Credit risk scoring with combined approach logistic regression and symbolic regression

Credit risk holds a special place among all different types of risks. Lack of attention to this risk may cause bankruptcy and waste of resources. Therefore, the development of models to differentiate different types of banking system customers is eminent in both theoretical and practical fields. This study predicts real customer credit risk of Qavamin Bank branches of Shiraz based on a logistic-symbolic regression approach.

In order to accomplish the purpose and objectives of the research, this study employed a survey research and correlation method. The statistical population of the study consisted of all real customers who received facilities from the branches of Qavamin Bank of Shiraz. Based on the population, the size of sample was calculated using Cochran formula. Then, the size of sample was estimated at 384 (number of facilities). Furthermore, the 351 cases were identified between 2011- 2017 using stratified random sampling method. The C5 criterion was used to select the independent variables. In addition, 17 explanatory variables including financial and nonfinancial variables were identified to classify customers as good (credit worthy) and bad (non-credit worthy), using a questionnaire. Finally, 5 variables influencing credit risk were selected from parent variables using forward parent selection technique. For this purpose, symbolic regression with 4 genes as well as optimum cut-off point was used to select the optimal cut-off point.

The results showed that among the 17 independent variables, the mean monthly income variables, the number of checks payable, bank debt history, account lifetime and type of collateral had the most significant effect on the dependent variable. Also, it was found that the accuracy of the hybrid logistic-symbolic regression model was 0.88 among good customers and 0.83 among bad customers.

accuracy of the results of logistic-symbolic regression model in the credit rating of real customers is better than the logistic regression model. Furthermore, the prediction accuracy of the hybrid logistic-symbolic regression model is adequate and satisfying.