A Multi-Modal Transformer Architecture Combining Sentiment Dynamics, Temporal Market Data, and Macroeconomic Indicators for Sturdy Stock Return Forecasting

Stock market prediction remains one of the most challenging problems due to the intrinsic complexity and evolving nature of financial markets. Most of the traditional models rely on using historical data and thus fail to capture the complex dynamics of the markets, leading to less-than-ideal predictions. This work proposes a novel multi-modal framework that combines sentiment analysis using real-time discussion forum sentiment scores with state-of-the-art graph neural network architectures to enhance stock market forecasting. Our approach uniquely fuses sentiment dynamics from social media and news sources with temporal market data and macroeconomic indicators to construct dynamic graph representations of interfirm relationships. Further, we employ state-of-the-art GNNs, such as temporal graph convolutions, that adapt to the changing market and significantly enhance prediction accuracy. Extensive experiments on benchmark datasets further confirm our model’s superior performance compared to traditional approaches. Our model has the potential to revolutionize AI-driven financial forecasting by building deeper insight into stock market behaviors.