Modeling the Probability of Bankruptcy of Listed Companies Using Classification Algorithms (Random Forest, XGBoost, SVM)

Abstract

The objective of this study is to develop and compare advanced machine learning models for predicting corporate bankruptcy among firms listed on the Tehran Stock Exchange in order to identify the most effective classification framework for early financial distress detection. This research employed a quantitative predictive design using financial data from 185 non-financial firms listed on the Tehran Stock Exchange over the period 2013–2023. A total of 1,850 firm-year observations were extracted from audited financial statements and official market databases. Key financial indicators covering liquidity, leverage, profitability, efficiency, growth, cash flow, and firm size were constructed and standardized. The dataset was divided into training and testing subsets, and three classification algorithms—Random Forest, XGBoost, and Support Vector Machine—were implemented. Hyperparameters were optimized using grid search with cross-validation, and model performance was evaluated using accuracy, precision, recall, F1-score, and area under the ROC curve. The XGBoost model achieved the highest predictive accuracy (93.1%) and the largest area under the ROC curve (0.972), followed by Random Forest (accuracy = 91.7%, AUC = 0.963) and Support Vector Machine (accuracy = 86.1%, AUC = 0.901). Feature importance analysis revealed that return on assets, operating cash flow to total debt, leverage ratios, and profitability indicators were the most influential predictors of bankruptcy risk. The results demonstrate that ensemble learning algorithms, particularly XGBoost, provide superior performance in bankruptcy prediction within the Tehran Stock Exchange, offering a reliable framework for early financial distress detection and enhanced financial risk management.