Machine learning-based insurance risk assessment pipeline for natural disaster prediction and claims estimation

This study develops a machine learning-based framework for disaster risk assessment, economic loss estimation, and insurance claims prediction using multi-source environmental, socioeconomic, and temporal data. The aim is to improve predictive accuracy and decision-making in insurance and disaster management systems.

A dataset of 68,485 disaster records (1953-2025) covering 10 disaster types and 49 engineered features was used. The methodology includes correlation analysis, synthetic target generation for economic losses, and a cascaded modeling pipeline consisting of classification (Random Forest, XGBoost, LightGBM), regression models (Extra Trees, Random Forest, XGBoost), and an insurance claims prediction layer. Feature engineering incorporates environmental (precipitation, elevation), socioeconomic, temporal, and program-based variables. Log-normal transformations, ensemble learning, and ablation analysis were applied for robustness.

Results demonstrate strong predictive capability of ensemble models in structured disaster risk pipelines. However, performance differences between random and temporal validation highlight dataset shift and potential synthetic bias in target construction. The framework provides a scalable decision-support system for insurance pricing, risk scoring, and early warning systems, with potential applications in climate risk analytics and catastrophe modeling.