تکنیک های داده کاوی برای کشف تقلب در صورت های مالی

This paper explores the effectiveness of Data Mining (DM) classification techniques in detecting firms that issue fraudulent financial statements (FFS) and deals with the identification of factors associated to FFS. In accomplishing the task of management fraud detection, auditors could be facilitated in their work by using Data Mining techniques. This study investigates the usefulness of Decision Trees, Neural Networks and Bayesian Belief Networks in the identification of fraudulent financial statements. The input vector is composed of ratios derived from financial statements. The three models are compared in terms of their performances.

Auditing nowadays has become an increasingly demanding task and there is much evidence that ‘book cooking’ accounting practices are widely applied. Koskivaara calls the year 2002, ‘the horrible year’, from a bookkeeping point of view and claims that manipulation is still ongoing (Koskivaara, 2004). Some estimates state that fraud costs US business more than $400 billion annually (Wells, 1997). Spathis, Doumpos, and Zopounidis (2002) claim that fraudulent financial statements have become increasingly frequent over the last few years. Management fraud can be defined as the deliberate fraud committed by management that causes damage to investors and creditors through material misleading financial statements. During the audit process, the auditors have to estimate the possibility of management fraud. The AICPA explicitly acknowledges the auditors’ responsibility for fraud detection (Cullinan & Sutton, 2002). In order to develop his/her expectations, the auditor employs analytical review techniques, which allow for the estimation of account balances without examining relevant individual transactions. Fraser, Hatherly, and Lin (1997) classify analytical review techniques as non-quantitative, simple quantitative and advance quantitative. Advance quantitative techniques include sophisticated methods derived from statistics and artificial intelligence, like Neural Networks and regression analysis. The detection of fraudulent financial statements, along with the qualification of financial statements, have recently been in the limelight in Greece because of the increase in the number of companies listed on the Athens Stock Exchange (and raising capital through public offerings) and the attempts to reduce taxation on profits. In Greece, the public has been consistent in its demand for fraudulent financial statements and qualified opinions as warning signs of business failure. There is an increasing demand for greater transparency, consistency and more information to be incorporated within financial statements (Spathis, Doumpos, & Zopounidis, 2003). Data Mining (DM) is an iterative process within which progress is defined by discovery, either through automatic or manual methods. DM is most useful in an exploratory analysis scenario in which there are no predetermined notions about what will constitute an “interesting” outcome (Kantardzic, 2002). The application of Data Mining techniques for financial classification is a fertile research area. Many law enforcement and special investigative units, whose mission it is to identify fraudulent activities, have also used Data Mining successfully. However, as opposed to other well-examined fields like bankruptcy prediction or financial distress, research on the application of DM techniques for the purpose of management fraud detection has been rather minimal (Calderon and Cheh, 2002, Koskivaara, 2004 and Kirkos and Manolopoulos, 2004). In this study, we carry out an in-depth examination of publicly available data from the financial statements of various firms in order to detect FFS by using Data Mining classification methods. The goal of this research is to identify the financial factors to be used by auditors in assessing the likelihood of FFS. One main objective is to introduce, apply, and evaluate the use of Data Mining methods in differentiating between fraud and non-fraud observations. The aim of this study is to contribute to the research related to the detection of management fraud by applying statistical and Artificial Intelligence (AI) Data Mining techniques, which operate over publicly available financial statement data. AI methods have the theoretical advantage that they do not impose arbitrary assumptions on the input variables. However, the reported results of AI methods slightly or occasionally outperform the results of the statistical methods. In this study, three Data Mining techniques are tested for their applicability in management fraud detection: Decision Trees, Neural Networks and Bayesian Belief Networks. The three methods are compared in terms of their predictive accuracy. The input data consists mainly of financial ratios derived from financial statements, i.e., balance sheets and income statements. The sample contains data from 76 Greek manufacturing companies. Relationships between input variables and classification outcomes are captured by the models and revealed. The paper proceeds as follows: Section 2 reviews relevant prior research. Section 3 provides an insight into the research methodology used. Section 4 describes the developed models and analyzes the results. Finally, Section 5 presents the concluding remarks.

5. Conclusions Auditing practices nowadays have to cope with an increasing number of management fraud cases. Data Mining techniques, which claim they have advanced classification and prediction capabilities, could facilitate auditors in accomplishing the task of management fraud detection. The aim of this study has been to investigate the usefulness and compare the performance of three Data Mining techniques in detecting fraudulent financial statements by using published financial data. The methods employed were Decision Trees, Neural Networks and Bayesian Belief Networks. The results obtained from the experiments agree with prior research results indicating that published financial statement data contains falsification indicators. Furthermore, a relatively small list of financial ratios largely determines the classification results. This knowledge, coupled with Data Mining algorithms, can provide models capable of achieving considerable classification accuracies. The present study contributes to auditing and accounting research by examining the suggested variables in order to identify those that can best discriminate cases of FFS. It also recommends certain variables from publicly available information to which auditors should be allocating additional audit time. The use of the proposed methodological framework could be of assistance to auditors, both internal and external, to taxation and other state authorities, individual and institutional, investors, the stock exchange, law firms, economic analysts, credit scoring agencies and to the banking system. For the auditing profession, the results of this study could be beneficial in helping to address its responsibility of detecting FFS. In terms of performance, the Bayesian Belief Network model achieved the best performance managing to correctly classify 90.3% of the validation sample in a 10-fold cross validation procedure. The accuracy rates of the Neural Network model and the Decision Tree model were 80% and 73.6%, respectively. The Type I error rate was lower for all models. The Bayesian Belief Network revealed dependencies between falsification and the ratios debt to equity, net profit to total assets, sales to total assets, working capital to total assets and Z score. Each of these ratios refers to a different aspect of a firm’s financial status, i.e., leverage, profitability, sales performance, solvency and financial distress, respectively. The Decision Tree model primarily associated falsification with financial distress, since it used Z score as a first level splitter. As usually happens, this study can be used as a stepping stone for further research. An important difference between the experiments is that the BBN model utilized discretized data due to software limitations. Data discretisation eliminates the effects of outliers at the cost of some information loss. Further research is required to address the topic of the impact of data discretisation on the models’ performance and the topic of optimal discretisation algorithms. Research is also needed to examine the circumstances under which DM methods perform better than other techniques. Our input vector solely consists of financial ratios. Enriching the input vector with qualitative information, such as previous auditors’ qualifications or the composition of the administrative board, could increase the accuracy rate. Furthermore a particular study of the industry could reveal specific indicators. We hope that the research presented in this paper will therefore stimulate additional work regarding these important topics.