SHAP for Actuaries: Explain any Model

Abstract

This tutorial gives an overview of SHAP (SHapley Additive exPlanation), one of the most commonly used techniques for examining a black-box machine learning (ML) model. Besides providing the necessary game theoretic background, we show how typical SHAP analyses are performed and used to gain insights about the model. The methods are illustrated on a simulated insurance data set of car claim frequencies using different ML models and different SHAP algorithms.

Keywords: XAI, explainability, machine learning, SHAP, Shapley values, regression modeling, interaction, partial dependence plot, motor insurance, claims frequency

JEL Classification: G22, C45, C18, C52, C55, C71