Non-Stationary Time Series Forecasting

This field studies how to forecast when trends, seasonality, volatility, or relationships between variables shift across time and regimes. Common approaches normalize or decompose changing signals, model time- and frequency-scale structure, and route predictions among regime-specific experts; online adaptation, retrieval, and continual learning use new observations or recent errors to update forecasts. Researchers evaluate methods on long-horizon and streaming tasks, including energy, finance, traffic, weather, and epidemiology, with growing attention to uncertainty and chronological benchmarks that test real temporal generalization. A central challenge is balancing quick responses to genuine shifts against stability, computational cost, and overreaction to noise or temporary changes.