A股上市公司财务造假预测模型研究

RESEARCH ON A-SHARE LISTED COMPANIES' FINANCIAL FRAUD

陶能发

11849092

硕士

金融

陈琨

2020-05-28

2020-06-30

哈尔滨工业大学

深圳

资本市场的良好运行离不开公开、透明和真实的信息披露，而财务造假层出不穷、手段多样，对资本市场的投资者带来了巨大的损失，也阻碍了资本市场的健康发展。如何通过公司特征提前判断出可能进行财务造假的公司就成为了实务界、资本市场监管机构和学术界共同关心的重要问题。为了建立财务造假预测模型，本文搜集了中国上市公司违规统计数据中，违规内容包含“虚增”、“虚减”、“收入”、“成本”、“利润”、“年报”和“年度报告”的公司，之后通过逐个检查，共获得 93 家在 2001 年至 2016 年对年报进行财务造假的公司。在进数据清理之后，共获得 175 个造假年样本及 608 个行业相对应的非造假年样本。本文首先将 Mscore 与 Fscore 模型提到的变量应用于本文的数据，并进行Logistic 回归，从回归结果中发现，Mscore 与 Fscore 模型对中国财务造假的预测结果不佳。其原因在于模型提出时间较早，且主要应用于美股，而中国 A 股市场与美股之间存在着差异。因此在后续研究中，本文深入研究了 93 家公司财务造假的原因，并从非财务数据出发，选取了年报文字数据、股票交易数据、年报发布时间、公司管理层变动情况等数据，总结成 12 个财务造假预测变量用于预测中国 A 股市场的财务造假。研究结果表明，本文提出的模型相对 Mscore 模型和 Fscore 模型更加准确。其中研报个数、发布时间差、涨跌停次数和投资比率与财务造假正向相关，年报相似度与财务造假负向相关。论文的主要创新点在于使用了一些较为新颖的变量用于预测财务造假，同时在模型回归过程中调整了财务造假与非财务造假的权重以消除造假与非造假样本之间的数量不相等所造成的影响。在变量的选取过程中，本文没有过多地使用财务数据，变量的选取也是经过的长时间的研究与思考所做出的决定，它们都没有跟公司的财务数据有直接的联系，但是却与公司的经营状况紧密相关，甚至相对于财务数据来说更加贴近公司的真实情况。

The regular operation of the capital market is inseparable from the open, transparent and true information disclosure, and financial fraud emerges in endlessly, with various means, which brings huge losses to investors in the capital market, and also hinders the healthy development of the capital market. It has become an important issue of common concern for the practitioners, capital market regulators and academia that how to judge the companies that may make financial fraud in advance through the characteristics of companies. In this paper, we collect the statistical data of Listed Companies in China, which include the companies of "virtual increase", "virtual decrease", "income", "cost", "profit", "annual report" and "annual report". This paper then get 93 companies that made financial fraud in the annual report from 2001 to 2016. After data cleaning, a total of 175 yearly fraud samples and 608 industry corresponding yearly non fraud samples were obtained. Firstly, this paper applies the variables mentioned in Mscore and Fscore models to the data of this paper, and carries out logistic regression. From the regression results, it is found that Mscore and Fscore models have poor prediction results for financial fraud in China. The reason is that the model was put forward earlier and was mainly applied to American stocks, while there are differences between Chinese A-share market and American stocks. Therefore, in the follow-up study, this paper deeply studies the causes of financial fraud of 93 companies, and selects the annual report text data, stock trading data, annual report release time, company management changes and other alternative data, and summarizes them to 12 financial fraud prediction variables to predict the financial fraud of China's A-share market. The results show that this paper’s model is more accurate than Mscore model and Fscore model. Among the variables, the number of research reports, the time difference between publication, the limit down times and the investment ratio are positively correlated with financial fraud, and the similarity of annual reports and limit up times is negatively correlated with financial fraud. The main innovation of this paper is that some new variables are used to predict financial fraud. At the same time, the weights of financial fraud and non-financial fraud are adjusted in the process of model regression to eliminate the effect of unequal quantity between fraud and non fraud samples. In the process of variable selection, this paper does not use too many financial variables. Variable selection is also a decision made after a long time of research and thinking. They are not directly related to the company's financial data, but they are closely related to the company's operating conditions, even closer to the real situation of the company compared with the financial data.

财务造假 虚增虚减 文本相似度 预测模型

financial fraud virtual increase and decrease text similarity predicting model

中文

联合培养

南方科技大学

陶能发. A股上市公司财务造假预测模型研究[D]. 深圳. 哈尔滨工业大学,2020.