The Construction Of The Credit Index System And The Optimization Of The Evaluation Model In The SMEs

As an important component of the national economy,the SMEs(Small and Medium Enterprises)play a significant role in increasing employment,promoting economic growth,accelerating technological innovation,and maintaining social stability and harmony.However,as a result of their relatively small scales,weak capital chains and imperfect products,their market is unstable and their profit is full of uncertainty.All these give rise to the difficulties in financing,which is the biggest bottleneck restricting the development of the SMEs.Therefore,in order to get rid of the dilemma,it is necessary for banks to improve the credit evaluation system and measurement models of the SMEs.In this paper,for the sake of finding out the most proper system and model for the SMEs,I consider from the different angles,then build up the different measurement models,and eventually compare and analyze all these models.Firstly,I select the companies of the SME board from Wind database and take them as samples.Then considering from the firm characteristics,profitability,solvency,cash convertibility,operational capability,and development capability,I build up the credit evaluation system of 29 indicators.And I process the data on the basis of the time factor.Secondly,I guess there may be some correlations among the indicators.So,I reduce the dimensions of the indicators on the basis of PCA and KPCA.Then I find that the correlation between the indicators is mainly the linear correlation,and there is less nonlinear information among the indicators.The result of the KPCA is worse than the result of PC A,so I extract a new index system on the basis of PCA.Thirdly,based on the new index system,I use the Smote algorithm to improve the unbalanced problem of the data set,and expand the samples to construct a new data set.Considering from the statistical angle,I use logistic regression and discriminant analysis to build up the measurement models,and then optimize the model for the latter.Based on the data set and index system in this paper,I find that logistic regression is not suitable,but QDA is on the contrary.Furthermore,the model of KNN has more balanced accuracies among categories and a lower total error rate,so it is more applicable.Fourthly,considering from the data mining,I use SVM,RF and ANN to build up the measurement models.And in each model,I adjust the parameters to the optimum.Based on the data set and index system in this paper,I find that SVM is not suitable,but RF is on the contrary,whose generalization ability is stronger and more robust.And ANN also shows a good classification effect,but the generalization ability is worse than RF.Finally,I summarize the conclusions of the paper,and point out the shortcomings and limitations.At last,I highlight the directions for further research.