یک مدل محاسباتی برای تشخیص کلاهبرداری گزارش های مالی

A computational fraud detection model (CFDM) was proposed for detecting fraud in financial reporting. CFDM uses a quantitative approach on textual data. It incorporates techniques that use essentially all of information contained in the textual data for fraud detection. Extant work provides a foundation for detecting deception in high and low synchronicity computer-mediated communication (CMC). CFDM provides an analytical method that has the potential for automation. It was tested on the Management's Discussion and Analysis from 10-K filings and was able to distinguish fraudulent filings from non-fraudulent ones. CFDM can serve as a screening tool where deception is suspected.

Corporate fraud has not been confined to the well-advertized cases of Enron, WorldCom, HealthSouth, etc. In all the AAERs we examined, the fraud detection was after years of abuse by senior management; and the Securities and Exchange Commission (SEC) did not detect it proactively. The continued pattern of fraud has shaken the confidence of the public in corporate America [28] both academics and auditing firms have been searching for ways to detect corporate fraud. While academic fraud research has examined many business areas [34], very little effort has been made to use quantitative approaches to examine textual data for automated financial reporting fraud detection. Phua et al. summarized the status of fraud research into four primary areas: internal, insurance, credit card, and telecommunications. In most of the internal fraud research, the object was to detect employee fraud or theft; financial reporting fraud involving senior management was not a major research focus. Most attempts to detect financial reporting fraud use financial ratios, applying various methodologies with varying results [6], [28], [30] and [31]. Phua et al. concluded that the use of unstructured data in fraud detection is essentially unexplored. This paper proposes a quantitative model for detecting fraudulent financial reporting. The model detects the attempt to conceal information and/or present incorrect information in annual filings with the US Securities and Exchange Commission (SEC). The model uses essentially all of the information contained in a text document for fraud detection. A consistent and accurate screening tool would provide decision support for early detection of fraud; and hopefully, early detection will provide a deterrent to the commission of fraud. In order to detect fraud, we must first define it. This would seem an easy task, but it is not always as straightforward as finding the dictionary definition [1]. We use the SEC's issuance of an Accounting and Auditing Enforcement Release (AAER) as a starting point for defining financial reporting fraud. An AAER is an administrative proceeding or litigations release that entails an accounting or auditing related violation of the securities laws as enforced by the Securities and Exchange Commission (SEC). In the period from 2000 to 2008, the SEC issued 1700 AAERs. In the period from 2006 to 2008, they issued 555 AAERs. In this analysis, we use the term fraud when referring to an AAER to be a litigation release of an accounting or auditing violation where the SEC used the word fraud in describing the violation. We examined a sample of 74 AAERs from this period that charged companies with fraud. They showed that the average time between the identified initial fraud and the SEC filing charges was 7.26 years with a range of 3.6 to 11.6 years. The SEC charged that these companies committed fraud for an average of 4.1 years with a range of 1 to 12 years. Churyk et al. [13] used qualitative content analysis of the required Management's Discussion and Analysis (MDA) part of the 10-K SEC filings to identify fraudulent filings. They were able to identify deceptive cues. SEC filings include several areas of text in addition to the MDA. All companies are required to explain anything in the operation that could have a significant impact on the future profitability of the company. Management's explanation is in the MDA; the accountant's explanation is in the notes to the consolidated financial statements. It has been proposed that if a company were to report information that would have potentially negative results on the company valuation, they would include it in the notes. The reason to include the information is that it is legally required, and failure to do so is a criminal act that could result in a long incarceration. Both fraud and failure to properly report are crimes; but reporting gives some protection to the accountants and auditors, which leaves only senior management at risk. The idea that information can be concealed in either the MDA or in the notes, as it can be said “in plain sight”, leads to the problem statement and subsequent research questions.

The CFDM demonstrates that it is possible to detect financial reporting fraud from the text of annual filings with the Security and Exchange Commission. The model is generalizable because it specifies automatable steps that can be adapted to other domains and genres. A potential application for CFDM is to screen companies for investigation of potential fraud by the SEC. This would allow effective use of SEC resources. Additional potential applications include investor analysis, e-mail spam detection, and business intelligence validation. This work can be criticized for limiting the training data set to sixty-nine companies. While the statistical power is over 90% for this sample size, further confirmation of the discriminatory power of the CFDM by increasing the sample size would be an area for future research. The domain of this work is limited to MDA in 10-K filings. We do not argue that it is directly generalizable to other text submissions, but that other domains are a fertile area for research using CFDM. The notes to the financial statements in the 10-K are a possible area to extend CFDM without changing the domain. The model opens several additional research areas: (1) deception detection in e-mail and other computer-mediated communication; (2) deception in business-to-consumer websites and in consumer-to-consumer websites; (3) increasing the understanding of the mechanisms present in asynchronous text deception. Further work analyzing the singular value decomposition and the respective document and text vectors may lead to improved understanding of the mechanisms present in deceptive communications. It would be interesting to see if the CFDM has the potential to be a tool for decision support through evaluation of the degree of the veracity in unstructured data provided by a business intelligence system.