Previsão do risco de insolvência empresarial no Brasil : uma abordagem estatística e de aprendizado de máquina

Abstract

This dissertation aims to analyze the predictive performance of statistical and machine learning techniques in forecasting the risk of insolvency of Brazilian non-financial firms listed on B3 (Brasil, Bolsa, Balcão) over the period from 2010 to 2024. The research is structured into three empirical studies that apply different methodologies to a set of traditional and dynamic financial indicators, mainly related to liquidity, leverage, profitability, and working capital management, in order to identify the variables with the greatest explanatory power in the classification of solvent and insolvent firms. The first study employs Discriminant Analysis, which presents robust performance, with an AUC of 0.955 and an accuracy of 93.3% in the test sample, highlighting the combined relevance of traditional indicators and the Fleuriet Dynamic Model in separating solvent and insolvent groups. The second study uses panel data Logistic Regression estimated by Random Effects, allowing the model to capture cross-firm variation, the temporal evolution of insolvency risk, and unobserved individual effects, achieving an AUC of 0.915 and an accuracy of 90.2% and emphasizing the importance of indicators related to capital structure, liquidity, working capital, and financial availability. The third study, of a comparative nature, evaluates six machine learning algorithms—Decision Tree, Random Forest, Support Vector Machine, K-Nearest Neighbors, Multilayer Perceptrons, and Extreme Gradient Boosting—and reveals the superiority of ensemble techniques, with the Extreme Gradient Boosting model obtaining the best overall performance, with an AUC of 0.981 and an accuracy of 95.4%, followed closely by Random Forest, with an AUC of 0.978 and an accuracy of 95.2%, indicating that profitability, capital structure, and liquidity indicators jointly explain corporate insolvency. In an integrated perspective, the results show that both statistical approaches and machine learning models are effective in anticipating corporate insolvency in the Brazilian context, while the analysis of variable importance confirms that traditional and dynamic indicators related to liquidity, leverage, profitability, and working capital are decisive in predicting and anticipating insolvency. From a scientific perspective, the study contributes by incorporating multiple techniques, using a large and recent sample, and integrating traditional and dynamic indicators into the field of insolvency research, thus filling gaps in the national literature. From a social, managerial, and institutional standpoint, the proposed models provide support for improving credit policies, investment appraisal, debt renegotiation, and preventive financial monitoring practices, assisting managers, creditors, investors, and regulators in mitigating insolvency risk and promoting economic stability. Keywords: Corporate Insolvency risk. Multivariate analysis. Machine learning models. Fleuriet's Dynamic Model. Traditional Financial Indicators.