Using Large Language Models to Generate New Features from Text Data for Loss Prediction

Insurance companies collect large volumes of unstructured text data, such as claims descriptions, adjuster notes, and customer feedback. While these data contain valuable context, their unstructured nature leads to underutilization in traditional actuarial models. This paper investigates the use of large language models (LLMs) to extract low-dimensional structured features from claims descriptions to improve loss prediction models. With structured prompts, we apply GPT-4o and Llama-3.2-3B to classify incidents into ordinal risk categories. The resulting labels are semantically meaningful, ordinally consistent, and predictive of average loss. Incorporating them into generalized linear models improves both in-sample fit and out-of-sample accuracy, with five-category labels from GPT-4o yielding the best performance. These results demonstrate that LLMs can effectively augment traditional actuarial risk models by extracting meaningful information from unstructured text data.