Investigating the impact of the ENSO index on flood risk and compensation claims in the USA

This thesis investigates the influence of ENSO (El Niño–Southern Oscillation) phenomena in accorance to extreme flood events in the United States, as well as their potential connection to flood insurance claims (National Flood Insurance Program (NFIP)). Recently an increasing frequency of extreme weather events has been observed, so this study aims to quantify the correlation between ENSO indicators and recorded economic data, at State and County level across the U.S.A. The thesis emphasizes on the analysis of the state of California, which is the most sensitive to El Niño events. The methodology is based on the use of multiple datasets (ENSO indices from NOAA, US-CAMELS streamflow data, COBE SST, DEM, NHD, OSM, US Census), with the aim of extracting geospatial and physical features such as hydrographic and transport (road) density, mean elevation, distance to the sea, county centroid coordinates, and population. These features were analyzed by statistical tools ,such as the Pearson correlation coefficient and the Threshold Overcome Analysis method, applied across various thresholds (90–99%). Also, this thesis includes the development of a machine learning model, aiming to predict insurance claim counts per 100,000 residents. The analysis reults show that the correlation between ENSO indices and streamflow data is significantly stronger than the correlation with insurance claim records. This is explained by the big impact of social and economic factors on the claims filling procedure. The state of California has the highest positive correlation between the maximum annual ENSO index (MAX_ENSO) and insurance claims (r ≈ 0.35). The CatBoost model, we developed, achieved good accuracy (R² = 0.638) with the use of static and dynamic features, as well. The study concludes that ENSO indices can contribute in the flood risk prediction process. Some good future proposals are the appliance of the same methodology to additional states or the whole USA and the addition of new input features, in order to maximize performance of the model.