Research map / Time series / Foundation models for time series
TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces TimeCAP, which uses two frozen GPT-4 agents (one contextualizes time series into text summaries, one predicts events from them) plus a trainable BERT-based multi-modal encoder that augments inputs and selects in-context examples for time series event prediction.
Problem
Time series event prediction often ignores contextual information that LLMs could supply; existing LLM-based time series methods use LLMs only as direct predictors on raw or embedded series.
Approach
Use LLMs as contextualizers that turn time series into text summaries, then feed those summaries both to a second LLM predictor and to a trainable multi-modal encoder, with mutual augmentation: summaries augment encoder inputs, and encoder-retrieved summaries serve as in-context examples for the predictor.
Key results
- TimeCAP achieves the best average rank (1.00 in ablation, 1.86 overall rank) and best F1 on most datasets, e.g., 0.676 F1 on New York vs 0.625 for the best baseline.
- The paper reports an average 28.75% F1 improvement, with up to 157% gain over state-of-the-art methods.
- TimeCP alone (zero-shot) outperforms PromptCast on all seven datasets, e.g., 0.949 vs 0.727 F1 on Test-Positive.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LLM contextualization before event prediction instead of directly prompting the LLM with the time series (ablation) | TimeCP with GPT-4 · 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 |
| Add text summaries to the time-series input for the multimodal encoder (ablation) | TimeCAP multimodal encoder · New York, San Francisco, Houston, S&P 500, Nikkei 225, Mortality, Test-Positive | F1: 0.592 -> 0.623; 0.542 -> 0.576; 0.593 -> 0.606; 0.373 -> 0.398; 0.391 -> 0.428; 0.695 -> 0.734; 0.841 -> 0.937 | Helped |
| Use multimodal-encoder-selected in-context examples instead of random examples (ablation) | TimeCAP prompt augmentation · New York, San Francisco, Houston, S&P 500, Nikkei 225, Mortality, Test-Positive | F1: 0.619 -> 0.641; 0.621 -> 0.626; 0.528 -> 0.578; 0.322 -> 0.341; 0.344 -> 0.401; 0.901 -> 0.947; 0.883 -> 0.961 | Mixed |
| Use multimodal-encoder-selected in-context examples instead of time-series-only-encoder examples (ablation) | TimeCAP prompt augmentation · New York, San Francisco, Houston, S&P 500, Nikkei 225, Mortality, Test-Positive | F1: 0.625 -> 0.641; 0.599 -> 0.626; 0.571 -> 0.578; 0.305 -> 0.341; 0.379 -> 0.401; 0.948 -> 0.947; 0.948 -> 0.961 | Mixed |
| Aggregate the multimodal encoder and LLM predictions instead of selecting one prediction source (ablation) | TimeCAP prediction fusion · New York, San Francisco, Houston, S&P 500, Nikkei 225, Mortality, Test-Positive | F1: 0.641 -> 0.676; 0.626 -> 0.632; 0.606 -> 0.614; 0.398 -> 0.398; 0.428 -> 0.428; 0.947 -> 0.947; 0.961 -> 0.962 | Mixed |
Surprising findings
- Zero-shot TimeCP outperforms several fully supervised time series models on some datasets, showing strong value from LLM contextualization alone.
Limitations
- Code is only available upon request, and the paper frames the problem only as multi-class classification rather than regression.
- Relies on GPT-4 API calls, which are costly and non-reproducible as a closed model.
- Improvements over some baselines (e.g., Crossformer on Test-Positive AUC) are small or mixed across datasets.
Experimental setup
Datasets: New York hourly weather San Francisco hourly weather Houston hourly weather S&P 500 daily finance Nikkei 225 daily finance Mortality weekly healthcare Test-Positive weekly healthcare
Benchmarks: New York San Francisco Houston S&P 500 Nikkei 225 Mortality Test-Positive
Metrics: F1 AUROC
Baselines: Autoformer Crossformer TimesNet DLinear TSMixer PatchTST FreTS iTransformer LLMTime PromptCast GPT4TS Time-LLM
Builds on: PromptCast PatchTST Dynamic Time Warping BERT
Closest papers by meaning
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.91)
- TimeXL: Explainable Multi-modal Time Series Prediction with LLM-in-the-Loop (NeurIPS 2025 · similarity 0.90)
- AutoTimes: Autoregressive Time Series Forecasters via Large Language Models (NeurIPS 2024 · similarity 0.89)
- TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment (AAAI 2025 · similarity 0.89)
- Are Language Models Actually Useful for Time Series Forecasting? (NeurIPS 2024 · similarity 0.89)
- STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.