Deep Learning-Based Prediction of Insurance Claim Severity Using Structured and Unstructured Data

Abstract

Accurate prediction of insurance claim severity remains a fundamental challenge in actuarial science due to nonlinear feature interactions, contextual variability, and heavy-tailed loss distributions. Traditional statistical models primarily rely on structured policy variables and often fail to capture semantic signals embedded in unstructured claim narratives. This study proposes a multimodal deep learning framework that integrates structured insurance data with contextual embeddings derived from claim descriptions to enhance severity regression performance. The proposed architecture employs independent encoding layers for structured and textual inputs, followed by an intermediate fusion mechanism that learns cross-modal feature interactions. Log transformation is applied to stabilize skewed loss distributions, and model training is conducted using a robust optimization strategy with cross-validation. The framework is benchmarked against classical actuarial models, tree-based ensemble methods, and structured-only neural networks. Experimental evaluation demonstrates that multimodal integration significantly improves predictive accuracy, reduces error dispersion, and enhances model calibration in high-severity cases. The results confirm that textual narratives provide incremental predictive value beyond traditional structured attributes. From a practical perspective, improved severity estimation supports more precise pricing, reserve allocation, and risk management decisions. This research establishes a scalable and production-oriented multimodal architecture for insurance analytics and highlights the strategic importance of combining structured and unstructured data in actuarial prediction systems.