Retrieval-Augmented Time Series Forecasting

Retrieval-augmented time series forecasting studies how past examples, external information, or stored model errors can complement forecasts learned from numerical sequences. Common approaches retrieve similar historical windows or input-output analogs, then use alignment, cross-attention, gating, or weighted aggregation to combine them with a forecast; some methods use spectral or channel-wise matching, text and metadata, or residual-based correction. These methods are evaluated on settings such as long-horizon and zero-shot forecasting, non-stationary data, and cold-start or cross-region prediction, often alongside time-series foundation models. A central tension is that retrieved context can help with regime changes or limited observations but can also mislead when examples are poorly matched, stale, or selected using information unavailable at decision time.