This thesis investigates the use of machine learning methods in non-life insurance pricing, with a par- ticular focus on claim frequency and claim severity modeling. The study compares the traditional Generalized Linear Model (GLM) with three machine learning methods: Random Forest (RF), Gra- dient Boosting Machine (GBM) and Neural Networks (NN). Three different insurance datasets were analyzed, consisting of insurance data from different countries.
The models were evaluated using Poisson deviance for claim frequency and Gamma deviance for claim severity. In addition, SHAP (SHapley Additive exPlanations) values were applied to investigate model interpretability.
The results show that GBM achieved the best or near-best predictive performance across the datasets. The ML methods performed the best, as expected, when the data includes stronger interaction effects and nonlinear patterns. The SHAP analyses showed that RF and GBM identified relatively consistent and interpretable relationships between risk factors and predicted claims, while the neural network produced more dispersed and less interpretable patterns.
To sum up, the machine learning methods can improve predictive performance in non-life insurance pricing but the traditional GLM remains relevant thanks to its interpretability and competitive per- formance. The ML methods performed better on two of three tested datasets.