The use of Machine Learning in non-life insurance: Literature review

Abstract

Insurance companies using risk modelling mainly focus on the mastery of Genelized linear models. Nevertheless, such models hinder constraints on the structure of risk and the interactions between the risk explanatory variables. Then, these limits can lead to a biased estimation of the insurance premium in certain populations of policyholders.

The traditional insurers have to face these existential challenges. Indeed, they need a focus on data strategy and implementation of statistical learning to achieve better pricing. In the last decades, computer performance has been continuously increasing, which has allowed a widespread application of the so-called statistical learning theory (Machine Learning) in many field. Non-life insurance pricing occupies as paradoxical place in actuarial science, hence the need for the application of different algorithms to evaluate the risks that insurance companies must face. Indeed, actuaries put forward the classical methods, linear algorithms mainly generalized linear model (GLM). Unfortunately, restrictions linked to this type of model, which can bias the estimation of the insurance premium, have pushed actuaries to opt for efficient algorithms, referred to as statistical learning models.  To do this, it is essential to look at the principals of classical GLM method, to identify their limitations and then to discuss the contributions of certain statistical learning methods in non-life insurance.

Keywords: Pricing, Non-life insurance, Generalized Linear Models GLM, Statistical Learning, Classification and Regression Trees CART, Random Forest, XGBoost, Neural Networks

Classification JEL: B23, C60

Paper Type: Theoretical research