Research map / Time series / Long-term forecasting
Lightweight long-term forecasters
Forecast long horizons with lean models exploiting periodicity, memory and mixing.
94 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Long-term forecasting, keyword cluster “long term · term time · term forecasting”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 6 tracked benchmarks.
23 papers in 2023–24 and 57 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 22.3% to 23.4%.
Open questions the data points to
- ETTh1 (MSE): best went from 0.405 (up to 2024) to 0.396 in 2025 — TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025)
- Papers report PatchTST on Traffic (MSE) anywhere from 0.394 to 0.555: comparisons against it depend on whose number you take.
- Papers report PatchTST on Weather (MSE) anywhere from 0.223 to 0.259: comparisons against it depend on whose number you take.
What looks settled
- Weather (MSE): best up to 2024 0.219, best in 2025–26 0.219 from 5 papers, at most 0.2% better, if at all.
- ETTm1 (MSE): best up to 2024 0.345, best in 2025–26 0.354 from 4 papers, at most 0.2% better, if at all.
- Traffic (MSE): best up to 2024 0.403, best in 2025–26 0.416 from 5 papers, at most 0.2% better, if at all.
- ETTh2 (MSE): best up to 2024 0.318, best in 2025–26 0.347 from 4 papers, at most 0.2% better, if at all.
- ETTm2 (MSE): best up to 2024 0.25, best in 2025–26 0.25 from 4 papers, at most 0.2% better, if at all.
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
- 2016 · Uncertainty Propagation in Long-Term Structured Regression on Evolving Networks (AAAI 2016 · 12 citations)
- 2017 · Nonlinear Dynamic Boltzmann Machines for Time-Series Prediction (AAAI 2017 · 70 citations)
- 2018 · Attend and Diagnose: Clinical Time Series Analysis Using Attention Models (AAAI 2018 · 456 citations)
- 2019 · Shallow RNN: Accurate Time-series Classification on Resource Constrained Devices (NeurIPS 2019 · 18 citations)
- 2020 · Tensorized LSTM with Adaptive Shared Memory for Learning Trends in Multivariate Time Series (AAAI 2020 · 40 citations)
- 2021 · Fully Exploiting Cascade Graphs for Real-time Forwarding Prediction (AAAI 2021 · 37 citations)
- 2022 · FiLM: Frequency improved Legendre Memory Model for Long-term Time Series Forecasting (NeurIPS 2022 · 104 citations)
- 2023 · TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis (ICLR 2023 · 439 citations)
- 2024 · TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting (ICLR 2024 · 93 citations)
- 2025 · WPMixer: Efficient Multi-Resolution Mixing for Long-Term Time Series Forecasting (AAAI 2025 · 77 citations)
- 2026 · Towards Non-Stationary Time Series Forecasting with Temporal Stabilization and Frequency Differencing (AAAI 2026 · 6 citations)
Earliest papers here
- Uncertainty Propagation in Long-Term Structured Regression on Evolving Networks (AAAI 2016 · 12 citations)
- Nonlinear Dynamic Boltzmann Machines for Time-Series Prediction (AAAI 2017 · 70 citations)
- Attend and Diagnose: Clinical Time Series Analysis Using Attention Models (AAAI 2018 · 456 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
ETTh1 · MSE (lower is better)
9 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0.405 | Chimera: Effectively Modeling Multivariate Time Series with 2-Dimensional State Space Models (NeurIPS 2024 · 10 citations) |
| 2025 | 0.396 | TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025) |
Weather · MSE (lower is better)
9 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0.219 | Chimera: Effectively Modeling Multivariate Time Series with 2-Dimensional State Space Models (NeurIPS 2024 · 10 citations) |
ETTm1 · MSE (lower is better)
8 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0.345 | Chimera: Effectively Modeling Multivariate Time Series with 2-Dimensional State Space Models (NeurIPS 2024 · 10 citations) |
Traffic · MSE (lower is better)
8 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0.403 | Chimera: Effectively Modeling Multivariate Time Series with 2-Dimensional State Space Models (NeurIPS 2024 · 10 citations) |
ETTh2 · MSE (lower is better)
7 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0.318 | Chimera: Effectively Modeling Multivariate Time Series with 2-Dimensional State Space Models (NeurIPS 2024 · 10 citations) |
ETTm2 · MSE (lower is better)
7 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0.25 | Chimera: Effectively Modeling Multivariate Time Series with 2-Dimensional State Space Models (NeurIPS 2024 · 10 citations) |
What has been tried
52 papers here report 572 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 7%, hurt 20%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add Future-Multipredictor-Mixing to combine predictions from multiple scales | M4 weighted average · OWA | 0.925 -> 0.840 | Helped | TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting (ICLR 2024 · 93 citations) |
| Add wavelet decomposition | ETT datasets · MSE | 0.388 -> 0.379; 0.311 -> 0.308; 0.339 -> 0.336; 0.247 -> 0.245 | Helped | WPMixer: Efficient Multi-Resolution Mixing for Long-Term Time Series Forecasting (AAAI 2025 · 77 citations) |
| Add long-term dependency capturing MLP | ETTm2, prediction length 96 · MSE | 0.165 -> 0.162 | Helped | HDMixer: Hierarchical Dependency with Extendable Patch for Multivariate Time Series Forecasting (AAAI 2024 · 72 citations) |
| remove memory mixing | Weather, horizon 96, look-back 336 · MSE | 0.149 -> 0.151 | Hurt | Unlocking the Power of LSTM for Long Term Time Series Forecasting (AAAI 2025 · 71 citations) |
| Vary the cluster ratio (number of clusters divided by number of channels) | ETTh1 and ETTm1 (horizon 96) · MSE | Mixed | From Similarity to Superiority: Channel Clustering for Time Series Forecasting (NeurIPS 2024 · 37 citations) | |
| add RCF to Linear | Electricity, H=96 · MSE | 0.197 -> 0.141 | Helped | CycleNet: Enhancing Time Series Forecasting through Modeling Periodic Patterns (NeurIPS 2024 · 34 citations) |
| Decrease p for Kruskal-Wallis sampling (epsilon=10) | Ross and SFC extreme points · RMSE_far | Ross: 28.7 -> 23.5; SFC: 129.0 -> 118.1 | Helped | Learning from Polar Representation: An Extreme-Adaptive Model for Long-Term Time Series Forecasting (AAAI 2024 · 18 citations) |
| add the Sparse technique | ETTh1, horizon 96 · MSE | 0.371 -> 0.359 | Helped | SparseTSF: Modeling Long-term Time Series Forecasting with *1k* Parameters (ICML 2024 · 10 citations) |
| use unidirectional recurrence along the variate dimension | ETTh1 · MSE | 0.405 -> 0.409 | Hurt | Chimera: Effectively Modeling Multivariate Time Series with 2-Dimensional State Space Models (NeurIPS 2024 · 10 citations) |
| Use both long-term and short-term branches instead of only the long-term branch | Long-range forecasting; five datasets · MSE | 0.2710 -> 0.2484 | Helped | PGN: The RNN's New Successor is Effective for Long-Range Time Series Forecasting (NeurIPS 2024 · 10 citations) |
| add phase space reconstruction (PSR) | ETTh2 (AVG) · MSE | 0.387 -> 0.376 | Helped | Attractor Memory for Long-Term Time Series Forecasting: A Chaos Perspective (NeurIPS 2024 · 7 citations) |
| Add the Initial, Local, and Global training phases instead of removing one or more phases | ETTh1, ETTh2, ETTm1, ETTm2, Weather · MSE | Helped | WaveletMixer: A Multi-Resolution Wavelets Based MLP-Mixer for Multivariate Long-Term Time Series Forecasting (AAAI 2025 · 5 citations) |
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 |
|---|---|---|---|---|
| PatchTST | Traffic | MSE | 0.394 – 0.555 | 2 papers, e.g. ICML 2025 and AAAI 2025 |
| PatchTST | Weather | MSE | 0.223 – 0.259 | 2 papers, e.g. ICML 2025 and AAAI 2025 |
| TimeMixer | Weather (Avg) | MSE | 0.222 – 0.245 | 2 papers, e.g. AAAI 2025 and ICLR 2025 |
| TimeMixer | ETTm1 (Avg) | MSE | 0.348 – 0.382 | 2 papers, e.g. AAAI 2025 and ICLR 2025 |
| TimeMixer | ETTm2 (Avg) | MSE | 0.256 – 0.277 | 2 papers, e.g. AAAI 2025 and ICLR 2025 |
| iTransformer | Electricity | MSE | 0.166 – 0.178 | 2 papers, e.g. ICML 2025 and NeurIPS 2024 |
What reviewers will expect
Datasets and baselines named in the 35 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ETTm1 · 21 of 35 papers
- Weather · 21 of 35 papers
- ETTh2 · 20 of 35 papers
- ETTh1 · 19 of 35 papers
- ETTm2 · 19 of 35 papers
- Traffic · 17 of 35 papers
- Electricity · 15 of 35 papers
- Exchange · 7 of 35 papers
Baselines
- PatchTST · 19 of 35 papers
- DLinear · 19 of 35 papers
- iTransformer · 16 of 35 papers
- TimesNet · 16 of 35 papers
- TimeMixer · 9 of 35 papers
- FEDformer · 9 of 35 papers
- Autoformer · 8 of 35 papers
- Crossformer · 8 of 35 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 80 papers; from OpenAlex and the papers' affiliations).
- University of Science and Technology of China · 5 papers
- Tsinghua University · 5 papers
- Hong Kong University of Science and Technology · 4 papers
- Huawei · 3 papers
- Aalborg University · 3 papers
- IBM · 2 papers
- Mohamed bin Zayed University of Artificial Intelligence · 2 papers
- Shenzhen University · 2 papers
- Korea Advanced Institute of Science and Technology · 2 papers
- South China University of Technology · 2 papers
Use it
Download BibTeX of 32 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
- Transformers for long-term forecasting (25 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).