Fairness in Non-Life Insurance Pricing: A Counterfactual Perspective on Fair Decisions

Specialeforsvar: Laura Lykke Amstrup, Mathilde Munch Ganderup

Titel:  Fairness in Non-Life Insurance Pricing: A Counterfactual Perspective on Fair Decisions

Abstract: This thesis explores different notions of fairness in decision-making, specifically focusing on fairness in non-life insurance pricing, and it suggests that fairness should be defined in a causal framework in terms of counterfactual fairness and in particular, path-specific counterfactual fairness. The main objective is to provide a theoretical justification and a simulation study-based investigation of the counterfactual fairness metrics. The study finds that counterfactual inference, and thus fair decision-making, can be obtained in the setting of structural causal models (SCMs), given that the assumptions about the causal
structure of the variables in the model are satisfied. In the study, three counterfactual fairness models and one path-specific counterfactual fairness model are implemented and their predictions are investigated in the simulation study. Additionally, two out of four approaches are applied in a real-world analysis of motor insurance pricing. The simulation study illustrates that obtaining counterfactual fairness requires that the model assumptions are well-specified and it investigates the consequences
of violated model assumptions. It concludes that if the model assumptions are violated, there is no guarantee that the obtained predictions are fair. Furthermore, it illustrates that fairness comes with the cost of lower predictive accuracy and that it is not clear how one should evaluate how well a model performs with respect to the counterfactual fairness metric. The four models investigated in the simulation study share a common limitation; they rely on a linear causal relationship between the variables in the SCM. In real-life scenarios, such as insurance pricing, this assumption is often violated, and therefore, a fifth model is
suggested. The model aims at obtaining path-specific counterfactual fairness via a Latent- Inference Projection Method and it can be applied in complex, non-linear scenarios. The model is only described from a theoretical viewpoint and to investigate whether it works well in real-life situations, further research is needed. The findings highlight the importance of ethical awareness, well-specified model assumptions and flexible models with efficient implementations.

Vejleder: Munir Hiabu
Censor:     Martin Møller Svensson