Credit scoring is a typical binary classification problem. Its significance to financial institutions has brought application of many quantitative methods. Most published research is focused on increasing classification performance by adjusting algorithms, generally without a corresponding analysis of intrinsic dataset difficulties. Prior research shows that these intrinsic difficulties cause all methods to yield less than perfect classification of testing samples in dataset. Hence, our discussion concentrates on the complexity of datasets. In this study, a new approach based on convex hull is suggested as a means to measure the classification complexity of credit scoring datasets. An empirical example is provided to demonstrate the efficiency of the new approach.