Research map / Time series / Causal and generative time series
Retrieval Augmented Time Series Forecasting
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 RAFT, an MLP-based time series forecaster that retrieves similar historical patches (with their future values) from the training series at multiple downsampled periods and feeds them alongside the input, reporting best MSE on most of ten benchmarks.
Problem
Long-horizon time series forecasting struggles with infrequent, temporally uncorrelated patterns that models must memorize in their weights; the paper asks whether explicit retrieval of historical patterns can supply useful inductive bias.
Approach
Retrieve top-m historical patches from the single training series whose inputs correlate most (Pearson, offset-removed) with the query, and use their subsequent values as extra input to a shallow linear/MLP predictor, applied across multiple average-pooled periods {1,2,4} and aggregated by summation.
Key results
- RAFT achieves the best average MSE on 8 of 10 datasets (e.g., ETTh1 0.420 vs TimeMixer 0.447), with an average win ratio of 86% over baselines.
- On synthetic data, retrieval reduces MSE by 9.2–14.7% for AR patterns and 16.0–31.5% for random-walk patterns depending on pattern occurrences.
- Adding the retrieval module to Autoformer lowers average MSE on ETTm1 from 0.588 to 0.454 and on ETTh1 from 0.496 to 0.471.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add retrieval to Autoformer predictions | Autoformer · ETTm1 | MSE: 0.588 -> 0.454 | Helped |
| add retrieval to Autoformer predictions | Autoformer · ETTh1 | MSE: 0.496 -> 0.471 | Helped |
| add retrieval to Autoformer predictions | Autoformer · ETTh2 | MSE: 0.450 -> 0.444 | Helped |
| add retrieval to Autoformer predictions | Autoformer · ETTm2 | MSE: 0.327 -> 0.326 | Helped |
| add retrieval instead of no retrieval (ablation) | RAFT without Retrieval · Synthetic AR patterns, occurrences 1 | MSE: 0.2590 -> 0.2209 | Helped |
| add retrieval instead of no retrieval (ablation) | RAFT without retrieval · Synthetic random-walk patterns, occurrences 1 | MSE: 0.2694 -> 0.1845 | Helped |
| use Pearson correlation rather than cosine similarity (ablation) | RAFT · ETTh1 univariate | MSE: 0.0561 -> 0.0559 | Helped |
| use Pearson correlation rather than cosine similarity (ablation) | RAFT · ETTm1 univariate | MSE: 0.0298 -> 0.0299 | Hurt |
| increase look-back window from 96 to 720 (ablation) | RAFT · ETTh1 | MSE: 0.387 -> 0.367 | Helped |
| increase look-back window from 96 to 720 (ablation) | RAFT · ETTm1 | MSE: 0.348 -> 0.302 | Helped |
| increase retrieval count from 1 to 20 (ablation) | RAFT · ETTh1 | MSE: 0.370 -> 0.367 | Helped |
| increase retrieval count from 1 to 20 (ablation) | RAFT · ETTm2 | MSE: 0.164 -> 0.164 | No clear effect |
Surprising findings
- Random retrieval or removing attention sometimes performs worse than no retrieval at all, per the ablation.
- RAFT is worse than TimeMixer on Exchange (0.441 vs 0.386) and PatchTST is best on Illness, so RAFT is not uniformly best.
Limitations
- Pre-computation of retrieval scales quadratically with training series length, requiring stride tuning for efficiency.
- Retrieval is restricted to a single time series; cross-series retrieval is out of scope.
- Hyperparameters (learning rate, m, lookback) are tuned per dataset and horizon via grid search on validation, and some tables mix univariate and multivariate settings.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Electricity Exchange Rate Illness Solar Traffic Weather synthetic AR and random-walk series
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Electricity Exchange Rate Illness Solar Traffic Weather
Metrics: MSE MAE
Baselines: TimeMixer PatchTST TimesNet MICN DLinear FEDformer Stationary Autoformer Informer
Builds on: TimeMixer NLinear DLinear PatchTST retrieval-augmented generation
Closest papers by meaning
- TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster (NeurIPS 2025 · similarity 0.89)
- SteerCast: Retrieval-Based Latent Steering for Decoder-Only Time Series Forecasting (NeurIPS 2026 · similarity 0.88)
- Retrieval-Augmented Diffusion Models for Time Series Forecasting (NeurIPS 2024 · similarity 0.88)
- Align-RAG: Alignment Is All You Need for TSFM In-Context Learning (NeurIPS 2026 · similarity 0.87)
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.86)
- Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization (ICML 2025 · similarity 0.86)
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.