Research map / Time series / Foundation models for time series
LLM-based time series forecasting
Repurpose language models for forecasting through prompting, retrieval and alignment.
90 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2023–2026) · an approach in Foundation models for time series, keyword cluster “shot · series foundation · llm”
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.
20 papers in 2023–24 and 70 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 18.7% to 27.6%.
Open questions the data points to
- Papers report PatchTST on ETTh2 (MSE) anywhere from 0.33 to 0.366: 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
- 2023 · One Fits All: Power General Time Series Analysis by Pretrained LM (NeurIPS 2023 · 156 citations)
- 2024 · GPT4MTS: Prompt-based Large Language Model for Multimodal Time-series Forecasting (AAAI 2024 · 97 citations)
- 2025 · TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment (AAAI 2025 · 96 citations)
- 2026 · OccamVTS: Distilling Vision Models to 1% Parameters for Time Series Forecasting (AAAI 2026 · 4 citations)
Earliest papers here
- One Fits All: Power General Time Series Analysis by Pretrained LM (NeurIPS 2023 · 156 citations)
- Large Language Models Are Zero-Shot Time Series Forecasters (NeurIPS 2023 · 96 citations)
- GPT4MTS: Prompt-based Large Language Model for Multimodal Time-series Forecasting (AAAI 2024 · 97 citations)
What has been tried
45 papers here report 467 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 6%, hurt 21%, 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 |
|---|---|---|---|---|
| use full hidden representations from textual and time-series inputs instead of only time-series hidden representations | GDELT, Average across 10 event types · MSE | 0.347 -> 0.360 | Hurt | GPT4MTS: Prompt-based Large Language Model for Multimodal Time-series Forecasting (AAAI 2024 · 97 citations) |
| Replace cross-modality alignment with concatenation | Eight forecasting datasets · MSE | Hurt | TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment (AAAI 2025 · 96 citations) | |
| Add feature regularization loss | ETTh1 · MSE | 0.446 -> 0.434 | Helped | CALF: Aligning LLMs for Time Series Forecasting via Cross-modal Fine-Tuning (AAAI 2025 · 66 citations) |
| Remove the LLM from Time-LLM | 13 datasets, averaged prediction lengths · forecasting wins | 0 -> 12 wins | Helped | Are Language Models Actually Useful for Time Series Forecasting? (NeurIPS 2024 · 53 citations) |
| increase training FLOPS across model sizes | Monash · Scaled MAE (Geometric Mean) | Helped | A decoder-only foundation model for time-series forecasting (ICML 2024 · 44 citations) | |
| Add diverse resolution sampling to Monash pre-training | Zero-shot, Monash pre-training · forecast performance | 37% | Helped | Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series (NeurIPS 2024 · 40 citations) |
| Remove prompt design | Zero-shot forecasting across reported datasets · MSE and MAE | Hurt | TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting (ICLR 2024 · 36 citations) | |
| replace the LLM with direct input tokens | ETTh1, Pred-96 · MSE | 0.360 -> 0.365 | Hurt | AutoTimes: Autoregressive Time Series Forecasters via Large Language Models (NeurIPS 2024 · 34 citations) |
| allow any frequency to use any patch size during training | Monash benchmark · normalized MAE | 0.655 -> 0.720 | Hurt | Unified Training of Universal Time Series Forecasting Transformers (ICML 2024 · 33 citations) |
| Choose text prototypes that better represent the full LLM text embedding space | UCR and UEA classification · accuracy | Helped | TEST: Text Prototype Aligned Embedding to Activate LLM's Ability for Time Series (ICLR 2024 · 20 citations) | |
| Add LLM contextualization before event prediction instead of directly prompting the LLM with the time series | New York, San Francisco, Houston, S&P 500, Nikkei 225, Mortality, Test-Positive · F1 | 0.499 -> 0.625; 0.510 -> 0.603; 0.412 -> 0.544; 0.276 -> 0.330; 0.333 -> 0.364; 0.695 -> 0.842; 0.727 -> 0.949 | Helped | TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents (AAAI 2025 · 19 citations) |
| Add agent-filtered news to the textual prompt instead of using no news | Electricity · RMSE | 336.41 -> 280.39 | Helped | From News to Forecast: Integrating Event Analysis in LLM-Based Time Series Forecasting with Reflection (NeurIPS 2024 · 13 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 | ETTh2 | MSE | 0.33 – 0.366 | 2 papers, e.g. ICLR 2024 and AAAI 2025 |
| PatchTST | ETTm1 | MSE | 0.351 – 0.381 | 2 papers, e.g. ICLR 2024 and AAAI 2025 |
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
- ETTh1 · 15 of 35 papers
- ETTh2 · 11 of 35 papers
- Weather · 11 of 35 papers
- ETTm1 · 10 of 35 papers
- ETTm2 · 10 of 35 papers
- Traffic · 10 of 35 papers
- Electricity · 7 of 35 papers
- ECL · 5 of 35 papers
Baselines
- PatchTST · 14 of 35 papers
- DLinear · 11 of 35 papers
- Time-LLM · 11 of 35 papers
- TimesNet · 11 of 35 papers
- iTransformer · 10 of 35 papers
- GPT4TS · 8 of 35 papers
- N-BEATS · 5 of 35 papers
- FEDformer · 5 of 35 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 90 papers; from OpenAlex and the papers' affiliations).
- Hong Kong University of Science and Technology · 4 papers
- Amazon · 4 papers
- University of Southern California · 3 papers
- Tongji University · 3 papers
- Peking University · 3 papers
- Griffith University · 3 papers
- IBM · 3 papers
- Google · 3 papers
- Tsinghua University · 3 papers
- Ant Group · 2 papers
Use it
Download BibTeX of 22 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
- Time series foundation models (71 papers)
- Time-series language reasoning (44 papers)
Elsewhere in the topic
- Time series representation learning (115 papers)
- Time series classification (61 papers)
- Satellite image time series (28 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).