基于BP-LVQ的组合神经网络舞弊风险识别模型研究

发布时间：2018-11-20 16:07

【摘要】：近年来,国内外上市公司舞弊丑闻层出不穷,给投资者带来巨大的投资风险和伤害的同时,沉重打击了社会公众对会计界和资本市场的信心。因此,如何有效识别企业舞弊行为成为会计理论界、实务界以及监管部门关注的重中之重。实证研究表明,模型舞弊识别效果优于舞弊案例分析,而有效的舞弊风险识别模型的构建离不开完善的舞弊识别指标和恰当的识别方法。目前,在舞弊识别指标方面的研究已经比较完善,但舞弊识别模型方面的研究较少。随着人工智能技术的不断发展和广泛应用,人工神经网络技术开始应用于舞弊识别领域。其中,以BP和LVQ神经网络在舞弊识别领域的应用最为广泛,舞弊识别率较高。本文在此背景下深入探究这两种神经网络技术,用同一舞弊样本检验这两种模型的舞弊识别率,并在此基础上提出优化的基于BP-LVQ的组合神经网络舞弊风险识别模型。本文查阅整理国内外相关文献后,在第二章文献综述部分简单阐述了国际上最为流行的六种管理舞弊动机与成因理论以及国内流行的舞弊动机与成因观点,梳理归纳了舞弊风险识别指标和舞弊风险识别方法的相关国内外文献资料,明确舞弊风险识别模型的研究现状、研究成果和现有的不足之处。在此基础上提出本文选用BP、LVQ神经网络技术作为舞弊风险识别模型的理由,并在第三章详细介绍了人工神经网络技术的特点与分类、BP神经网络和LVQ神经网络的结构及运作机制。第四章主要是样本选取和舞弊风险识别指标筛选。本文选取2010年到2014年发生舞弊的506家上市公司作为舞弊样本,按照Beasley原则一比一确定非舞弊的配对样本公司506家,以此作为研究样本。将根据文献梳理出的识别效果较好的舞弊风险识别指标作为最初的指标体系,通过配对样本T检验以及主成分分析消除共线性问题后,最终删选出识别效果最好的10个指标。第五章主要对BP、LVQ神经网络模型的舞弊风险识别效果进行检验,并对两种识别模型的舞弊判别效果进行分析。第六章在分析BP、LVQ神经网络模型各自优缺点的基础上提出构建基于BP-LVQ的组合神经网络模型,介绍了组合模型的构建原理和思路,用同一研究样本检验组合模型的舞弊识别效果,发现组合模型的舞弊判别率显著高于单个神经网络模型,用2015年舞弊样本数据进行稳定性检验后发现,组合模型的舞弊识别效果更好而且判别率稳定。最后,根据上文理论分析和实证研究总结全文,分析本次研究中的不足之处,并对未来舞弊风险识别模型的发展提出展望。
[Abstract]:In recent years, fraud scandals of listed companies at home and abroad have emerged one after another, which bring huge investment risks and injuries to investors, and at the same time, have dealt a heavy blow to the public's confidence in accounting and capital markets. Therefore, how to effectively identify corporate fraud has become the most important concern of accounting theory, practice and regulatory authorities. Empirical research shows that the effectiveness of fraud identification model is better than that of fraud case analysis, and the effective fraud risk identification model can not be constructed without perfect fraud identification index and appropriate identification method. At present, the research on fraud identification index has been perfect, but the research on fraud identification model is less. With the development and wide application of artificial intelligence technology, artificial neural network technology has been applied to the field of fraud recognition. Among them, BP and LVQ neural networks are the most widely used in the field of fraud identification, fraud recognition rate is high. In this context, this paper explores the two neural network technologies, and uses the same fraud sample to test the fraud recognition rate of the two models, and on this basis, an optimized combined neural network fraud risk identification model based on BP-LVQ is proposed. In the second chapter of literature review, the author briefly expounds the six most popular theories of management fraud motivation and cause, and the domestic popular theories of fraud motivation and cause of formation. Combing and summarizing the relevant domestic and foreign literature data of fraud risk identification index and fraud risk identification method, clarifying the research status of fraud risk identification model, research results and existing deficiencies. On this basis, the reason for choosing BP,LVQ neural network technology as the fraud risk identification model is put forward. In chapter 3, the characteristics and classification of artificial neural network technology are introduced in detail. The structure and operation mechanism of BP neural network and LVQ neural network. The fourth chapter is mainly about sample selection and fraud risk identification index selection. In this paper, 506 listed companies with fraud from 2010 to 2014 are selected as fraud samples, and 506 non-fraudulent matched sample companies are determined according to Beasley principle. In this paper, the fraud risk identification index, which has good identification effect according to the literature, is taken as the initial index system. After eliminating the collinearity problem by paired sample T test and principal component analysis, the best 10 indexes are selected. Chapter 5 mainly tests the effect of fraud risk identification of BP,LVQ neural network model, and analyzes the fraud discrimination effect of two kinds of identification models. In chapter 6, on the basis of analyzing the advantages and disadvantages of BP,LVQ neural network model, the author proposes to construct the combined neural network model based on BP-LVQ, and introduces the construction principle and train of thought of the combined neural network model. Using the same sample to test the fraud identification effect of the combined model, it is found that the fraud discrimination rate of the combined model is significantly higher than that of the single neural network model. The combined model has better effect of fraud identification and stable discriminant rate. Finally, according to the theoretical analysis and empirical research summarized the full text, analyzes the shortcomings of this study, and puts forward the future development of fraud risk identification model.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F275;F832.51

【参考文献】