Research map / Time series / Foundation models for time series
GPT4MTS: Prompt-based Large Language Model for Multimodal Time-series Forecasting
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.
Proposes GPT4MTS, a prompt-based GPT-2 framework that feeds LLM-summarized news text as soft prompts alongside patched time series, and releases a GDELT-based multimodal forecasting dataset.
Problem
Time series forecasting usually uses only numerical history, ignoring contextual textual information, and there are few multimodal time series datasets or effective ways to fuse text with numerical series.
Approach
Collect textual summaries per event type/region via an LLM-based pipeline, embed them with BERT as trainable soft prompts prepended to patched, RevIN-normalized time series inputs fed into a frozen GPT-2, and use only the time-series hidden states for a linear forecasting head.
Key results
- GPT4MTS achieves average MSE 0.347 vs 0.362 for GPT4TS (about 4.14% reduction) and average MAE 0.394 vs 0.398 across 10 event types.
- Using only time-series hidden states for the output layer outperforms using full hidden states (full selection MSE 0.360), which the authors attribute to textual representations polluting the forecast.
- Linear models (DLinear, NLinear) outperform Transformer-based forecasters (Informer, Autoformer, Transformer) on this dataset.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add textual information as prompts to time-series forecasting | GPT4TS · GDELT, Average across 10 event types | MSE: 0.362 -> 0.347 | Helped |
| add textual information as prompts to time-series forecasting | GPT4TS · GDELT, Average across 10 event types | MAE: 0.398 -> 0.394 | Helped |
| use full hidden representations from textual and time-series inputs instead of only time-series hidden representations (ablation) | GPT4MTS · GDELT, Average across 10 event types | MSE: 0.347 -> 0.360 | Hurt |
| use full hidden representations from textual and time-series inputs instead of only time-series hidden representations (ablation) | GPT4MTS · GDELT, Average across 10 event types | MAE: 0.394 -> 0.402 | Hurt |
Surprising findings
- Including full hidden states from textual inputs in the output layer degrades performance relative to GPT4TS, so textual inputs work best only as prompts.
- Linear models outperform several Transformer-based models on the GDELT dataset.
Limitations
- Textual summaries for some regions were produced by T5 instead of ChatGPT-3.5 due to budget, leading to some mal-summarized texts.
- Over-reliance on textual information can harm forecasting performance, per the ablation.
- Evaluated on a single news-based dataset with three numerical variables.
Experimental setup
Datasets: GDELT
Benchmarks: GDELT-based multimodal time series dataset
Metrics: MSE MAE
Baselines: FEDformer Autoformer Informer PatchTST DLinear NLinear Transformer LLaMA GPT4TS
Builds on: GPT4TS PatchTST RevIN BERT
Closest papers by meaning
- ChatTime: A Unified Multimodal Time Series Foundation Model Bridging Numerical and Textual Data (AAAI 2025 · similarity 0.91)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.90)
- TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment (AAAI 2025 · similarity 0.90)
- AutoTimes: Autoregressive Time Series Forecasters via Large Language Models (NeurIPS 2024 · similarity 0.88)
- TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting (ICLR 2024 · similarity 0.88)
- $S^2$IP-LLM: Semantic Space Informed Prompt Learning with LLM for Time Series Forecasting (ICML 2024 · similarity 0.88)
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.