Estimating bankruptcy using neural networks trained with hidden layer learning vector quantization

• Autor(a) principal: Neves, João Carvalho das
• Data de Publicação: 2004
• Outros Autores: Vieira, Armando
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.5/2189
• topic: Bankruptcy Prediction; Neural Networks; Discriminant Analysis; Ratio Analysis

The Hidden Layer Learning Vector Quantization (HLVQ), a recent algorithm for training neural networks, is used to correct the output of traditional MultiLayer Preceptrons (MLP) in estimating the probability of company bankruptcy. It is shown that this method improves the results of traditional neural networks and outperforms substantially the discriminant analysis in predicting one-year advance bankruptcy. We also studied the effect of using unbalanced samples of healthy and bankrupted firms. The database used was Diane, which contains financial accounts of French firms. The sample is composed of all 583 industrial bankruptcies found in the database with more than 35 employees, that occurred in the 1999-2000 period. For the classification models we considered 30 financial ratios published by Coface available from Diane database, and additionally the Beaver (1966) ratio of Cash Earnings to Total Debt, the 5 ratios of Altman (1968) used in his Z-model and the size of the firms measured by the logarithm of sales. Attention was given to variable selection, data pre¬processing and feature selection to reduce the dimensionality of the problem.