Hybrid deep learning for bankruptcy prediction: An optimized LSTM model with harmony search algorithm

Abstract Accurate bankruptcy prediction is essential for financial stability, risk management, and decision making. Traditional statistical models cannot capture complex nonlinear relationships in financial data; hence, it is essential to use advanced deep learning techniques. This study proposes a hybrid long short-term memory (LSTM) model, optimized using the harmony search algorithm (HSA) for enhanced predictability. A financial dataset of Polish companies from the emerging markets information service was utilized, which consisted of financial indicators spanning several years. For feature selection, principal component analysis was employed to achieve dimensionality reduction without loss of pertinent financial data. The new HSA-LSTM model was compared with benchmark classifiers such as fuzzy logic neural network, kernel extreme learning machine, extreme learning machine, and support vector machine in terms of key performance measures. The HSA-LSTM model outperformed all of the baselines with 90.80% accuracy, 89.46% precision, 90.22% recall, and an F -score of 90.54%. Statistical verification using ANOVA and Tukey’s HSD test confirmed that the improvement was significant ( p < 0.05). These findings highlight deep learning with hyperparameter optimization for financial distress prediction. This study contributes to early bankruptcy prediction and financial risk modeling and offers a high-accuracy, scalable prediction model. Future research could explore additional metaheuristic optimization techniques, explainable AI (XAI), and macroeconomic indicators to further enhance predictive accuracy.