Analysis Of Credit Rating Based On A New RBF Neural Network

Credit rating has been regarded as a core appraisal tool in credit risk management during the last few decades, in this paper we will force on the application of radial basis function neural network (RBFNN) in credit rating. In the area of credit rating, the most important data is the profile of the borrowers, so most of the data is categorical data, however, the traditional RBFNN just treat these data type as continual data an so as to get a pool result in credit rating; furthermore, the traditional RBFNN is sensitive to the initial clustering center, which makes the model get a pool robustness.To overcome the problems above, we develop a new RBFNN based on the fuzzy K-Prototypes cluster algorithm; we also improve the quality of initial clustering centers. The new RBFNN can deal with categorical as well as continual data, and it is less sensitive to the noisy data. The result on the credit data indicates that the new RBFNN has a higher accuracy and robustness.The structure of the thesis is as follows.In character1, we will give a brief introduction to the credit rating, ranging from the background to the definition of credit, and a comparison of traditional credit rating method and advanced credit rating model based on statistical models.In character2, we will introduce some common statistical models used in the construction of credit models, which contains linear regression model, generalized linear models, discriminant analysis, k-nearest neighbors, etc.In character3, a brief introduction of radial basis function neural network is arranged, which contains the fundamental mathematical knowledge of RBFNN, the principle in physiology of RBFNN, and the learning algorithm for each part of RBFNN.In character4, we introduce the fuzzy K-Prototypes clustering algorithm and try to develop a new RBFNN based on it, we also develop a new method for selecting the initial clustering center.Character5is an empirical analysis based on the individual credit rating data, a comparison of different models is given in this part, and character6is a conclusion for the paper.