A Statistical Perspective on Fairness in Insurance Pricing

Fairness is context dependent and multifaceted. Many jurisdictions prohibit insurers from discriminating on sensitive characteristics, such as race or gender, in insurance pricing. However, some direct or indirect discrimination is inevitable: Allowed rating factors and protected characteristics are associated. As solidarity (risk pooling) and actuarial fairness (premiums proportional to expected losses) have always formed the core tension in insurance pricing, causality has recently emerged as a third essential fairness dimension. We present the dominant unfairness mitigation approaches: autocalibration, which reduces cross-subsidies; discrimination-free insurance pricing and its extensions, which mitigate proxy effects; and optimal transport approaches, which eliminate disparities. We situate the methods within a broader taxonomy of fairness methodologies. We discuss some statistical challenges to implementing antidiscrimination regulation, including the choice of sensitive characteristics, measurement under data privacy, diagnosis of unfairness in finite sample, and generalization beyond an insurer's portfolio.