Study On The Application Of Combined Voting Prediction Model To Personal Credit Evaluation

personal credit evaluation in commercial bank has been not only the urgent problem need to improve in China's current economic situation, but also the focus of academic research. At present, China's commercial banks have been practicing some of the credit risk assessment approach, but due to the constraints of many real problems, such as lacking of a sound personal credit database, and index system confusion, and commercial banks into the system for each individual credit assessment, Lacking of uniform standards, etc., these problems are seriously hampering a scientific and reasonable credit rating system established. In previous studies, many scholars have discussed the combined forecast model, and also done a lot of empirical research in the area. They found that combined model is superior in performance than single prediction models.But in recent years, the development of combined model has slowed down, which was mainly due to the determination of weight . The existence of weights makes interaction between a single model, if you can not get the weight of scientific and rational, it will make the model in the combined loss of accuracy in the process, and affect the final prediction.Based on this situation, an important issue of the National Natural Foundation named "study on the nonlinear combined model and its dynamic optimization to personal credit scoring " proposed the establishment of non-linear and dynamic optimization prediction model. As a part of this project ,this paper proposed a combined voting model based on Bayesian algorithm, and chose three models which are most popular and used more in the area of credit evaluation in domestic and international, and are currently recognized as mature credit single models, they are Logistic regression, cluster analysis and neural networks, with their vote results as input of the combination. Since the combination method was established after the decision-making of single models, every single model has the same voting rights, and they are independently of each other, which give full play to the advantages of a single model, and reducing the error caused by weight determination. And because the results based on Bayesian posterior probability classification, taking advantage of prior knowledge, so it can amend prior probability, and make sure the decision-making is more scientific and credible. By empirical findings, the combination is really better than single models, indicating the combined model has obvious performance advantages.