Research map / Time series / Foundation models for time series
TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster
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 TS-RAG, which augments a frozen time series foundation model (Chronos-Bolt) with embedding-based retrieval from a knowledge base and a learnable Adaptive Retrieval Mixer, improving zero-shot forecasting on seven benchmarks.
Problem
Time series foundation models struggle with non-stationarity, distribution shifts, and lack mechanisms for dynamic external knowledge, limiting zero-shot forecasting and interpretability.
Approach
Retrieve top-k similar context windows and their future horizons from a knowledge base using a pretrained TSFM encoder, then fuse the projected retrieved horizons with the backbone's query embedding via an attention-based Adaptive Retrieval Mixer while keeping the backbone frozen.
Key results
- TS-RAG (Chronos-Bolt backbone) reduces average MSE by 3.54% and MAE by 1.43% versus Chronos-Bolt across seven benchmarks, with average MSE 0.1940 vs 0.2008.
- On Exchange Rate, TS-RAG achieves MSE 0.0627 vs 0.0673 for Chronos-Bolt, a 6.84% reduction.
- TS-RAG reports 9.62 ms/iter total inference versus 3474 ms/iter for RAF, and average MSE 0.1940 vs RAF's 0.2318 (or 0.2320 in text).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add retrieval augmentation with ARM to Chronos-Bolt | Chronos-Bolt · Seven benchmark average | MSE: 0.2008 -> 0.1940 | Helped |
| Add retrieval augmentation with ARM to Chronos-Bolt | Chronos-Bolt · Seven benchmark average | MAE: 0.2525 -> 0.2492 | Helped |
| Replace gated fusion with ARM (ablation) | TS-RAG · Seven benchmark average | MSE: 0.1982 -> 0.1940 | Helped |
| Use an in-domain rather than cross-domain retrieval knowledge base (ablation) | TS-RAG · ETTm1 | MSE: 0.3046 -> 0.2906 | Helped |
| Use an in-domain rather than multi-domain retrieval knowledge base (ablation) | TS-RAG · ETTm2 | MSE: 0.1513 -> 0.1466 | Helped |
| Increase retrieval lookback length from 64 to 512 (ablation) | TS-RAG · ETTm2 | MSE: 0.1502 -> 0.1466 | Helped |
| Increase retrieval lookback length from 64 to 512 (ablation) | TS-RAG · ETTh1 | MSE: 0.3540 -> 0.3557 | Mixed |
| Extend forecasting with retrieval augmentation (ablation) | TS-RAG · ETTm2, forecast length 720 | MSE: 0.4162 -> 0.4164 | Hurt |
| Use a Chronos, TTM, or MOMENT retriever encoder (ablation) | TS-RAG · ETTh1, ETTh2, ETTm1, ETTm2 | MSE | No clear effect |
| Use cosine rather than Euclidean retrieval distance (ablation) | TS-RAG · ETTh1 | MSE: 0.3557 -> 0.3558 | No clear effect |
| Use DTW rather than Euclidean retrieval distance (ablation) | TS-RAG · ETTh1 | MSE: 0.3557 -> 0.3606 | Hurt |
| Increase the number of retrieved sequences (ablation) | TS-RAG · Seven zero-shot evaluation datasets | MSE | Mixed |
Surprising findings
- Retrieval from a multi-domain knowledge base did not beat the in-domain knowledge base on most datasets, though in-domain results were lowest MSE overall.
- DTW-based retrieval in raw time-series space gave only modest gains and sometimes underperformed the no-retrieval baseline.
Limitations
- Relies on a fixed forecasting length of 64 with rolling for longer horizons, and retrieval per step adds overhead.
- Lookback length for retrieval shows no consistent best setting; longer is not always better.
- Evaluated only on seven benchmark datasets with Chronos-Bolt as the main backbone.
Experimental setup
Datasets: Chronos pretraining dataset ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Exchange Rate
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Exchange Rate
Metrics: MSE MAE
Baselines: Chronos-Bolt MOMENT TTM Moirai TimesFM Chronos Time-MoE RAF
Builds on: Chronos-Bolt RAF retrieval-augmented generation Chronos
Closest papers by meaning
- Align-RAG: Alignment Is All You Need for TSFM In-Context Learning (NeurIPS 2026 · similarity 0.90)
- Retrieval Augmented Time Series Forecasting (ICML 2025 · similarity 0.89)
- Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.89)
- Less is More: Unlocking Specialization of Time Series Foundation Models via Structured Pruning (NeurIPS 2025 · similarity 0.89)
- STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · similarity 0.89)
- Multi-Scale Finetuning for Encoder-based Time Series Foundation Models (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.