Amir Khorrami; Mahmoud Dehghan Nayeri; Ali Rajabzadeh
Abstract
This study assesses a new simulation-optimization method for credit scoring and bank loan parameter optimization. The proposed approach encompasses data preparation, credit scoring, and simulation-optimization stages. Initially, data regarding bank loans and company financial statements are collected ...
Read More
This study assesses a new simulation-optimization method for credit scoring and bank loan parameter optimization. The proposed approach encompasses data preparation, credit scoring, and simulation-optimization stages. Initially, data regarding bank loans and company financial statements are collected and relevant features are calculated. The Minimum Redundancy Maximum Relevance (MRMR) algorithm selects the most critical features. Subsequently, classification methods including logistic regression (LR), K-nearest neighbor (KNN), artificial neural network (ANN), adaptive boosting (AdaBoost), and random forest (RF) are employed to address the credit scoring problem. These models' performance is evaluated using accuracy, F1-score, and area under curve (AUC) criteria, with the best-performing model selected for subsequent stages. During simulation-optimization, optimal loan features are determined to minimize default rates by treating loan size, interest rate, and repayment period as optimization variables. The memetic algorithm (MA) solves this optimization problem in four cases, utilizing a pre-trained credit scoring model to estimate client default probability. A case study involving 1000 legal clients of an Iranian commercial bank demonstrated that 11 features were selected from 30 defined features for credit scoring. The RF method outperformed other credit scoring models. The simulation-optimization approach reduced default rates from 38% to 20% through decreased loan size and interest rates, coupled with extended repayment periods. These results confirm the method's effectiveness in reducing banking credit risk.
