Copula approaches gained their importance due to Sklar’s Theorem (Sklar
1959). It states that every multivariate distribution
F with margins
\(F_1, \dots , F_d\) can be written as:
$$\begin{aligned} F(x_1, \dots , x_d)= C(F_1 (x_1), \dots , F_d (x_d)) \end{aligned}$$
(1)
with
C as a
d-dimensional copula, a multivariate distribution on the unit hypercube
\([0,1]^d\) with uniform marginal distributions.
F is continuous with strictly increasing continuous marginals
\(F_1, \dots , F_d\) such that:
$$\begin{aligned} f(x_1,\dots ,x_d)= \left[ \prod _{k=1}^d f_k(x_k)\right] c(F_1(x_1 ),\dots ,F_d(x_d)) \end{aligned}$$
(2)
where
c is the copula density,
f, the multi-dimensional probability density function of
F, and
\(f_k\), the marginal probability density function of
\(F_k\),
\(k=1,\dots ,d\).
d in our example refers to the number of breadbasket states in India. The advantage of this approach is that the copula, and thereby the dependence structure, can be chosen independently from the marginal distributions. There are different copula families capturing different dependence structures (Jaworski et al.
2010) but generally speaking, any multivariate distribution can be used as the basis for a copula (Aas
2007). To structure our multivariate copula model, we used regular vines (RVines) (Aas et al.
2009; Bedford and Cooke
2002; Kurowicka and Cooke
2006; Dißmann et al.
2013), a flexible graphical model that consists of a cascade (or tree structure) of conditional and unconditional bivariate copulas that decompose the multivariate probability density. A
d-dimensional RVine model consists of
\(d-1\) trees. Each tree consists of nodes and edges joining the nodes. The first tree identifies
\(d-1\) pairs of variables with a directly modeled distribution. The second tree identifies
\(d-2\) pairs with a distribution conditional on a single variable, modeled by a pair-copula. Continuing in that way, the last tree consists of a single pair-copula with a distribution conditional on all remaining variables. For a graphical representation of a RVine tree, see (Dißmann et al.
2013) or (Brechmann and Schepsmeier
2013). The RVine tree structure in this study is selected using the maximum spanning tree approach with Kendall’s
\(\tau \) as edge weights (Dißmann et al.
2013):
$$\begin{aligned} \max \sum _{\begin{array}{c} E=\{m,n\} \\ \text { in the} \\ \text {spanning tree} \end{array}} |{\hat{\tau }}_{m,n} | \end{aligned}$$
(3)
with
\({\hat{\tau }}_{m,n}\) as pairwise empirical Kendall’s
\(\tau \) (that is, a rank correlation coefficient, see Embrechts et al.
1997). As will be discussed in the results section, this approach was adopted for India on the state level, which showed large tail dependencies for some regions and therefore is an excellent example for showing the advantages of a copula approach when it comes to determine systemic risks.