Ensemble KNNs for Bankruptcy Prediction

The business failure has been widely researched, trying to identify the various determinants that can affect the existence of firms. However, the variety of models as well as the variety of the theoretical frameworks, illustrates the lack of consensus on how to understand the phenomenon and the difficulties in formulating a general model interpretation. One hotspot nowadays is the prediction of the bankruptcy, the final stage of failure, which has been regarded as a classification problem. Thus, this paper presents a global methodology to classify bankrupt companies and healthy ones. The computational time of this method is extremely small since it use k nearest neighbors (KNN) to build several classifiers, each of the classifiers use different nearest neighbor on different subset of input variables and try to minimize the mean square error. Finally a linear combination of these classifiers is calculated to get even better performance. On the other hand, this method is robust because the ensemble of classifiers has smaller variance than each single classifier. The method is tested using a real world data, which comprises 41 financial variables measured on 500 companies (250 healthy and 250 bankrupt) from the year 2002 and 2003. The result confirms that the advantage of this method, which is that it is robust while it provides good performance and a comparatively simple model at extremely high learning speed.