Using machine learning, neural networks and statistics to predict bankruptcy]

Recent literature strongly suggests that machine learning approaches to classification outperform "classical" statistical methods. We make a comparison between the performance of linear discriminant analysis, classification trees, and neural networks in predicting corporate bankruptcy. Linear discriminant analysis represents the "classical" statistical approach to classification, whereas classification trees and neural networks represent artificial intelligence approaches. A proper statistical design is used to be able to test whether observed differences in predictive performance are statistically significant. The data set consists of a collection of 576 annual reports from Belgian construction companies. We use stratified 10–fold cross–validation on the training set to choose "good" parameter values for the different learning methods. The test set is used to obtain an unbiased estimate of the true prediction error. Using rigorous statistical testing, we cannot conclude that in the case of the data set studied, one learning method clearly outperforms the other methods.