A Multivariate Energy-Based Fairness Adjuster for Premiums

Fairness in insurance pricing has received increasing attention, particularly in light of regulations calling for nondiscriminatory premium estimation such as the European Union (EU) Gender Directive (2012). This research focuses on removing structural bias in predicted premiums as a solidarity-driven fairness concept. We propose a gradient-based distributional adjustment designed to mitigate disparities across protected groups. The method relies on the Energy distance, a multivariate metric that enables the joint alignment of predicted premium distributions across multiple sensitive attributes (including nonbinary ones) simultaneously. To maintain overall prediction balance, a post hoc autocalibration step corrects biases in the total predicted number of claims. In addition, the method supports fairness adjustments for new policyholders without requiring retraining of the underlying fairness model. We evaluate the proposed methodology in a car insurance pricing setting, where demographic factors such as age and gender are commonly used for risk assessment and premium determination. Results show that the method effectively reduces group-level disparities while retaining a significant share of predictive accuracy.