Research map / Time series / Foundation models for time series

Align-RAG: Alignment Is All You Need for TSFM In-Context Learning

NeurIPS 2026 · Time series › Foundation models for time series · approach: Time series foundation models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Align-RAG, a training-free retrieval-augmented forecasting method that applies closed-form amplitude rescaling and integer-lag phase alignment to retrieved past-future windows before feeding them as in-context demonstrations to a frozen TSFM.

Problem

Retrieval-augmented forecasting for frozen time series foundation models currently relies on trained fusion modules; the paper asks whether frozen TSFMs can use retrieved context directly if it is aligned to the query.

Approach

Align each retrieved past-future pair to the query via a Wiener-shrunk affine amplitude fit (applied to both past and future) and an integer-lag cross-correlation phase shift, then pack the aligned pairs as in-context demonstrations for a frozen backbone with no learned parameters.

Key results

• On frozen Chronos-Bolt, Align-RAG beats TS-RAG on 7/7 datasets in MSE (avg -3.75%) and 6/7 in MAE, with no learned parameters.
• Across four additional frozen TSFMs, Align-RAG reduces zero-shot MSE by 2.5% to 13.7% per backbone; Moirai gains the most at avg -13.73%.
• A retriever x alignment ablation shows that turning alignment off costs over 20 pp of MSE while swapping trained retrieval for random retrieval costs 0.84 pp.

Surprising findings

• Random retrieval with alignment performs nearly as well as the trained TS-RAG retriever, suggesting alignment rather than retrieval ranking drives most of the gain.
• Shuffling futures within the demonstration set reverses the MSE gain on ETTm1 (from -4.2% to +12.2%) while the ridge-like signature stays high.

Limitations

• The mechanism evidence is correlational: the rho_GD signature is directional, not an equivalence to ridge regression.
• Evaluation uses the univariate TS-RAG benchmark; multivariate covariates and distribution shift are untested.
• The retriever x alignment ablation and mechanism analysis are only run on Chronos-Bolt (and partly Chronos-2).

Experimental setup

Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Exchange Electricity

Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Exchange Electricity

Metrics: MSE MAE

Baselines: TS-RAG RAFT TimeRAF RATD TimesFM-ICF Naive RAF Advanced RAF TimeRAG

Builds on: TS-RAG Chronos-Bolt Maximal Marginal Relevance RAFT

Closest papers by meaning

• TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster (NeurIPS 2025 · similarity 0.90)
• Less is More: Unlocking Specialization of Time Series Foundation Models via Structured Pruning (NeurIPS 2025 · similarity 0.87)
• TRACE: Grounding Time Series in Context for Multimodal Embedding and Retrieval (NeurIPS 2025 · similarity 0.87)
• Retrieval Augmented Time Series Forecasting (ICML 2025 · similarity 0.87)
• Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting (NeurIPS 2026 · similarity 0.87)
• Context-Alignment: Activating and Enhancing LLMs Capabilities in Time Series (ICLR 2025 · similarity 0.86)