Bankruptcy Prediction Using Ex Ante Neural Networks and Realistically Proportioned Testing Sets

The objective of this study is to assess the viability of a neural network in the prediction of bankruptcy, using a more realistic setting than has generally been examined in the past. In order for neural network bankruptcy prediction models to be useful and relevant, they must be designed to predict in realistic settings, rather than carefully designed matched-pair settings. The data sets which are tested in this study reflect the true proportion of firms that actually fail, which is less than one percent of all firms. Further, all models are developed ex ante to test subsequent years' data sets. Comparison of neural network predictive power with that from traditional logit models indicate that logit models remain a viable alternative to neural networks. The results also indicate that need for continuous updating of bankruptcy prediction models, a task to which neural networks are well adopted.