To refer to fraudulent techniques that are generally accepted, here the ten fraudulent techniques from (Beasley et al. 1999) are used. That is, there are three basic types of fraudulent techniques: Improper Revenue Recognition, Overstatement
of Assets, and Others. Improper Revenue Recognition includes recording fictitious
revenues (FT1), recording revenues prematurely (FT2), and no description/overstated revenues (FT3). Overstatement of Assets includes overstating existing assets (FT4), recording fictitious assets or assets not owned (FT5), and capitalizing items that should be expensed (FT6). Others includes understatement of expenses/liabilities (FT7), misappropriation of assets (FT8), inappropriate disclosure (FT9), and other miscellaneous techniques (FT10).For demonstration purposes, we take merely the three leaf nodes, #11, #14-21, and
#14-24 to illustrate the parts of uncovering the regularity of fraudulent techniques
10
from the corresponding indictments and sentences for major securities crimes issued by the Department of Justice. Table 6 summarizes the fraudulent techniques
commonly adopted by companies clustered in these three leaf nodes. The code and year in the first two column of Table 6 lists the company code and the year of each clustered financial statement.
<Insert Table 6 here>
As shown in Table 6, common fraudulent techniques found in leaf node #11 are FT1, FT6 and FT8; in leaf node #14-24 are FT1, FT4 and FT8; and in leaf node
#14-21 are FT4 and FT8. Note that leaf nodes #14-24 and #14-21 have same parent (leaf node #14) and they share a certain similarity in common fraudulent techniques.
In sum, Table 6 shows that the observed common fraudulent techniques in different leaf nodes are distinctive even though samples are clustered based upon corporate financial situations proxied by input variables (i.e., the eight variables identified from discriminant analysis).
Compared to the traditional fraudulent technique classification scheme, such a contrast demonstrates the advantage of our approach since our classification outcomes appear to be more delicate. For instance, some fraud samples in leaf node #11 were found using FT1 via creating fictitious transactions and defrauding export drawbacks from the Internal Revenue Service by reporting fictitious export sales. Moreover, some fraud samples used FT8 by processing the receipt and payment in advance. In contrast, some fraud samples in leaf node #14-24 were found to have been using FT4 through purchasing intangible asset/long-term investment with high premiums. Some fraud samples used FT8 through related party transactions and merger and acquisition activities to misappropriate cash.
Conclusion
In the data preprocessing stage, a sample set comprised of 113 fraud samples and 467 non-fraud samples is used to identify eight significant variables regarding FFR via the discriminant analysis. Based upon the (identified) variables as inputs, GHSOM clusters 113 fraud samples into 13 (small-sized) leaf nodes. Distinguishing this study from others of feature extraction is that, for each leaf node, common fraud techniques are disclosed with the assistance of expert knowledge in examining corresponding
11
FFR indictments and sentences (exogenous information) of clustered samples without referring to the attributes of input variables. With acknowledging that different leaf nodes have distinctive common fraudulent techniques, the study confirms that
GHSOM can provide implicitly a relationship between common fraud techniques (an exogenous variable) and input variables. To go further to uncover the relationship between common fraud techniques and input variables is one of future works. The study also demonstrates the abilities of GHSOM to (1) extract features from exogenous information that are more abundant and more informative than input variables and (2) classify exogenous variables in terms of input variables.
The exogenous abilities of GHSOM can contribute to the FFR literature at least as follows. Exogenous FFR features uncovered in each leaf node describe the observed regularity of corporate behavior in that subgroup and are applicable to all samples clustered in that leaf node. For each leaf node, this principle and any pre-warning signals provided by exogenous FFR features can result in some FFR audit guideline.
For instance, with the assistance of experts with domain knowledge on common fraudulent techniques, we can identify the financial indicators revealing the potential fraudulent activities as pre-warning signals. Take leaf node #14-21 as an illustration.
The expert with domain knowledge on common fraudulent techniques summarizes the primary causes of utilizing both FT4 and FT8 fraudulent techniques as the bad cash flow condition of the firms and high financial pressure from management and derive the relevant pre-warning signals shown in Table 7. When a new sample is imported into the obtained GHSOM and the distance deviation between the input vector and the weight vector of leaf node #14-21 is less than a predefined threshold, it is assigned as the potential (fraud) member of that subgroup. Based upon the assignment, the relevant pre-warning signals shown in Table 7 can help auditors perform prudent audit planning and audit judgment.
<Insert Table 7 here>
In addition, with distinctive exogenous FFR features extracted from different leaf nodes and tons of leaf nodes, a further analysis of associations between (exogenous and endogenous) FFR features and corresponding clustered samples can provide more insights of FFR.
12
Other future works are suggested as follows: (1) to refine the GHSOM to get a better classification mechanism or to identify better ways in extracting exogenous FFR features from the outcomes of GHSOM; (2) to investigate the generality of our approach using data from other countries; and (3) to examine the prediction ability extended from the study.
13
References
ACFE. 2008. 2008 Report to the nation on occupational fraud & abuse, Association of Certified Fraud Examiners, Austin, TX.
Altman, E.I. 1968. "Financial ratios, discriminant analysis and the prediction of corporate bankruptcy,"
Journal of Finance (23:4), pp 589-609.
Beasley, M.S., Carcello, J.V., and Hermanson, D.R. 1999. Fraudulent financial reporting: 1987-1997 COSO, New York.
Bell , T.B., and Carcello, J.V. 2000. "A decision aid for assessing the likelihood of fraudulent financial reporting," Auditing: A Journal of Practice & Theory (19), pp 169-184.
Dechow, P.M., Ge, W., Larson, C.R., Sloan, R.G., and Investors, B.G. 2007. "Predicting material accounting manipulations," AAA 2007 Financial Accounting and Reporting Section (FARS) Paper [Electronic Version] (1001), December 13, 2007 p48109. Retrieved December 13, 2007 from http://ssrn.com/abstract=997483.
Dechow, P.M., Sloan, R.G., and Sweeney, A.P. 1996. "Causes and Consequences of Earnings Manipulation: An Analysis of Firms Subject to Enforcement Actions by the SEC," Contemporary accounting research (13:1), pp 1-36.
Dittenbach, M., Merkl, D., and Rauber, A. 2000. "The growing hierarchical self-organizing map,"
Proceedings of the IEEE-INNS-ENNS International Joint Conference on Neural Networks, IJCNN, pp. 15-19.
Fanning, K.M., and Cogger, K.O. 1998. "Neural network detection of management fraud using published financial data," Intelligent Systems in Accounting, Finance & Management (7:1), pp 21-41.
Green, B.P., and Choi, J.H. 1997. "Assessing the risk of management fraud through neural network technology," Auditing (16), pp 14-28.
Hoogs, B., Kiehl, T., Lacomb, C., and Senturk, D. 2007. "A genetic algorithm approach to detecting temporal patterns indicative of financial statement fraud," Intelligent Systems in Accounting, Finance & Management (15:1-2), pp 41-56.
Kirkos, E., Spathis, C., and Manolopoulos, Y. 2007. "Data Mining techniques for the detection of fraudulent financial statements," Expert Systems with Applications (32:4), pp 995-1003.
Kohonen, T. 1982. "Self-organized formation of topologically correct feature maps," Biological cybernetics (43:1), pp 59-69.
La Porta, R., Lopez-de-Silanes, F., Shleifer, A., and Vishny, R. 1999. "Corporate ownership around the world," Journal of Finance (54:2), pp 471-517.
Lee, T.S., and Yeh, Y.H. 2004. "Corporate governance and financial distress: evidence from Taiwan,"
Corporate Governance: An International Review (12:3), pp 378-388.
Ngai, E.W.T., Hu, Y. , Wong , Y.H., Chen, Y., and Sun, X. 2010. "The application of data mining techniques in financial fraud detection: A classification framework and an academic review of
14
literature", Decision Support Systems, In Press, Accepted Manuscript, Available online 19 August 2010.
Persons, O.S. 1995. "Using financial statement data to identify factors associated with fraudulent financial reporting," Journal of Applied Business Research (11), pp 38-46.
Rauber, A., Merkl, D., and Dittenbach, M. 2002. "The growing hierarchical self-organizing map:
exploratory analysis of high-dimensional data," IEEE Transactions on Neural Networks (13:6), pp 1331-1341.
Rumelhart, D.E., and Zipser, D. 1985. "Feature discovery by competitive learning*," Cognitive Science (9:1), pp 75-112.
Schweighofer, E., Rauber, A., and Dittenbach, M. 2001. "Automatic text representation, classification and labeling in European law," ACM, p. 87.
Séverin, E. 2010. "Self organizing maps in corporate finance: Quantitative and qualitative analysis of debt and leasing," Neurocomputing (73:10-12), pp. 2061-2067
Shih, J., Chang, Y., and Chen, W. 2008. "Using GHSOM to construct legal maps for Taiwan's securities and futures markets," Expert Systems with Applications (34:2), pp 850-858.
Soriano-Asensi, A., Martin-Guerrero, J., Soria-Olivas, E., Palomares, A., Magdalena-Benedito, R., and Serrano-Lopez, A. 2008. "Web mining based on Growing Hierarchical Self-Organizing Maps:
Analysis of a real citizen web portal," Expert Systems with Applications (34:4), pp 2988-2994.
Stice, J.D. 1991. "Using financial and market information to identify pre-engagement factors associated with lawsuits against auditors," Accounting Review (66:3), pp 516-533.
Summers , S.L., and Sweeney, J.T. 1998. "Fraudulently misstated financial statements and insider trading: An empirical analysis," Accounting Review (73:1), pp 131-146.
Virdhagriswaran, S., and Dakin, G. 2006. "Camouflaged fraud detection in domains with complex relationships," ACM, pp. 941-947.
Yeh, Y., Lee, T., and Woidtke, T. 2001. "Family control and corporate governance: Evidence from Taiwan," International Review of finance (2:1 2), pp 21-48.
15
Table 1: Research methodology and findings in nature-related FFR studies.
Research Methodology Findings
(Beasley et al.
1999)
• Case study
• Descriptive statistics
• Nature of companies involved
– Companies committing financial statement fraud were relatively small.
– Companies committing the fraud were inclined to
experience net losses or close to break-even positions in periods before the fraud.
• Nature of the control environment
– Top senior executives were frequently involved.
– Most audit committees only met about once a year or the company had no audit committee.
• Nature of the frauds
– Cumulative amounts of fraud were relatively large in light of the relatively small sizes of the companies involved.
– Most frauds were not isolated to a single fiscal period.
– Typical financial statement fraud techniques involved the overstatement of revenues and assets.
• Consequences for the company and individuals involved – Severe consequences awaited companies committing
fraud.
– Consequences associated with financial statement fraud were severe for individuals allegedly involved.
(ACFE 2008) • Case study
• Descriptive statistics
• Occupational fraud schemes tend to be extremely costly.
The median loss was $175,000. More than one-quarter of the frauds involved losses of at least $1 million.
• Occupational fraud schemes frequently continue for years, two years in typical, before they are detected.
• There are 11 distinct categories of occupational fraud.
Financial statement fraud was the most costly category with a median loss of $2 million for the cases examined.
• The industries most commonly victimized by fraud in our study were banking and financial services (15% of cases), government (12%) and healthcare (8%).
• Fraud perpetrators often display behavioral traits that serve as indicators of possible illegal behavior. In financial statement fraud cases, which tend to be the most costly, excessive organizational pressure to perform was a particularly strong warning sign.
16
Table 2: Research methodology and findings in FFR empirical studies.
Author Methodology Variable Sample Findings
(Dechow – Financial ratios – Other actions by the SEC
• To attract external financing at low cost was found an
important motivation for earnings manipulation
• Firms manipulating earnings are more likely to have:
- insiders dominated boards, - Chief Executive Officer
simultaneously serves as Chairman of the Board.
(Persons 1995)
Stepwise logistic model
• 9 financial ratios
• Z-score
Matched- pairs design
The study found four significant indicators: financial leverage, capital turnover, asset composition and firm size (Fanning
• Financial ratios
• Other indicators:
corporate
• Neural network is more effective
• Financial ratios such as debt to equity, ratios of accounts receivable to sales, trend variables etc are significant indicators.
Logistic regression model outperformed auditors for fraud samples, but were equally performed for non-fraud samples.
(Kirkos et al. 2007)
• Decision tree
• Back-propagatio n neural network
• Bayesian belief network
• Training dataset: neural network is the most accurate
• Validation dataset: Bayesian belief network is the most accurate
(Hoogs et al. 2007)
Genetic Algorithm • 38 financial ratios
• 9 qualitative indicators
51 fraud samples vs. 51 non-fraud samples
Integrated pattern had a wider coverage for suspected fraud companies while it remained lower false classification rate for non-fraud ones
17
Table 3: Variable definition and measurement
Variable Definition Literature Measurement
Dependent variable:
FRAUD (Persons 1995)
If a company’s financial statements for specific years are confirmed to be fraudulent by the indictments and sentences for major securities crimes issued by the Department of Justice, the firm-year data are classified into fraud
observations, and the variable FRAUD will be set to 1, 0 otherwise.
Independent variable Profitability
Gross profit margin
(GPM) (Dechow et al. 2007) OperatingOperatingincome-incomeOperatingcosts
Operating profit ratio
(OPR) (Green and Choi 1997) Operatingincome-OperatingOperatingincomecosts-Operatingexpenses Return on assets
(ROA)
(Hoogs et al. 2007; Persons
1995) Average totalassets
rate)
Growth rate of net sales
(GRONS)
(Dechow et al. 2007; Stice 1991; Summers and
Growth rate of net income
(Green and Choi 1997)
receivable
18 Total asset turnover
(TAT)
(Kirkos et al. 2007; Persons
1995) Totalassets
sales
Growth rate of inventory to gross sales
(GRARTGS)
(Summers and Sweeney
1998) t-1
Inventory to gross sales
(GRITGS)
(Summers and Sweeney
1998) t-1 to total assets
(ARTTA)
(Green and Choi 1997;
Persons 1995; Stice 1991) Totalassets receivable Accounts
Inventory to total assets
(ITTA)
(Persons 1995; Stice 1991)
assets
(Kirkos et al. 2007; Persons
1995) Totalassets
s liabilitie Total
Long-term funds to fixed assets
Cash flow ability Cash flow ratio
(CFR) (Dechow et al. 2007)
Cash flow adequacy ratio
19
(Altman 1968; Fanning and Cogger 1998; Stice 1991;
Summers and Sweeney
1998) 0.6 (Book valueMarket valueof totalofequitydebt) 1.0 TAT Stock Pledge ratio
(SPR)# (Lee and Yeh 2004)
Sum of percentage of major
shareholders’
shareholdings (SMLSR)
(Beasley et al. 1999) Σ (Percentage of shareholdings >10%)
Deviation between CR and CFR (DBCRCFR)
(La Porta et al. 1999; Lee and
Yeh 2004) Voting rights - Cash flow rights Deviation between
CBS and CFR (DBCBSCFR)
(Lee and Yeh 2004; Yeh et al. 2001)
Percentage of board seats controlled - Cash flow rights
#: According to the rule issued from the Securities and Futures Commission (SFC) of Taiwan, directors, supervisors, managers and large shareholders (that own 10 per cent or more of a company’s outstanding shares) in public companies are obliged to report to the SFC the percentage of their shareholdings that are pledged for loans and credits. These data matter, since pledging for loans effectively reduces the personal funds required for shareholding. In other words, the degree of personal leverage expands and the over-investments in the stock market by the largest shareholder also create risk for the companies to a certain degree. (Lee and Yeh 2004)
20
Table 4: Descriptive Statistics of variables
Fraud Sample (N=113) Non-fraud Sample (N=467)
Variable Mean Median 25 Percentiles
75 Percentiles
Mean Median 25
Percentiles
75 Percentiles
Z
GPM 11.85 10.65 4.99 19.41 15.51 14.47 8.12 22.77 -3.19 OPR -5.39 0.32 -7.26 6.92 -34.49 3.81 -0.24 8.60 -3.98 ROA -13.45 -2.76 -23.48 5.29 3.40 4.19 0.39 7.97 -6.53 GRONS 8.30 7.84 -15.47 24.99 38.73 5.23 -7.77 19.89 -0.08 GRONI 47.23 -71.97 -636.91 24.49 -41.32 14.30 -44.89 80.07 -6.74 CR 109.83 104.68 60.98 141.48 190.94 150.01 110.02 210.00 -7.00 QR 57.79 45.54 21.84 77.09 110.36 75.73 38.09 124.66 -5.16 ART 7.10 4.62 3.16 7.34 8.91 5.36 3.75 8.94 -2.51 TAT 0.61 0.48 0.31 0.74 0.75 0.64 0.41 0.93 -3.69 GROAR 39.67 -5.73 -37.06 34.73 68.97 6.03 -15.15 33.86 -2.42 GROI 13.85 -1.02 -28.82 23.66 27.03 2.18 -14.80 31.14 -1.67 GRARTGS -0.17 -1.04 -7.95 3.30 2.13 0.22 -2.75 3.11 -2.46 GRITGS 24.91 -0.34 -5.40 3.43 23.96 0.00 -3.37 4.80 -1.11 ARTTA 12.02 10.11 4.79 18.37 13.70 10.84 5.05 20.33 -1.33 ITTA 16.72 11.36 5.96 19.49 19.94 13.57 5.82 24.67 -1.74 DR 64.02 60.23 48.10 71.40 48.17 45.03 33.67 56.75 -7.59 LFTFA 452.26 165.79 95.29 399.96 482.48 225.20 146.73 427.05 -3.48 CFR -14.91 -6.88 -21.21 6.54 13.41 8.12 -5.96 29.70 -6.26 CFAR -18.56 -6.54 -27.97 8.65 9.36 14.52 -17.16 54.56 -5.53 CFRR -46.73 -2.69 -14.70 3.74 0.37 2.03 -4.17 7.56 -4.59 SPR 37.44 33.44 1.83 63.26 19.32 3.58 0.00 32.49 -5.67 SMLSR 13.97 11.98 3.72 20.38 10.83 7.89 0.09 16.96 -3.16 DBCRCFR 3.47 0.47 0.00 2.76 3.62 0.56 0.00 4.09 -0.66 DBCBSCFR 46.00 45.58 22.87 67.41 44.26 43.68 26.99 63.69 -0.59 Z-Score 31.45 79.60 -91.69 166.17 198.67 194.70 120.89 270.95 -8.68
21
Table 5: Empirical results of discriminant analysis.
Variable Coefficient F-value Significance
GPM 0.14 3.51 0.061
OPR -0.03 0.16 0.688
ROA 0.77*** 105.82 0.000
GRONS 0.06 0.63 0.427
GRONI -0.02 0.05 0.822
CR 0.34*** 20.59 0.000
QR 0.28*** 13.42 0.000
ART 0.09 1.58 0.210
TAT 0.19 6.38 0.012
GROAR 0.03 0.12 0.731
GROI 0.07 0.90 0.344
GRARTGS 0.00 0.00 0.997
ARTTA 0.11 2.25 0.134
ITTA 0.12 2.37 0.125
DR -0.42*** 30.46 0.000
LFTFA 0.02 0.09 0.764
CFR 0.33*** 19.21 0.000
CFAR 0.24*** 9.89 0.002
CFRR 0.19 6.41 0.012
SPR -0.47*** 38.85 0.000
SMLSR -0.19 6.18 0.013
DBCRCFR 0.02 0.04 0.835
DBCBSCFR -0.05 0.41 0.524
Z-score 0.64*** 72.74 0.000
Wilks' Λ value 0.77 p-value 0.000
χ2 151.10 p-value 0.000
22
Table 6: Common fraudulent techniques adopted of leaf nodes #11, #14-21 & #14-24.
Code year FT1 FT2 FT3 FT4 FT5 FT6 FT7 FT8 FT9 FT10 leaf node #11 (12/9)
2505 1998 ●
2529 1998 ● ●
8716 1999 ● ●
2334 1999 ● ●
3039 2004 ●
1601 1998 ●
1221 2002 ● ● ●
1221 2003 ● ● ●
2014 2003 ● ●
5901 1997 ● ●
5901 1998 ● ●
5901 1999 ● ●
leaf node #14-24 (12/9)
2206 1999 ●
2350 1998 ●
2407 2002 ● ● ● ● ● ●
2407 2003 ● ● ● ● ● ●
2407 2004 ● ● ● ● ● ●
2490 2000 ● ●
2490 2002 ● ●
8295 1998 ● ●
1221 2001 ● ●
8723 1998 ● ● ●
2017 1997 ● ●
5007 1999 ● ●
leaf node #14-21 (7/7/)
5504 1999 ●
2328 1998 ● ●
2334 1998 ● ●
1505 1997 ●
5007 1998 ● ●
2614 1999 ● ● ● ●
1466 1998 ● ●
FT1:recording fictitious revenues;FT2:recording revenues prematurely;FT3:no description /overstated about revenues; FT4: overstating existing assets; FT5: recording fictitious assets or assets not owned; FT6: capitalizing items that should be expensed; FT7: understatement of expenses/liabilities; FT8: misappropriation of assets; FT9: inappropriate disclosure; FT10:
other miscellaneous techniques.
23
Table 7: Relevant pre-warning signals for leaf node #14-21.
leaf node Fraudulent techniques Relevant pre-warning signals
#14-21
Overstating existing assets +
Misappropriation of assets via manipulated cash flow
Cash flow ratio
Cash flow adequacy ratio Investment cash flow Free cash flow
Related party transaction (disposal of assets related)
Cash flow reinvestment ratio Stock pledge ratio
24
Figure 1: The GHSOM structure adapted from (Dittenbach et al. 2000)
Figure 2: Horizontal growth of individual SOM. The notation e indicates the error node and d the dissimilar neighbor. Source: (Dittenbach et al. 2000)
Figure 3: The sample distribution in the obtained GHSOM, in which leaf nodes are marked in taint. In each node, the numbers within the parenthesis indicate the number of fraudulent financial statements and the number of (fraud) firms.
#11 (12/9)
#12 (28/18)
#13 (25/18)
#14 (48/15) Layer 1
Layer 2
#12-21
(3/3) #12-23
(13/9)
#12-22 (11/7)
#12-24 (1/1)
#13-21
(6/5) #13-23
(4/4)
#13-22 (9/7)
#13-24 (6/3)
#14-21
(7/7) #14-23 (21/12)
#14-22 (8/8)
#14-24 (12/9) #01
(113/58)