Credit Rating Based On Support Vector Machine

Finance plays a key role in the modern economy.It is important for promoting the operation of the economic and social system.Moreover,risk management is the core of all financial businesses.With the development of non-governmental financial institutes and the innovation of financial tools,the fraud risk has become more and more series.As noted,the best way to solve this problem is to accurately assess the credit risk of the customers.Credit rating is a primary tool for credit risk assessment.Essentially,for a customer,it uses a mathematical model to calculate his/her insolvency and default loss rate,then classifies him/her into different risk levels.The classification accuracy has a very important impact on the stability of the financial market.When the classification is inaccurate,it may mislead the judgment of the user and cause a large-scale financial crisis.At the same time,credit rating can alleviate the asymmetry of market information,reduce the transaction cost and improve the ability of identifying the default risk for the Chinese market.Therefore,it is very important for the financial institutes to construct a credit rating system.Nowadays,with the rapid development of the Internet,the data resources are extremely rich and huge.Big data is considered as an important factor to promote the economic growth.Based on these big data resources,the new tendency of risk assessment is to establish and optimize the new models through novel technologies,and apply efficient algorithms and flexible strategies to improve the assessing accuracy.In recent years,the government has released several documents to prevent and control financial risks.It encourages the financial institute to establish and improve their financial risk monitoring models through new big data technologies.As noted,traditional credit risk rating applies various methods,such as scorecard model,linear regression,logistic regression and so on.All these methods have achieved some good results and effectively promoted the development of this area.In recent years,as a popular machine learning method,Support Vector Machine(SVM)has been widely used and attracted a lot of attention because of its excellent performance.SVM applies the idea of structural optimization to deal with the binary classification.It is well-known for its ability to overcome the classical difficulties of those traditional machine learning models such as "dimension disaster" and "over-fitting".Traditional SVM models have to introduce the kernel functions into their models to deal with the nonlinearity in the data.The performances of these SVM models largely depend on the choices of the kernel functions and their corresponding parameters.For a certain data set,there is no general rule to guide the users how to select the proper kernel function and its parameters.Therefore,to get a good model,users need to spend a lot of time to check what kind of kernel function and its parameters are more suitable.In order to overcome this problem,we apply the most cutting-edge research results – non-kernel Quadratic Surface Support Vector Machine(QSSVM)for the credit risk rating in our thesis.It is worth pointing out that the new model directly generates a quadratic surface in the original space to do the classification.Hence,it skips the difficult task of choosing the kernels.In this thesis,we take the data from a real small loan company.Then we establish a credit index system according to the credit risk characteristics of individual customer.Moreover,we apply SVM models to dig the relationship between credit index and default status.Besides,we also discuss about how to select the effective features of the data and how to measure the classification accuracy.Then we conduct our numerical experiment based on the sample data of more than 90,000 loans from this small loan company.In the numerical experiment,to comprehensively check the performance of our model,we use three methods in the comparison which includes the logistic regression,traditional support vector machine and non-kernel quadratic surface support vector machine.The experimental results definitely show that our model achieves the best performance both in accuracy and stability.Therefore,our model demonstrates a big potential in some real applications in future.