1、上海交通大学 硕士学位论文 基于人工神经网络的虚假财务报表警示系统 姓名：陆金国 申请学位级别：硕士 专业：工商管理（MBA） 指导教师：徐博艺 20041017 NEURAL NETWORK BASED FALSE FINANCIAL STATEMENTS SPOTTING SYSTEM ABSTRACT This paper is about a computer assisted false financial statements of listed companies spotting system with artificial neural networks (ANN). The
2、 emphasis is to use data mining technology on financial analysis. ANN is one of the most representative methods of AI. Its application is very popular in classical engineering field such as pattern recognition, signal procession, control system and etc. Recently it is also very active in some Wester
3、n countries in social science area, such as banking, insurance and other business administration areas. An artificial neural network is set up through a large number of samples, so it has a remarkable learning ability due to its characteristics of self-organization, self-adaptation and self-learning
4、. ANNs are good at processing large volume data, especially suitable for dealing with multi-factor projects with vague, imperfect information. Spotting and recognizing the false financial statements through financial ratio analysis is technologically feasible. Usually there are a lot of contradictio
5、ns and doubtful points in a false financial statement. Most cheating companies are suffering from cash deficiency. These companies are eager to defraud of listing qualification through making up profits, or to seek exorbitant profit through handling stock second market price of company. And because
6、those false profits lack support of corresponding cash, they usually appear in various forms of “soft assets” such as the account receivable, inventory, construction in process, prepaid expenses, and etc. The enormous amount involved in false financial statements offered technological feasibility of
7、 quantitative methods of financial statement assisting detection system. In summary, the listed company which made up false profits in the financial statements usually has the following uncommon characteristics. a) The growing rate of account receivable is much higher than sales growing rate. The ac
8、count receivable recycling period and rate are very low b) Account receivable accounts for a very large proportion of the total assets. c) Inventory grows far faster than sale, selling cost and account payable assets. The proportion of inventory value to the total assets is much higher than the indu
9、stry average level d) Major product margin is unusually higher than industry average level. e) The cash flow fails to catch up with the pace of sale revenue. f) Fixed asset value rises sharply comparing to total asset value. g) Selling cost drops suddenly comparing to selling revenue h) Proportion o
10、f money and monetary equivalent to the total asset value is lower than industry average level. i) Total corporation loan grows, while the interest expenses drops. The input factors of the artificial neural network based financial shenanigans detection system are mainly selected according to the abov
11、e-mentioned characteristic points found in false financial statements of listed companies. The system input factors are divided into three parts of static ratios, companys historical data, and industry average ratios. A static ratio consists of the basic data items of the financial statements in cur
12、rent year, including total assets, core business income, core business cost, fixed assets, inventory, monetary fund, account receivable, account payable, construction in process, operational cash flow and prepaid expenses, and etc. It is used in the static analysis. Companys historical data consist
13、of companys last three years basic data items, and are used in trend analysis. An industry average ratio is used in comparison with other company in the same industry. The developing tool of the project is iData Analyzer (iDA) software package included in the book Data Mining - A Tutorial Based Prim
14、er (by Richard J. Roiger and Michael W. Geatz). It is an add-on component upon Microsoft Excel, it could take the full advantage of the extremely abundant powerful functions and convenient interface of data manipulation and preprocess. Its demonstration edition is very suitable for prototype develop
15、ment of a data mining project. Artificial neural network false financial statement spotting system has been a preliminary success project already, though it has only very limited training and testing data set. It needs a lot of time and efforts to make it a better system. KEY WORDS: artificial neural network, data mining, artificial intelligence, financial statement analysis, information technology 1 2004 10 17 2 2004 10 17 2004 10 17 MBA 1 1 1 ” ” ( 6) 1-1 7 MBA 2 MBA MBA MBA 3 GAL IDEA (Backpropagation Algorithm, BP ) BP MBA 4 1 2 (3 BP (4) BP 4.1 iDA MBA 5 BP
