基于BP神经网络的商业银行信用风险模型改进探究  

A Study on Improvement of Commercial Banks' BP Neural Network-based Credit Risk Management Model

作　　者：宿玉海[1] 彭雷[1] 郭胜川[1] 

机构地区：[1]山东财经大学金融学院,山东济南250014

出　　处：《山东财政学院学报》2012年第2期12-19,共8页Journal of Shandong Finance Institute

基　　金：国家社科基金资助项目"企业金融衍生业务风险测度及管控研究"(10BGJ054)

摘　　要：BP神经网络信用风险管理模型以其较强的逼近非线性函数的能力而适应商业银行信用风险管理的要求,但是其自身的权值调整方式存在的缺陷影响了模型的应用。分别采用Adaboost算法和遗传算法对BP神经网络信用风险模型进行了改进,通过对200家上市公司的财务数据指标进行考察,比较了两种模型的优劣:经过Adaboost算法改进后,模型可以平稳地达到系统判别的最小误差,但运行时间较长;遗传算法采用变异操作可以迅速达到系统判别的最小误差,但由于权值改变过于激烈,可能造成系统过于注重权值的改变而忽视了原始数据指标的特性。With its relatively strong ability to approximate nonlinear functions, the BP neural networkbased credit risk management model is up to the credit risk management demands of commercial banks. However, there exist defects with the weight adjusting manner of the model itself, which have hindered its application. In this paper, the author improved the BP neural networkbased credit risk management model by using Adaboost algorithm and Ge netic algorithm. Through an examination of the financial data of 200 listed companies, the author compared the pros and cons of the model after using the two algorithms： After being improved with Adaboost algorithm, the model can steadily reach the minimum error of system identification, only that its running time is longer; while with Genetic algorithm＇ s mutation operation, the model can quickly reach the minimum error of system identification, but because the change of the weights is too sharp, the system may put too much emphasis on the change of the weights to neglect the features of the original data index.

关 键 词：BP神经网络 ADABOOST算法 遗传算法 

分 类 号：F832.332[经济管理—金融学]