مقايسه‌ي مدل‌هاي مختلف شبكه‌ي عصبي در رتبه‌بندي اعتباري سيستم بانكي و معرفي بهترين مدل (1390-1385)

Comparison of neural network models in the credit rating of the banking system and providing the best )optimal( model

در نگاهي ساده، حوزه‌ي فعاليت بانك‌ها تجهيز و تخصيص منابع است. لذا، بانك‌ها با در نظر گرفتن ريسك اعتباري مشتريان، به تقاضاهاي آن‌ها مبني بر اخذ تسهيلات جامه‌ي عمل مي‌پوشانند. اين پژوهش، با هدف انتخاب متغيرهاي اثرگذار و مدل بهينه، به‌منظور رتبه‌بندي اعتباري مشتريان بانكي با استفاده از مدل‌هاي شبكه‌هاي عصبي با الگوريتم پس‌انتشار خطا، شبكه‌هاي عصبي GMDH، شبكه‌هاي عصبي با الگوريتم شعاع‌محور، مدل‌هاي لاجيت، پروبيت و تحليل مميزي ارايه شده است. لذا 200 نفر از مشتريان حقيقي يكي از بانك‌هاي دولتي در فواصل سال‌هاي 1390-1385 انتخاب شده‌‎اند كه از اين تعداد 105 نفر خوش‌حساب و 95 نفر از مشتريان بدحساب بوده‌اند. در مرحله‌ي اول 9 متغير به‌‌عنوان متغيرهاي بي‌اثر در وضعيت اعتباري مشتريان تشخيص داده شد كه 5 متغير حذف شدند. نهايتاً مقايسه‌ي اين مدل‌ها با يكديگر نشان داد كه شبكه‌هاي عصبي با الگوريتم شعاع‌محور و شبكه‌هاي عصبي GMDH بالاترين دقت را در پيش‌بيني رفتار اعتباري مشتريان دارد.

In simple terms, banks operate in two areas of equipment and allocation of resources. Meanwhile, by taking into account the credit risk of customers, banks provide customer demands based on their requested facilities. One of the most important problem of management of loan portfolio is bankruptcy and bank failure. So, one of the most important techniques of financial and banking sections, conspicuously noticed, is the technique of risk management. With the aim of selecting the optimal model and effective variables to rank customers credit, this study is presented by the models used in this research including Neural Networks back propagation the error, neural network algorithms, neural networking "GMDH", neural network algorithm with radius axis, "Logit" model, "Probit" model, and discriminate analysis model. This paper analyzes the internal and external factors influencing credit risk of “Ayandeh bank”. For this purpose, 200 cases of actual customers of the state-owned banks were selected during seasonal intervals of 2006-2011 (1385-1390( to provide an effective strategy for reducing the risk and help to improve the implementation of the decision-making in “Ayandeh bank”. In this data, we have 200 customers whom 105 of them were “creditworthy customers” and 95 of them were “uncreditworthy customers”. In the first phase, 9 variables were recognized as ineffective and five of them were removed. Finally, the comparison of these models show that neural network algorithms and neural network-centric radius "GMDH" have the highest accuracy in predicting the credit behavior of banking customers.