THE IMPORTANCE OF EARNINGS MANAGEMENT DETECTION MODELS TO
IDENTIFY FRAUD: A CASE FROM ITALIAN LISTED FIRMS
Eugenio D’Amicoa , Elisabetta Mafrollab *
Rome Third University, Department of Public Affairs – 00134 Rome, via Capranica
University of Foggia, Department of Economics – 71122 Foggia, via Caggese n. 1
Financial statement frauds is a recurrent and unsettled problem of nowadays economic and financial
landscapes. We examine frauds due to misstatement in financial reporting or misappropriation of
companies’ assets occurred in the last years (1990-2009) amongst Italian listed firms. We test the
validity of the accruals based analysis to predict the fraudulent event and measure the accuracy of
the prediction, using (as control variables) other appropriate indicators individuated by main
financial literature on bankruptcy. We measure the relevance of earnings manipulation affecting
total accruals (measured following the literature) in fraud firms and in a peer group of non-fraud
firms. Surprisingly, most traditional instruments adopted to explain total accruals have no statistic
significance in our sample. Conversely, we find that the binary variable indicating the presence of
fraud is the most statistically significant element, and is able to explain a portion of total accruals
II. THEORY AND HYPOTHESIS DEVELOPMENT
Fraud is an attracting but inaccessible field for accounting researchers. Relevant problems affect the
identification of the fraudulent event and the accessibility of related information (Goode & Lacey,
2010). Consequences of difficult (and late) fraud detection on investors’ confidence make it the
question more and more relevant both on an academic and practitioner point of view, fostering
interest on the identification of the proper technical instruments able to detect the fraudulent
behavior (Bolton & Hand, 2002; Rezaee, 2005).
Our research question is located in the field of earnings management studies, targeted to test the
hypothesis of total accruals analysis as a consistent instrument to detect frauds in Italy. Following
the dominant literature (Beneish, 1997; Dechow et al.,1996; Jones et al., 2007) we believe earnings
management displayed in total accruals is higher for firms experiencing financial statement
fraudulent events then for the not fraudulent comparable companies. Thus, we aim to confirm
literature adding a case from Italy.
III. SAMPLE SELECTION AND DESCRIPTION
Corresponding author. Tel.: 0039 339 8613473; e-mail: firstname.lastname@example.org ; Skype: elisabetta mafrolla
The sample used to test our hypothesis is made up of 40 Italian publicly traded firms. 20 of them
represent the sample group of fraud firms, because each of them incurred in financial statement
fraud, publicly reported during the period 1990-2009. A peer group of 20 non-fraud firms is
selected, matching each fraud firm with a comparable non fraud one.
Selection of fraud firms
As usually happens in the fraud and bankruptcy research field, it is difficult to individuate
objectively an exhaustive database of fraudulent listed firms, due to some reasons.
First, in Italy there is no comprehensive list of pronounced frauds and it is difficult to access to the
single judgment passed. Furthermore, the fraudulent event may be juridically ascertained even
many years after the fraud is committed. Thus, browsing through the main national economic press
we selected a list of public companies (not exhaustive and liable to be integrated) involved in
bankruptcy, whose CEOs, CFOs or managers have been accused and sentenced of fraud, with the
impeachment for fraudulent financial reporting or misappropriation of assets.
In order to collect the larger number of firms, we analyze a wide period of time, starting with
discovered frauds affecting the reports on the companies in the Nineties and getting to events
ascertained involving the annual reports for the year 2009
Along these lines, we extracted totally 24 companies involved in fraudulent events. Successively,
we excluded some of them due to relevant reasons. First, in order to undertake our analysis, we had
to collected at least 5 years of annual statements before fraud occurrence. Then, we excluded firms
with no available financial statement data for 5 years or more (2 companies). Then, to adopt an
homogeneous criterion and interpret our analysis, we exclude banks and insurance firms (2
Hence, we analyze a sample of 20 Italian firms listed in the Main National Stock Market in the
years 1990-2009, whose managers or directors are accused and challenged for financial statement
Selection of non-fraud firms
To verify our hypothesis, we choose a group of peer firms. We select a matching group of 20 listed
companies (Beasley, 1996), and undertake the analysis on them as a control group. Comparability is
stated in selecting for matching group the firms that are more close to the fraud sample in some
criteria. Thus, companies are similar in stock market (Milan Stock Exchange); industry 3 (if
available) or 2 digits US-SIC; and size (number of employees, total sales, total assets).
Description of the sample
We don’t select frauds occurring after 2009 to guarantee a sufficient lapse of time to verify that non-fraud firms have
not experienced fraudulent events (Beasley, 1996). As well, we select data regarding the no-fraud group from the
financial statements of the years 2005 to 2009.
Our sample (N=40) is composed of firms different in size and operating in various industries. The
fraud group and the non-fraud group are consistently comparable. We check the comparability
adopting the stock market, size and industry criterion, as shown in table 1.
Insert table 1 about here
Our dataset of variables is mainly composed of items selected in the Datastream database by
Thompson. Where the Thompson’s database is lacking we tried to fill blanks using Osiris database
by Bureau van Dijk.
IV. RESEARCH DESIGN
Our research design involves panel regression analysis, following traditional models to measure
discretionary accruals. Hence, following Jones Model and Modified Jones Models, we measure the
residuals from equations (1), (2), (3) and (4) as proxies for discretionary accruals in our sample
(n=40). We add to the traditional models the binary variable FRAUDi, to control the fraudulent
event and test its influence on the discretional earnings management analysis.
Jones Model (1991):
TAit = β0 + β1ATit-1-1+ β2ΔREVit+ β3 PPEit+ FRAUDi + uit (1)
where TAit is total accruals, calculated as the difference between operating income and operating
cash flow; ATit-1 is total assets at the beginning of the year, ΔREVit is change in total sales from year
t-1 to year t; PPEit is gross property, plant and equipment (i.e.: operating fixed assets) and uit
measures residuals for the regression. TAit, ΔREVit and PPEit are scaled with ATit-1.
Dechow et al. (1995) Modified Jones Model:
TAit = β0 + β1ATit-1-1+ β2(ΔREVit - ΔARit) + β3 PPEit+ β4FRAUDi + uit (2)
where ΔARit is change in accounts receivable from year t-1 to year t, and all other symbols assume
the meaning described above.
Larcker and Richardson (2004) Modified Jones Model:
TAit = β0 + β1ATit-1-1+ β2(ΔREVit - ΔARit) + β3 PPEit+ β4 MBit+ β5 CFOit+ β6FRAUDi + uit (3)
where MBit is market to book ratio and CFOit is operating cash flow over ATit-1.
Kothari et al. (2005) Modified Jones Model:
TAit = β0 + β1ATit-1-1+ β2(ΔREVit - ΔARit) + β3 PPEit+ β4 ΔROit+ β5FRAUDi + uit (4)
where the Author adds to the previous models the annual the operating income (RO) as a measure
of the performance of the firm‡.
Table 2 reports the correlations between the explanatory variables used in our study. It shows that
only moderate levels of collinearity exist between the explanatory variables. The highest correlation
is between ΔREV and ΔREV – ΔAR of 0,6977 (p-value= 0,000). Moreover, we also calculate the
variance inflation factors (VIFs), when estimating the base pooled OLS model, to avoid multi-
collinearity among regressors. Our (unreported) results confirm that no VIFs exceed 5 for any
variable, so multi-collinearity does not affect our analysis.
Insert table 2 about here
We pool (1), (2), (3) and (4) for both the fraud sample and the control group. Regression analysis
are shown in table 3. Normality of residuals from the regressions is even checked and found (but
not reported) in our analysis.
Insert table 3 about here
Traditional instruments adopted to predict earnings management through total accruals seem to be
inconsistent testing them on our sample. The discretionary portion of accruals is supposed to be
expressed in the residuals from regressions (1), (2), (3) and (4) when the non-discretionary portion
of accruals is statistically tested in the explanatory variables adopted.
In our sample only 1 of the 4 models can be used to predict earnings management. Jones Model and
the Dechow et al. Modified Jones Model have no statistic significance, thus we cannot infer any
relation among variables (neither about residuals) adopting those well-known and experienced
accrual detection models.
Pooling the Kothari et al. Model only the control variable (ΔROit) has a relevance (β= 0,8227; p-
value = 0.000), while the overall meaning of the regression is quite weak (Adj-R2 = 0,378).
We find significant relations in the Larcker and Richardson Model (Adj-R2 = 0,695).
First, in the Larcker and Richardson Model, the variable MBit has a slightly negative coefficient (β
=-0,0006; p-value = 0,0378), meaning that total accruals increase when the market to book value of
a firm decrease. Thus, over-stating of income occurs in firms that have a market value closer to
We pooled first the model separately with ROit-1 and with ROit and then we tried to measure the relevance of the
change ΔROit. Using the variation over time (ΔROit.), TAit seems to be better predicted. Hence, we adopt only this
their book value. This is consistent with the circumstance that the manager who is less prudent in
his accounting estimates, on the left hand side will overstate income (enhancing TAit) but, on the
right hand side will over-evaluate even other assets, enlarging the book value of the firm, but
reducing the market to book ratio. Hence, market prices lead earnings in the detection of the
discretional accrual manipulation and market value is not relevantly affected by earnings
manipulations (Kothari and Zimmerman, 1995).
Second, even the coefficient of the variable CFOit is significantly negative (β= -0,7654; p-value =
0,000). This finding is consistent with Literature (Burgstahler and Dichev, 1997; Degeorge et al.,
1999), stating that managers try to hide negative performances writing more favorable earnings. So,
the worse the cash flow performance, the larger the earnings manipulation to over-state net income.
An important finding of our research deals with the introduction of the FRAUDi binary independent
variable in the total accruals detection models. When the fraud befalls the total accruals are much
more dispersed from the zero level than in the cases on no-fraud firms, as clearly shown in graph 1.
Insert graph 1 about here
Statistic significance of this finding is high in most of the accruals detection models we adopt. In
Jones Model β= -0,05913; p-value = 0,0143; in Dechow et al. Modified Jones Model β= -0,05913;
p-value = 0,0143; in Larcker and Richardson Model β= -0,05824; p-value = 0,000; in Kothari
Model the statistic significance of the variable is too low.
Literature (Jones et al., 2008) generally adopts a two steps research design to study the discretion
used by managers in earnings manipulation. Even the relevance of the fraudulent event on earnings
management is traditionally traced back through a two steps approach.
In a first step, the net discretionary accruals (NDAit) for various earnings management detection
models is measured pooling the fraud sample and no-fraud peer group separately, and then
calculating the difference between residuals from the fraud group and residuals from the non-fraud
Then (second step), researchers generally adopt a logit or probit model, where the dependent
variable (FRAUDit) is 1 in the fraud firms and 0 in the non-fraud control group. The independent
variables of this model are total accruals (TAit), the portion of residuals considered to be
anomalously discretional (NDAit) and a various range of control variables.
We verify the existence of a strong relation between FRAUDit and TAit, but we cannot undertake the
exposed logit or probit analysis, due to a relevant reason. The portion of non-discretionary TAit,
which traditionally is explained by ATit-1-1, ΔREVit, ΔARit and PPEit is not significant in our sample.
Hence, the residuals from our regression may contain even a portion of non-discretionary accruals.
Probably, fraudulent events analyzed in our Italian sample are basically undetectable through
traditional instruments because the fraud is not simply earnings management, but consists of
fraudulent creation of revenues and cash flows (e.g.: Parmalat, one of the most striking cases of
fraud in Italy, wrote millions of euro of inexistent sales, credits and cash in banks).
For further developments in our study, consistently with previous researches, we know that a wide
range of control variables could be added to the analysis in order to test the efficacy of our findings,
measuring both financial and governance patterns.
The influence of leverage ratio on fraudulent bankruptcy could be an important measure how much
the fraudulent firms depend on external financial provisions of liquidity. Generally, the higher the
weight of third parties’ capital, the higher the probability of reimbursement requests that could
cause financial distress to the firm. Nevertheless, the presence of a powerful and influent financing
third party could protect from fraudulent behavior, representing an armful controlling party. Even
the relevance of intangibles assets could contain information on one of the most risky component of
the assets of a firm. In the case of improvise default (and liquidation) of the firm, selling intangible
assets is usually more difficult then selling property plants and equipments, especially for some
particular categories of items (e.g.: R&D investments). Thus, the portion of the capital invested in
general intangibles is probably not going to produce revenues in the case of failure and separate
selling of the assets of the firm; hence, the intangible assets become a loss in the case of improvise
bankruptcy. Another meaningful control variable could be the Altman Z score, generally considered
one of the most powerful instruments to predict bankruptcy (Altman, 1968). Even non-financial
variables could be important to improve our fraud prediction model. As an example, Literature
demonstrates the relevance of the choice of a Big4-Auditor, of the CEO-duality, of the presence of
Independent Directors in the Board.
Altman, E. 1968. Financial ratios, discriminant analysis and the prediction of corporate bankruptcy.
Journal of Finance (september): 189-209.
Beneish, M. D. 1997. Detecting GAAP violation: implications for assessing earnings management
among firms with extreme financial performance. Journal of Accounting and Public Policy, 16 (3):
Bolton, R. J., D. J. Hand. 2002. Statistical fraud detection: a review. Statistical Science, 17 (3): 235-
Burgstahler D., I. D. Dichev. 1997. Earnings management to avoid earnings decreases and losses.
Journal of Accounting and Economics, 24 (2): 99-126.
Dechow, P. M., R. G. Sloan, A. P. Sweeney. 1995. Detecting earnings management. The
Accounting Review, 70 (2): 193-225.
Dechow, P. M. 1996. Causes and consequences of earnings manipulation: an analysis of firm
subject o enforcement actions by SEC. Contemporary Accounting Research, 13 (1): 1-36.
Dechow, P. M., I. D. Dichev. 2002. The quality of accruals and earnings: the role of accrual
estimation errors. The Accounting Review, 77 (supplement): 35-99.
Degeorge, F., J. Patel, R. Zeckhauser, Earnings management to exceed thresholds, Journal of
Business, 72 (1): 1-33.
Goode, S., D. Lacey. 2011. Detecting complex account fraud in the enterprise: The role of technical
and non-technical controls. Decision Support Systems, 50 (4): 702-714.
Jones, J. 1991. Earnings management during import relief investigations. Journal of Accounting
Research, 29 (2): 193-228.
Jones, K. L., G. V. Krishnan, K. D. Melendrez. 2008. Do models of discretionary accruals detect
actual cases of fraudulent and restated earnings? An empirical analysis. Contemporary Accounting
Research, 25 (2): 499-531.
Kothari, S. P., A. J. Leone, C. Wasley. 2005. Performance matched discretionary accrual measures.
Journal of Accounting and Economics, 39 (1): 163-197.
Kothari, S. P., J. L. Zimmerman. 1995. Price and return models. Journal of Accounting and
Economics, 20 (2): 155-192.
Larcker, D. F., S. A. Richardson. 2004. Fees paid to audit firms, accrual choices, and corporate
governance. Journal of Accounting Research, 42 (3): 625-656.
Rezaee, Z. 2005. Causes, consequences, and deterence of financial statement fraud, Critical
Perspectives on Accounting, 16 (3): 277–298.
Zahra, S., R.L. Priem, A.M.A. Rasheed. 2005. The antecedents and consequences of top
management fraud. Journal of Management, 31 (6).
FRAUD FIRMS GROUP AND NON-FRAUD FIRMS PEER GROUP
Patterns Fraud firms Non-fraud firms
Size Number of employees 4174 1668
Mean  
[Median] (8849) (2527)
(Standard Deviation) Total Sales 1128400 420060
Th.€  
Total Assets 1283900 1291000
Th.€  
Market Value 478290 158470
Th.€  
Year of default (if occurred) 1990-1996 6
i.e.: last year of annual 1997-2002
report used in this research
N 20 20
AT-1 ΔREV PPE Δ(REV-AR) MB RO OCF FRAUD
AT-1 1,0000 -0,0762 -0,1763 0,0733 0,0252 -0,0824 0,0279 0,0878
ΔREV 1,0000 0,1007 0,6977 *** 0,2366 0,1178 0,0282 0,0427
PPE 1,0000 -0,6431 *** -0,0146 -0,2125 -0,0765 -0,0510
Δ(REV-AR) 1,0000 0,1963 0,2437 *** 0,0745 0,0871
MB 1,0000 0,1039 0,4985 *** -0,0771
RO 1,0000 0,2124 *** -0,1571
OCF 1,0000 -0,1285
* significance at the 0.1 level; ** significance at the 0.5 level; *** significance at the 0.01 level.
TOTAL ACCRUAL REGRESSION MODELS
TAit = β0 + βjXit+ βl FRAUDi + uit
variable (1) (2) (3) (4)
Coeff. St.err. p-value Coeff. St.err. p-value Coeff. St.err. p-value Coeff. St.err. p-value
const 0.0311 0.0569 0.5860 0.0311 0.0569 0.5860 0.0373 0.0151 0.0157 ** -0.0038 0.0185 0.8371
AT-1 -3001.1 2210.3 0.1769 -3001.1 2210.3 0.1769 12.800 1815.8 0.9944 -1206.6 1265.3 0.3420
ΔREV 0.0381 0.0409 0.3531
PPE -0.0437 0.1503 0.7716 -0.0055 0.1408 0.9684 0.0156 0.0349 0.6553 0.0510 0.0576 0.3776
Δ(REV-AR) 0.0381 0.0409 0.3531 0.0164 0.0143 0.2562 -0.000 0.0398 0.9803
MB - 0.0003 0.0378 **
OCF -0.7654 0.1178 0.0000 ***
RO 0.8227 0.1395 0.0000 ***
FRAUD -0.0591 0.0238 0.0143 ** -0.0591 0.0238 0.0143 ** -0.0582 0.0106 0.0000 *** -0.0323 0.0212 0.1309
Adj-R2 0.005 0.005 0.964 0.378
Schwarz 4.911 4.911 -206 -73.9
Akaike -134 -134 -344 -216
βjXjit represents j-eme coefficient related to the j-eme independent variable. X varies amongst different models.
* significance at the 0.1 level; ** significance at the 0.5 level; *** significance at the 0.01 level.
Standard errors are corrected using the White (1980) HAC procedure.
TOTAL ACCRUALS AND FRAUDULENT EVENTS
TA rispetto a FRAUD (con retta dei minimi quadrati)
Y = 0,00465 - 0,0555X