Using genetic programming as a feature selector and classifier to implement bankruptcy prediction models

Genetic Programming (GP) was used as a feature selector and classifier to implement bankruptcy prediction models for medium-sized companies. Two sets of input variables were used for the prediction models: one using a large number of exclusively financial variables and the other incorporating variables from the economic environment, which allows analyzing the capability of the latter to improve performance. Two strategies were defined for GP as a feature selector, based on the statistical relevance of the selected features in the GP process, with a novel proposal based on a progressive reduction of the set of selected variables and with the aim of minimizing the risk of eliminating relevant features. An analysis is performed of the improvement obtained with feature selection with both GP-based methods in comparison with the use of complete sets of variables and using GP as a classifier. With the selected variables, we also compared GP as a classifier with respect to other standard classifiers, using automatic parameter adjustment with AutoWeka for these classifiers. The best results are obtained with the synergy of using GP as a feature selector and as a classifier, with the advantage of the direct interpretability that GP provides in the application.