Align-RAG Boosts Time Series Forecasting with Training-Free Context Alignment.

Mohammad Asadi, Soheil Hor, Bardiya Akhbari, Jack W. O'Sullivan, Tahoura Nedaee, Layne C. Price, Raviteja Anantha, Euan Ashley, Ehsan Adeli· August 7, 2026 View original

Key takeaways

  • Align-RAG improves TSFM forecasting with training-free context alignment.
  • It uses closed-form amplitude rescaling and phase shifting for retrieved data.
  • Frozen TSFMs can dynamically incorporate context without learned fusion modules.
  • Align-RAG outperforms state-of-the-art trained adapters, reducing MSE.

Who benefits

FinanceSupply ChainEnergyManufacturingRetail

Summary

Align-RAG is a training-free method that significantly improves Time Series Foundation Models (TSFMs) by applying closed-form amplitude rescaling and phase shifting to retrieved past-future windows before in-context learning. It outperforms learned fusion modules, demonstrating that frozen TSFMs can dynamically incorporate context without additional training.

Retrieval-augmented forecasting for Time Series Foundation Models (TSFMs) typically relies on learned fusion modules to integrate retrieved examples, assuming frozen backbones cannot handle context dynamically. This paper challenges that assumption with Align-RAG, a novel training-free method. Align-RAG enhances in-context learning by applying simple, closed-form amplitude rescaling and integer-lag phase shifts to retrieved past-future windows before feeding them into a frozen TSFM. This direct alignment allows the TSFM to dynamically incorporate context without any learned parameters. Empirical results show Align-RAG outperforming state-of-the-art trained retrieval adapters on multiple datasets, reducing MSE by an average of 3.75% on Chronos-Bolt. It also improves zero-shot MSE across various frozen TSFMs. The research suggests that frozen TSFMs inherently support dynamic in-context use of retrievals, making closed-form alignment a strong baseline before considering complex fusion modules.

Why it matters

Professionals in data science and engineering can achieve significant improvements in time series forecasting accuracy with Align-RAG, without the computational cost and complexity of fine-tuning or training additional fusion modules.

How to implement this in your domain

  1. 1Apply Align-RAG to existing Time Series Foundation Models for improved forecasting accuracy without fine-tuning.
  2. 2Implement closed-form amplitude rescaling and phase shifting for retrieved time series data in your RAG pipelines.
  3. 3Evaluate the performance of training-free alignment methods before considering more complex learned fusion modules for TSFMs.
  4. 4Integrate Align-RAG into your time series prediction workflows to leverage dynamic in-context learning.

Original post by Mohammad Asadi, Soheil Hor, Bardiya Akhbari, Jack W. O'Sullivan, Tahoura Nedaee, Layne C. Price, Raviteja Anantha, Euan Ashley, Ehsan Adeli

"arXiv:2608.05571v1 Announce Type: new Abstract: Retrieval-augmented forecasting promises to adapt frozen Time Series Foundation Models (TSFMs) to new domains without fine-tuning, but recent methods typically rely on learned fusion modules, i.e., trained adapters that merge retrie…"

View on X

Originally posted by Mohammad Asadi, Soheil Hor, Bardiya Akhbari, Jack W. O'Sullivan, Tahoura Nedaee, Layne C. Price, Raviteja Anantha, Euan Ashley, Ehsan Adeli on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses