Financial time series prediction is extremely challenging due to the intertwined effects of market narratives and complex inter-asset relationships. Traditional prediction models often fail to distinguish similar price patterns driven by different underlying causes, limiting their predictive accuracy in practical scenarios. To address these limitations, this study proposes the Dual-stream Alpha Factor Fusion Network (DAFF-Net), an innovative deep learning framework that integrates event-driven temporal pattern extraction with multi-dimensional relationship-aware channel soft clustering. The event-driven temporal pattern extractor employs an event-aware router to fuse time series data with contextual event information encoded from news, corporate announcements, and macroeconomic data, enabling the model to understand the underlying narratives behind market fluctuations. The multi-dimensional relationship-aware channel soft clustering module constructs a comprehensive asset relationship network through adaptive fusion of frequency-domain, fundamental, and knowledge graph relationships, which is more effective than single-relationship approaches and better captures complex cross-asset dependencies. We validated our approach primarily on Amazon stock data covering the period from 2010 to 2025, with additional cross-asset validation on four stocks from different sectors (healthcare, financial, energy, and electric vehicle sectors). Results demonstrate that DAFF-Net significantly outperforms eight representative baseline models including ARIMA, LSTM, Transformer, and DUET across multiple prediction time horizons. Specifically, compared to the strongest baseline, DAFF-Net achieves 7.4%-15.2% improvement in MSE and 7.0%-21.4% enhancement in $$\text {R}^{2}$$ metrics, showing particularly outstanding advantages in long-term prediction tasks. These results prove the effectiveness of integrating event information and multi-dimensional relationships in financial prediction, providing a new technical paradigm for quantitative investment and risk management applications.

Enterprise bankruptcy prediction is a critical task in financial risk management. Traditional methods, such as logistic regression and decision trees, rely heavily on structured financial data, which limits their ability to capture complex relational networks and unstructured industry information. Heterogeneous graph neural networks (HGNNs) offer a solution by modeling multiple relationships between enterprises. However, current models struggle with financial risk graph data challenges, such as the oversimplification of internal financial features and the lack of dynamic imputation for missing external topological features. To address these issues, we propose HGNN-EBP, an enterprise bankruptcy prediction algorithm that integrates both internal and external features. The model constructs a multi-relational heterogeneous graph that combines structured financial data, unstructured textual information, and real-time industry data. A multi-scale graph convolution network captures diverse relationships, while a Transformer-based self-attention mechanism dynamically imputes missing external topological features. Finally, a multi-layer perceptron (MLP) predicts bankruptcy probability. Experimental results on a dataset of 32,459 Chinese enterprises demonstrate that HGNN-EBP outperforms traditional models, especially in handling relational diversity, missing features, and dynamic financial risk data.