Scholarlink

<p>This study presents a two-model approach for predicting hurricane wind building loss using insurance policy and claims data from four recent hurricanes (Matthew, Florence, Dorian, and Isaias) in eastern North Carolina. The first model predicts the probability a building experiences a loss; the second estimates the dollar value of the incurred loss. Four types of statistical and machine learning models were compared for each step, and the performance of the models was evaluated using a suite of metrics at different spatial scales. Using an XGBoost loss occurrence model and an XGBoost-Tweedie loss amount model—identified as the best models—we find that for a future hurricane, we are able to predict the total number of claims in the study area with an error of 2–12% and the total expected loss within approximately 1–16%. Results also suggest that wind speed and total precipitation are the most important features, and older and larger homes are more likely to experience a loss. More valuable, larger homes east of the Intracoastal Waterway tend to experience higher losses when they do incur a loss. These models can be useful for directly predicting loss in future hurricanes and for calibrating and validating loss simulation models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hurricane wind loss modeling using insurance claims data

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Nii Otu Tackie-Otoo,
  • Mohammad Askari,
  • Patrick Hadinata,
  • Rachel A. Davidson,
  • Ertugrul Taciroglu,
  • Gina Hardy

摘要

This study presents a two-model approach for predicting hurricane wind building loss using insurance policy and claims data from four recent hurricanes (Matthew, Florence, Dorian, and Isaias) in eastern North Carolina. The first model predicts the probability a building experiences a loss; the second estimates the dollar value of the incurred loss. Four types of statistical and machine learning models were compared for each step, and the performance of the models was evaluated using a suite of metrics at different spatial scales. Using an XGBoost loss occurrence model and an XGBoost-Tweedie loss amount model—identified as the best models—we find that for a future hurricane, we are able to predict the total number of claims in the study area with an error of 2–12% and the total expected loss within approximately 1–16%. Results also suggest that wind speed and total precipitation are the most important features, and older and larger homes are more likely to experience a loss. More valuable, larger homes east of the Intracoastal Waterway tend to experience higher losses when they do incur a loss. These models can be useful for directly predicting loss in future hurricanes and for calibrating and validating loss simulation models.