Machine learning models for predicting flood insurance claims through recorded streamflow and historical-claims data integration

Floods rank among the most financially devastating natural hazards, imposing significant risks on communities, insurers, and policymakers. In recent years, intensive advancements in artificial intelligence and machine learning research have enhanced the ability to mitigate these risks but also for predicting vulnerable areas, claim amounts, and patterns of flood impact. In this context, our study explores the potential of machine learning models to predict flood insurance claims based on historical streamflow data, actual flood claim records, and regional characteristics. To this respect, we integrate the US-CAMELS dataset, which provides detailed streamflow timeseries, with Federal Emergency Management Agency’s National Flood Insurance Program (NFIP) Redacted Claims dataset, containing millions of flood-related insurance claims across the contiguous USA. This integration yields a composite dataset featuring streamflow metrics—such as intensity, duration, and recurrence—alongside FEMA variables, including claim history, flood frequency, and policy characteristics.
Our approach employs machine learning models to predict outcomes such as expected aggregated insurance claims and the likelihood of claim occurrences across different regions, while simultaneously evaluating model’s performance. Through this methodology, we identify critical predictors of flood-related insurance claims, providing valuable insights for risk assessment, enhancing the non-structural elements of early warning systems and economic resilience in flood-prone areas, thus, contributing to the development of proactive and data-driven insurance strategies.