Welfare Cost of Fair Prediction and Pricing in Insurance Market

Abstract

While the fairness and accountability in machine learning tasks have attracted attention from practitioners, regulators, and academicians for many applications, their consequence in terms of stakeholders' welfare is under-explored, especially via empirical studies and in the context of insurance pricing. General insurance pricing is a complicated process that may involve cost modeling, demand modeling, and price optimization, depending on the line of business and jurisdiction. Fairness and accountability regulatory constraints can be applied at each stage of the insurers’ decision-making. The field so far lacks a framework to empirically evaluate these regulations in a unified way. In this paper, we develop an empirical framework covering the entire pricing process to evaluate the impact of fairness and accountability regulations on both consumer welfare and firm profit, as the link between the predictive accuracy of cost modeling and its welfare consequence is theoretically undetermined for insurance pricing. Applying the empirical framework to a dataset of the French auto insurance market, our main results show that (1) the accountability requirement can incur significant costs for the insurer and consumers; (2) fairness-aware ML algorithms on cost modeling alone cannot achieve fairness in the market price or welfare, while they significantly harm the insurer's profit and consumer welfare, particularly of females; (3) the fairness and accountability constraints considered on the cost modeling or pricing alone cannot satisfy the EU gender-neutral insurance pricing regulation unless we combine the price optimization ban with particular individual fairness notions in the cost prediction.

Keywords: Welfare Cost; General Insurance; Fairness-aware Machine Learning; Accountability in Machine Learning; Price Discrimination

JEL Classification: G22, J16