Pair-copula constructions of multiple dependence (Q80563)

The authors use a cascade of pair-copulae to model complex patterns of dependence in the tails of non-normal multivariate families. In a sense every joint distribution function contains both a description of the marginal behaviour of the individual variables and a description of their dependency structure. Copulae provide a way of isolating the description of their dependency structure. Each joint density function can be decomposed into a product of pair copulae and marginal densities. For high-dimensionals distributions, there are a significant number of possible pair-copulae constructions. To organize them, a graphical model denoted as `the regular vine' is introduced. The model construction is hierarchical in nature. Various levels of hierarchy correspond to incorporation of more variables in the conditioning sets using pair-copulae as simple building blocks. Assuming conditional independence may reduce the number of levels and hence simplify the construction. The methodology is applied to a financial data set containing daily data about four Norwegian indices. Algorithms that allow inference on the parameters of the pair-copulae on the various levels of the construction are developed. The approach is the first step towards the development of an unsupervised algorithm that explores the space of possible pair-copula models, that also can be applied to huge data sets automatically.