Classification of hundreds of classes: A case study in a bank internal control department

Abstract

The main function of the internal control department of a bank is to inspect the banking operations to see if they are performed in accordance with the regulations and bank policies. To accomplish this, they pick up a number of operations that are selected randomly or by some rule and, inspect those operations according to some predetermined check lists. If they find any discrepancies where the number of such discrepancies are in the magnitude of several hundreds, they inform the corresponding department (usually bank branches) and ask them for a correction (if it can be done) or an explanation. In this study, we take up a real-life project carried out under our supervisory where the aim was to develop a set of predictive models that would highlight which operations of the credit department are more likely to bear some problems. This multi-classification problem was very challenging since the number of classes were enormous and some class values were observed only a few times. After providing a detailed description of the problem we attacked, we describe the detailed discussions which in the end made us to develop six different models. For the modeling, we used the logistic regression algorithm as it was preferred by our partner bank. We show that these models have Gini values of 51 per cent on the average which is quite satisfactory as compared to sector practices. We also show that the average lift of the models is 3.32 if the inspectors were to inspect as many credits as the number of actual problematic credits.