Extensions to dependent compound loss models

Develop extensions of the conditional Wasserstein generative adversarial network framework to compound loss models incorporating copula dependence or additional latent effects.

Background

The paper develops a conditional Wasserstein generative adversarial network for posterior approximation in Poisson frequency and Pareto severity models under conditional independence. In the frequency model, the sum of observed counts is a one-dimensional sufficient statistic, which substantially reduces the conditioning dimension.

The authors explicitly identify copula dependence and additional latent effects as extensions not treated by the current formulation. Such features would generally require richer conditioning summaries and would extend amortized posterior approximation beyond the conditionally independent benchmark models considered in the paper.