Tarification à l'aide des modèles mixtes hiérarchiques

Details

Supervisors

Denuit, Michel

Faculty

Faculté des sciences

Degree label

Master [120] en sciences actuarielles, à finalité spécialisée

Abstract

enInsurance pricing models must accurately capture risk heterogeneity while remaining operationally feasible. Traditional generalized linear models often struggle to account for hierarchical risk structures commonly observed in insurance portfolios. This master thesis examines the application of hierarchical credibility models to workers’ compensation insurance, comparing a benchmark Tweedie GLM with a hierarchical GLM and a Tweedie GLMM. Using a Belgian portfolio organized by industry sectors and nested sub-sectors, we evaluate predictive performance in terms of risk classification and aggregate calibration. The results show that hierarchical models substantially improve risk differentiation while maintaining acceptable portfolio-level calibration. In this application, the hierarchical GLM provides performance comparable to the mixed-effects model while offering greater computational efficiency, making it a practical alternative for actuarial pricing.