Detection Of Fraudulent Financial Information Of Listed Company Based On Data Mining Techniques

Finance fraud of campanies is an international difficult problem with a long history. The British South-Sea Company Event happened in 1720 was the first case of Finance fraud. The event resulted in the close of stock company for nearly 100 years and the appearance of public independent audit composed of certified public accountants mainly. But the phenomenon of finance fraud never disappeared. The far finance fraud case is McKesson & Robbins event which resulted in the collapse of American stock market. The nearer one is the Enron event. In China, scandals about the finance fraud of Yinguangxia, ST Limin, Lantian Stock were disclosed just after Qiongminyuan and PT Zhengbaiwen events. The finance fraud problem is concerned by lots of people. Some researchers make a lot of qualitative or quantitative researches and get some valuable conclusions. This paper makes a review about current research from fraud theory, fraud characteristic and fraud detection model. On the base of the review, the paper makes the following research.First, this paper analyzes the reasons and means of finance fraud. The reasons can be classified as the external and the internal. External reasons include:the problems of government function, the limitation of management structure of company, incompleteness of stock market policy, inadequate monitoring system. Internal reasons or motives include:to get the qualification of stock issue, to increase the price of stock issue, to get the qualification of stock rations, to avoid special treating or cancel the special treating, to maintain or increase the price of stock, to get the credit funds or credit standing, to improve the achievements assessment, to satisfy the qualification of bond issue and to decrease the tax. The means of making FFS is maily adjusting the profit by connected transation, assets reorganization, selecting unsuitable accounting policies, changing accounting policies and accounting estimates, assets evaluation, etc.Second, this paper analyzes the choice of recognition variables and experimental samples. We choose the listed companies which were operated continuously and not canceled the qualification in the market during 1999-2002 as the base of training sample and the companies during 2003-2006 as the base of testing sample. We select the financial statements which are offered the audit opnion of denial or disclaimer by the CPA as the base of choosing the fraudulent financial statements (FFS). And we select the control sample according some specific standards considering the characteristic of Chinese stock market. On the base of prior research, we select 47 financial indexes as the candidates of recognition variables including business solvency and management efficiency ratios, profit and cash creating ratioes, expense rationality ratios, business growth ratios, business development durative ratios.Third, this paper designs a detection framework of FFS based on classification according to the characters of most classification methods and financial statements. Then it makes an empirical analysis in four methods including logistic regression, neural networks, support vector machine and naive bayesian classification according to the framework. During the empirical analysis, we use train sample and some recognition variables to training the model, then use testing sample to test the classification results of different classification model. After the comparing of the results of the four models, we find that neural netork and naive bayesian classification is better than the others on the accuracy rate and stability based on expected misclassification cost.Fourth, a clustering method is proposed to carry out the clutering analysis of FFS. In the current context of audit, we can not insure the correctness of class label of the cases in training sample just based on the audit opinion of CPA, especially when the financial statements is recognized as true. So there may exist some errors in the classification result of the supervised model. Because clustering is an unsupervised data mining technique, namely it need not know the class labels of the cases in the sample, it may provide more help in the detection of FFS when we can not get convictive training sample.In this paper, we design a clustering model V-KSOM combining SOM and K-means clustering based on a cluster validity measure Silhouette index. The model inherits the advantage (unsupervised self-learning) of SOM, and applies k-means clustering to the results of SOM avoiding one of the disadvantages (unclear clustering boundaries of nodes) of SOM. Besides, the clustering results are not consistent each time due to the influence of the initial value of nodes and inputting order of cases, so the model applies Silhouette index to measure the validity of different clustering results. We apply the V-KSOM model to the experimental data and experimental results show that the clustering model is effective.