基于反向传播算法神经网络的信用评分系统预测力研究  

Study on Prediction Ability of Credit Scoring System Based on Neural Network Using Back Propagating Algorithm

作　　者：朱晓明[1] 程建[1] 刘治国[1] 钟经樊[1] 

机构地区：[1]西安交通大学金禾经济研究中心,西安710049

出　　处：《西安交通大学学报》2006年第12期1405-1409,共5页Journal of Xi'an Jiaotong University

摘　　要：为了提高信用评分系统的预测准确性和稳定性,建立了基于反向传播(BP)算法神经网络的信用评分系统,并提出信用评分系统预测力和预测稳定性验证的新方法.结合信用评分问题的实际特点建立了模型并确定了参数,然后采用一种正向选入法确定输入变量,进行模型训练,并通过引入接收器操作特征曲线的分析理论、曲线面积(AUC)值及信息理论等评价方式,对所构造的神经网络信用评分系统预测力进行评价,最后利用自抽样法构造出多个验证样本来评估信用评分系统的稳定性.与传统的逻辑信用评分系统的比较结果表明,BP神经网络信用评分系统具有更高的预测准确性和稳定性,其AUC值平均提高0.036 7,AUC值的标准误差平均降低0.005.A credit scoring system based on BP（back propagation） neural network model is pres ented in order to improve the accuracy and stability of the prediction. A new method for evalua ting the accuracy and stability of the system is also presented. Combining with the features of credit scoring problems, a neural network model is established and the parameters are deter mined. Then the input variables are decided by a forward selecting method to train the model. The prediction ability of the proposed credit scoring system is valuated through introducing theory of ROC （receiver operating characteristics） curves, AUC（area under curve） value and information theory, and compared to the traditional logistic credit scoring system. Finally, the checking samples are created by means of bootstrap method to validate the stability of the system. The ex perirnent results show that the credit scoring system based on BP neural network is more accurate and stable than the logistic credit scoring system, and averagely, the AUC value is increased by 0. 036 7 and its standard error is decreased by 0. 005.

关 键 词：神经网络 反向传播算法 信用评分 曲线面积值 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术] F244.1[自动化与计算机技术—计算机科学与技术]