From Predictive Pricing to Fair Pricing: A Causal Fairness Audit Framework for Insurance Models

Background: The pervasive adoption of Machine Learning in insurance pricing raises fundamental questions: when does a predictive variable become discriminatory? How can we distinguish spurious correlations from legitimate causal relationships? The EU AI Act classifies credit scoring—and implicitly insurance pricing—as a high-risk system, imposing stringent transparency and non-discrimination requirements.Objective: We propose an operational audit framework enabling actuaries to systematically assess predictive model fairness, distinguishing between variables with legitimate causal links to risk and discriminatory proxies masquerading as predictors.Methodology: Using the French Motor Third-Party Liability dataset (freMTPL2freq), we develop a three-phase structured audit:

1. Calibration Audit: Statistical sufficiency tests verifying that, for a given predicted premium, expected risk is uniform across demographic groups (age, gender, geographic area)
2. Proxy Detection via XAI: SHAP analysis to quantify each variable’s contribution and identify potential proxies for protected attributes through residual correlations
3. Counterfactual Fairness Check: Counterfactual scenario simulation to distinguish “option luck” (modifiable behaviors such as vehicle type, coverage choice) from “brute luck” (immutable or semi-immutable characteristics)Expected Results: A replicable audit protocol producing: (i) group-stratified calibration metrics, (ii) stratified SHAP attribution maps, (iii) operational recommendations for mitigating identified biases, with documentation compliant with EU AI Act requirements.Contribution: The framework evolves the actuary’s role from loss ratio optimizer to “causal nexus curator,” providing concrete tools to balance technical sustainability, social equity, and regulatory compliance.Keywords: Algorithmic fairness, Insurance pricing, Machine Learning, Explainable AI, SHAP, Counterfactual fairness, EU AI Act, Actuarial science