D-HAN-Net: A Hybrid Dual-Stream Architecture for Corporate Bankruptcy Prediction via Multimodal Fusion

Early detection of corporate bankruptcy is essential for maintaining financial market stability and reducing systemic risk. However, existing predictive models often struggle under conditions of extreme class imbalance. Traditional approaches either analyze financial ratios or textual disclosures in isolation, while conventional cross-modal fusion strategies tend to dilute rare distress signals when integrating structured accounting metrics with qualitative management narratives. Consequently, subtle indicators of impending corporate failure are frequently overshadowed by dominant non-distress patterns, limiting the effectiveness of existing predictive systems. To bridge this methodological limitation, this study proposes D-HAN-Net, a hybrid dual-stream deep learning architecture. This framework is in-tended to address minority class suppression by dynamically balancing structured financial ratios with unstructured textual disclosures. Specifically, the model processes numerical indicators through a gated residual network to capture complex patterns. Simultaneously, it extracts semantic cues from corporate reports via a FinBERT-driven bidirectional GRU. These dual modalities are then aligned using a learnable cross-modal attention fusion gate, jointly optimized with focal loss. Experimental evaluations on a comprehensive multimodal dataset, utilizing stratified splits demonstrate that D-HAN-Net significantly outperforms state-of-the-art baselines, achieving a predictive accuracy of 94.00%, an F1-score of 88.00%, and an AUC of 0.9734. Practically, this framework equips investors, financial institutions, and regulatory authorities with a decisive early warning system. It enables proactive risk management by detecting subtle distress signals before corporate failure becomes irreversible. Furthermore, extensive stability testing and ablation analysis confirm that the model’s superior predictive reliability is highly robust against sampling uncertainty, fundamentally relying on the synergistic integration of all its architectural modules.