Abstract
Regulators increasingly require insurers to demonstrate that their rating algorithms do not unfairly discriminate with respect to protected attributes, but they have not specified how such assessments should be conducted. This paper develops an operative criterion and a diagnostic framework for this task. We first discuss why, in insurance ratemaking, unfairness originates primarily in the information the fitted model uses rather than in biased outcome labels or historical human decisions. We then propose a residual-distribution fairness criterion: conditional on the admissible rating variables, the distribution of pricing errors should not depend on protected group membership. Comparing distributions rather than averages matters because insurance transfers the residual risk from the policyholder to the insurer, so that groups with equal average pricing errors can face economically different residual lotteries. We implement the criterion using surrogate residuals, which share a common standard normal reference distribution under correct model specification. The resulting diagnostic distinguishes misfit shared across groups from group-specific disparities and from differences in risk composition, localizes these disparities, and indicates the direction of mispricing. Using automobile insurance data, we demonstrate how the framework can be used to detect and localize disparities and to generate testable hypotheses about their sources.