SERAF adds semantic retrieval to time-series forecasting

A new arXiv paper proposes SERAF, a semantics-enhanced retrieval-augmented framework for time-series forecasting. The method retrieves both historical time-series segments and textual descriptions of those patterns.

The authors argue that relying only on numerical similarity can fail when time series are non-stationary. Adding semantic descriptions gives the model another way to find relevant historical context.

The work extends the RAG idea beyond text generation into forecasting, showing how retrieval can support model performance in structured prediction tasks.