Predicting japanese corporate bankruptcy in terms of financial data using neural network

This study investigated the effectiveness of a neural network with one hidden layer as a bankruptcy prediction system using only financial data for 1 and 3 years prior to failure to show equivalence of both techniques for listed firms and for unlisted firms.

Abstract

The study investigated the effectiveness of a neural network with one hidden layer as a bankruptcy prediction system using only financial data for 1 and 3 years prior to failure. 8 and 4 sets of input data items were used for listed and unlisted firms respectively. 3 levels of number of units in hidden layer were used for both listed and unlisted. For listed firms only five ratios were sufficient to obtain zero misclassification. For unlisted data 9 input items were needed to obtain the best 15.2% misclassification rate of nonfailed firms for 1 year data. Comparison with discriminant analysis showed equivalence of both techniques for listed firms. Whereas for unlisted firms neural nets proved to be superior than the other.