Advancing Zero-Inflated Tweedie Models and Evaluating Gradient Boosting Libraries for Auto Claims
Banghee So, Min Deng · North American Actuarial Journal · 2025
The Tweedie model is extensively employed in the insurance industry to forecast premiums for policyholders. However, the right-skewed and zero-inflated traits of claims data in property and casualty (P&C) insurance pose distinct challenges. This research extends the work of So and Valdez by introducing additional functional links between the expected claim amount μ and the zero inflation probability q, greatly improving the model’s efficiency, fit accuracy, and predictive precision. To overcome the constraints of traditional generalized linear models (GLMs) with intricate data interactions, we employ sophisticated machine learning approaches, particularly gradient boosting methods. We conduct an in-depth assessment of three leading gradient boosting libraries—XGBoost, LightGBM, and CatBoost—scrutinizing their effectiveness in managing zero-inflated Tweedie models and adapting to the distinct features of insurance claim data. Our study uses synthetic telematics and French Motor Third-Party Liability datasets to test the proposed model enhancements. This article advances actuarial science by providing a more comprehensive understanding of zero-inflated models and enriching the toolkit for forecasting insurance claims with advanced statistical methods and machine learning.