Pricing from experience: Predictive analytics for dynamic pricing in non-life insurance

The class of Generalized Linear Models has been widely embraced by the non-life insurance literature and industry to construct premium rate structures. However, increasingly sophisticated statistical techniques as well as new and richer sources of information have emerged that open up possibilities to improve upon classical methods. This thesis explores how we can utilize these advancements to further improve and refine the pricing of non-life insurance risks. First, we exploit the increasing availability of a posteriori information in non-life insurance companies to introduce a multivariate and dynamic form of risk classification. The resulting multi-product approach substantially outperforms the classical, static pricing method in a property and casualty insurance portfolio and it is able to identify profitable cross-selling opportunities from the observed multi-product claims experience. Secondly, we propose a novel joint experience rating approach to allow for a positive or negative dynamic frequency-severity dependence. A Motor Third Party Liability insurance application demonstrates that this joint approach is able to identify customer risk profiles with distinctive claiming behavior and that these data-driven profiles, in turn, substantially improve the classical pricing method under independence. Thirdly, we present a causal inference framework, which controls for the fact that premium renewal offers are typically based on a customer’s level of risk in non-life insurance, to measure and account for a customer's price sensitivity. This framework identifies the policy's competitiveness in the market as fundamental driver of these price sensitivities in an automobile insurance portfolio and it can yield considerably more expected profit than currently realized.