Attribute Selection in Bankruptcy Prediction: Application and Evaluation Using Recent Brazilian Data

• Autor(a) principal: Horta, Rui Américo Mathiasi
• Data de Publicação: 2013
• Outros Autores: Alves, Francisco Jose dos Santos, de Carvalho, Frederico A.
• Tipo de documento: Artigo
• Idioma: por
• Título da fonte: RAM. Revista de Administração Mackenzie
• Texto Completo: https://editorarevistas.mackenzie.br/index.php/RAM/article/view/4052
• topic: Índices econômico-financeiros; Previsão de insolvência; Mineração de dados; Seleção de atributos; Abordagens filtro e wrapper.

Bankruptcy prediction may have great utility to financial and nonfinancial institutions with regard to take in advance the best possible decisions regarding loans or investments. In specific literature, many bankruptcy prediction models have made use of Data Mining. The preprocessing step is important to select good quality data for use in mining operations. Still, although the selection of attributes can be very beneficial to pre-select representative data to improve the forecast performance end, it is not known which method is the best selection. This work has as main objective to compare two approaches for evaluating subsets of attributes - Filter and Wrapper. Despite being based on data mining techniques and widely used in the step of feature selection in bankruptcy prediction models, these techniques are rarely used to treat data from financial statements of Brazilian companies. Therefore the empirical basis of this study consists of a sample of Brazilian industrial and commercial enterprises, collecting data for the period 2004 to 2011. The results indicated that, in this sample, the filter approach was more efficient, providing better classification results both for logistic regression (91,80%) and for neural networks (93,98%). It was shown also the importance of making explicit the evaluation stage of the selection of attributes for achieving better results in applications of data mining techniques to predict insolvency. A specific conclusion about the advantages of the filter approach shows that it may be preferred to assess the attributes that will make predictive models.