Leveraging Weather Dynamics in Insurance Claims Triage Using Deep Learning

Abstract

In property insurance claims triage, insurers often use static information to assess the severity of a claim and to identify the subsequent actions. We hypothesize that the pattern of weather conditions throughout the course of a loss event is predictive of the insured losses, and hence appropriate use of weather dynamics improves the operation of insurers’ claim management. To test this hypothesis, we propose a deep learning method to incorporate dynamic weather information in the predictive modeling of the insured losses for reported claims. The proposed method features a hierarchical network architecture to address the challenges in claims triage due to the nature of weather dynamics. In the empirical analysis, we examine a portfolio of hail damage property insurance claims obtained from a major U.S. insurance carrier. When supplemented by dynamic weather information, the deep learning method exhibits substantial improvement in the hold-out predictive performance. We further design a cost-conscious decision strategy for triaging claims using the probabilistic forecasts of the insurance claim amounts. We show that leveraging weather dynamics in claims triage leads to a reduction of up to 9% and 6% in operational costs compared to when the triaging decision is based on forecasts without any weather information and with only static weather information, respectively. Supplementary materials for this article are available online.

Keywords:

• Finite mixture model
• Longitudinal data
• Neural network
• Probabilistic forecasting
• Property insurance

Supplementary Materials

Supplementary materials summarize additional data features and analytical results. The sample data and program codes are also provided.

Acknowledgments

We thank Editor Prof. Michael Stein, the Associate Editor, and anonymous reviewers for their invaluable and insightful comments that have significantly improved the quality of our paper. The first author also acknowledge the generous support for this research from the Data Science Institute of University of Wisconsin-Madison and the Wisconsin Alumni Research Foundation.

Disclosure Statement

No potential conflict of interest was reported by the author(s).