Modeling NFIP-insured housing losses from Texas and Florida flood disasters: A comparative analysis of hurricane Harvey, the 2016 tax day flood, and hurricane Irma

Source abstract

Reliable flood loss models can support rapid mitigation and recovery decisions, but their value depends on whether relationships learned from one disaster are useful in another. We evaluate tract-level National Flood Insurance Program (NFIP)-insured housing losses across 6065 census tracts affected by the 2016 Tax Day Flood, Hurricane Harvey, and Hurricane Irma. Mean normalized loss ratio (Mean_NLR) and the probability of observed NFIP-insured loss are modeled with a parsimonious hazard-exposure-vulnerability (HEV) specification driven by precipitation, Special Flood Hazard Area (SFHA) share, insurance penetration, population density, social vulnerability, building age, and Community Rating System discounts. We used nested grouped spatial cross-validation to tune Random Forest, XGBoost, and logistic classification models while retaining OLS and spatial lag models as benchmarks. Tuned regression performance is modest, with out-of-fold R 2 peaking at 0.34 for Harvey and lower values for Tax Day and Irma. Binary classification is stronger, with within-event AUC of 0.81–0.90 and Harvey PR-AUC up to 0.71. Transfer is high between the two Texas events, with XGBoost AUC of 0.92–0.95. Although population density and precipitation are consistently influential, AUC declines when Texas-trained models are applied to Irma because hazard mechanisms, exposure patterns, and predictor distributions differ across regions. Because the NFIP claims used to train these models capture insured and claimed building losses with capped payments, the models are appropriate for screening observed NFIP-insured losses. These models should not be used to estimate total physical damage or as universal flood loss transfer functions.