A PREDICTIVE MODEL FOR STOCK PRICES BASED ON TRANSFORMER AND UTILIZING MULTIMODAL DATA

Stock market prediction necessitates effective multimodal data integration and robust uncertainty quantification. This paper proposes a novel Transformer-based architecture addressing two critical limitations of existing approaches: static cross-modal interaction and deterministic output assumptions. Our framework introduces (1) a multimodal subspace attention mechanism that projects numerical and textual features into orthogonal subspaces, enabling disentangled learning of modality-specific interactions through multiple attention heads, and (2) a dynamic gated recalibration module that adaptively adjusts modality contributions using time-variant weights. Evaluated on Technology Select Sector SPDR Fund (XLK) data with market sentiment feeds, the model achieves higher directional accuracy than conventional Transformers while reducing volatility period prediction errors. The integrated uncertainty quantification module further provides statistically reliable confidence intervals, verified through backtesting.