fragility index of block tailed vectors
TRANSCRIPT
Contents lists available at SciVerse ScienceDirect
Journal of Statistical Planning and Inference
Journal of Statistical Planning and Inference 142 (2012) 1837–1848
0378-37
doi:10.1
n Corr
E-m
journal homepage: www.elsevier.com/locate/jspi
Fragility index of block tailed vectors
Helena Ferreira a, Marta Ferreira b,n
a Department of Mathematics University of Beira Interior, Covilh ~a, Portugalb Department of Mathematics – Center of Mathematics, University of Minho, Braga, Portugal
a r t i c l e i n f o
Article history:
Received 23 December 2011
Accepted 30 January 2012Available online 7 February 2012
Keywords:
Multivariate extreme value theory
Tail dependence
Fragility index
Extremal coefficients
58/$ - see front matter & 2012 Elsevier B.V. A
016/j.jspi.2012.01.021
esponding author.
ail address: [email protected] (M. F
a b s t r a c t
Financial crises are a recurrent phenomenon with important effects on the real
economy. The financial system is inherently fragile and it is therefore of great
importance to be able to measure and characterize its systemic stability. Multivariate
extreme value theory provide us such a framework through the fragility index (Geluk
et al., 2007; Falk and Tichy, to appear-a, to appear-b). Here we generalize this concept
and contribute to the modeling of the stability of a stochastic system divided into
blocks. We will find several relations with well-known tail dependence measures in the
literature, which will provide us immediate estimators. We end with an application to
financial data.
& 2012 Elsevier B.V. All rights reserved.
1. Introduction
In the last decade, dependencies between financial asset returns have increased, mostly as a consequence ofglobalization effects and relaxed market regulation. Therefore, the concept of tail dependence has been discussed infinancial applications related to market or credit risk, e.g., Hauksson et al. (2001), Ane and Kharoubi (2003), Junker andMay (2005), Straetmans et al. (2008) and Embrechts and Puccetti (2006). The natural framework to model extremaldependence turns out to be the multivariate extreme value theory. The study of systemic stability is an important issuewithin this context of extreme risk dependence. The fragility of a system has been addressed to the fragility index (FI)introduced in Geluk et al. (2007). More precisely, consider a random vector X¼ ðX1, . . . ,XdÞ and Nx :¼
Pdi ¼ 1 1fXi 4xg the
number of exceedances among X1, . . . ,Xd above a threshold x. The FI corresponding to X is the asymptotic conditionalexpected number of exceedances, given that there is at least one exceedance, i.e., FI¼ limx-1 EðNx9Nx40Þ. The stochasticsystem fX1, . . . ,Xdg is called fragile whenever FI41. Theoretical developments, namely, the asymptotic distribution of Nx
conditional to Nx40, can be seen in Falk and Tichy (to appear-a, to appear-b).In this work we generalize some properties of the FI presented in the references above, contributing to the modeling of
the stability of a stochastic system divided into blocks.We shall state some notation that will be used throughout the paper.Consider D¼ fI1, . . . ,Isg a partition of D¼ f1, . . . ,dg. For the random vector X¼ ðX1, . . . ,XdÞ, let XIj
be a sub-vector of Xwhose components have indexes in Ij, with j¼ 1, . . . ,s. If F denotes the d.f. of X, then FIj
denotes the d.f. of sub-vector XIj,
j¼ 1, . . . ,s, and Fi the marginal d.f., i¼ 1, . . . ,d. Let xIjbe a vector of length 9Ij9 with components equal to x 2 R. We will
say that XIjis the jth block of random vector X and denote by Nx the number of blocks where it occurs at least one
ll rights reserved.
erreira).
H. Ferreira, M. Ferreira / Journal of Statistical Planning and Inference 142 (2012) 1837–18481838
exceedance of x, i.e.
Nx ¼Xs
j ¼ 1
1fXIj� xIjg:
All operations and inequalities on vectors are meant componentwise.
Definition 1.1. The fragility index (FI) of a random vector X¼ ðX1, . . . ,XdÞ relative to partition D is
limx-1
EðNx9Nx40Þ ð1Þ
whenever the limit exists and is denoted FIðX,DÞ.
If we consider Ij ¼ fjg, j¼ 1, . . . ,d, we find the FI introduced in Geluk et al. (2007) and latter study by Falk and Tichy(to appear-a, to appear-b). This partition will be denoted as Dn, i.e., Dn ¼ fIj ¼ fjg : j¼ 1, . . . ,dg.
We give particular emphasis to random vectors in the domain of attraction of a multivariate extreme value distribution(MEV) and consider either the case of identically distributed margins or tail equivalent margins in the sense considered inFalk and Tichy (to appear-b). In Section 2 we present some asymptotic properties of the distribution of Nx conditionalto Nx40, and find generalizations of results in Falk and Tichy (to appear-b). We prove that FIðX,DÞ exists and relates withthe extremal coefficients E of Tiago de Oliveira (1962) and Smith (1990) in the case of identically distributed margins(Section 3). We define generalized versions of the multivariate tail dependence coefficients of Li (2009), and extend someof its results (Section 4). In Section 5 we relate this latter coefficients with FIðX,DÞ.
For independent margins we have an unit FI. However, the stability of a stochastic system at higher levels can also becharacterized by asymptotic independence (Geluk et al., 2007). Asymptotic independence means that the dependencywhen present vanishes at extreme quantiles and the system is said to be weakly fragile, albeit possibly correlated (e.g.,Gaussian vectors). We extend the concept of asymptotic independent FI in Geluk et al. (2007) for blocks. A second measureis also presented by extending the 2-blocks asymptotic independent coefficient in Ferreira and Ferreira (2011) to the caseof s-blocks, with s42. This issue is considered in Section 6.
Our results relating FIðX,DÞ with well-known tail dependence measures, for which estimators and respective propertieshave already been study in the literature, will provide us immediate estimators (Section 7). We end with an application tofinancial data.
2. Asymptotic properties of Nx
In this section we present some asymptotic properties of the distribution of Nx conditional to Nx40. We start to relatethis latter with EG
A , the extremal coefficients (Tiago de Oliveira, 1962; Smith, 1990) of the sub-distribution functions of theMEV G corresponding to margins in A, i.e., by assuming G has unit Frechet margins
EGA ¼�log Gð1�1
A ð1Þ, . . . ,1�1A ðdÞÞ, ð2Þ
where, for all A � D, x 2 R
1�1A ðxÞ ¼
1, x 2 A,
1, x =2 A:
(
They may be written through the stable tail dependence function (Huang, 1992)
lGðx�11 , . . . ,x�1
d Þ ¼�log Gðx1, . . . ,xdÞ ¼�log CGðe�1=x1 , . . . ,e�1=xd Þ, ð3Þ
where CG is the copula of G, i.e.
Gðx1, . . . ,xdÞ ¼ CGðG1ðx1Þ, . . . ,GdðxdÞÞ, ðx1, . . . ,xdÞ 2 Rd: ð4Þ
In the sequel we will use notation I ðAÞ ¼ [j2AIj.
Proposition 2.1. If X has d.f. F with identically distributed continuous margins and belongs to the domain of attraction of a MEV
G with unit Frechet margins then, for each k 2 f1, . . . ,sg, we have
limx-1
PðNx ¼ k9Nx40Þ ¼1
EGD
XS�f1,...,sg;9S9 ¼ k
XT�S
ð�1Þ9T9þ1EGI ðT[SC
Þ:
Proof. We have, successively
PðNx ¼ k9Nx40Þ ¼1
1�PðNx ¼ 0Þ
XS�f1,...,sg;9S9 ¼ k
P \j2S
XIj�xIj
,\j=2S
XIjrxIj
� �
¼1
1�PðNx ¼ 0Þ
XS�f1,...,sg;9S9 ¼ k
XT�S
ð�1Þ9T9P \j2T[SC
XIjrxIj
!
H. Ferreira, M. Ferreira / Journal of Statistical Planning and Inference 142 (2012) 1837–1848 1839
¼1
1�FðxÞ
XS�f1,...,sg;9S9 ¼ k
XT�S
ð�1Þ9T9þ1ð1�Fðx1�1
I ðT[ScÞð1Þ, . . . ,x1
�1I ðT[Sc
ÞðdÞÞÞ ð5Þ
sinceP
T�Sð�1Þ9T9þ1¼ 0. Assuming w.l.o.g. that F has unit Pareto marginals, we obtained
PðNx ¼ k9Nx40Þ ¼1
1�CF 1�1
x1
� � XS�f1,...,sg;9S9 ¼ k
XT�S
ð�1Þ9T9þ1 1�CF 1�1
x1�1I ðT[Sc
Þð1Þ, . . . ,1�1
x1�1I ðT[Sc
ÞðdÞÞ
� �:
�ð6Þ
By hypothesis, F belongs to the domain of attraction of a MEV G, which is equivalent to de Haan and de Ronde (1998):
limt-1
1�CF ð1�y1=x, . . . ,1�yd=xÞ
1=x¼�log CGðe
�y1 , . . . ,e�yd Þ, ðy1, . . . ,ydÞZ0: ð7Þ
Taking limits in (5) and dividing both members by 1=x, conditions (6) and (7) lead us to
limx-1
PðNx ¼ k9Nx40Þ ¼1
�log CGðe�1, . . . ,e�1Þ
XS�f1,...,sg;9S9 ¼ k
XT�S
ð�1Þ9T9þ1ð�log CGðe
�1I ðT[Sc Þð1Þ, . . . ,e�1I ðT[Sc ÞðdÞÞÞ
¼1
EGD
XS�f1,...,sg;9S9 ¼ k
XT�S
ð�1Þ9T9þ1EGI ðT[SC
Þ: &
The previous result can be generalized to random vectors X with equivalent marginal distributions, in the sense thatthere exists a d.f. H such that
limx-wðHÞ
1�FiðxÞ
1�HðxÞ¼ gi 2 ð0,1Þ, i¼ 1, . . . ,d, ð8Þ
where w(H) is the right-end-point of H. In this case it is no longer possible an interpretation based on extremal coefficients,as can be seen in the following result.
Proposition 2.2. If X has d.f. F with equivalent marginal distributions in the sense of (8), and belongs to the domain of
attraction of a MEV G with unit Frechet margins then, for each k 2 f1, . . . ,sg, we have
limx-wðHÞ
PðNx ¼ k9Nx40Þ ¼1
log CGðe�g1 , . . . ,e�gd Þ
XS�f1,...,sg;9S9 ¼ k
XT�S
ð�1Þ9T9þ1log CGðe�g11I ðT[Sc Þð1Þ, . . . ,e�gd1I ðT[Sc ÞðdÞÞ:
Proof. Observe that
1�Fðx1�1I ðT[Sc
Þð1Þ, . . . ,x1�1I ðT[Sc
ÞðdÞÞ ¼ 1�CF ð1�ð1�F1ðx1�1I ðT[Sc
Þð1ÞÞÞ, . . . ,1�ð1�F1ðx1�1I ðT[Sc
ÞðdÞÞÞÞ
¼ 1�CF ð1�tI ðT[ScÞ
1 ðxÞ, . . . ,1�tI ðT[ScÞ
d ðxÞÞ, ð9Þ
where
tAi ðxÞ ¼ 1AðiÞ
1�FiðxÞ
1�HðxÞð1�HðxÞÞ, i¼ 1, . . . d:
Applying (8), we have
limx-wðHÞ
1�CF ð1�tI ðT[ScÞ
1 ðxÞ, . . . ,1�tI ðT[ScÞ
d ðxÞÞ
1�HðxÞ¼ �log CGðe
�g11I ðT[Sc Þð1Þ, . . . ,e�gd1I ðT[Sc ÞðdÞÞ: ð10Þ
The result follows by retaking expression in (5) and considering (9) and (10). &
If in particular we consider Ij ¼ fjg, j¼ 1, . . . ,d, we find the result of Falk and Tichy (to appear-b).The results above can also be obtained through the relation between CG and D-norms presented in Aulbach et al.
(to appear). However, we have chosen to present self-contained proofs using the usual arguments of multivariate extremevalue theory that are more familiar.
3. The fragility index for blocks
In this section we compute the FI for blocks given in (1), whenever X has equally distributed or tail equivalent marginsin the sense of (8), belonging to the domain of attraction of a MEV G with unit Frechet margins.
H. Ferreira, M. Ferreira / Journal of Statistical Planning and Inference 142 (2012) 1837–18481840
Proposition 3.1. If X has d.f. F with identically distributed continuous margins and belongs to the domain of attraction of a MEV
G with unit Frechet margins, we have
FIðX,DÞ ¼Ps
j ¼ 1 EGIj
EGD
:
Proof. Observe that
FIðX,DÞ ¼Xs
j ¼ 1
limx-1
PðXIj�xIj
9Nx40Þ ¼Xs
j ¼ 1
limx-1
1�Fðx1�1Ijð1Þ, . . . ,x1�1
IjðdÞÞ
1�Fðx, . . . ,xÞ
¼Xs
j ¼ 1
limx-1
1�CF 1�1
x1Ijð1Þ, . . . ,1�
1
x1IjðdÞ
� �=ð1=xÞ
1�CF 1�1
x, . . . ,1�
1
x
� �=ð1=xÞ
¼1
EGD
Xs
j ¼ 1
EGIj: &
In particular, for partitions corresponding to the family of all margins we obtain the known result FIðX,DnÞ ¼ d=EGD.
Hence, the FI of the system divided into blocks is smaller than the system itself, i.e., FIðX,DÞrFIðX,DnÞ.
Remark 3.1. We can relate FIðX,DÞ with the fragility indexes of the whole system and of each block. More precisely,FIðX,DÞ is a convex linear combination of the ratios FIðX,DnÞ=FIðXIj
,Dn
IjÞ, since we can write
FIðX,DÞ ¼Xs
j ¼ 1
9Ij9d
FIðX,DnÞ
FIðXIj,DnÞ
:
Furthermore, it is also a weighted mean of those ratios:
FIðX,DÞ ¼ 1
s
Xs
j ¼ 1
9Ij9sd
FIðX,DnÞ
FIðXIj,DnÞ
:
Proposition 3.2 (inter-blocks dependence). Under the conditions of Proposition 3.1, we have
(i)
1rFIðX,DÞrPs
j ¼ 1 EGIjWs
j ¼ 1 EGIj
:
(ii)
FIðX,DÞ ¼ 1 if and only if XIj, j¼ 1, . . . s are independent random vectors.P W(iii)
FIðX,DÞ ¼ sj ¼ 1 EGIj= s
j ¼ 1 EGIj
if and only if XIj, j¼ 1, . . . s are totally dependent random vectors in the sense
GðxÞ ¼Vs
j ¼ 1 GIjðxIjÞ.
Proof. By the inequalitiesYs
j ¼ 1
GIjðxIjÞrGðxÞr
s
j ¼ 1
GIjðxIjÞ
we have, successively
ðGf1gðxÞÞ
Ps
j ¼ 1EG
Ij rGEG
D
f1gðxÞr ðGf1gðxÞÞWs
j ¼ 1EG
Ij and_s
j ¼ 1
EGIjrEG
DrXs
j ¼ 1
EGIj: &
Proposition 3.3 (intra-blocks dependence). Under the conditions of Proposition 3.1, we have
(i)
sEGD
rFIðX,DÞr d
EGD
:
(ii)
FIðX,DÞ ¼ d=EGD if and only if the sub-vectors XIj, j¼ 1, . . . s, only have independent r.v.’s.
(iii) FIðX,DÞ ¼ s=EGD if and only if the sub-vectors XIj, j¼ 1, . . . s, only have totally dependent random r.v.’s.
Proof. Just observe that 1rEIjr9Ij9, with the lower and upper bounds corresponding to, respectively, complete
dependence and independence of r.v.’s within sub-vectors XIj, j¼ 1, . . . s. &
Next result presents an extremal coefficient for the amount of dependence between YIj, j¼ 1, . . . ,s, Y� G, through the FI of
X� F in the domain of attraction of the MEV G.
H. Ferreira, M. Ferreira / Journal of Statistical Planning and Inference 142 (2012) 1837–1848 1841
Proposition 3.4. Under the conditions of Proposition 3.1, we have
GðxÞ ¼ ðGI1ðxI1Þ, . . . ,GIs
ðxIsÞÞ
1=FIðX,DÞ:
Proof. Observe that
GðxÞ ¼ G1ðxÞEG
D ¼ ðG1ðxÞ
Ps
j ¼ 1EG
Ij Þ1=FIðX,DÞ
¼ ðG1ðxÞEG
I1 � � �G1ðxÞEG
Is Þ1=FIðX,DÞ
¼ ðGI1ðxI1Þ, . . . ,GIs
ðxIsÞÞ
1=FIðX,DÞ: &
If we consider partition Dn in the previous result, we obtain the known relation (Smith, 1990)
GðxÞ ¼ ðG1ðxÞdÞ1=FIðX,DnÞ
¼ ðG1ðxÞdÞEG=d¼ G1ðxÞ
EG
:
Observe also that we can write FIðY,DÞ instead of FIðX,DÞ with X in the domain of attraction of the MEV distribution of Ysince Y belongs to the same domain of attraction. We finish this section with a generalization of Proposition 3.1 to the caseof equivalent margins and two illustrative examples.
Proposition 3.5. If X has d.f. F with equivalent marginal distributions in the sense of (8), and belongs to the domain of
attraction of a MEV G with unit Frechet margins then, for each k 2 f1, . . . ,sg, we have
FIðX,DÞ ¼Ps
j ¼ 1 log CGðe�g11Ij
ð1Þ, . . . ,e
�gd1Ij ðdÞÞ
log CGðe�g1 , . . . ,e�gd Þ:
Example 3.1. We consider a random vector X of Example 3.2 in Falk and Tichy (to appear-a), i.e., having componentsXi ¼
Pmk ¼ 1 likYk, where Y1, . . . ,Ym are independent r.v.’s with Pareto(a) distribution, a40, and lijZ0 such thatPm
k ¼ 1 laik ¼ 1, i¼ 1, . . . ,d. Taking for H any of the distributions of the margins of X, the equivalence condition (8) holds
with gi ¼ 1, i¼ 1, . . . ,d. The distribution of X belongs to the domain of attraction of
GðxÞ ¼ exp �Xm
k ¼ 1
3di ¼ 1
lik
xi
� �a !
, x40
with Frechet margins, GiðxÞ ¼ expð�x�aÞ. Hence, for u¼ ðu1, . . . ,udÞ 2 ð0,1Þ,
CGðuÞ ¼ exp �Xmk ¼ 1
3di ¼ 1l
aikð�log uiÞ
!:
By Proposition 3.5, we have
FIðX,DnÞ ¼
Pdj ¼ 1
Pmk ¼ 1 3
di ¼ 1l
aik1fjgðiÞPm
k ¼ 1 3di ¼ 1l
aik
¼
Pdj ¼ 1
Pmk ¼ 1 l
ajkPm
k ¼ 1 3di ¼ 1l
aik
¼dPm
k ¼ 1 3di ¼ 1l
aik
as obtained in Falk and Tichy (to appear-a). For any partition D, we have
FIðX,DÞ ¼Ps
j ¼ 1
Pmk ¼ 1 3
di ¼ 1l
aik1IjðiÞPm
k ¼ 1 3di ¼ 1l
aik
:
To illustrate, consider d¼ 3¼m, a¼ 1 and weights
l11 ¼ 4=8, l12 ¼ 2=8, l13 ¼ 2=8,
l21 ¼ 1=8, l22 ¼ 1=8, l23 ¼ 6=8,
l31 ¼ 3=8, l32 ¼ 2=8, l33 ¼ 3=8:
We have
FIðX,DnÞ ¼3
4=8þ2=8þ6=8¼
24
12¼ 2
and, for D¼ ff1;2g,f3gg
FIðX,DÞ ¼ ð4=8þ2=8þ6=8Þþð3=8þ2=8þ3=8Þ
4=8þ2=8þ6=8¼
20
12o2:
Example 3.2. If G has copula
CGðu1, . . . ,udÞ ¼ exp �Xd
i ¼ 1
ð�log uiÞ1=a
!a !, 0oar1
H. Ferreira, M. Ferreira / Journal of Statistical Planning and Inference 142 (2012) 1837–18481842
(symmetric logistic model) then, for any partition D, we have
FIðX,DÞ ¼Ps
j ¼ 1 9Ij9a
da¼Xs
j ¼ 1
9Ij9d
� �a
and FIðX,DnÞ ¼ d1�a as already stated in Geluk et al. (2007). In the symmetric model the FI is only a function of the blockssize. If we consider the more general asymmetric logistic model, whose copula is given by
CGðu1, . . . ,udÞ ¼ exp �Xq
k ¼ 1
Xd
i ¼ 1
ð�bki log uiÞ1=ak
!ak( )
,
where bki are non-negative constants such thatPq
k ¼ 1 bki ¼ 1, i¼ 1, . . . ,d, 0oakr1, k¼ 1, . . . ,q, we obtain
FIðX,DÞ ¼Ps
j ¼ 1
Pqk ¼ 1ð
Pi2Ij
b1=ak
ki ÞakPq
k ¼ 1ðPd
i ¼ 1 b1=ak
ki Þak
:
4. Tail dependence for blocks
In the following we always consider that X has continuous marginal d.f.’s. Consider notation MðIjÞ ¼W
i2IjFiðXiÞ, j 2 D.
Definition 4.1. The upper-tail dependence coefficients of X corresponding to partition D of D are defined by, for eachSD! f1, . . . ,sg
tFS ¼ lim
um1P
\j =2 S
MðIjÞ4u9\j2S
MðIjÞ4u
0@ 1A ð11Þ
when the limit exists.
If we consider partition Dn, then SD! f1, . . . ,dg and we find the definition of Li (2009). Further, the case s¼2 lead us todefinition of Ferreira and Ferreira (2011).
Consider
lFS :¼ lim
um1
PT
j2SMðIjÞ4u� �
1�uð12Þ
for each SD! f1, . . . ,sg. Hence, we can write
tFS ¼
lFf1,...,sg
lFS
: ð13Þ
Observe that lFS corresponds to the multivariate upper-tail dependence coefficient LUð1SÞ in Schmidt and Stadtmuller
(2006), where 1S denotes the unit vector with dimension 9S9. In particular, for partition Dn, lFfi,jg ¼LUð1;1Þ corresponds to
the well-known bivariate tail dependence concept (Sibuya, 1960; Joe, 1997).Before we relate the FI with the tail dependence coefficients corresponding to a partition, we present in this section
some extensions of the results in Li (2009).
Proposition 4.1. If X has MEV distribution G with standard Frechet margins and spectral measure� W defined on the
d-dimensional unit sphere Sd then, for each SD! f1, . . . ,sg, we have
tGS ¼
RSd
Vsj ¼ 1
Wi2Ij
wi dWðwÞRSd
Vj2s
Wi2Ij
wi dWðwÞ: ð14Þ
Proof. From the spectral representation of G we obtain, for u sufficiently close to 1
P\s
j ¼ 1
MðIjÞ4u
0@ 1A¼ 1�X
|aS�f1,...,sg
ð�1Þ9S9þ1G �1�1I ðSÞð1Þ
log u, . . . ,�
1�1I ðSÞðdÞ
log u
!
� 1�X
|aS�f1,...,sg
ð�1Þ9S9þ1 1þ log u
ZSd
_i2I ðSÞ
wi dWðwÞ
0@ 1A¼ 1�
X|aS�f1,...,sg
ð�1Þ9S9þ1 1�ð�log uÞ
ZSd
_j2S
_i2Ij
wi dWðwÞ
0@ 1A: ð15Þ
H. Ferreira, M. Ferreira / Journal of Statistical Planning and Inference 142 (2012) 1837–1848 1843
Since X|aS�f1,...,sg
ð�1Þ9S9þ1¼ 1
and X|aS�f1,...,sg
ð�1Þ9S9þ1_j2S
aj ¼^
j2f1,...,sg
aj
expression in (15) becomes
P\s
j ¼ 1
MðIjÞ4u
0@ 1A� ð�log uÞ
ZSd
s
j ¼ 1
_i2Ij
wi dWðwÞ
0@ 1A:Analogously, we obtain
P\j2S
MðIjÞ4u
0@ 1A� ð�log uÞ
ZSd
^j2S
_i2Ij
wi dWðwÞ
0@ 1A: &
For the particular case Dn, the previous result is the one found in Li (2009). Note also that the numerator of (14) can beexpressed through extremal coefficients as follows:X
|aS�f1,...,sg
ð�1Þ9S9þ1Z
Sd
_j2S
_i2Ij
wi dWðwÞ ¼X
|aS�f1,...,sg
ð�1Þ9S9þ1EGIðSÞ ¼ E
GI1þ � � � þEG
Is�ðEG
I1[I2þEG
I1[I3þ � � �Þ� � � � þð�1Þ9S9þ1EG
I1[���[Is
a generalization of result (15) in Ferreira and Ferreira (2011) where s¼2.The next result highlights the connections between tail dependence and extremal coefficients.
Corollary 4.1. Under the conditions of Proposition 4.1, we have
(i)
lGS ¼X
|aT�S
ð�1Þ9T9þ1EGIðTÞ,
(ii)
tGS ¼
P|aT�f1,...,sgð�1Þ9T9þ1EG
I ðTÞP|aT�Sð�1Þ9T9þ1EG
I ðTÞ:
We end this section with a generalization of Theorem 2.6 in Li (2009), by adapting the arguments to subsets of D thatcorrespond to unions of blocks in D.
Proposition 4.2. If F belongs to the domain of attraction of a MEV G with unit Frechet margins then, for any partition D and
|aS � f1, . . . ,sg, the non-null upper-tail dependence coefficients tFS are the same as the corresponding ones of G.
5. Fragility index and tail dependence for blocks
In this section we shall see that the asymptotic d.f.’s of the conditional probability of k exceedances between blocks,I1, . . . ,Is, can be derived through the tail dependence coefficients given in (12). More precisely, if the d.f. F of X has taildependence coefficients lF
S corresponding to partition D, we can obtain limx-1 PðNx ¼ k9Nx40Þ from this latter. Inparticular, for F in the domain of attraction of a MEV G, besides the representations presented in the previous results, wecan write those limiting probabilities through tail dependence coefficients lG
S .
Proposition 5.1. Let X be a random vector with d.f. F with continuous and identically distributed margins. Let D be a partition
of D for which the tail dependence coefficients lFS corresponding to F exist for each S � f1, . . . ,sg. Then
(i)
For each k 2 f1, . . . ,sglimx-1
PðNx ¼ k9Nx40Þ ¼
PS�f1,...,sg,9S9 ¼ k
PT�SC ð�1Þ9T9lF
T[SPsk ¼ 1
PS�f1,...,sg,9S9 ¼ k
PT�SC ð�1Þ9T9lF
T[S
as long as the numerator is non-null.
H. Ferreira, M. Ferreira / Journal of Statistical Planning and Inference 142 (2012) 1837–18481844
(ii)
If F belongs to the domain of attraction of a MEV G with unit Frechet marginals, the limits in (i) exist and coincide for bothdistributions, F and G.
Proof. Just observe that
limx-1
PðNx ¼ k9Nx40Þ ¼ limum1
PS�f1,...,sg,9S9 ¼ k
PT�SC ð�1Þ9T9Pð
TT[SMðIjÞ4uÞPs
k ¼ 1
PS�f1,...,sg,9S9 ¼ k
PT�SC ð�1Þ9T9Pð
TT[SMðIjÞ4uÞ
:
Now divide both terms of the ratio by 1�u. &
Analogously, we obtain the FI through lFS or lG
S .
Proposition 5.2. Under the conditions of Proposition 5.1, we have
(i)
FIðX,DÞ ¼Ps
j ¼ 1 lFfjgPs
k ¼ 1
PS�f1,...,sg,9S9 ¼ k
PT�SC ð�1Þ9T9lF
T[S
as long as the numerator is non-null.
(ii) If F belongs to the domain of attraction of a MEV G with unit Frechet marginals, FIðX,DÞ exists and the the expression in (i)coincides for both distributions, F and G.
The statements in (ii) of the two propositions are consequences of Proposition 4.2.
6. Asymptotic independence
If X has independent margins Xi, i¼ 1, . . . ,d, we have an unit FI. As Geluk et al. (2007) observed, in this case we might havean asymptotic independence characterized by a dependency that vanishes at extreme quantiles. Asymptotic independencemeans that the system is weakly fragile, albeit possibly correlated (e.g., Gaussian vectors). Geluk et al. (2007) have defineda fragility index for asymptotic independence (AIFI) given by
ZD ¼1
dlimx-1
Pdi ¼ 1 log PðXi4xÞ
log PðX14x, . . . ,Xd4xÞ: ð16Þ
In case d¼2 we find the Ledford and Tawn coefficient of asymptotic independence (Ledford and Tawn, 1996, 1997) and, ifd42, a multivariate extension of this latter (Ferreira and Ferreira, 2012).
Here we consider an extension of the AIFI in (16) for blocks, in the same spirit of the FI in Proposition 3.1, i.e., by relatingthe AIFI within blocks with the AIFI of the whole vector.
Let ZA be the AIFI of sub-vector XA of X, with A � D, i.e.
ZA ¼1
9A9limx-1
Pi2Alog PðXi4xÞ
log PðXA4xAÞ: ð17Þ
Definition 6.1. Let X¼ ðX1, . . . ,XdÞ be a random vector with FIðX,DÞ ¼ 1. Then the AIFI of X¼ ðX1, . . . ,XdÞ relative topartition D of D is
1
slimx-1
Psj ¼ 1 log PðXIj
4xIjÞ
log PðX14x, . . . ,Xd4xÞð18Þ
whenever the limit exists and is denoted ZðX,DÞ.
Proposition 6.1. Let X¼ ðX1, . . . ,XdÞ be a random vector with FIðX,DÞ ¼ 1. Assume that (16) holds and that (17) holds for all
Ij 2 D, j¼ 1, . . . ,s, with limit given by ZIj, respectively. If X has identically distributed or equivalent margins in the sense of (8),
then
ZðX,DÞ ¼ ZD
1
s
Xs
j ¼ 1
1
ZIj
: ð19Þ
Proof. Observe that by (16) and (17), then
ZðX,DÞ ¼ limx-1
ZD
1
s
Psj ¼ 1
1
ZIj9Ij9
Pi2Ij
log PðXi4xÞ
1
d
Pdi ¼ 1 log PðXi4xÞ
H. Ferreira, M. Ferreira / Journal of Statistical Planning and Inference 142 (2012) 1837–1848 1845
and by (8), we have
ZðX,DÞ ¼ limx-1
ZD
1
s
Psj ¼ 1
1
ZIj
1
9Ij9
Pi2Ij
log giþ logð1�HðxÞÞ
!1
d
Pdi ¼ 1 log giþ logð1�HðxÞÞ
: &
In particular, we have ZðX,DnÞ ¼ ZD and hence, the AIFI of the system divided into blocks is larger than the AIFI of thesystem itself, i.e., ZðX,DÞZZðX,DnÞ.
Remark 6.1. In the particular case of FIðX,DnÞ41 we have ZðX,DÞ ¼ 1.Observe also that, by Proposition 6.1, we can relate ZðX,DÞ with the fragility indexes of the whole system and of each
block. More precisely, ZðX,DÞ is the arithmetic mean of the ratios ZðX,DnÞ=ZðXIj,DnÞ.
Example 6.1. Consider X¼ ðX1, . . . ,XdÞ a standard d-variate Gaussian random vector with d.f. Fdð�; ðSi,jÞi,j2DÞ having positivedefinite correlation matrix ðSi,jÞi,j2D. We have, for all A � D, Z�1
A ¼ 1AðSi,jÞ�1i,j2A1T
A, the sum of all elements of the sub-matrixðSi,jÞ
�1i,j2A (Geluk et al., 2007; Hua and Joe, 2011). For illustration, consider dimension d¼4, constant correlation r, and take
s¼3 with I1 ¼ f1;2g, I2 ¼ f3g and I3 ¼ f4g. We have ZD ¼ ð3rþ1Þ=4, ZI1¼ ðrþ1Þ=2 and ZI2
¼ ZI3¼ 1. Hence
ZðX,DÞ ¼ 1
3
3rþ1
4
2
rþ1þ1þ1
� �¼ð3rþ1Þðrþ2Þ
6ðrþ1Þ:
In the sequel we consider positive/negative association of a random vector in the sense of Ledford and Tawn (1996,1997).
Proposition 6.2 (inter-blocks asymptotic independence). Under the conditions of Proposition 6.1, we have
(i)
ZðX,DÞr1=s in case of positive association between sub-vectors XIj, j¼ 1, . . . s, and ZðX,DÞZ1=s in case of negativeassociation.
(ii)ZðX,DÞZVs
j ¼ 1 ZIj
s
Xs
j ¼ 1
1
ZIj
:
(iii)
ZðX,DÞ ¼ 1=s if and only if the sub-vectors XIj, j¼ 1, . . . s, are independent.V P(iv)
ZðX,DÞ ¼ ð sj ¼ 1 ZIj=sÞ sj ¼ 1ð1=ZIj
Þ if and only if the sub-vectors XIj, j¼ 1, . . . s, are totally dependent.
Proof. Observe that PðX14x, . . . ,Xd4xÞrVs
j ¼ 1 PðXIj4xIjÞ, with the upper bound corresponding to total dependence
between sub-vectors XIj, j¼ 1, . . . ,s. Under positive association we have PðX14x, . . . ,Xd4xÞZ
Qsj ¼ 1 PðXIj
4xIjÞ and
negative association otherwise, with the bounds corresponding to independence between sub-vectors XIj, j¼ 1, . . . ,s. &
Proposition 6.3 (intra-blocks asymptotic independence). Under the conditions of Proposition 6.1, we have
(i)
ZðX,DÞrZDd=s in case the sub-vectors XIj, j¼ 1, . . . ,s, only have positively associated r.v.’s and ZðX,DÞZZDd=s in case ofnegative association.
(ii) ZðX,DÞZZD.(iii)
ZðX,DÞ ¼ ZDd=s if and only if the sub-vectors XIj, j¼ 1, . . . ,s, only have independent r.v.’s.(iv)
ZðX,DÞ ¼ ZD if and only if the sub-vectors XIj, j¼ 1, . . . ,s, only have totally dependent random r.v.’s.Proof. For each j¼ 1, . . . ,s, under positive association of the r.v.’s in XIjwe have PðXIj
4xIjÞZQ
i2IjPðXi4xÞ, and hence
1=ZIjr9Ij9, with the upper bound corresponding to independence. For negative association we have
PðXIj4xIjÞrQ
i2IjPðXi4xÞ. Observe also that total dependence within each block means PðXIj
4xIjÞ ¼V
i2Ijgið1�HðxÞÞ. &
As already mentioned, the definition of the AIFI in (18) measures the asymptotic independent fragility of a system dividedinto blocks (sub-vectors), by relating the asymptotic independent fragility within the blocks and in the whole system. If in(16) we generalize for blocks the concept of an exceedance of a r.v., Xi4x, through events XIj
� xIj, we obtain another
coefficient for asymptotic tail independence. In this way, we extend the coefficient of asymptotic tail independence that wasconsidered in Ferreira and Ferreira (2011) for the particular case of a partition D¼ fI1,I2g of D¼ f1, . . . ,dg.
H. Ferreira, M. Ferreira / Journal of Statistical Planning and Inference 142 (2012) 1837–18481846
Definition 6.2. Let X¼ ðX1, . . . ,XdÞ be a random vector with FIðX,DÞ ¼ 1. The coefficient of asymptotic independence ofX¼ ðX1, . . . ,XdÞ relative to partition D of D is
limx-1
1
s
Psj ¼ 1 log PðXIj
�xIjÞ
log PðTs
j ¼ 1fXIj�xIjgÞ
ð20Þ
whenever the limit exists, and is denoted ZðI1 ,...,IsÞ.
The following result is therefore an immediate extension of Proposition 2.4 in Ferreira and Ferreira (2011).
Proposition 6.4. Let X¼ ðX1, . . . ,XdÞ be a random vector with FIðX,DÞ ¼ 1. Assume that the limit in (20) exists and, for all
|aKj � Ij, j¼ 1, . . . ,s, Eq. (17) holds for [sj ¼ 1Kj. Then
ZðI1 ,...,IsÞ¼maxfZfi1 ,...,isg
: ij 2 Ij, j¼ 1, . . . ,sg: ð21Þ
Similar to ZðX,DÞ, coefficient ZðI1 ,...,IsÞis also based on the coefficient of Ledford and Tawn or multivariate extensions of
this latter.In the example below, one can see that the asymptotic tail independent coefficients, ZðX,DÞ and ZðI1 ,...,IsÞ
, are different.
Example 6.2. Consider fVigiZ1 an i.i.d. sequence of unit Pareto r.v.’s. Let ðX1,X2,X3Þ be a random vector such that,X1 ¼minðV1,V2Þ, X2 ¼minðV2,V3Þ and X3 ¼minðV3,V4Þ. If D¼ fI1,I2g, with I1 ¼ f1;2g and I2 ¼ f3g, we have PðXi4xÞ ¼ x�2,i¼ 1;2,3, PðX14x,X24xÞ ¼ PðX24x,X34xÞ ¼ x�3 and PðX14x,X34xÞ ¼ PðX14x,X24x,X34xÞ ¼ x�4. Therefore
FIðX,DÞ ¼ limx-1¼
PðS2
i ¼ 1fXi4xgÞþPðX34xÞ
PðS3
i ¼ 1fXi4xgÞ¼
2x�2�x�3þx�2
3x�2�2x�3�x�4þx�4¼ 1,
ZðX,DÞ ¼ 1
2limx-1
P2j ¼ 1 log PðXIj
4xIjÞ
log PðX14x,X24x,X34xÞ¼
1
2limx-1
log PðX14x,X24xÞþPðX34xÞ
log PðX14x,X24x,X34xÞ¼
1
2
�3�2
�4¼
5
8
and
ZðI1 ,I2Þ¼
1
2limx-1
log PðXI1�xI1Þþ log PðXI2
�xI2Þ
log PðXI1�xI1
,XI2�xI2Þ
¼1
2limx-1
log ðPðX14xÞþPðX24xÞ�PðX14x,X24xÞÞþ log PðX34xÞ
log ðPðX14x,X34xÞþPðX24x,X34xÞ�PðX14x,X24x,X34xÞÞ
¼1
2limx-1
log ð2x�2�x�3Þþ log x�2
log ðx�4þx�3�x�4Þ¼ 2=3:
Observe that the same results are obtained if we apply Propositions 6.1 and 6.4, respectively. More precisely, we have
ZD ¼ Zf1;2,3g ¼1
3limx-1
P3i ¼ 1 log PðXi4xÞ
log PðX14x,X24x,X34xÞ¼
1
3
3 log x�2
log x�4¼
1
2,
Zf1;3g ¼1
2
2 log x�2
log x�4¼
1
2and Zf2;3g ¼
1
2
2 log x�2
log x�3¼
2
3:
Obviously ZI2¼ 1 and
ZI1¼
1
2limx-1
P2i ¼ 1 log PðXi4xÞ
log PðX14x,X24xÞ¼
1
2
2 log x�2
log x�3¼
2
3:
Hence, by Propositions 6.1 and 6.4, respectively
ZðX,DÞ ¼ 12 ð
12 ð3=2þ1ÞÞ ¼ 5
8
and
ZðI1 ,I2Þ¼maxfZf1;3g,Zf2;3gg ¼maxf1=2;2=3g ¼ 2=3:
7. Estimation of the fragility index for blocks
In the previous sections we have related the FI for blocks with well-known tail dependence measures. This will allow to obtainimmediate estimators for our index through the estimators of those coefficients that are already studied in the literature.
Proposition 3.1 presents an estimation procedure for the FI based on the extremal coefficients of Tiago de Oliveira(1962) and Smith (1990) given in (2). Observe that they can be expressed through the stable tail dependence function, lG,defined in (3). There are several references in the literature on the estimation of the stable tail dependence function. In aparametric framework, a model for lG must be imposed. Non-parametric estimators are usually based on an arbitrarilychosen parameter, corresponding to the number of top order statistics to use in order to provide the best trade-off
H. Ferreira, M. Ferreira / Journal of Statistical Planning and Inference 142 (2012) 1837–1848 1847
between bias and variance, which is not an easy task. For a survey, we refer Krajina (2010) or Beirlant et al. (2004). Themore recent work in Ferreira and Ferreira (2011) presents a simpler non-parametric estimator based on sample means,which we shall adopt here.
For A � D, denote MðAÞ ¼W
i2AFiðXiÞ. Consider
bEGD ¼
MðDÞ
1�MðDÞand bEG
Ij¼
MðIjÞ
1�MðIjÞ, ð22Þ
where MðAÞ is the sample mean
MðAÞ ¼1
n
Xn
i ¼ 1
_j2A
bF jðXðiÞj Þ ð23Þ
and bF j, j 2 A, is the (modified) empirical d.f. of Fj
bF jðuÞ ¼1
nþ1
Xn
k ¼ 1
1fXðkÞ
jrug
:
The denominator nþ1 instead of n in the ordinary empirical d.f. concerns estimation accuracy and other modifications canbe used (see, for instance, Beirlant et al., 2004). Based on Proposition 3.1, and for a partition D of D, we consider estimator
bFIðX,DÞ ¼Ps
j ¼ 1bEG
IjbEGD
, ð24Þ
which is consistent given the consistency of estimators bEGD and bEG
Ijalready stated in Ferreira and Ferreira (2011).
By Proposition 5.2, we can also estimate FIðX,DÞ based on the tail dependence coefficients in (12). As alreadymentioned, these correspond to multivariate upper-tail dependence coefficients considered in Schmidt and Stadtmuller(2006), for which non-parametric estimators have been studied. We remark that these estimators are also based on asimilar procedure as described above for the stable tail dependence function, i.e., it comprises the choice of a number oftop order statistics to be used, not too large neither too small to avoid, respectively, large bias and large variance.
Remark 7.1. In case of asymptotic independence considered in Section 6, immediate estimators for the asymptoticindependent coefficients, ZðX,DÞ and ZðI1 ,...,IsÞ
, can be derived from the found relations with the Ledford and Tawncoefficient (or multivariate extensions), whose estimation has already been studied in the literature (see, for instance,Draisma et al., 2004 or, for a survey Beirlant et al., 2004). In particular, the Ledford and Tawn coefficient can be estimatedas the extreme value index of r.v. minð1=ð1�F1ðX1ÞÞ,1=ð1�F2ðX2ÞÞÞ and, similarly, its multivariate versions in (17) can beestimated as the extreme value index of r.v. mini2Að1=ð1�FiðXiÞÞÞ.
7.1. An application to financial data
We now illustrate the estimation of the FI for blocks through an application to data analyzed by Ferreira and Ferreira(2011). The data are the series of negative log-returns of the closing values of the stock market indexes, CAC 40 (France),FTSE100 (UK), SMI (Swiss), XDAX (German), Dow Jones (USA), Nasdaq (USA), SP500 (USA), HSI (China), Nikkei (Japan). Theperiod covered is from January 1993 to March 2004. Since we do not have a sample of maximum values, we consider themonthly maximums in each market. We group the indexes in Europe (CAC 40, FTSE100, SMI, XDAX), USA (Dow Jones,Nasdaq) and Far East (HSI, Nikkei). The presence of dependence within these groups was already evidenced in Ferreira andFerreira (2011). We are interested in assessing the fragility within the system of the financial stock markets whenever
grouped in the three big world markets referred: Europe, USA and Far East. To this end, we use estimator bFIðX,DÞ in (24). In
Table 1 are the obtained estimates, as well as, the estimates of the extremal coefficient (bEG) within each group and in the
whole system (we denote the whole system, i.e., the vector of all observations as ‘‘Global’’). The estimates of the FI withineach financial market group and in the whole system are also presented. One can see that the USA is the most fragile
Table 1
Estimates of the extremal coefficient (bEG) within each block (Europe, USA, Far East) and in the whole system (Global)
comprising the three blocks, as well as, estimates of the FI (bFI) within each block, in the whole system and within the
system divided into blocks (bFIðX,DÞ).
Block bEG bFI bFIðX,DÞ
Europe 2.243980009 1.78254707 1.370940479
USA 1.590711176 1.885948905
Far East 1.673156121 1.195345715
Global 4.017568517 2.240160924
H. Ferreira, M. Ferreira / Journal of Statistical Planning and Inference 142 (2012) 1837–18481848
financial system with bFI ¼ 1:885948905. Observe also that the FI of the whole system is almost twice the FI of the systemdivided into blocks.
Acknowledgement
Helena Ferreira was partially supported by the research unit ‘‘Centro de Matematica’’ of the University of Beira Interiorand the research project PTDC/MAT/108575/2008 through the Foundation for Science and Technology (FCT) co-financed byFEDER/COMPETE.
Marta Ferreira was partially supported by FEDER funds through ‘‘Programa Operacional Factores de Competitividade -COMPETE’’, by Portuguese funds through FCT - ‘‘Fundac- ~ao para a Ciencia e a Tecnologia’’, within the Project Est-C/MAT/UI0013/2011 and by PTDC/FEDER grants from Portugal.
References
Ane, T., Kharoubi, C., 2003. Dependence structure and risk measure. Journal of Business and Economic Statistics 76, 411–438.Aulbach, S., Bayer, V., Falk, M. A multivariate piecing-together approach with an application to operational loss data. Bernoulli, to appear.Beirlant, J., Goegebeur, Y., Segers, J., Teugels, J., 2004. Statistics of Extremes: Theory and Application. John Wiley.de Haan, L., de Ronde, J., 1998. Sea and wind: multivariate extremes at work. Extremes 1, 7–45.Draisma, G., Drees, H., Ferreira, A., de Haan, L., 2004. Bivariate tail estimation: dependence in asymptotic independence. Bernoulli 10, 251–280.Embrechts, P., Puccetti, G., 2006. Bounds for functions of dependent risks. Finance and Stochastics 10 (3), 341–352.Falk, M., Tichy, D. Asymptotic conditional distribution of exceedance counts, Adv. in Appl. Probab. (to appear-a).Falk, M., Tichy, D. Asymptotic conditional distribution of exceedance counts: fragility index with different margins, Ann. Inst. Statist. Math. (to appear-b).Ferreira, H., Ferreira, M., 2011. Extremal Dependence: Some Contributions. /arXiv:1108.1972v1 [math.ST]S.Ferreira, M., Ferreira, H., 2012. Tail dependence between order statistics. Journal of Multivariate Analysis 105 (1), 176–192.Geluk, J.L., De Haan, L., De Vries, C.G., 2007. Weak and Strong Financial Fragility. Tinbergen Institute Discussion Paper, TI 2007-023/2.Hauksson, H., Dacorogna, M., Domenig, T., Mueller, U., Samorodnitsky, G., 2001. Multivariate extremes, aggregation and risk estimation. Quantitative
Finance 1, 79–95.Hua, L., Joe, H., 2011. Tail order and intermediate tail dependence of multivariate copulas. Journal of Multivariate Analysis 102 (10), 1454–1471.Huang, X., 1992. Statistics of Bivariate Extreme Values. Ph.D. Thesis, Tinbergen Institute Research Series 22, Erasmus University Rotterdam.Joe, H., 1997. Multivariate Models and Dependence Concepts. Chapman & Hall, London.Junker, M., May, A., 2005. Measurement of aggregate risk with copulas. Econometrics Journal 8, 428–454.Krajina, A., 2010. An M-Estimator of Multivariate Tail Dependence. Tilburg University Press, Tilburg.Ledford, A., Tawn, J.A., 1996. Statistics for near independence in multivariate extreme values. Biometrika 83, 169–187.Ledford, A., Tawn, J.A., 1997. Modelling dependence within joint tail regions. Journal of the Royal Statistical Society, Series B: Statistical Methodology 59,
475–499.Li, H., 2009. Orthant tail dependence of multivariate extreme value distributions. Journal of Multivariate Analysis 100 (1), 243–256.Schmidt, R., Stadtmuller, U., 2006. Nonparametric estimation of tail dependence. Scandinavian Journal of Statistics 33, 307–335.Sibuya, M., 1960. Bivariate extreme statistics. Annals of the Institute of Statistical Mathematics 11, 195–210.Smith, R.L., 1990. Max-stable Processes and Spatial Extremes, Preprint. Univ. North Carolina, USA.Straetmans, S., Wolff, C., Verschoor, W., 2008. Extreme U.S. stock market fluctuations in the wake of 9/11. Journal of Applied Econometrics 23 (1), 17–42.Tiago de Oliveira, J., 1962/1963. Structure theory of bivariate extremes, extensions. Estudos de Matemdtica, Estatistica e Econometrica 7, 165–195.