Credit risk scoring based on machine learning models: a comparative assessment

Commercial banks primarily generate revenue from secured and unsecured loans. However, despite the use of risk assessment tools, many borrowers are unable to repay their financial obligations for various reasons, which can impact the profitability and reputation of financial institutions. Therefore, credit approval processes should be carefully considered. Credit risk scoring was developed to assess customers before approving loans in order to improve the performance of credit assessment. This research aims to develop a credit risk scoring model using machine learning model with feature selection techniques. The dataset used in this research is the home credit default risk dataset from Kaggle. LightGBM was compared to Decision Tree and Random Forest, which were enhanced with feature engineering and statistical methods. Feature selection was performed to identify the optimal number of features. The model performance was evaluated using the AUC score. The number of features was reduced from 624 to 75 after performing feature selection process, and the results showed that reducing the number of features did not impact the model performance while it improves training time. LightGBM had the best AUC score at 78.00%, and Decision Tree had the best computational time reduction at 68.22%