UniTime: A Language-Empowered Unified Model for Cross-Domain Time Series Forecasting

Source

• Raw Markdown: paper_unitime-2023.md
• PDF: paper_unitime-2023.pdf
• Preprint: arXiv 2310.09751
• Official code: liuxu77/UniTime

Core Claim

Domain instructions can help a unified cross-domain time-series forecasting model distinguish between heterogeneous datasets and align time-series tokens with language-model representations.

Key Contributions

• Proposes UniTime for cross-domain multivariate time-series forecasting.
• Uses channel independence so the model can handle datasets with different numbers of variables.
• Patches time-series channels into tokens and projects them into a shared hidden space.
• Prepends natural-language domain instructions before time-series tokens so causal attention can use the instruction while processing the series.
• Uses masking to reduce domain convergence-speed imbalance and improve cross-domain learning.

Method Notes

UniTime is an early language-empowered time-series model. Its text is mostly domain instruction, not arbitrary natural-language reasoning over a time series.

The input-order detail matters: because GPT-style causal masking prevents later tokens from attending backward only to previous tokens, UniTime places the instruction before the time-series tokens so the numeric sequence can attend to the text.

Evidence And Results

The paper reports state-of-the-art cross-domain forecasting performance and zero-shot transferability to unseen domains. For the wiki, its durable contribution is the interface pattern: domain text as a conditioning prefix for time-series tokens.

Alex Notes

• Alex flagged this as the first Multi-Modal LLM plus time-series model that follows instructions, albeit poorly.

Limitations

• Instruction following is narrow compared with modern time-series MLLMs.
• Channel independence improves cross-domain flexibility but can miss native multivariate channel dependence.
• It is still passive forecasting: the domain instruction is context, not an action, control input, or intervention.

Links Into The Wiki

• Time-Series Foundation Models
• Context-Aided Forecasting
• Unified Multimodal Models
• Position: What Can LLMs Tell Us about Time Series Analysis
• ChatTS

Open Questions

• How much of UniTime’s gain comes from domain text versus cross-domain masking and channel-independent design?
• Can instruction conditioning preserve calibration when the text is noisy or misleading?
• What is the minimal language interface needed for context-aided forecasting?