Time Series Forecasting via Large Language Model-Based Textual Reasoning

• 1、School of Computer Science, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：Time series forecasting has been widely applied across various domains. In recent years, with the advancement of multimodal technologies, research on integrating textual information into time series forecasting tasks through large language models (LLMs) has progressed rapidly. However, leveraging external textual data to assist forecasting still faces significant challenges: textual data and time series often lack direct correlation, and existing modality fusion approaches struggle to effectively exploit intrinsic characteristics of time series, such as periodicity. To address these issues, this study proposes a novel model that more effectively utilizes external text to enhance time series forecasting. Unlike existing methods that either treat LLMs as the backbone of the forecasting model or merely use them to summarize textual inputs, our approach harnesses the knowledge and reasoning capabilities of LLMs to generate prediction-oriented textual descriptions highly relevant to the target time series, thereby providing direct guidance for forecasting trend changes. Additionally, we introduce a low-complexity, multi-scale modality fusion mechanism that captures correlations between time series and text across different frequency-domain levels, improving forecasting accuracy while maintaining computational efficiency. Extensive experiments on real-world datasets demonstrate that the proposed model significantly enhances time series forecasting performance.

Keywords： Computer Application Technology Time Series Forecasting Multimodal Fusion