Reducing Lending Risk: SVM Model Development with SMOTE for Unbalanced Credit Data

Abstract

Lending is an important activity for banks in managing available funds. However, lending is also an activity that has a high risk, because not all customers who borrow funds can fulfill the responsibilities of the existing agreement. Because of this, it is necessary to have a method that can predict creditworthiness to customers in order to minimize the risks that arise. This research uses machine learning method, namely Support Vector Machine (SVM) in predicting creditworthiness. This method is applied and compared before and after the Synthetic Minority Oversampling Technique (SMOTE) on historical bank credit data BPR NBP 16 Rantau Prapat, North Sumatra and find the best parameters with grid search. According to the results of the analysis based on Area Under the Receiver Operating Characteristic Curve (AUC-ROC), SVM with SMOTE shows better results, namely 96%, than SVM without SMOTE, namely 56%.