A Dynamic Weighted Fusion-Based Hybrid XGBoost-LSTM Model for Financial Distress Prediction

Financial distress prediction (FDP) is pivotal for corporate risk management and financial ecosystem stability. Traditional methods such as the Z-Score and logistic regression are constrained by their inability to effectively address temporal dependencies and limited model interpretability. Although XGBoost excels in static feature modeling and LSTM captures temporal dynamics, their isolated implementations fail to harmonize static and temporal feature interactions. This study proposes a novel dynamically weighted fusion-based hybrid XGBoost-LSTM ensemble model to address these gaps. By integrating gradient-boosted trees for nonlinear static feature analysis and LSTM with attention mechanisms for sequential pattern mining, our model employs adaptive weight allocation via a learnable gating network, where XGBoost and LSTM contributions are dynamically assigned weights of 0.6 and 0.4, respectively. Evaluated on a Kaggle-sourced dataset of 2,836 firms with 12-quarter financial records, the model achieves a ROC AUC of 0.8809, surpassing standalone XGBoost by 2.1% and LSTM by 1.2%, with a recall of 0.9962, minimizing false negatives. Enhanced interpretability is achieved through XGBoost's feature importance ranking and LSTM's attention visualization, providing dual explanations for risk alerts. The proposed model provides a deployable solution for real-time financial monitoring systems, demonstrating practical utility in early risk detection.