Electronic copy available at: http://ssrn.com/abstract=1998387

Fraud Detection and Expected Returns

Messod D. Beneish, Charles M.C. Lee, D. Craig Nichols**

January 31, 2012

Abstract An accounting-based model has strong out-of-sample power not only to detect fraud, but also to predict cross-sectional returns. Firms with a higher probability of manipulation (MSCORE) earn lower returns in every decile portfolio sorted by: Size, Book-to-Market, Momentum, Accruals, and Short-Interest. We show that the predictive power of MSCORE is related to its ability to forecast the persistence of current-year accruals, and is most pronounced among low-accrual (ostensibly high earnings-quality) stocks. Most of the incremental power derives from measures of firms’ predisposition to manipulate, rather than their level of aggressive accounting. Our evidence supports the investment value of careful fundamental analysis, even among public firms.

 

                                                            ** Beneish (dbeneish@indiana.edu) is Sam Frumer Professor of Accounting at Indiana University, Lee (clee8@stanford.edu) is Joseph McDonald Professor of Accounting at Stanford University, and Nichols (dcnichol@syr.edu) is Assistant Professor of Accounting at Syracuse University. An earlier version of this paper was presented at the June 2007 Corporate Ethics and Investing Conference of the Society of Quantitative Analysts, the May 2008 LSV-Penn State Conference, and workshops at Cornell University, Indiana University, the University of Maryland, Notre Dame University, the University of Arizona, and Stanford University. The authors would like to thank the participants at these workshops, as well as C. Harvey, P. Hribar, J. Salamon, and C. Trzincka, S. Bhojraj, D. Givoly, J. Lakonishok, M. Lang, A. Leone, R. Morton, B. Swaminathan, P. von Hippel, N. Yehuda, and X. Zhang, for helpful discussions and comments. We also thank M. Drake, L. Rees, and E. Swanson, for kindly sharing their short-selling data with us.

Electronic copy available at: http://ssrn.com/abstract=1998387

1  

1. Introduction

Financial economists have long recognized that when information is costly, mispricing and

arbitrage must co-exist in equilibrium. On the one hand, the pricing efficiency of markets

depends on the forces of arbitrage. On the other hand, sufficient mispricing must remain in

equilibrium to ensure arbitrageurs are rewarded for their efforts. As Grossman and Stiglitz

(1980) show in a rational expectation framework, the cost of information determines the

equilibrium ratio of information quality between informed and uninformed traders.

While the distinction between informed and uninformed traders is often invoked in the literature,

far less is known about the nature of the information that distinguishes the two. In particular,

academics and practitioners seem sharply divided on the investment value of financial analysis.

If markets for public equity are “semi-strong efficient” (Fama (1965)), attempts to best a passive

index with publicly available financial information are futile. Yet each year trillions of dollars

continue to be controlled by professional asset managers who claim to do exactly that. Is there

any investment value to careful financial analysis? And if so, what type of financial information

is particularly valuable to investors? Answers to these two questions remain surprisingly elusive.

This study investigates the investment value of a particular form of financial analysis known as

“Forensic Accounting.” In recent years, this phrase has acquired currency as a moniker for the

art and science of carefully parsing a company’s financial records with a view toward forecasting

its future prospects. Closely related to the “Quality of Earnings” analysis popularized by

O’Glove (1987), Kellogg and Kellogg (1991) and Siegel (1991), forensic accountants pour over

company’s financial statements looking for inconsistencies, irregularities, and other signs of

trouble. While these efforts have yielded individual success stories, the evidence to date has

been largely anecdotal.1

The statistical model we examine (Beneish (1999)) represents a systematic distillation of

forensic accounting principles described in the practitioner literature. Using model

coefficients that were estimated in a prior time period, we provide clear out-of-sample

evidence that between 1998 and 2009, these techniques have had substantial investment

                                                            1 See Schilit (2002) for a list of case studies. Foster (1979) provides similar evidence in an earlier era.

2  

value.2 Specifically, we show that this model correctly identified, in advance of public

disclosure, a large majority (71%) of the most famous accounting fraud cases that

surfaced subsequent to the model’s estimation period.3

While relatively few firms are actually indicted for accounting fraud, the probability of

manipulation generated by the model (MSCORE) should be informative of a firm’s future

prospects. This is because the profile of a “typical earnings manipulator” as defined by Beneish

(1999) is a firm that: (1) is growing quickly (extremely high year-over-year sales); (2) is

experiencing deteriorating fundamentals (as evidenced by a decline in asset quality, eroding

profit margins, and increasing leverage); and (3) is adopting aggressive accounting practices

(receivables growing much faster than sales; large income-inflating accruals; decreasing

depreciation expense).

In this study, we conjecture that firms which share common traits with past earnings

manipulators (i.e. those who “look like manipulators”) represent a particularly vulnerable type of

growth stock to investors. Because of their high recent growth trajectory, these firms are more

likely to be richly priced. At the same time, they exhibit a number of other potentially

problematic characteristics (either lower earnings quality or more challenging economic

conditions).4 Although the accounting games they engage in might not be serious enough to

warrant prosecution, we posit that their earnings trajectory is more likely to disappoint investors

(i.e., they have lower “earnings quality”). To the extent that the pricing implications of these

accounting-based indicators are not fully transparent to investors, firms that “look like” past

earnings manipulators will also earn lower future returns.

                                                            2 By using the originally published coefficients, we avoid data-snooping and peek-ahead issues. Although these weights are unlikely to be optimal for either fraud detection or returns prediction, our goal is not so much to optimize these outcomes as to demonstrate the out-of-sample efficacy of forensic accounting techniques. 3 On average, the model flags 17.7% of all firms as potential manipulators; therefore, the “hit rate” of 71% among these large fraud cases is highly statistically significant. 4 The term “earnings quality” has been used in academic literature to describe seemingly divergent ideas, such as the ability of earnings to faithfully represent past performance, or to more accurately predict future performance. In this paper, we use earnings quality to describe its “persistence”. To us, “higher quality” refers to earnings that are less transitory, and more likely to persist (or endure) in the future. Consistent with common usage among analysts, higher quality earnings by our definition deserve a higher price-to-earnings multiple.

3  

We find that firms with a higher probability of manipulation (MSCORE) earn lower

returns in every decile portfolio sorted by Size, Book-to-Market, Momentum, Accruals,

and Short-Interest. These returns are economically significant (averaging just below 1%

per month on a risk-adjusted basis), and survive the usual risk controls. We further

show that a large proportion of the abnormal return is earned in the short three-day

windows centered on the next four quarterly earnings releases, suggesting our results are

due to a delayed reaction to earnings-related news rather than risk-based factors. The

robustness of these results, even among highly liquid firms, suggests they are unlikely to

be fully explained by transaction costs.

Having documented MSCORE’s ability to predict risk-adjusted returns, we aim to better

understand the nature of the information it conveys. In particular, extensive evidence

exists that Accruals, as well components and variants of Accruals, predict one-year-

ahead returns (e.g., Sloan 1996; Fairfield, et al. 2003; Richardson, et al. 2005; Cooper et

al. 2008; Hirshleifer, et al. 2010). Given its close ties to Accruals, we wish to better

understand: (1) How is MSCORE different from Accruals in terms of each variable’s

predictive ability; and (2) What is the source and nature of MSCORE’s incremental

predictive power over Accruals?

We perform three sets of analyses. First, we conduct detailed tests on the joint ability of

Accruals and MSCORE to predict returns. We find that the dominance of MSCORE

over Accruals is evident in both independent sorts and nested sorts. When firms are

sorted on these two variables independently, MSCORE is particularly effective in

predicting returns among low Accruals firms (i.e. firms that have “high earnings quality”

according to their accrual ranking). In the lowest Accrual quintile, the spread in size-

adjusted returns between high MSCORE firms and low MSCORE firms is -19.9% over

the next 12 months. Among firms in the second lowest Accrual quintile, the spread is -

10.8% per year. In general, after controlling for MSCORE, Accruals exhibits limited

predictive power, which is concentrated primarily in the mid-MSCORE quintiles.

4  

Second, we analyze individual components of the model in specific subpopulations. Noting that

MSCORE is particularly effective in separating future winners from losers among low accrual

firms, we design a difference-in-difference test that sheds light on the element(s) of the model

that most contributed to this effect. Our analysis shows that the variables relate to the

predisposition to commit fraud, rather than the variables associated with the level of aggressive

accounting, are the primary drivers of the incremental power of the model. Specifically, the

model’s incremental predictive power in the low-accrual group is most directly related to its

ability to identify fast-growing firms that have recently experienced some recent economic head-

wind. Elements of the model associated with aggressive accounting (i.e. those components most

closely aligned with Accruals), provide no incremental predictive power.

Third, we show that the Beneish model’s efficacy is associated with its ability to predict

the persistence of firms’ current year accruals (i.e. whether the accrual component of

this year’s earnings will continue into next year, or will disappear). Specifically, we

find that high MSCORE firms (firms that “look like manipulators”) have income-

increasing accruals that are much more likely to disappear next year, and income-

decreasing accruals that are much more likely to persist, or re-appear, next year. We

observe the exact opposite among low MSCORE firms (firms that look “least like

manipulators”). In other words, MSCORE is providing useful information about the

future change in accruals – i.e. it is informative about the “quality” of the accrual

component of current-year earnings.

This study is related to several strands of existing research. First, our findings help to shed light

on a persistent puzzle in financial markets – i.e., the role and value of fundamental analysis in

active investing. Our evidence shows that in recent years, careful fundamental analysis can

provide investment value, even among pubic firms. To the extent that some of the strategies

used by informed traders are correlated with the model we test, our findings help to explain the

continued existence of active managers who practice this form of analysis.

Second, our study is related to a growing literature on the effective use of firm’s financial

information. Aside from fraud detection, academics have used financial information in different

5  

contexts to better understand: earnings quality (e.g., Lev and Thiagarajan (1993), Sloan (1996)),

Richardson et al. (2005)), bankruptcy risk (e.g., Altman (1968), Ohlson (1980)), the direction of

future earnings (e.g., Ou and Penman (1998)), and future returns (e.g., Holthausen and Larcker

(1992), Piotroski (2000), Beneish, Lee and Tarpley (2001), Mohanram (2005)). We build on,

and extend, this line of research by documenting the usefulness of fraud detection techniques for

both earnings quality assessment and returns prediction. In addition, by parsing the different

elements of the model, we provide new insights on how and why fraud detection techniques

work. These insights point to new directions for assessing firms’ earnings quality, and should

enhance future efforts to identify potential over/under valuations.5

Finally, our findings extend the literature on market learning and the limits of arbitrage (e.g.,

Berk and Green (2004), Bebchuk et al. (2011), Green et al. (2011)). At the heart of issue is the

definition of “publicly available” information. While all the individual components of the model

are publicly available, our evidence suggests that public availability alone does not ensure these

elements are fully integrated into price in a timely manner. Learning takes time. Indeed, our

results suggest that even publication in a journal does not guarantee the usefulness of a strategy

is transparent to the market immediately.6

Overall, our analyses provide substantial support for the use of forensic accounting in equity

investing. We show a model built and estimated in the early 1990’s has been effective in

detecting some of the most famous fraud cases that occurred in a subsequent period. Moreover,

we find this model has incremental ability to predict stock returns out-of-sample, beyond the

usual suspects. Our evidence indicates the efficacy of the model derives from its ability to

separate firms whose accruals are more likely to persist from those whose accruals are more

likely to reverse. Since Beneish developed this model using forensic accounting principles, and

the parameters for the model were estimated in a prior period, our findings offer out-of-sample

validation for the approach advanced by forensic accountants.

                                                            5 In Section 2, we provide a more detailed discussion of how our research is related to, but distinct from, these prior studies. 6 A similar pattern was observed with the original Accrual anomaly documented by Sloan (1996), which had strong out-of-sample predictive power for returns for a number of years after publication.

6  

The remainder of the paper proceeds as follows. In the next section, we discuss related

literature. In Section 3, we examine the Beneish model and its conceptual underpinnings. In

Section 4 we explore its ability to predict cross-sectional returns. Finally, in Section 5, we

summarize and discuss the implications our findings.

2. Related Literature

Our paper is related to two main bodies of work: (1) the large literature on Accruals and future

returns, and (2) studies that use financial statement variables in other decision contexts. In this

section we relate the Beneish (1999) fraud detection model, in broad terms, to these prior studies.

In Section 3, we discuss the model in more detail.

2.1 Accruals and Future Returns

In a seminal study, Sloan (1996) shows that firms with higher (more income inflating) accruals

earn lower returns than firms with lower accruals. He provides evidence that the predictive

power of accruals derives from the fact that the cash-based component of earnings is more

persistent – i.e., is of “higher quality”, than the accrual-based component. Sloan’s original study

spawned a large literature seeking to explain, confirm, refute, or recast, his findings.7 Although

most of these studies confirm or extend the original findings, a few raise questions as to its

robustness or interpretation.8 Most recently, the focus has shifted to limits to arbitrage and

market learning related explanations.9

Although our study is related to the accrual literature, the model we test was designed for a

different purpose, and the individual components of the model reflect this broader mandate. In

the Appendix and in the next section, we present a detailed description of the Beneish (1999))

model, as well as the intuition behind each of the eight accounting-based variables used in its

estimation. In broad terms, the profile of an earnings manipulator that emerges from the Beneish

                                                            7 For example, Chan et al. (2001), Collins and Hribar (2002), Fairfield, et al. (2003), Hirshleifer, et al. (2004), Richardson, et al. (2005), Cooper et al. (2008), and Hirshleifer et al. (2011).   Other related studies examine whether different market constituents understand the implications of Accruals (e.g., Bradshw et al. (2011), Collins et al. (2003), Barth and Hutton (2004), and Beneish and Vargus (2002)). See Zach (2011), chapter 2, for a good summary. 8 For example, Desai et al. (2004) question whether the accrual anomaly is a manifestation of the value-glamour anomaly, and Khan (2005) examines a model misspecification explanation for accrual mispricing. 9 For example, Mashruwala et al. (2006), Lev and Nissim (2006), and Green et al. (2011).

7  

model is a firm that is: (1) growing extremely fast (Sales Growth Index); (2) experiencing some

economic headwind (Asset Quality Index; Gross Margin Index; SGA Index; Leverage Index);

and (3) practicing aggressive accounting (Days in Receivables; Depreciation Index; Accruals to

total assets).10

Of these three categories, only the last group is aligned with the Accrual measures. Explanatory

variables from the first two groups are designed primarily to detect the propensity to commit

fraud (i.e., exceptionally high year-over-year sales growth, eroding profit margins, a decline in

asset quality, and an increase in leverage), rather than the effect of aggressive accounting. In

other words, five out of eight variables in Beneish’s fraud detection model are actually not

associated with various manifestations of accruals.

In fact, our test results show that the incremental power of the model over Accruals stem mainly

from variables in the first two categories, and is not a function of aggressive accounting per se.

Specifically, our findings show that the efficacy of MSCORE is associated with its ability to

predict changes in current-year Accruals. In other words, MSCORE is providing additional

information about the quality of earnings beyond the current year level of reported Accruals.

2.2 Other Financial Analysis Research

Our study is also related to prior studies that examine the usefulness of financial statement

information in other decision contexts. Broadly speaking, these studies fall into three categories:

(1) Distress Analysis: studies that aim to predict bankruptcy risk and financial distress (e.g.,

Altman (1968), Beaver (1966), Ohlson (1980), Beaver et al. (2005)); (2) Basic Ratio Analysis:

studies that combine a large set of financial ratios in a less structured manner, to predict either

future earnings or stock returns (e.g., Ou and Penman (1989), Holthausen and Larcker (1992));

and (3) Contextual Analysis: studies that apply financial analysis in targeted settings, such as

among value stocks (Piotroski (2000), growth stocks (Mohanram (2005)), or extreme performers

                                                            10 We use three broad, and not necessarily mutually exclusive, categories to help explain the main motivations behind the eight variables. In fact, some variables (such as Days in Receivables and Asset Quality) could be indicators of either economic challenges or aggressive accounting, and probably contain elements of both.

8  

(Beneish et al. (2001)). While all these studies use accounting information, each group is

designed for a different purpose, and feature different financial variables.11

The primary distinguishing characteristic of Beneish (1999), compared to these studies, is its

close allegiance to the fraud detection literature espoused by financial practitioners. For

example, the model is related to Foster (1979)’s analysis of Abraham Briloff, an accounting

professor who successfully used public reports to identify accounting irregularities. Other

sources that influenced the original model include O’Glove (1987), Kellogg and. Kellogg (1991),

Siegel (1991), and an earlier edition of Schilit (2010). Unlike Ou and Penman (1989) and

Holthausen and Larcker (1992), Beneish (1999) does not use a large number of ratios; and unlike

the contextual studies, this model was not designed to predict returns in particular subpopulations

of firms. Rather, the model was built by carefully observing the financial information most often

discussed by experts in forensic accounting, who are concerned with detecting accounting

irregularities, particularly among fast growing companies.12

In sum, prior studies have examined both the usefulness of financial information in a variety of

contexts. Our study extends this literature by focusing on a model built on forensic accounting

principles, and demonstrating its strong out-of-sample ability to both identify famous fraud cases

and predict cross-sectional stock returns. We also provide new evidence on why forensic

accounting techniques work, and how they inform us about firms’ quality of earnings.

3. The Earnings Manipulation Detection Model

Beneish (1999) profiles firms that manipulate earnings (firms either charged with manipulation

by the SEC, or admitted to manipulation in the public press) and develops a statistical model to

discriminate manipulators from non-manipulators. The model presented in Beneish (1999) relies

exclusively on financial statement data and is thus useful in assessing fraud potential in firms

without appeal to security prices (for example, in pricing an initial public offering). In the

                                                            11 The bankruptcy prediction models, for example, are concerned with a firm’s “distance from default” and feature variables that measure profitability, leverage, and overall health. The “contextual analysis” studies are concerned with indicators of improving fundamentals for different types of stocks. 12 In taking seriously the work of financial practitioners, Beneish (1999) is similar in spirit to Lev and Thiagarajan (1993) and Abarbanell and Bushee (1997). However, unlike these studies, the Beneish model was built specifically to detect the both the propensity to commit accounting fraud, and the effect of such activities.

9  

original paper, this model was estimated using data from the period 1982-1988 and its holdout

sample performance assessed in the period 1989-1992.

Since the publication of the original study, this model has attained some notoriety after flagging

Enron well in advance of its eventual demise.13 It has been featured in financial statement

analysis textbooks (e.g., Fridson 2002, Stickney et al. 2003) and in articles directed at auditors,

certified fraud examiners, and investment professionals (e.g., Cieselski 1998, Merrill Lynch

2000, Wells 2001, DKW 2003, Harrington 2005). However, evidence of its out-of-sample

performance is ad hoc and anecdotal.

The Appendix provides a detailed description of each variable, loadings on these variables, and a

tabulation of the sample distribution over time and across industries. Table A.1 provides

descriptive statistics for these variables and Table A.2 provides information on the incidence of

manipulation by time period, as well as by industry. These reveal an increasing frequency of

SEC Accounting and Auditing Enforcement Actions over the sample period, and, unsurprisingly,

a higher concentration of manipulators in software, hardware and retail concerns (13.5%, 9.5%

and 6.8% of the sample manipulators).

As evidenced by Table A.1, compared to a same-industry control group, potential manipulators

are growing extremely fast (58.1% year-over-year sales growth; compared to 13.0% for control

firms). Yet despite this ultra-fast sales growth, receivables as a percentage of sales have

increased even faster (average growth in Days in Receivables of 41.2%; compared to 3.0% for

control firms). Manipulators also have deteriorating fundamentals (for example, their Asset

Quality Index shows, on average, a 22.8% increase in the proportion of “soft assets” to total

assets, compared to 3.0% for control firms; their Gross Margin Index shows gross profit margins

declined by an average of 15.9%, compared to 1.7% for the control firms). Finally, they exhibit

more aggressive accounting (Depreciation expense as a percentage of gross PP&E decreased by

7.2% compared to 0.7% for control firms; Accruals are income-inflating to the tune of 4.9% of

                                                            13 The model gained widespread recognition when a group of MBA students at Cornell University posted the earliest warning about Enron’s accounting manipulation score using the Beneish (1999) model a full year before the first professional analyst reports (Morris 2009). This episode in American financial history is preserved in the Enron exhibit at Museum of American Finance, New York (www.moaf.org) and is also recounted in Gladwell (2009).

10  

total assets, compared to 1.5% for control firms). In short, the manipulators as a group appear to

be growth firms that are running into some potential problems.

To compute a probability of manipulation (MSCORE), Beneish estimated a Probit regression

using a portion of his sample, and validated the model’s efficacy through a holdout sample.14 As

in all statistical discriminate analyses of this genre, Type I classification error (the probability of

“missing a culprit”) needs to be traded off against Type II classification error (the probability of

“nabbing an innocent firm”). Beneish (1999) contains an extensive discussion of the relative

implicate cost of the two error types for alternative MSCORE threshold (or “cutoff”) values. His

results show that the model performs best relative to a naïve model when the relative cost of

Type I to Type II error is between 20:1 and 30:1 (Beneish; Table 5). This relative cost function

corresponds to a MSCORE cut-off value of -1.78 (i.e. firms with a MSCORE that exceeds -1.78

would be flagged as potential manipulators).15 Using this cut-off value, the author demonstrated

that the model flagged approximately 13% of the holdout sample firms as manipulators.

Strikingly, those 13% of the sample firms that were flagged actually included approximately half

of all manipulators.

In this study, we closely replicate the model as published in Beneish (1999). We use the exact

coefficients as estimated for the model (Beneish; Table 3). We also used the same cut-off value

implied by the original study to classify firms as manipulators (i.e. MSCORE values exceeding -

1.78). For seven of the eight explanatory variables, we follow exactly the same construction as

the original paper. For the Accruals variable, the original paper used a Balance Sheet estimation

method (i.e. Sloan (1996)). For this study, consistent with the evolution in the accruals literature

since 1999, we derived the same conceptual measure as Beneish (1999), but used information

from the Statement of Cash Flows rather than the noisier Balance Sheet estimate.16

                                                            14 Beneish (1999) estimated an unweighted and weighted Probit regression. Our discussion refers to the unweighted Probit model. 15 The issue of classification errors is less pertinent to this study as for the purpose of return prediction, MSCORE often appears as a continuous variable. Nevertheless, it seems to us that a cost ratio of Type I to Type II error in the neighborhood of 20:1(or 30:1) is quite reasonable for investment purposes as well. This is because to most active asset managers, the cost of erroneously including an earnings manipulator in a given portfolio is much higher than the cost of erroneously omitting an innocent firm. 16 This option was not available to Beneish (1999) because the Statement of Cash Flows only became required disclosure in the U.S. in 1987, and much of his sample pre-dates the adoption of the statement. As a practical matter, in the absence of major business acquisitions, the two methods yield very similar estimates.

11  

In Table 1, we demonstrate the continued relevance of the model to detect fraud by examining its

performance for well-known fraud cases from 1998-2002 (as reported by auditintegrity.com).

This period was marked by an unusual number of high profile fraud cases that helped spur

forensic accounting to prominence. In its aftermath, Auditintegrity.com listed 20 firms that it

deemed to be the most egregious examples of accounting earnings manipulation. Two of these

firms are financial stocks (to which this model does not apply); one is not an actual fraud case

(Motorola did not manipulate its own earnings, it only abetted Adelphia). Since the holdout

sample for the original Beneish model ended in February 1993, the remaining 17 firms identified

by Auditintegrity.com represent an entirely out-of-sample test for the model.

As Table 1 shows, the model predicted the fraud in 12 of the 17 firms, including Cendant, Enron,

Global Crossing, Qwest and several other famous cases. On average, the model detected the

fraud using financial information that was available a year and a half before the public revelation

of its problems. Of particular note, the model received attention subsequent to the Enron scandal

as the investing public discovered that the model had flagged Enron prior to the debacle.17

In sum, the model appears to have performed quite well in identifying the most famous

accounting fraud cases that surfaced after its publication. In the next section, we examine its

ability to predict one-year-ahead stock returns in a broad cross-section of firms.

4. Does MSCORE predict future returns?

4.1 Sample

We select the initial sample from the Compustat Industrial, Research, and Full Coverage files for

the period 1993 to 2007. We eliminate (1) financial services firms (SIC codes 6000 – 6899), (2)

firms with less than $100,000 in sales (Compustat #12) or in total assets (Compustat #6), (3)

firms with market capitalization of less than $50 million at the end of the fiscal period preceding

portfolio formation, and (4) firms without sufficient data to compute the probability of

                                                            17 On January 25th, 2002, the Wall Street Journal reported that in seizing e-mails at Arthur Andersen, Congress found evidence that the Chicago office of Arthur Andersen had issued two “alerts” to the Houston office in the spring of 2001 with respect to earnings manipulation at Enron. The alerts came from a tailored version of the model that Beneish had estimated under a consulting relationship with Andersen. (“Andersen Knew of `Fraud' Risk at Enron --- October E-Mail Shows Firm Anticipated Problems Before Company's Fall”, 01/25/2002, A3).

12  

manipulation. Following Beneish (1999), we winsorize the predictive variables in the model at

the 1 percent and 99 percent levels each year in our sample period to deal with problems caused

by small denominators and to control for the effect of potential outliers.

We compute size-adjusted returns following a slightly modified version of the procedures

outlined in Lyon, Barber, and Tsai (1999).18 To form reference portfolios, we first identify

decile portfolio breakpoints based on all NYSE firms. We then assign all NYSE, AMEX, and

Nasdaq firms to portfolios based on those breakpoints. The smallest portfolio has a

disproportionately large number of stocks, so we further sort those stocks into five portfolios

based on market cap. The end result is 14 size-based portfolios. We then accumulate returns for

12 months starting with the first day of the next month following portfolio assignment. If a firm

delists, we include returns to the delist date as well as any delisting return reported by CRSP. If

a delist return is missing, we estimate it using the procedures outlined in Beaver, McNichols, and

Price (2007). As in Lyon, Barber, and Tsai (1999), from the month following delisting to the

end of the holding period, we assume the proceeds from delisting, if any, were invested in the

CRSP size-based portfolio to which the firm belongs.

To compute size-adjusted returns, we accumulate returns for twelve months starting with the

fifth month after year end using the same delisting procedures described above, if necessary. We

use the stock’s market cap at the end of the fourth month following the fiscal year end to identify

its reference portfolio. We then subtract the return for the reference portfolio from the return for

the firm.

To ensure that the trading strategies that we examine are implementable, we require all firms

used in our rankings to have stock return data available in the CRSP tapes at the time rankings

are made, and use prior year decile cut-offs to assign firms to deciles of the ranking variable

(e.g., the probability of manipulation, accruals, momentum, etc.) in the current year. Our trading

                                                            18 Although Lyon, Barber, and Tsai (1999) form reference portfolios once per year, we perform our sorts and form reference portfolios monthly. This is because the return windows for our stocks are not aligned by calendar date (i.e. they begin in the fifth month after the end of the fiscal year for each stock).

13  

strategy return computations are based on taking positions four months after the end of the fiscal

year. The final sample consists of 41,544 firm-year observations from 1993 to 2009.19

4.2 MSCORE and future returns

Although Table 1 and prior research (e.g., Beneish 1999) demonstrate the ability of the Beneish

model to identify firms that commit fraud, very few fraud cases actually lead to indictment.20

The Beneish model, however, flags 17.7% of firm-year observations as potential frauds. As

discussed earlier, the much higher number of potential frauds flagged by the model is a function

of a trade-off between the costs of Type I and Type II errors. Based on results reported in the

original study, as well as an intuitive assessment of costs to an asset manager of missing a

manipulating firm, we adopted a threshold (“cut-off”) value for MSCORE of -1.78.21

Table 2 compares returns for firms that are flagged as probable manipulators to the returns of

firms that are not flagged using this cut-off value. Overall for the full sample, flagged firms

generate one-year-ahead size-adjusted returns of -7.5%, while firms that are not flagged

experience positive returns of 3.2%. Both of these average returns are statistically significant.

Firms that are not flagged outperformed flagged firms by 10.7% on average, and this is also

statistically significant.

Table 2 also compares flagged and not-flagged firms by year. The spread in returns across not-

flagged and flagged firms is negative in only four years (1994, 2002, 2004, and 2008), but is not

significantly negative in any year. Firms that are not flagged significantly outperform flagged

firms in eleven years. Overall, Table 2 suggests flagged firms (firms that merely “look like a

manipulator”) are associated with lower expected returns.

                                                            19 Returns from CRSP are not available after December 2010. A one year return window along with a four month period between fiscal period end and the start of the holding period results in August, 2009 as the latest date for financial statement data. Thus, 2009 has only 428 observations. 20 Beneish (1999) examines 74 cases of fraud from 1982 to 1993. Dechow, Ge, Larson, and Sloan (2011), who investigate Accounting and Audit Enforcement Actions (AAERs), report less than 0.5% of the firm-years in their sample are associated with fraud. 21 We also computed results for a wide range of cut-off values around -1.78, and the results are quite similar. In later tests, we use MSCORE as a continuous variable and find robust predictive power.

14  

4.3 Distinguishing MSCORE from other predictors of future returns

Prior research shows that a number of characteristics are correlated with subsequent returns: (1)

the difference between earnings and cash flows from operations (Accruals), following Sloan

(1996); (2) price momentum (Momentum), following evidence in Jegadeesh and Titman (1993)

that past 3 to 12 month returns tend to continue in the subsequent year; (3) firm size (MVE),

following evidence in, among others, Fama and French (1992); (4) the book-to-price ratio

(BTM), following evidence in Davis (1994), Haugen and Baker (1996), and Fama and French

(1992); and (5) the short interest ratio (SIRatio) following evidence in Drake et al. (2011) that

firms with high short interest ratios subsequently earn lower returns.

In Table 3, we report the correlation matrix for these characteristics. Pearson correlations are

above the diagonal and Spearman correlations are below the diagonal. Correlations of MSCORE

with three variables are noteworthy. First, MSCORE and Accruals are highly correlated

(correlation = 0.444, p < 0.001). Many observers speculate that earnings management is an

important reason why the persistence of accounting accruals differs from those of cash flows,

suggesting that earnings management misleads investors. Thus, it is possible that both

MSCORE and Accruals measure earnings manipulation and that little incremental value exists in

studying MSCORE. Second, the negative correlations between MSCORE and both Momentum

and BTM suggest that firms with high probability of earnings overstatement have “momentum”

(-0.039, p-value<0.001) and “glamour” characteristics (low BTM, -0.021, p-value<0.001)).

Third, the correlation between MSCORE and Short Interest Ratio is positive and significant

(0.038, p-value<0.001), consistent with high MSCORE firms attracting the attention of short

sellers.

To examine whether the returns to a strategy based on MSCORE are subsumed by other

potential predictors of future returns, we estimate Fama-MacBeth cross-sectional regressions of

one-year-ahead buy and hold size-adjusted returns (BHSARt+1) on scaled decile ranks of several

predictors:

BHSARt+1 = a0 + a1MSCOREt + a2 Accrualst + a3Momentumt + a4MVEt

+ a5BTMt + a6 SIRatiot + et+1 (1)

15  

In Table 4 Panel A, we report the coefficients from 17 annual cross-sectional regressions based

on equation (1), as well as their time-series average. The dependent variable is one-year-ahead

size-adjusted returns computed using reference portfolio returns computed as outlined in Lyon et

al. (1999). The independent variables are scaled decile ranks, so the values range from 0 to 1.

The results indicate that scaled MSCORE ranks are negatively correlated with one-year-ahead

abnormal returns (-0.093, t-statistic=-2.62), and that Momentum is positively correlated with

one-year-ahead abnormal returns (0.091, t-statistic=2.63). BTM is also positively correlated with

future abnormal returns (0.075, t-statistic=2.00). The remaining variables, including Accruals,

MVE, and SIRatio do not attain significance. This suggests that after controlling for Accruals

and other variables associated with future returns, a long-short portfolio strategy based on

extreme MSCORE deciles earns a 9.3% abnormal return one-year-ahead.22

To better understand the source of the abnormal return from a MSCORE strategy, we also

estimate time-series regressions of monthly excess returns on the Fama-French (1993) factors as

well as the momentum factor from Carhart (1997):

ExRett = a0 + a1MKTt + a2 SMBt + a3HMLt + a4WML + et+1 (2)

Where ExRet is the monthly value-weighted MSCORE portfolio return in excess of the return on

the one-month T-Bill, MKT denotes the value-weighted market index return in excess of the

one-month T-Bill, SMB denotes returns on a factor mimicking portfolio for the size factor (small

minus big), HML denotes the book-to-market factor (high minus low), and WML denotes the

momentum factor (winners minus losers). To calculate ExRet, firms are sorted into portfolios

each month based on the firm’s most recent MSCORE and the prior year MSCORE cutoffs. We

calculate value-weighted returns for each portfolio each month using market value of equity at

the beginning of the month. We then subtract the return on the one-month T-Bill to calculate the

                                                            22 As with all of our analyses, we do not explicitly incorporate transactions costs and short-selling constraints. These costs vary by market participant, so we leave it to the reader to consider their impact.

16  

excess return (ExRet). Each portfolio consists of 210 monthly observations from July 1993 to

December 2010.

In Table 5 Panel A, we report results for the three Fama-French factors and we augment the

model with the momentum factor in Panel B. The results are similar for both panels. Focusing

on Panel B, our results show that, on average, a MSCORE strategy involves a negative bet on

market beta (average coefficient on MKT of -0.20) as well as a positive exposure to differential

expected returns based on firm size (average coefficient on SMB of 0.70). On average, a

MSCORE strategy has no significant exposure to either the Value (HML) or Price Momentum

(WML) factor.

More importantly, after controlling for monthly correlations with all four factor mimicking

portfolios, a hedged portfolio of extreme MSCORE decile firms (D1-D10) has a positive

intercept (i.e. monthly alpha) of around 0.94% (11.3% per annum), which is highly statistically

significant. This effect is most pronounced in the highest MSCORE decile – i.e. most of the

return derives from a negative bet against growth firms that subsequently perform much worse

than expected. As a group, these firms underperform their size peers by around 75 basis-points

per month.

To further evaluate MSCORE’s ability to predict returns in various sub-populations of stocks, we

separate the firms in each decile of MVE, BTM, Momentum, SIRatio and Accruals into high and

low MSCORE, where high MSCORE includes the 7,357 observations in our sample that are

flagged as potential manipulators. The results, reported in Table 6 and Figure 1, are striking. The

performance of the flagged sub-sample is worse than that of its not-flagged counterpart in all 50

decile breakdowns. Furthermore, the average size-adjusted return of the flagged firms is

negative in 48 of the 50 deciles.

In Panel A, a size-based trading strategy that buys small firms (decile 1) and shorts large firms

(decile 10) yields 2.7% per year. By combining MSCORE with size, e.g., buying small not-

flagged firms and selling short large flagged firms, the strategy yields 15.2% per year. Similarly,

in Panel B, we show that the improvement to a BTM strategy is also quite substantial. Buying

17  

Value (decile 10) and shorting Glamour (decile 1) yields 9.2%. By combining with MSCORE –

i.e., buying Value firms that are not-flagged and selling Glamour firms that are flagged firms –

the strategy’s yield improves to 16.7% per year.

In Panel C, a momentum based trading strategy that buys high (decile 10) and shorts low (decile

1) yields 10.7%. Combining with MSCORE – i.e. buying high momentum firms that are not-

flagged and selling short low momentum firms that are flagged – the strategy’s yield improves to

25.3%. Similarly in Panel D, trading on extreme short interest ratio deciles yields an abnormal

return of 5.3% per year, which improves to 15.8% per year by superimposing MSCORE.

Finally in Panel E, Accruals alone returns 8.0% and this yield can be improved to 14.1% per year

by selling short high accrual flagged firms and buying low accrual firms that are not flagged.

Note that MSCORE is especially helpful when accruals provide a positive signal of earnings

quality but MSCORE offers a conflicting signal. When low accrual (generally regarded as “high

quality”) firms are flagged, they return an average abnormal return of -19.8%. As a group, they

underperform low accrual firms that are not flagged by 26.1%.

Given MSCORE’s high correlation with Accruals (Pearson Correlation of 0.444, per Table 3),

we conduct more detailed tests to assess the joint ability of Accruals and MSCORE to predict

returns. In Table 7 Panel A, we report average size-adjusted returns when firms are sorted

independently into quintiles by both Accruals and MSCORE. In Panels B and C of the same

table, we report the results of nested (i.e. sequential) sorts. The strong correlation observed in

Table 3 is apparent in Table 7 Panel A, as approximately 25% of the sample observations reside

in two of the twenty-five portfolios (upper-left and bottom-right).

Again, we see that MSCORE is particularly effective in predicting returns among low accrual

firms. The positive returns for firms with low accruals are concentrated among the low

MSCORE firms. Firms with low accruals and low MSCORE generate returns of 6.4%. In

contrast, firms with low accruals but high MSCORE have strong negative returns (-13.5%). For

firms in the lowest accrual quintile, firms with low MSCORE outperform high MSCORE firms

by 19.9%. In quintile 2, low MSCORE firms outperform by 10.8%. On the other hand, accruals

18  

do not distinguish firms in any of the MSCORE quintiles. The only exception is the high

MSCORE quintile, where the high accrual firms actually outperform low accrual firms.

In Panel B, we further isolate the effect of Accruals and MSCORE by sorting firms on MSCORE

within each Accruals quintile. This allows us to “spread-out” the variation in MSCORE across

firms that have relatively similar Accruals rankings. For low (income-decreasing) accrual firms,

low MSCORE firms outperform high MSCORE firms by 10.4%. For high accrual firms, the

spread is smaller, such that low MSCORE firms outperform high MSCORE firms by 8.1%.

Returns are also significant for the second and fourth accrual quintiles. Interestingly, firms with

large differences in accruals do not have differences in returns when MSCORE is extremely high

or low, but we do observe differences in returns for the three intermediate quintiles.

In panel C, we first sort on MSCORE and then sort on Accruals. Results are consistent across all

MSCORE portfolios: extreme differences in accruals do not result in differences in returns once

firms are sorted on MSCORE. The only exception is the fourth quintile. In contrast, low

MSCORE firms outperform high MSCORE firms across all accrual sorts, and the spread in

returns is strikingly consistent. Overall, Table 7 confirms the findings in Table 5, Panel E (as

well as the evidence on accruals and future returns in Table 4), and demonstrates that MSCORE

dominates accruals as a predictor of future returns.

4.4 The incremental usefulness of individual MSCORE components

The evidence presented thus far indicates that MSCORE has significant ability to predict one-

year-ahead cross-sectional returns. Our results show that this predictive power does not come

from its correlation with Value, Momentum, Size, Accruals, or Short Interest. In the next three

sections, we seek to provide more direct evidence on the nature of the information contained in

MSCORE.

Table 8 compares the mean for each of the seven components of the Beneish model (other than

accruals) in various subpopulations of our sample. To construct this table, we independently

sorted firms into quintiles according to Accruals and MSCORE (same as Table 7 Panel A). We

have seen from Table 7 that the Beneish model is particularly effective in separating out winners

19  

from losers among low accrual firms. In Table 8, we focus sharply on firms in the top right and

top left corners of Table 7 Panel A, to better understand which individual component of the

model is contributing to the efficacy of MSCORE for return prediction among low accrual firms.

The first two rows of Table 8 report the means for each variable for firms in the lowest Accrual

quintile that are also either in the highest or the lowest MSCORE quintile. We then subtract the

mean of the high MSCORE firms from the mean of the low MSCORE firms, and report the

results in row three. As expected, we find that high MSCORE firms have higher means for

variables entering the model with a positive coefficient (DSR, AQI, GMI, SGI, and DEPI), and

have lower means for the variables entering the model with a negative coefficient (SGAI, LEVI).

On their own, these findings are not particularly informative, because we formed the two

portfolios on the basis of MSCORE.

For comparison, we employ a difference-in-difference test design. Specifically, Table 7 shows

that MSCORE has little incremental predictive power for firms in the highest Accrual (quintiles

3, 4, and 5). Exploiting this fact, we first group firms in these three Accrual quintiles, then

further separate them according to their MSCORE score. Once again, we compute the mean for

the same seven variables, and report the results in rows four through six of Table 8. Finally, in

the bottom row, we report descriptive statistics and the result of a statistical significance test for

the difference-in-difference across the two high-low MSCORE portfolios.

The results show that, compared to their counterparts in the bottom half of the table, the high

MSCORE firms in the upper half of the table had significantly higher sales growth (SGI), change

in asset quality (AQI), and increase in leverage (LEVI). In other words, the incremental

predictive power of MSCORE among low accrual firms is driven largely by these three factors.23

                                                            23 Of the three, perhaps the LEVI result is least intuitive, and therefore merits an explanation. When we compute coefficients for High MSCORE minus Low MSCORE firms, SGAI and LEVI should be negative, because they have negative coefficients in the model. Our difference-in-difference test asks whether LEVI is more negative when we are better able to predict returns. The answer is no. In fact, we have greater prediction power in the sample where LEVI is relatively more positive. In other words, the model is more effective in predicting returns when LEVI is relatively more positive for the firms we long (as compared to the firms we short). In short, a sharper increase in a firm’s leverage is bad news, controlling for other factors.

20  

Interestingly, all three factors are primarily indicators of a predisposition to misstate earnings,

rather than the result of the misstatement itself.

4.5 MSCORE and the persistence of earnings components

Many researchers and practitioners speculate that the Accruals variable predicts returns because

it provides information about earnings quality that market participants fail to fully utilize. In this

section, we pursue this line of reasoning, and examine the extent to which MSCORE is an

indicator of a firm’s earnings quality. Specifically, we examine whether and how MSCORE

might help us forecast firms’ future earnings.

The notion that MSCORE might be useful in predicting future earnings is intuitive. As we have

seen, high MSCORE firms are growing rapidly in terms of top-line sales. Moreover, these firms

face operating environments that are becoming more challenging (e.g., declining margins and

decreasing asset quality). Finally, these firms have adopted more aggressive accounting

practices in the most recent reporting period (higher receivables to sales, more income-inflating

accruals, lower depreciation expense). Future earnings for these firms will be lower (i.e. current

earnings will be less persistent) if either of two conditions occurs: (a) current period accounting

distortions are corrected or reversed; or (b) the firm succumbs to the difficult economic

conditions that are beginning to appear in its financial results.

We focus on the incremental ability of MSCORE to predict the persistence of firms’ earnings.

Prior studies have demonstrated that the cash flow component of earnings is more persistent than

the accruals component (see, for example, Sloan (1996) and Richardson et al. (2005)). We

extend this analysis by exploring the differential persistence of one-year-ahead accruals for high

and low MSCORE firms. If the forensic accounting principles that underpin this model are

useful in separating firms with relatively high/low quality earnings, this fact should be evidenced

in a difference in the persistence of future accounting accruals.

Specifically, we predict that income-increasing accruals for firms with high (low) MSCORE

should be less (more) persistent, while income-decreasing accruals for firms with high (low)

MSCORE should be more (less) persistent. In other words, for firms whose accrual component

21  

increases current year income, we expect higher MSCORE to be associated with decreased

accrual persistence (leading to lower income next year). Conversely, for firms whose accrual

component decreases current year income, we expect higher MSCORE to increase accrual

persistence (leading to lower income next year).

To examine whether MSCORE contains such incremental information, we estimate the

following relation between future earnings and current earnings components

EARNt+1 = a0 + a1CFOt + a2 ACCPOSt + a3ACCNEGt + a4ACCPOSt*SPMt

+ a5ACCNEGt*SPMt + a6SPMt + et+1 (3)

Where EARN denotes operating earnings before depreciation, CFO is cash flows from

operations, ACCPOS (ACCNEG) is working capital accruals when these are positive (negative)

and zero otherwise, SPM denotes MSCORE ranked into deciles and scaled to range from 0

(lowest MSCORE) to +1 (highest MSCORE), and all earnings and earnings components are

deflated by average assets in year t.24

We begin Table 9 by providing a frame of reference. In the first two columns, we report that the

current year’s earnings have a persistence coefficient of regression of 0.796 and that cash flows

are more persistent than accruals (0.938 and 0.493). The results for cash flows are similar to

those documented by Sloan (1996) [0.860] but the persistence of accruals is smaller than his

[0.765], largely due to the fact that we use working capital accruals and thus exclude

depreciation. In the third column, we partition accruals into positive and negative samples. Our

results show that, consistent with Beneish and Vargus (2002), the persistence of both positive

and negative accruals are significantly lower than that of cash flows.

In the last column, we report the results of tests examining the persistence of accruals conditional

on MSCORE. We again find strong persistence for CFO (coefficient = 0.975). The coefficients

on ACCPOS and ACCNEG reflect the persistence of positive and negative accruals of firms in

the lowest MSCORE decile (SPM=0). For example, the coefficient of 0.996 on ACCPOS means

                                                            24 Key results are qualitatively identical when the dependent variable is deflated by year t+1 average total assets.

22  

that positive accruals for low MSCORE firms have exceptionally high persistence, indicating

they have a much higher likelihood of repeating. Conversely, the coefficient on ACCNEG is

0.213, indicating that for low MSCORE firms, negative (or income deflating) accruals are much

less likely to repeat.

The coefficients on ACCPOS*SPM and ACCNEG*SPM capture the incremental effect of

MSCORE on accrual persistence as we move from the lowest decile (SPM=0) to the highest

decile (SPM=1). ACCPOS*SPM is negative and significant (coefficient = -0.353) suggesting

that for firms in the highest MSCORE decile (SPM=1), income-increasing accruals are much

less likely to repeat (estimate coefficient= 0.996-0.353= 0.643). For comparison, recall that the

income-increasing accruals for firms in the lowest MSCORE decile are highly persistent (0.996).

Similarly, ACCNEG*SPM is positive and significant (coefficient = 0.555), indicating that for

firms in the highest MSCORE decile, income-decreasing accruals are much more likely to

persist (estimate coefficient = 0.213+0.555= 0.768). In other words, for high MSCORE firms,

any income-decreasing accruals this year will have a higher probability of repeating next year,

leading to lower future earnings. Finally, SPM itself is positive and significant (coefficient =

0.032), suggesting high MSCORE firms (i.e. high growth firms) have higher future earnings

after controlling for differences in accrual persistence. This may reflect the fact that high

MSCORE firms, as a group, are continuing to grow more rapidly than the control firms;

alternatively, it is possible that their inflated earnings in the current year do not fully reverse one-

year-ahead.

In sum, we find that MSCORE is incrementally informative about the persistence of the accrual

component of earnings. First, the one-year-ahead persistence of income-increasing accruals is

significantly lower for high MSCORE firms. Second, any income-decreasing accrual reported in

the current year is much more likely to persist for high MSCORE firms than for low MSCORE

firms. Both results show that MSCORE is useful in predicting future earnings. This is

consistent with MSCORE possessing the ability to either: (a) anticipate the reversal of transitory

distortions in current years’ reported accruals, or (b) predict worsening economic conditions that

will manifest themselves in subsequent accruals.

23  

4.5 Predicting Short-Window Earnings Announcement Returns

Finally, to better understand the information content of MSCORE, we examine abnormal returns

in three-day windows centered on subsequent earnings announcements. The main purpose of

this test is to further discriminate between the market-inefficiency and risk-based explanations

for the predictive power of MSCORE. If MSCORE’s predictive power is due to a delayed

market response to current information about future earnings realizations, and this misperception

is corrected when further information about future earnings is released, then subsequent

abnormal returns should cluster around the future announcements of earnings news. Conversely,

if MSCORE’s predictive power is due to omitted risk variables, we should not observe higher

abnormal returns concentrated around the release of subsequent earnings news.25

Table 10 reports the cumulative abnormal returns (CAR) around the next four earnings

announcements after firms are sorted into MSCORE-based portfolios. Firms are added to

portfolios on the first day of the fifth month following the end of the fiscal year. CAR denotes

the cumulative raw return over days -1, 0, and +1 relative to the earnings announcement date

minus the cumulative return to the benchmark size portfolio to which the firm belongs. To

construct Panel A we sort firms into Flagged and Not Flagged categories based on MSCORE.

We report short-window announcement period CAR for each of the next four quarters, as well as

the total CAR for all four quarterly announcements.

Panel A shows firms that are “Not Flagged” earn positive CAR over each of the next four

quarters, while “Flagged” firms experience on average negative returns. The quarter-by-quarter

results show that short-window returns are reliably more positive for “Not Flagged” (low

MSCORE) firms every quarter. The last column shows that the difference in total CAR over the

next 4 quarters is 2.51%. In other words, approximately one-quarter (25%) of the annual hedge

return to this simple MSCORE strategy is earned in the 12 trading-days (5% of all trading days)

around the next four earnings announcements. This evidence is consistent with a delayed market

reaction to earnings-related news contained in MSCORE, and inconsistent with risk-based

explanations.

                                                            25 Short-window returns are far less susceptible to risk-model misspecifications.

24  

Panels B and C report average earnings announcement returns to alternative hedge strategies

based on MSCORE deciles. To construct this panel, we first sort firms into deciles based on

their current MSCORE and decile cutoffs from the prior year MSCORE distribution. We then

calculate CAR for each decile portfolio (Panel B), as well as for alternative hedge portfolios

(Panel C). In Panel C, we start with the most extreme portfolios (deciles 1 and 10) and

progressively add less extreme MSCORE firms to the long and short positions.

Panel B shows that in each of the next four quarters, earnings announcement returns for low

MSCORE firms (Deciles 1 through 5) are consistently more positive than those for high

MSCORE firms (Deciles 6 through 10). Over our sample period, firms generally received

positive announcement period returns. However, firms in the top two MSCORE deciles

experience consistently negative earnings announcement returns over the next two quarters.

The first row of Panel C shows that when we long decile 1 (low MSCORE) firms, and short

decile 10 (high MSCORE) firms, we generate CAR of 0.66% to 0.86% over each of the next four

quarters. The total CAR over the next four quarters is 2.85%, which is slightly higher than the

strategy reported in Panel A. The following 4 rows show that the hedge returns are not sensitive

to the decile cutoffs. As expected, the average hedge return decreases as we expand the number

of firms included in the strategy. However, the CARs for these hedge strategies are all reliably

positive. Moreover, the earnings announcement returns are, once again, disproportionately large

relative to total annual returns for every hedge portfolio.

5. Summary

Fraudulent financial reporting imposes large costs on financial markets. For example,

shareholders of the firms listed in Table 1 collectively lost over $180 billion dollars when these

accounting ‘irregularities’ were announced.26 Perhaps even more important than the investor

wealth losses are the large welfare costs imposed by fraudulent financial reporting when

resources are misdirected from their most productive use. These accounting misrepresentations

                                                            26Beneish (1999a) and Karpoff et al. (2008) provide evidence of large market value losses to public revelations of accounting manipulation.

25  

increase transactions costs by eroding investor confidence in the integrity of the capital market.

In recent years, we have seen how accounting misrepresentations triggered action by regulators,

who impose (often costly) regulation on firms and markets. In short, when it comes to reporting

frauds, many must pay for the transgressions of a relative few.

Efforts to combat accounting fraud involve both public and private initiatives. On the one hand,

accounting and security market regulators can help curb the practice through legislation and

enforcement actions. On the other, private parties, such as more sophisticated investors, play a

role by identifying firms that are likely to have manipulated earnings, and holding these firms

accountable through market-based disciplining mechanisms.

In this study, we have explored the implications of an earnings manipulation detection model for

equity investors. Using the Beneish (1999) model, which was estimated using data from the

period 1982-1988 and its holdout sample performance assessed in the period 1989-1992, we

show forensic accounting has significant out-of-sample ability to not only detect fraud, but also

predict stock returns. Moreover, we provide evidence that the efficacy of the model derives

substantially from its ability to predict in advance, the likely persistence (or reversal) of the

accrual component of current year earnings.

A key feature of the Beneish model is its focus, not only on the results of aggressive accounting,

but also on managerial predisposition to undertake such action in the first place. As Schrand and

Zechman (2011) observed in their detailed analysis of 49 fraud cases, sometimes a firm’s initial

misstatement reflects not so much a deliberate intent to deceive as simply an optimistic

managerial bias. Our evidence is consistent with this observation, in that much of the

incremental predictive power of the model derives from variables that indicate deteriorating

fundamentals in fast-growing firms, rather than aggressive accounting, per se. These findings

point to new directions for future research on earnings quality. Our hope and expectation is that

these results will spur further work in the area of forensic accounting.

We end with a cautionary note to readers who may be interested in exploiting the return

prediction patterns documented in this study. Looking ahead, it appears likely to us that as the

26  

methods for fraud detection become more sophisticated, so will the techniques of the

perpetrators. For example, the evidence in Kama and Melumad (2011) shows that in recent (post

Sarbanes-Oxley) years, U.S. firms seem to have adopted new methods – such as the factoring of

receivables – to mask the effect of their earnings manipulations. In the case of the Beneish

model, the factoring of receivables will directly affect the usefulness of the DSR variable. Over

time, one should reasonably expect evolving adaptations of this nature to diminish the overall

efficacy of the model for returns prediction.

   

27  

REFERENCES

Abarbanell, J., and B. Bushee. 1997. Fundamental analysis, future earnings, and stock prices. Journal of Accounting Research 35 (Spring): 1-24. Altman, E. I. 1968. Financial ratios, discriminant analysis, and the prediction of corporate bankruptcy. Journal of Finance 23: 589-609. Beaver, W., M. McNichols, and R. Price. 2007. Delisting returns and their effect on accounting-based market anomalies. Journal of Accounting and Economics 43(2-3): 341-368. Bebchuk, L., A. Cohen, and C. Wang. 2011. Learning and the disappearing association between governance and returns. Journal of Financial Economics, forthcoming. 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): 271-309. Beneish, M.D. 1999. The detection of earnings manipulation. Financial Analysts Journal (September/October): 24-36. Beneish, M.D. 1999a. Incentives and penalties related to earnings overstatements that violate GAAP. The Accounting Review (74): 425-457. Beneish M.D., and M. E. Vargus. 2002. Insider trading, earnings quality, and accrual mispricing. The Accounting Review 77(4): 755-791. Beneish, M. D., C. M. C. Lee, and R. Tarpley. 2001. Contextual fundamental analysis through the prediction of extreme returns. Review of Accounting Studies 6: 165-189. Berk, J. B. and R. C. Green. "Mutual Fund Flows and Performance in Rational Markets." Journal of Political Economy 112 (2004), 1269-1295. Bernard, V.L., and J.K. Thomas. 1989. Post-earnings-announcement drift: Delayed price response or risk premium? Journal of Accounting Research 27 (Supplement): 1-36. Carhart, M. M., 1997. On persistence in mutual fund performance. Journal of Finance 52 (March): 57-82. Chan, L.K., N. Jegadeesh, and J. Lakonishok. 1996. Momentum strategies. Journal of Finance 51 (December): 1681-1713. Ciesielski, J., 1998. What’s Happening to the “Quality of Assets”? The Analyst’s Accounting Observer 7(3), February 19. Cooper, M.J., H. Gulen, and M.J. Schill2008.Assetgrowthandthecross‐sectionofstockreturns.JournalofFinance63 August :1610‐1651.

28  

Davis, J. L. 1994. The cross-section of realized stock returns: The pre-COMPUSTAT evidence. Journal of Finance 49 (December): 1579-1593. Dechow, P. M., Ge, W., C.R. Larson and R. G. Sloan. 2011. Predicting material accounting misstatments. Contemporary Accounting Research 28(1): 1-16. Desai, H., S. Rajgopal, and M. Venkatachalam 2004. Value-Glamour and Accruals Mispricing: One Anomaly or Two? The Accounting Review 79 (April): 355-385. Drake, M.S., Rees, L. and E.P. Swanson. 2011. Should investors follow the prophets or the bears? Evidence on the use of public information by analysts and short sellers. The Accounting Review 86 (1): 101-130. Dresdner, Kleinwort, Wasserstein (DKW). 2003. Earnings Junkies. Global Equity Research, October 29, London, U.K. Fama, E. F., and K. R. French. 1992. The cross-section of expected stock returns. Journal of Finance 47 (June): 427-465. Fama, E. F., and K. R. French. 1993. Common Risk Factors in the Returns on Stocks and Bonds. Journal of Financial Economics 33 (1): 3–56. Foster, G. 1979. Briloff and the Capital Market. Journal of Accounting Research 17(1): 262-274. Fridson, M.S. 2002. Financial Statement Analysis: A Practitioner’s Guide. New York: John Wiley & Sons. Gladwell, M. 2009. What the Dog Saw: And Other Adventures. New York: Little, Brown and Company. Gleason, C. A., and C. M. C. Lee. 2003. Analyst forecast revisions and market price discovery. The Accounting Review 78 (1): 193-225. Green, J., J. R. M. Hand, and M. Soliman. 2011. Going, going, gone? The apparent demise of the accruals anomaly. Management Science 57 (5), 797-816. Harrington, C. 2005. Analysis of Ratios for Detecting Financial Statement Fraud. Fraud Magazine. (March/April): 24-27. Haugen, R. A., and N. L. Baker. 1996. Commonality in the determinants of expected stock returns. Journal of Financial Economics 41 (July): 401-439. Hirshleifer, D., Hou, K., Teoh, S., and Zhang, Y., 2004, Do investors overvalue firms with bloated balance sheets, Journal of Accounting and Economics 38, 297-331.

29  

Hirshleifer, D., Hou, K., and Teoh, S., 2011, The accrual anomaly: risk or mispricing? Management Science, forthcoming. Holthausen, R. W., and D. F. Larcker. 1992. The prediction of stock returns using financial statement information. Journal of Accounting and Economics 15: 373-411. Jegadeesh, N. 1990. Evidence of predictable behavior of security returns. Journal of Finance 45 (July): 881-898. Jegadeesh, N., and S. Titman. 1993. Returns to buying winners and selling losers. Journal of Finance 48 (March): 65-91. Kellogg, I. and L.B. Kellogg. 1991. Fraud, Window Dressing, and Negligence in Financial Statements. New York: McGraw-Hill. Khan, M. 2005. Are accruals really mispriced? Evidence from tests of an intertemporal capital asset pricing model. MIT working paper Kraft, A., A.J. Leone, and C. Wasley. 2006. An analysis of the theories and explanations offered for the mispricing of accruals and accrual components. Journal of Accounting Research 44 (May 2006): 297-339. Lev, B., and R. Thiagarajan. 1993. Fundamental information analysis. Journal of Accounting Research 31 (2): 190-215. Lev B. and D. Nissim. 2006. The persistence of the accruals anomaly. Contemporary Accounting Research, 193-226. Lyon, J., B. Barber, C. Tsai. 1999. Improved methods for tests of long-run abnormal stock returns. Journal of Finance 54 (1): 165-201. Mashruwala, C.; S. Rajgopal; and T. Shevlin. 2006, “Why is the accrual anomaly not arbitraged away?” Journal of Accounting & Economics 42, 3–33. Merrill Lynch, 2000. Financial Reporting Shocks. March 31, New York. Mishkin, F. 1983. A Rational Expectations Approach to Macroeconomics, Chicago: University of Chicago Press. Mohanram, P. 2005. Separating winners from losers among low book-to-market stocks using financial statement analysis. Review of Accounting Studies 10: 133-170. Morris, G. D. L. 2009. Enron 101: How a group of business students sold Enron a year before the collapse. Financial History Spring/Summer: 12-15. (www.moaf.org) O'Glove, T. L. 1987. Quality of Earnings. New York: The Free Press.

30  

Ou, J. A. and S. H. Penman. 1989. Financial statement analysis and the prediction of stock returns. Journal of Accounting & Economics 11 (4): 29-46.

Piotroski, J. 2000. Value investing: The use of historical financial statement information to separate winners from losers. Journal of Accounting Research 38: 1-41.

Richardson, S. A., R. G. Sloan, M. T. Soliman, and I. Tuna. 2005. Accrual reliability, earnings

persistence and stock prices. Journal of Accounting & Economics 39 (3): 437-485. Schilit, H. 2010. Financial Shenanigans: How to detect accounting gimmicks and fraud in

financial reports. 3rd Edition, New York: McGraw-Hill. Siegel, J. G. 1991. How to Analyze Businesses, Financial Statements, and the Quality of

Earnings. 2nd Edition, New Jersey: Prentice Hall. Sloan, R.G. 1996. Do stock prices fully reflect information in accruals and cash flows about future earnings? The Accounting Review 71 (July): 289-315. Stickney, C., P. Brown, and J. Wahlen. 2003. Financial Reporting, Financial Statement Analysis, and Valuation, 5th Edition. Thomson-Southwestern. Wells, J. T. 2001. Irrational ratios. Journal of Accountancy. New York: Aug : 80-83. Zachs, L., Editor. 2011. The Handbook of Equity Market Anomalies: Translating Market Inefficiencies into Effective Investment Strategies. New York: Wiley. 

 

31  

Appendix --The Probability of Manipulation 1. Estimating MSCORE The sample in Beneish (1999) consisted of 74 firms that manipulated earnings and 2,332 non-manipulators matched by industry over the period 1982-1992. On average, manipulators were smaller, less profitable, more levered and experienced faster growth than industry controls. We estimate the probability of manipulation to overstate earnings (which we denote MSCORE for ease of exposition) using the following (unweighted) PROBIT model, as published in Beneish (1999): MSCORE= -4.84 + .920*DSR + .528*GMI + .404*AQI + .892*SGI + .115*DEPI-.172*SGAI

+4.679*ACCRUALS - .327*LEVI27

Where:

Variable Name

Description Rationale

DSR (Receivablest[2]/Salest[12])/(Receivablest-1/Salest-1) [Numbers in squared brackets are COMPUSTAT codes]

Captures distortions in receivables that can result from revenue inflation

GMI Gross Margint-1/ Gross Margint, where Gross Margin is 1 minus Costs of Goods Sold [#8]/ Sales

Deteriorating margins predispose firms to manipulate earnings

AQI [1-(PPEt+CAt)/TAt] /[1-(PPEt-1+CAt-1)/TAt-1], where PPE is net [#8], CA are Current Assets [#4]and TA are Total Assets [#6]

Captures distortions in other assets that can result from excessive expenditure capitalization

SGI Salest[12]/Salest-1 Managing the perception of continuing growth and capital needs predispose growth firms to manipulate sales and earnings.

DEPI Depreciation Ratet-1/ Depreciation Ratet, where depreciation rate equals Depreciation [#14-#65]/(Depreciation+PPE [#8])

Captures declining depreciation rates as a form of earnings manipulation.

SGAI (SGAt[189]/Salest[12])/(SGAt-1/Salest-1) Decreasing administrative and marketing efficiency ( larger fixed SGA expenses) predisposes firms to manipulate earnings

LEVI Leveraget /Leveraget-1 where Leverage is calculated as debt to assets [(#5+#9)/#6]

Increasing leverage tightens debt constraints and predisposes firms to manipulate earnings

Accruals to Total Assets28

(Income Before Extraordinary Items [18]- Cash from Operations[308])/ Total Assetst[6]

Capture where accounting profits are not supported by cash profits.

                                                            27 Five of the eight variables in the multivariate estimation are statistically significant (DSR, GMI, AQI, SGI, and ACCRUALS); the remaining three (DEPI, SGAI, LEVI) are not (see Beneish 1999, Table 3). To gain additional insight on the relative importance of the individual inputs, Beneish (1999) re-estimated this model 100 times using 100 random estimation samples. At the 5% level, DSR and SGI were significant in all 100 estimations, Accruals in 95 of the 100 estimations, GMI and AQI in 84 of the 100 estimations. In contrast, DEPI, SGAI and LEVI were only significant in 18, 12, and two estimations respectively (see Beneish 1999, Table 4). 28 Beneish (1999) uses a total accruals variable that is computed slightly differently but yields similar results. This is because before the current presentation of the statement of cash flows became effective (pre-1987) few firms

32  

2. Intuition behind the Eight Variables Broadly speaking, the profile of a “typical earnings manipulator” as defined by Beneish (1999) is a firm that: (1) is growing extremely quickly; (2) is experiencing deteriorating fundamentals (as evidenced by a decline in asset quality, eroding profit margins, and increasing leverage); and (3) is adopting aggressive accounting practices (receivables growing much faster than sales; large income-inflating accruals; decreasing depreciation expenses). More specifically, the model consists of eight financial ratios that can capture either financial statement distortions that result from earnings manipulation (DSR, AQI, DEPI and Accruals) or indicate a predisposition to engage in earnings manipulation due to certain economic conditions (GMI, SGI, SGAI, LEVI).29 Descriptive statistics for these ratios appear in Table A.1 below. Each is constructed so that a higher number increases the likelihood of manipulation. The following is a summary of the intuition behind these variables. Not all eight are individually important, but collectively they create a “composite sketch”, or profile, of a potential earnings manipulator. (1) Rapid Sales Growth

A striking characteristic of the manipulator population is its rapid revenue growth. Table A.1 shows that the mean sales growth rate for the sample of manipulators (SGI) is 1.581, indicating that these firms on average increased year-over-year sales by 58%, compared to a 13.3% rate for an industry control group. Sales growth per se, is not a negative, but in this context, a firm’s high past growth trajectory may increase its predisposition to engage in manipulative behavior.

(2) Deteriorating Fundamentals Several variables are designed to capture deteriorating economic conditions. Specifically, manipulators are hypothesized to have deteriorating gross margins (GMI) and increasing SG&A expenses (SGAI). Their debt-to-asset ratio is increasing (LEVI), and a greater proportion of their total assets reflect non-current and non-PPE investments – i.e. they have increased their proportion of “soft assets” (AQI).

(3) Aggressive Accounting The year-over-year ratio of the manipulators’ receivables-to-sales (DSR) shows their receivables are growing rapidly as a percentage of sales (1.412 for manipulators versus 1.03 for the control group). This indicates an unusual buildup in receivables despite the rapidly increasing sales (a sign of potential revenue inflation). DEPI indicates manipulators tend to have slowed down their depreciation expense as a percentage of their gross PP&E. Finally, Accruals indicates that the reported accounting profits of the manipulators are less supported by cash profits that those of non-manipulators.   

                                                                                                                                                                                                reported cash flow from operations. Our current implementation follows the evolution of this variable in the accruals literature. 29 Some variables, such as DSR and AQI could be indicative of either deteriorating economic conditions or aggressive accounting (i.e. may play a dual role).

33  

Table A.1

Potential Predictive Variables: Descriptive Statistics for the Sample of 74 Manipulators and 2332 Industry-Matched Non-Manipulators in the Period 1982-1992

Manipulators Controls Wilcoxon-Z Median Characteristic Mean Median Mean Median P-Valuea P-Valuea

Days in Receivables 1.412 1.219 1.030 0.995 0.001 0.001 Gross Margin Index 1.159 1.028 1.017 1.001 0.019 0.078 Asset Quality Index 1.228 1.000 1.031 1.000 0.035 0.824 Sales Growth Index 1.581 1.341 1.133 1.095 0.001 0.001 Depreciation Index 1.072 0.977 1.007 0.972 0.346 0.638 SGA Index 1.107 1.028 1.085 0.990 0.714 0.098 Leverage Index 1.124 1.035 1.033 1.000 0.107 0.039 Accruals to total assets 0.049 0.026 0.015 0.012 0.001 0.018

a. The Wilcoxon Rank-Sum test and the Median test compare the distribution of sample

firms' characteristics to the corresponding distribution for non-manipulators. The reported p-values are two tailed and indicate the smallest probability of incorrectly rejecting the null hypothesis of no difference.

34  

3. Incidence of Manipulation The distribution of sample manipulators over the sample period suggests an increasing frequency of SEC Accounting and Auditing Enforcement Actions time is as follows:

Years 1981-1985 1986-1989 1990-1993 Total Number of Firms 8 35 31 74

The distribution of manipulators by two-digit SIC suggest the highest concentration is in Business Services (10 firms, 13.5%), followed by Industrial Products (7 firms, 9.5%), and both Electronic Manufacturing and Wholesales-trade tied at (5 firms, 6.8%).

Table A.2 Manipulators by Two-Digit Industry

SIC Industry Description N % 1 Agricultural Production 1 1.4% 10 Metal Mining 1 1.4% 13 Oil and Gas Extraction 1 1.4% 15 General Building Contractors 1 1.4% 20 Food and Kindred Products 1 1.4% 22 Textile Mill Products 3 4.1% 23 Apparel and Other Textile Products 1 1.4% 24 Lumber and Wood Products 1 1.4% 27 Printing and Publishing 2 2.7% 28 Chemicals and Allied Products 4 5.4% 30 Rubber and Misc. Plastics Products 1 1.4% 34 Fabricated Metal Products 1 1.4% 35 Industrial and Related Products 7 9.5% 36 Electronic & Other Electric Equipment 5 6.8% 37 Transportation Equipment 2 2.7% 38 Instruments and Related Products 2 2.7% 45 Transportation by Air 1 1.4% 47 Transportation Services 1 1.4% 48 Communications 1 1.4% 49 Electric, Gas, and Sanitary Services 3 4.1% 50 Wholesale Trade-Durable Goods 5 6.8% 51 Wholesale Trade-Nondurable Goods 1 1.4% 52 Building Materials & Garden Supplies 1 1.4% 54 Food Stores 1 1.4% 56 Apparel and Accessory Stores 1 1.4%

35  

57 Furniture and Home furnishings Stores 3 4.1% 58 Eating and Drinking Places 1 1.4% 59 Miscellaneous Retail 2 2.7% 70 Hotels and Other Lodging Places 1 1.4% 73 Business Services 10 13.5% 75 Auto Repair, Services, and Parking 2 2.7% 78 Motion Pictures 3 4.1% 80 Health Services 1 1.4% 82 Educational Services 2 2.7%

74 100.0%

36  

Table 1. Performance of Model for High‐Profile Fraud Cases during 1998‐2002  This table reports the 20 companies identified by auditintegrity.com as the “highest profile” fraud cases uncovered during the 1998 to 2002 time period.*  We examine the probability‐of‐manipulation score (MSCORE) for each firm based on financial statement information reported by the firm during the period of alleged manipulation but prior to public discovery.  Firms are flagged as manipulators if MSCORE exceeds ‐1.78 at any time during the period of alleged (or admitted) violation.  We compute MSCORE = ‐4.84 + .920*DSR + .528*GMI + .404*AQI + .892*SGI + .115*DEPI ‐ .172*SGAI + 4.679*ACCRUALS ‐ .327*LEVI.  DSR denotes the ratio of receivables to sales in year t divided by the same ratio in year t‐1. GMI denotes the ratio of gross margin to sales in period t‐1 to the same ratio in period t. SGA denotes the ratio of selling, general, and administrative expense to sales in period t divided by the same ratio in period t‐1. SGI equals sales in t divided by sales in t‐1. DEPI denotes the ratio of depreciation to depreciable base in t‐1 divided by the same ratio in t. AQI equals all non‐current assets other than PPE as a percent of total assets in t divided by the same ratio in t‐1. ACC equals income before extraordinary items minus operating cash flows divided by average total assets. LEVI equals the ratio of long‐term debt +current liabilities to total assets in t divided by the same ratio in t‐1. Year flagged refers to the first year the firm is flagged by the MSCORE model as a manipulator.  Year discovered refers to the year in which the fraud was first publicly revealed in the business press.  Market cap lost denotes the change in market capitalization during the three months surrounding the month the fraud was announced (i.e., months ‐1, 0, +1). Market cap lost (%) denotes the market capitalization lost in the three months surrounding the fraud announcement month, as a percentage of market capitalization at the beginning of month ‐1.  

  Flagged as  Year  Year Market Cap  Market Cap 

Company Name   manipulator?  Flagged  Discovered Lost ($B)  Lost (%) 

         

Adelphia Communications  Yes  1999  2002  4.82  96.8% 

American International Group, Inc.  N/A ‐ Financial     

AOL Time Warner, Inc.  Yes  2001  2002  25.77  32.2% 

Cendant Corporation  Yes  1996  1998  11.32  38.1% 

Citigroup  N/A ‐ Financial     

Computer Associates International, Inc.  Yes  2000  2002  7.23  36.4% 

Enron Broadband Services, Inc.  Yes  1998  2001  26.04  99.3% 

Global Crossing, Ltd  Yes  1999  2002  (Delisted due to bankruptcy) 

HealthSouth Corporation  No    2002  2.31  57.3% 

JDS Uniphase Corporation  Yes  1999  2001  32.49  61.0% 

Lucent Technologies, Inc  Yes  1999  2001  11.15  24.7% 

Motorola  N/A – Only abetted Adelphia     

Qwest Communications International  Yes  2000  2002  9.84  41.8% 

Rite Aid Corporation  Yes  1997  1999  2.83  59.1% 

Sunbeam Corporation  Yes  1997  1998  1.28  58.8% 

Tyco International   No    2002  37.55  58.2% 

Vivendi Universal   No    2002  1.28  27.9% 

Waste Management Inc  Yes  1998  1999  20.82  63.6% 

WorldCom Inc. ‐ MCI Group  No    2002  1.03  69.8% 

Xerox Corporation  No    2000  7.73  43.8%      

      Mean  10.89  51.94% 

      Median  8.79  57.75% 

 * This five‐year period was marked by a large number of corporate accounting scandals.  It also represents an out‐of‐sample test for the Beneish (1999) model, which was estimated using data from 1982‐1988 and tested on a holdout sample from 1989‐1992.   We have no affiliation with AuditIntegrity.com.

37  

Table 2. Year‐by‐year Size‐Adjusted Returns to Flagged Firms  The table reports the year‐by‐year size‐adjusted returns for firms flagged by the Beneish (1999) model and those that were not.  BHSAR denotes annual buy‐and‐hold returns to an equal‐weighted portfolio formed at the start of the first day of the fifth month following the end of the fiscal year, less the returns to a portfolio of firms from the same NYSE/AMEX/NASDAQ size decile (size decile membership determined at the beginning of return window). For firms that delist, any proceeds upon delisting are reinvested in the size portfolio to which the company belongs.  Flagged denotes firms that fit the profile of an earnings manipulator based on the MSCORE model in Beneish (1999) and a cutoff of ‐1.78.  ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.  

Not Flagged  Flagged 

Year N Percent BHSAR   Percent  BHSAR   Spread

1993 2304 82.4% 3.5%***  17.6% ‐5.1%**  8.6%*** 

1994 2474 79.0% 1.4% 21.0% 2.7% ‐1.3%

1995 2786 79.8% 1.2% 20.2% ‐15.4%***  16.5%*** 

1996 2972 77.1% ‐0.7% 22.9% ‐8.7%***  8.0%*** 

1997 3139 76.6% 0.4% 23.4% ‐9.4%***  9.9%*** 

1998 2798 78.1% 9.1%***  21.9% 5.9% 3.2%

1999 2789 77.2% 6.6%***  22.8% ‐21.6%***  28.2%*** 

2000 2616 72.8% 4.1%***  27.2% ‐28.1%***  32.2%*** 

2001 2480 86.6% ‐1.5%*  13.4% ‐17.5%***  16.0%*** 

2002 2278 89.6% 5.7%***  10.4% 11.0%**  ‐5.3%

2003 2550 88.1% 0.8% 11.9% ‐5.5%**  6.3%** 

2004 2571 84.7% 2.2%**  15.3% 5.2% ‐3.0%

2005 2534 87.3% 1.1% 12.7% ‐3.3% 4.4%** 

2006 2480 86.3% 5.3%***  13.7% 1.5% 3.8%* 

2007 2348 86.8% ‐0.7% 13.2% ‐3.5% 2.9%

2008 1997 88.6% 3.1%**  11.4% 6.5%  ‐3.3% 

2009 428 91.1% 6.5%**  8.9% ‐10.3%   16.8%* 

Full Sample  41544 82.3% 3.2%***  17.7% ‐7.5%***  10.7%***  

38  

Table 3. Correlation matrix  This table reports Pearson (above diagonal) and Spearman (below diagonal) correlations for sample variables. MSCORE denotes the probability of earnings manipulation based on the Beneish (1999) model. See the notes to Table 1 for a description of the MSCORE model. Accruals denotes earnings before extraordinary items less cash flows from operations scaled by average assets. Momentum denotes size‐adjusted returns for the six months prior to the BHSAR return window. Market value of equity (MVE) is measured as of end of the fiscal year. Book‐to‐market (BTM) denotes market value of equity divided by common equity. Short interest ratio (SIRatio) denotes the number of shares sold short as a percentage of the number of shares outstanding. BHSAR denotes annual returns starting the first day of the fifth month following the end of the fiscal year, less the returns to a portfolio of firms with comparable size. For firms that delist, any proceeds upon delisting are reinvested in the size portfolio to which the company belongs. ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.  

MSCORE Accruals Momentum Ln(MVE) BTM SIRatio BHSAR

MSCORE  0.444***  ‐0.039***  ‐0.041***  ‐0.021***  0.038***  ‐0.063*** 

Accruals  0.640***  ‐0.035***  ‐0.017***  ‐0.020***  ‐0.009*  ‐0.033*** 

Momentum  ‐0.066***  ‐0.037***  0.036***  ‐0.118***  ‐0.030***  0.025*** 

Ln(MVE)  ‐0.046***  ‐0.048***  0.116***  ‐0.159***  0.126***  ‐0.012** 

BTM  ‐0.115***  ‐0.002 ‐0.281***  ‐0.283***  ‐0.036***  0.028*** 

SIRatio  0.021***  ‐0.041***  ‐0.048***  0.390***  ‐0.158***  ‐0.023*** 

BHSAR  ‐0.090***  ‐0.029***  0.061***  0.064***  0.056***  ‐0.029***   N = 41,544      

39  

 

Table 4. Multivariate Cross‐Sectional Regressions  This table reports the time‐series mean from 15 annual cross‐sectional (Fama‐MacBeth) regressions.  The dependent variable is the firm‐specific one‐year‐ahead buy‐and‐hold size‐adjusted return.  The independent variables are MSCORE (see the notes to Table 1 for a description), Accruals (earnings before extraordinary items less cash flows from operations, all scaled by average assets), Momentum (size‐adjusted returns for the six months prior to the BHSAR window), MVE (market value of equity), SIRatio (short interest ratio, the number of shares sold short as a percentage of the number of shares outstanding) and BTM (book‐to‐market). Observations are assigned to ten portfolios based on prior year cutoff values. Portfolio assignments are then scaled to range from 0 to 1. T‐statistics are based on the time‐series distribution of the parameter estimates. 

 

Intercept MSCORE Accruals MVE BTM Momentum SIRatio Adj. R‐sq

1993  0.074**  ‐0.127***  ‐0.002 ‐0.028 0.033 0.071***  ‐0.034 1.5%

1994  0.127**  ‐0.171***  0.020 ‐0.136*** ‐0.213*** 0.238***  0.063 3.0%

1995  ‐0.002 ‐0.145***  0.039 0.038 0.141*** 0.046*  ‐0.135***  2.7%

1996  ‐0.017 ‐0.015 ‐0.077**  0.011 0.096*** ‐0.006 ‐0.040 0.5%

1997  ‐0.008 ‐0.162***  ‐0.015 ‐0.010 0.017 0.166***  ‐0.009 1.9%

1998  0.256**  0.016 ‐0.319*** ‐0.311*** ‐0.283*** 0.302***  0.373***  3.4%

1999  ‐0.060 ‐0.348***  0.151*** 0.117**  0.277*** 0.081*  ‐0.105**  3.2%

2000  ‐0.138***  ‐0.370***  0.120*** 0.046 0.219*** 0.340***  ‐0.158***  12.7%

2001  ‐0.045 ‐0.158***  0.106*** 0.109*** 0.005 0.116***  ‐0.176***  4.4%

2002  0.099 0.143**  ‐0.256*** ‐0.121**  0.183*** ‐0.053 ‐0.008 2.0%

2003  ‐0.235***  0.107***  ‐0.052 0.086*** 0.246*** 0.179***  ‐0.055**  4.5%

2004  ‐0.099**  0.063 ‐0.012 ‐0.045 0.057*  0.106***  0.069**  0.4%

2005  ‐0.059**  0.014 ‐0.004 0.079*** 0.153*** ‐0.022 ‐0.068***  2.2%

2006  0.013 ‐0.060*  0.081*** 0.013 ‐0.003 0.080***  ‐0.031 0.6%

2007  0.007 ‐0.032 0.070*** ‐0.018 ‐0.034*  0.007 ‐0.022 0.3%

2008  0.078 ‐0.074 ‐0.086 0.010 0.238*** ‐0.264***  ‐0.005 3.8%

2009  0.032 ‐0.269**  0.176*  ‐0.175**  0.148 0.157**  ‐0.011 3.8%

Average  0.001 ‐0.093**  ‐0.003 ‐0.020 0.075**  0.091**  ‐0.021 3.0%

t‐statistic  0.05 ‐2.62***  ‐0.11 ‐0.72 2.00**  2.63***  ‐0.70

         

40  

Table 5. Time‐series asset pricing regressions  To construct this table firms are sorted into deciles each month based on the most recent MSCORE for the firm and prior year MSCORE decile cutoffs. See the notes to Table 1 for a description of the MSCORE model. Value‐weighted returns are calculated each month based on market value of equity at the beginning of the month. Value‐weighted portfolio returns are then regressed on the excess market return (MKT), the size factor (SMB), the book‐to‐market factor (HML), and the momentum factor (WML). The table also reports returns for a hedge portfolio, calculated as the return for the Low MSCORE portfolio minus the return for the High MSCORE portfolio. Each portfolio includes a time‐series of 210 monthly observations from July 1993 through December 2010. ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.  Panel A. Three factor model 

Intercept (%) MKT SMB HML Adj. R‐sq.

Low  0.159 1.130***  0.005 0.239***  71.3%

2  0.239 1.002***  0.094*  0.039 77.4%

3  0.210*  0.854***  0.220***  0.007 83.8%

4  0.162 0.944***  0.010 ‐0.067 80.8%

5  0.081 0.874***  ‐0.116***  ‐0.098**  82.0%

6  0.243*  0.838***  ‐0.071*  ‐0.058 81.4%

7  ‐0.099 0.950***  ‐0.168***  ‐0.055 79.8%

8  ‐0.294 1.021***  ‐0.136**  0.121**  76.5%

9  ‐0.402**  1.254***  ‐0.347***  0.284***  85.1%

High  ‐0.786***  1.334***  ‐0.692***  0.354***  85.6%

Hedge  0.945***  ‐0.204***  0.697***  ‐0.115 29.1%

N = 210 months for each portfolio   Panel B. Four factor model 

Intercept (%) MKT SMB HML WML Adj. R‐sq.

Low  0.189 1.113***  ‐0.007 0.244***  ‐0.039 71.3%

2  0.215 1.015***  0.104*  0.034 0.031 77.4%

3  0.208*  0.856***  0.221***  0.006 0.004 83.7%

4  0.168 0.940***  0.007 ‐0.066 ‐0.008 80.7%

5  0.099 0.864***  ‐0.123***  ‐0.095**  ‐0.022 82.0%

6  0.224*  0.848***  ‐0.063 ‐0.061 0.024 81.4%

7  ‐0.125 0.965***  ‐0.157***  ‐0.060 0.034 79.8%

8  ‐0.285 1.015***  ‐0.140**  0.123**  ‐0.012 76.4%

9  ‐0.329*  1.214***  ‐0.378***  0.298***  ‐0.095*** 85.5%

High  ‐0.747***  1.313***  ‐0.708***  0.361***  ‐0.051 85.6%

Hedge  0.936***  ‐0.199***  0.701***  ‐0.117 0.012 28.8%

N = 210 months for each portfolio   

41  

Table 6. Size‐adjusted returns to decile portfolios conditional on MSCORE  

To construct this table, firms are first sorted into decile portfolios by Market‐value‐of‐equity, Book‐to‐Market, Momentum, and Accrual each year based on prior year cutoff values, then grouped by MSCORE.  Flagged (Not Flagged) denotes firms that fit (do not fit) the profile of an earnings manipulator based on the MSCORE model from Beneish (1999) and a cutoff of ‐1.78. Momentum denotes size‐adjusted returns for the six months prior to the BHSAR return window. Short interest ratio denotes the number of shares sold short as a percentage of the number of shares outstanding, and is measured in the month before portfolio formation. Accruals denotes earnings before extraordinary items less cash flows from operations scaled by average assets. BHSAR denotes annual size‐adjusted returns starting the first day of the fifth month following the end of the fiscal year. For firms that delist, any proceeds upon delisting are reinvested in the size portfolio to which the company belongs. ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.  Panel A. MVE (Market Value of Equity) Portfolios 

Full Sample  Not Flagged  Flagged  Not Flagged 

Portfolio  N BHSAR   N BHSAR   N BHSAR   Less Flagged

1  4009 2.6%*  3236 4.6%*** 773 ‐5.7% 10.3%*** 

2  4127 2.5%*  3251 6.2%*** 876 ‐11.5%*** 17.7%*** 

3  4119 3.8%***  3225 5.4%*** 894 ‐2.0% 7.4%** 

4  4151 0.8% 3243 3.6%*** 908 ‐9.1%*** 12.7%*** 

5  4119 1.3% 3291 3.5%*** 828 ‐7.6%*** 11.1%*** 

6  4180 0.6% 3439 3.0%*** 741 ‐10.4%*** 13.3%*** 

7  4206 0.9% 3500 2.9%*** 706 ‐8.8%*** 11.7%*** 

8  4217 0.0% 3624 1.3%*  593 ‐7.5%** 8.8%*** 

9  4074 0.5% 3557 1.2%*  517 ‐4.2% 5.4%** 

10  4342 ‐0.1% 3976 0.8% 366 ‐10.6%*** 11.5%*** 

Spread  2.7%*  3.8%** 5.0% Panel B. Book‐to‐Market Portfolios 

Full Sample  Not Flagged  Flagged  Not Flagged

Portfolio  N BHSAR   N BHSAR   N BHSAR   Less Flagged

1  2936 ‐5.0%***  2085 ‐3.0%** 851 ‐10.0%*** 7.0%** 

2  4260 ‐2.3%**  3274 0.4% 986 ‐11.3%*** 11.7%*** 

3  4159 0.2% 3311 2.5%** 848 ‐9.0%*** 11.5%*** 

4  4197 0.9% 3415 2.5%** 782 ‐6.1%** 8.6%*** 

5  4305 1.7%*  3586 2.7%*** 719 ‐3.2% 5.8%** 

6  4124 2.8%***  3482 4.4%*** 642 ‐5.9%** 10.3%*** 

7  4178 1.1% 3582 2.1%** 596 ‐5.2%** 7.3%*** 

8  4205 3.9%***  3658 4.9%*** 547 ‐2.7% 7.6%*** 

9  4359 3.1%***  3782 4.9%*** 577 ‐8.6%*** 13.5%*** 

10  4821 4.2%***  4167 6.7%*** 654 ‐11.6%*** 18.3%*** 

Spread  ‐9.2%***  ‐9.7%*** 1.6%   

42  

Panel C. Momentum Portfolios 

Full Sample  Not Flagged  Flagged  Not Flagged

Portfolio  N BHSAR   N BHSAR   N BHSAR   Less Flagged

1  4326 ‐3.2%**  2990 2.7% 1336 ‐16.6%*** 19.3%*** 

2  4146 ‐2.3%**  3307 0.2% 839 ‐12.3%*** 12.4%*** 

3  4202 ‐1.9%**  3536 ‐0.1% 666 ‐11.8%*** 11.7%*** 

4  4086 ‐0.6% 3511 0.4% 575 ‐6.3%*** 6.7%*** 

5  4116 0.7% 3593 1.7%** 523 ‐6.1%** 7.9%*** 

6  4092 2.4%***  3604 3.3%*** 488 ‐4.8%** 8.2%*** 

7  4169 2.2%***  3633 3.6%*** 536 ‐7.1%** 10.7%*** 

8  4148 3.1%***  3533 4.6%*** 615 ‐5.7%*** 10.3%*** 

9  3945 5.2%***  3315 6.6%*** 630 ‐2.4% 9.0%*** 

10  4314 7.4%***  3320 8.7%*** 994 3.1% 5.7%

Spread  ‐10.7%***  ‐6.0%** ‐19.6%***

 Panel D. Short Interest Ratio Portfolios  

Full Sample  Not Flagged  Flagged  Not Flagged 

Portfolio  N BHSAR   N BHSAR   N BHSAR   Less Flagged

1  3751 2.2%**  3189 4.0%*** 562 ‐7.6%*** 11.6%*** 

2  3889 4.5%***  3275 4.8%*** 614 2.5% 2.4%

3  3877 1.8%**  3322 3.3%*** 555 ‐6.9%** 10.2%*** 

4  3920 3.5%***  3333 5.7%*** 587 ‐9.2%*** 15.0%*** 

5  3918 2.9%**  3356 3.8%*** 562 ‐2.7% 6.4%* 

6  4084 1.2% 3450 3.1%*** 634 ‐8.9%*** 12.0%*** 

7  4165 2.5%**  3512 4.6%*** 653 ‐8.9%*** 13.5%*** 

8  4678 ‐0.4% 3878 1.0% 800 ‐7.2%** 8.2%*** 

9  4655 ‐0.9% 3672 1.5% 983 ‐9.8%*** 11.3%*** 

10  4607 ‐3.0%***  3355 0.2% 1252 ‐11.8%*** 12.1%*** 

Spread  5.3%***  3.7%** 4.3%    

43  

Panel E. Accrual Portfolios 

Full Sample  Not Flagged  Flagged  Not Flagged 

Portfolio  N BHSAR   N BHSAR   N BHSAR   Less Flagged

1  4193 3.3%**  3727 6.2%*** 466 ‐19.8%*** 26.1%*** 

2  4121 3.2%***  3735 4.5%*** 386 ‐9.7%** 14.2%*** 

3  4136 2.5%**  3796 3.3%*** 340 ‐6.9%** 10.2%*** 

4  4128 2.3%***  3787 3.2%*** 341 ‐7.0%*  10.1%*** 

5  3999 3.4%***  3627 4.3%*** 372 ‐4.6% 8.8%*** 

6  4018 1.2% 3630 1.7%*  388 ‐3.1% 4.8%* 

7  4008 1.2% 3530 2.4%** 478 ‐7.5%*** 9.8%*** 

8  4224 1.1% 3570 2.0%** 654 ‐3.7% 5.7%** 

9  4334 ‐0.3% 3270 1.8%*  1064 ‐6.7%*** 8.5%*** 

10  4383 ‐4.7%***  1670 0.6% 2713 ‐7.9%*** 8.5%*** 

Spread  8.0%***  5.7%** ‐11.9%***

               

   

44  

Table 7. Size‐adjusted returns to accrual and MSCORE quintile portfolios  To construct Panel A, firms are independently sorted on accruals and MSCORE based on prior year cutoff values. Panels B and C report nested sorts. In Panel B (C) firms are sorted on accruals (MSCORE) first and, within each accrual (MSCORE) portfolio, further sorted into MSCORE (accrual) portfolios. For the first‐pass sorts in Panels B and C, firms are sorted into portfolios based on prior year cutoff values. The second pass sorts in Panels B and C are based on current year cutoff values. See the notes to Table 1 for a description of the MSCORE model. Accruals (or Acc) denotes earnings before extraordinary items less cash flows from operations scaled by average assets. BHSAR denotes annual size‐adjusted returns starting the first day of the fifth month following the end of the fiscal year. For firms that delist, any proceeds upon delisting are reinvested in the size portfolio to which the company belongs. ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.   Panel A. Independent sorts 

Low MSCORE  2  3  4  High MSCORE

Portfolio  N BHSAR   N BHSAR   N BHSAR   NBHSAR   N BHSAR  Spread

Low Acc  5244 6.4%***  1113 3.3%*  537 7.6%*  478 ‐3.5% 942 ‐13.5%*** 19.9%***

2  2248 4.5%***  3128 4.0%***  1286 1.0% 796 1.0% 806 ‐6.2%*** 10.8%***

3  672 0.2% 2599 3.6%***  2520 2.3%**  1365 2.7%*  861 0.0% 0.2%

4  237 ‐0.8% 981 4.2%**  2918 3.0%*** 2723 0.7% 1373 ‐3.8%*  3.1%

High Acc  72 ‐5.6% 199 8.3% 759 2.7% 3060 ‐1.2% 4627 ‐4.7%*** ‐1.0%

Spread  12.1%**  ‐5.0% 4.8% ‐2.4% ‐8.8%***

  Panel B. MSCORE sorted within Accruals portfolios 

Low MSCORE  2  3  4  High MSCORE

Portfolio  N BHSAR   N BHSAR   N BHSAR   NBHSAR   N BHSAR  Spread

Low Acc  1656 5.7% 1667 8.7%***  1664 5.5%*** 1667 1.1% 1660 ‐4.7%**  10.4%***

2  1646 3.7%**  1656 5.8%***  1656 2.3%**  1656 2.9%**  1650 ‐2.8%*  6.5%***

3  1596 3.9%***  1607 1.9%*  1605 2.1%*  1607 2.4% 1602 1.4% 2.4%

4  1641 3.4%**  1650 2.0%*  1647 3.1%**  1650 1.4% 1644 ‐4.3%**  7.7%***

High Acc  1736 1.9% 1749 0.7% 1743 ‐3.3%**  1749 ‐5.5%*** 1740 ‐6.2%*** 8.1%***

Spread  3.8% 8.0%***  8.8%*** 6.7%***  1.5%    

45  

 Panel C. Accruals sorted within MSCORE portfolios  

Low Acc  2  3  4  High Acc 

Portfolio  N BHSAR   N BHSAR   N BHSAR   NBHSAR   N BHSAR  Spread

Low MSCORE  1686 5.6% 1697 8.8%***  1699 4.3%*** 1697 5.0%*** 1693 1.9% 3.6%

2  1597 4.1%**  1607 4.3%***  1608 3.5%*** 1607 3.6%*** 1599 4.0%*** 0.1%

3  1598 3.5%**  1606 2.8%***  1607 2.1% 1606 3.0%**  1601 2.2%*  1.3%

4  1677 0.2% 1687 0.8% 1689 1.2% 1687 2.5% 1679 ‐4.0%*** 4.3%** 

High MSCORE  1715 ‐6.7%***  1725 ‐4.9%***  1724 ‐2.6% 1725 ‐6.2%*** 1719 ‐5.5%*** ‐1.2%

Spread  12.2%***  13.7%***  7.0%*** 11.2%***  7.4%***

 

46  

 Table 8. Comparing MSCORE components for high and low accrual firms This table compares the mean for each of the seven components of the Beneish model (other than accruals) in various subpopulations of our sample.   To construct this table, we independently sorted firms into quintiles according to Accruals and MSCORE (same as Table 7 Panel A).  To focus on the subpopulation of firms where MSCORE exhibited the strongest incremental predictive power over Accruals, we examine firms in the lowest Accrual quintile that are also either in the highest or the lowest MSCORE quintile.  We then compare and contrast the mean for each of the seven variables in the model across high and low MSCORE firms.  The results are reported in the upper three rows of this table.  For comparison, we also group firms in the three highest Accrual quintiles (quintiles 3, 4, and 5) and further separate them according to their MSCORE score.  Once again, we compute the mean for the same seven variables, and report the results in rows four through six.  Finally, in the bottom row, we report descriptive statistics and the result of a significance test for the difference‐in‐difference across these firms.     Each column of the table pertains to one of the seven model variables, where: DSR denotes the ratio of receivables to sales in year t divided by the same ratio in year t‐1. GMI denotes the ratio of gross margin to sales in period t‐1 to the same ratio in period t. SGA denotes the ratio of selling, general, and administrative expense to sales in period t divided by the same ratio in period t‐1. SGI equals sales in t divided by sales in t‐1. DEPI denotes the ratio of depreciation to depreciable base in t‐1 divided by the same ratio in t. AQI equals all non‐current assets other than PPE as a percent of total assets in t divided by the same ratio in t‐1. LEVI equals the ratio of long‐term debt +current liabilities to total assets in t divided by the same ratio in t‐1. ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.   Accrual  Quintile 

MSCORE Quintile  N DSR   GMI   AQI   SGI   SGAI   DEPI   LEVI

Lowest  Lowest  5244 0.867 1.030 0.948 1.084 1.075 0.966 1.167

Highest  942 1.404   1.167   5.272   2.381   0.929   1.071   1.064

High ‐ Low  0.537 0.137 4.324 1.297 ‐0.146 0.105 ‐0.103

3, 4 and 5  Lowest  981 0.759 0.908 0.857 0.980 1.110 1.011 1.216

Highest  6861 1.331   1.071   2.330   1.570   0.949   1.108   0.991

High ‐ Low  0.572 0.163 1.473 0.590 ‐0.161 0.097 ‐0.225

Low Accruals ‐ High Accruals ‐0.035 ‐0.025 2.851***  0.707***  0.015 0.008 0.122*** 

  

47  

Table 9. Regression of future earnings on current period earnings components   This table reports the results from a series of pooled (cross‐sectional and time‐series) regressions.  The dependent variable is always year t+1 earnings (defined as income before extraordinary items excluding depreciation, divided by average total assets in year t).  The independent variables are various components of current period (year t) earnings, a Scaled Probability of Manipulation (SPM) measure, and interaction terms.  EARN denotes income before extraordinary items excluding depreciation; CFO denotes cash from operations; ACC denotes income before extraordinary items excluding depreciation less CFO;  ACCPOS denotes ACC if positive, 0 otherwise; ACCNEG denotes ACC if negative, 0 otherwise.  All these variables are divided by average total assets in year t.  SPM denotes MSCORE ranked into deciles and scaled to range from 0 (lowest MSCORE) to +1 (highest MSCORE). ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.  

Model 1 Model 2  Model 3 Model 4

Intercept  0.025***  0.013***  ‐0.003*  ‐0.017***

EARN  0.796***          

CFO     0.938***  0.974*** 0.975***

ACC     0.493***       

ACCPOS        0.769*** 0.996***

ACCNEG        0.340*** 0.213***

ACCPOS*SPM           ‐0.353** 

ACCNEG*SPM           0.555***

SPM           0.032***

           

Adj R‐sq  43.84% 48.16% 48.85% 49.35%

N  38,008    

48  

Table 10. Earnings announcement returns by quarter  

This table reports the cumulative abnormal returns (CAR) around the next four earnings announcements for various MSCORE‐based portfolios.  Firms are added to portfolios on the first day of the fifth month following the end of the fiscal year.  CAR denotes the cumulative raw return over days ‐1, 0, and +1 relative to the earnings announcement date, minus the cumulative return to the benchmark size portfolio to which the firm belongs.  To construct Panel A we sort firms into Flagged and Not Flagged categories based on MSCORE and a cutoff of ‐1.78.  To construct the next two panels, we first sort firms into deciles based on their current MSCORE score and decile cutoffs from the prior year MSCORE distribution.  We then calculate CAR for each decile portfolio (Panel B), as well as for alternative hedge portfolios (Panel C).  In Panel C, we start with the most extreme portfolios (deciles 1 and 10) and progressively add less extreme MSCORE firms to the long and short positions. We report short‐window announcement period CAR for each of the next four quarters, as well as the total CAR for all four quarterly announcements.  ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.  Panel A. Earnings announcement returns by quarter for flagged and not flagged firms  

Quarter t+1  Quarter t+2  Quarter t+3  Quarter t+4  Total CAR

N CAR N CAR N CAR N CAR For 4 Qtrs

Not Flagged  33658 0.35%***  32939 0.11%**  32017 0.16%***  28411 0.51%***  1.03%*** 

Flagged  7263 ‐0.44%***  7099 ‐0.38%***  6848 ‐0.70%***  5844 ‐0.01% ‐1.48%*** 

Difference  0.79%***  0.49%***  0.85%***  0.52%***  2.51%*** 

  

Panel B. Earnings announcement returns by quarter and MSCORE decile 

Quarter t+1  Quarter t+2  Quarter t+3  Quarter t+4  Total CAR

Decile  N CAR N CAR N CAR N CAR For 4 Qtrs

1  4297 0.23% 4200 0.31%* 4062 0.01% 3540 0.58%*** 1.01%***

2  4025 0.25%*  3937 0.21% 3839 0.07% 3426 0.60%*** 1.03%***

3  3896 0.47%***  3807 0.21% 3707 0.24% 3348 0.55%*** 1.38%***

4  4006 0.29%**  3914 0.23%* 3824 0.41%*** 3431 0.62%*** 1.44%***

5  3791 0.46%***  3703 0.17% 3597 0.45%*** 3202 0.47%*** 1.45%***

6  4104 0.34%***  4026 ‐0.08% 3920 0.27%* 3506 0.71%*** 1.13%***

7  4091 0.59%***  3998 0.12% 3882 0.16% 3444 0.42%*** 1.21%***

8  4214 0.16%  4142 ‐0.19% 4025 ‐0.19% 3516 0.30%* 0.05%

9  4202 ‐0.13%  4104 ‐0.19% 3956 ‐0.49%*** 3421 0.06% ‐0.72%**

10  4295 ‐0.51%***  4207 ‐0.55%*** 4053 ‐0.77%*** 3421 ‐0.08% ‐1.84%***

49  

 Panel C. Earnings announcements returns to alternative MSCORE‐based hedge portfolios  

Total CAR

Long  Short  Qtr t+1 CAR  Qtr t+2 CAR  Qtr t+3 CAR  Qtr t+4 CAR  For 4 Qtrs 

Decile 1  Decile 10  0.74%***  0.86%***  0.78%***  0.66%**  2.85%*** 

1 to 2  9 to 10  0.56%***  0.63%***  0.67%***  0.59%***  2.31%*** 

1 to 3  8 to 10  0.29%**  0.45%***  0.43%***  0.40%***  1.98%*** 

1 to 4  7 to 10  0.22%**  0.42%***  0.39%***  0.30%***  1.55%*** 

1 to 5  6 to 10  0.25%***  0.41%***  0.44%***  0.28%***  1.31%*** 

 

50  

Figure 1A: Market value of equity portfolios 

  Figure 1B: Book‐to‐market portfolios 

 

51  

Figure 1C: Momentum portfolios 

  Figure 1D: Short Interest Portfolios 

 

52  

 Figure 1E: Accrual portfolios