ENSO impacts on flood risk and insurance claims in the United States: a machine learning approach

This research investigates the influence of the El Niño–Southern Oscillation (ENSO) on extreme flood events in the United States and its potential connection to flood insurance claims from the National Flood Insurance Program (NFIP). Given the recently observed increase in the frequency of extreme weather events, this study aims to quantify the correlation between ENSO indicators and recorded economic losses at state and county levels across the USA. Emphasis is particularly placed on the state of California, which is highly sensitive to El Niño events.

The methodology is based on the integration of multiple datasets, including ENSO indices from NOAA, US-CAMELS streamflow data, COBE sea surface temperature (SST), digital elevation models (DEM), National Hydrography Dataset (NHD), OpenStreetMap (OSM), and US Census data. From these datasets, geospatial and physical features were extracted, such as hydrographic and road network density, mean elevation, distance to the coastline, county centroid coordinates, and population. These features were analyzed using statistical tools, including the Pearson correlation coefficient and Threshold Exceedance Analysis, applied across multiple percentile showing thresholds (90–99%).

In addition, a machine learning model was developed to predict flood insurance claims per 100,000 residents. The results indicate that correlations between ENSO indices and streamflow data are significantly stronger than those between ENSO indices and insurance claim records, highlighting the substantial influence of socioeconomic factors on the insurance claim filing process. California exhibits the highest positive correlation between the maximum annual ENSO index and insurance claims (r ≈ 0.35). The developed CatBoost model can be used to predict a high percentage (>60%) of their variability, using both static and dynamic features.

The study concludes that ENSO indices can contribute meaningfully to flood risk prediction frameworks. Future work will focus on extending the analysis to additional states or the entire USA and incorporating new explanatory features to further improve model performance.