Position: What Can Large Language Models Tell Us about Time Series Analysis
Source
- Raw Markdown: paper_llms-time-series-analysis-2024.md
- PDF: paper_llms-time-series-analysis-2024.pdf
- Preprint: arXiv 2402.02713
Core Claim
The paper argues that LLMs can expand time-series analysis beyond numeric prediction into modality switching, question answering, reasoning, and natural-language interaction. For this wiki, it is a position source for why time-series models may need language interfaces, not only better forecast heads.
Key Contributions
- Maps time-series analysis tasks onto LLM-era capabilities such as prompting, agents, tool use, and multimodal alignment.
- Reviews early approaches that adapt LLMs to time series through prompting, reprogramming, soft prompts, or lightweight adapters.
- Frames time-series question answering and modality switching as important future interfaces.
- Emphasizes trust, interpretability, and the practical integration of LLM technologies with time-series analysis.
Method Notes
This is not a new forecasting architecture. It is a position and survey-like paper, useful for organizing the space of LLM-assisted time-series analysis.
The paper should be read alongside context-aided forecasting work. A general LLM interface is broader than forecasting with essential textual context, but it also risks over-claiming if numeric calibration is not preserved.
Evidence And Results
The source is mostly conceptual. Its value is in synthesis and roadmap framing rather than a benchmark result. It names the main interface patterns: direct prompting over serialized values, LLM backbone adaptation, time-series encoders aligned to language space, and agent/tool pipelines.
Limitations
- Position paper rather than a controlled model comparison.
- Many LLM-for-time-series systems remain brittle, expensive, or weakly calibrated.
- Natural-language interaction does not automatically make a model action-conditioned or causal.
Foundation TSFM Relevance
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Context interface | adjacent | The paper frames language, prompts, tools, modality switching, and time-series QA as interfaces around numeric sequences. | Needs a typed split between channel context, general context, and numeric state. |
| Benchmark level | warning | It argues for tasks beyond forecasting, including QA and reasoning over time series. | Many cited systems are roadmap-level or brittle, so benchmark claims need controlled numeric calibration tests. |
| Causal structure, counterfactuals, and control | insufficient evidence | Agent/tool discussion can support operational workflows. | It does not define action-conditioned time-series dynamics or causal intervention evaluation. |
Links Into The Wiki
- Time-Series Foundation Models
- Foundation Time-Series Model Research Agenda
- Context-Aided Forecasting
- Unified Multimodal Models
- Number Tokenization
- ChatTS
- TimeOmni-1
Open Questions
- Which time-series tasks need an LLM, and which only need better numeric encoders?
- How can a time-series LLM preserve probabilistic calibration while following instructions?
- Should language be used as a domain/context interface, a reasoning interface, or the primary modeling substrate?