A Hybrid Machine Learning Framework for Predicting Hurricane Losses in Parametric Insurance with Highly Imbalanced Data

This paper proposes a novel methodology, based on machine learning and statistical models, for predicting hurricane-related losses to specific assets. Our approach uses three critical storm parameters typically tracked by meteorological agencies: maximum wind speed, minimum sea level pressure, and radius of maximum wind. The system categorizes potential damage events into three insurance-relevant classes: non-payable, partially payable, and fully payable. Three triggers for final payouts were designed: hybrid framework, standalone regression, and standalone non-linear regression. The hybrid framework combines two classification models and a non-linear regression model in an ensemble specifically designed to minimize the absolute differences between predicted and actual payouts (Total Absolute Error or TAE), addressing highly imbalanced and partially compensable events. Although this complex approach may not be suitable for all current contracts due to limited interpretability, it provides an approximate lower bound for the minimization of the absolute error. The standalone non-linear regression model is structurally simpler, yet it likewise offers limited transparency. This hybrid framework is not intended for direct deployment in parametric insurance contracts, but rather serves as a benchmarking and research tool to quantify the achievable reduction in basis risk under highly imbalanced conditions. The standalone linear regression provides an interpretable linear regression model optimized for feature selection and interaction terms, enabling direct deployment in parametric insurance contracts while maintaining transparency. These three approaches allow analysis of the trade-off between model complexity, predictive performance, and interpretability. The three approaches are compared using comprehensive hurricane simulation data from an industry-standard catastrophe model. The methodology is particularly valuable for parametric insurance applications, where rapid assessment and claims settlement are essential.