Global Forecasting of Extreme Weather and Insurance Losses Using an LSTM-Based, Audit-Ready Framework

Abstract. The property and casualty insurance industry increasingly relies on deep neural networks to quantify weather-driven risks. This study develops a forecasting framework based on long short-term memory networks to estimate the global impact of extreme weather on insurance claims by integrating authoritative meteorological and financial datasets. Specifically, we leverage globally consistent records of temperature, precipitation, and snowfall from NOAA’s National Centers for Environmental Information (1995–2025) and insured-loss statistics from Swiss Re’s global catastrophe reports, rather than focusing on local or regional case studies. The model captures long-range temporal dependencies without manual feature engineering and employs adaptive moment estimation to stabilize training and reduce prediction errors. A fully connected layer with rectified linear unit activation enhances nonlinear fitting, while post-hoc Shapley additive explanations clarify how weather variables and recent claims shape predicted losses. Benchmarks against classical baselines—random forest, support vector machine, and autoregressive integrated moving average—demonstrate consistent accuracy gains. Using three decades of data, including a decade reserved for out-of-sample evaluation, the framework delivers accurate forecasts with transparent attributions that support pricing, reinsurance planning, and catastrophe response under climate risk. This integrates extreme-weather signals with insurance losses, based on globally aggregated datasets, to provide reproducible, regulator-auditable global insights.