Previsão de insolvência corporativa: uma análise de empresas brasileiras de capital aberto por meio de aprendizado de máquina

Abstract

This research aimed to analyze the effectiveness of machine learning models in predicting corporate bankruptcy, considering the context of the Brazilian stock market, with data from 2010 to 2020. I considered companies from all sectors. Moreover, I analyzed financial firms jointly and separately, given their specificities. Considering the annual average, 299 Brazilian publicly traded companies had their data analyzed by the study. I measured the risk of financial distress in two ways: using the Z’’ Score model, and using the K-means, an unsupervised machine learning model. I tested the following supervised machine learning models: random forest, naive bayes, logit, K-NN, SVM (linear, polynomial and radial basis), bagging and boosting. When I carried out the analysis with all companies, the random forest showed the best results, with an accuracy between 90.94% and 95.44%. When I analyzed only financial firms, the best predictive models were as following: polymonial SVM, logit or boosting, depending on the performance metrics. If the decisive metric for choosing the best model is AUC, random forest shows marginal differences compared to the best models, which means that it can be used in financial firms without loss of performance. The results of this work support the decisions of investors, who capture information to rationally define the most advantageous assets; of executives, who are concerned with the credibility of organizations; of credit provider institutions, which will provide loans and financing with more assertiveness; of researchers, who will contribute with improvements for the prediction of the risk of insolvency; and other stakeholders, who will use the findings according to their informational needs.