Advancing Zero-Inflated Tweedie Models and Evaluating Gradient Boosting Libraries for Auto Claims (Q6868529)

The study's focus is on insurance claim prediction, which is essential for the insurance industry in calculating premiums for policyholders. The study is based on the Tweedie model's widespread use in the insurance industry to forecast premiums for policyholders. Another crucial element of the study is the fact that the right-skewed and zero-inflated traits of claims data in property and casualty insurance pose specific difficulties. The paper extends previous results in the actuarial literature by introducing functional links between the expected claim amount and the probability of zero inflation; in addition, sophisticated machine learning approaches, specifically gradient boosting methods, are utilized to overcome the limitations of traditional generalized linear models. Within this framework, after an overview on the gradient-boosted decision trees algorithm, comparisons are made between the performance of well-known gradient boosting libraries, i.e. XGBoost, LightGBM, and CatBoost. Then, the zero-inflated Tweedie boosted tree models are discussed. The paper concludes with the application of the models presented to two datasets of auto insurance claims. The Appendices contain detailed information about the training algorithm and the dataset.