Economic Modelling 35 (2013) 199–206

Contents lists available at ScienceDirect

Economic Modelling
journal homepage: www.elsevier.com/locate/ecmod

Analyzing the dependence structure of various sectors in the Brazilian

market: A Pair Copula Construction approach
Marcelo Brutti Righi ⁎, Paulo Sergio Ceretta
Federal University of Santa Maria, Department of Business, Santa Maria, Rio Grande do Sul, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: In this paper we estimate the dependence structure between economic sectors in the Brazilian financial mar-
Accepted 5 June 2013 ket through Pair Copula Construction. We use daily data from indices which represent telecommunications,
energy, industrials, consumer, financial, basic materials and real estate sectors in BM&F/Bovespa. Results in-
JEL classification:
dicate predominance of student's t copula in structure. BB1, BB7, BB8, Frank and Joe copulas also fit into some
C00
relationships. Regarding dependence, tail measures obtain relevant values in most relationships. Lower tail
C5
G2
dependence exceeds absolute, measured by Kendall's Tau, and upper tail in many cases, reflecting the asym-
metry in some relationships. Further, in order to give robustness to these results, we forecast daily Value at
Keywords: Risk, considering distinct significance levels, of a portfolio composed of studied sectors through the estimated
Dependence structure. Results allow one to conclude that VaR predictions are correct. These results permit business indus-
Risk management try participants to construct portfolios with assets of these sectors under a proper diversification structure.
Pair Copula Construction Moreover, from an international point of view, investors who are interested in diversification could perform
Sectors more sophisticated strategies in this country rather than simply trading the index.
Brazilian market
© 2013 Elsevier B.V. All rights reserved.

1. Introduction but recent studies have ascertained the superiority of copulas to

Since the introduction of portfolio selection mathematical theory
and Capital Asset Pricing Model (CAPM), the issue of dependence
has been of fundamental importance to financial economics. In the
context of international diversification, there is a need of minimizing
risk of specific assets through optimal resource allocation. Therefore,
it is necessary to understand multivariate relationships between dif-
ferent markets. Thus, we need a statistical model able to measure
temporal dependence between shocks of different assets.
Corroborating with the increasing importance of financial asset
dependence, emerges the fact that the extent of extreme shocks and
potential damaging consequences continuously attract attention
among economists and policymakers. These extreme dependencies,
which represent probability function tails, beyond any fundamental
link, for a while have been an issue of interest to academics, fund
managers and traders, as it has important implications for portfolio
allocation and asset pricing.
An inappropriate model for dependence can lead to suboptimal
portfolios and inaccurate assessments of risk exposures. Traditionally,
correlation is used to describe dependence between random variables,
⁎ Corresponding author at: Department of Administration, University of Santa Maria,
Santa Maria, Rio Grande do Sul, 97105-900, Brazil. Tel./fax: +55 55 3220 9258.
E-mail addresses: marcelobrutti@hotmail.com (M.B. Righi), ceretta10@gmail.com
(P.S. Ceretta).
model dependence, as they offer much more flexibility than the correla-
tion approach, because a copula function can deal with non-linearity,
asymmetry, serial dependence and also the well-known heavy-tails of
marginal and joint probability distribution of financial assets.
A copula is a function that links univariate marginals to their mul-
tivariate distribution. Since it is always possible to map any vector of
random variables into a vector with uniform margins, we are able to
split margins of that vector and a digest of dependence, which is the
copula. A great part of the research on copulas is still limited to the bi-
variate case. Thus, to construct higher dimensional copulas is the natu-
ral next step, despite its hardship. Apart from the multivariate Gaussian
and Student, selection of higher-dimensional parametric copulas is still
rather limited (Genest et al., 2009).
Developments in this area tend to hierarchical, copula-based
structures. It is very possible that the most promising of these is the
Pair Copula Construction (PCC). Originally proposed by Joe (1996),
it has been further discussed and explored in the literature for ques-
tions of inference and simulation. PCC is based on a decomposition
of a multivariate density into bivariate copula densities, of which
some are dependence structures of unconditional bivariate distribu-
tions, and the rest are dependence structures of conditional bivariate
distributions.
Regarding practical issues, a fundamental concern in financial
markets is diversification of investments in distinct economic sectors.
In that sense, dependence of shocks from one sector to another was

0264-9993/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.econmod.2013.06.012
200 M.B. Righi, P.S. Ceretta / Economic Modelling 35 (2013) 199–206
documented by Ewing (2002), among others. The finding of these
spillovers brings a whole new set of implications. Additionally, since
different financial assets are traded based on these sector indices, it
is important for financial market participants to understand the
transmission across sectors in order to make an optimal portfolio al-
location decision (Hassan and Malik, 2007).
Based on this perspective, this paper aims to estimate the depen-
dence structure between economic sectors in the Brazilian financial
market through PCC. For that, we collected daily data from BM&F/
Bovespa indices which represent the telecommunications, energy, in-
dustrials, consumer, financial, basic materials and real estate sectors.
The structure estimated allows one to calculate the tail dependence
of each bivariate relationship between sectors, which gives investors
relevant complementary information. Further, in order to give robust-
ness to these estimates, we predicted the daily Value at Risk (VaR) of
a portfolio composed of studied sectors.
The main contribution of this paper is to obtain information about
the dependence of the economic sectors in the Brazilian market in
order to aid market participants to construct portfolios with assets
of these sectors under a proper diversification structure. Moreover,
from an international point of view, Brazil has become an important
alternative for diversification, especially since the flight of capital
occasioned by recent financial crises, it is crucial for international in-
vestors to have a mapped scenario of the internal market of this coun-
try. This would allow international investors in Brazil to have more
sophisticated strategies rather than simply negotiating the index.
The sequence of this paper is structured in the following way:
Section 2 briefly exposes the background about copulas and PCC;
Section 3 presents materials and methods, exposing data and proce-
dures to achieve the paper objective; Section 4 presents results and
their discussion; Section 5 concludes the paper.
2. Background
This section is subdivided on: i) Copulas, which briefly explains
about definition and properties of this class of function; ii) Pair Copula
Construction, which succinctly exposes the concepts of this structure.
2.1. Copulas
Dependence between random variables can be modeled by cop-
ulas. A copula returns the joint probability of events as a function of
marginal probabilities. This property makes copulas attractive, as uni-
variate marginal behavior of random variables can be modeled sepa-
rately from their dependence (Kojadinovic and Yan, 2010).
The concept of copula was introduced by Sklar (1959). However,
only recently its applications have become clear. A detailed treatment
of copulas as well as their relationship to concepts of dependence is
given by Joe (1997) and Nelsen (2006). A review of the applications
of copulas in finance can be found in Embrechts et al. (2003) and in
Cherubini et al. (2004).
For ease of notation we restrict our attention to the bivariate case. The
extensions to the n-dimensional case are straightforward. A function
C : [0,1]2 → [0,1] is a copula if, for 0 ≤ x ≤ 1 and x1 ≤ x2, y1 ≤ y2, (x1,
y1), (x2,y2) ∈ [0,1]2, it fulfills the following properties:
C ðx; 1Þ ¼ C ð1; xÞ ¼ x; C ðx; 0Þ ¼ C ð0; xÞ ¼ 0:
C ðx2 ; y2 Þ−C ðx2 ; y1 Þ−C ðx1 ; y2 Þ þ C ðx1 ; y1 Þ≥0:
Property Eq. (1) means uniformity of the margins, while Eq. (2),
the n-increasing property means that P(x1 ≤ X ≤ x2, y1 ≤ Y ≤ y2) ≥ 0
for (X, Y) with distribution function C.
In the seminal paper of Sklar (1959), it was demonstrated that a
Copula is linked with a distribution function and its marginal distribu-
tions. This important theorem states that:
(i) Let C be a copula and F1 and F2 univariate distribution functions.
Then Eq. (3) defines a distribution function F with marginals F1
and F2.
2
F ðx; yÞ ¼ C ðF 1 ðxÞ; F 2 ðyÞÞ; ðx; yÞ∈R : ð3Þ
(ii) For a two-dimensional distribution function F with marginals F1
and F2, there exists a copula C satisfying Eq. (3). This is unique if
F1 and F2 are continuous and then, for every (u,v) ∈ [0,1]2:


−1 −1
C ðu; vÞ ¼ F F 1 ðuÞ; F 2 ðvÞ : ð4Þ
In Eq. (4), F−1
1 and F−1
2 denote the generalized left continuous in-
verses of F1 and F2.
However, as Frees and Valdez (1998) note, it is not always obvious
to identify the copula. Indeed, for many financial applications, the
problem is not to use a given multivariate distribution but consists
in finding a convenient distribution to describe some stylized facts,
for example the relationships between different asset returns.
2.2. Pair Copula Construction
The PCC is a very flexible construction, which allows for free
specification of n(n − 1)/2 bivariate copulas. This construction was
proposed by the seminal paper of Joe (1996), and it has been
discussed in detail, especially, for applications in simulation and in-
ference (Bedford and Cooke, 2001; Bedford and Cooke, 2002;
Kurowicka and Cooke, 2006). The PCC is hierarchical in nature. The
modeling scheme is based on a decomposition of a multivariate density
into n(n − 1)/2 bivariate copula densities, of which the first n − 1 are
dependence structures of unconditional bivariate distributions, and
the rest are dependence structures of conditional bivariate distributions
(Aas and Berg, 2011).
PCC is usually represented in terms of density. The two main types of
PCC that have been proposed in the literature are C (canonical)-vines
and D-vines. In this paper we focus on D-vine estimation, which accord-
ingly to Aas et al. (2009) has the density as in formulation Eq. (5).
8
 9
< F xi jxiþ1 ; ⋯; xiþj−1 ; =
f ðx1 ; ⋯; xn Þ ¼
n n−1 n−j
∏k¼1 f ðxk Þ∏j¼1 ∏i¼i c
 : ð5Þ
: F x jx ; ⋯; x ;
iþj iþ1 iþj−1
In Eq. (5), x1, ⋯, xn are variables; f is the density function; c(⋅,⋅) is a
bivariate copula density and the conditional distribution functions are
computed, according to Joe (1996), by formulation Eq. (6).
n

o
∂C x;vj jv−j F xjv−j ; F vj jv−j
ð1Þ F ðxjvÞ ¼
 : ð6Þ
∂F vj jv−j
ð2Þ
In Eq. (6) C x;vj jv−j is bivariate conditional distribution of x and vj
dependency structure conditioned on v− j, where the vector v− j is
the vector v excluding the component vj. To become possible to use
D-vine construction to represent a dependence structure through
 M.B. Righi, P.S. Ceretta / Economic Modelling 35 (2013) 199–206
Fig. 1. Four-dimensional Pair Copula Construction.
copulas, we must assume that univariate margins are uniform in (0, 1).
As an illustration, we present in Eq. (7) a four-dimensional case, and its
graphical representation in Fig. 1.
C ðu1 ; u2 ; u3 ; u4 Þ ¼ C 12 ðu1 ; u2 Þ⋅C 23 ðu2 ; u3 Þ⋅C 34 ðu3 ; u4 Þ⋅
C 13j2 ðF ðu1 ju2 Þ; F ðu3 ju2 ÞÞ⋅C 24j3 ðF ðu2 ju3 Þ; F ðu4 ju3 ÞÞ⋅
C 14j23 ðF ðu1 ju2 ; u3 Þ; F ðu4 ju2 ; u3 ÞÞ:
Fig. 2. Daily log-returns of the sector indices and Ibovespa from January, 2007 to August, 2011 (in-sample period).
201
Where,
F ðu1 ju2 Þ ¼ ∂C 12 ðu1 ; u2 Þ=∂u2 ;
F ðu3 ju2 Þ ¼ ∂C 23 ðu2 ; u3 Þ=∂u2 ;
F ðu2 ju3 Þ ¼ ∂C 23 ðu2 ; u3 Þ=∂u3 ;
F ð
u4 ju3 Þ ¼∂C 34 ðu3 ; u4 Þ=∂u3 ;
F u1 ju2; u3 ¼ ∂C 13j2 ðF ðu1 ju2 Þ; F ðu3 ju2 ÞÞ=∂F ðu3 ju2 Þ;


F u4 ju2; u3 ¼ ∂C 24j3 ðF ðu4 ju3 Þ; F ðu2 ju3 ÞÞ=∂F ðu2 ju3 Þ:
Thus, conditional distributions involved at one level of construc-
tion are always computed as partial derivatives of bivariate copulas
at previous level (Aas and Berg, 2011). Since only bivariate copulas
are involved, partial derivatives may be obtained relatively easy for
most parametric copula families. It is worth to note that copulas involved
in Eq. (5) do not have to belong to the same family. Hence, we should
choose for each pair of variables, the parametric copula that best fits
the data.
3. Empirical method
We collected daily log-returns of BM&F/Bovespa indices that rep-
resent telecommunications (ITEL), energy (IEE), industrials (INDX),
consumer (ICON), financial (IFNC), basic materials (IMAT) and real
estate (IMOB) sectors from January, 2007 to December, 2011, totalizing
1235 observations. The period was chosen considering the existence of
the most recent index, ICON. These sectors were chosen because they
have representative stocks traded in Brazilian market and do not have
stocks in common, that could cause some bias on the results. The last
100 observations (September to December of 2011) were separated
for posterior validation.
ð7Þ We modeled log-returns marginal through a generalized auto-
regressive conditional heteroscedastic — GARCH (1, 1) model with
202 M.B. Righi, P.S. Ceretta / Economic Modelling 35 (2013) 199–206

Table 1
Descriptive statistics of the daily log-returns of the sector indices and Ibovespa from January, 2007 to August, 2011 (in-sample period).

Statistics ITEL IEE INDX ICON IFNC IMAT IMOB Ibovespa

Minimum −0.1371 −0.0721 −0.1182 −0.0967 −0.1285 −0.1445 −0.1629 −0.1210

Maximum 0.1315 0.1160 0.1183 0.1183 0.1900 0.1339 0.1777 0.1368
Mean 0.0004 0.0007 0.0002 0.0004 0.0004 0.0001 0.0002 0.0002
Median 0.0004 0.0010 0.0014 0.0013 0.0007 0.0011 0.0006 0.0012
St. dev. 0.0187 0.0139 0.0195 0.0169 0.0235 0.0243 0.0272 0.0209
Skewness 0.1028 0.1345 −0.2239 −0.0657 0.4669 −0.1424 −0.1332 −0.0363
Kurtosis 6.5551 8.4378 5.5193 5.6592 6.8265 4.3204 5.2860 5.8004

student innovations, introduced by Bollerslev (1986) and largely (U1j, …,Uij) through ranks as Uij = Rij/(n + 1). We ordered variables
applied in finance along the last decades, in order to consider the by decreasing order of the sum of non-linear dependence, measured
well-known conditional heteroscedastic heavy-tailed behavior of through Kendall's tau, with the other variables. Subsequently, to
the financial assets (Longin and Solnik, 2001). The estimated model choose the copula that best fits each bivariate pair of variables we
is represented in Eqs. (8) to (10). employed the AIC criterion. A more detailed presentation of the cop-
ula families present in this selection is given in the Supplementary
r i;t ¼ εi;t ; ð8Þ data.
To validate the choice of a D-vine PCC, we compared the estimated
ε i;t ¼ hi;t zi;t ; zi;t et v ; ð9Þ model with its counterpart C-vine by a test proposed by Clarke
(2007). For this, let C1 and C2 be two competing vine copulas in
2 2 2
hi;t ¼ ci þ ai εi;t−1 þ bi hi;t−1 : ð10Þ terms of their densities and with estimated parameter sets θ1 and
θ2. The null hypothesis of two model statistical indifference is:
Where ri,t is the log-return of asset i in period t; h2i,t is the condi-
 
tional variance of asset i in period t; ai, bi and ci are parameters; εi,t C 1 ðui jθ1 Þ
H 0 : P ðmi N0Þ ¼ 0:5; mi ¼ log ; ∀i ¼ 1; ⋯; n: ð12Þ
is the innovation in conditional mean of asset i in period t; zi,t is a v C 2 ðui jθ2 Þ
degrees of freedom student distributed white noise. The models
were validated through verification of serial correlation in linear
We used fitted PCC in order to calculate dependence measures.
and squared standardized residuals through Ljung and Box (1978),
Given estimated bivariate copula C, lower and upper tail dependence
Q statistic, represented for Eq. (11).
is represented by Eqs. (13) and (14), respectively. The absolute de-
^ 2k pendence calculated with Kendall's Tau through conversion of bivar-
h ρ
Q ¼ nðn þ 2Þ∑k¼1 : ð11Þ iate copulas is exposed in Eq. (15).
n−k

^ 2k is the autocorrelation of
In Eq. (11), n is the size of sample; ρ C ðu; uÞ
λL ¼ limu→0þ : ð13Þ
sample in lag k; h is the number of lags being tested. The Q statistics u
follows a chi-squared (χ2) distribution with h degrees of freedom.
After, we estimated the PCC composed of sector indices. To that, 1−2u þ C ðu; uÞ
λU ¼ limu→1− : ð14Þ
we standardized GARCH residuals into pseudo-observations Uj = 1−u

Table 2
Estimated parameters⁎ and diagnostics⁎⁎ of the linear and squared residuals of the estimated GARCH models.

ITEL IEE INDX ICON IFNC IMAT IMOB Ibovespa

Parameters
c 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0159 0.3425 0.0795 0.1008 0.1256 0.0906 0.0041 0.0979
a 0.1093 0.1138 0.1250 0.1231 0.0910 0.1051 0.1283 0.0971
0.0000 0.0004 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
b 0.8629 0.8754 0.8607 0.8595 0.8993 0.8833 0.8421 0.8888
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
Shape 8.3026 8.4474 9.1716 8.0152 9.7775 9.0099 10.0758 7.8045
0.0000 0.0000 0.0001 0.0000 0.0002 0.0001 0.0001 0.0000

Residuals
Q(10) 10.43 10.72 5.78 6.43 10.04 5.38 16.06 6.56
0.40 0.38 0.83 0.78 0.44 0.86 0.10 0.77
Q(15) 18.32 11.90 6.85 7.30 11.30 5.82 20.50 7.085
0.25 0.69 0.96 0.95 0.73 0.98 0.15 0.96
Q(20) 23.50 23.28 21.32 24.63 19.32 21.68 24.32 16.32
0.26 0.27 0.38 0.22 0.50 0.36 0.21 0.70

Sqn residuals
Q(10) 10.53 12.34 11.07 6.26 5.87 10.91 6.69 5.58
0.39 0.26 0.35 0.79 0.83 0.36 0.75 0.85
Q(15) 13.40 16.24 13.81 9.80 8.12 14.55 10.33 9.54
0.57 0.37 0.5403 0.83 0.92 0.48 0.80 0.85
Q(20) 15.34 19.48 15.16 13.54 14.43 15.73 20.73 10.73
0.76 0.49 0.7673 0.85 0.81 0.73 0.41 0.95

*a, b, and c are defined in (10), Shape is the number of degrees of freedom of the student's t conditional distribution; **Q(k) is the statistic of Ljung–Box for k lags; Bold values are
significant at the 5% level.
 M.B. Righi, P.S. Ceretta / Economic Modelling 35 (2013) 199–206 203

Fig. 3. Estimated and predicted daily conditional volatilities for the indices ITEL, IEE, INDX and ICON. The vertical line represents the out-sample period.

1 1
τðx; yÞ ¼ 4∫0 ∫0 C ðu; vÞdC ðu; vÞ−1:
Where notation follows that presented in Section 2. Further, to
validate the estimated PCC, we constructed a portfolio, attributing
equal weights to sector indices, and predicted its one-day VaR at dif-
ferent significance levels for the last 100 observations of 2011. The
ð15Þ procedure, adapted from Aas and Berg (2011), is descripted below.
For each day t in prediction period:
We computed one-step ahead forecast of conditional standard de-
viation σj,t of each asset through estimated GARCH models; we simu-
lated 10,000 samples u1, u2, u3, u4 through the estimated PCC; each set
of simulations was converted to z1, z2, z3, z4 samples through the in-
version of their density probability (student t). For each asset j, we
determine 10,000 simulations of daily log returns by rj,t = σj,tzj. We

Fig. 4. Estimated and predicted daily conditional volatilities for the indices IFNC, IMAT, IMOB and Ibovespa. The vertical line represents the out-sample period.
204 M.B. Righi, P.S. Ceretta / Economic Modelling 35 (2013) 199–206

Table 3
Kendall's Tau⁎ dependence matrix of the daily log-returns of the sector indices from
January, 2007 to August, 2011 (in-sample period).

ITEL IEE INDX ICON IFNC IMAT IMOB

ITEL 1.0000 0.4955 0.5063 0.5096 0.4497 0.4608 0.4135

IEE 0.4955 1.0000 0.5406 0.5462 0.5007 0.4876 0.4448
INDX 0.5063 0.5406 1.0000 0.6783 0.6158 0.8046 0.6063
ICON 0.5096 0.5462 0.6783 1.0000 0.5707 0.5436 0.5292
IFNC 0.4497 0.5007 0.6158 0.5707 1.0000 0.5704 0.5270
IMAT 0.4608 0.4876 0.8046 0.5436 0.5704 1.0000 0.5518
IMOB 0.4135 0.4448 0.6063 0.5292 0.5270 0.5518 1.0000
Sum 3.8355 4.0154 4.7518 4.3776 4.2343 4.4188 4.0726
⁎ The Kendall's Tau measure was chosen because it can identify non-linear dependence,
unlike the traditional linear correlation.

computed the return of the portfolio as log-returns mean of each

asset, being 10,000 simulations in each period. For each significance
level q = {0.05, 0.01, 0.005} we computed one-day VaRqt as the qth
Fig. 5. Observed log-returns of the portfolio composed of the sector indices and the predicted
quantile of the portfolio returns distribution.
one-day VaR for the 5%, 1% and 0.5% significance levels, through the GARCH-PCC construction
If the observed log-return was below the predicted VaR, then a vio- from September, 2011 to November, 2011.
lation occurred. To test the significance of difference between realized
and expected number of violations, we used the likelihood ratio statistic
proposed by Kupiec (1995), represented in Eq. (16). Complementing this initial descriptive analysis, some statistics are
exhibited in Table 1.
h
x x
x n−x i  x n−x  Plots in Fig. 2 indicate that there is a similar behavior in all sectors.
2 ln 1− −2 ln α ð1−α Þ : ð16Þ
n n This pattern follows Brazilian market evolution as a whole, represent-
ed by Ibovespa log-returns plot. There was a huge volatility cluster
In Eq. (16), the null hypothesis is that the expected proportion of around observations 400–600, representing 2007/2008 sub-prime
violations is equal to α; x is the number of occurred violations; n is crisis. The end of this in-sample period shows a considerable fall in
the length of the sample. This statistic is asymptotically distributed all indices. This is a possible vestige of 2011 Euro zone crisis. Further,
as χ2(1). We computed p-values of the null hypothesis for each sectors represented in Fig. 1 first column appear to be less volatile
quantile q. than those represented in the second column.
Results in Table 1 indicate that sectors had great range in their
4. Results and discussion log-returns during in-sample period, as evidenced by minimum and
maximum values. This result, associated with the calculated standard
Firstly, we present in Fig. 2 plots of studied sector indices of daily deviations corroborates with plots in Fig. 1 in the sense that the first
log-returns during in-sample period. The main index in Brazilian mar- four sectors are less volatile than the others. This can be explained
ket temporal evolution (Ibovespa) was added to serve as benchmark. by an intense relation of these turbulent sectors with the financial
market, especially because in-sample period considers sub-prime cri-
Table 4
sis. Regarding skewness, INDX, ICON, IMAT and IMOB presented neg-
Pair Copula Constructions⁎ of the daily log-returns of the sector indices from January, ative asymmetry, just as the Ibovespa, while ITEL, IEE and IFNC
2007 to August, 2011 (in-sample period). exhibited a positive one. Further, all assets are leptokurtic, as pointed
by kurtosis.
Relationship Parameters Dependence
After this initial descriptive analysis, the first step in order to esti-
Copula Family First Second Lower Upper Tau
mate the dependence structure of Brazilian economic sectors was to
C12 BB1 1.0463 3.4136 0.8236 0.7749 0.8083 model indices as marginal. This step was performed through GARCH
C23 BB7 1.8959 1.5267 0.6351 0.5586 0.5046 models. The results of these estimations as well as their diagnostic
C34 Student 0.7822 3.3261 0.5046 0.5046 0.5640
C45 BB7 1.8736 1.3954 0.6085 0.5523 0.5258
tests, are presented in Table 2. Complementing, Figs. 3 and 4 exhibit
C56 BB1 0.6498 1.3707 0.4593 0.3419 0.4509 estimated and predicted conditional volatilities.
C67 BB1 0.7176 1.4687 0.5181 0.3969 0.5008 Results in Table 2 indicate that there are similarities between sector
C13|2 Student 0.7773 18.1023 0.1383 0.1383 0.5667 log-returns. Estimated conditional variance of all markets is impacted
C24|3 BB7(180°) 1.3757 0.4415 0.3449 0.2081 0.3065
by past squared shocks (around 0.09–0.12) and lagged conditional
C35|4 Student 0.3804 5.4159 0.1375 0.1375 0.2498
C46|5 Student 0.4554 6.0002 0.1496 0.1496 0.3021 variance (around 0.84–0.89). The estimated models were validated
C57|6 Student 0.3047 7.6591 0.0614 0.0614 0.1967 through diagnostics for serial correlation in linear and quadratic stan-
C14|23 Student 0.0706 9.5447 0.0121 0.0121 0.0530 dardized residuals. Q statistics for 10, 15 and 20 lags did not reject the
C25|34 Student 0.3363 4.0299 0.1745 0.1745 0.2259
C36|45 Student 0.4015 5.7376 0.1353 0.1353 0.2685
C47|56 Frank 1.1228 – 0.0000 0.0000 0.1341
C15|234 Frank 1.8013 – 0.0000 0.0000 0.1968 Table 5
C26|345 Student 0.1292 21.3243 0.0004 0.0004 0.0929 Back test of the predicted one-day VaR for the log-returns of the portfolio constructed
C37|456 BB8(180°) 2.2699 0.6869 0.0000 0.0000 0.1786 through the PCC of the sector index for the out-sample period from September, 2011 to
C16|2345 Student −0.0147 11.7081 0.0032 0.0032 −0.0032 December, 2011.
C27|3456 Student 0.1283 11.2915 0.0093 0.0093 0.0841
Significance Violated Expected Test p-value
C17|23456 Joe(180°) 1.0359 – 0.0474 0.0000 0.0209
AIC −10273.3 5% 4 5 0.2253 0.6350
Clark test 518.0 p-value 0.0036 1% 2 1 0.7827 0.3763
0.5% 2 0.5 2.5679 0.1091
⁎ Selected families and their estimated parameters. These parameters were converted
in the lower tail, upper tail and Kendall's Tau dependence measures. * None of the test values were significant at the 5% level.
 M.B. Righi, P.S. Ceretta / Economic Modelling 35 (2013) 199–206
null hypothesis, indicating absence of dependence with past informa-
tion. Thus, the selected models properly fitted log-returns marginal.
Besides this parametric analysis, plots in Figs. 3 and 4 reinforce
some previous results. All indices exhibited very similar temporal
evolution for conditional volatility. A peak of turbulence occurred in
the sub-prime crisis period. Again, ITEL, IEE and ICON appear to be
less volatile than the others. An economic reasoning for this result is
that these sectors are more controlled by actions of the Federal gov-
ernment such as monetary policies and fiscal incentives. Moreover,
these sectors represent basic services which can hardly suffer de-
mand change. IMOB obtained the maximum value for daily volatility,
almost 0.10, because it is the sector intrinsically connected with
2007/2008 crisis. Regarding out-sample period, it is perceptible that
models properly predicted turbulence occasioned by Euro zone crisis
in the end of 2011. These correct forecasts are fundamental to VaR
posterior prediction.
Subsequent to this marginal specification, we calculated the ma-
trix of dependence of sector indices through Kendall's Tau, aiming
to select their order in PCC estimation. The criterion adopted was
the absolute sum of dependence between each index with all others.
Results are presented in Table 3. There is considerable dependence
between the sectors analyzed (54% in mean). The greater association
is between INDX and IMAT (around 80%). This result can be explained
by the economic association between the sector of materials and the
industry per se, once the first is generally one level before the second
in the supply chain, i.e., the sector of material gives inputs for the in-
dustrial sector. The established order for estimation of the PCC was:
INDX, IMAT, ICON, IFNC, IMOB, IEE and ITEL. Sectors which are more
sensible to country economic situation were precisely those with
more absolute dependence in the structure.
With variables' order established, we estimated that the PCC with
GARCH models standardized residuals. The results of this estimation,
containing selected copula families and their estimated parameters,
which were converted in lower tail, upper tail and Kendall's Tau de-
pendence measures, are presented in Table 4.
Results in Table 4 initially indicate that there is a predominance of
student's t copula in bivariate relationships. This result corroborates
with previous research, such as that performed by Marshal and
Zeevi (2003) among others, that have shown that this copula family
is generally superior to other copulas for financial data fitting. BB1
and BB7 copulas also obtained best fit to some (three) bivariate rela-
tionships, especially in vine first level. Frank, BB8 and Joe copulas
appeared in vine posterior levels. These copulas assign, in some cer-
tain degree, importance to joint probability distribution tails. This
fact clarifies that there is more dependence in extreme events than
normally expected. This corroborates with studies that appoint to
an increase of the dependence between financial assets during pe-
riods of great shocks (we can cite here, among others: Hon et al.,
2007; Kenourgios et al., 2011).
Regarding dependence, all measures exhibited decreasing behav-
ior in direction of initial levels to final ones, which was expected
once that this is the nature of the hierarchical construction. Further,
tail measures obtained relevant values in most relationships, except
those in vine last levels where even absolute dependence was very
low. Lower tail dependence exceeded absolute one in many cases.
All cases were in vine first levels. Dependence in lower tails was
equal or greater than upper tail, reflecting the presence of asymmetry
in some bivariate relationships. Further, the estimated PCC rejected
Clark test null hypothesis, which states that there is significant dis-
tinction in fit of utilized D-vine approach and C-vine construction,
emphasizing advantage in D-vine construction choice.
These results highlight the importance of risk management in
terms of sector diversification. This is because such concentration of
joint probability in tails, in particular for lower values, indicates em-
pirically that shocks of these economic sectors have dependence
above normally expected in extreme values. This increased joint
205
probability in tails indicates that it can be difficult to minimize portfo-
lio risk based on asset allocation in these sectors, especially in times of
negative innovations, such as a crisis, where it is exactly when active
managers most need to protect their investments.
Moreover, the results in Table 4 fundamentally emphasize the
need for a proper dependence structure estimation. This procedure
allied with precise marginal estimation, should lead to a trustable
portfolio dynamic risk prediction. In this sense, to give robustness
for the estimated PCC, we exhibit in Fig. 5 observed log-returns and
predicted one-day VaRs, as explained in Section 3. This portfolio is com-
posed of studied economic sectors, considering out-sample period from
September, 2011 to December, 2011, totalizing 100 observations.
As can be visually perceptible by plots in Fig. 4, there was a vola-
tility cluster in portfolio observed log-returns. This turbulence repre-
sents some Eurozone crisis vestiges. The estimated PCC precision in
modeling dependence structure is notable. Just a few days the ob-
served portfolio log-return was below the predicted VaR. To statically
test robustness in portfolio daily risk prediction, we applied the
backtest presented in Section 3. Results of this test are presented in
Table 5.
Results in Table 5 were explicit that the null hypothesis of violations'
expected proportion equal to VaRs significance level was not rejected
for none quantile. This result confirms the robustness of the estimated
PCC. This fact frizzes the relevance of a precise and real dependence
specification in portfolios' risk management.
5. Conclusions
Results indicated the predominance of student's t copula in sector in-
dices relationships. BB1, BB7, BB8, Frank and Joe copulas also obtained
the best fit to some relationships. These copulas assign relevant impor-
tance to joint probability distribution tails. Tail measures obtained rele-
vant values in most relationships, except those in the last levels of the
vine where even the absolute dependence was very low. The lower tail
dependence exceeded the absolute and upper tail ones in many cases,
reflecting the asymmetry in some relationships. With PCC we predicted
portfolio daily risk. Out-sample period exhibited some Eurozone crisis
vestiges, but even in this turbulent period the backtest null hypothesis
was not rejected for none adopted quantile.
These results stress the importance of efficient risk management
in terms of sectorial diversification. Such concentration of joint prob-
ability in the tails, in particular for lower values, difficult portfolio risk
minimization based on asset allocation in these sectors, especially in
times of negative innovations, such as a crisis, is exactly when active
managers need to protect their investments. Thus, the relevance of
a precise and real specification of the dependence between financial
assets, beyond conditional variance dynamic behavior portfolio risk
management is reinforced.
Appendix A. Supplementary data
Supplementary data to this article can be found online at http://
dx.doi.org/10.1016/j.econmod.2013.06.012.
References
Aas, K., Berg, D., 2011. Modeling Dependence Between Financial Returns Using PCC. In:
Kurowicka, D., Joe, H. (Eds.), (Org.), Dependence Modeling: Vine Copula Handbook.
World Scientific, pp. 305–328.
Aas, K., Czado, C., Frigessi, A., Bakken, H., 2009. Pair-copula constructions of multiple
dependence. Insurance: Mathematics and Economics 44, 182–198.
Bedford, T., Cooke, R., 2001. Probability density decomposition for conditionally dependent
random variables modeled by vines. Annals of Mathematics and Artificial Intelligence
32, 245–268.
Bedford, T., Cooke, R., 2002. Vines: a new graphical model for dependent random variables.
The Annals of Statistics 30, 1031–1068.
Bollerslev, T., 1986. Generalized autoregressive conditional heteroscedasticity. Journal
of Econometrics 31, 307–327.
206 M.B. Righi, P.S. Ceretta / Economic Modelling 35 (2013) 199–206
Cherubini, U., Luciano, E., Vecchiato, W., 2004. Copula Methods in Finance. John Wiley
Finances Series, Chichester, U.K.
Clarke, K., 2007. A simple distribution — free test for nonnested model selection. Political
Analysis 15, 347–363.
Embrechts, P., Lindskog, F., McNeil, A., 2003. Modelling dependence with copulas and
applications to risk management. In: Rachev, S. (Ed.), Handbook of Heavy Tailed
Distributions in Finance. Elsevier, New York, pp. 329–384.
Ewing, B., 2002. The transmission shocks among S&P500 indexes. Applied Financial
Economics 12, 285–290.
Frees, E., Valdez, E., 1998. Understanding relationships using copulas. North American
Actuarial Journal 2, 1–25.
Genest, C., R'emillard, B., Beaudoin, D., 2009. Omnibus goodness-of-fit tests for
copulas: a review and a power study. Insurance: Mathematics and Economics
44, 199–213.
Hassan, S., Malik, F., 2007. Multivariate GARCH modeling of sector volatility transmis-
sion. The Quarterly Review of Economics and Finance 47, 470–480.
Hon, M., Strauss, J., Yong, S., 2007. Deconstructing the Nasdaq bubble: a look at conta-
gion across international stock markets. Journal of International Financial Markets,
Institutions and Money 17, 213–230.
Joe, H., 1996. Families of m-variate distributions with given margins and m(m-1)/2
bivariate dependence parameters. In: Rueschendorf, L., Schweizer, B., Taylor, M.D.
(Eds.), Distributions with Fixed Marginals and Related Topics. IMS Lecture Notes —
Monograph Series. Institute of Mathematical Statistics, Hayward, CA, pp. 120–141.
Joe, H., 1997. Multivariate Models and Dependence Concepts. Chapman Hall, London
412.
Kenourgios, D., Samitas, A., Paltalidis, N., 2011. Financial crises and stock market contagion in
a multivariate time-varying asymmetric framework. Journal of International Financial
Markets, Institutions and Money 21, 92–106.
Kojadinovic, I., Yan, J., 2010. Modeling multivariate distributions with continuous
margins using the copula R package. Journal of Statistical Software 34, 1–20.
Kupiec, P., 1995. Techniques for verifying the accuracy of risk measurement models.
Journal of Derivatives 2, 73–84.
Kurowicka, D., Cooke, R.M., 2006. Uncertainty Analysis with High Dimensional Dependence
Modelling. Wiley, New York 284.
Ljung, G., Box, G., 1978. On a measure of a lack of fit in time series models. Biometrika
65, 297–303.
Longin, F., Solnik, B., 2001. Extreme correlation of international equity markets. Journal
of Finance 56, 649–676.
Marshal, R., Zeevi, A., 2003. Beyond correlation: extreme comovements between financial
assets. Working Paper. Columbia University.
Nelsen, R., 2006. An Introduction to Copulas, second ed. Springer, New York 269.
Sklar, A., 1959. Fonctions de Repartition á n Dimensions et leurs Marges. Publications
de l'Institut de Statistique de l'Université de Paris 8, 229–231.
Marcelo Brutti Righi graduated in Business Administration at the Federal University of
Santa Maria (2011). Currently he is a doctoral student on Finance at the Federal
University of Santa Maria. He has experience in the area of finance, acting mainly in
quantitative risk management.
Paulo Sergio Ceretta graduated in Business Administration at the Federal University of
Santa Maria (1992), a Master in Production Engineering at the Federal University of
Santa Maria (1995) and a Doctor in Production Engineering at the Federal University
of Santa Catarina (2001). He is currently an Associate Professor at the Federal Univer-
sity of Santa Maria acting on the following topics: finance, stock market, predictability,
time series, financial econometrics, and quantitative and computational methods.