Eric Wolsztynski, University College Cork

Premium pricing models for non-life insurance can lead to discrimination of policyholders based on any individual, socio-economical or demographic attributes, overcharging or undercharging specific categories of clients. We carried out an in-depth benchmark of leading machine learning (ML) techniques to assess whether (and how) they could address such ethical issues, or on the contrary introduce or enhance bias.

Machine Learning methodologies were compared against the industry reference in terms of model goodness-of-fit, risk prediction performance and bias. The models were also evaluated in terms of the information they use, and in terms of pricing fairness with respect to specific policyholder information deemed sensitive by recent European guidelines. The models were assessed on two large, open-access motor insurance datasets, and pricing performance and fairness were assessed simultaneously (to see whether a reliable estimate was also a fair one).

Our findings show that Machine Learning pricing models can yield more appropriate pricing across various sensitive variables such as gender and age, compared to the reference pricing methodology currently in use in general insurance.