SurvFormer: A Transformer Based Framework for Survival Analysis in Insurance Underwriting

Insurance claim prediction is a critical task in the underwriting process of health insurance. It helps keep insurance programs financially stable and ensures that people have fair access to coverage. In industrial insurance systems, predictive modeling based on a large amount of user data accumulated from online insurance platforms can significantly improve underwriting accuracy. However, large amounts of policies are newly underwritten especially for some newly developed insurance products. The time lag between a policy being underwritten and a corresponding insurance claim can range from several months to years. Thus, newly acquired data may lack accurate label information when used by the traditional classification algorithm, presenting a data censoring challenge in the prediction of insurance claims. Moreover, the timing of insurance claims can be significantly influenced by distinct features associated with the policyholder's behavior and circumstances. In addressing this problem, we incorporate survival analysis into our modeling framework. We propose SurvFormer, an innovative transformer-decoder-based approach designed to estimate temporal survival probabilities. SurvFormer utilizes attention mechanisms to highlight relevant user behaviors and effectively captures inter-dependencies among temporal survival states. Furthermore, to address the prevalent issue of event-censoring imbalance in survival data, we design a survival loss called imbalance-aware censoring loss (IACL) to prevent the model from being overwhelmed by the prevalence of samples that lack precise labels. Through comprehensive experiments on three real-world data sets, SurvFormer achieves significant performance improvement and maintains high efficiency compared to the state-of-the-art baseline methods.