Predicting stock price movements using hyperbolic space representation learning with cross-attention and multimodal data fusion

Abstract Accurate prediction of stock price movements remains a central challenge in developing effective investment strategies. However, the stock market is inherently a complex system influenced by diverse factors, requiring an integrated representation of multimodal data sources such as time series prices, news articles, and social media commentary. While prior studies have explored multimodal data fusion to some extent, they have largely overlooked the need for architectures specifically suited to capturing the hierarchical and non-linear structures embedded in market dynamics. To address this gap, we propose CHARMED (Cross-attention Hyperbolic Architecture for Representation of Multimodal Embeddings and Data-fusion), a novel multimodal fusion framework based on hyperbolic geometry. CHARMED integrates three distinct modalities–stock price sequences, financial news and Twitter commentary, and price covariance graphs–by leveraging self-attention and cross-attention mechanisms in the Poincaré ball model, followed by gated fusion to dynamically synthesize information. This enables effective learning of latent market structures that are both hierarchical and non-linear. Empirical evaluations on multiple real-world financial datasets demonstrate that CHARMED consistently outperforms conventional stock prediction baselines. Its use of hyperbolic embeddings allows for improved predictive accuracy and generalization, particularly under low-dimensional settings. Furthermore, quantitative analyses of hierarchy and scale-freeness in the learned embeddings confirm that CHARMED captures intrinsic structural patterns of the financial market.