Discrimination-free Insurance Pricing with Privatized Sensitive Attributes

A team of researchers has proposed a statistical framework that allows insurers to calculate discrimination-free premiums without directly accessing sensitive customer attributes such as gender or race, addressing a growing tension between fair pricing and privacy regulations. The study, submitted to arXiv on 16 Apr 2025 and revised through 12 Jun 2026, introduces methods for estimating premiums that remove both direct and indirect effects of protected characteristics while preserving actuarial consistency [1]. The work responds to a regulatory environment in which insurers face increasing pressure to demonstrate fairness in machine-learning-driven pricing models, even as privacy laws restrict the collection and use of sensitive personal data [2]. In the European Union, for instance, the General Data Protection Regulation defines personal data broadly to include any information related to an identifiable natural person, a category that encompasses the very attributes fairness mandates seek to neutralize [7]. The proposed solution operates within a multi-party data setting. Insurers hold non-sensitive attributes and claims outcomes, while a trusted third party retains a privatized or noise-perturbed version of the sensitive attributes generated through a privacy mechanism [5]. The insurer transmits transformed non-sensitive data and the response variable to the third party, which then trains a model and returns the resulting discrimination-free premium [5]. The framework is designed so that the insurer never observes the true sensitive attribute, only its privatized form [9]. The researchers examine two scenarios. In the first, the trusted third party knows the full privacy mechanism, including the noise rate. In the second, the mechanism is known but the noise rate is not [5]. For both cases, the paper establishes theoretical guarantees for the estimators and validates the approach through numerical experiments and empirical applications using synthetic and real-world datasets [5]. The authors note that errors in estimating the noise rate can affect model performance, an insight they explore in their analysis [5]. The concept of a discrimination-free premium has gained traction in actuarial literature as a way to eliminate bias while maintaining the statistical soundness of risk-based pricing [2]. Pricing itself is a fundamental business function, the only revenue-generating element among the four components of the marketing mix, and automated pricing systems require careful setup to avoid errors [6]. The new framework adds a layer of fairness without requiring insurers to breach privacy barriers, a combination that the authors argue is straightforward to implement and provides robust statistical guarantees [5]. The paper was also submitted to the 2025 International Conference on Learning Representations under the primary area of alignment, fairness, safety, privacy, and societal considerations [4]. The authors are Tianhe Zhang, Suhan Liu, and Peng Shi [4].