Prior research raised questions about the information value of some of the variables included in Altman's 1968 seminal bankruptcy prediction model. Answering these questions is of great importance since the original Altman model and variations on it are still used to provide bankruptcy risk signals in accounting and audit practice.
This study applied genetic programming to Altman's original data set in order to examine the issue of variable significance. Two parsimonious models employing either one or two variables were developed that equaled the accuracy rate of Altman's five variable model when tested on Altman's original data set. The two variable model also equaled or exceeded the Altman model both when error rates were compared based on prior probabilities of bankruptcy and when relative misclassification costs were considered. The accuracy levels and parsimony of the two genetic programming models supports prior research and confirms that some of the variables in the original model were an artifact of discriminant analysis.
