基于数据挖掘技术的上市公司财务舞弊识别研究

发布时间：2018-04-23 18:59

本文选题：财务舞弊识别 + 数据挖掘　；参考：《杭州电子科技大学》2013年硕士论文

【摘要】：进入21世纪以来，世界各国的上市公司数量都在飞速增长。以A股为例，从2002年到2012年这十年间，A股的上市公司数量从1138家增长到2472家，翻了一倍。在这些成功上市或正准备上市的公司中，大多数都是希望通过上市的方式来获得更多的社会资源并最终以此提高自己在整个市场中的竞争力，但是也有部分公司因为各种各样的动机通过财务舞弊来谋取一些不正当的利益。这部分上市公司的财务舞弊行为对证券市场的各个参与者都有着重大影响。对于投资者来说，如果根据存在舞弊迹象的财务报告对相关上市公司做出了错误的投资，势必会造成损失。对于监管层来说，保证公开透明的证券市场环境是其本职工作，若对舞弊行为不加以制止，则会降低证券市场的资源配置效率甚至会造成证券市场的混乱。对于其他相关者，虚假的财务报告会使得利益相关者无法准确清晰的了解公司实际经营状况进而无法做出正确决策。正因如此，研究出能够有效识别上市公司财务舞弊行为的方法就变得很有实际意义。本文在对上市公司财务舞弊相关理论深入理解的基础上，综合利用数据挖掘技术对2000-2008年的有舞弊迹象的上市公司及相应的控制样本建立模型。本文的研究成果体现在两个方面，一方面是通过特征选择结果得出对上市公司舞弊行为影响最大的几个指标，从而根据这几大指标对如何防止舞弊提出相应的建议；另一方面是通过对备选模型的建模效果进行比较，找出最适合这一问题的模型。论文的思路如下：首先，阐述财务舞弊的概念和动因、数据挖掘的产生背景等相关理论，并对文章所使用的数据挖掘工具R语言进行了介绍。然后，结合财务舞弊动因理论，初选了22个与财务舞弊行为相关的指标。通过Relief和信息增益这两种特征选择算法对2000-2008年的舞弊公司与控制样本进行指标筛选。综合两种特征选择算法的结果，对财务舞弊行为影响最大的几个指标分别为资产负债率、资产报酬率、应收账款周转率、销售期间费用率和流动资产周转率。其次，经过Relief和信息增益两种特征选择算法筛选后的指标组合与未经指标筛选的原22个指标组合总共形成了三个数据集。将这三个数据集与最近邻、朴素贝叶斯和随机森林这3种具有代表性的数据挖掘算法相结合建立模型。在多种评价方法评价之后，发现预测效果最优的为经过信息增益算法筛选后的指标组合与最近邻算法结合的模型。再次，考虑到所得的信息增益与最近邻算法结合的模型在舞弊公司的查全率上依然不尽如人意。将代价敏感学习引入到模型组合中，并提出了一种基于代价敏感学习的随机森林算法。该算法在测试集的舞弊公司查全率上达到89%，大大提高了整个舞弊识别模型的效果。最后，结论部分从经过特征选择选择后的指标和模型应用两方面对全文进行了总结，，并对进一步规范我国上市公司的行为给出了一定的建议。
[Abstract]:Since twenty-first Century, the number of Listed Companies in the world has been growing rapidly. In the ten years from 2002 to 2012, for example, the number of Listed Companies in A shares has doubled from 1138 to 2472, doubling the number of companies that have successfully listed or are ready to go on the market, most of which are expected to get more by the way of listing. Social resources and ultimately improve their competitiveness in the whole market, but there are also some companies who make some unjustified interests through financial fraud for all kinds of motives. The financial fraud of this part of the listed company has a significant impact on the various participants in the stock market. For investors, if According to the financial reports of the signs of fraud, they make the wrong investment to the related listed companies. It is bound to cause loss. For the regulators, it is its job to ensure the open and transparent securities market environment. If the fraud is not stopped, it will reduce the efficiency of the resource allocation of the stock market and even cause the mix of the stock market. For others, false financial reports can make it impossible for the stakeholders to understand the actual situation of the company and make the right decision.
On the basis of in-depth understanding of the related theories of financial fraud of listed companies, this paper makes a comprehensive use of data mining technology to set up a model for 2000-2008 years of listed companies with signs of fraud and corresponding control samples. The research results of this paper are embodied in two aspects, on the one hand, through the results of characteristic selection, the fraud of the listed companies is obtained. In order to influence the most important indexes, the corresponding suggestions on how to prevent fraud are put forward according to these major indexes; on the other hand, by comparing the modeling effect of the alternative model, the model which is most suitable for this problem is found.
First, it expounds the concept and motivation of financial fraud, the background of data mining and other related theories, and introduces the data mining tool R language used in the article.
Then, combined with the theory of financial fraud motivation, 22 indexes related to financial fraud are selected. Through the two characteristic selection algorithms of Relief and information gain, 2000-2008 years of fraud companies and control samples are selected. The results of comprehensive two feature selection algorithms have the greatest impact on financial frauds. It should not be assets liability ratio, assets return rate, accounts receivable turnover rate, sales period cost rate and current assets turnover rate.
Secondly, a total of three data sets are formed by the combination of two feature selection algorithms selected by the Relief and information gain selection algorithm and the original 22 index combinations without index selection. The three data sets are combined with the nearest neighbor, simple Bias and random forest, which are combined with 3 representative data mining algorithms to establish a model. After the evaluation of the price method, it is found that the best prediction result is the combination of the index combination and the nearest neighbor algorithm after the information gain algorithm is selected.
Thirdly, considering the information gain combined with the nearest neighbor algorithm is still unsatisfactory in the recall company's recall rate, the cost sensitive learning is introduced into the model combination, and a stochastic forest algorithm based on cost sensitive learning is proposed. The algorithm has reached 89% on the recall rate of the fraud company of the test set, which is greatly raised. The effect of the entire fraud identification model is higher.
Finally, the conclusion is summed up in the conclusion of the full text from two parties after the selection of features and models, and gives some suggestions for further standardizing the behavior of the listed companies in China.

【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】：2013
【分类号】：F275;F224;F276.6

【相似文献】