基于自适应LASSO变量选择的Logistic信用评分模型研究

发布时间：2018-09-09 10:28

【摘要】：本文主要研究的是信用评分模型中的变量选择问题,变量选择不仅是建模过程中需要考虑的问题,也是统计学研究中很重要的问题之一。在建模中不仅要确保选择的变量集合能反映所要研究的对象特性,还要保证得到的结果具有较好的拟合性和预测性。本文主要工作如下：一系统介绍了变量选择的方法：子集选择法和系数压缩法。子集选择法包括AIC值、BIC值和逐步回归法,系数压缩法包括岭回归、LASSO以及自适应LASSO.分析了信用评分模型的发展,变量选择在信用评分模型建立中的关键作用,以及国内外研究现状。二研究了Logistic模型的自适应LASSO变量选择和参数估计,核心是讨论将变量的信息价值作为自适应LASSO的惩罚项权重,以及证明其得到的参数估计量具有Oracle性质。由此构建了基于信息权重的Logistic信用评分模型。三进行了实证分析,主要通过对深交所148家中小企业以及30个信用指标体系变量进行以信息价值为惩罚项权重的自适应LASSO变量选择,建立Logistic信用评分模型,并利用发生比率来解释估计量的含义,并对选择的变量进行经济意义检验,可知模型包含的变量基本涵盖了企业的各项能力指标。最后并通过对比逐步回归法和LASSO两种变量选择法建立的logistic信用评分模型,利用ROC曲线、KS值、WGRP和CIER指标来比较验证三种模型的风险区分能力和估计准确性,得到自适应LASSO建立的Logistic模型要好于其他两种方法得到的模型,也表明了包含信息价值的自适应LASSO变量选择方法在实证检验中具有很好的表现。
[Abstract]:In this paper, the problem of variable selection in credit scoring model is mainly studied. Variable selection is not only a problem to be considered in the modeling process, but also one of the most important problems in statistical research. In modeling, it is necessary not only to ensure that the selected set of variables can reflect the characteristics of the object to be studied, but also to ensure that the obtained results have good fitting and predictability. The main work of this paper is as follows: firstly, the methods of variable selection, subset selection and coefficient compression are introduced systematically. Subset selection includes AIC and stepwise regression, coefficient compression includes ridge regression and adaptive LASSO.. This paper analyzes the development of credit scoring model, the key role of variable selection in the establishment of credit scoring model, and the current research situation at home and abroad. Secondly, the adaptive LASSO variable selection and parameter estimation of Logistic model are studied. The key is to consider the information value of the variable as the weight of the penalty term of adaptive LASSO, and to prove that the parameter estimator has Oracle property. Based on this, the Logistic credit scoring model based on information weight is constructed. Thirdly, through the empirical analysis, this paper establishes the Logistic credit scoring model by selecting the adaptive LASSO variables with the weight of information value as punishment item, mainly through 148 SMEs and 30 credit index system variables of Shenzhen Stock Exchange. Using the ratio of occurrence to explain the meaning of the estimator and to test the economic significance of the selected variables, it can be seen that the variables contained in the model basically cover the various capability indicators of the enterprise. Finally, by comparing the logistic credit rating model established by stepwise regression method and LASSO variable selection method, using ROC curve KS value and CIER index to compare and verify the risk differentiation ability and estimation accuracy of the three models. The Logistic model established by adaptive LASSO is better than that obtained by the other two methods. It also shows that the adaptive LASSO variable selection method with information value has a good performance in the empirical test.
【学位授予单位】：东南大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：F832.4

【相似文献】