Olivier Côté: A causal perspective on fairness in insurance pricing

Abstract: In many jurisdictions, insurers may not differentiate prices based on protected characteristics. Yet even when variables such as race or gender are excluded, models trained on rich demographic data can reproduce discriminatory patterns. The lack of clear definitions, especially for indirect discrimination, and the underuse of causal reasoning still hinder rigorous assessment. We review causal inference notions and introduce a causal graph tailored for fairness in insurance. Leveraging this framework, we discuss potential sources of bias, formally define indirect discrimination, and study the theoretical properties of fairness methodologies. A five-point spectrum of premiums (best-estimate, unaware, aware, hyperaware, and corrective) is constructed, each reflecting distinct fairness goals and trade-offs. Acknowledging that fairness metrics often lack actuarial relevance, we define local metrics to quantify the monetary impact of unfairness at the policyholder level. We apply our framework through a large-scale case study assessing the fairness of a pseudoprice relative to credit score, using proprietary data from 768,000 vehicles insured in Québec (Canada) in 2016-17. Zooming out, we study market fairness when each insurer computes "fair" premiums from its own selection-biased piece of the market puzzle.