New Method Prevents "Text Collapse" in Multimodal Time Series Forecasting

Multimodal time series forecasting aims to enhance predictions by combining numerical data with relevant textual reports. However, a significant issue termed "text collapse" has been identified, where the text processing component fails to contribute meaningful information, essentially becoming content-independent. This problem arises because the numerical data often strongly correlates with the output, making the numerical model inherently dominant and underutilizing the complementary textual information. To counteract this, a novel approach called REST-TS (Residual-Exclusive Supervision for Text in Time Series) has been developed. This method leverages the inherent asymmetry by allowing the numerical model to make its primary forecast. The text branch is then exclusively trained to predict the structured components of the residual, which represents the prediction errors the numerical model cannot explain. By forcing the text branch to focus on these unexplained gaps, REST-TS ensures it extracts genuine, discriminative content from the input text. Evaluations across various real-world scenarios and model architectures demonstrate that REST-TS achieves state-of-the-art performance and significantly increases the utilization of textual data compared to existing frameworks.