Research map / Time series / Foundation models for time series

Position: What Can Large Language Models Tell Us about Time Series Analysis

ICML 2024 · Time series › Foundation models for time series · approach: LLM-based time series forecasting · 17 citations (OpenAlex)

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.

A position paper arguing LLMs can serve as the central hub for time series analysis in three roles (enhancers, predictors, agents), supported by a literature review and a small GPT-3.5 zero-shot HAR case study.

Problem

Time series models are task-specific and rely on domain knowledge and heavy tuning; the paper asks how LLMs could move time series analysis toward general-purpose, universal intelligence.

Approach

Frames LLMs as a central hub for time series analysis, organizing prior work into three roles (data/model enhancer, predictor with tuning and non-tuning variants, and agent) and identifying open challenges, with a small zero-shot GPT-3.5 demonstration on HAR classification.

Key results

• In the zero-shot HAR demo with GPT-3.5 over four activities (ten instances per class), all Stand instances were classified correctly.
• The GPT-3.5 demo showed systematic errors: Lay instances were consistently misclassified as Sit and Stand.
• The paper reports no quantitative benchmark comparison against domain-specific time series models.

Surprising findings

• When asked to perform data augmentation, GPT-3.5 returned copies of the given instances while claiming to continue trends.
• When asked to justify misclassifications, GPT-3.5 produced plausible-sounding but unsupported explanations, indicating fabrication.

Limitations

• Empirical evidence is a small qualitative demo (ten instances per class, one LLM) rather than a systematic benchmark.
• Argument is largely a synthesis of existing literature with no new method or quantitative evaluation.

Experimental setup

Datasets: HAR

Benchmarks: HAR

Metrics: classification accuracy confusion matrix

Builds on: OFA Time-LLM LLMTime Chain-of-thought prompting

Closest papers by meaning

• Are Language Models Actually Useful for Time Series Forecasting? (NeurIPS 2024 · similarity 0.91)
• Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.91)
• Can LLMs Understand Time Series Anomalies? (ICLR 2025 · similarity 0.90)
• AutoTimes: Autoregressive Time Series Forecasters via Large Language Models (NeurIPS 2024 · similarity 0.89)
• STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · similarity 0.89)
• TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents (AAAI 2025 · similarity 0.88)