Abstract-Since the granting of banking facilities in recent years has faced problems such as customer credit risk and affects the profitability directly, customer credit risk assessment has become imperative for banks and it is used to distinguish good applicants from those who will probably default on repayments. In credit risk assessment, a score is assigned to each customer then by comparing it with the cut-off point score which distinguishes two classes of the applicants, customers are classified into two credit statuses either a good or bad applicant. Regarding good performance and their ability of classification, generalization and learning patterns, Multilayer Perceptron Neural Network model trained using various Back-Propagation (BP) algorithms considered in designing an evaluation model in this study. The BP algorithms, Levenberg-Marquardt (LM), Gradient descent, Conjugate gradient, Resilient, BFGS Quasinewton, and One-step secant were utilized. Each of these six networks runs and trains for different numbers of neurons within their hidden layer. Mean squared error (MSE) is used as a criterion to specify optimum number of neurons in the hidden layer. The results showed that LM algorithm converges faster to the network and achieves better performance than the other algorithms. At last, by comparing classification performance of neural network with a number of classification algorithms such as Logistic Regression and Decision Tree, the neural network model outperformed the others in customer credit risk assessment. In credit models, because the cost that Type II error rate imposes to the model is too high, therefore, Receiver Operating Characteristic curve is used to find appropriate cut-off point for a model that in addition to high Accuracy, has lower Type II error rate.

An interesting tool for non-linear multivariable modeling is the Artificial Neural Network (ANN) which has been developed recently. The use of ANN has been proved to be a cost-effective technique. It is very important to choose a suitable algorithm for training a neural network. Generally Backpropagation (BP) algorithm is used to train the neural network. While these algorithms prove to be very effective and robust in training many types of network structures, they suffer from certain disadvantages such as easy entrapment in a local minimum and very slow convergence. In this paper, to improve the performance of ANN, the adjustment of network weights using Particle Swarm Optimization (PSO) was proposed as a mechanism and the results obtained were compared with various BP algorithms such as LevenbergMarquardt and gradient descent algorithms. Each of these networks runs and trains for different learning ratios, activation functions and numbers of neurons within their hidden layer. Among different criteria Mean Square Error (MSE) and Accuracy are the main selected criteria used for evaluating both models. Also the MSE was used as a criterion to specify optimum number of neurons in hidden layer. The results showed that PSO approach outperforms the BP for training neural network models.