Comparison Between Neural Network, Genetic Algorithm and Logit Models in Evaluating Consumer Credit Risk

The purpose of this study is to assess the credit rating methods of real customers (micro-credit recipients) of banks, by reviewing the financial records and characteristics of the applicant's characteristics. In this research, the effectiveness of some methods (logit model, neural network, and genetic algorithm) is evaluated for accurate measurement of the Defaults. For this purpose, the information and financial and qualitative data of a random sample of 399 customers who have received facilities during the years 1387 to 1391 have been investigated. After reviewing the credit records of each of the customers, 12 explanatory variables were identified which, based on the logit test variables, credit history, six-month average account, employment status, amount of credit, monthly installments and repayment period, had a significant effect on default. The results of the evaluation of credit rating methods indicate that the performance of the neural network is much better than the Genetic and Logit models because the sensitivity is 82.92% and the specificity is 76.92%, and in general, this model has been able to 80% Predict default or non-default. Therefore, in order to reduce the bank's credit risk, it is suggested that a structural adjustment based on the creation of a customer validation system based on the neural network is proposed.