Reproduce “TimeRAF: Retrieval-Augmented Foundation model for Zero-shot Time Series Forecasting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Huanyu Zhang, Chang Xu, Yi-Fan Zhang, Zhang Zhang, Liang Wang, Jiang Bian, Tieniu Tan
TimeRAF enhances zero-shot time series forecasting by integrating retrieval augmentation with Time Series Foundation Models (TSFMs). The system achieves improved accuracy by dynamically incorporating external knowledge from a curated knowledge base via a learnable retriever and a novel Channel Prompting mechanism.
Key insights
- Dynamically retrieving and integrating external time series knowledge can significantly enhance the zero-shot generalization capabilities of Time Series Foundation Models (TSFMs).
- Training a retriever based on the utility of the retrieved information for the downstream forecasting task (utility-based supervision) is more effective than relying solely on embedding similarity.
- Efficiently enhancing TSFMs is possible by keeping the foundation model backbone frozen and only training lightweight retrieval and integration modules.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.