Application of explainable machine learning in non-life insurance pricing: a profitable segment finder tool

This paper aims to present an application in non-life insurance of the existing literature on explainable machine learning. A framework of analysis and visualization methods is proposed to put in place a profitable segment finder tool based on loss ratio modeling. It is compared to the current methods used in practice to analyse profitability in an insurance portfolio. In addition to demonstrate the use of this framework for multiple analysis purposes, the methods are also intended to be applicable in pure premium pricing. The offered methodology is composed of efficient and scalable gradient boosting machine XGBoost working in parallel with model interpretability techniques to understand the model prediction on variable importance, global effects, interactions and decomposition in main order effects. Besides these methods of global interpretability, local interpretability on individual predictions is also performed to provide the deepest understanding of the data.