Modelos de classificação aplicados à previsão de insolvência de empresas brasileiras de capital aberto

Abstract

This study aims to apply five different classification techniques to build insolvency prediction models. The methods used were Linear Discriminant Analysis, Logistic Regression, k-Nearest Neighbors, Classification Trees and Artificial Neural Networks. To this end, it was used accounting-based information of Brazilian joint-stock companies, which compounded a sample of 87 firms. As secondary objectives, this research compared the results that were obtained by each method and then tried to indicate which method performed better, and also which indicators were the most useful to the analysis. The results obtained to the five models indicate that it is possible to identify with good confidence, which companies are becoming insolvent. The method that achieved the best performance was the Artificial Neural Network. The most important indicators were the Equity to Total Liabilities, Net Profit Margin and the Working Capital to Total Assets.