Bayesian CART for insurance pricing
An insurance portfolio offers protection against a specified type of risk to a collection of policy-holders with various risk profiles. Insurance companies use risk factors to group policyholders with similar risk profiles in tariff classes. Premiums are set to be equal for policyholders within the same tariff class which should reflect the inherent riskiness of each class. Both accuracy and interpretability of the model used are essential in (non-life) insurance pricing. Tree-based methods, like the classification and regression tree (CART), have gained popularity as they can in some cases give good performance and be easily interpretable. In this talk, we discuss a Bayesian approach applied to CART models. The two basic components of this approach consist of prior specification and stochastic search using MCMC. The idea is to have the prior induce a posterior distribution that will guide the stochastic search towards more promising trees. We shall introduce the Bayesian CART models for insurance claims data, which include various frequency models and a frequency-severity model. The proposed method aims to search for trees which can better classify the policyholders into risk groups. Some simulation and real data examples will be discussed.
Seminar organized by Prof. Enkelejd Hashorva