Research map / Time series / Long-term forecasting
Spatiotemporal forecasting models
Model time series with spatiotemporal graphs, convolutions, state-space and diffusion pretraining.
77 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in Long-term forecasting, keyword cluster “graph · multivariate time · spatial”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Few recent reports None of the tracked benchmarks has a report from 2025–26.
18 papers in 2023–24 and 41 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 17.5% to 16.8%.
No longer reported by papers here: PeMSD4 (MAE, last in 2023); PeMSD8 (MAE, last in 2023).
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
- 2018 · Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting (ICLR 2018 · 509 citations)
- 2019 · Memory in Memory: A Predictive Neural Network for Learning Higher-Order Non-Stationarity From Spatiotemporal Dynamics (CVPR 2019 · 481 citations)
- 2020 · Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting (NeurIPS 2020 · 714 citations)
- 2021 · Inductive Graph Neural Networks for Spatiotemporal Kriging (AAAI 2021 · 166 citations)
- 2022 · SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction (NeurIPS 2022 · 246 citations)
- 2023 · Spatio-Temporal Meta-Graph Learning for Traffic Forecasting (AAAI 2023 · 311 citations)
- 2024 · Efficient and Effective Time-Series Forecasting with Spiking Neural Networks (ICML 2024 · 5 citations)
- 2025 · RhythmMamba: Fast, Lightweight, and Accurate Remote Physiological Measurement (AAAI 2025 · 18 citations)
- 2026 · MedSpaformer: A Transferable Transformer with Multi-Granularity Token Sparsification for Medical Time Series Classification (AAAI 2026 · 2 citations)
Earliest papers here
- Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting (ICLR 2018 · 509 citations)
- Multivariate Convolutional Sparse Coding for Electromagnetic Brain Signals (NeurIPS 2018 · 35 citations)
- Memory in Memory: A Predictive Neural Network for Learning Higher-Order Non-Stationarity From Spatiotemporal Dynamics (CVPR 2019 · 481 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.
PeMSD4 · MAE (lower is better)
4 papers, 2020–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 19.83 | Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting (NeurIPS 2020 · 714 citations) |
| 2021 | 19.5 | Z-GCNETs: Time Zigzags at Graph Convolutional Networks for Time Series Forecasting (ICML 2021 · 46 citations) |
| 2022 | 17.58 | Time-Conditioned Dances with Simplicial Complexes: Zigzag Filtration Curve based Supra-Hodge Convolution Networks for Time-series Forecasting (NeurIPS 2022 · 2 citations) |
PeMSD8 · MAE (lower is better)
4 papers, 2020–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 15.95 | Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting (NeurIPS 2020 · 714 citations) |
| 2021 | 15.76 | Z-GCNETs: Time Zigzags at Graph Convolutional Networks for Time Series Forecasting (ICML 2021 · 46 citations) |
| 2022 | 13.86 | Time-Conditioned Dances with Simplicial Complexes: Zigzag Filtration Curve based Supra-Hodge Convolution Networks for Time-series Forecasting (NeurIPS 2022 · 2 citations) |
What has been tried
42 papers here report 463 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 7%, hurt 19%, 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 diff-fusion | MMPD · MAE | 5.71 -> 3.16 | Helped | RhythmMamba: Fast, Lightweight, and Accurate Remote Physiological Measurement (AAAI 2025 · 18 citations) |
| Add motif-induced matrices | Air Quality, MCAR · MAE | 17.49 ± 0.32 -> 15.43 ± 0.32 | Helped | Motif-aware Graph Neural Networks for Networked Time Series Imputation (AAAI 2025 · 5 citations) |
| Use a convolutional spike encoder instead of a delta spike encoder | Metr-la and Electricity · R^2 | +0.09 on average | Helped | Efficient and Effective Time-Series Forecasting with Spiking Neural Networks (ICML 2024 · 5 citations) |
| Add relational message passing | LA, 15/30/60 min · MAE | 2.80 { ± .01} -> 2.68 { ± .01}; 3.14 { ± .01} -> 3.02 { ± .01}; 3.47 { ± .02} -> 3.40 { ± .01} | Helped | Graph-based Time Series Clustering for End-to-End Hierarchical Forecasting (ICML 2024 · 4 citations) |
| replace Chazelle MST with Kruskal MST | ImageNet-1K Tiny · top-1 accuracy | 83.6 -> 83.5 | No clear effect | GG-SSMs: Graph-Generating State Space Models (CVPR 2025 · 3 citations) |
| Add TMP topological features to spatial graph convolution features | PeMSD4, first 1,000 networks · RMSE | 153.75±0.22 -> {\bf 147.57}±{\bf 0.23} | Helped | Time-Aware Knowledge Representations of Dynamic Objects with Multidimensional Persistence (AAAI 2024 · 3 citations) |
| Add the autoregressive mechanism | ETTm2 · MSE | 0.281 -> 0.257 | Helped | TimeDART: A Diffusion Autoregressive Transformer for Self-Supervised Time Series Representation (ICML 2025 · 2 citations) |
| Use a non-learnable numeric timestamp feature or decomposed calendar features instead of the proposed temporal embedding | TabReD · average performance improvement | Num: -0.05\%; Time: -0.39\% | Hurt | Understanding the Limits of Deep Tabular Methods with Temporal Shift (ICML 2025 · 0 citations) |
| change the ACN channel-similarity metric among cosine, ℓ1, and ℓ2 | 8 datasets · MSE | .309, .309, .308 | No clear effect | Channel Normalization for Time Series Channel Identification (ICML 2025 · 0 citations) |
| Use local rather than global merging at the fastest setting | Dummy Mouse Enhancers Ensembl · accuracy | 69.4% -> 74.0% | Helped | Efficient Time Series Processing for Transformers and State-Space Models through Token Merging (ICML 2025 · 0 citations) |
| add Correlation Mixing | Electricity, Traffic, Weather, ETTh1, ETTh2, ETTm1, ETTm2 (horizon 96) · MSE | 0.137 -> 0.129; 0.385 -> 0.367; 0.171 -> 0.144; 0.362 -> 0.361; 0.276 -> 0.274; 0.308 -> 0.292; 0.168 -> 0.167 | Helped | CMoS: Rethinking Time Series Prediction Through the Lens of Chunk-wise Spatial Correlations (ICML 2025 · 0 citations) |
| Add inter-pool to TVNet | Weather, Electricity, Traffic; prediction lengths 96, 192, 336, 720 · MSE | Weather 0.156 -> 0.147; Electricity 0.147 -> 0.142; Traffic 0.377 -> 0.367 | Helped | TVNet: A Novel Time Series Analysis Method Based on Dynamic Convolution and 3D-Variation (ICLR 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 18 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Weather · 8 of 18 papers
- Electricity · 7 of 18 papers
- Traffic · 6 of 18 papers
- ETTh1 · 6 of 18 papers
- Exchange · 5 of 18 papers
- ETTh2 · 5 of 18 papers
- ETTm1 · 5 of 18 papers
- ETTm2 · 4 of 18 papers
Baselines
- iTransformer · 5 of 18 papers
- TimesNet · 4 of 18 papers
- PatchTST · 4 of 18 papers
- FEDformer · 3 of 18 papers
- DLinear · 3 of 18 papers
- S-Mamba · 2 of 18 papers
- RMLP · 2 of 18 papers
- Transformer · 2 of 18 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 59 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 3 papers
- Politecnico di Milano · 3 papers
- Imperial College London · 3 papers
- University of Science and Technology of China · 2 papers
- Hong Kong University of Science and Technology · 2 papers
- Fudan University · 2 papers
- Universita della Svizzera Italiana · 2 papers
- Università della Svizzera Italiana · 2 papers
- Nanjing University · 2 papers
- Tencent · 2 papers
Use it
Download BibTeX of 39 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
- Attention for time series forecasting (123 papers)
- Graph-based multivariate forecasting (87 papers)
Elsewhere in the topic
- Lightweight long-term forecasters (94 papers)
- Transformers for long-term forecasting (25 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).