Credibility theory is the usual framework in actuarial science for reinforcing individual experience by transferring rates estimated from collective information. Based on the paradigm of transfer learning, this article presents the idea that a machine learning (ML) model pretrained using a rich market data portfolio improves rate predictions for individual insurance portfolios. This framework first trains several ML models on a market portfolio of insurance data. The pretrained models provide valuable information on relationships between features and predicted rates. Furthermore, features shared with the company dataset improve rate prediction compared with the same ML models trained on the insurer’s dataset alone. We applied classical ML models using an anonymized dataset that included both market data and data from a European non-life insurance company; this approach is comparable with a hierarchical Bühlmann-Straub credibility model. We observed that the transfer learning strategy of combining company data with external market data significantly improved prediction accuracy compared with an ML model trained on the insurer’s data alone and provided competitive results compared with those of hierarchical credibility models. This work has been sponsored by the Casualty Actuarial Society (CAS) and the Society of Actuaries (SOA) Individual Grants Competition for 2020.
Dutang et al. (Fri,) studied this question.