Identification Analysis Of The Closure Of Mini,Small And Medium-sized Enterprises

The construction of government big data platform provides data support for analysis of mini,small and medium-sized enterprises.With the deepening of big data,the scope of identification analysis can be extended to small or even mini-enterprises.According to standard of the Bureau of Statistics,we extracted the mini,small and medium-sized enterprises.Through quantitative analysis,we get the reasons for the closure of enterprises;the distribution of company in the dimensions of time,space and industry;enterprise financing situation and credit status;production and operation conditions.Because the Logit model is sensitive to multicollinearity,it is necessary to reduce the dimension of the database.In this paper,the nonparametric kernel smoothing method is used to select variables.The bandwidth of each variable is selected by Cross Validation.The local linear least-squares method is used to confirm whether the mechanism of the correlation variables is linear or non-linear.Since X and Y are nonlinearly related,we use semi-parametric Logit model.It contains both parametric and non-parametric components,on the one hand,the deterministic trend is embodyed from parametric components,On the other hand,non-linear terms are used to show the relationship between the dependent variable is not clear or the independent variables of interest,which combines the parameters III and non-parametric model of the advantages of both the full use of data in the information,but also to prevent the model set error.For the unknown function of semi-parametric Logit model,we use the Penalty Spline estimation method.The smoothing parameters are determined by Generalized Cross Validation.And the predicted error rate is less than 5%.The semi-parametric model is compared with the traditional Logit model on error rate and ROC.The results show the semi-parametric model is superior to the traditional model.For the stability test of the semi-parametric Logit model,We set five sample sizes and each perform 100 predictions,The conclusion is as follows:the highest error rate and the highest error rate of the model have strong stability.