Examination of tree-based techniques in imbalanced data sets: The example of credit risk

Credit risk is a type of risk that may endanger the existence of the bank as a result of the loan customer's failure to pay the funds borrowed from the bank in the specified time. For this reason, the correct identification of customers that may cause credit risk is of great importance for the continuity of the bank's activities. At this point, banks that take action to lend their customers must carry out their lending processes with great care in order to eliminate the possible credit risk. Otherwise, a weakness or negligence that will occur in the process may cause the credit risk to arise and the financial activities of the bank to be interrupted or stopped. Machine learning techniques, which have become widespread especially with the development of technology, have been integrated into lending activities in the banking sector and have helped to develop mechanisms that automatically decide whether customers who request credit carry a risk or not. In this context, different statistical and machine learning techniques have been sought over time to make the decision mechanisms created more reliable and more effective, based on the fact that the correct determination of the customer's ability to pay the loan is important for the financial and physical existence of the bank. In this study, it is aimed to create a decision mechanism in which banks can perform their lending activities automatically and at the same time increase their profitability by reducing the credit risk. The study was conducted on open source German, Australian and HMEQ (Home Equity) loan datasets. Eleven different machine learning classification techniques, nine different resampling vii techniques and one feature selection technique were used to create the targeted mechanism. As a result, the best performing models in the study were determined as tree-based models. In this direction, it has been found that preparing the mechanisms to be developed in response to the credit risk problem with combinations of over-sampling technique and tree-based model and at the same time operating this mechanism with datasets that do not apply the feature selection technique will enable banks to make more accurate decisions in their lending processes.