Research map / Time series / Causal and generative time series
Robust forecasting and evaluation
Study robustness, benchmarking pitfalls and task-based optimization in time series forecasting.
91 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Causal and generative time series, keyword cluster “conformal · causal discovery · predictions”
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.
27 papers in 2023–24 and 30 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 17.4% to 13.1%.
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 · Temporal Regularized Matrix Factorization for High-dimensional Time Series Prediction (NeurIPS 2016 · 320 citations)
- 2017 · Regularization in Hierarchical Time Series Forecasting with Application to Electricity Smart Meter Data (AAAI 2017 · 25 citations)
- 2018 · Accurate Uncertainties for Deep Learning Using Calibrated Regression (ICML 2018 · 352 citations)
- 2019 · RobustSTL: A Robust Seasonal-Trend Decomposition Algorithm for Long Time Series (AAAI 2019 · 222 citations)
- 2020 · RiskOracle: A Minute-Level Citywide Traffic Accident Forecasting Framework (AAAI 2020 · 125 citations)
- 2021 · Revisiting Time Series Outlier Detection: Definitions and Benchmarks (NeurIPS 2021 · 53 citations)
- 2022 · Reinforcement Learning Based Dynamic Model Combination for Time Series Forecasting (AAAI 2022 · 58 citations)
- 2023 · WildfireSpreadTS: A dataset of multi-modal time series for wildfire spread prediction (NeurIPS 2023 · 11 citations)
- 2024 · Motion Forecasting in Continuous Driving (NeurIPS 2024 · 18 citations)
- 2025 · Cherry-Picking in Time Series Forecasting: How to Select Datasets to Make Your Model Shine (AAAI 2025 · 7 citations)
- 2026 · TianQuan-S2S: A Subseasonal-to-Seasonal Global Weather Model via Incorporate Climatology State (ICLR 2026 · 1 citation)
Earliest papers here
- Temporal Regularized Matrix Factorization for High-dimensional Time Series Prediction (NeurIPS 2016 · 320 citations)
- Data Poisoning Attacks against Autoregressive Models (AAAI 2016 · 217 citations)
- Seeing the Unseen Network: Inferring Hidden Social Ties from Respondent-Driven Sampling (AAAI 2016 · 8 citations)
What has been tried
68 papers here report 642 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 6%, hurt 21%, 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 |
|---|---|---|---|---|
| Reorganize scenes into continuous data | Argoverse 2 validation set · minFDE_1 | 4.499 -> 4.397 | Helped | Motion Forecasting in Continuous Driving (NeurIPS 2024 · 18 citations) |
| Use time series data instead of randomly sampled pairwise data | Street-view change detection test set · Accuracy | 85.99 -> 88.85 | Helped | CityPulse: Fine-Grained Assessment of Urban Change with Street View Time Series (AAAI 2024 · 8 citations) |
| add time encoding | Santos storm tide forecast, 0% missing data · IoA (SSH) | 0.94±0.10 -> 0.96±0.09 | Helped | Early Detection of Extreme Storm Tide Events Using Multimodal Data Processing (AAAI 2024 · 5 citations) |
| Add multi-task joint training instead of training separate single-task models | DroughtSet — Soil Moisture · MAE (× 10⁻³) | 19.39 -> 21.39 | Hurt | DroughtSet: Understanding Drought Through Spatial-Temporal Learning (AAAI 2025 · 4 citations) |
| add temporally contiguous strips rather than point-based masks | synthetic data sets · mask discreteness DM | Helped | CGS-Mask: Making Time Series Predictions Intuitive for All (AAAI 2024 · 1 citation) | |
| Use categorical cross-entropy instead of binary cross-entropy | Tabular discounted state occupancy estimation · estimation error | Helped | Contrastive Difference Predictive Coding (ICLR 2024 · 1 citation) | |
| Use MAE instead of MSE loss | Electricity, Missing anomaly η=0.3 · MAE | 0.249 -> 0.194 | Helped | RobustTSF: Towards Theory and Design of Robust Time Series Forecasting with Anomalies (ICLR 2024 · 1 citation) |
| Add river-network topology rather than isolating gauges | LamaH-CE Danube A · NSE | 84.12 ± 1.88 (isolated) -> 85.56 ± 1.41 (learned bidirected) | No clear effect | The Merit of River Network Topology for Neural Flood Forecasting (ICML 2024 · 1 citation) |
| Add GloFAS input to U-TAE | GFF Track 1 · F1 | 0.77 -> 0.76; 0.72 -> 0.73 | Mixed | Off to new Shores: A Dataset & Benchmark for (near-)coastal Flood Inundation Forecasting (NeurIPS 2024 · 1 citation) |
| Use B-APP loss instead of MSE or MAE | Internal factual forecasting · Precision | 5.3% higher precision than MSE/MAE | Helped | GCF: Estimating Unobserved Demand Using Graph Causal Forecasting (AAAI 2025 · 0 citations) |
| Use r = 2 instead of r = 1 | Hydraulic system data · AUC | 0.927 -> 0.933 | Mixed | Hubness Change Point Detection (AAAI 2025 · 0 citations) |
| Remove Sentinel-2 bands B01 and B09 | Europe GEDI validation · L1 (m) | 1.992 ± 0.003 -> 1.991 ± 0.002 | No clear effect | Capturing Temporal Dynamics in Large-Scale Canopy Tree Height Estimation (ICML 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 15 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ETTh1 · 2 of 15 papers
- ETTh2 · 2 of 15 papers
- ERA5 · 2 of 15 papers
- METR-LA · 2 of 15 papers
Baselines
- Transformer · 3 of 15 papers
- PatchTST · 3 of 15 papers
- DLinear · 3 of 15 papers
- ARIMA · 2 of 15 papers
- DeepAR · 2 of 15 papers
- Informer · 2 of 15 papers
- iTransformer · 2 of 15 papers
- TimesNet · 2 of 15 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 57 papers; from OpenAlex and the papers' affiliations).
- University of California, San Diego · 4 papers
- Alibaba Group · 3 papers
- Massachusetts Institute of Technology · 3 papers
- Google · 2 papers
- Stanford University · 2 papers
- The Ohio State University · 2 papers
- Amazon · 2 papers
- Carnegie Mellon University · 2 papers
- Inria · 2 papers
- University of Copenhagen · 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
- Probabilistic and conformal forecasting (121 papers)
- State-space and latent dynamics forecasting (94 papers)
- Dynamical system reconstruction (90 papers)
- Generative temporal point processes (53 papers)
Elsewhere in the topic
- Irregular series and imputation (59 papers)
- Diffusion models for time series (55 papers)
- Neural differential equations (52 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).