Two-stage hybrid learning techniques for bankruptcy prediction*

Many machine learning-based techniques have been used for the prediction of bankruptcy. They can be divided into single, ensemble, and hybrid learning techniques. This paper focuses on a two-stage hybrid learning approach for bankruptcy prediction where, in the first stage, a clustering algorithm is used to perform the instance selection task in order to filter out a certain number of unrepresentative training data. The clustering results output from the first stage are used with a classification algorithm to construct the prediction model. The results of experiments based on five different country datasets show that the best support vector machine (SVM) classifier performance is obtained using instance selection by affinity propagation (AP) and k-means individually. Moreover, we also find that although the best AP/k-means and SVM combination is dataset dependent, the criteria for selecting representative training data are specific. This should become a guideline for developing bankruptcy prediction systems based on the hybrid learning approach.