Machine Learning–Based Financial Distress Prediction: Evidence from Borsa Istanbul Manufacturing Firms

Accurate financial distress prediction is crucial for investors, regulators, and firm managers to mitigate risk and enhance early warning systems. This study investigates the effectiveness of alternative machine learning approaches for financial distress prediction using firm-level financial ratios. Financial distress is modeled as a binary classification problem, where profitability, liquidity, leverage, and activity ratios are employed as explanatory variables. To assess the impact of alternative modeling approaches, three complementary classification models are implemented: Logistic Regression as a linear model, Support Vector Machine with Principal Component Analysis as a kernel-based nonlinear approach, and Random Forest as a tree-based ensemble method. Model performance is evaluated using accuracy, macro F1-score, recall for the distressed class, and the area under the ROC curve, alongside stratified cross-validation to assess robustness and generalization ability. The empirical results indicate that while LR provides strong and interpretable baseline performance, the RF model consistently outperforms alternative approaches across all key evaluation metrics, achieving the highest macro F1-score and ROC–AUC values. Feature importance analysis reveals that profitability indicators are the most influential predictors of financial distress, followed by liquidity measures. The findings demonstrate that ensemble-based machine learning models offer substantial improvements in financial distress prediction by capturing nonlinear relationships and complex interactions among financial ratios. The results highlight the importance of profitability and liquidity in explaining financial distress and support the use of RF models as a robust and effective framework for early warning systems.