Özet:
In this thesis, credit scoring ability of several machine learning techniques were investigated such as multi-layer perceptron (MLP), radial basis function (RBF), knearest neighbor (k-NN) and support vector machines (SVM). Statistical technique logistic regression was also used for the purpose of comparison. In the second part, a two layer cascading model methodology called SVM-Reject was proposed which does not classify the instances under a threshold with its first layer model SVM. Experiments were performed by using on German credit data set and model comparisons are based on accuracy, error cost and ROC Analysis. Results show that SVM is a good option for credit scoring applications and SVM-Reject is the most accurate model. In the last part of this study, PD (probability of default) model building by using machine learning techniques was discussed in a comparative manner with logistic regression. This part is mostly from a bank's point of view and includes practical information as well.
