Predicting business risk using combined case-based reasoning in Euclidean space

We attempt to investigate the possibility of combining the thought of multiple classifiers system with the classifier of case-based reasoning (CBR) in Euclidean space to predict business risk. We use the classical CBR algorithm as the basic classifier, which is based on Euclidean metric and k-nearest neighbors. The three commonly used feature selection methods, i.e., stepwise method of multi-variant discriminant analysis (MDA), stepwise method of Logit, and ANOVA, are used to generate three optimal feature subsets, which are inputs of the classical CBR. Thus, there are three classifiers that could be generated, i.e., MDA-CBR, Logit-CBR, and ANOVA-CBR. Results of the three CBR classifiers are combined by majority voting to produce the final predictive result for a company´s financial state. The ensemble system is named as M-CBR-E-MV. In the experiment, data were collected from two Stock Exchanges in China. The two statistical models of MDA and Logit were employed as comparative models. The strategy of hold-out method was repeated for 30 times to generate 30 series of predictive results, on the basis of which, statistical indices and significance test were employed to assess the performances of MDA-CBR, Logit-CBR, Logit-CBR, M-CBR-E-MV, MDA, and Logit. The finding is that this ensemble system improves predictive stability of the systems more than it does in decreasing predictive precise than the best single classifier. Meanwhile, the ensemble system outperforms the two statistical models of MDA and Logit significantly.