Learning Inter-Annual Flood Loss Risk Models from Historical Flood Insurance Claims

Flooding is one of the most disastrous natural hazards responsible for substantial economic losses. To address this issue, developing a predictive model for flood-induced financial damages holds immense value in various areas, such as climate change adaptation planning and insurance underwriting. This research evaluates the predictive capabilities of different regression models, including neural networks (Conditional Generative Adversarial Networks), decision trees (Extreme Gradient Boosting), and kernel-based regressors (Gaussian Process), utilizing the National Flood Insurance Program (NFIP) dataset for the years between 2000 and 2020. By leveraging these models, the study identifies the most informative predictors for regression, shedding light on crucial factors influencing flood-related financial damages. To enhance the predictive performance of the regressors, a Burr distribution is employed along with a bias correction scheme. A study of the coastal counties in eight US Southern states resulted in an R2 = 0.807. Additionally, a more detailed analysis of 11 counties with significant claims in the NFIP dataset reveals that Extreme Gradient Boosting yields the most favorable results among the evaluated models. Moreover, bias correction significantly improves the similarity between the predicted claim amount distributions and the reference distribution, enhancing the model’s accuracy and reliability