Predicting financial statement manipulation in South Africa: A comparison of the Beneish and Dechow models

Figures & data

Table 1. Sample size determination

	No. of companies^1	No. of firm-year observations
Population	330	2 775
Listed after year-end but before the release of AFS	(1)	(26)
Change of year-end	(3)	(104)
Other anomalous situations	(0)	(5)
Missing data for M- or F-score	(27)	(272)
Companies with only one observation	(25)	(25)
Sample size	274	2 343

Note: ¹The removal of firm-year observations exceeds the removal of companies as, for some companies, not all observations were removed.

(Source: Researchers’ own construction)

Table 2. Summary of classification between manipulated and non-manipulated observations

	Current study		Beneish (1999)		Dechow et al. (2011)
	No. of observations	% of observations	No. of observations	% of observations	No. of observations	% of observations
Manipulated observations	23	0.98	74	3.08	494	0.37
FSCA enforcement actions	14	0.60
FRIP restatements	5	0.21
Applicable qualified audit opinions	4	0.17
Non-manipulated observations	2 320	99.02	2 332	96.92	132 967	99.63
Total observations	2 343	100.00	2 406	100.00	133 461	100.00

Note: (Source: Researchers’ own construction, as well as data obtained from Beneish (1999) and Dechow et al. (2011))

Table 3. Industry classifications and descriptive statistics

	Manipulated observations	Non-manipulated observations	Total
Industry classification	23	2 320	2 343
Basic materials	2	580	582
Consumer goods	16	238	254
Consumer services	2	473	475
Healthcare	0	79	79
Industrials	1	728	729
Oil and gas	0	35	35
Technology	2	143	145
Telecommunications	0	44	44
Size (R’000)	31 665 723.61	35 788 356.25	35 747 886.53
Return on assets (%)	3.52	9.86	9.80

Note: (Source: Researchers’ own construction)

Table 4. Distribution of underlying M- and F-score variables for manipulators and non-manipulators

Panel A: Beneish (1999) M-score
	Current sample			Beneish (1999)
Characteristic	Manipulators mean	Non-manipulators mean	Wilcoxon Z	Manipulators mean	Non-manipulators mean	Wilcoxon Z p-value^1
DSRI	1.1386	1.0710	−1.2680	1.465	1.031	0.000
GMI	0.8986	1.0044	0.2970	1.193	1.014	0.006
AQI	1.0529	1.1464	0.0970	1.254	1.039	0.096
SGI	1.0944	1.1278	0.5050	1.607	1.134	0.000
DEPI	1.0788	1.0571	−1.5100	1.077	1.001	0.307
SGAI	1.0170	1.1006	−0.1050	1.041	1.054	0.271
LVGI	1.0217	1.0469	−0.3730	1.111	1.037	0.394
TATA_BS	−0.0042	−0.0287	−2.1900**	0.031	0.018	0.000
TATA_IS^2	−0.0061	−0.0006	0.0910
Panel B: Dechow et al. (2011) F-score
	Current sample			Dechow et al. (2011)
Characteristic	Manipulators mean	Non-manipulators mean	Wilcoxon Z^3	Manipulators mean	Non-manipulators mean	Pairwise t-test
RSST	0.0313	0.0369	−0.4190	0.135	0.044	3.85***
REC	0.0185	0.0173	−0.7230	0.071	0.028	6.12***
INV	0.0079	0.0122	−0.0600	0.046	0.023	4.22***
SASS	0.5925	0.5588	−0.5710	0.647	0.611	3.87***
CSALES	0.0558	0.1245	0.2260	0.466	0.257	3.88***
ROA	−0.0048	−0.0085	0.5930	−0.032	−0.025	−0.38
AISS	0.9565	0.8039	−1.8400*	0.938	0.869	4.75***

Note: ¹Beneish (1999) did not present the Wilcoxon Z, but only the associated p-value.

²Total assets to total accruals based on the income statement approach was not used in the original Beneish (1999) study, so the means and p-value are unavailable.

³Given the nature of the underlying data, a Wilcoxon rank sum test was more appropriate than the pairwise t-test used by Dechow et al. (2011). A t-test did not yield different conclusions to the Wilcoxon rank sum test.

*Statistical significance <0.10, **Statistical significance <0.05, ***Statistical significance <0.01

Note: (Source: Researchers’ own construction, as well as data obtained from Beneish (1999) and Dechow et al. (2011))

Table 5. Classification performance of the M- and F-scores in the SA context

	M-score (BS) (Cut-off = -1.49)	M-score (BS) (Cut-off = -1.78)	M-score (BS) (Cut-off = -1.89)	M-score (IS) (Cut-off = -1.49)	M-score (IS) (Cut-off = −1.78)	M-score (IS) (Cut-off = −1.89)	F-score (UEM = 0.0098)	F-score (UEM = 0.0037)
Predicted manipulators	224	342	408	261	459	553	62	926
True positive	1	4	6	1	3	5	0	12
False positive	223	338	402	260	456	548	62	914
Predicted non-manipulators	2 119	2 001	1 935	2 082	1 884	1 790	2 281	1 417
True negative	2 097	1 982	1 918	2 060	1 864	1 772	2 258	1 406
False negative	22	19	17	22	20	18	23	11
Accuracy	89.54%	84.76%	82.12%	87.96%	79.68%	75.84%	96.37%	60.52%
Per original estimation sample^1	91.41%^2	85.84%^2	82.31%^2					63.71%
Per original holdout sample^1	95.63%^2	92.17%^2	90.38%^2					61.73%
Sensitivity	4.35%	17.39%	26.09%	4.35%	13.04%	21.74%	0.00%	52.17%
Per original estimation sample^1	58.00%	74.00%	76.00%					68.62%
Per original holdout sample^1	37.50%	50.00%	56.10%					73.83%
Precision	0.45%	1.17%	1.47%	0.38%	0.65%	0.90%	0.00%	1.30%
Per original estimation sample^1	18.24%^2	13.55%^2	11.28%^2					0.70%^2
Per original holdout sample^1	13.64%^2	9.30%^2	8.36%^2					1.15%^2
F-measure	0.87%	2.19%	2.78%	0.70%	1.24%	1.74%	N/A^4	2.53%
Per original estimation sample^1	27.75%^2	22.91%^2	19.64%^2					1.38%^2
Per original holdout sample^1	20.00%^2	15.69%^2	14.55%^2					2.27%^2
Type I error	95.65%	82.61%	73.91%	95.65%	86.96%	78.26%	100.00%	47.83%
Per original estimation sample^1	42.00%^2	26.00%^2	24.00%^2					31.38%
Per original holdout sample^1	62.50%^2	50.00%^2	43.90%^2					26.17%
Type II error	9.61%	14.56%	17.32%	11.20%	19.65%	23.61%	2.67%	39.38%
Per original estimation sample^1	7.60%	13.80%	17.50%					36.31%
Per original holdout sample^1	3.50%	7.20%	9.10%					38.35%
Area under the receiver operator curve3	0.5936			0.4789			0.6067

Note: ¹The performance measures for the original studies relate to Beneish (1999) for the M-score and Dechow et al. (2011) for the F-score. No original performance measures were presented for the M-score based on the income statement approach or for the F-score based on the UEM of 0.0098, as these were not included in the original study.

²These performance measures were not presented in the original studies by Beneish (1999) and Dechow et al. (2011), but they have been recalculated based on the data presented in these studies.

³The results for the area under the receiver operator curve were based on the underlying M- or F-score rather than a specific cut-off point or unconditional expectation of misstatement. The original studies did not include a calculation of the AUC, nor was it possible to recalculate based on the data presented in the studies.

⁴Due to both precision and sensitivity being equal to zero, it was impossible to compute the F-measure.

(Source: Researchers’ own construction, as well as data obtained from Beneish (1999) and Dechow et al. (2011)).

Table 6. Comparison of earnings management characteristics of false positives to true manipulators and non-manipulators

	M-score (BS) (Cut-off = -1.49)	M-score (BS) (Cut-off = -1.78)	M-score (BS) (Cut-off = -1.89)	M-score (IS) (Cut-off = -1.49)	M-score (IS) (Cut-off = -1.78)	M-score (IS) (Cut-off = -1.89)	F-score (UEM = 0.0098)	F-score (UEM = 0.0037)
Panel A: False positives compared to manipulators
DISCRETIONARY ACCRUALS
False positive mean	0.0836	0.0732	0.0706	0.0500	0.0454	0.0470	0.1460	0.0145
Manipulator mean	0.0280	0.0280	0.0280	0.0280	0.0280	0.0280	0.0280	0.0280
Wilcoxon Z	2.0060**	2.1000**	2.1690**	1.3940	1.2940	1.2250	3.4410***	−0.0920
ABSOLUTE DISCRETIONARY ACCRUALS
False positive mean	0.1396	0.1239	0.1157	0.1067	0.0974	0.0959	0.1623	0.0755
Manipulator mean	0.0684	0.0684	0.0684	0.0684	0.0684	0.0684	0.0684	0.0684
Wilcoxon Z	2.3240**	2.241**	2.084**	1.6390	1.4790	1.2600	2.8700***	0.1200
MEET OR JUST BEAT − 1 CENT
False positive mean	0.0448	0.0414	0.0423	0.0423	0.0417	0.0365	0.0000	0.0186
Manipulator mean	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
Pearson Chi^2	1.0751^1	0.9911^1	1.0132^1	1.0124^1	0.9979^1	0.8699^1	N/A	0.4357^1
MEET OR JUST BEAT − 2 CENT
False positive mean	0.0583	0.0621	0.0597	0.0538	0.0570	0.0511	0.0192	0.0339
Manipulator mean	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
Pearson Chi^2	1.4156^1	1.5173^1	1.4553^1	1.3029^1	1.3867^1	1.2358^1	0.4483^1	0.8068^1
MEET OR JUST BEAT − 3 CENT
False positive mean	0.0718	0.0799	0.0746	0.0692	0.0768	0.0693	0.0769	0.0481
Manipulator mean	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
Pearson Chi^2	1.7650^1	1.9858^1	1.8468^1	1.7005^1	1.9045^1	1.7086^1	1.8689^1	1.1618^1
Panel B: False positives compared to true negatives
DISCRETIONARY ACCRUALS
False positive mean	0.0836	0.0732	0.0706	0.0500	0.0454	0.0470	0.1460	0.0145
True negative mean	−0.0105	−0.0142	−0.0164	−0.0079	−0.0128	−0.0162	−0.0044	−0.0115
Wilcoxon Z	9.4700***	12.2510***	13.8000***	8.4250***	11.0160***	12.0470***	6.9380***	7.3400***
ABSOLUTE DISCRETIONARY ACCRUALS
False positive mean	0.1396	0.1239	0.1157	0.1067	0.0974	0.0959	0.1623	0.0755
True negative mean	0.0637	0.0620	0.0618	0.0666	0.0646	0.0635	0.0692	0.0683
Wilcoxon Z	8.1390***	9.7460***	9.8110***	6.1660***	7.4610***	7.1930***	5.2560***	1.8630*
MEET OR JUST BEAT − 1 CENT
False positive mean	0.0448	0.0414	0.0423	0.0423	0.0417	0.0365	0.0000	0.0186
True negative mean	0.0224	0.0217	0.0209	0.0223	0.0204	0.0209	0.0251	0.0284
Pearson Chi^2	4.2315**	4.6879**	6.3707**	3.8446**	6.9231***	4.2590**	1.3398^1	2.2425
MEET OR JUST BEAT − 2 CENT
False positive mean	0.0583	0.0621	0.0597	0.0538	0.0570	0.0511	0.0192	0.0339
True negative mean	0.0448	0.0434	0.0433	0.0451	0.0435	0.0446	0.0467	0.0541
Pearson Chi^2	0.8313	2.3050	2.0386	0.3972	1.5319	0.4035	0.8743^1	5.1057**
MEET OR JUST BEAT − 3 CENT
False positive mean	0.0718	0.0799	0.0746	0.0692	0.0768	0.0693	0.0769	0.0481
True negative mean	0.0582	0.0560	0.0563	0.0583	0.0553	0.0564	0.0591	0.0669
Pearson Chi^2	0.6635	2.9428*	1.9934	0.4974	3.0263*	1.2469	0.2892^1	3.4684*

Note: ¹The statistical significance of the Pearson Chi² for these characteristics was based on Fisher’s exact rather than the standard p-value due to the small sample size. No differences in the significance level would have arisen had the standard p-value been used.

***, ** and * represent statistical significance at 1%, 5% and 10% respectively. (Source: Researchers’ own construction)

Table 7. Re-estimated coefficients for the M- and F-scores in the SA context

Panel A: Re-estimated M-score			Panel B: Re-estimated F-score
	BS	IS
DSRI	0.0776 (0.1622)	0.1194 (0.1506)	RSST	−0.4474 (1.7727)
GMI	−0.1463 (0.1497)	−0.1355 (0.1412)	REC	0.9622 (3.9236)
AQI	−0.0692 (0.1237)	−0.0722 (0.1267)	INV	−2.8220 (5.9092)
SGI	−0.2178 (0.2997)	−0.1280 (0.2819)	SASS	0.7275 (0.9417)
DEPI	0.0307 (0.1848)	0.0347 (0.1783)	CSALES	−0.5816 (0.7008)
SGAI	−0.2042 (0.2180)	−0.1875 (0.2070)	ROA	0.8299 (2.1063)
TATA_BS	1.4912 (1.0488)		AISS	1.7646* (1.0284)
TATA_IS		−0.3350 (0.9243)	Constant	−6.5314*** (1.1503)
LVGI	−0.0603 (0.2713)	−0.1538 (0.2819)
Constant	−1.6897*** (0.6040)	−1.7944*** (0.5898)
Observations	2 343	2 343	Observations	2 343
LR Chi^2	5.10	3.19	LR Chi^2	6.68
Prob > Chi^2	0.7464	0.9219	Prob > Chi^2	0.4631
Pseudo R^2	0.0197	0.0123	Pseudo R^2	0.0258

Note: No evidence of multicollinearity was identified. Standard errors are presented in parentheses. The re-estimated M-score coefficients were based on probit estimation, while the re-estimated F-score coefficients were based on logit estimation, in line with the original studies.

*** and * represent statistical significance at 1% and 10%, respectively.

(Source: Researchers’ own construction)

Table 8. Classification performance for re-estimated M- and F-scores (estimation sample)

	M-score (BS) (Cut-off = -1.7910)	M-score (BS) (Cut-off = -1.9653)	M-score (BS) (Cut-off = -2.0407)	M-score (IS) (Cut-off = -1.4539)	M-score (IS) (Cut-off = -1.9735)	M-score (IS) (Cut-off = -2.1641)	F-score (UEM = 0.0098)	F-score (UEM = 0.0037)
Predicted manipulators	10	28	63	1	22	105	1 347	1 889
True positive	1	2	3	0	1	4	17	22
False positive	9	26	60	1	21	101	1 330	1 867
Predicted non-manipulators	2 333	2 315	2 280	2 342	2 321	2 238	996	454
True negative	2 311	2 294	2 260	2 319	2 299	2 219	990	453
False negative	22	21	20	23	22	19	6	1
Accuracy	98.68%	97.99%	96.59%	98.98%	98.16%	94.88%	42.98%	20.27%
Sensitivity	4.35%	8.70%	13.04%	0.00%	4.35%	17.39%	73.91%	95.65%
Precision	10.00%	7.14%	4.76%	0.00%	4.55%	3.81%	1.26%	1.16%
F-measure	6.06%	7.84%	6.98%	N/A^1	4.44%	6.25%	2.48%	2.30%
Type I error	95.65%	91.30%	86.96%	100.00%	95.65%	82.61%	26.09%	4.35%
Type II error	0.39%	1.12%	2.59%	0.04%	0.91%	4.35%	57.33%	80.47%
AUC^2	0.5339			0.4419			0.5546

Note: ¹Due to both precision and sensitivity being equal to zero, it was impossible to compute the F-measure.

²The AUC was based on 10-fold cross-validation.

(Sources: Researchers’ own construction)

Table 9. Classification sensitivity per manipulator category

SCORE	CATEGORIES OF MANIPULATORS
	FSCA enforcement action	FRIP restatement	Qualified audit opinion
M-score (BS)
Cut-off = −1.49	0.00%	0.00%	25.00%
Cut-off = −1.78	7.14%	20.00%	50.00%
Cut-off = −1.89	7.14%	60.00%	50.00%
M-score (IS)
Cut-off = −1.49	0.00%	20.00%	0.00%
Cut-off = −1.78	7.14%	20.00%	25.00%
Cut-off = −1.89	7.14%	40.00%	50.00%
F-score
UEM = 0.0098	0.00%	0.00%	0.00%
UEM = 0.0037	50.00%	80.00%	25.00%

Table

Variable	Description	Calculation
DSRI	Day’s sales receivable index	$\frac{AccountsReceivabl{e}_t/Revenu{e}_t}{AccountsReceivabl{e}_{t-1}/Revenu{e}_{t-1}}$
GMI	Gross margin index	$\frac{\left(Revenu{e}_{t-1}-Costofsale{s}_{t-1}\right)/Revenu{e}_{t-1}}{\left(Revenu{e}_t-Costofsale{s}_t\right)/Revenu{e}_t}$
AQI	Asset quality index	$\frac{1-(\left(Currentasset{s}_t+PP{E}_t\right)/Totalasset{s}_t)}{1-(\left(Currentasset{s}_{t-1}+PP{E}_{t-1}\right)/Totalasset{s}_{t-1})}$
SGI	Sales growth index	$\frac{Revenu{e}_t}{Revenu{e}_{t-1}}$
DEPI	Depreciation index	$\frac{Depreciatio{n}_{t-1}/\left(Depreciatio{n}_{t-1}+PP{E}_{t-1}\right)}{Depreciatio{n}_t/\left(Depreciatio{n}_t+PP{E}_t\right)}$
SGAI	Sales, general and administrative expenses index	$\frac{Sales,generalandadminexpense{s}_t/Revenu{e}_t}{Sales,generalandadminexpense{s}_{t-1}/Revenu{e}_{t-1}}$
TATA_BS	Total accruals (based on the balance sheet approach) to total assets	$\frac{\left(\begin{array}{c}\mathrm{\Delta }Currentasset{s}_t-\mathrm{\Delta }Cas{h}_t-\mathrm{\Delta }Currentliabilitie{s}_t-\\ \mathrm{\Delta }Currentmaturitiesoflongtermdeb{t}_t-\\ \mathrm{\Delta }Incometaxpayabl{e}_t-Depreciationandamortisatio{n}_t\end{array}\right)}{Totalasset{s}_t}$
TATA_IS	Total accruals (based on the income statement approach) to total assets	$\frac{\left(Incomebeforeextraordinaryitem{s}_t-Cashfromoperation{s}_t\right)}{Totalasset{s}_t}$
LVGI	Leverage index	$\frac{\left(Longtermdeb{t}_t+Currentliabilitie{s}_t\right)/Totalasset{s}_t}{\left(Longtermdeb{t}_{t-1}+Currentliabilitie{s}_{t-1}\right)/Totalasset{s}_{t-1}}$

Table

Variable	Description	Calculation
RSST	Accruals as measured by Richardson et al. (2005)	$\frac{\mathrm{\Delta }WC+\mathrm{\Delta }NCO+\mathrm{\Delta }FIN}{Averagetotalassets}$ Where: $\mathrm{\Delta }WC=\left(Currentassets-Cashandshortterminvestments\right)-\left(Currentliabilities-Debtincurrentliabilities\right)$ $\mathrm{\Delta }NCO=\left(Totalassets-Currentassets-Investmentsandadvances\right)-\left(Totalliabilities-Currentliabilities-Longtermdebt\right)$ $\mathrm{\Delta }FIN=\left(Shortterminvestments+Longterminvestments\right)-\left(Longtermdebt+Debtincurrentliabilities+Preferredstock\right)$
ΔREC	Change in receivables	$\frac{\mathrm{\Delta }Accountsreceivable}{Averagetotalassets}$
ΔINV	Change in inventory	$\frac{\mathrm{\Delta }Inventory}{Averagetotalassets}$
SASS	Percentage soft assets	$\frac{Totalasset{s}_t-PP{E}_t-Cashandcashequivalent{s}_t}{Totalasset{s}_t}$
ΔCSALES	Change in cash sales	Percentage change in cash sales, where cash sales is measured as $\left(Sales-\mathrm{\Delta }Accountsreceivable\right)$
ΔROA	Change in return on assets	$\frac{Earning{s}_t}{Averagetotalasset{s}_t}-\frac{Earning{s}_{t-1}}{Averagetotalasset{s}_{t-1}}$
AISS	Actual issuance	Indicator variable coded to 1 if the firm issued securities during year t.

Beneish, M. D. (1999). The detection of earnings manipulation. Financial Analysts Journal, 55(5), 24–36. https://doi.org/10.2469/faj.v55.n5.2296

 Google Scholar

Dechow, P. M., Ge, W., Larson, C., & Sloan, R. G. (2011). Predicting material accounting misstatements. Contemporary Accounting Research, 28(1), 17–82. https://doi.org/10.1111/j.1911-3846.2010.01041.x

 Web of Science ®Google Scholar