改进支持向量机在虚假财务报告识别中的应用  

Improved Support Vector Machine Algorithm for Fraudulent Financial Statements Detection

作　　者：阚宝奎[1] 刘志新[1] 宋晓东[1] 杨众[1] 

机构地区：[1]北京航空航天大学经济管理学院,北京100191

出　　处：《管理评论》2012年第5期144-153,共10页Management Review

基　　金：国家自然科学基金重点项目(70821061);第六届全国大学生创新创业训练计划项目

摘　　要：针对虚假财务报告识别实证研究的不足,本文提出了一种新的支持向量机方法对公司财务报告的真伪进行判别。第一,针对两类训练样本存在的"重叠"问题,建立双隶属支持向量机模型,通过基于谱聚类方法的隶属度模型来确定样本点对于两类样本的隶属程度;考虑到人们对于两种判别错误的"厌恶程度差异",在模型训练时,对训练样本进行了"非对称"处理,来降低虚假财务报告未识别的错误率。第二,在模型的输入财务指标选择上,本文通过浮动顺序搜索算法得到了虚假财务报告识别的全局最优财务指标组合。第三,实证结果表明,改进模型的判别准确率和泛化能力显著优于普通的支持向量机和BP神经网络,而公司治理指标的加入会提升模型的判别能力。Empirical researches for fraudulent financial statements still remain in the early stage. This paper develops an improved SVM algorithm for the detection of financial statements. First, to solve the ＂overlap＂ problem between two classes of training samples, we build a dual membership support vector machine model and use the membership model based on spectral clustering to calculate the membership value of each sample belonging to the two classes; considering the difference between people＇s aversion to the two types of classification error, we do the unbalanced classifier training to reduce the error rate of fraudulent financial statements detection. Second, in the selection of financial indicators, we use the floating sequential search model to obtain the global optimal portfolio of financial indicators. Third, the empirical results show that the discrimination accuracy and generalization ability of our improved model are significantly better than traditional support vector machine and BP neural network; furthermore, the addition of corporate governance indicators will increase the detection capabilities.

关 键 词：虚假财务报告 支持向量机 隶属度模型 非对称训练 指标组合选择 

分 类 号：F233[经济管理—会计学] F224[经济管理—国民经济]