Hurricane loss modeling using insurance claims data
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
This study presents the predictive modeling of hurricane wind damage and loss using insurance claims data from four recent hurricanes [Matthew (2016), Florence (2018), Dorian (2019), and Isaias (2020)] in the coastal region of North Carolina. Four statistical/machine learning models—Ordinary Least Squares Regression (OLS), Tweedie Regression (Tweedie), Extreme Gradient Boosting Regressor (XGBoost), and Extreme Gradient Boosting Tweedie Regressor (XGBoost-Tweedie)—are developed to predict hurricane loss amount and loss ratio. The models are evaluated using different evaluation metrics at the total study area, county level, and individual building level. The results of the out-of-sample predictive performance of the models show that, although the Tweedie model provides good predictions with less than 6% error at the county and level and total study area for all the hurricanes combined, the XGBoost-Tweedie model performs better overall when the models are evaluated for each hurricane individually at all three spatial scales for both response variables. Key variables influencing hurricane loss were identified, and the OLS model was used to understand the underlying relationships between the predictors and hurricane loss.
Description
Keywords
Hurricane, Insurance claims, Loss modeling, Machine learning, Natural disaster