MM-iTransformer: A Multimodal Approach to Economic Time Series Forecasting with Textual Data
Shangyang Mou, Qiang Xue, Jinhui Chen, Tetsuya Takiguchi, Yasuo Ariki
Abstract
This paper introduces a novel multimodal framework for economic time series forecasting, integrating textual information with historical price data to enhance predictive accuracy. The proposed method employs a multi-head attention mechanism to dynamically align textual embeddings with temporal price data, capturing previously unrecognized cross-modal dependencies and enhancing the model’s ability to interpret event-driven market dynamics. This enables the framework to model complex market behaviors in a unified and effective manner. Experimental results across multiple financial datasets, including the foreign exchange (Forex) and Gold-price datasets, demonstrate that incorporating textual information significantly enhances forecasting accuracy. Compared to models relying solely on historical price data, the proposed framework achieves a substantial reduction in Mean Squared Error (MSE) loss, with improvements of up to 26.79%. This highlights the effectiveness of leveraging textual inputs alongside structured time series data in capturing complex market dynamics and improving predictive performance.