An Interpretable Deep Learning Model for General Insurance Pricing

Background The most popular statistical model used in modeling general insurance claims is the Generalized Linear Model (GLM), introduced by Nelder and Wedderburn (1972). GLMs allow actuaries to incorporate a wide range of statistical distributions that are commonly adopted in actuarial analytics, and the underlying linearity assumption provides an explainable framework for claims modeling (Wüthrich and Merz, 2023). This model has been shown to work well in practice; however, deep learning--the subset of machine learning focusing on artificial neural network models--has been gaining substantial ground in recent years. Applications of deep learning and other novel machine learning techniques in claim modeling have shown an improvement in prediction accuracy compared to classical methods such as the GLM (Noll et al., 2020; Wüthrich and Buser, 2023). Nevertheless, the integration of such advanced techniques as the primary pricing method among actuaries has been slow since they are often perceived as "black boxes", where the intricacies of the inner workings remain obscured, making it challenging to decipher the rationale behind the models' predictions (Harris et al., 2024). Corresponding author Email address: tupho289@gmail.com