ARTIFICIAL INTELLIGENCE–DRIVEN TIME-SERIES RISK MODELING IN PROPERTY & CASUALTY INSURANCE: THEORY, METHODS, AND PRACTICAL PATHWAYS FOR ROBUST UNDERWRITING AND REAL-TIME DECISIONING

Abstract

This article synthesizes and extends contemporary theoretical and applied knowledge on the use of artificial intelligence (AI), machine learning (ML), and deep learning (DL) for time-series risk modeling in Property & Casualty (P&C) insurance. It presents a unified conceptual framework that integrates distributional time-series forecasting, event-sourced streaming analytics, hybrid statistical–deep learning architectures, and domain-aware engineering practices for underwriting, pricing, reserving, and loss forecasting. The abstract outlines the research scope, the methodological orientation, and the principal contributions: (1) a taxonomy of time-series risk problems in P&C insurance; (2) a detailed, text-based methodology combining classical time-series models (ARIMA family), support vector regression, and stateful neural architectures (LSTM, dendritic neuron models) with wavelet and singular spectrum transforms; (3) practical approaches to event sourcing and Kafka-based real-time risk pipelines; (4) evaluation and interpretability techniques tailored to insurer needs; and (5) a rigorous discussion of governance, deployment, and limitations. The paper shows how theoretical advances in distributional forecasting (Dobronets et al., 2021), ARIMA-deep hybrids (Li et al., 2020), wavelet-LSTM approaches (Tang et al., 2021), and Kafka stream analytics (Kesarpu & Dasari, 2025) collectively enable insurers to achieve more resilient, granular, and timely risk assessment, while also identifying open problems—particularly around regulatory explainability, model robustness to nonstationarity, and data governance. The article is intended for academic researchers, actuarial practitioners, and risk engineering teams seeking a deep, actionable synthesis of literature and methods for modern P&C risk modeling.