Phase-type mixture-of-experts regression for loss severities

TL;DR

This paper introduces a flexible phase-type mixture-of-experts regression model for claim severity data, extending traditional models to better handle heterogeneous and catastrophic loss data in insurance.

Contribution

It generalizes mixture-of-experts models to phase-type distributions with Markov jump processes, providing a novel, interpretable, and computationally efficient approach.

Findings

Model demonstrates high flexibility in fitting complex data.

Estimation procedure effectively handles real-world datasets.

Model maintains computational efficiency comparable to simpler methods.

Abstract

The task of modeling claim severities is addressed when data is not consistent with the classical regression assumptions. This framework is common in several lines of business within insurance and reinsurance, where catastrophic losses or heterogeneous sub-populations result in data difficult to model. Their correct analysis is required for pricing insurance products, and some of the most prevalent recent specifications in this direction are mixture-of-experts models. This paper proposes a regression model that generalizes the latter approach to the phase-type distribution setting. More specifically, the concept of mixing is extended to the case where an entire Markov jump process is unobserved and where states can communicate with each other. The covariates then act on the initial probabilities of such underlying chain, which play the role of expert weights. The basic properties of such a model are computed in terms of matrix functionals, and denseness properties are derived, demonstrating their flexibility. An effective estimation procedure is proposed, based on the EM algorithm and multinomial logistic regression, and subsequently illustrated using simulated and real-world datasets. The increased flexibility of the proposed models does not come at a high computational cost, and the motivation and interpretation are equally transparent to simpler MoE models.