Customer Credit Risk Assessment And System Design Based On Bayesian Network

Credit Risk, also named Default Risk, is one of the three primary risks, also the most important risk the bank may encounter. Credit Risk Assessment is the key process in the management of bank’s risks. The bank can take corresponding actions of escaping from, transferring or hedging the risks, to avoid the risks from getting worse into loss. With the development of information technology, the concept of information science is also adopted by the banks and used in the assessment of Credit Risk. The method is no more the pure mathematical statistics, but also includes the technologies of data mining and knowledge finding. The data mining method extends Probability Theory and Mathematical Statistics, combining mathematics and computer science. Bayes Network (BN) is a data mining technology which can make the classification and prediction. With the conditional probability in Probability Theory and the computing power of computers, BN displays people the global dependent relations with visualized graphics. Because BN is a "White-Box" model, it has a character of good explanation, and is easy to be accepted by people.Researches about BN include structure learning, parameter learning and BN inference. This paper introduces some usual technologies and methods about BN learning and inference systematically, including Bayes scoring, Likelihood scoring, orientation by collision, Gibbs Sampling and so on. Data mining system is based on data mining technology, including data pretreatment module, visualization module. This paper makes the BN model as the core to design a data mining system which can reflect the Prohibit of Default (PD), and the module of data pretreatment is designed in detail.In this paper, we firstly get the minimum set of indicators which reduces the time complexity and improves the learning accuracy based on rough set theory, and then we use the Greedy Searching (GS) strategy to find the BN model whose posterior probability is max. Then we learn the Conditional Probabilistic Table (CPT) on incomplete dataset with Expectation-Maximization (EM) method to achieve a BN reflecting customer’s PD. On the PDBN, we continue to get every customer’s PD with exact inference method. In the practical application, we classify the customers into different categories based on PD. In the end, we compare the results of PDBN classification with that of Logistic Regression Model and Neural Network (NN) which is commonly used in risk assessment areas, also we take loss matrix into consideration, and we can find that the BN method has the more highly rate of accuracy and better characteristic of explanation.