The Application of Ensemble Methods to Bankruptcy Prediction on the Example of Russian Manufacturing Companies

In research literature there are number of studies addressed to the problem of bankruptcy prediction. In search of better model authors provide a pool of methods with a single classifier, among which statistical, machine learning and more advanced techniques are applied. Recent studies outbreak that multiple classifier models, e.g. ensemble methods, demonstrate better performance in comparison with single classifier models. In this study we applied three common ensemble methods (bagging, boosting and stacking) based on three the most popular base learning techniques (support-vector machine, logit regression and decision tree) in order to check the comparative performance of ensemble methods over models with single classifier. We carried out an experiment on the sample of Russian manufacturing companies, which has high number of bankruptcy cases, and made predictions in one-year horizon. We found that on average ensemble methods improve the results demonstrated by the methods with a single classifier, and, in particular, Stacking outperforms other ensemble techniques, Bagging demonstrates better performance than Boosting over the base learners and application of these three ensemble methods allow to significantly decrease type II error for support-vector machine. We also overcame the limitation of removing necessary instances of majority class by random under sampling function.