The effect of class balancing methods on machine learning techniques: Example of credit risk

Credit risk arises as a result of defaulting on liabilities at the end of the specified maturity of funds ceded to customers to be reimbursed by banks. This situation causes the credit customer to default and negatively affects their reputation. With technological development, machine-learning methods used in different areas facilitated the identification of risky customers with the system adapted to the lending process of banks. Thus, the risk situation of customers applying for loans can be determined quickly and more accurately. The Identification of these risky customers is important in terms of both reducing the credit risk and maintaining the financial activities of the banks. In this study, it is aimed to investigate the effects on machine learning by balancing the datasets with different resampling methods in order to find a solution to the class imbalance problem encountered in unbalanced datasets so that banks can make the most appropriate assessment in terms of credibility. In the implementation phase of the study, German, Australian and HMEQ real-life problem credit datasets obtained from an open source website were used. In the study, different machine learning classification methods such as K-Nearest Neighbor, Naive Bayes, Logistic Regression, Support Vector Machines, Multilayer Perceptron, Decision Trees, Random Forests, Gradient Boosting Decision Trees, Extra Trees, Voting Hard and Soft were used to detect risky customers. The class imbalance problem has been balanced with resampling and hybrid methods such as RUS, ROS, SMOTE-ENN, SMOTE-Tomek and balanced bagging Classifier. In this direction, four different scenarios were examined in three different sets of credit data. As a result, it was observed that the hybrid method, in which over-and-under sampling methods are used together for the class balancing problem, can help machine learning techniques to achieve the best classification performance. This method will provide a great advantage in identifying risky customers and will help banks to reduce their credit risk. Keywords: Credit Risk, Machine Learning, Ensemble Learning, Classification Algorithms, Resampling, Class Balancing