Resumo

Corporate bankruptcy forecasting, a field of study that has been intensively explored since the 1960s, is vital to various stakeholders. From managers and investors to employees and customers, these forecasts are crucial in mitigating financial risks and maintaining global economic stability. This study evaluates six machine learning models to predict bankruptcies in companies listed on the Brazilian stock exchange (B3), continuing the historical journey of bankruptcy prediction.

The Brazilian market's unique economic characteristics necessitate adaptive forecasting models. This research explores the applicability of Random Forest, XGBoost, LightGBM, CatBoost, SVM, and Logistic Regression in predicting bankruptcies. The objective is to adapt these models to different industrial sectors in Brazil, enhancing their accuracy and robustness in predicting financial distress within the specific economic context of Brazilian companies, thereby providing practical solutions for stakeholders in the Brazilian financial market.

The transition from traditional statistical methods to modern machine learning techniques marks a significant evolution in bankruptcy prediction. This evolution, from traditional approaches like discriminant analysis and logistic regression to advanced models such as ensemble methods, neural networks, and SVM, offers improved accuracy by capturing nonlinear relationships and complex interactions within financial data. This progress, as demonstrated in studies by Barboza et al. (2017) and Chen and Guestrin (2016), is a testament to the innovation in the field of bankruptcy prediction.

The study analyzed financial data from 564 companies listed on B3, covering 2000 to 2023. The dataset included various financial indicators critical for predicting bankruptcy. Six machine learning models were trained and tested using a split of 76% for training and 24% for testing. The SMOTE technique was employed to address the class imbalance. Hyperparameters were optimized using Grid Search and Bayesian optimization, and model performance was evaluated using accuracy, precision, recall, specificity, and AUC metrics.

The CatBoost model, combined with SMOTE, achieved the highest accuracy (99%) and specificity (100%) at specific cutoff points. SHAP value analysis revealed that market value, current liabilities, and gross debt were significant bankruptcy predictors. Sectoral dummy variables indicated that industry-specific factors significantly impact bankruptcy prediction. These findings align with the studies by Wang et al. (2023) and Sulistiani et al. (2022), highlighting the importance of tailored models for specific economic contexts.

Machine learning models, particularly those utilizing boosting techniques like CatBoost, proved highly effective in predicting bankruptcies in the Brazilian market. The study underscores the critical role of financial and sector-specific variables in enhancing prediction accuracy. Although hyperparameter optimization was explored, default parameters were often robust. Future research should incorporate qualitative data and explore these techniques in diverse economic and industrial contexts to improve model robustness and applicability.

Ferren, F., & Kurniadi, F. I. (2023). We are enhancing Bankruptcy Prediction with Feature Selection in the AdaBoost Algorithm. 2023 10th International Conference on ICT for Smart Society (ICISS), 1-4. https://doi.org/10.1109/ICISS59129.2023.10291988 Sulistiani, I., Widodo, & Nugraheni, M. (2022). Comparison of Bankruptcy Prediction Models Using Support Vector Machine and Artificial Neural Network. 2022 11th Electrical Power Electronics Communications Controls and Informatics Seminar (EECCIS), 316-321. https://doi.org/10.1109/EECCIS54468.2022.9902935