Tree-based machine learning methods for predicting vehicle insurance claim size

Area of Science:

• Actuarial science
• Machine learning
• Statistical modeling

Background:

• Accurate modeling of vehicle insurance claim severity is crucial for handling skewed and heterogeneous loss data.
• Classical parametric regression models and modern ensemble learning approaches are key methods for this challenge.

Purpose of the Study:

• To empirically compare parametric regression models with tree-based ensemble learning for predicting conditional motor insurance claim severity.
• To evaluate model performance using out-of-sample metrics and assess predictor importance.

Main Methods:

• Ordinary Least Squares (OLS), Tweedie Generalized Linear Model (GLM), Bagging, Random Forests (RFs), and Gradient Boosting were implemented.
• Cross-sectional conditional severity framework applied to real-world motor insurance data.
• Performance evaluated using Root Mean Square Error (RMSE); variable importance assessed predictor contributions.

Main Results:

• Tree-based ensemble methods showed modest improvements in predictive accuracy over classical parametric models.
• The Tweedie GLM performed competitively as a flexible parametric benchmark for skewed positive claim amounts.
• Premium and insured value were consistently identified as key determinants of claim severity.

Conclusions:

• Ensemble learning methods can enhance traditional actuarial models by providing flexibility for non-linear effects.
• These methods maintain comparable predictive performance for moderate-complexity severity data.
• The study highlights the value of integrating advanced machine learning techniques into actuarial practice.