Multivariate risks and depth-trimmed regions (Q1003339)

The authors present a unified approach to the definition of risk measures considering risk measures as maps that take values in a certain partially ordered cone, which may be, e.g., the real line or the Euclidian space or a family of convex sets in the Euclidean space. They outline the basic properties of the so defined risk measures and then describe the main technical constructions that allow to produce new risk measures from existing ones while respecting their properties, such as homogeneity or coherence. It is demonstrated that this approach naturally includes the classical risk measures, set-valued risk measures and provides a natural definition of vector-valued risk measures. Several main constructions of risk measures are described in the mentioned axiomatic framework. Also it is shown that the multivariate setting for risk measures has a number of common features with the concept of central (or depth-trimmed) regions from multivariate statistics. In particular, the half-space trimming corresponds to value-at-risk, while the zonoid trimming yields the expected shortfall. In abstract framework, it is shown how to establish a two-way correspondence between risk measures and depth-trimmed regions. It is also demonstrated how the lattice structure of the space of risk values influences this relationship.