The effect of training set selection when predicting defaulting small and medium-sized enterprises with unbalanced data

We focus on classification methods to separate defaulting small and medium sized enterprises from nondefaulting ones. In this framework, a typical problem occurs because the proportion of defaulting firms is very close to zero, leading to a class imbalance. Moreover, a form of bias may affect the classification because models are often estimated on samples of large corporations that are not randomly selected. We investigate how different criteria for sample selection may affect the accuracy of the classification and how this problem is strongly related to class imbalance.