It has been initially mentioned in this research that because of increasing diversity of credit institutions and increasing fields of activity and their relationship with people, these institutions require updated and useful methods to manage their credit risk in a scientific method. Research questions include how to measure credit risk of the customers and methods of credit risk measurement. General concepts on credit risk and methods of credit rating have been expressed and have been selected by using genetic algorithm of existing, limited data, and 11 ultimate features. Then the results have been expressed in tables and diagrams in the results section by putting the data in two Mamdani and Sugeno neuro-fuzzy networks, and training network in Matlab software. These results indicate that firstly, rate of interest and facilities has the greatest effect on customers’ credit risk measurement and whatever the number of features is fewer, network accuracy is higher. They also indicate that Sugeno neuro-fuzzy network has more accuracy despite closeness of the results of the two networks’ analysis.
risk, credit risk measurement, neuro-fuzzy network, Mamdani structure, Sugeno structure
Today credit industry plays an important role in the countries’ economy. Increase of credit demand has created more competition for the creditor institutions. Banks are considered as one of the fundamental pillars of this system. When a bank can develop a desirable system to evaluate and assess specific credits of each customer can be considered successful (Gupta, 2000). Parametric, nonparametric, and data mining methods can be mentioned among credit rating methods (Alborzi, Pourzandi, and Khanbabaee, 2010).
Fuzzy model, support vector, and hybrid model of neural networks and genetic algorithm can be also mentioned among hybrid methods (Janu et al, 2011; Kuzni, 2015). Each of these methods has advantages and disadvantages compared to each other. For example, the old methods such as statistical logistic regression are not suitable for large and small inputs. This assumption is often considered that the relationship between variables is linear, while
Nonlinear relationship is practically between variables in most cases (Khales and Shekuhi, 2010).
Neuro-fuzzy networks method has been used in this research to examine all aspects of the existing methods. Simplicity and intelligibility of fuzzy systems are great advantages of them that make them very simple to use and understand. Approximate inference capability and adjustable accuracy in dealing with the problems are considered of other their advantages. But these systems do not have learning capability, so they encounter with problem when facing with complex systems (Zhang et al., 1997).
The aim of conducting this research is to provide an appropriate model for measuring credit risk of customers as well as identifying and assorting the credit characteristics of customers by using genetic algorithm.
- We see when a genetic algorithm with binary chromosome is used for feature selection that this algorithm has selected the main feature and applies it to sorter network. We can found from the percentage of accuracy obtained for training and testing the network that facilities type and the interest rate features have a high frequency in terms of repetition. Since inputs are used again in the result of Sugeno structure and its number of training parameters is more than Mamdani structure, so it has relatively better performance.
- Revision of conventional techniques of customers’ credit assessment, which are based on subjective judgments, seems essential due to increasing expansion of variety and number of facilities provided by banks, as well as intense competition among financial institutions to attract customers. So, turning to the mentioned advantages in this study, implementation of credit rating models in banking system make faster, more accurate and with lower cost decision-making possible and meet also increasing trend of credit demand.
- Neural networks do not require any distribution assumptions for input data. For this reason they can be used for a wider range of problems compared to statistical techniques and regression and discriminant analysis.
- Neural networks have this property to allow new data to be added to training neural network, and to update the previous trained resources, while new and old data in statistical techniques should be entered together into a single class to yield new results.
- These networks are an estimator without the need for model. This feature allows them to be able to record data interactive effect on each other without need to formulate an explicit model
270 total views, 1 views today