Optimal Feature Selection in order to Bank Customer Credit Risk Determination

Credit risk interprets as the probability of obligations non-repayment by customer in due date is considered as one of causes financial institutions bankruptcy. For this purpose, data mining techniques such as neural networks, Decision Tree, Bayesian networks, Support Vector Machine is used for customer segmentation to high-risk and low-risk groups. In this paper, we present the hybrid Imperialist Competitive optimization algorithm and neural network for increasing classification accuracy in evaluation and measurement credit risk of bank customers. The proposed method identifies the optimistic features and eliminates unnecessary features decreases problem dimension and increases classification accuracy. To validate this method, it implements on UCI dataset and also on a reality dataset of a private Iranian bank. The experimental results show this method is more satisfactory than other data mining techniques. The neural network error for the test set decreases with selection of effective features and elimination of low-impact features by the Binary Imperialist Competitive Optimization Algorithm. In addition test data error rate remains at acceptable level for other used classification methods. This article is the first use of algorithms Imperialist Competitive for credit risk assessment of bank customers.