TS-RAG Brings Retrieval Augmentation to Time-Series Forecasting via Purpose-Built Reference Tokens

arXiv 2608.06223·low signalXBlueskyLinkedInCopy link

TS-RAG (2608.06223, submitted 2026-08-06) notes RAG remains largely unexplored in time-series forecasting, and that the language-model recipe does not transfer: forecasting models have limited training data, smaller parameter counts, and no generative capacity, so simply concatenating retrieved similar sequences into the input does not work. The method instead introduces designed reference tokens that fuse the input sequence with retrieved similar sequences, aiming to capture complex temporal dynamics more robustly. The paper claims consistent state-of-the-art results across several real-world forecasting benchmarks, though the abstract names neither the datasets nor the margins — single-source and unquantified, so verify before adopting.