Predicting default probability in credit risk with machine learning algorithms

Failure to repay the loans provided by banks and various financial foundations by the customer, entails both the capital loss of the lending institution and various risk factors that may occur in the general economy. In this context, financial control institutions such as the Basel Committee and BRSA (Turkish Banking Regulatory and Supervision Agency) have determined various regulatory policies during the phase of lending decision of the lending institutions in order to ensure the appropriate management of loan risk, which have critical importance, and to ensure international financial stability. In addition, lending institutions develop credit evaluation models via analytical risk units and calculate the credit risk score of customers. In this study, it is aimed to determine the algorithm that makes the most successful estimation that can be used in credit scoring systems with the machine learning method. Within this scope, models for algorithms with Gradient Boosting, Artificial Neural Networks, Logistic Regression, Random Forest, Decision Tree, Support Vector Machines, K-Nearest Neighbor and WOE transformations Logistic Regression were established and Gradient Boosting algorithm has shown the best classification performance for defaulters and non-defaulters. In analytical data quality and model development processes, SAS Enterprise Guide and SAS Enterprise Miner software programs were used. Key Words: Credit Risk, Machine Learning, Gradient Boosting, Neural Network, Logistic Regression, Random Forest, Decision Tree, Support Vector Machine, K-Nearest Neighbor