New RAG Method Boosts Time Series Forecasting Accuracy

Time series forecasting often benefits from large pretrained foundation models, but adapting them to specific domains with limited historical data usually requires either computationally expensive fine-tuning or Retrieval Augmented Generation (RAG) techniques. This research focuses on RAG, specifically introducing a new method called Retrieval-Augmented Extended Forecasting (RAEF) as a more efficient alternative. RAEF builds upon existing Retrieval Augmented Forecasting (RAF) by incorporating two key improvements. First, it performs direct retrieval in the input space rather than the embedding space, which significantly reduces inference overhead. Second, it uses a concatenation-based aggregation strategy that better preserves the temporal structure of the data, as opposed to simple averaging. Empirical evaluations across various benchmark datasets demonstrate that RAEF not only surpasses RAF in both accuracy and computational efficiency but also achieves performance comparable to or even superior to fine-tuning methods. This positions RAEF as a practical and scalable solution for adapting time series foundation models to new domains without the heavy computational burden of full fine-tuning.