Research On Selective Ensemble Alogirithm Models For Credit Scoring

As the economy develops, personal credit, credit economy and credit society were built up, which represent the scale effect. Banking business and profit from personal customer accumulate. Meanwhile, the default risks increase. When individuals apply for a loan from banks or financial institutions in the process of credit regulation, credit rationing issues arise, which caused by information asymmetry and adverse selection. The prime loans fall short of optimal distribution ratio. High-risk loans increase the risk of default probability. Therefore, it is important means for banks to process effective management of personal credit risk, recognize potential default, calculate and predict credit risk.Literature review of domestic and foreign personal credit default risk research reveal that credit risk management related with each steps of optimization. For example, to discover the reason of loans imbalance, to mechanically analyze multi-dimensional factors(ratio of index weight, sample distribution, sample size and method for credit scoring, etc.), to choose the reasonable algorithms for credit scoring, or to predict default risk. Based on that, credit default risk could be effectively managed, as well, banks increase income and profit. In such cases, how to optimize the credit system, how to choose a profit model have become the issues for academia and the field of finance.The process of credit scoring systematically introduced. There are several steps, including problem definition, theory of mechanism analysis, data acquisition and processing, establishing personal credit scoring model, model test, interpretation and application. First of all, credit rationing theory demonstrates the relationship of credit scoring and credit ratio. Some problems of information asymmetry, adverse selection and moral hazard are solved by optimization solutions, for instance, index weight, sample distribution, sample scale. Based on that, single model of personal credit scoring is built up. Model test proposed through accuracy, stability and adaptability. Definition of differentiation and selective ensemble algorithms are discussed and built up in credit system. Meanwhile, Three real world data from countries are presented and tested for empirical analysis.The main research is as follows.Firstly, the relations of credit scoring and credit ratio are discussed based on the theory of credit rationing. The theory are used in analysis of microcosmic mechanism and the mechanism of action for credit scoring. Information asymmetry, adverse selection and moral hazard cased credit rationing imbalance. From the view of the economics of supply and demand, the problems are solved by credit rationing theory in the process of personal credit loans. Through theoretical analysis of credit scoring optimization mechanism, the study lay a theoretical foundation.To solve the promlems of credit rationing, optimization of index system and sample structure is analyzed. Unbiased index system of personal credit scoring is build up. Structural equation model is set up, then the indicator variables of interaction coefficient are proposed through the model of the path coefficient. The indexes weights variables are given. The results of accuracy and stability are optimized through the method. Besides, sample size and sample distribution are optimized. Comparative analysis of models for credit scoring through different sample size and sample distribution are put forward. The resonable sample structure of size and distribution are calculated.Additionally, in pre-process of model setting up for credit scoring with selective ensemble algorithm, diversity measurement is proposed for better classification. Single model for credit scoring are processd by differentiation, through Q-statistics method. Four representative single method are selected from fifteen models for the selective ensemble algorithrm. Credit scoring model is established based on selective ensemble algorithm. Three models are built based on various technique, including GSEAN optimization algorithm, OO ranking method and FCM-CFP cluster model.Dataset is from China, German and Austria for selective ensemble model testing. Three evaluation criterions are put forward, concerning estimated performance, selection time and size of ensemble classifier. Consequently, the three methods possess advantages. OO and FCM-CFP methods have best estimated performance. Operation time is OO ranking method is shortest. Quantity of basic classifiers are least in GSEAN credit scoring.Finally, the comparisons of credit scoring models are proposed. Models with different technique are compared, between selective ensemble model, single model and ensemble model for credit scoring. In summary, selective ensemble model processes better accurate rate and more stability.