimedean copulas by adding
Inspired from Marshall and Olkin (1988), the sampling pro- ramifications such that their dependence structures can be repre-
cedure for a hierarchical Archimedean copula C defined with a sented as multi-level trees.
multivariate compound distribution follows from its construction. Let U = (U1 , . . . , Ud ) be a random vector with a cdf defined by
The algorithm given just below aims to simulate samples of a a multi-level hierarchical Archimedean copula C , such that each
random vector U = (U1 , . . . , Ud ) (with Ui = (Ui,1 , . . . , Ui,ni ), for subgroup Ui , i = 1, . . . , ni , characterized by its genetic code (the
i = 1, . . . , d) whose multivariate cdf is the copula C . path in the hierarchy) ei and its dimension ni , can be written as
Ui = Uei ,j , j = 1, . . . , ni .
( )
Algorithm 1. Let C be a hierarchical Archimedean copula with d
subgroups and root M. Define U as a vector of standard uniformly This genetic code is represented by a vector in which the first
distributed rvs with cdf C . element corresponds to the root of the tree (0 by default) and the
1. Sample M; kth element represents the position of a node in the (k 1)th level,
2. For each subgroup (i = 1, . . . , d 2): within its mothers direct descendants, assuming that the leaves
are always at the far right. A value of 0 at the last position of the
2.1. Sample Bi,k for k = 1, . . . , M; vector represents a direct link between a leaf and a mother node.
M
2.2. Return i = k=1 Bi,k ;
The design of the hierarchical structure comes from considering
2.3. Sample Ri,j Exp(1) for j = 1, . . . , ni ; the primary distribution of the compound rvs as a random sum.
2.4. Return Ui,j = Li (Ri,j /i ) for j = 1, . . . , ni ; Let Ui be a subgroup with genetic code ei , and let be a right
truncated vector of ei , and k the (dim( ) + 1)th element of ei , where
3. Return U = (U1,1 , . . . , U1,n1 , . . . , Ud,1 , . . . , Ud,nd ). dim(.) is the dimension of a vector. On one hand, if dim( ) <
An application of Algorithm 1 is provided in the following dim(ei ) 1, then the rv M ( ,k) is defined as a random sum, i.e,
M ( ) ( ,k) ( ,k)
example. M ( ,k) = j=1 Nj where Nj is distributed as a discrete rv
N ( ,k) which plays the same role as the rv Bi in Section 2.1. On the
Example 4. Let C be the Geometric-Gamma hierarchical Archime- other hand, if dim( ) = dim(ei ) 1, then the procedure is similar
dean copula presented in Example 1 with d = 2 and n1 = n2 = to the one presented in Section 2.1.
2. To illustrate the pairwise dependence between components of A typical hierarchical structure can be depicted by a tree where
U = (U1,1 , U1,2 , U2,1 , U1,2 ), we give, in Fig. 2, a graph of sampled the root is always M (0) and all leaves are of type u. A node of type
values of U, where the parameters of the copula C are q = 0.1, M, at a given level of the tree, can have children of all available
1 = 0.04 and 2 = 0.2. Note that , in Fig. 2, corresponds to the types, meaning M, and u. Further, nodes can only have
Kendalls tau. u children (leaves), while leaves have no children at all.
H. Cossette et al. / Insurance: Mathematics and Economics 76 (2017) 113 7

Fig. 2. Pairwise graph of the 10 000 4-dimensional vectors of realizations sampled from the hierarchical Archimedean copula with multivariate compound distributions in
Example 4.

1. Fix j = 2
2. For each genetic code with dim(ei ) = j do:

2.1. Sample M (ei )


2.2. If the last element of ei is zero i.e. ei,dim(ei ) = 0:


i. Sample R Exp(1)
ii. Set the corresponding component of Ui as L (e )
( ) M i
R
M (ei )
Fig. 3. Example of a multi-level hierarchical structure.
2.3. Else:
i. Sample (ei )
To illustrate the notation, let us consider the tree construction ii. Sample R Exp(1)
depicted in Fig. 3: iii. Set the
The tree in Fig. 3 corresponds to the dependence structure of ( ) corresponding component of Ui as L (ei )
R
a random vector U = (U1 , . . . , U8 ) characterized by the following (ei )
genetic codes:
3. Set j = j + 1
e1 = (0, 1, 1) e2 = (0, 1, 2) e3 = (0, 1, 3) e4 = (0, 1, 0)
4. Repeat from 2
5. Return U.
e5 = (0, 2) e6 = (0, 3, 1) e7 = (0, 3, 0) e8 = (0, 0)
Algorithm 2 proceeds as follows: we start by simulating the
Note that, one can deduce a tree from any given set of genetic
root, which is always of type M. Next, we perform a series of
codes. This set of codes constitutes a simple tool to summarize all
the information in regard to the dependence structure depicted by simulations in the following order: simulate all u-children of the
the tree. root, simulate all -children of the root then simulate all M-
For any given multi-level hierarchical copula, it is possible to de- children of the root. Finally, we move the root to M and repeat the
rive a tailored sampling procedure by adapting the one described same logic over again.
in Algorithm 1. Algorithm 2 provides a sampling procedure for In the following illustration, we provide an example of a 2-level
several hierarchical levels, assuming the three different types of hierarchical Archimedean copula.
children.
Example 5 (Two level tree). We consider a 2-level hierarchical
Algorithm 2. Let C be a multi-level hierarchical copula with root Archimedean copula C , for which the associated 2-level tree repre-
M (0) and d subgroups. Define U as a vector of standard uniformly sentation is given in Fig. 4. The 6-dimensional copula C is the mul-
distributed rvs with cdf C . Let ei be the genetic code of each tivariate cdf of the random vector U = (U(0,2),1 , U(0,2),2 , U(0,1,1),1 ,
subgroup i and let ei be the vector ei without the last element. U(0,1,1),2 , U(0,1,2),1 , U(0,1,1),2 ). Let the rvs (0,2) , (0,1,1) , and (0,1,2)
8 H. Cossette et al. / Insurance: Mathematics and Economics 76 (2017) 113

copulas based on multivariate compound distributions provide


very flexible combinations of Archimedean copula families with-
out such restrictions. Given their similar nature, it is of interest
to push further their comparison. For that purpose, we use the
common mixture representation of a copula to write the proposed
hierarchical Archimedean copula. This will also prove useful for
further investigation of aggregation methods.
Using (20) and (21), the common mixture representation of a hi-
Fig. 4. Two level tree structure. erarchical Archimedean copula based on a multivariate compound
distribution is given by

n1 nd

be associated to the pairs (U(0,2),1 , U(0,2),2 ), (U(0,1,1),1 , U(0,1,1),2 ) and



1
u1,j , . . . , 1
( ) ( ) ( )
C u = L1 ,...,d L1 Ld ud,j
(U(0,1,2),1 , U(0,1,2),2 ). j=1 j=1
(0,1)
M (0) (0,1) (0,2)
M (0) (0,2)


We define M = k=1 Nk , = k=1 Bk and n1 nd
M (0,1) (0,1,i)

1
( )
1
( )
(0,1,i) = k=1 Bk , for i = 1, 2. = PM LB1 L u1,j LBd
1
L ud,j d
By first conditioning on M (0,1) and then on M (0) , the LST of j=1 j=1
(0,1,i) can be written as

d
ni

L ui,j fM (m) .
1
Lm
( )) ( )
= (29)
(
B
LB(0,1,i) (t ) .
i i
L (0,1,i) (t) = PM (0) PN (0,1) m=1 i=1 j=1

Since LmBi is the LST of a positive rv which we denote by Vi , we can


Similarly, we obtain the trivariate LST of ( (0,2) , (0,1,1) , (0,1,2) ), write the copula given in (29) as
which is given by
d (

(ui,j )
(n )
1
i L
)
L (0,2) , (0,1,1) , (0,1,2) (t1 , t2 , t3 )
vi

dFVi (vi ) fM (m).
( )
j=1
C u = e i (30)
0
( )) m=1 i=1
(
= PM (0) LB(0,2) (t1 ) PN (0,1) LB(0,1,1) (t2 ) LB(0,1,2) (t3 ) .
From the common mixture representation given in (30), we adapt
Algorithm 1 and provide below a more efficient one.
Thus, the 6-dimensional copula C can be written as
Algorithm 3. Let C be a hierarchical Archimedean copula with d
C (u(0,2),1 , u(0,2),2 , u(0,1,1),1 , u(0,1,1),2 , u(0,1,2),1 , u(0,1,2),2 )
subgroups and root M0 . Define U as a vector of standard uniformly
2
( distributed rvs with cdf C .
= L (0,2) , (0,1,1) , (0,1,2) L
1
(u
(0,2) (0,2),k
),
k=1
1. Sample M0 ;
2 2 2. For each subgroup (i = 1, . . . , d 2):
)
L
1
(u
(0,1,1) (0,1,1),k
), L
1
(u
(0,1,2) (0,1,2),k
) . (28) 2.1. Sample Vi with LVi (t) = LB 0 (t);
M
i
k=1 k=1 2.2. Sample Ri,j Exp(1) for j = 1, . . . , ni ;
We assume that M (0) ShiftedGeo(0.5), N (0,1) ShiftedGeo(0.1) 2.3. Return Ui,j = Li (Ri,j /Vi ) for j = 1, . . . , ni ;
and B(0,2) B(0,1,1) B(0,1,2) Gamma( = 30 1
, = 1). In 3. Return U = (U1,1 , . . . , U1,n1 , . . . , Ud,1 , . . . , Ud,nd ).
Fig. 5, considering that B(0,2) , B(0,1,1) and B(0,1,2) are iid, we clearly
see the impact of M (0) and M (0,1) on their respective dependence In order to confirm the efficiency of Algorithm 3, let us consider
structure. Using the numbered boxes with coordinates (x, y) in a simple one level tree hierarchical Archimedean copula with two
Fig. 5, the scatter plot associated to the pair (U(0,2),1 , U(0,1),2 ) (see groups where M Geo(q), B1 Gamma(2.5, 1) and B2
box (5, 1)) has a shape that is the closest to the one of a Clayton Gamma(2.5, 1). The top graph of Fig. 6 serves as an illustration to
copula with a dependence parameter of 30. This phenomenon is compare computation times (in seconds) between Algorithms 1
explained by the fact that the couple depends only on M (0) with a and 3 for 100 000 realizations with respect to q, while the bottom
one compares both algorithms with respect to the sampling size.
parameter of 0.5, which in turn has little to no impact on (0,2) .
Evidently, Algorithm 3 is far more efficient. In fact, Algorithm 3 has
The scatter plot for the pair (U(0,1,1),1 , U(0,1,1),2 ) (see box (1, 5)),
a very small computation time since we sample directly from the
however, shows a dependence restrained by M (0,1) , which has
known distribution of the sum in step 2.1 of Algorithm 3.
a much bigger impact due to its overall higher values. We can
also see that boxes {(1, 1), (1, 2), . . . , (4, 1), (4, 2)} only depend 4. Comparisons with other construction methodologies of hi-
on M (0) , which implies a very small dependence. However, boxes erarchical Archimedean copulas
{(1, 3), (1, 4), (2, 3), (2, 4)} depend on M (0,1) , a rv compounded on
M (0) . As discussed in Section 2.5, the copula generated from M (0,1) 4.1. Links with nested Archimedean copulas
is more concordant than the one generated from M (0) and hence
M (0,1) M (0) . Now that hierarchical Archimedean copulas with multivariate
compound distributions are written as common mixtures, we aim
3. Representation of the copula as a common mixture to compare them to nested Archimedean copulas. To do so, we con-
sider a simple example of a one level partially nested Archimedean
The functional symmetry of Archimedean copulas lead to the copula with d children copulas i.e.
introduction by Joe (1997) of nested Archimedean copulas. This
C u = C C u1 ; 1 , . . . , C ud ; d ; 0
( ) ( ( ) ( ) )
popular class of copulas results from nesting Archimedean copulas
into each other, allowing asymmetries and multiple hierarchy d ni

= 0 01 i i1 ui,j ,
( )
levels. Such constructions are however limited by nesting condi- (31)
tions that must be fulfilled. In contrast, hierarchical Archimedean i=1 j=1
H. Cossette et al. / Insurance: Mathematics and Economics 76 (2017) 113 9

Fig. 5. Pairwise graph of 10 000 6-dimensional vectors of realizations sampled from the hierarchical Archimedean copula with multivariate compound distributions in
Example 5.

where 0i (t ; r) = exp r 01 i (t) . The representation in


( )
(32) is a proper copula if and only if 0i is the LST of a positive rv,
say Wi , i.e.
d (

i1 (ui,j )
(n )
i
)
wi

dFWi (wi ) fR (r).
( )
j=1
C u = e (33)
r =1 i=1 0

Remark. It is important to highlight that verifying that 0i is a LST,


and hence that the sufficient nesting condition is verified, is not an
easy task. Nesting together different Archimedean families has its
limitations and parameter restrictions (see e.g. Hofert, 2010).
We can clearly see that the representation in (30) of the hierar-
chical Archimedean copula with multivariate compound distribu-
tions and the one in (33) of the nested Archimedean copula have
the same general form. In order to compare these two representa-
tions, we rewrite Lm
B (t) as i

1
mL
Lm
Bi (t) =e M
Li (t)
= 0i (t ; m). (34)
Fig. 6. Top: Computation time (sec) for Algorithms 1 and 3 with respect to
q. Bottom: Computation time (sec) for Algorithms 1 and 3 with respect to the Expressing Lm Bi as 0i allows us to represent the hierarchical
sampling size (n) with q = 0.01. Archimedean copula with multivariate compound distributions in
(30) as the partially nested Archimedean copula in (32), where
the mother copula C0 is generated from the rv M, and the child
copula Ci is generated from i . Note that Ci is an Archimedean
where is the generator of the Archimedean copula C . If we copula derived from i =
M
j=1 Bi,j , which is in general not a well
suppose that 0 is the LST of a discrete strictly positive rv R, then, known distribution and expressed in terms of the rv M. Given that
(31) becomes 0i = Lm Bi , we have that 0i solely depends in our structure on
the distribution of Bi unlike in the nested Archimedean copulas in
d

i1 (ui,j ) which both the mother and the child copulas affect 0i , leading
(n )
r 01 i i
( )
j=1
C u = e fR (r) to a restrictive nesting condition. Also, since Lm Bi depends only on
r =1 i=1 the parameters of Bi regardless of the distribution of the rv M,

d

ni
the choice of Ms distribution is very flexible in comparison to the
= 0i i1 (ui,j ); r fR (r), (32) nested Archimedean copulas, where nesting different families is a
r =1 i=1 j=1 recurrent problem.
10 H. Cossette et al. / Insurance: Mathematics and Economics 76 (2017) 113

The above remarks imply that the hierarchical Archimedean Table 1 of Hering et al. (2010) provides a short list of possible
copula C , as defined in (20) can be represented as a nested distributions for V , which can be either discrete or continuous. The
Archimedean copula as follows: random time rv V is used as a common factor to define the vector
of rvs
C (u) = C (C (u1 ; L1 ), . . . , C (ud ; Ld ); LM ). (35)
(V ) = (1) (V ) , . . . , (d) (V ) .
( )
Since the hierarchical Archimedean copulas proposed in this
paper can be written as a nested Archimedean copula, one may Given V , the rvs (i) (V ), for i = 1, . . . , d, are assumed to be
wonder if we can produce similar copulas from both approaches. conditionally independent. This implies that the multivariate LST
To do so, we have to find the distribution of Bi for which (35) is a of (V ) is obtained as follows :
known nested Archimedean copula. (1)
[ (d)
]
Let M be a positive and discrete rv and i (i = 1, . . . , d) be L(V ) (x1 , . . . , xd ) = E ex1 (V ) exd (V )
strictly positive rvs generating known Archimedean copulas. In [ [ (1 ) (d)
]]
this case, the distribution of the rvs Bi can be deduced from = EV E ex1 (V ) exd (V ) |V
= E eV 1 (x1 ) eV d (xd )
( )
1
[ ]
L Li (t)
1
.
( )
M
LBi (t) = PM Li (t) = e (36) [ V ( (x )++ (x )) ]
=E e 1 1 d d
From (34), (36) becomes
= LV (1 (x1 ) + + d (xd )) . (40)
Lm
Bi (t) = 01 (t ; m), (37)
Following a similar procedure to the one presented in Sec-
which means that tion 2.2, Hering et al. (2010) propose to construct multivariate
m
mixed exponential distributions using (V ). Let
Wi = Bi ,
Y = Y1,1 , . . . , Y1,n1 , . . . , Yd,1 , . . . , Yd,nd
( )
k=1

where Wi is the rv generating 01 (.; m). be a vector of n1 + + nd rvs. Given (V ) = , it is assumed that
To illustrate this idea, let (35) be a nested AMHAMH copula. Y1,1 | (V ) = , . . . , Y1,n1 | (V ) = , . . . ,
( ) ( )
To obtain such a copula with our approach, we must have M
Yd,1 | (V ) = , . . . , Yd,nd | (V ) =
( ) ( )
Geo(1
( ) 0 ) and i Geo(1 i ) which implies that Bi Geo
) given (V ) = =
1i
10
, 0 i (i = 1, . . . , d). Similarly, in order for (35) to be a are conditionally independent. Also,
( 1 , . . (.i), d ) , the)rv Yi,j | (V ) = is identically distributed as
(
(
nested JoeJoe copula, our approach requires that M
( Sibuya( 0 )
Yi,j | (V () = i for j = 1), 2, . . .(, ni , i = 1, 2, . . .), d. We as-
)
i
and i Sibuya(i ) which implies that Bi Sibuya 0
, i
sume that Yi,1 |(i) (V ) = i , . . . , Yi,ni |(i) (V ) = i are expo-
0 (i = 1, . . . , d).
nentially distributed with parameter i , for i = 1, . . . , d. Then, Y
The representation in (35) can be easily generalized to multi-
follows a multivariate mixed exponential distribution with multi-
level hierarchical structures by adapting the notation presented in
variate survival function given by
Section 2.8. For example, (28) can also be written as
F Y y1,1 , . . . , y1,n1 , . . . , yd,1 , . . . , yd,nd
( )
C (u1 , u2 , u3 ) = C C C u1 ; L (0,1,1) , C u2 ; L (0,1,2) ; LM (0,1) ,
( ( ( ) ( ) )
(1) (V )y (1) (V )y
[
C u3 ; L (0,2) ; LM (0) , = E e 1 ,1 e 1,n1
( ) )
(38)
where u1 = (u(0,1,1),1 , u(0,1,1),2 ), u2 = (u(0,1,2),1 , u(0,1,2),2 ) and (d) (V )y (d) (V )y
]
e d,1 e d,nd
. (41)
u3 = (u(0,2),1 , u(0,2),2 ). Notation wise, (38) can be simplified to
Due to (40), we can write (41) as
C (u1 , u2 , u3 )
F Y y1,1 , . . . , y1,n1 , . . . , yd,1 , . . . , yd,nd
( )
= C (0) C (0,1) C (0,1,1) u1 , C (0,1,2) u2 , C (0,2) u3 .
( ( ( ) ( )) ( ))
(39)
= L(V ) y1,1 + + y1,n1 , . . . , yd,1 + + yd,nd
( )

= 0 1 y1,1 + + y1,n1 + + d yd,1 + + yd,nd .


( ( ) ( ))
4.2. Approach based on Lvy subordinators and random time
(42)
In this section, we examine the link between the approach
proposed by Hering et al. (2010) to construct hierarchical Also, the univariate survival function of Yi,j is
Archimedean copulas and the one described within this paper. [ (i) (V )y
]
F Yi,j yi,j = E e = L(i) (V ) yi,j = LV i yi,j
( ) ( ) ( ( ))
i ,j
The construction approach by Hering et al. (2010) is based on
increasing Lvy processes, also called Lvy subordinators. They are
= 0 i yi,j = 0 i yi,j ,
( ( )) ( )
(43)
notably important components for building models in financial
mathematics (see { e.g. Cont and}Tankov, 2003). for yi,j 0, j = 1, 2, . . . , ni , and i = 1, . . . , d.
Let (i) = (i) (t ) , t 0 be a Lvy subordinator, for i = Then, using Sklars Theorem, the copula associated to F Y is given
1, 2, . . . , d. For all x, t 0, the LST of (i) (t ) is given by by
(i)
[ ]
L(i) (t ) (x) = E ex (t ) = et i (x) ,
( 1 1
C u = F Y F Y1,1 u1,1 , . . . , F Y1,n u1,n1 , . . . ,
( ) ( ) ( )
1

where the function i corresponds to the Laplace exponent of (i) , 1 1 ))


F Yd,1 ud,1 , . . . , F Yd,n ,
( ) (
ud,nd (44)
for i = 1, 2, . . . , d. d

Let a strictly positive rv V be a random time with LST defined 1


where F Yi,j is obtained from (43) i.e.
by
1
LV (x) = E eVx = 0 (x) ,
[ ]
F Yi,j ui,j = L1
ui,j = i1 01 ui,j = i1 01 ui,j ,
( ) ( ) ( ( )) ( )
(i) (V )
for x 0. (45)
H. Cossette et al. / Insurance: Mathematics and Economics 76 (2017) 113 11

for ui,j [0, 1], j = 1, 2, . . . , ni , and i =( 1), . . . , d. Combining (44) strategies appear similar. Both approaches can lead to the same
and (45) with (42), the expression for C u becomes hierarchical Archimedean copula in some cases. However, others
( can only be obtained with one of the two methods. For all the
C u = L(V ) L1
u1,1 + + L1
u1,n1 , . . . ,
( ) ( ) ( )
(1) (V ) (1) (V )
popular L vy subordinators provided in Table 2 of Hering et al.
)) (2010), it is possible to define the distribution of Bi such that its
Laplace exponent i is equal to the Laplace exponent i of (i) ,
1 1
ud,1 + + L
( ) (
L
(d) (V ) (d) (V )
u1,nd
( d
) i = 1, 2, . . . , d. For example, if Bi follows a gamma distribution
with parameters and 1, then i (x) = ln (1 + x) as in (ii) of
= 0 i i 0 ui,1 + + i 0 ui,ni .
( 1 1
( ) 1 1
( ))
Table 2 ( is a scaling factor) then the expressions for L(i) (V ) and
i=1
Li are identical (assuming LV = LM ). If B( i follows an inverse
In Hering et al. (2010), conditions for admissible Lvy subordina-
)
Gaussian with parameters 1

and 1, i (x) = 2x
2
+ 1 1 as in
tors are given in Theorem 2.1 and a list of popular Lvy subordina-
tors can be found in Table 2. (iii) of Table 2 ( is a scaling factor) then it follows that L(i) (V )
We may now compare both approaches and the resulting hi- and Li have the same expression (assuming LV = LM ). Also,
erarchical Archimedean copulas. Under both approaches, a hier- under the assumption LV = LM (i.e. V is a discrete rv) and since
archical Archimedean copula is identified from the joint survival the rv Bi can follow many distributions including the ones listed
function of a multivariate mixed exponential distribution. In both in Table 1 of Hering et al. (2010), the proposed method provides
cases, it leads to natural generic sampling algorithms for hier- a larger range of possible hierarchical Archimedean copulas in
archical Archimedean copulas. The dependence structure of the comparison to Hering et al. (2010)s construction method. If V is
multivariate mixed exponential distribution is defined through a a continuous rv, the resulting hierarchical Archimedean copula
vector of dependent mixing rvs, obtained with Hering et al. (2010)s approach cannot be obtained
with our approach.
(V ) = (1) (V ) , . . . , (d) (V )
( )
Finally, inspired from Hering et al. (2010), the expression of the
hierarchical Archimedean copula constructed under our proposed
in Hering et al. (2010), and through
approach given in (20) becomes
= (1 , . . . , d )
1 u1,1 + + L1 u1,n1 , . . . , Ld ud,1
1 1 1
C u = L L
( ) ( ( ) ( ) ( )

within this paper. Even if these two vectors of rvs are defined + + L 1
( ))
d ud,nd
through different probabilistic arguments (Lvy subordinators ( d )
with a common random time and random sums with a common (
i i LM ui,1 + + i LM ui,ni
1 1 1 1
.
( ) ( ))
counting rv), both approaches lead to identical generic structures = LM
for L(V ) and L and for L(i) (V ) and Li . The expression for L(i) (V ) i=1

is
4.3. Estimation procedure and determination of the tree structure
L(i) (V ) (x) = LV (i (x)) ,

for i = 1, . . . , d. Let us define the function i associated to LBi such The main objective of the present work is to propose an al-
that ternative approach to construct hierarchical Archimedean copu-
las. The interestingness of these new copulas also relies on the
LBi (x) = exp (i (x)) , capability to estimate them. Recently, different research works
or, equivalently, have appeared in the literature on the estimation of hierarchical
Archimedean copulas, more specifically on the structure determi-
i (x) = ln LBi (x) ,
( )
nation and the parameter estimation. Okhrin et al. (2013) appears
to be the first paper to address simultaneously both of these
for i = 1, 2, . . . , d. Then, the expression for Li becomes
tasks through a multi-stage procedure in a bottom-up manner.
Li (x) = PM LBi (x) = LM ln LBi (x) = LM (i (x)) ,
( ) ( )
Maximum-likelihood estimation is used for the parameters and the
inversion of Kendalls tau is also suggested. The structure deter-
for i = 1, . . . , d. Note that mination is investigated in several ways, with notably approaches
L(V ) (x1 , . . . , xd ) = LV (1 (x1 ) + + d (xd )) based on a goodness-of-fit test or binary trees. In Grecki et al.
(2016a) and Grecki and Holea (2013), the estimator for both
and the structure and the parameters of hierarchical Archimedean
L (x1 , . . . , xd ) = LM (1 (x1 ) + + d (xd )) . copulas is based on Kendalls tau, more precisely on the inversion of
Kendalls tau estimator. They use agglomerative hierarchical clus-
The rvs V and M play similar roles as a common factor. As tering (with three different definitions for the distance between
mentioned earlier, the rv V can be either discrete or continuous. clusters) in which the relationship between two rvs is established
Under the approach proposed within this paper, M is a strictly through their Kendalls tau. Other research papers, such as Uyt-
positive discrete rv. If V is a strictly positive discrete rv, than LV and tendaele (2016) and Segers and Uyttendaele (2014) estimate the
LM are identical. Also, both LV and LM correspond to the generator structure differently by considering either trivariate structures or
0 in the definition of a nested Archimedean copula. supertrees.
Finally, both i and i correspond to the function 01 1 which These works mainly focus on hierarchical Archimedean copulas
can be found in the definition of the nested Archimedean copula in which all copulas in the tree are from the same Archimedean
and for which it is difficult to verify the conditions of admissibility. family. An excellent paper, Grecki et al. (2016b), paves the way
In conclusion, the approaches proposed both in Hering et al. (2010) to the estimation of the structure and the parameters of hierar-
and within the present paper circumvent this difficulty. Note that chical Archimedean copulas involving different Archimedean fam-
i plays the same role as i (the Laplace exponent of the Lvy ilies. This new estimation procedure, which is in part based on
subordinator (i) ). goodness-of-fit tests, adopts estimation algorithms from previous
The construction method provided in this paper and Hering papers in a way that guarantees the verification of the nesting
et al. (2010)s approach have their own advantages even if their condition. In the same context of different family generators, Zhu
12 H. Cossette et al. / Insurance: Mathematics and Economics 76 (2017) 113

et al. (2016) recently proposed an approach using a three stage Table 2


estimation approach based on a clustering procedure to choose the Values of the variance, VaR and TVaR of S = X1,1 + + X1,40 + X2,1 + + X2,40
where the joint cdf FX1,1 ,...,X1,40 ,X2,1 ,...,X2,40 is as defined in Example 6.
optimal hierarchical structure as an estimation of a Lvy subordi-
nated hierarchical Archimedean copula. Risk measures Cossette et al. (2017) Sampling (1M)
Given the similarities of our construction with the one of Hering Var(S) 1157.4461 1156.8011
et al. (2010) discussed in Section 4.2, the approach suggested in Zhu VaR0.9 (S) 193.0000 193.0000
TVaR0.9 (S) 214.4829 214.4334
et al. (2016) is evidently a good starting point for the estimation of VaR0.99 (S) 240.0000 240.0000
our structure based on maximum likelihood. This should be com- TVaR0.99 (S) 252.1244 252.2950
pared with the method proposed in Grecki et al. (2016b), based on VaR0.999 (S) 267.0000 268.0000
Kendalls tau estimator, which allows the use of Archimedean cop- TVaR0.999 (S) 276.1494 276.4880
ulas from different copula families as the proposed construction VaR0.9999 (S) 287.0000 288.0000
TVaR0.9999 (S) 293.5822 294.8900
here. Contrarily to the nested Archimedean copulas approach, we
will not be confronted with the difficulties surrounding the nesting
condition. However, the flexibility of our new structure leads to d
new challenges in regard to finding the optimal solution among all Let S = i=1 Si , where Si = Xi,1 + + Xi,ni , for i = 1, 2, . . . , d.
possible structures. This mainly concerns for us, firstly, the choice To find the desired cdf of S, one may apply the approach proposed
of the appropriate LST of M and then the choice of the LSTs of in Cossette et al. (2017). To do so however, we must assume Vi to be
the rvs Bi (i = 1, . . . , d). This will be followed by the estimation a strictly positive discrete rv defined on {1, 2, . . .}, with probability
of the parameters of the hierarchical Archimedean copulas. These mass function fVi (vi ) = Pr(Vi = vi ), Vi = 1, 2, . . . In such a case,
questions are at the core of another ongoing research project. (47) becomes
FX x1,1 , . . . , xd,nd
( )
5. Aggregation method ( n )
d
i
( ))
fVi (vi ) fM (m).
(
Risk aggregation comes into play for tasks such as the analysis = FXi,j |M =m,Vi =vi xi,j
of a risk portfolio and regularly capital calculations within and m=1 i=1 vi =1 j=1
between risk categories. In this section, we examine the behav- We provide next an example to illustrate the applicability of the
ior of aggregated dependent risks with multivariate cdf defined aggregation procedure proposed in Cossette et al. (submitted for
through a hierarchical Archimedean copula based on multivariate publication) in the context of the new construction given in this
compound distributions. paper. Since the method of Cossette et al. (submitted for publica-
Let X = (X1,1 , . . . , X1,n1 , . . . , Xd,1 , . . . , Xd,nd ) be a random vec- tion) leads to exact results, the following example also shows the
tor with the following multivariate cdf defined with an underlying precision of the values obtained with Algorithm 3.
hierarchical Archimedean copula as given in (21) and marginals
FXi,j (for i = 1, . . . , d and j = 1, . . . , ni ) a portfolio of 80 risks X = X1,1 , . . . , X1,40 ,
(
Example 6. Consider
X2,1 , . . . , X2,40 with a multivariate cdf defined as in (46) with
)
FX x1,1 , . . . , x1,n1 , . . . , xd,1 , . . . , xd,nd
( )
d = 2 and n1 = n2 = 40. We assume M Logarithmic(0 .5),)
B1 Geo(0.8) and B2 Geo(0.9). Let Xs,i Binom 10, qs,i
( (
= C FX1,1 x1,1 , . . . , FX1,n1 x1,n1 , . . . ,
( ) ( )
where qs,i = 0.05 s + 0.005i, s = 1, 2 and i = 1, 2, . . . , 40. It
)) 2 40
FXd,1 xd,1 , . . . , FXd,n xd,nd implies that E [S ] = 142. Relevant measures of S = i=1 Xs,i
( ) (
s=1
d
can be obtained with the approach of Cossette et al. (submitted for
n1 nd publication) or with MC simulations using Algorithm 3. Note that

L1 FX1,i x1,i , . . . , Ld FXd,j xd,j .
1 1
( ( )) ( ( ))
= L (46) the distributions of M and Bi are not from the same families. Both
i=1 j=1 risk measures, VaR and TVaR, are given in Table 2 to illustrate rele-
vant results for both methods as a mean of comparison. Note that
Since the vectors of rvs (Xi,1 | i = i ), . . . , (Xi,ni | i = i ),
the approach of Cossette et al. (submitted for publication) always
for i 1, . . . , d, are conditionally independent given i = i and
provides exact results for discrete marginals FXi,j . Moreover, the
(X1,j1 | M = m), . . . , (Xd,jd | M = m), for ji 1, . . . , ni , are also
simulation results are very close to the actual results, considering
conditionally independent given M = m, FX can be represented as
1 million simulations were done.
a common mixture of conditional cdfs as

FX x1,1 , . . . , x1,n1 , . . . , xd,1 , . . . , xd,nd


( )
6. Conclusion


FX |M =m x1,1 , . . . , x1,n1 , . . . , xd,1 , . . . , xd,nd fM (m)
( )
= A new hierarchical Archimedean copulas construction method
m=1 involving multivariate compound distributions was presented. The
ni absence of nesting and marginal conditions enlarges the possi-
d

( )

FXi,j |M =m,i =i xi,j dF i (i ) fM (m).
( )
= (47) bilities of nested copulas by improving the flexibility in both the
m=1 i=1 0 j=1 choice of families and parameters. Moreover, new copulas with
multiple parameters were derived allowing for a large variety of
To find the expression of the conditional cdfs FXi,j |M =m,i =i , we dependence structures. In addition to the new parametric copu-
use the common mixture representation given in (30). We obtain
las, well-known Archimedean copulas can be obtained as special
cases. Closed form expressions have been derived for the copulas.
FX x1,1 , . . . , xd,nd
( )
To further complement our theoretical results, efficient sampling
d (

(xi,j ) algorithms have been developed for computational applications.
(n ( ))
1
i L
)
vi FX

= e j=1 i i,j dFVi (vi ) fM (m). (48)
0
m=1 i=1 Acknowledgments
( )
Given (47) and (48), we have FXi,j |M =m,Vi =vi xi,j =
This work was partially supported by the Natural Sciences
FXi,j (xi,j )
( ( ))
1
vi L
e i , for i = 1, . . . , d 2 and j = 1, . . . , ni . and Engineering Research Council of Canada (Cossette: 054993;
H. Cossette et al. / Insurance: Mathematics and Economics 76 (2017) 113 13

Marceau: 053934) and by the Chaire en actuariat de lUniversit Hering, C., Hofert, M., Mai, J.-F., Scherer, M., 2010. Constructing hierarchical
Laval (Cossette, Gadoury, Marceau, and Mtalai: FO502323). Helpful archimedean copulas with lvy subordinators. J. Multivariate Anal. 101 (6),
14281433.
comments by two anonymous referees on earlier versions of the
Hofert, M., 2010. Sampling Nested Archimedean Copulas with Applications to cdo
manuscript have been highly appreciated. Pricing. (Ph.D. Thesis), Universitt Ulm.
Hofert, M., 2011. Efficiently sampling nested Archimedean copulas. Comput. Statist.
References Data Anal. 55 (1), 5770.
Hofert, M., 2012. A stochastic representation and sampling algorithm for nested
Bedford, T., Cooke, R.M., 2002. Vines: A new graphical model for dependent random archimedean copulas. J. Stat. Comput. Simul. 82 (9), 12391255.
variables. Ann. Statist. 10311068. Joe, H., 1997. Multivariate Models and Multivariate Dependence Concepts. CRC
Brechmann, E.C., 2014. Hierarchical kendall copulas: Properties and inference. Press.
Canad. J. Statist. 42 (1), 78108. Joe, H., 2014. Dependence Modeling with Copulas. CRC Press.
Cont, R., Tankov, P., 2003. Financial Modelling with Jump Processes, vol. 2, 2003. Kimberling, C.H., 1974. A probabilistic interpretation of complete monotonicity.
Chapman & Hall/CRC, London. Aequationes Math. 10 (2), 152164.
Cossette, H., Marceau, E., Mtalai, I., Veilleux, D., 2017. Archimedean copulas: aggre- Marshall, A.W., Olkin, I., 1988. Families of multivariate distributions. J. Amer. Statist.
gation, capital allocation and other applications. Available At http://Ssrn.Com/ Assoc. 83 (403), 834841.
Abstract=2957181. McNeil, A.J., 2008. Sampling nested archimedean copulas. J. Stat. Comput. Simul.
Devroye, L., 2009. Random variate generation for exponentially and polynomially 78 (6), 567581.
tilted stable distributions. ACM Trans Modeling Comput. Simul. 19 (4), 18. Nelsen, R.B., 2007. An Introduction to Copulas. Springer Science & Business Media.
Feller, W., 1971. An Introduction to Probability and Its Applications, Vol. II. Wiley, Okhrin, O., Okhrin, Y., Schmid, W., 2013. On the structure and estimation of hierar-
New York. chical archimedean copulas. J. Econometrics 173 (2), 189204.
Grecki, J., Hofert, M., Holea, M., 2016a. An approach to structure determination Segers, J., Uyttendaele, N., 2014. Nonparametric estimation of the tree structure of
and estimation of hierarchical archimedean copulas and its application to a nested archimedean copula. Comput. Statist. Data Anal. 72, 190204.
bayesian classification. J. Intell. Inf. Syst. 46 (1), 2159. Sklar, M., 1959. Fonctions de rpartition n dimensions et leurs marges. Universit
Grecki, J., Hofert, M., Holea, M., 2016b. On structure, family and parameter esti- Paris 8.
mation of hierarchical archimedean copulas. ArXiv Preprint arXiv:1611.09225. Uyttendaele, N., 2016. On the estimation of nested archimedean copulas: A theo-
Grecki, J., Holea, M., 2013. Structure determination and estimation of hierarchical retical and an experimental comparison.
archimedean copulas based on kendall correlation matrix. In: International Zhu, W., Wang, C.-W., Tan, K.S., 2016. Structure and estimation of Lvy subordinated
Workshop on New Frontiers in Mining Complex Patterns. Springer, hierarchical Archimedean copulas (LSHAC): Theory and empirical tests. J. Bank.
pp. 132147. Finance 69, 2036.