Predicting bankruptcy risk with bidirectional long short-term memory and convolutional neural networks

Predicting the threat of bankruptcy risk and estimating company survival rate, especially under many challenges that might unfold, is of high economic importance. In this study, we aimed to predict the one-year-ahead and two-year-ahead bankruptcy status of companies using financial features (predictor variables) selected by the recursive feature elimination technique. We proposed the use of deep learning techniques, bidirectional long short-term memory (BiLSTM), and convolutional neural network (CNN) for bankruptcy prediction using temporal data. We compared prediction performance of models using the selected features with the original Altman Z-score financial features. To evaluate the proposed deep learning models, we used a sample of 4,884 companies (150 bankrupt and 4,734 non-bankrupt) drawn from the COMPUSTAT database (1994–2020) and compared them against six benchmark models, reweighted Altman Z-score, linear discriminant analysis, regularized logistic regression, random forest, XGBoost, and LightGBM, over 100 random 70/30 company-level splits, four feature sets, and two prediction horizons. The results showed that no single model dominates across all evaluation criteria: At the conventional decision threshold (t = 0.5), BiLSTM achieved the highest geometric mean of recall and specificity on the Altman feature set (0.653 and 0.615 for the one- and two-year-ahead horizons, respectively), whereas random forest attained the highest AUC on data-driven feature sets (up to 0.790). At that threshold, the tree ensembles flagged almost no failing companies because their calibrated probabilities rarely exceeded 0.5 under severe class imbalance; thus, the proposed deep models, and BiLSTM in particular, were the only ones that detected bankrupt companies reliably without tuning the classification threshold to the data. When the decision threshold was tuned near the bankruptcy base rate, tree-ensemble models recovered comparable balanced detection, underscoring the importance of evaluating bankruptcy prediction models using multiple performance criteria and operating thresholds rather than a single accuracy metric.