ZeroTS: Zero-shot time series prediction via multi-party data-model interaction
Keywords: Time-series forecasting; Retrieval Augmented Generation; Large Language Model; Zero-shot prediction
TL;DR: We propose a Zero-shot Time-Series learning framework, to bridge open-world knowledge and structured data regularity via enabling multi-party data-model interactions.
Abstract: Time series forecasting (TSF) is a fundamental task in artificial intelligence, with applications ranging from weather prediction, stock market analysis to electricity demand forecasting. While existing models, particularly large language models (LLMs) tailored for TSF, primarily focus on improving accuracy and generalization through pre-training and fine-tuning, zero-shot prediction without task-specific fine-tuning, still remains underexplored. This limitation arises from the restricted scalability and flexibility of current LLMs for TSF, which struggle to fully capture the interactions between data and model. In this work, we introduce ZeroTS, a novel approach that bridges open-world knowledge with inherent data regularities by constructing multi-party interactions between data and models. On the data side, we propose a TS-RAG (Retrieval-Augmented Generation for Time Series), which efficiently retrieves both meta and series information, enabling diverse domain-specific time series to be used as prompts. On the model side, we develop a reinforcement learning framework that treats ground-truth as environments, providing error feedback to optimize a smaller model and harnessing the capabilities of LLMs. This allows ZeroTS to incrementally approach inherent data regularities while iteratively refining its outputs. We validate ZeroTS via extensive experiments on zero-shot and long-horizon forecasting. ZeroTS achieves best or second best results with comparative parameters, 1/4 memory and 1/7 inference speed, demonstrating its efficiency and effectiveness. Our results highlight the potential of Data-LLM interactions for zero-shot learning with acceptable parameters, opening new avenues on research of this underexplored area.
Primary Area: learning on time series and dynamical systems
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 705
Loading