Author(s)

Haotong Chen

Corresponding Author

Haotong Chen

Abstract

Financial markets are nonlinear, dynamic, and complex in nature with structural, temporal, and textual interactions. Nevertheless, the current prediction models are based on either single modality or naive fusion, and cannot reflect inter-company risk spreading and cross-modal effects. To mitigate this weakness, this paper suggests TriFusionRiskNet, the single multi- modal neural architecture for simultaneous prediction of stock returns and risk variances. The structure incorporates three dualistic modalities: a temporal graph encoder, which models the changing relationship between firms; a time-series encoder, which models the multi-scale dynamics of historical trading; and a BERT-based text encoder, which parses sentiment-driven market information in financial news. To selectively direct the information between modalities with structural guidance, a graph-guided cross-modal attention mechanism is proposed, and a dynamic graph learning module is proposed to continually adjust according to actual regime changes in the market. Considerable testing of CSI300 A-share data in 2016-2023 shows that TriFusionRiskNet is much better in predictive accuracy and risk stability compared to the state-of-the-art baselines, especially in volatile market situations. The suggested framework is structurally based, interpretable and deployment-ready multimodal financial forecasting.

Keywords

Financial Risk Forecasting; Graph neural network; Stock Return and Volatility Prediction