Research map / Time series / Long-term forecasting
Transformers for long-term forecasting
Adapt transformers with patching, decomposition and sharpness-aware training for long-term forecasting.
25 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in Long-term forecasting, keyword cluster “long term · term time · term forecasting”
State History Tried Baselines Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
14 papers in 2023–24 and 9 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 13.6% to 3.7%.
Open questions the data points to
- Papers report Dlinear on Exchange (avg.) (MSE) anywhere from 0.297 to 0.354: comparisons against it depend on whose number you take.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2021 · Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting (NeurIPS 2021 · 501 citations)
- 2022 · FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting (ICML 2022 · 548 citations)
- 2023 · Are Transformers Effective for Time Series Forecasting? (AAAI 2023 · 3,150 citations)
- 2024 · Are Self-Attentions Effective for Time Series Forecasting? (NeurIPS 2024 · 34 citations)
- 2025 · Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting (AAAI 2025 · 41 citations)
- 2026 · MixLinear: Extreme Low Resource Multivariate Time Series Forecasting with $0.1K$ Parameters (ICLR 2026 · 0 citations)
Earliest papers here
- Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting (NeurIPS 2021 · 501 citations)
- FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting (ICML 2022 · 548 citations)
- Are Transformers Effective for Time Series Forecasting? (AAAI 2023 · 3,150 citations)
What has been tried
21 papers here report 212 measured changes (a component added, removed or swapped, compared with a baseline): helped 69%, no clear effect 3%, hurt 19%, mixed 9%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add MPE | ECL · MSE | 0.183 -> 0.171 | Helped | Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting (AAAI 2025 · 41 citations) |
| Replace self-attention with a linear embedding layer | forecasting horizons across datasets (all reported cases) · forecasting performance | Helped | Are Self-Attentions Effective for Time Series Forecasting? (NeurIPS 2024 · 34 citations) | |
| Add IDMB | ETTh1 (avg.) · MSE | 0.423 -> 0.411 | Helped | Affirm: Interactive Mamba with Adaptive Fourier Filters for Long-term Time Series Forecasting (AAAI 2025 · 24 citations) |
| Replace the linear mapping with a three-layer nonlinear MLP mapping | Eight LTSF datasets, four horizons · MSE | 0.323 vs 0.325 | Mixed | Rethinking Fourier Transform from A Basis Functions Perspective for Long-term Time Series Forecasting (NeurIPS 2024 · 10 citations) |
| replace discrete token representation with continuous features | Electricity, Traffic, and Taxi (prediction length 48) · CRPS_sum | 0.025(.002) vs 0.327(.004); 0.034(.001) vs 0,467(.011); 0.166(.005) vs 0.861(.006) | Hurt | HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting (AAAI 2025 · 8 citations) |
| add instance normalization | ERA5-S average, input-3072-pred-96 · MSE | 0.168 -> 0.172 | Hurt | Timer-XL: Long-Context Transformers for Unified Time Series Forecasting (ICLR 2025 · 6 citations) |
| Use delay embedding rather than folding | FRED-MD · marg | .012 -> .017 | Hurt | Utilizing Image Transforms and Diffusion Models for Generative Modeling of Short and Long Time Series (NeurIPS 2024 · 6 citations) |
| add periodic components rather than constructing the pyramid from patches | ETTh1 long-term forecasting · MSE | 0.734 -> 0.409 | Helped | Peri-midFormer: Periodic Pyramid Transformer for Time Series Analysis (NeurIPS 2024 · 5 citations) |
| Add the selection mechanism to SCNN | ECL and Traffic · MSE and MAE | approximate 8% improvement | Helped | Parsimony or Capability? Decomposition Delivers Both in Long-term Time Series Forecasting (NeurIPS 2024 · 5 citations) |
| Use signal decay-based loss instead of MSE | ETTm1 · MSE | 0.390 -> 0.383 | Helped | CARD: Channel Aligned Robust Blend Transformer for Time Series Forecasting (ICLR 2024 · 4 citations) |
| add complementary sequences (K ≥ 3) to PatchTST | Exchange, Weather, ETTm1, and ETTh1 · forecasting error | Helped | Sequence Complementor: Complementing Transformers for Time Series Forecasting with Learnable Sequences (AAAI 2025 · 3 citations) | |
| Train the transformer with SAM | Synthetic validation data · validation loss | Helped | SAMformer: Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness-Aware Minimization and Channel-Wise Attention (ICML 2024 · 1 citation) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| Dlinear | Exchange (avg.) | MSE | 0.297 – 0.354 | 2 papers, e.g. AAAI 2025 and AAAI 2025 |
| FEDformer | Weather, horizon 96 | MSE | 0.217 – 0.238 | 2 papers, e.g. AAAI 2023 and ICLR 2023 |
What reviewers will expect
Datasets and baselines named in the 6 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Traffic · 5 of 6 papers
- Weather · 4 of 6 papers
- ETTh1 · 3 of 6 papers
- ETTh2 · 3 of 6 papers
- ETTm1 · 3 of 6 papers
- ETTm2 · 3 of 6 papers
- Electricity · 3 of 6 papers
- Exchange · 2 of 6 papers
Baselines
- iTransformer · 4 of 6 papers
- PatchTST · 4 of 6 papers
- DLinear · 4 of 6 papers
- Crossformer · 3 of 6 papers
- RLinear · 3 of 6 papers
- TimesNet · 3 of 6 papers
- TimeMixer · 2 of 6 papers
- SCINet · 2 of 6 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 23 papers; from OpenAlex and the papers' affiliations).
- University of Science and Technology of China · 3 papers
- University of Technology Sydney · 2 papers
Use it
Download BibTeX of 25 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Lightweight long-term forecasters (94 papers)
Elsewhere in the topic
- Attention for time series forecasting (123 papers)
- Graph-based multivariate forecasting (87 papers)
- Spatiotemporal forecasting models (77 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).