Research map / Time series / Causal and generative time series
Diffusion models for time series
Generate and forecast time series using conditional diffusion models.
55 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in Causal and generative time series, keyword cluster “differential · continuous · equations”
State History Tried 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 37 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 9.0% to 16.2%.
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
- 2020 · Modeling Continuous Stochastic Processes with Dynamic Normalizing Flows (NeurIPS 2020 · 18 citations)
- 2021 · Autoregressive Denoising Diffusion Models for Multivariate Probabilistic Time Series Forecasting (ICML 2021 · 43 citations)
- 2022 · Generative Time Series Forecasting with Diffusion, Denoise, and Disentanglement (NeurIPS 2022 · 36 citations)
- 2023 · Predict, Refine, Synthesize: Self-Guiding Diffusion Models for Probabilistic Time Series Forecasting (NeurIPS 2023 · 18 citations)
- 2024 · Latent Diffusion Transformer for Probabilistic Time Series Forecasting (AAAI 2024 · 45 citations)
- 2025 · Auto-Regressive Moving Diffusion Models for Time Series Forecasting (AAAI 2025 · 22 citations)
- 2026 · DiM-TS: Bridge the Gap Between Selective State Space Models and Time Series for Generative Modeling (AAAI 2026 · 1 citation)
Earliest papers here
- Modeling Continuous Stochastic Processes with Dynamic Normalizing Flows (NeurIPS 2020 · 18 citations)
- Autoregressive Denoising Diffusion Models for Multivariate Probabilistic Time Series Forecasting (ICML 2021 · 43 citations)
- Generative Time Series Forecasting with Diffusion, Denoise, and Disentanglement (NeurIPS 2022 · 36 citations)
What has been tried
29 papers here report 301 measured changes (a component added, removed or swapped, compared with a baseline): helped 72%, no clear effect 5%, hurt 16%, mixed 7%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add self-conditioning | Solar, Electricity, and Traffic · CRPS-sum | 0.301(.001) -> 0.253(.002); 0.024(.000) -> 0.021(.001); 0.050(.003) -> 0.040(.000) | Helped | Latent Diffusion Transformer for Probabilistic Time Series Forecasting (AAAI 2024 · 45 citations) |
| replace sliding intermediate-state generation with interpolation | seven forecasting datasets · MSE | Hurt | Auto-Regressive Moving Diffusion Models for Time Series Forecasting (AAAI 2025 · 22 citations) | |
| Add retrieval references instead of using no references | Exchange, Wind, Electricity, Weather · MSE | 0.077 -> 0.013; 1.066 -> 0.784; 0.379 -> 0.151; 0.356 -> 0.281 | Helped | Retrieval-Augmented Diffusion Models for Time Series Forecasting (NeurIPS 2024 · 15 citations) |
| add Fourier-based training loss | 24-step generation across Sines, Stocks, ETTh, MuJoCo, Energy, and fMRI · Context-FID | 0.008±.001 -> 0.006±.000 (Sines); 0.216±.031 -> 0.147±.025 (Stocks); 0.219±.021 -> 0.116±.010 (ETTh); 0.017±.002 -> 0.01 | Helped | Diffusion-TS: Interpretable Diffusion for General Time Series Generation (ICLR 2024 · 12 citations) |
| add high-frequency inflation | AR1, Stock, Energy, Temperature, and ECG block maxima · JS Divergence | 0.0006+0.0001 -> 0.0004±0.0001; 0.0898+0.0054 -> 0.0700±0.0061 | Helped | FIDE: Frequency-Inflated Conditional Diffusion Model for Extreme-Aware Time Series Generation (NeurIPS 2024 · 10 citations) |
| Add a temporal channel | Sines, Stocks, and Energy · FDDS | 0.0005, 1.8838, 0.5254 vs 0.0003, 0.0588, 0.0442 | Helped | Population Aware Diffusion for Time Series Generation (AAAI 2025 · 5 citations) |
| Add diffusion step embedding to a linear-schedule model | Solar forecasting · CRPS | 0.326 -> 0.399 | Hurt | ANT: Adaptive Noise Schedule for Time Series Diffusion Models (NeurIPS 2024 · 5 citations) |
| replace similarity-driven quantization with distance-driven quantization | ETTh · Discriminative Score | 0.003±.001 -> 0.010±.005 | Hurt | SDformer: Similarity-driven Discrete Transformer For Time Series Generation (NeurIPS 2024 · 3 citations) |
| Add TST-based completion before masked diffusion | Energy, length 24, 30% missing · discriminative score | 0.157 -> 0.048 | Helped | A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking (NeurIPS 2025 · 1 citation) |
| Use autocorrelated rather than uncorrelated noise across forecast times | t850 hourly forecasts · temporal difference | Helped | Continuous Ensemble Weather Forecasting with Diffusion models (ICLR 2025 · 1 citation) | |
| smooth the ECG spectra with increasingly wide Gaussian kernels, reducing frequency localization | ECG variants · sliced Wasserstein distance | around at \sigma \approx 7 | Mixed | Time Series Diffusion in the Frequency Domain (ICML 2024 · 1 citation) |
| Add the LSNM assumption (variable conditional variance) | ETTh1 · QICE | 2.821 -> 1.470 | Helped | Non-stationary Diffusion For Probabilistic Time Series Forecasting (ICML 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 23 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 · 7 of 23 papers
- Electricity · 6 of 23 papers
- Exchange · 5 of 23 papers
- ETTh1 · 5 of 23 papers
- ETTh2 · 5 of 23 papers
- ETTm1 · 5 of 23 papers
- ETTm2 · 5 of 23 papers
- Sines · 4 of 23 papers
Baselines
- CSDI · 8 of 23 papers
- Diffusion-TS · 7 of 23 papers
- TimeGrad · 5 of 23 papers
- PatchTST · 4 of 23 papers
- TimeVAE · 4 of 23 papers
- D3VAE · 3 of 23 papers
- DLinear · 3 of 23 papers
- TimeGAN · 3 of 23 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 51 papers; from OpenAlex and the papers' affiliations).
- Nanyang Technological University · 3 papers
- Tencent · 2 papers
- Peking University · 2 papers
- University of British Columbia · 2 papers
- Hong Kong University of Science and Technology · 2 papers
- Amazon · 2 papers
- Tsinghua University · 2 papers
- Harbin Institute of Technology · 2 papers
Use it
Download BibTeX of 27 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
- Irregular series and imputation (59 papers)
- Neural differential equations (52 papers)
Elsewhere in the topic
- Probabilistic and conformal forecasting (121 papers)
- State-space and latent dynamics forecasting (94 papers)
- Robust forecasting and evaluation (91 papers)
- Dynamical system reconstruction (90 papers)
- Generative temporal point processes (53 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).