Non-life insurance pricing under ethical constraints (interpretability, non-discrimination and fairness)

Details

Supervisors

Pechon, Florian

Faculty

Faculté des sciences

Degree label

Master [120] en sciences actuarielles, à finalité spécialisée

Abstract

When implementing a predictive model for non-life insurance pricing, it is legitimate for this model to respect several ethical constraints. Firstly, having an interpretable model will make it easier for the people working on it to understand its behaviour, and it will also be easier to explain to policyholders why they are paying a certain premium. Moreover, non-discrimination and fairness are also ethical constraints that receive significant attention. An illustration is the European Council directive stating that a woman and a man having the same risk profile must pay the same premium for their insurance product. Regarding these three ethical constraints, this master thesis focuses on two research questions. Does an interpretable model such as EBM (explainable boosting machine) have prediction performances similar to RF (random forests) and GBM (gradient boosting machine) models ? Is it possible to have a model for non-life insurance pricing that satisfies simultaneously the three ethical constraints stated earlier: interpretability, non-discrimination and fairness ? To answer these questions, this master thesis first studies the EBM models. Then, different non-discrimination methods and fairness criteria are implemented. This will allow to see the implications or not between the non-discrimination methods and the fairness criteria.