Bank Failure Prediction Using Hybrid Classifier Ensembles of Random Sub-Spaces and Bagging

Assoc. Prof. Halil İbrahim Erdal (speaker), Dr. Aykut Ekinci
Turkey, Turkish Cooperation and Coordination Agency, F. A. Hayek Visiting Scholar, George Mason University
In this study, the use of hybrid classifier ensembles of artificial neural networks (ANN) is investigated. In other words, an attribute-base ensemble learning method (i.e., random sub-spaces RS) and an instance-base ensemble learning method (i.e., bagging B) are employed to enhance
the prediction accuracy of conventional ANN models for bank failure prediction. Two common artificial neural networks methods (i.e. multilayer perceptron MLP; voted perceptron VOP) are used as the base learners and benchmark models. The extent of this study encompasses 37 privately owned commercial banks (17 failed, 20 non-failed) that were operating in Turkey for the period of 1997-2001. Initially, four single ensemble models are built by using two ensemble learning methods and two conventional ANN (i.e., B-MLP; B-VOP; RS-MLP; RS-VOP). Next, in this study, two different hybrid ensemble learning approaches are proposed: (i) hybrid 1 approach: instance-based ensemble learning method + attribute-based ensemble learning method + artificial neural networks (i.e., B-RS-MLP, B-RS-VOP); (ii) hybrid 2 approach: attribute-based ensemble learning method + instance-based ensemble learning method + artificial neural networks (i.e., RS-B-MLP, RS-B-VOP). The obtained results indicate that ensemble models are superior to the base models and hybrid 2 approach classifier ensembles are the best models among 10 predictive models for bank failure prediction in terms of average accuracy.