Actuarial fairness is a core topic in the insurance and reinsurance industries. In 2021, the European insurance regulator EIOPA published a report on AI governance principles that underlined the importance of digital ethics in insurance. In this context, the actuarial community is a key stakeholder to help structure a framework for fair insurance.

To produce a fair pricing of insurance products and a fair risk selection, insurers need to adapt simple or advanced models. If the simple removal of protected attributes prevents any 'direct' discrimination, models are still able to 'indirectly' discriminate between individuals thanks to the latent interactions between variables. In this presentation, SCOR proposes a model-agnostic pre-processing technique based on linear algebra to reduce 'indirect' discrimination and move towards fairness, as measured by a selected statistical definition. This method was applied to a mortality dataset in a risk selection case, with the goal of measuring the impact on both the fairness and the performance of the model.

For more detailed information, Scor will release a research article on Biais and Fairness in 2023.