MultiStream-FinBERT: A Hybrid Deep Learning Framework for Corporate Financial Distress Prediction Integrating Accounting Metrics, Market Signals, and Textual Disclosures

Financial distress prediction represents a critical challenge in corporate finance, with significant implications for investors, creditors, and regulatory bodies. This paper introduces MultiStream-FinBERT, a novel hybrid deep learning framework that integrates accounting metrics, market signals, and textual disclosures to enhance the accuracy and timeliness of financial distress prediction. The proposed architecture employs specialized processing modules for each data stream, with a sophisticated cross-attention mechanism facilitating effective information fusion across modalities. We construct and validate our model using a comprehensive dataset of 3,582 publicly traded companies spanning 2010-2023, with 426 experiencing financial distress. Extensive experiments demonstrate that MultiStream-FinBERT achieves 94.73% accuracy and 96.84% AUROC, substantially outperforming existing approaches including LSTM-Attention (91.86%, 94.18%) and traditional statistical models (79.24%, 81.56%). Ablation studies confirm the critical contribution of each data stream, with the accounting stream providing the strongest individual signal. The model maintains strong predictive performance up to 9 months before distress events, offering stakeholders extended warning periods for intervention. Feature importance analysis reveals distinct patterns across industry sectors and prediction horizons, with a shift from immediate liquidity indicators at shorter horizons toward structural factors at longer timeframes. The proposed framework offers significant advancements in financial risk assessment through its multimodal approach and enhanced interpretability.