Selection of accounting features in the prediction of bankruptcy in Brazilian companies – A comparison of approaches

Abstract

Insolvency prediction has been a topic of study that has gained much attention in business analysis because of the importance of accurate and timely information on strategic business decisions. This is because the incorrect decision-making in institutions can generate financial difficulties besides causing huge social costs that affect the owners or shareholders, managers, employees, creditors, suppliers, customers, community, government, etc. As a result, bankruptcy prediction has been one of the most challenging tasks and an important research topic in accounting, finance and computer science and data mining techniques have been applied to solve problems in bankruptcy prediction. The selection of attributes is important to select the most representative data from a set of accounting ratios derived from financial statements of Brazilian companies; this step aims to improve the performance of the final prediction step. The main objective of this paper is to compare three approaches to feature selection, viz. Filter, wrapper and principal component analysis, in data selected for the development of insolvency prediction models. This research is of an empirical, descriptive and quantitative nature, comprising companies classified at SERASA and BOVESPA as insolvent in the period of 2005-2007. This work demonstrated, for the sample used, that the wrapper approach is the most effective one; it obtained the best classification results in the techniques of logistic regression (89,88%), decision tree (93,45%) and support vector machine (97,02%).

Keywords: selection of attributes, insolvency prediction, accounting ratios, data mining.