Explainable Machine Learning for Motor Insurance Pricing Using a SHAP Based Framework for Actuarial Model Governance Validation and Fairness
TL;DR
A SHAP-based governance framework for evaluating explainable machine learning in motor insurance pricing, with particular attention to ASOP 41 and ASOP 56, which demonstrates that SHAP can provide useful evidence for model understanding and actuarial communication, but cannot by itself establish compliance with ASOP 41 or ASOP 56.
Abstract
Machine learning models can improve insurance pricing accuracy but create challenges for actuarial transparency and model governance. This paper develops a SHAP-based governance framework for evaluating explainable machine learning in motor insurance pricing, with particular attention to ASOP 41 and ASOP 56. The framework is evaluated using the real freMTPL2freq portfolio of 677,991 French motor third-party-liability policies. A Poisson GLM and gradient-boosted Poisson model are compared using out-of-sample deviance, cross-validation, calibration, and Gini discrimination measures. The gradient-boosted model achieves a 2.2–2.4% reduction in Poisson deviance relative to the GLM and substantially higher risk-ranking discrimination (Gini 0.158 versus 0.089). SHAP analysis identifies bonus-malus level, driver age, and population density as important predictors and indicates an age-by-vehicle-power interaction. A geographic proxy-discrimination audit finds material variation in predicted risk across areas and regions, while residual association with population density is weak after controlling for conventional rating factors. The results demonstrate that SHAP can provide useful evidence for model understanding and actuarial communication, but cannot by itself establish compliance with ASOP 41 or ASOP 56. The proposed framework therefore integrates explainability with validation, calibration, stability, and fairness diagnostics to support broader actuarial model governance.