Aplicação de redes neurais na previsão de dificuldades financeiras em empresas listadas na B3

Abstract

The ability to anticipate financial distress is a pillar of risk management and investment analysis, as traditional statistical models show significant limitations in volatile economic environments. This study's central objective was to evaluate the application and effectiveness of Artificial Neural Networks (ANN), more specifically a Multilayer Perceptron (MLP), in predicting financial distress for companies listed on the B3, with a one-year predictive horizon. The methodology was based on a rigorous pipeline that started with 9,025 quarterly observations (2010-2024), including the construction of a time-lagged target variable and a hybrid feature engineering approach that combined financial, macroeconomic, and sectoral control indicators. The MLP's performance was compared directly to a Logistic Regression benchmark and to international historical models, such as Altman's z-score, and national models, such as those of Elizabetsky and Kanitz, through temporal cross-validation and a final test on recent data. The results indicate that hybrid feature engineering is the primary driver of performance, elevating the Logit model's AUC to 0.9070. However, the non-linear architecture of the MLP demonstrated a statistically significant advantage in the final test, achieving an AUC of 0.9229. The interpretability analysis confirmed that both models agree on the most important features, validating the hybrid set and suggesting that the MLP's advantage lies in its ability to capture complex nonlinear interactions. It is concluded that the MLP is the superior-performing tool, but its success is fundamentally dependent on feature engineering that incorporates macroeconomic context.