View in EDS

TS2Lang: A Co-occurrence pattern-driven translation mechanism for zero-shot time series forecasting with LLMs.

Saved in:
Bibliographic Details
Title: TS2Lang: A Co-occurrence pattern-driven translation mechanism for zero-shot time series forecasting with LLMs.
Authors: Xun, Haoran1 (AUTHOR) xhaoran@bupt.edu.cn, Zhou, Wen'an1 (AUTHOR) zhouwa@bupt.edu.cn, Tao, Liwen1 (AUTHOR) taoliwen@bupt.edu.cn, Zhong, Yingyi1 (AUTHOR) zyy2018@bupt.edu.cn
Source: Journal of Intelligent Information Systems. Apr2026, Vol. 64 Issue 2, p597-620. 24p.
Subjects: Sequential pattern mining, Vector quantization, Machine learning, Forecasting, Software frameworks, Language models, Autoregressive models
Abstract: With the rapid advancement of large language models (LLMs) in natural language processing and multimodal tasks, their potential in time series forecasting has attracted increasing attention. However, the representational differences between time series and text limit the effectiveness of transferring LLMs to temporal tasks. Existing approaches attempt to bridge this gap by learning discrete representations of time series, but these representations constitute a "foreign language" to LLMs, requiring additional continual pre-training and thus struggling to adapt to low-resource settings. To address this, we propose TS2Lang, a plug-and-play framework for time series forecasting with LLMs. TS2Lang first learns a general discrete representation of time series based on VQ-VAE and incorporates frequent sequence mining (FSM) to extract high-frequency patterns, effectively "translating" the time series into a form directly interpretable by LLMs. The resulting token sequences are then fed into a pre-trained LLM for autoregressive prediction, requiring no fine-tuning and supporting flexible replacement of the language model. In zero-shot settings, TS2Lang achieves the best average performance using far fewer trainable parameters and data than the strongest baseline, TimesFM, with MAE and MSE improved by up to approximately 13.4% and 7.9%, respectively. Moreover, adaptation with a small amount of target-domain data can further enhance performance, demonstrating the method's effectiveness and practical utility. [ABSTRACT FROM AUTHOR]
Copyright of Journal of Intelligent Information Systems is the property of Springer Nature and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:With the rapid advancement of large language models (LLMs) in natural language processing and multimodal tasks, their potential in time series forecasting has attracted increasing attention. However, the representational differences between time series and text limit the effectiveness of transferring LLMs to temporal tasks. Existing approaches attempt to bridge this gap by learning discrete representations of time series, but these representations constitute a "foreign language" to LLMs, requiring additional continual pre-training and thus struggling to adapt to low-resource settings. To address this, we propose TS2Lang, a plug-and-play framework for time series forecasting with LLMs. TS2Lang first learns a general discrete representation of time series based on VQ-VAE and incorporates frequent sequence mining (FSM) to extract high-frequency patterns, effectively "translating" the time series into a form directly interpretable by LLMs. The resulting token sequences are then fed into a pre-trained LLM for autoregressive prediction, requiring no fine-tuning and supporting flexible replacement of the language model. In zero-shot settings, TS2Lang achieves the best average performance using far fewer trainable parameters and data than the strongest baseline, TimesFM, with MAE and MSE improved by up to approximately 13.4% and 7.9%, respectively. Moreover, adaptation with a small amount of target-domain data can further enhance performance, demonstrating the method's effectiveness and practical utility. [ABSTRACT FROM AUTHOR]
ISSN:09259902
DOI:10.1007/s10844-025-01015-6