Research On Ensemble Model For Credit Scoring

Due to the rapid development of China’s credit business, financial institutions credit risk management system is facing a severe test. In recent years, personal credit business has been an important part of the banks’ businesses. With the development of credit cards, personal car loans, loans for durable goods, housing mortgage loans, student loans and other consumer loans business, banks and other financial institutions have to be of the ability of high speed processing personal credit business. Personal credit scoring technology is of the characteristics such as standardization and quantization, and can fast and efficiently process credit business, which makes personal credit scoring methods have been widely studied.Although the credit scoring technology in foreign countries has well developed, it is still at the starting stage in china. There is not a uniform standard for financial institutions how to choose models and establish domestic index system because each financial institution has its own index system and how to use the system is strictly confidential. The research method of domestic personal credit scoring model is mainly divided into statistical and non statistical method, and main research areas include the accuracy and stability of model. The reason that many models can not be extended is bad unification of the model accuracy and stability. After analyzing the difference to choose personal credit score index systems between domestic and foreign institutions, and combining the real sample data, we establish appropriate indicators of credit scoring system. Then we select the commonly used model of personal’credit scoring, use sample data to empirically analyse each model, and compare the effect of each model prediction. The analysis shows that the prediction accuracy of Logistic regression model is not better than neural network model, but the stability of model is good; the prediction accuracy of neural network is the highest, but the model is not stable; the accuracy and stability of Bayes discriminant are good, but limitation of the model is more stringent; prediction accuracy and stability of the tree is the worst in the four models.In order to obtain high accuracy and stability of a good model, we first use Odds-matrix to choose a hybrid model, estimates weights coefficient of the model which is of better accuracy and stability than a single model. Compared with other methods, the hybrid model is not sensitive to the change of the odds ratio, is simple and easy to operate, and its weights can be updated at any time. For a large sample of data, the accuracy of Odds-Matrix method is more than the weight method, least square method and regression method.