Predicting corporate bankruptcy using modular neural networks

Abstract

The paper reports on the use of modular neural networks for predicting corporate bankruptcy. We obtained our financial, as well as, political and economic data from The London Stock Exchange, JORDANS financial database of major British public and private companies, and the Bank of England. In the past, various statistical techniques, such as univariate and multivariate discriminant analysis have been used in the modelling of corporate bankruptcy prediction. We use domain expert knowledge to select, and organise data in the modular neural network architecture constructed for this study. There are three sub-networks representing the periods, 1994, 1995, and 1996. Each sub-network is made of five adjacent networks representing the Balance Sheet network, the Profit and Loss network, the Financial Summary network, the Key Financial Ratios network, and the Economic and Political factors network. These adjacent networks although coupled but not linked at the input level represent five facets of failure in predicting corporate bankruptcy. The training sets represent data for 2500 companies selected randomly from a population of 270000 sample. The trained neural network will access 435000 data records before making a prediction for the particular company. The results obtained shows that neural networks outperform statistical techniques in modelling corporate failure prediction.