Evaluating XGBoost for Competitive Insurance Pricing: A Case Study on Motor Third-Party Liability Insurance

In numerous studies, the Gradient Boosting Machine (GBM) has shown strong out-of-sample performance on insurance claim data. This evidence has led to increased interest in enhanced versions of GBM, such as XGBoost. Some studies highlight the superiority of XGBoost over GBM on insurance claim data. However, many of them focus only on statistical model fit rather than assessing the model's effectiveness as a pricing tool, especially its economic value. Our study takes a different approach by evaluating the potential of XGBoost as a pricing model through both out-of-sample statistical performance and model lift. The experiment was conducted using Motor Third-Party Liability (MTPL) claim dataset obtained from a Belgian insurer in 1997. In terms of statistical performance, results show that XGBoost and GBM perform similarly on the frequency component, outperforming other models like Random Forest or Generalized Additive Model (GAM). For the severity component, neither XGBoost nor GBM surpasses the simple GAM. However, XGBoost is still generally a better severity model than GBM in most trials. Model lift reveals that XGBoost offers an economic advantage over GBM, especially when assessing low-risk policyholders. According to the Gini index, XGBoost is also identified as a model with minimal susceptibility to adverse selection, making it a promising choice for insurance pricing. Overall, our findings suggest that XGBoost is effective both statistically and economically. This supports its potential as a powerful pricing model.