Early warning systems (EWS) for firm-level financial distress are essential for identifying potential bankruptcies or insolvencies before their realization. While traditional statistical models such as Z-score and logistic regression offer interpretability, they lack the ability to capture nonlinear and temporal dependencies in financial data. Recent deep learning approaches improve predictive accuracy but often sacrifice interpretability. The purpose of this study is to develop and evaluate a novel deep learning-based early warning model for firm-level financial distress that integrates temporal attention with parametric survival analysis to improve both predictive accuracy and interpretability. Therefore, this study proposes an AFT-Attentive BiLSTM model that integrates a Bidirectional Long Short-Term Memory (BiLSTM), a temporal attention mechanism, and a log-normal Accelerated Failure Time (AFT) survival framework. The model predicts time-to-distress distributions rather than binary outcomes, enabling probabilistic early warnings with calibrated survival probabilities. Empirical results demonstrate that the proposed model outperforms Cox Proportional Hazards, DeepSurv, and prior AFT-BiLSTM models without attention. The inclusion of temporal attention improves concordance index (C-index), Integrated Brier Score (IBS), and time-dependent AUC, and provides interpretable insights by identifying critical financial periods preceding distress. Kaplan–Meier analysis confirms strong separation between high- and low-risk groups. The findings suggest that combining temporal attention with parametric survival modeling enhances both predictive accuracy and interpretability in financial distress early warning systems.
Chohan et al. (Thu,) studied this question.