Speaker
Description
This work focuses on the estimation of multivariate generalized gamma convolutions (MGGC), a class of distributions widely used in risk modeling for which no closed-form density is available. In practice, only their characteristic functions are known, which makes standard estimation methods such as maximum likelihood inapplicable. To overcome this difficulty, we adopt an RKHS-based approach and use the Kernel Stein Discrepancy (KSD) as an estimation criterion. More precisely, we develop a method to identify a Stein operator for multivariate MGGC from their characteristic function only, leading to a tractable expression of the associated Stein kernel and KSD. We also highlight the link between the Stein operator, the underlying subordinator, and the Lévy measure of GGC distributions. Finally, we illustrate the relevance of the proposed approach through several numerical experiments.
| Classification | Mainly methodology |
|---|---|
| Keywords | Generalized Gamma Convulsions -RKHS - Stein Operator |