Discrimination vs. calibration in machine learning: Implications for insurance premium calculation

Details

Supervisors

Denuit, Michel

Faculty

Faculté des sciences

Degree label

Master [120] en sciences actuarielles, à finalité spécialisée

Abstract

enThis thesis addresses the challenge of integrating machine learning into actuarial pricing while preserving fairness, stability, and regulatory compliance. Although machine learning often boosts predictive accuracy, it does not inherently ensure that total premiums align with actual claims, which limits its practical application. To bridge this gap, a post‐processing autocalibration framework is introduced. By adjusting model outputs to achieve both portfolio‐wide (global) and subgroup‐level (local) balance, this approach lets actuaries leverage machine learning’s strengths without sacrificing pricing consistency. The framework is grounded in theory using performance curves and calibration metrics to assess effectiveness and is validated on real insurance data, demonstrating its utility as a practical tool for improving actuarial model reliability.