CrossRAG: Retrieval-Augmented Framework for Time Series Forecasting

A recent study published on arXiv (2607.29459) presents CrossRAG, a framework that enhances long-term multivariate time series forecasting utilizing diverse IoT sensors through retrieval augmentation. This approach tackles the shortcomings of current foundational time series models, which depend on fixed parametric knowledge and do not dynamically access historical patterns during inference. CrossRAG combines Shape-Aware Memory (SAM) with RevIN normalization for robust shape-level retrieval, employs Future-Consistent Contrastive (FCC) learning to differentiate valuable references from challenging negatives with similar pasts but different futures, and utilizes Cross-Attention Temporal Fusion to merge retrieved references with the input data. The study addresses challenges in applying retrieval-augmented generation (RAG) to time series, such as variations in magnitude and the inconsistency between historical similarity and future outcomes. It aims to tackle these issues by emphasizing shape-level patterns and future consistency. This cross-type announcement is accessible on arXiv.