Research map / Time series / Foundation models for time series
ChatTime: A Unified Multimodal Time Series Foundation Model Bridging Numerical and Textual Data
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.
Treats time series as a foreign language by discretizing and tokenizing values into an expanded LLM vocabulary, then continual pre-training and LoRA instruction tuning of LLaMA-2-7B to yield ChatTime, a zero-shot multimodal time series foundation model.
Problem
Existing time series approaches either train foundation models from scratch (costly, text-incapable), or fine-tune LLMs per dataset (no zero-shot, no text output), so a multimodal time series model with zero-shot forecasting and bimodal text/time-series input-output is lacking.
Approach
Model time series as a foreign language: min-max normalize, bin values into 10K discrete centers, wrap each value with "###" marks so each value is one token in an extended LLM vocabulary; then continual pre-training and instruction fine-tuning of an existing LLM with LoRA, without changing architecture.
Key results
- On zero-shot forecasting over eight benchmarks, ChatTime reaches average MAE 0.2515, matching Chronos (0.2512) with only 4% of pre-training data.
- On context-guided forecasting, ChatTime achieves the best average MAE (0.5884) and rank (2.58) versus dataset-specific and shared baselines, beating TGForecaster (0.5955).
- On time series QA, ChatTime achieves 0.7605 average accuracy versus 0.5567 for GPT-4 and ranks first in all feature groups.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ChatTime instead of Chronos | Zero-shot forecasting · ZSTSF (8 datasets) | Avg. MAE: 0.2512 -> 0.2515 | Hurt |
| Use ChatTime instead of Chronos | Zero-shot forecasting · ZSTSF (8 datasets) | Avg. Rank: 4.8438 -> 4.4688 | Helped |
| Add textual context to the forecast input | ChatTime · CGTSF (MSPG, LEU, PTF) | Avg. MAE: 0.6160 -> 0.5884 | Helped |
| Use ChatTime instead of GPT4 | Time series question answering · TSQA (4 features) | Avg. Acc: 0.5567 -> 0.7605 | Helped |
| Remove autoregressive continuous pre-training (ablation) | ChatTime · ZSTSF | forecasting performance | Hurt |
| Replace autoregressive continuous pre-training with instruction fine-tuning data (ablation) | ChatTime · CGTSF and TSQA | performance | Helped |
| Replace clustered high-quality time-series slices with randomly sampled slices (ablation) | ChatTime · ZSTSF, CGTSF, and TSQA | task performance | Hurt |
| Remove text question answering data from instruction fine-tuning (ablation) | ChatTime · ZSTSF, CGTSF, and TSQA | task performance | Hurt |
Surprising findings
- Simple linear model DLinear performs comparably to complex SOTA models like GPT4TS, suggesting unimodal methods may be saturating.
- Adding LLMs to GPT4TS and TimeLLM brings only minor gains over DLinear.
Limitations
- Zero-shot advantage over Chronos is not consistent; ChatTime underperforms GPT4TS and DLinear on several unimodal datasets.
- Context-guided and TSQA datasets are small and self-collected or synthetic; instruction tuning for CGTSF uses training splits of the same datasets, deviating from zero-shot.
- Discretization into 10K bins is lossy and limited to the fixed vocabulary; ChatTime has not reached saturation given resource constraints.
Experimental setup
Datasets: Monash TFB Alpaca KernelSynth-generated TSQA MSPG LEU PTF ETTh1 ETTh2 ETTm1 ETTm2 Electric
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Electric Exchange Traffic Weather MSPG LEU PTF Synthetic TSQA
Metrics: MAE average rank accuracy
Baselines: DLinear iTransformer GPT4TS TimeLLM TimeGPT Moirai TimesFM Chronos TGForecaster GPT-4 GPT-3.5 GLM4
Builds on: LLMTIME LLaMA-2 KernelSynth Unsloth
Closest papers by meaning
- GPT4MTS: Prompt-based Large Language Model for Multimodal Time-series Forecasting (AAAI 2024 · similarity 0.91)
- A decoder-only foundation model for time-series forecasting (ICML 2024 · similarity 0.91)
- Empowering Time Series Analysis with Large-Scale Multimodal Pretraining (NeurIPS 2026 · similarity 0.90)
- Time-VLM: Exploring Multimodal Vision-Language Models for Augmented Time Series Forecasting (ICML 2025 · similarity 0.90)
- TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment (AAAI 2025 · similarity 0.89)
- AutoTimes: Autoregressive Time Series Forecasters via Large Language Models (NeurIPS 2024 · 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.