Kurumsal Kredi Skorlamasında Klasik Yöntemlerle Yapay Sinir Ağı Karşılaştırması
Şahap KavcıoğluBankaların, müşterilerinin kredi değerliliğini doğru bir şekilde analiz etmemeleri yıkıcı sonuçlar doğurmaktadır. Bu nedenle, bankacılık sektöründe kredi skorlamasının önemi son yıllarda büyük bir araştırma alanı haline gelmiştir. Kredi değerliliğinin skorlanması için lojistik regresyon, doğrusal regresyon, diskriminant analizi ve yapay sinir ağları gibi yöntemler mevcuttur. Bu araştırmanın konusu makine öğrenmesi ve lojistik regresyon modellerinin kredi skorlaması modelindeki performanslarınnı kıyaslama yoluyla değerlendirmektir. Bu çalışma ile klasik yöntemlerle yapay sinir ağlarını karşılaştırarak, bankaların kredi riskine en az düzeyde maruz kalabilecekleri bir skorkart modeli geliştirilmesi amaçlanmıştır. Literatürde kredi skorlaması modellerinin kıyaslanmasına ilişkin çalışmalar mevcut olmakla birlikte, çalışmalar perakende portföyler üzerinden ve en fazla 4 yılı kapsayan bir örneklem üzerinden yapılmıştır. Araştırma literatürdeki çalışmalardan farklı olarak kurumsal firmalar üzerinden ve literatürdeki çalışmalara göre daha geniş bir örneklem üzerinden ele alınmıştır. Çalışma sonucunda geliştirme örnekleminde daha yüksek başarı sergileyen yapay sinir ağlarının, örneklem dışı veri seti üzerinde lojistik regresyondan daha düşük bir performans sergilediği görülmüştür. Böylece yapay sinir ağları yüksek performans gösterse de, lojistik regresyonun daha tutarlı sonuçlar verdiği bulgusuna ulaşılmakla birlikte yapay sinir ağlarının iterasyon süreçlerinde optimizasyon yapılması ile daha tutarlı sonuçlar üretebileceği düşünülmektedir.
A Comparison of the Artificial Neural Network with Classical Methods in Corporate Credit Scoring
Şahap KavcıoğluThe failure of banks to correctly analyze the credit worthiness of their customers has devastating consequences. Therefore, the importance of credit scoring in the banking sector has become a major field of research in recent years. There are some methods such as logistic regression, linear regression, discriminant analysis and artificial neural networks for credit scoring. The subject of this research is to evaluate the performance of machine learning and logistic regression models on credit scoring by comparison. In this study, it is aimed to develop a scorecard model in which banks can be exposed to a minimum level of credit risk by comparing the logistic regression and artificial neural network methods which are two of these methods. Although there are studies on the comparison of credit scoring models in the literature, the studies have been conducted through retail portfolios and a sample that covers a maximum of 4 years. Unlike the studies in the literature, this research was conducted through corporate firms and a larger sample than the studies in the literature. The result of the study indicated that artificial neural networks which have higher success than logistic regression on the development sample, saw lower success on the out of sample data. Thus, while artificial neural networks show higher performance, it is concluded that logistic regression provides more consistent results, and it is thought that artificial neural networks can produce more consistent results by optimization of the iteration processes.
In the banking sector, credit risk is one of the most important risk types that needs to be managed by banks. Banks have applied many different methods in order to measure credit risk. In this context, various statistical methods have been used in recent years in order to quickly and objectively measure the credit risk of customers. Regulators have set various standards and regulations for the use and spread of these statistical methods.
In the first part of the study, an overview of the traditional model methodologies is given and in the next part, general information about the artificial neural networks which is one of the machine learning methodologies is given. Linear, logistic regression and discriminant analysis methods used in credit scoring are mentioned on a general level. Artificial neural networks are discussed in detail, the general structure of the model is explained and artificial neural network classifications in the literature are mentioned.
The aim of this study is to develop a scorecard model in which banks can be exposed to a minimum level of credit risk. To this end, scorecard models were developed using logistic regression and artificial neural network methods, which are credit scoring methods, and the obtained model results were compared. In this study, 8 years of data which includes the financial information and repayment habits of companies operating in the manufacturing, service and trade sectors in the corporate segment between 2010 and 2017 was used. The models in the study were developed through SPSS Modeler program.
The modular structure was preferred for logistic regression and financial and behavioral modules were created using the stepwise logistic regression method with 5% margin of error. The integration of modules was made by the enter method of logistic regression, thus it was possible that both modules could take part in module integration.
Unlike the logistic regression method, the modular structure was not used while developping the artificial neural network method. The variables in two different modules used in logistic regression were combined and the artificial neural network model was established over this single variable list. In the study, 80% of the data was allocated as learning and 20% of the data was allocated as test in the whole data set. After various analyses, 18 financial and 11 behavioral input variables were used in the model. In addition, the model consisted of 3 layers: 1 input layer, 1 hidden layer and 1 output layer. Considering the model’s technical aspects, the Sigmoid function was used as an activation function. The learning coefficient and the momentum coefficient were determined automatically by SPSS Modeler.
In the last part, ROC curves and Gini coefficients were compared to evaluate the model developed by the logistic regression and artificial neural network methods. The evaluation of the models was carried out both by the development sample and the out of sample data. In the development sample, the artificial neural network model which performed with a 0.76 Gini value showed a higher success than the logistic regression model which performed with a 0.56 Gini value. However, it was observed that artificial neural networks performed with a 0.60 Gini value which is less than the logistic regression 0.69 Gini value on the out of sample data set that was not included in the modeling process. According to this result, logistic regression was more consistent. However, it is thought that artificial neural networks can produce more consistent processes with the optimization of iteration processes.
As a result of this study, it is foreseen that traditional methods such as logistic regression will continue to be preferred in credit scoring and adopted by regulation as they are more comprehensible and produce consistent results. It is thought that artificial neural networks, which is one of the systems based on machine learning, will be developed and used in both credit scoring and customer service related areas.
