Application Of Sparse Group Lasso Method In Personal Credit Risk Evaluation

YWith the rapid development of the Internet finance and other industries,people's consumption concepts have undergone tremendous changes,and credit consumption has become a trend,at the same time,more and more Internet finance companies have developed and launched their own personal credit consumer products,such as Jingdong white bars,wayward pay(Suning Tesco),ant Ant Check Later(Alibaba)and so on.China is an economic power with a population of 1.4 billion,thus forming a huge credit market.According to the "2017 China Consumer Finance Innovation Report",China's annual consumption credit amount by 2020 can reach an astonishing RMB 12 trillion,which will be the largest consumer finance market in the world.Since China's Internet finance is still in its infancy and the credit risk management model is not yet mature,how to properly evaluate the credit of customers applying for loans is the key for major financial institutions to avoid risks.And establishing a scientific personal credit risk assessment model can help financial institutions effectively avoid potential risks.There are many factors that may affect the results of personal credit risk assessment,such as age,nationality,family income,academic qualifications,etc.Correctly selecting the main influencing factors as variables is a prerequisite for improving model prediction accuracy and enhancing model interpretability.Currently,the most commonly used methods for variable selection are the optimal subset selection method and the coefficient shrinkage method.The lasso method can be used as a representative of the coefficient shrinkage method,and it can continuously perform variable selection.It also obtains the parameter estimation while completing the variable selection process.So it can effectively overcome the disadvantages of optimal subset selection method such as instability when selecting variables.However,the lasso method cannot deal with variable selection problems with group effects,the group lasso method can only perform group-level variable selection and cannot perform intra-group variable selection.To solve this problem,Friedman integrated the sparse group lasso method,and this method can not only complete the variable selection process between groups,but also make the variables in the group sparse.In this thesis,we first introduce the two main methods used to solve the variable selection problem,the optimal subset selection method and the coefficient shrinkage method.Then we study the sparse group lasso method in the framework of linear regression model and logistic regression framework,and compared with the logistic regression model based on the lasso method and the group lasso method respectively by making the simulation experiment to demonstrate the superiority of the sparse group lasso method in model selection.Finally,we applied the sparse lasso-logistic model to personal credit data,by analyzing the empirical results,we show that sparse lasso-logistic model has good performance in personal credit risk prediction.