LLM-based Online Time Series Forecasting with Frequency-driven Pattern Recognition

ICLR 2026 Conference Submission16051 Authors

Keywords: Time Series Forecasing, Online Learning, Continual Learning

Abstract:

Online Time Series Forecasting (OTSF) task has been consistently studied due to its practicality in multiple domains. Considering the sequential and evolving nature of time series, OTSF models must be robust to distribution shifts and possess long-term adaptability for practical scenarios. However, existing research falls short due to lack of explicit handling of time series patterns and limitations of memory buffer-based retrieval strategies. In this paper, we propose a novel LLM-based online time series forecaster, called LLM4OT, which excels not only in continuous distribution shifts, but also in extended online scenarios. Our main idea can be summarized in two points: (1) By representing time series as a combination of frequency bases, and encoding the knowledge of each basis into prompts that guide the data distribution, our model can effectively adapt to unobserved patterns. (2) By collaboratively employing pretrained LLM with time series backbone, we enhance the model’s adaptation to data-scarce online scenarios. Additionally, we provide text-based descriptions that the LLM can easily understand, enriching the sparse data and maximizing the LLM’s adapting ability without requiring training. Our extensive
experiments on various real-world datasets demonstrate superiority and practicality of LLM4OT in various scenarios, including cross-dataset scenarios that maximize distribution shifts and scenarios with an extended online phase. Our code is available at https://anonymous.4open.science/r/LLM4OTSF-38FE/.

Primary Area: learning on time series and dynamical systems

Submission Number: 16051