MoAT: Multi-Modal Augmented Time Series Forecasting

TL;DR — Multi-modal augmentation with text data is effective for time series forecasting.

Abstract

Time series forecasting plays a pivotal role in various domains, facilitating optimized resource allocation and strategic decision-making. However, the scarcity of training samples often hinders the accuracy of the forecasting task. To address this, we explore the potential of leveraging information from different modalities that are commonly associated with time series data. In this paper, we introduce MoAT, a novel multi-modal augmented time series forecasting approach that strategically integrates both feature-wise and sample-wise augmentation methods to enrich multi-modal representation learning. It further enhances prediction accuracy through joint trend-seasonal decomposition across all modalities and fuses the information for the final prediction. Extensive experiments show that MoAT outperforms state-of-the-art methods, resulting in a substantial reduction in mean squared error ranging from 6.5% to 71.7%, which demonstrates the effectiveness and robustness in addressing the limitations imposed by data scarcity. The datasets and code are available at https://anonymous.4open.science/r/MoAT-201E.