Credit ranking of natural clients of banks using various neural network models: a case study of a private Iranian bank

Abstract

Previously, decision about granting facilities to clients of banks in Iran were made based on personal judgment about the risk of failure in reimbursement. But, increasing demand for bank facilities by economic firms and households in one hand, and increasing extensive commercial competition among banks and economic- credit companies in the country and their efforts to alleviate the risk of failure in reimbursement of facilities on the other hand, have resulted in using modern methods such as statistical methods in this area. Today, to predict the possibility of failure in reimbursement of facilities and to classify their applicants, banks use the credit ranking of their clients. Savings in time and costs, removing personal judgments and increasing the accuracy of evaluating applicants of various facilities are some of the benefits gained in this method. There are various statistical methods such as audit analysis, logistic regression, nonparametric smoothing and other methods including neural networks which have been used in ranking the credits. Among these methods, neural network method is of higher flexibility and has attracted more attention in recent years due to its ability to classify, generalize and learn the patterns. In this paper, firstly we select some of the important criteria in granting various credit facilities such as financing loan, civil partnership, installment sale and unilateral contract to natural clients of a private bank in the country using questionnaire and the opinions of elite people in the field of banking. Then, we classify them by presenting four models of neural networks namely MOE, MLP, LVQ and RBF and evaluate the accuracy of the ranking of these models. The obtained results indicate that MOE model is more accurate compared to MLP and RBF models and LVQ has not acceptable accuracy for ranking the credits of bank applicants.

Keywords

• credit ranking
• neural network
• multi-layer perceptron (mlp)
• mixture of experts (moe)
• radial basis function (rbf)
• learning quantization (lvq)