Business Model Shocks and Abnormal Accrual Models Model Shocks and Abnormal Accrual Models1 Edward Owens Joanna Shuang Wu Jerold Zimmerman Simon School of Business, University of Rochester,

Business Model Shocks and Abnormal Accrual Models1

Edward Owens

Joanna Shuang Wu Jerold Zimmerman

Simon School of Business, University of Rochester, Rochester, NY 14627

June 2013 Abstract Insights from economics challenge the development of better specified accrual models. Since rent seeking firms pursue differentiated business strategies, firms in the same industry have heterogeneous accrual generating processes. Moreover, technological innovation, regulatory changes, and entry of new firms (i.e., business model shocks) frequently force existing firms to revise their extant business models. We present evidence that such business model shocks are widespread, propagate through several years of financial statements, reduce accrual models’ goodness of fit, and result in unrealistically large unsigned “abnormal” accruals. There is a spillover effect among firms in the same industry in that one firm’s abnormal accrual is affected by business model shocks experienced by the other firms in the industry. We show that business model shocks not only add noise to abnormal accruals, but can also introduce biases into both unsigned and signed discretionary accruals. The effect of business model shocks on estimating abnormal accruals cannot be eliminated by excluding a few major corporate events, such as mergers and acquisitions, as is common in the literature. We conclude that further refinements of accrual models likely will prove difficult due to the random, diverse, and persistent affects business model shocks have on firms’ accrual generating processes.

1 We gratefully acknowledge the financial support provided by the Simon School at the University of Rochester and the comments from Dan Amiram, Dan Collins, Mark Evans, Shane Heitzman, Jerry Warner, Charles Wasley and seminar participants at the University of Rochester and the Penn State Accounting Research Conference.

1

1. Introduction

Since Healy (1985), the accounting literature has pursued an accruals-based measure of

accounting discretion. Jones (1991) introduced a methodology to extract “discretionary accruals”

from total accruals by modeling “non-discretionary accruals” as a function of firm

characteristics, with the residuals then interpreted as "discretionary" or "abnormal" accruals.

Despite significant effort seeking better specified accrual models, many concerns regarding

accrual model misspecification, the implausibly large magnitude of discretionary accruals,

power, and bias remain. A couple of representative quotations illustrate the lack of progress:1

“The (positive correlation between measures of abnormal accruals and total accruals) raises concerns about whether abnormal accruals reflect accounting distortions or whether they instead are the result of poorly specified accruals models and include a component that measures fundamental performance.” (Dechow et al. 2010 p. 358)

“A powerful cocktail of authors’ strong priors, strong ethical and moral views, limited knowledge of the determinants of accruals in the absence of manipulation, and willingness to ignore correlated omitted variables in order to report a result, seems to have fostered a research culture that tolerates grossly inadequate research designs and publishes blatantly false positives.” (Ball 2013).

As a profession, we have very limited theory of the accrual generating process and how

fundamental firm performance maps into discretionary accruals (McNichols 2000; Dechow et al.

2010). In the face of these problems, several authors encourage research to develop improved

discretionary accrual models (e.g., Holthausen et al. 1995; Guay et al. 1996; Dechow et al.

2010), and there is a general belief that such improvements can indeed be discovered. In this

study, based on research in economics we offer more texture as to the theoretical challenges

accounting researchers face in trying to develop better specified discretionary accruals models.2

We argue that certain key insights from economics challenge the fundamental assumptions

underlying “normal” (non-discretionary) accruals, and hence innately thwart the ability of

researchers to parse discretionary accruals from total accruals.

Estimation of either cross-sectional or firm-specific accrual models relies on two

assumptions: firm stationarity (i.e., firms should have reasonably stable accrual generating

1 Also see Bernard and Skinner (1995), Dechow et al. (1994), Subramanyam (1996), Guay, et al. (1996), Ball and Shivakumar (2008), and Dichev et al. (2012). 2 We rely on research from industrial organization, finance, and strategy. For parsimony, we refer to these literatures broadly as “economics,” and review these literatures in section 2.1.

2

processes) and intra-industry homogeneity (i.e., industry peers should have similar accrual

generating processes). 3 However, a key economic principle dictates that a firm should seek rents

by developing a unique, sustainable market niche that differentiates its business strategy from

other firms (Brickley and Zimmerman 2010). Profit maximization via innovation and strategy

differentiation presents an inherent challenge to the intra-industry homogeneity assumption.

Further, technological innovations, regulatory changes, and entry by new (or existing) firms

(what we call “business model shocks”) cause firms to frequently alter their existing business

strategies. These frequent business model revisions/shocks present an inherent challenge to the

firm stationarity assumption, and likewise decrease intra-industry homogeneity if the shock

causes further strategy differentiation.4

The central issue we examine in this study is the extent to which business model shocks

affect accrual models and the residuals from these models (discretionary accruals). To illustrate

the complexities involved, consider that firms respond differently to a given random shock to

their business strategies and it can take several years for firms to adjust to the shock. A given

shock will cause some firms’ total accruals to increase and other firms’ total accruals to decrease

as managers change their business strategies, cash operating cycles, and their accrual generating

processes. Since firms in the same industry react differently to common business model shocks,

including an instrument for a shock that occurs in year t in estimating a cross-sectional accrual

model in year t only captures the average effect of the shock for all firms in the industry-year t.

Moreover, these shocks likely impact several years of operations as managers react to the shock

by implementing new strategies (e.g., Gerakos and Kovrijnykh 2012). Including an instrument

for the business model shock fails to capture how the shock in year t affects accruals in years t+1

and t+2 for all firms in the industry. If firm-specific accrual models are estimated, a business

model shock in year t can affect accruals in years t, t+1, and t+2 in different and unpredictable

ways as this firm and other firms in the same market dynamically adjust their business strategies.

3 Dopuch et al. (2011) conclude that substantial heterogeneity exists in accrual generating processes in some industries and that this heterogeneity generates “a large noise component in abnormal accruals.” 4 Cross-sectional accrual models rely not only on intra-industry homogeneity but also the firm stationarity assumption. For example, if large business model shocks last year caused large accruals last year and these accruals reverse, then the cross-sectional model this year will be estimated with less precision. Likewise, firm-specific time-series accrual models rely on intra-industry homogeneity in addition to firm stationarity. Large business model shocks reduce intra-industry homogeneity, which affects the nature of competition in the industry. Firm-specific accrual models will be estimated less precisely because the accrual generating process is affected by the changing industry competition.

3

For example, suppose a new technological innovation causes one firm (either with a patent or

complementary technology) to introduce a new product with a different accrual generating

process than its existing products (i.e., Amazon enters the e-reader market with its Kindle).

Unless the researcher can specify ex ante the evolution of this firm’s accrual generating process,

the time-series regression for this firm’s normal accruals fails to correctly capture how the

accrual generating process changes dynamically. Absent such knowledge of the evolution, some

“non-discretionary” accruals will be misclassified as “discretionary” accruals.

Because the same business model shock can increase some firms’ value and decrease

others’ and firms react to the same shock with different strategy responses, we focus on unsigned

abnormal accruals in this study, although in supplemental tests (Section 4.6.3) we also

investigate the implications for signed abnormal accruals. Unsigned abnormal accruals have been

used as instruments for earnings management (when the research study does not have a signed

prediction about earnings management) and as an instrument for “earnings quality.” Firms

generally do not disclose intentional changes to their business models as shocks occur (likely due

to competitive reasons) and rarely disclose immediately shocks that adversely impact their

business strategies. The effect of these adverse shocks on firms’ strategies usually gets reflected

in the financial statements with a lag. Hence, we use a measure of large, firm-specific abnormal

returns as our primary indicator for the presence of business model shocks.

We begin by documenting that 55% of all firm-years from 1988 to 2010 have at least one

monthly unsigned abnormal stock return of 20% or more. Large abnormal returns reflect both

realizations of past activities and changes in expectations of future activities (including changes

in the firm’s business model). Moreover, large abnormal returns provide signals to managers

about the efficacy of their and their competitors’ business strategies. Hence, large abnormal

returns capture both changes in growth rates of the firm’s current business model as well as

expected revisions to the firm’s business model. To help validate this return-based shock proxy,

we next identify firm-years with large acquisitions, four-digit SIC code changes, and large

restructuring charges or special items. Not surprisingly, the incidence of these large operational

shocks is statistically and economically associated with large contemporaneous and lagged

unsigned abnormal returns. Large unsigned abnormal returns appear to be capturing large

business model shocks, not just changes in growth rates of firms with stable business models.

Moreover, our finding that one- and two-year lagged unsigned abnormal returns are correlated

4

with contemporaneous operational shocks suggests that business model shocks can affect several

years of financial statements, including accruals.

To address the primary focus of this paper, i.e., the impact of business model shocks on

estimation of abnormal accruals, we compute traditional abnormal accruals using the Jones

(1991) model via annual industry cross-sectional estimation, along with modifications to the

original Jones model suggested by Ball and Shivakumar (2006) (i.e., a nonlinear accrual model)

and Kothari et al. (2005) (i.e., with performance control). Consistent with prior literature, we find

that for the median firm in our sample, the Jones model produces unsigned abnormal accruals

that are 67% of unsigned net income, which implies roughly one part management discretion and

12 parts “noise” assuming auditors apply a materiality threshold of a 5% net income (62%/5%).

We document that the existence of business model shocks materially affects the

estimation of abnormal accruals for all three accrual models. Specifically, contemporaneous and

lagged business model shocks are positively correlated with large unsigned abnormal accrual

estimates from cross-sectional accrual model estimations. These results suggest that unsigned

abnormal accruals capture shocks to firms’ business models, and therefore likely overstate the

existence of unsigned discretionary earnings management. Further, we document a spillover

effect on a firm’s unsigned abnormal accruals from business model shocks experienced by the

other firms in the same industry-year due to the effects of these shocks on the overall fit of the

accrual model. We also show that the variability in firms’ cash operating cycles is positively

correlated with large operational shocks and large unsigned abnormal returns. This finding

establishes a direct causal link between our proxy for business model shocks and firms’ accruals

generating processes.

Next, we examine how the inclusion of firm-year observations identified as having large

business model shocks (proxied by the existence of a large unsigned monthly abnormal stock

return during the year or the previous two years) impacts the magnitude of unsigned abnormal

accruals. As more “contaminated” observations (i.e., those with large business model shocks) are

included in the industry-year Jones model estimation, the mean unsigned abnormal accrual

increases from 3.3% of total assets when all contaminated observations are removed to 8.1% of

total assets when all contaminated observations are included. Moreover, including only a

relatively small number of contaminated observations significantly affects the outputs of the

5

abnormal accrual model. In other words, the magnitude of unsigned abnormal accruals remains

large unless a substantial number of contaminated observations are removed.

To directly investigate the effect of business model shocks on empirical inferences we

conduct simulation tests. Our results suggest that when studying the relationship between

“earnings quality” (proxied by unsigned abnormal accruals) and a partitioning variable, the null

of no relation between the two variables is over rejected even when modest correlation exists

between the partitioning variable and business model shocks. The results hold after controlling

for various firm characteristics including the operating volatility measures in Hribar and Nichols

(2007). Furthermore, the above results hold for both firms’ own business model shocks and the

shocks experienced by the other firms in the same industry-year, again confirming a spillover

effect. We further utilize the setting examined by Bharath et al. (2008) to demonstrate that

relations between abnormal accruals and dependent variables of interest (e.g., cost of debt) may

capture effects of operational risk, rather than effects of discretionary accounting choices.

Specifically, we show that a statistically significant association between unsigned abnormal

accruals and cost of private debt only obtains for observations where business model shocks

exist. Finally, we examine how business model shocks affect inferences regarding earnings

management based on signed abnormal accruals. There, our simulation tests show that business

model shocks can exacerbate the problem of over-rejecting the null of no earnings management

in various subsamples based on firm characteristics such as market-to-book, firm size and sales

growth.5

Our paper makes several contributions. First, extant literature recognizes that normal

accruals and earnings quality depend on both the firm’s fundamental performance and on the

accounting system that measures that performance. Dechow et al. (2010 p. 345) conclude that

“we have relatively little evidence about how fundamental performance affects earnings quality.”

Our analysis provides an economics-based framework and associated evidence on how one

driver of fundamental performance, business model shocks, affects abnormal accruals and hence

those earnings quality proxies that rely on reported earnings. Second, researchers point to several

methods for improving discretionary accrual models, such as using the balance sheet approach to

5 In our primary empirical specifications we estimate accrual models using total accruals, but our inferences remain intact when using working capital accruals. Further, although we do not impose any explicit size filters in our sample, our results continue hold when we exclude smaller firms (for example, the bottom 10% of firms with market capitalization of less than $10 million, or the bottom 50% of firms with market cap of less than $250 million, alternatively).

6

compute accruals (Collins and Hribar, 2002) and reducing various other forms of model

misspecification (e.g., Kothari et al., 2005, Ball and Shivakumar, 2006). Our analysis suggests

that such attempts are likely to prove ineffective, as they do not purge abnormal accruals of non-

discretionary components that arise from pervasive business model shocks. The reason for our

pessimism is that these shocks occur randomly, their effects on accruals can persist over several

years, observable proxies for these shocks are very noisy, and firms in the same industry respond

to shocks in different ways and with different implications for abnormal accruals. Better

understanding of the limitations of accrual models will allow researchers to construct more

compelling tests involving earnings management and earnings quality.

The remainder of the paper proceeds as follows: in Section 2 we discuss related literature

and our empirical conjectures. Section 3 describes our sample selection. Section 4 presents our

research design and main empirical findings. Section 5 concludes.

2. Related literature and empirical conjectures

This section first summarizes the economics, strategy and finance literatures that question

the firm stationarity and intra-industry homogeneity assumptions underlying the estimation of

accrual models. Then we summarize related accounting studies that have recognized and tried to

address the misspecification problems in accrual models. Finally, we derive the empirical

conjectures that follow from these literature reviews

2.1. Firm stationarity and intra-industry homogeneity assumptions - economic insights

An early, influential concept in economics was Schumpeter’s (1942) "creative

destruction," whereby entry by entrepreneurs capitalizing on new technology or regulatory

changes disrupts existing markets and destroys the value of established companies (e.g., digital

imaging destroying traditional photography). Schumpeter’s “creative destruction” forces both

incumbent firms and new entrants to modify their business strategies in response to entry and to

adjustments in business models by competitors. If the rate of technological and/or regulatory

change is high, creative destruction casts doubt on the firm stationarity assumption.

Successful firms search for unique market niches that differentiate and shield themselves

from competition by other firms in the industry. Porter (1979) describes the various competitive

forces affecting business strategies (i.e., threat of entry, bargaining power, and current

contestants revising their strategies). So, changes in any of these competitive forces, not just new

technologies, generate shocks to firms’ extant business models. The shocks include (i) managers

7

intentionally innovating in response to new technologies or changes they observe in their

customers’, suppliers’, or competitors’ strategies, and (ii) unintentional shocks to their business

as competitors enter or alter their strategies (e.g., Schumpeter’s “creative destruction”). Porter

and other business strategists emphasize establishing differentiated products, solidifying

customer relationships, vertical integration, and technological leadership. Prahalad and Hamel

(1990) argue that firms differentiate themselves by developing and exploiting their unique core

competencies (coordinating diverse production skills and integrating multiple technologies that

are difficult to imitate). The strategic management literature predicts that firms competing within

the same product markets will exhibit heterogeneous firm characteristics, thereby challenging the

intra-industry homogeneity assumption (Brickley and Zimmerman 2010).

The following empirical regularities further challenge the validity of the firm-stationarity

and intra-industry-homogeneity assumptions:

considerable heterogeneity exists across industries in firm survival, entry and exit

rates, and market structures (Berry and Reiss 2007);

firm size distributions within industries are highly skewed (Schmalensee 1989);

as new markets develop, the number of firms tends to rise and later falls and these

trends vary widely across product markets (Sutton 2007);

firm-specific growth rates have become more volatile (Comin and Mulani, 2006);

various operating characteristics of firms such as idiosyncratic risk and survival rates

of newly listed firms have changed over time (Fama and French, 2004, Irvine and

Pontiff, 2006, and Brown and Kapadia, 2007).

Another general view from both economics (Sutton 2007) and strategic management that

questions both the firm stationarity and intra-industry homogeneity assumptions involves the

dynamic nature of business models. Business models are “likely over time to be replaced by an

improved model that takes advantage of further technological or organizational innovations. The

right business model is rarely apparent early on in emerging industries. Of course, once a

business model is successfully established, changing technology and enhanced competition will

require more than defenses against imitation. It is also likely that even successful business

models will at some point need to be revamped, and possibly even abandoned” (Teece, 2010).

Successful implementation of a new business model often requires changes in real

investment and operating policies and an organizational architecture (decision rights partition,

8

performance evaluation, and compensation schemes) that provides managers with incentives to

execute the strategy (Brickley et al. 2009). Likewise, firms experiencing shocks to their business

model from competitors also alter their real investment, operating, and organizational structures

to survive. Milgrom and Roberts (1995) argue that complementarities exist between strategy and

organizational design variables. Because business model shocks usually involve changing the

firm’s core competencies, location of the firm in the value chain, its asset base, capital structure,

and organizational architecture, these shocks usually impact the firm’s operating cycle and hence

its accrual generating process.6 In fact, Dichev et al. (2012) report that CFOs list their firm’s

business model as being the most important factor affecting their company’s earnings quality.

As further illustration of the questionable validity of the firm stationarity and intra-

industry homogeneity assumptions, in Figure 1 we plot the annual operating cycles of seven

major airlines (a relatively homogenous industry) over the period 1990-2011. As revealed in the

figure, there is significant variation in operating cycle not only over time within a single airline,

but also in the industry cross-section. Specifically, the average pair-wise correlation between any

two airlines is only 0.18. Further, Fama and French (2004) report that the ten-year survival rate

for seasoned firms falls from 61% for the 1973 cohort to 47% for the 1991 cohort, and the

likelihood that a newly listed firm survives its first ten years falls from 61% for the 1973 cohort

to 37% for the 1991 cohort. Hence, firms do not appear very stationary, and firm stationarity

appears to have declined.

2.2. Related accounting studies of abnormal accruals

Various accrual regression models exist to parse “discretionary accruals” from total

accruals. The regression residuals usually represent “discretionary accruals” and are used as

proxies for earnings management (signed residuals) or earnings quality (unsigned residuals).

Jones (1991) pioneered the accruals regression framework by modeling “normal” accruals as a

linear function of current period change in sales and PPE.7 Dechow and Dichev (2002, hereafter

DD) represents another significant model innovation, where working capital accruals are

regressed on past, current, and future periods’ operating cash flows and the standard deviation of 6 As an example of how business model revisions affect the accrual generating process consider Amazon.com. It started with few accounts receivables because customers paid with third-party credit cards, which amount to cash sales to Amazon. As it introduced its own credit card and began selling services to other merchants, Amazon’s receivables increased, affecting its operating cycle. In addition, as Amazon entered into the hardware business (Kindle devices), inventory turnover slowed, impacting its operating cycle. 7 Dechow et al. (1995) develop the modified Jones model where credit sales growth are backed out of sales changes due to concerns about potential manipulations of credit sales.

9

the regression residuals is interpreted as a measure of “earnings quality.” 8 While accrual models

play an important role in earnings management/earnings quality studies, many researchers voice

concerns about accrual model misspecifications. Some studies point out that the magnitudes of

the accrual model residuals are implausibly large and hence difficult to attribute solely, or even

mostly, to management discretion (e.g., Ball and Shivakumar, 2008, Dopuch et al., 2011). The

large residuals suggest poor fit of the models and are consistent with significant violations of the

firm stationarity and intra-industry homogeneity assumptions due to business model shocks.

Further, accrual model misspecifications not only introduce noise but can also lead to

biased inferences. Dechow et al. (1995) find excessive rejection rates in favor of the existence of

earnings management in firms with extreme financial performance. McNichols (2002)

conjectures that both the Jones and DD models are likely misspecified due to inadequate

considerations of firm economic fundamentals such as uncertain economic environments and

sales growth. In a recent survey, Dechow et al. (2010) similarly emphasize the importance of

firm fundamentals by stating that “(t)he literature often inadequately distinguishes the impact of

fundamental performance on (earnings quality) from the impact of the measurement system.”

Efforts have been made to better incorporate the effects of firm fundamentals, especially

firm performance, into the estimation of “normal” accruals. Kothari et al. (2005) develop a

performance matching procedure where discretionary accruals (estimated by industry-year) of

treatment firms are measured relative to those of control firms in the same industry and with

similar current or lagged ROA. However, as recognized by the authors, the success of their

matched-firm approach depends on “the homogeneity in the relation between performance and

accruals for the matched and the sample firm.” In other words, Kothari et al. (2005) employ the

standard assumption of intra-industry homogeneity to estimate normal accruals and to choose

their matched control firms. Collins et al. (2012) further match firms based on sales growth in

addition to ROA. Zhang and Zhuang (2012) add current period signed stock returns to the

standard accrual models to capture anticipated future firm performance. Again, these recent

papers continue to rely on the standard accrual model assumptions of firm stationarity and intra-

industry homogeneity, which are likely violated when business model shocks are present.

8 DD estimate firm-specific accrual models and require at least eight annual observations per firm. Besides imposing a survivorship bias on their results, their approach assumes business model stationarity, and data restrictions reduce their final sample by roughly 75%.

10

Ball and Shivakumar (2006) argue that the standard linear accrual models are

misspecified because accruals likely reflect economic losses faster than economic gains due to

accounting conservatism. They introduce a nonlinear accrual model to allow asymmetric

associations of accruals with gains relative to losses. Their model explicitly accounts for one

important aspect of firm performance, i.e., economic gains versus losses, in accrual estimations.

However, within the subsample experiencing either gains or losses their model still relies on the

assumption of intra-industry homogeneity, which is likely violated due the presence of business

shocks within that subsample. While the prior literature’s attempts to incorporate fundamental

performance in the accrual models have resulted in improved model specifications and

regression goodness-of-fit, these models continue to suffer from the potential effects of business

model shocks. Adding performance controls and/or performance interaction terms to the accrual

models are unlikely to be effective because business shocks are unpredictable and have different

effects on firms’ accrual generating processes. The same shock likely elicits different strategy

reactions (and hence different accounting accrual effects) from firms in the same industry, and

this shock will impact several years of firms’ financial statements.

The poor fit of the accrual models due to business model shocks has implications for

research on “earnings quality,” which relies on unsigned accrual model residuals to measure

accounting quality. Hribar and Nichols (2007) point out that the mean of the unsigned accrual

model residual is a positive function of the standard deviation of the signed residual and is

associated with firm operating volatility. They show that when researchers associate “earnings

quality” based on the unsigned residual with a partitioning variable, biased inferences result

when the partitioning variable is correlated with firm operating volatility. Similar problems of

biased inferences also arise for signed abnormal accruals. Prior studies such Kothari et al. (2005)

and Collins et al. (2012) show that the null of no earnings management tends to be over-rejected

in samples with extreme firm characteristics, such as market-to-book, size, and sales growth.

However, none of these studies considers how the prevalence of business model shocks affects

these biases, which we examine in our subsequent analyses.

Finally, our paper is related to Collins and Hribar (2002), who provide evidence that

accrual models are better specified using accrual data from the cash flow statement than from the

balance sheet because of the existence of certain "non-articulation" events (e.g., mergers and

acquisitions and discontinued operations). Many subsequent studies using cash flow statement

11

data do not exclude major events such as M&A because the common perception is that these

events are problematic only for balance sheet data.9 A key message from our study is that major

events remain problematic even after using cash flow statement data to compute accrual models

because strategy shocks violate the firm stationarity and intra-industry homogeneity assumptions

implicit in these regressions. Further, those studies that remove certain events do not go far

enough, because as we show later, business model shocks are much more pervasive than those

few events that are often excluded from studies.10

2.3. Empirical predictions

We start with the following generic cross-sectional accrual model:

01

N

ijt jt njt nijt ijtn

TA X

(1)

where TAijt is total accruals for firm i in industry j in year t, Xnijt is the nth firm-specific variable

thought to explain the accrual generating process, βnjt is the estimated coefficient on the nth

variable in industry j in year t, and ijt is the residual. In the literature the signed residual (the

prediction error), ijt , is often used as a proxy for earnings management and the unsigned

residual, ijt , a measure of “earnings quality” or sometimes a measure of earnings management

if the researcher predicts earnings management but not its direction. However, due to model

misspecifications, ijt likely contains both non-discretionary and discretionary accruals:

ijt ijt ijt ijtDA NDA (2)

where DAijt is discretionary accruals for firm i in industry j in year t, NDAijt is non-discretionary

accruals for firm i in industry j in year t, and ijt is white noise arising from accounting errors. In

this study we focus on the implications of business models shocks on the properties of the

unsigned residual ijt .

9 For example, papers that use the cash flow statement data and do not make adjustments for major events include Dechow and Dichev (2002), Wysocki (2008), Ecker et al. (2011), and Dechow et al. (2012). 10 McNichols (2002) uses the cash flow statement data and excludes M&A and discontinued operations because, as she points out, accruals in one period and cash flows in another period may be for different economic entities. Ball and Shivakumar (2006) use the cash flow statement data and exclude acquisitions.

12

We expect business models shocks to negatively affect the fit of the accrual models for

two reasons. First, business model shocks are associated with changes in firm strategies,

operations, and accrual generation processes. Liu and Zhang (2008) report that firms with large

lagged returns (either positive or negative), our proxy for business model shocks, are more

sensitive to industrial production growth in the following period. This suggests that these firms

are more likely to make investments and divestitures and other major business decisions.11 These

real changes cause firms to restructure their long-term assets and working capital, which can then

alter their operating cash flow cycles and generate large non-discretionary abnormal accruals

(NDA ≠ 0 in Eq. 2) as GAAP requires write-offs, write-downs and the recognition/de-recognition

of deferred tax assets and so forth. Furthermore, extreme stock returns in year t signal additional

business model uncertainty that manifests itself in real investment and operating changes in the

following years. The Liu and Zhang (2008) findings provide evidence consistent with the

dynamic nature of business model shocks.12

Second, because firms in the same industry have unique core competencies, each will

react to the same shock with a different strategy response. The same shock can cause some

firms’ total accruals to increase and other firms’ total accruals to decrease, which makes it

difficult to control for the effect of the shocks in the accrual models. For example, a new

scientific advance will provide new opportunities for those firms with complementary

technologies to the new advance and reduce the value of firms with similar products without

complementary technologies. Large coal mines were made better off by tighter mine-safety

regulations because the increased regulation shut down small, marginal mines (Henderson 1977).

The invention of the automobile opened vast new markets for Standard Oil because it had the

gasoline cracking technology necessary to produce large quantities of gasoline (Chernow 2007).

Ball (2013) argues that two firms experiencing a similar negative demand shock could respond

very differently with one firm seeing an increase in inventory and the other a decrease. This

suggests that business model shocks can affect firms’ accrual generating processes in

11 Liu and Zhang (2008) suggest that the growth in industrial production is a priced risk factor and firms with extreme stock performances are more sensitive to this risk. 12 Strategy changes can take several years to implement. For example, some firms adapt to a shock by adding new competencies via acquisitions. If the strategy fails, the acquirer may sell or close the business. Mitchell and Lehn (1990) document that firms are more likely to become subsequent takeover targets if they made prior value-destroying acquisitions; and later, these firms divested or restructured to thwart their own takeover.

13

unpredictable ways that are difficult to model and control for ex ante, reducing the fit of the

accrual regression. The preceding discussion leads to the following predictions:

Prediction 1: The unsigned residual ijt in Eq. (1) for firm i in year t is positively

associated with business model shocks experienced by firm i in years t, t-1, and t-2.

Because business model shocks experienced by a particular firm affect the overall fit of

the accrual model in the same and subsequent industry-years, we expect to observe a spillover

effect of the shocks on other firms’ accrual residuals in the same and subsequent industry-years.

Prediction 2: The unsigned residual ijt in Eq. (1) for firm i in industry j and year t is

positively associated with business model shocks experienced by firm i’(i’ i) in industry j and years t, t-1, and t-2.

We expect the goodness of fit of the accrual model to be more negatively affected when

more firms in a given industry-year experience business model shocks.

Prediction 3: The goodness of fit (R2s) of the regressions estimated in Eq. (1) in industry j year t is negatively associated with the percentage of firms experiencing business model shocks in industry j and years t, t-1, and t-2.

Business model shocks can inflate the magnitude of Jones-type abnormal accruals by

increasing the volatility in firms’ operating cycles. Dechow et al. (1998) develop an algebraic

model of working capital accruals, where the accrual is approximated by the product of a firm’s

operating cycle and the sales shock. This implies that a firm’s operating cycle affects the strength

of the relation between accruals and sales changes (assuming sales follow a random walk).

Cross-sectional accrual models impose the same slope coefficient (i.e., the same operating cycle)

on all firm-year observations in that regression. However, if operating cycles are affected by

business model shocks and vary by firm and over time, such variation can be one mechanism

that leads to the large absolute abnormal accruals that we predict earlier.

Prediction 4: The change in firm i’s operating cycle volatility between years t-1 and t is positively associated with business model shocks to firm i in year t.

Empirical evidence on the above predictions benchmarks the pervasiveness of the

potential accrual model misspecifications that arise from violating the firm stationarity and intra-

industry homogeneity assumptions. We next investigate whether business model shocks can bias

14

researchers’ inferences regarding the determinants and effects of “earnings quality” when ijt is

used to proxy for ijtDA as a measure of earnings quality.

From Hribar and Nichols (2007) (hereafter HN) we know (i) the mean and variance of

ijt increase in the variance of ijt in Eq. (1), and (ii) operating volatility is associated with the

variance of ijt . To the extent business model shocks in year t cause firms to change real

investment and operating decisions in years t, t+1, and t+2, and these real changes result in large

unsigned non-discretionary accruals ( ijtNDA ), then business model shocks create bias in

unsigned abnormal accruals ( ijt ) as an instrument for unsigned discretionary accruals ( ijtDA ).

HN document that ijt in year t are correlated with the variances of cash flow from operations

and revenue, where the later variances are computed over years t-5 to t-1. While the HN

variances will capture some business model shocks, implicit in their methodology is the

assumption of firm stationarity (past cash flow and revenue volatility predict future abnormal

accrual volatility). While the likelihood of business model shocks in period t is correlated with

cash flow and revenue volatilities from t-5 to t-1, business model shocks in year t will cause

large unsigned non-discretionary accruals ( ijtNDA ) in years t, t+1, and t+2. So, controlling for

past cash flow and revenue volatility in computing unsigned abnormal accruals in year t does not

undo all the bias caused by business model shocks that occur in years t-2, t-1, and t that have not

yet manifested in higher cash flow or revenue volatility computed over t-5 to t-1.13

Business model shocks likely increase ijtNDA and ijt due to the inability of the accrual

models to capture these shocks. When researchers investigate the effect of a particular variable

(i.e., partitioning variable) on “earnings quality” as proxied by ijt , false inferences can result if

the partitioning variable is correlated with business model shocks. We make this prediction after

controlling for the HN operating volatility measures.

13 For example, suppose Firm A introduced a new product based on a patented technology that gains considerable market acceptance, and Firm A’s success obsoletes Firm B’s business model. Further assume it takes several years before Firm B’s revenues are adversely affected. Firm B’s stock price adjusts as the market (i) learns of Firm A’s success and comes to understand the implications of Firm A’s success on Firm B’s value, and (ii) how Firm B changes its business model in response to Firm A’s success. Hence, business model shocks to Firm B (from Firm A) in year t are not captured by the volatility of Firm B’s cash flow and revenue computed over years t-5 to t-1.

15

Prediction 5: Tests of the relation between earnings quality (proxied by ijt ) and a

partitioning variable (PART) are biased in favor of finding a positive relation between the two variables when PART is correlated with business model shocks.

Because business model shocks affect the overall fit of the accrual model, we predict a

spillover effect among firms in the same industry-year, again after controlling for the HN

operating volatility measures.

Prediction 6: Tests of the relation between earnings quality (proxied by ijt ) and a

partitioning variable (PART) are biased in favor of finding a positive relation between the two variables when PART is correlated with business model shocks experienced by firm i’(i‘ i) in industry j.

3. Data and Sample Selection

Our sample consists of the intersection of the annual Compustat file and the CRSP

monthly return file for fiscal years 1988 through 2010. We begin the sample in 1988 because our

analysis examines estimated abnormal accruals using data from the statement of cash flows

(unavailable until 1988) to avoid problems associated with the use of balance sheet data (Collins

and Hribar 2002). For each fiscal year observation we further require non-missing industry

identifiers and twelve non-missing CRSP monthly return observations, leaving a sample size of

125,964 firm-year observations, which hereafter we refer to as our "full" sample. Our accrual

model estimations impose additional data requirements. Specifically, we require non-missing

observations for accruals, cash flows, sales, PP&E, return-on-assets, and total assets. Next,

because we estimate our abnormal accrual model in industry-year cross sections, we delete

observations with fewer than twenty-five observations in a fiscal-year-industry group, resulting

in a sample of 90,822 observations.

Table 1 presents descriptive statistics of the variables in our analyses. Notably, the

median value of unsigned "abnormal" accruals from the original Jones model, nonlinear accrual

model, and performance control accrual model is 4.9%, 4.0%, and 4.4% of total assets,

respectively. In untabulated analyses we also note that the median unsigned "abnormal" accruals

are roughly the same magnitude for both negative and positive original Jones model accruals

(5.0% and 4.8% of total assets, respectively). Table 2 presents correlations among the variables

in our analyses, with Pearson (Spearman) correlations reported above (below) the diagonal.

16

Consistent with Hribar and Nichols (2007), unsigned abnormal accruals are positively correlated

with both cash flow volatility and revenue volatility, which suggests the need to control for these

variables in our multivariate analyses. We also note that the probability of a firm experiencing a

business model shock is associated with numerous firm characteristics including size, market-to-

book, and leverage.

4. Research design and empirical findings

4.1. Business model shocks

As described by Teece (2010), business models are dynamic. Unfortunately, few firms

specifically disclose changes in their business strategies. Some managers only learn that their

current business model needs revamping after their financial performance deteriorates. Because

business model shocks arise for various reasons (e.g., technology, regulatory changes, entry and

exit) and are often difficult to identify empirically from financial statements (particularly

contemporaneously), we utilize a market-based measure as our primary empirical shock proxy.

In part to help validate this market-based proxy, we also consider a financial statement-based

proxy generated from events observable in Compustat. Each proxy has advantages and

disadvantages.

4.1.1 Market-based proxy for business model shocks (RetShock)

We consider firm-years which contain an unsigned monthly market-adjusted return in

excess of 20% to have experienced a large business model shock during that year. Using stock

price data from the CRSP monthly file, we create a firm-year variable called MaxMUARi,t, which

equals the largest monthly unsigned abnormal return (i.e., firm return less the value-weighted

CRSP index return) experienced by firm i in year t. Figure 2 documents that it is relatively

common for firms to experience at least one month in a year with a large unsigned abnormal

return. The modal largest unsigned monthly abnormal return in a year is 15%. We define an

indicator variable RetShocki,t that equals one if firm i experienced at least one month during year

t with an unsigned abnormal return of greater than 20%, and zero otherwise. 55% of all firm-

years experience such an event.

We recognize that abnormal returns may also capture non-business model shocks (growth

shocks to the firm’s existing business model, liquidity shocks, and investor sentiment), which

add noise to using abnormal returns to proxy for business model shocks (Savor, 2012 and

17

Barberis, 1998).14 To examine the usefulness of using large unsigned abnormal returns to capture

business model shocks we randomly sampled 100 firm-months with unsigned monthly abnormal

return greater than 20%, read all the news stories on Factiva for that firm-month, and in

particular identified the particular day(s) with large daily abnormal returns in that month and the

stories released on that day. We also noted the unsigned Jones-model residual corresponding to

that fiscal year and the previous and following years. Appendix A provides four examples from

the 100 firms examined, with a few additional observations. Based on our examination of these

100 random large monthly abnormal returns we conclude that it is very difficult to pinpoint the

exact date when firms experience business model shocks. Business models seem to evolve over

time and shocks get impounded into stock prices as tangible evidence, usually in the form of

earnings, is released. While large abnormal monthly returns capture some of these business

model shocks, it does so with considerable error. And, large unsigned abnormal accruals are not

concentrated only in the year of the large abnormal returns. Rather, large business model shocks

(as captured by returns) tend to propagate through several years of abnormal accruals.

4.1.2. Operational proxies for business model shocks (OpShock)

As a secondary measure, we develop an operational shock proxy that relies on five

Compustat-reported data items to capture business model shocks: (i) the extent of acquisitions

(sale_fn), (ii) discontinued operations (do), (iii) four-digit SIC industry changes, (iv)

restructuring charges (rcp), and (v) special items (spi). To capture substantial business strategy

shocks using these operational variables, we create the following indicator variables:

LargeMergAcqi,t equals one if Compustat footnote data item “sale_fn” = "AB" (i.e., sales have

been "restated for/reflects a major merger or reorganization resulting in the formation of a new

company") (0.1% of sample, or 127 observations) and equals zero otherwise; LargeDiscOpsi,t

equals one if the magnitude of the income effect of discontinued operations is greater than five

percent of sales (i.e., |do/sale| > 0.05) (2.8% of sample, or 3,516 observations) and equals zero

otherwise; IndChangei,t equals one if firm i’s four digit SIC differs in years t-1 and t (3.4% of

sample, or 4,328 observations) and equals zero otherwise; LargeRestruci,t equals one if the

magnitude of restructuring charges is greater than five percent of sales (i.e., |rcp/sale| > 0.05)

(1.1% of sample, or 1,415 observations) and equals zero otherwise. LargeSpecItemi,t equals one

14 Growth shocks to the firm’s existing business models also can affect the estimation of abnormal accruals if firms adjust contemporaneous working capital in anticipation of future growth. Working capital adjustments and sales changes occurring in different years reduce the accrual model’s goodness of fit.

18

if the magnitude of special items is greater than five percent of sales (i.e., |spi/sale| > 0.05)

(12.2% of sample, or 15,324 observations) and equals zero otherwise. Further, we define an

indicator OpShocki,t that equals one if at least one of the preceding five event indicator variables

equals one (16.6% of sample, or 20,899 observations), and equals zero otherwise (i.e., OpShock

equals one if firm i experienced at least one of the above described five large operational

shocks).

About 17% of all firm years experience at least one of the five operational shock proxies.

A couple of observations are warranted. These five operational shock proxies will to varying

degrees be mechanically associated with abnormal accruals. For example, restructuring charges

in year t are mechanically related to abnormal accruals in year t. While these five operational

variables have some intuitive appeal as proxies for business model shocks, the five variables and

the cutoffs for defining “large” are ad hoc. Further, these five operational shock proxies unlikely

capture all business model shocks.

4.1.3. Association between market-based and operational shock proxies

The rationale for the tests in this section is two-fold: (i) to establish that large market-

adjusted returns are capturing changes in future cash flows due to business models shocks

(acquisitions, divestitures, restructuring charges, etc.), and are not merely capturing changes in

growth rates of firms with stable business models, and (ii) to show that it can take several years

from when the market first learns of the business strategy revision to when the change manifests

in one of the operational shock proxies. We estimate the association between our market-based

proxy (RetShock) and the five operational shocks discussed above using the following logit

model:

,

1Pr( 1) ,

1i l zOpShock

e

(3)

0 1 , 2 , 1 3 , 2 ,i t i t i tz RetShock RetShock RetShock

where all variables are as defined above.

Table 3 presents the results of separately estimating Eq. (3) for the five individual

components of OpShock, and for the aggregate OpShock indicator variable. The results are

consistent with our conjecture that contemporaneous and prior two-year lagged large unsigned

abnormal returns are associated with operational shocks. Note that our set of operational shocks

is but a subset of business model shocks. In addition, consistent with Gerakos and Kovrijnykh

19

(2012) the statistically significant coefficients on the one- and two-year lagged RetShock indicate

that business model shocks can take several years from when the market first learns of the shock

until it manifests in the financial statements. The findings in Table 3 suggest that simple cross-

sectional accrual models are likely unable to capture the complex dynamic and persistent

financial effects induced by business model shocks.

4.2. Abnormal accrual models

The above analysis suggests that business model shocks are common in our sample. We

next examine how the prevalence of business model shocks affects the estimation of abnormal

accruals. For simplicity, we begin by estimating abnormal accruals using the original model from

Jones (1991) in the cross-section by industry-year, as follows:

, 0 1 , 2 , ,i t i t i t i tTotalAccruals SalesChange PPE (4)

where TotalAccruals is taken from the statement of cash flows (Collins and Hribar 2002). As

mentioned earlier, our inferences are unchanged when using working capital accruals instead of

total accruals. SalesChange equals the change in sales from year t-1 to t, and PPE equals gross

property, plant and equipment. All variables are scaled by beginning-of-period total assets, and

Winsorized at the top and bottom one percent by two-digit SIC. As is standard, we define a

measure of abnormal accruals, AbAccruali,t, as the residual from estimating Eq. (4). Further, we

define an unsigned form of this residual, UAA, as |AbAccrual|.

Next, we repeat our analyses using a nonlinear modification to the Jones model as

proposed by Ball and Shivakumar (2006). Specifically, we estimate the following model in the

cross-section by industry-year:

, 0 1 , 2 , 3 , 4 , 5 ,

6 , 7 , 8 , ,

*

* ,i t i t i t i t i t i t

i t i t i t i t

TotalAccruals SalesChange PPE CF DCF DCF CF

ABNRET DABNRET DABNRET ABNRET

(5)

where CF is operating cash flows scaled by average total assets, DCF is an indicator that equals

one if CF is less than zero and equals zero otherwise, ABNRET is firm i's abnormal stock return

during fiscal year t (based on the CRSP equal-weighted market index), DABNRET equals one if

ABNRET < 0 and zero otherwise, and all other variables are as defined above. We denote the

residual from Eq. (5) as AbAccrual_NL, and the corresponding unsigned residual as UAA_NL.

Finally, we use a variant of the original Jones model that adds a control for firm

performance, as suggested by Kothari et al. (2005):

20

, 0 1 , 2 , 3 , , ,i t i t i t i t i tTotalAccruals SalesChange PPE ROA (6)

where ROA is calculated as net income divided by average total assets, and all other variables are

as previously defined. We denote the residual from Eq. (6) as AbAccrual_PF, and the

corresponding unsigned residual as UAA_PF.

Figure 3 plots histograms of the R2 from the 878 industry-year estimations of Eqs. (4) -

(6). There is wide variation in the model fit across estimations. In the original Jones model, most

industry-years have R2s of less than 10%, suggesting that most industry-year models are

estimated relatively imprecisely. While there is substantially improved model fit using both the

nonlinear and performance control accrual models from Eqs. (5) and (6), the sample mean and

median values of the abnormal accruals from these models are still relatively large (see Table 1).

For example, 42% of sample firm-years (untabulated) have an abnormal accrual from the

nonlinear model (i.e., UAA_NL) greater than 5% of total assets.

The literature generally attributes the entire abnormal accrual to "discretion," or earnings

management.15 However, given the large magnitude of the mean and median unsigned abnormal

accruals, such earnings management would be difficult to disguise from the external auditors

(Ball and Shivakumar, 2008). In applying generally accepted auditing standards auditors seek

reasonable assurance that the financial statements are free from material misstatement whether

from errors, fraud, or management’s use of unreasonable estimates. Since a number of accounts

require management judgment (deferred taxes, inventory obsolescence, uncollectible receivables,

warranty obligations, pension liabilities, impaired assets), auditors must test the reasonableness

of these estimates, and “indicate a possible bias on the part of the entity's management.” (AU

Sections 312.37 and 9312A PCAOB).16 Auditors often base their materiality assessment on a 5%

of net income rule, which holds that “a reasonable investor would not be influenced in

15 For example, Cohen, et al. (2008) use unsigned abnormal accruals to measure accounting-based earnings management and report median modified-Jones unsigned abnormal accruals of 6% of total assets over their sample period 1987-2005. 16 PCAOB standard AU 9312A reads, “The auditor should also consider whether the difference between estimates best supported by the audit evidence and the estimates included in the financial statements, which are individually reasonable, indicate a possible bias on the part of the entity's management. For example, if each accounting estimate included in the financial statements was individually reasonable, but the effect of the difference between each estimate and the estimate best supported by the audit evidence was to increase income, the auditor should reconsider the estimates taken as a whole. In these circumstances, the auditor should reconsider whether other recorded estimates reflect a similar bias and should perform additional audit procedures that address those estimates.”

21

investment decisions by a fluctuation in net income less than or equal to 5%” (Vorhies 2005).17

In contrast, the median Jones model abnormal accrual is 67% of unsigned net income.18 These

unsigned abnormal accruals clearly exceed the auditor’s materiality threshold of 5% of net

income, and hence are likely to have been examined by the external auditor and deemed to be

free from material misstatement (including unreasonable management bias). On the assumptions

that auditors detect and prevent material misstatements (including unreasonable management

estimates) and that they apply a 5% of net income materiality threshold, then if a firm has an

abnormal accrual of 67% of unsigned net income it indicates that abnormal accruals as a proxy

for managerial accounting discretion have a noise-to-signal ratio of roughly 12 (62% ÷ 5%).

4.3. Association between unsigned abnormal accruals and business model shocks

We next examine whether the existence of business model shocks affect the estimation of

abnormal accruals. We first test Prediction 1 on the relation between the magnitude of abnormal

accrual estimates and business model shocks in the current and previous years. We focus on

unsigned abnormal accruals because business model shocks can increase or decrease accruals.

We begin by estimating the following OLS model:

, 0 1 , 2 , 1 3 , 2

4 , 5 , , ,i t i t i t i t

i t i t i t

UAA BMShock BMShock BMShock

CFO REV

(7)

where UAA is unsigned Jones model abnormal accruals and BMShock is, alternately, OpShock

and RetShock, i.e., the indicators for the presence of business model shocks, as discussed above.

σCFO and σREV capture operating volatility, measured as the standard deviation of cash flow

and sales, respectively, scaled by assets over the current and previous four years (Hribar and

Nichols, 2007). We also estimate a logit specification that tests for a relation between business

model shocks and large unsigned abnormal accruals, as follows:

,

1Pr( 1) ,

1i t zLUAA

e

(8)

17 In SAB 99 the SEC cautions auditors to avoid “exclusive reliance on certain quantitative benchmarks to assess materiality in preparing financial statements and performing audits.” Moreover, “the staff has no objection to such a ‘rule of thumb’ as an initial step in assessing materiality. But quantifying, in percentage terms, the magnitude of a misstatement is only the beginning of an analysis of materiality; it cannot appropriately be used as a substitute for a full analysis of all relevant considerations.” The 5% net income rule of thumb is consistent with Dichev’s et al. (2012) finding that CFOs believe about 10% of reported earnings is managed for firms that manage earnings. 18 Because we are interested in the magnitude of abnormal accruals relative to the magnitude of net income, we calculate |abnormal accrualsi,t | ÷ |net incomei,t |. Inferences are similar using the nonlinear and performance control accrual models in Eqs. 5 and 6.

22

0 1 , 2 , 1 3 , 2

4 , 5 , ,i t i t i t

i t i t

z BMShock BMShock BMShock

CFO REV

where LUAA is an indicator that equals one if UAA is greater than the sample median (i.e., the

absolute value of abnormal Jones model accruals is greater than 5% of total assets) and equals

zero otherwise. All other variables are as previously defined. We also repeat the estimation of

Eqs. (7) and (8) using both the nonlinear and the performance controlled accrual models.

Before estimating Eqs. (7) and (8), Figure 4 presents time-series plots of the proportion of

sample firms each year with business model shocks and large abnormal accruals (i.e., OpShock =

1, RetShock = 1, LUAA = 1, LUAA_NL = 1, and LUAA_PF = 1) to give a sense of associations

between time trends in these variables. Figure 4 shows that 40% of all firms consistently have

unsigned abnormal accruals in excess of 5% across all model specifications. Further, our

instruments for business model shocks indicate that in excess of 20% and 60% of firms have

operational shocks and large return shocks, respectively, in most years. Moreover, these

variables are strongly correlated through time. For example, the correlation (untabulated)

between the percentage of firms with OpShock = 1 and RetShock = 1 is 0.71 and the correlation

between the percentage of firms with LUAA = 1 and RetShock = 1 is 0.72.

Table 4 reports estimation of Eqs. (7) and (8). As shown in Panel A, contemporaneous

operational shocks (OpShock) are positively associated with the magnitude of unsigned abnormal

accruals and the presence of a large abnormal accrual for all three abnormal accrual models we

estimate. Results are generally weaker for lagged operational shocks. As shown in Panel B,

contemporaneous and both one and two-year lagged stock return shocks are positively associated

with both measures of abnormal accrual magnitude across all models, where the strength of the

association generally diminishes with the lag of the return shock. These results support

Prediction 1. The operating volatility control variables are positively associated with the

likelihood of a firm having a large unsigned Jones model abnormal accrual, consistent with

Hribar and Nichols (2007). So even after controlling for cash flow (σCFO) and revenue (σREV)

volatility, which likely also capture business model shocks, our operational shock (OpShock) and

market shock (RetShock) proxies (including lagged variables) continue to explain both the

magnitude and the presence of large unsigned abnormal accruals.

To test Prediction 2 on the spillover effect from other firms in the same industry-year, we

estimate the following OLS model:

23

, 0 1 , 2 , 1 3 , 2

4 , 5 , 1 6 , 2

7 , 8 , , ,

i t i t i t i t

i t i t i t

i t i t i t

UAA MaxMUAR MaxMUAR MaxMUAR

OtherMaxMUAR OtherMaxMUAR OtherMaxMUAR

CFO REV

(9)

where MaxMUARi,t is firm i's maximum monthly unsigned absolute market-adjusted return

during year t and OtherMaxMUARi,t is the average MaxMUAR across all other firms in firm i's

industry during year t. The results are presented in Table 4 Panel C. We find similar results

regarding a firm’s own business model shocks, in that MaxMUAR is positively associated with

large abnormal accrual magnitudes. These results confirm those in Panel B and support

Prediction 1. Turning to the coefficients on OtherMaxMUAR, i.e., shocks experienced by the

other firms, we find uniformly positive and significant coefficients on OtherMaxMUARi,t across

the three accrual models, indicating a spillover effect from the business model shocks

experienced by other firms due to their effect on the accrual model fit. These results support

Prediction 2. The coefficients on the lagged OtherMaxMUAR variables are insignificant.

4.4. Effect of sample "business model shock" contamination on abnormal accruals

To further illustrate the dynamics behind our key results in Table 4, in this section we

examine how inclusion of an increasing number of "contaminated" firm-year observations with

large business model shocks affect the estimation of accrual models. For parsimony, we focus on

how including or excluding observations with business strategy shocks in either years t, t-1 and t-

2 impact estimated abnormal accruals where the existence of a shock is proxied by an unsigned

monthly abnormal stock return during the year in excess of 20% (RetShock =1).

We begin by estimating the abnormal accrual models of Eqs. (4)-(6) by industry-year

using only "uncontaminated" observations (i.e., RetShockt = RetShockt-1 = RetShockt-2 = 0), again

requiring at least 25 observations in each industry-year regression. The resulting

"uncontaminated" sample consists of 10,844 observations. We report the average unsigned

abnormal accrual (UAA, UAA_NL, UAA_PF) and the percent of observations with a large

unsigned abnormal accrual (LUAA, LUAA_NL, LUAA_PF). Then, we iteratively repeat the

estimations after increasing the number of "contaminated" observations in the sample.

Specifically, in stepwise fashion we randomly select 12.5%, 25%, 37.5%, … 100% of the

contaminated observations to include in the sample, recording the average unsigned abnormal

accrual and the percent of observations with a large unsigned abnormal accrual in each iteration.

We predict that each of these statistics will increase with the number of contaminated

24

observations included in the sample. Each iteration tracks these statistics separately for the

original uncontaminated 10,844 observations to separately document how inclusion of

contaminated observations affects abnormal accruals of the uncontaminated sample.

Table 5 presents the results from this analysis. Panel A reports the sample composition

across the nine iterations. The number of “uncontaminated” observations increases as more

“contaminated” observations are included because more industry-years acquire the 25

observations required to estimate the regression. Panel B tabulates the average unsigned

abnormal accruals, percent of observations with a large abnormal accrual, and median industry-

year R2 from estimations using all three abnormal accrual models. Figure 5 presents these

statistics in graphical form. For brevity, we will narrate the abnormal accrual results only for the

Jones model, as the results from the other models are inferentially equivalent. Focusing on

columns (1) and (2) in Panel B1, the sample of purely uncontaminated observations iteration (1)

has an average UAA of 0.033 (i.e., abnormal accrual of 3.3% of total assets), where 19% of the

observations have a large abnormal accrual (i.e., UAA > 0.05). In stark contrast, estimation with

the full sample (iteration 9) generates an average UAA of 8.1%, where 49% of the observations

have a large abnormal accrual. As can be seen in both Panel B and Figure 5, consistent with our

expectation there is a monotonically increasing pattern in both the average UAA and the

percentage of observations with a large abnormal accrual (LUAA) as the number of contaminated

observations in the sample increases. This is further evidence in support of Prediction 1.

Interestingly, the increasing pattern is concave, which reveals that it only takes a relatively small

number of contaminated observations to significantly affect the outputs of the accrual model.

Stated differently, if one starts with the fully contaminated sample and begins removing

contaminated observations, little is accomplished unless a substantial number of contaminated

observations are removed. Therefore, studies that simply exclude one or a few specifically

identified events such as mergers are likely not substantially altering model performance.

Columns (3) and (4) of Panel B present the effect on the initial 10,848 uncontaminated

observations of adding an increasing number of contaminated observations to estimating the

accrual model. A similar pattern emerges, in that the constant sample of uncontaminated

observations likewise exhibits a monotonic increase in both the average UAA and the percent of

observations with a large abnormal accrual as the number of contaminated observations in the

sample increases. Thus, not only do contaminated observations themselves display problematic

25

statistics from estimation of accrual models, but their inclusion in the model estimation likewise

affects the efficiency with which accrual models are estimated for the uncontaminated

observations. Whereas 19% of the original 10,844 uncontaminated observations have a large

abnormal accrual when the model is estimated without contamination (iteration 1), 27% of the

same 10,844 observations have a large abnormal accrual when the full sample, including all

contaminated observations, are included in the model estimation (iteration 9). The above findings

are consistent with Prediction 2’s spillover effect.

Column (5) of Panel B in Table 5 reports the median industry-year R2 across sample

contamination iterations. In both the Jones and nonlinear models, there is a nearly monotonic

decrease in industry-year R2 as sample contamination increases, which reinforces the inference

that contamination reduces the fit of the accrual models, i.e., there is a decreasing pattern in R2 as

the number of contaminated observations in the sample increases, and supports Prediction 3.

Interestingly, as shown in Panel B3, when ROA is included in the accrual model, industry-year

R2 increases with the degree of sample contamination. This is inconsistent with Prediction 3 and

is likely caused by the correlation between return shocks and ROA. Nonetheless, even in this

case, the average unsigned abnormal accrual and the proportion of observations with large

unsigned abnormal accruals increases as sample contamination increases.

To further examine the relation between the presence of shocks and accrual model fit, we

regress R2 from individual industry-year accrual models on the percentages of firms in that

industry-year regression that that have RetShock = 1 in years t, t-1, and t-2. In untabulated results

we find negative coefficients on the percentage of firms experiencing business strategy shocks in

year t for the Jones and nonlinear accrual models, but a positive coefficient on the performance

control accrual model (consistent with Table 5 Panel B3). Also, the coefficient on the percentage

of firms with large return shocks in t-2 is negative and statistically significant in the nonlinear

accrual model.

4.5. Cash operating cycle variability and business model shocks

In this section, we test Prediction 4 on whether business model shocks are associated

with greater volatility in firm operating cycles by estimating the following regression.

,

1Pr( 1) ,

1i t zOPCYCVOLUP

e

(10)

0 1 ,i tz BMShock

26

where OPCYCVOLUPt is an indicator that equals one if firm i's 4-quarter operating cycle

volatility increased from year t-1 to t, and equals zero otherwise. To calculate operating cycle

volatility in year t, we first compute the operating cycle for each quarter in year t as "days sales

outstanding" + "days inventory outstanding" – "days payables outstanding." We then take the

standard deviation of these four quarterly operating cycles as year t's operating cycle volatility. If

this standard deviation increases from year t-1 to t, then OPCYCVOLUPt = 1.19

We report the results from Eq. (10) in Table 6, with the operational shocks in column (1)

and return shocks in column (2). The results in both columns confirm that business model shocks

in year t are associated with a higher likelihood of increases in firm operating cycle volatility,

consistent with Prediction 4. As firms’ operating cycles become more volatile due to business

model shocks, the empirical accrual models’ assumptions of intra-industry homogeneity and firm

stationarity are likely violated, reducing the fit of the models.

4.6. Business model shocks and inferences using abnormal accruals

4.6.1. Rejection frequencies from simulations – unsigned abnormal accruals

This section tests Predictions 5 and 6 on how business model shocks affect researchers’

inferences regarding earnings management/earnings quality measured using unsigned Jones

model abnormal accruals. We follow the methodology in HN, where we construct a partitioning

variable (PART) as a weighted combination of MaxMUARi,t and a random variable. We draw 250

random samples of 1,000 observations each from our sample and run a regression of unsigned

Jones model abnormal accruals on PART for each random sample. We record the number of

times out of 250 the t-statistic on the PART rejects the null hypothesis of no correlation between

unsigned abnormal accruals and PART at the 5% level in a one-tailed test. We repeat this process

11 times at different levels of correlation between MaxMUARi,t and PART. Within a randomly

selected sample, we expect no systematic association between earnings management/earnings

quality and large return shocks, suggesting rejection rates of around 5% in properly designed

tests. The actual rejection rates are plotted in Figure 6 Panel A, with the diamond-dotted line for

regressions with no control variables and the square-dotted line for regressions with the

following controls: LnSize, MktToBook, Leverage, and the HN operating volatility measures of

19 We do not include the lagged business model shocks in Eq. 10 due to ambiguous predictions on the relation between lagged shocks and the change in operating cycle volatility. For example, a large positive shock in year t-1 likely leads to higher operating cycle volatilities in year t-1 and it may also increase the operating cycle volatility in year t if the shock affects multiple years’ accruals generating processes. As a result, the implication of the lagged shock on the change in operating cycle volatility from year t-1 to year t is unclear.

27

σCFO and σRev. Focusing first on the diamond line with no control variables, we find excessive

rejection rates (i.e., > 5%) in all cases when there is a positive correlation between MaxMUARi,t

and PART, starting from 0.1. Adding the control variables substantially lowers the rejection

rates. However, the test still over-rejects the null even at modest levels of correlations such as 0.2

after controlling for the HN volatility measures. These findings support Prediction 5 that

business model shocks can create false inferences in research on “earnings quality” due to

correlations between researchers’ partitioning variables and firms’ business model shocks.

We test Prediction 6 by repeating the above analysis replacing MaxMUARi,t with

OtherMaxMUARi,t. The rejection rates are plotted in Panel B of Figure 6, again with the

diamond-dotted line for regressions with no control variables and the square-dotted line for

regressions with the same control variables in Panel A. When no control variables are included,

the rejection rate is greater than 5% at 0.1 correlation and climbs quickly as the correlation rises.

Even with the controls, the test over rejects the null at very modest levels of correlation and the

rejection rates quickly become elevated. These results support the spillover effect in Prediction 6

and suggest a new dimension of potentially correlated omitted variables not considered in prior

literature – business model shocks experienced by other firms in the industry affect the overall fit

of accrual models.

4.6.2. Illustrative example

To further illustrate the issues with drawing inferences based on the use of unsigned

abnormal accruals to measure “discretion” in the presence of frequent business model shocks, we

revisit the setting examined in Bharath et al. (2008) (hereafter BSS). A key question addressed

by BSS is how "accounting quality" affects the interest rate paid on private debt, i.e., bank loans.

The authors state that "we measure accounting quality using the magnitude of operating accruals

to proxy for the influence of discretionary accounting choices. Large abnormal operating

accruals ... make it harder for the lenders to reliably estimate future operating cash flows" (p. 2).

To measure accounting quality, the authors use the first principle component from three

measures of abnormal operating accruals, one being the modified Jones model estimated by

industry-year. Among other things, BSS provides evidence that a higher level of accounting

quality (i.e., lower abnormal operating accruals) leads to lower cost of private debt.

As our study focuses on abnormal accruals from the Jones model, we re-examine the

question of BSS using this single proxy for discretionary accruals, rather than their principle

28

component approach using multiple measures. As such, this example is not intended to challenge

that study's key inference that accounting quality is associated with cost of bank debt. Rather, we

use their setting to illustrate issues associated with making inferences based on the use of accrual

model residuals as a specific proxy for “discretionary accruals.”

We first obtain a sample of bank loan facilities issued during the period 1988-2010 from

Dealscan, and similar to BSS we exclude very short-term loans (i.e., maturities less than three

months). We match the Dealscan loan facilities with our Jones model sample using the most

recent annual accounting data available prior to the issuance date for a given loan, and delete

observations with missing data required by our analysis. The final loan sample contains 23,738

distinct loan facilities. Using this loan sample, we estimate the following OLS specification used

by BSS in column (1) of their Table 5 (p. 17), where we likewise follow BSS and cluster

standard errors at the firm level:

, 0 1 , 2 , 3 , 4 ,

5 , 6 , 7 , 8 ,

9 , 10 , , .

i l i t i t i t i t

i t i t i m i l

i l i l i l

Spread UAA Leverage LnAssets Tangibility

CurrRatio MktToBook DefaultRisk LnFacility

LnMaturity Secured

(11)

Spread is the loan basis point spread over LIBOR, Tangibility is net PP&E divided by total

assets, LnAssets is the natural log of total assets, and CurrRatio is current assets divided by

current liabilities. DefaultRisk is a market-based measure of default likelihood for the most

recent month prior to loan initiation, calculated based on the Black-Scholes-Merton equation

closely following the approach in Hillegeist et al. (2004). LnFacility is the natural log of the loan

amount, LnMaturity is the natural log of the loan maturity in months, and Secured is an indicator

that equals one if the loan has a collateral requirement and equals zero otherwise. All variables

are described more fully in Appendix B.

We present the results from estimating Eq. (11) in column (1) of Table 7. Inferences are

consistent with the key findings in BSS. Specifically, higher values of UAA (i.e., lower

"accounting quality") are associated with higher loan spreads (coefficient estimate of 80.58 with

a t-statistic of 4.83). This result would lead to the same inference made by BSS that low

accounting quality (as proxied by large unsigned abnormal accruals) leads lenders to charge a

higher interest rate because of "information risk." The signs and significance of the control

variables are generally consistent with those reported in BSS.

29

Next, we introduce an indicator variable for the presence of business model shocks (i.e.,

"contaminated" observations), and its interaction with UAA, to examine whether the association

between UAA and spread is different for contaminated versus uncontaminated observations:

, 0 1 , 2 , 3 4 , 5 ,

6 , 7 , 8 , 9 ,

10 , 11 , 12 , ,

*

.

i l i t i t i t i t

i t i t i t i m

i l i l i l i l

Spread Contam UAA Contam UAA Leverage LnAssets

Tangibility CurrRatio MktToBook DefaultRisk

LnFacility LnMaturity Secured

(12)

Columns (2) and (3) of Table 7, respectively, report results from estimating Eq. (12)

using our RetShock and OpShock proxies for the presence of business model shocks within the

current or previous two years. In these specifications, the association between UAA and Spread

are only significant for those observations that have experienced a business model shock.

Focusing on column (2), the coefficient estimate on UAA (25.18; t-statistic 0.66) captures the

insignificant relation between UAA and Spread for firms that have not experienced a business

model shock. In contrast, the coefficient sum UAA + Contam*UAA captures the relation between

UAA and Spread for firms experiencing a business model shock. As indicated by the associated

χ2 statistic, the sum of 73.25 is statistically significant at the p < 0.0001 level. These findings

suggest that the significant relation documented in column (1) between UAA and Spread is

driven by firms with business model shocks within the three years prior to loan inception. This

raises questions about attributing the findings in column (1) to "information risk" arising from

discretionary accounting choices. Alternatively, the relation between UAA and Spread may

simply reflect firms undergoing business model shocks being riskier than firms not experiencing

a strategy shock in a manner not captured by market-based default risk measures.

4.6.3. Rejection frequencies from simulations – signed abnormal accruals

Kothari et al. (2005) show that accrual models are likely to be misspecified, i.e.,

researchers tend to over-reject the null of no earnings management, for firms with extreme

characteristics. Following their methodology, we group all 90,822 observations in our Jones

model sample into quartiles each year based on the following variables -- market-to-book, sales

growth and firm size. We combine the observations in each quartile across the different years

and randomly select 200 firms without replacement from each quartile and calculate the mean

and t-statistic of the signed abnormal accruals from the Jones model. We repeat this process 250

times for each quartile, from which we compute the rejection rates for the null hypothesis of

AbAccrual = 0 for both the alternative hypotheses (Ha) of AbAccrual < 0 and Ha of AbAccrual >

30

0 at the 5% level using one-sided tests. The null rejection rates for the market-to-book quartiles

are reported in Table 8 Panel A, those for size are in Panel B, and sales growth in Panel C.

Rejection rates that are significantly different from the 5% nominal significance level (i.e., below

2% or above 8%) are bolded in the table. Below each row that presents the proportion of random

samples in each quartile where the null is rejected, we tabulate the actual number of random

trials (out of the 250) where the null is rejected. For brevity we will discuss only the results in

Panel A (partitions based on market-to-book), as inferences in Panels B and C are similar.

Again focusing on the full sample (Columns 1-6), in random sampling from the entire

pool of 90,822 sample observations we find that rejection rates are not statistically different from

5%, similar to findings in prior studies (Column 1). However, we find that firms in the extreme

market-to-book quartiles tend to exhibit excessive rejection rates on the negative side, i.e. 33.6%

(16.8%) for the lowest (highest) market-to-book quartile, supporting Ha of AbAccrual < 0. This

is consistent with inferences from Kothari et al. (2005) for the Jones model in these extreme

portfolios. Even though Kothari et al. (2005) do not report the rejection rates for the two middle

quartiles, we show excessive rejection rates there (42% and 46.8%, respectively) supporting Ha

of AbAccrual > 0. These findings suggest that excessive rejection of the null is not limited to the

extreme portfolios. Column (6) reveals that for the 1,000 combined random trials across the four

quartiles, AbAccrual > 0 (AbAccrual < 0) is accepted 226 (126) times, in contrast to an expected

acceptance frequency of 50 in a well specified model (i.e., 5% of 1,000 trials).

To investigate the effect of business model shocks, we next exclude all observations with

large absolute abnormal returns in the current year (resulting in a much smaller subsample of

33,027 firm-year observations) and re-estimate the Jones model by industry-year and then repeat

the above rejection tests, and present results in Columns (7)-(12). This analysis shows that the

over-rejection problem is greatly mitigated when observations with business model shocks are

excluded from the analysis, and the cell-by-cell change in the rejection rates is generally

statistically significant based on a two-sample Wald test. To get an overall sense of the

improvement in model specification that comes from removal of shock observations, it is useful

to compare Columns (6) and (12). Whereas for the full sample 226 (126) out of 1,000 trials result

in the inference AbAccrual > 0 (AbAccrual < 0), for the no-shock subsample, only 78 (66) out of

1,000 trials results the inference AbAccrual > 0 (AbAccrual < 0). Accordingly, over-rejection of

the null in both directions is greatly mitigated when shock observations are excluded.

31

The above analysis indicates that the presence of business model shocks likely generates

false inferences when signed abnormal accruals are used to capture earnings management. We

emphasize that even though the reduced sample produces better specified tests, throwing out

observations with large absolute returns is not a panacea, as it can lead to greatly reduced sample

sizes. The reduced sample in Table 8, which excludes observations with large returns in the

current year only, is about 1/3 the size of the overall sample. Further excluding firms with large

returns in prior years (because business model tend to persist through several years’ financial

statements) would have contributed to even further loss of observations.

5. Conclusion

In this paper we argue that economic insights pose serious challenges to the two

maintained assumptions underlying empirical accrual models: intra-industry homogeneity and

firm-stationarity. In particular, economics teaches that technological innovation, changes in

regulation, and entry by new (or existing firms) create shocks to existing firms’ business models.

These shocks cause revisions in strategies and hence changes in real investment, operating, and

organizational policies, and eventually different accrual generating processes.

Business model shocks cause firms to revise their business strategies by undertaking real

transactions such as acquisitions, divestitures, outsourcing, insourcing, and reorganizations, all of

which generate accruals. Because firms in the same industry can react to the same shock with

diverse business strategy revisions (because they are seeking differentiated strategies and have

different core competencies), a given business model shock does not have the same effect on

total accruals for all firms in the industry. Therefore, including an instrument for the business

model shock in the cross-sectional industry-year accrual model only captures the average effect

of the shock in “normal” or “non-discretionary” accruals. The firm-specific effect of the shock

on total accruals ends up in the regression residual and is mischaracterized as “discretionary”

earnings management. Moreover, since it often takes firms several years to adapt to a business

model shock, past business model shocks propagate through several years of firm’s accruals and

cause instability of a firm’s accrual generating process.

We provide evidence that business model shocks are pervasive, and have consequences

for the firm-stationarity and intra-industry-homogeneity assumptions underlying empirical

accrual models. In particular, large unsigned abnormal accruals are associated with

contemporaneous and lagged proxies for business model shocks and with business model shocks

32

experienced by other firms in the same industry-year. Business model shocks reduce accrual

model goodness of fit, and it takes relatively few firm years contaminated by a business model

shock to substantially affect the outputs of accrual models. As more “contaminated” observations

are included in the sample, the Jones model R2 declines monotonically in a convex fashion, and

the magnitude of abnormal accruals rises monotonically in a concave pattern. Finally, we

document that the inclusion of business model shock observations in industry-year abnormal

accrual estimations can lead to spurious inferences in studies that use either unsigned or signed

abnormal accruals as proxies for earnings management / earnings quality. We further note that

little is accomplished in terms of goodness of fit and smaller abnormal accruals unless a

substantial number of contaminated observations are removed from the sample. Studies that

simply exclude one or a few specifically identified events, such as mergers or acquisitions,

unlikely alter substantially model performance.

We see no clear roadmap to resolve the accrual model misspecification problems we raise

because business model shocks have large effects on the accrual generating process, they occur

randomly, their effects on accruals can persist over several years, observable proxies for these

shocks are noisy, and firms in the same industry respond to shocks in diverse ways causing

different implications for signed abnormal accruals. Contributing to the difficulty of developing

better specified accrual models that explicitly recognize intra-industry heterogeneity and firm-

non-stationarity is the absence of satisfactory theoretical and empirical industrial organizational

models that explain entry, growth, and survival across different markets and regulatory regimes

(Sutton, 2007). Although our investigation has focused specifically on abnormal accruals, our

results have implications for a broader set of “earnings quality” measures. These include the

Dechow and Dichev (2002) accrual quality measure because business model shocks can

similarly affect their model’s goodness-of-fit and their measure of earnings quality. Other

earnings quality measures such as earnings persistence and earnings smoothness that rely on

time-series properties of accounting earnings may also suffer from the confounding effects due

to volatilities induced by business model shocks.

Finally, our analysis has implications for the stream of literature that investigates the

consequences of having lower “earnings quality,” including documenting an association between

“earnings quality” and costs of capital or firm investment behavior (Dechow et al., 2010 review

this literature). The association between “earnings quality” and some variable of interest such as

33

cost of capital often is explained by “information risk” caused either by managers reporting

lower quality earnings to obfuscate the true nature of the firm’s underlying economic

performance, or by the innate information risk due to the complexity of the firm or its

environment. Our evidence suggests that firms face inherent risk in their business models and

that many measures of “earnings quality” likely proxy for this risk. In other words, in most

accounting studies of “earnings quality,” business model risk is a correlated omitted variable and

the various instruments for “earnings quality” capture this risk. If proxies for information risk are

indeed driven by inherent business model risk, then one needs robust models of the relation

between business model risk and innate information risk to estimate the discretionary portion of

information risk (Zimmerman 2013).

34

Appendix A

Examples of Firms with Large Unsigned Monthly Abnormal Returns, the Associated News Stories in that Month, and the Lead, Lagged, and Contemporaneous Unsigned Abnormal Accruals We randomly selected 100 firm-year observations with unsigned monthly abnormal returns of at least 20%. We searched all news stories on Factiva for the corresponding month with the large unsigned monthly abnormal return. The following are four examples. 1. Mercury Computer Systems Inc (Market cap: $246 million) Large monthly abnormal return: 128% Month of large unsigned monthly abnormal return: 12/08 Lagged unsigned Jones-model abnormal accrual: 0.01 Contemporaneous unsigned Jones-model abnormal accrual: 0.08 Lead unsigned Jones-model abnormal accrual: 0.13 Various News stories (12/15/08): “Visage Imaging Receives FDA 510(k) Clearance for its Latest Thin Client” (12/20/08) “Mercury Computer Systems Inc., said this week its Visage unit received Food and Drug Administration clearance to sell a medical- imaging unit. The Visage CS 3.1 performs a range of imaging functions, including CAT scans and magnetic resonance imaging. In addition, the Chelmsford manufacturer of computer, signal, and image- processing systems and software is launching the Echotek Series of Virtex digital receivers.”

Abnormal returns: 16% (12/2/08), 11% (12/3/08), 13% (12/9/08), 20% (12/19/08)

2. CompuCredit Corporation (Market cap: $325 million) Large monthly abnormal return: -52% Month of large unsigned monthly abnormal return: 1/02 Lagged unsigned Jones-model abnormal accrual: 0.00 Contemporaneous unsigned Jones-model abnormal accrual: 0.38 Lead unsigned Jones-model abnormal accrual: 0.24

News story (1/29/02): “CompuCredit announced preliminary fourth quarter 2001 net income of $5.6 million, or $0.12 per share. Wall Street analysts on average were expecting the Company to earn $0.28 per share in the same period, according to Multex Global Estimates. The Company cited reduced loan growth and increased expenses as the primary reasons for its expected earnings shortfall. The Company also announced that it has reduced its workforce by approximately 70 employees. CompuCredit is a credit card company that uses analytical techniques, including sophisticated computer models, to market general-purpose credit cards and related fee-based products and services.”

Abnormal return on 1/30/02: -42%

3. Tenet Healthcare Corp. (Market cap: $5.1 billion) Large monthly abnormal return: -25% Month of large unsigned monthly abnormal return: 1/04 Lagged unsigned Jones-model abnormal accrual: 0.08 Contemporaneous unsigned Jones-model abnormal accrual: 0.08 Lead unsigned Jones-model abnormal accrual: 0.04

35

News story (1/28/04): “Tenet Healthcare plans to sell more than a quarter of its hospitals this year and its dim near-term outlook signaled that a turnaround is more distant than anticipated.”

Abnormal return on 1/28/04: -17%

4. Geo Group (Market cap: $741 million) Large monthly abnormal return: 46% Month of large unsigned monthly abnormal return: 3/06 Lagged unsigned Jones-model abnormal accrual: 0.04 Contemporaneous unsigned Jones-model abnormal accrual: 0.10 Lead unsigned Jones-model abnormal accrual: 0.12

News story (3/31/06): “The GEO Group, Inc. today increased its previously issued earnings guidance for the first quarter of 2006 by $0.10 per share to a range of $0.39 to $0.41 per share. This increased guidance is primarily attributable to increased occupancy at GEO's facilities, including those formerly operated by Correctional Services Corporation. GEO is increasing its first quarter 2006 revenue guidance by $2.0 million to a range of $184 million to $188 million. … GEO has previously announced its intention to restructure its relationship with CentraCore Properties Trust ("CPT"), from whom GEO leases 11 of its correctional and detention facilities (the "Leased Facilities"). … GEO has previously stated that it may elect to not exercise its exclusive option to renew certain of the Expiring Leases in favor of the construction and development through government-sponsored bonds or other third party financing of new replacement facilities in close proximity to the facilities covered by the Expiring Leases. In furtherance of this strategy, GEO is announcing that it has acquired three properties in close proximity to certain of the expiring Leased Facilities, and has entered into an agreement to buy a fourth property adjacent to an expiring Leased Facility. GEO is also in negotiations to purchase additional properties in close proximity to the Leased Facilities. … GEO is a world leader in the delivery of correctional, detention, and residential treatment services to federal, state, and local government agencies around the globe. GEO offers a turnkey approach that includes design, construction, financing, and operations. GEO represents government clients in the United States, Australia, South Africa, Canada, and the United Kingdom. GEO's worldwide operations include 61 correctional and residential treatment facilities with a total design capacity of approximately 49,000 beds.”

Abnormal return on 1/30/02: 14%

Additional observations: Very few stories corresponding to the large daily abnormal returns specifically mentioned that the firm was revising its business strategy, although a few did (see GEO Group). Many of the stories associated with large daily abnormal returns involved earnings releases. Within these stories the firm offers more texture about the success or failure of its current business model. For example, CompuCredit reported lower earnings citing reduced loan growth, and announced layoffs. Its stock price fell 42% on that day. Such a story provides the market with tangible information about the success (or failure in this case) of the firm’s strategy. Interestingly, the current and following years’ unsigned abnormal accruals are 38% and 24% of total assets, respectively. Other large monthly abnormal returns are associated with new products, joint ventures, and FDA drug approvals. For example, Mercury Computer’s stock price rose 128% in December 2008. Various stories started to appear during December that the FDA approved the sales of its medical imaging system. Mercury’s lagged, contemporaneous, and lead abnormal accruals were 1%, 8%, and 13% of total assets, respectively.

36

Appendix B Variable Definitions σCFOi,t Standard deviation of cash flow from operations (Compustat oancf)

deflated by total assets (Compustat at) over the current and prior four years

σREVi,t Standard deviation of sales (Compustat sale)deflated by total assets (Compustat at) over the current and prior four years

AbAccruali,t Firm i's year t abnormal accrual, estimated as the residual from the estimation of the cross-sectional version of the original Jones model by industry-year.

AbAccrual_NLi,t Firm i's year t abnormal accrual, estimated as the residual from the estimation of the cross-sectional version of the nonlinear accrual model (Ball and Shivakumar 2006) by industry-year.

AbAccrual_PFi,t

Firm i's year t abnormal accrual, estimated as the residual from the estimation of the cross-sectional version of the accrual model with ROA performance control (Kothari et al., 2005) by industry-year.

ABNRETi,t Firm i's cumulative abnormal stock return during fiscal year t (using the CRSP monthly returns file), where expected return is measured by the equal-weighted CRSP index.

CFi,t Firm i's cash flow from operations (Compustat oancf) in year t, scaled by average total assets (Compustat at).

ContamFSi,t An indicator variable that equals one if OpShockt=1 or OpShockt-1=1 or OpShockt-2=1 .

ContamReti,t An indicator variable that equals one if RetShockt=1 or RetShockt-1=1 or RetShockt-2=1 .

CurrRatioi,t Firm i's quarter t current ratio, measured as current assets divided by current liabilities (Compustat act/lct).

DABNRETi,t An indicator variable that equals one if ABNRET < 0, and equals zero otherwise.

DCFi,t An indicator variable that equals one if CF < 0, and equals zero otherwise.

DefaultRiski,l A market-based measure of firm i's probability of default as of the end of the month immediately preceding the inception of loan l. Based on the Black-Scholes-Merton model.

37

IndChangei,t An indicator that equals one if firm i changed four digit sic industries from year t-1 to year t, and equals zero otherwise.

LUAAi,t An indicator that equals one if firm i has a "large" unsigned abnormal accrual in year t, and equals zero otherwise, where “large” is defined as an abnormal accrual with magnitude greater than 5% of total assets, which is roughly the median UAA (UAA > 0.05)

LUAA_NLi,t An indicator that equals one if firm i has a "large" unsigned abnormal accrual (based on the nonlinear accrual model) in year t, and equals zero otherwise, where “large” is defined as an abnormal accrual with magnitude greater than 5% of total assets (UAA_NL > 0.05)

LUAA_PFi,t An indicator that equals one if firm i has a "large" unsigned abnormal accrual (based on the ROA performance control accrual model) in year t, and equals zero otherwise, where “large” is defined as an abnormal accrual with magnitude greater than 5% of total assets (UAA_PF > 0.05)

LargeDiscOpsi,t An indicator that equals one if firm i has discontinued operations (Compustat do) greater than five percent of sales in year t, and equals zero otherwise.

LargeMergAcqi,t An indicator that equals one if firm i in year t engaged in a large merger/acquisition (as indicated by Compustat sales footnote code "AB"), and equals zero otherwise.

LargeRestruci,t An indicator that equals one if firm i has restructuring charges (Compustat rcp) greater than five percent of sales in year t, and equals zero otherwise.

LargeSpecItemi,t An indicator that equals one if firm i has special items (Compustat spi) greater than five percent of sales in year t, and equals zero otherwise.

Leveragei,t Firm i's quarter t leverage, measured as total liabilities (Compustat lt) divided by total assets (Compustat at).

LnAssetsi,t The natural log of firm i's quarter t ending total assets (Compustat at).

LnFacilityi,l The natural log of the face amount of firm i's loan facility l (Dealscan facilityamt), in millions of U.S. dollars.

LnMaturityi,l The natural log of the maturity in months of firm i's loan facility l (Dealscan maturity).

LnSizei,t The natural log of firm i's quarter t ending market value of common equity (Compustat prcc_c*csho).

38

MaxMUARi,t The maximum monthly abnormal return experienced by firm i during year t, where abnormal return is measured as CRSP firm return less the return on the CRSP value weighted index.

MktToBooki,t Firm i's market-to-book ratio at the end of quarter t (Compustat [prcc_c*csho]/ceq)

OpCycVolUPi,t An indicator that equals one if firm i's four-quarter operating cycle volatility (i.e., std. deviation) increased from year t-1 to t. We measure operating cycle as "days sales outstanding" (90/(saleq/avg_rectq)) + "days inventory outstanding" (90/(cogsq/avg_invtq)) - "days payables outstanding" (90/(cogsq/avg_apq)).

OpShocki,t An indicator variable that equals one if firm i in year t experiences at least one of five specifically identified operational shocks (from financial statements), and equals zero otherwise; specifically, this variable equals one if at least one of the following variables equals one: IndChange, LargeDiscOps, LargeMergAcq, LargeRestruc, LargeSpecItem.

OtherMaxMUARi,t The average maximum monthly abnormal return experienced by all other firms in firm i's industry in year t, where abnormal return is measured as CRSP firm return less the return on the CRSP value weighted index.

PPEi,t Firm i's quarter t gross property, plant and equipment balance, scaled by beginning-of-period total assets.

RetShocki,t An indicator variable that equals one if firm i in year t experiences a large monthly abnormal stock return, defined as MaxMUAR > 20%.

ROAi,t Firm i's year t return on assets, computed as net income divided by average total assets.

SalesChangei,t Firm i's change in sales from year t-1 to t, scaled by year t-1 total assets.

Securedi,l An indicator variable that equals one if loan facility l requires collateral, and equals zero otherwise (DL-secured).

Spreadi,l interest rate on loan facility l in excess of LIBOR, in basis points (DL-allindrawn)

Tangibility,i,t Asset tangibility, measured as property, plant and equipment, scaled by total assets (Compustat ppent/at).

TotalAccrualsi,t Firm i's year t total accruals, taken from the statement of cash flows.

UAAi,t Firm i's year t unsigned abnormal accrual, calculated as |AbAccrual|.

39

UAA_NLi,t Firm i's year t unsigned abnormal accrual, calculated as |AbAccrual_BS|.

UAA_PFi,t Firm i's year t unsigned abnormal accrual, calculated as |AbAccrual_PF|.

40

REFERENCES

Ball, R. 2013, Accounting Informs Investors and Earnings Management is Rife: Two Questionable

Beliefs, forthcoming Accounting Horizons. Ball, R., Shivakumar, L., 2006. The role of accruals in asymmetrically timely gain and loss recognition.

Journal of Accounting Research 44 (2), 207-242. Ball, R., Shivakumar, L., 2008. Earnings quality at initial public offerings. Journal of Accounting and

Economics 45, 324–349. Barberis, N., A. Shleifer, R. Vishny, 1998. A model of investor sentiment, Journal of Financial

Economics 49, 307-343. Berry, S., P. Reiss, 2007. Empirical models of entry and market structure, M. Armstrong and R. Porter

eds. Handbook of Industrial Organization 3 (Elsevier) 1848-1886. Bernard, V., Skinner, D., 1996. What motivates managers' choice of discretionary accruals? Journal of

Accounting and Economics 22, 313 325. Bharath, S., Sunder, J., Sunder, S., 2008. Accounting quality and debt contracting. The Accounting

Review 83 (1), 1-28. Bresnahan, T., Greenstein, S., Henderson, R., 2011. Schumpeterian competition and diseconomies of

scope; illustrations from the histories of Microsoft and IBM. Harvard Business School working paper 11-077.

Brickley, J., Smith, C., Zimmerman, J., 2009. Managerial Economics and Organizational Architecture 5th

edition (Boston: McGraw-Hill). Brickley, J. and Zimmerman, J., 2010. Corporate governance myths: Comments on Armstrong, Guay, and

Weber, Journal of Accounting and Economics 50, 235–245. Brown, G. and Kapadia, N., 2007. Firm-specific risk and equity market development, Journal of Financial

Economics 84, 358–388. Chernow, R., 2007. Titan: The Life of John D. Rockefeller, Sr., (Knopf Doubleday Publishing Group.

Kindle Edition.) Cohen, D., Dey, A., Lys, T., 2008. Real and accrual-based earnings management in the pre- and post-

Sarbanes-Oxley periods, The Accounting Review 83 (3), 757-787. Collins, D., Hribar, P., 2002. Errors in estimating accruals: Implications for empirical research. Journal of

Accounting Research 40, 105-134. Collins, D., R. S. Pungaliya, and A. Vijh. 2012. The effects of firm growth and model specification

choices on tests of earnings management in quarterly settings. Working paper, University of Iowa. Comin, D. and Mulani, S., 2006. Diverging Trends in Aggregate and Firm Volatility, The Review of

Economics and Statistics 88, 383–388.

41

Corrado, C. and Hulten, C., 2010. How Do You Measure a “Technological Revolution?” American

Economic Review: Papers & Proceedings 100, 99–104. Daniel D., Pungaliya R., and Vijh, A., 2012. The effects of firm growth and model specification choices

on tests of earnings management. Working paper. University of Iowa. Dechow, P. and Dichev, I., 2002. The quality of accruals and earnings: the role of accrual estimation

errors. The Accounting Review 77, 35–59. Dechow,P., W. Ge, and C. Schrand, 2010. Understanding earnings quality: A review of the proxies, their

determinants and their consequences. Journal of Accounting and Economics 50, 344-401. Dechow, P., Hutton, A., Kim, J., Sloan, R., 2012. Detecting earnings management: a new approach.

Journal of Accounting Research, Vol. 50 (2), 275-334. Dechow, P., S.P. Kothari, R. Watts, 1998. The relation between earnings and cash flows. Journal of

accounting and Economics 25, 133-168. Dechow, P., Sloan, R., Sweeney, A., 1995. Detecting earnings management. The Accounting Review 70,

193–225. Dichev, I., Graham, J., Harvey, C., and Rajgopal, S., 2012. Earnings Quality: Evidence from the Field,

forthcoming Journal of Accounting and Economics.

Dichev, I. and Tang, V., 2008. Matching and the changing properties of accounting earnings over last 40 years, Accounting Review 83, 1425-60.

Donelson. D., Jennings, R., and McInnis, J., 2011. Changes over time in the revenue-expense relation:

Accounting or Economics, Accounting Review 86, 945-974. Dopuch, N., Mashruwala, R., Seethamraju, C., and Zach, T., 2011. The impact of a heterogeneous

accrual-generating process on empirical accrual models. Journal of Accounting, Auditing & Finance 1, 1–26.

Ecker, F., Francis, J., Olsson, P., and Schipper, S, 2011. Peer firm selection for discretionary accruals

models. Working paper. Duke University. Fama, E. and French, K., 2004. New lists: fundamentals and survival rates, Journal of Financial

Economics 73, 229-269. Fields, T., Lys, T., Vincent, L., 2001. Empirical research on accounting choice, Journal of Accounting

and Economics 31, 255–307. Gerakos J. and A. Kovrijnykh, 2012. Performance shocks and misreporting. Working paper, University of

Chicago. Guay, W., Kothari, S., and Watts, R., 1996. A market-based evaluation of discretionary accruals models,

Journal of Accounting Research 34, 83-105. Healy, P., 1985. The effect of bonus schemes on accounting decisions. Journal of Accounting and

Economics 7, 85–107.

42

Henderson, D., 1977. The economics of federal coal mine safety regulation: Safety or monopoly?" unpublished working paper, University of Rochester.

Hillegeist, S., Keating, E., Cram, D., Lundstedt, K., 2004. Assessing the probability of bankruptcy.

Review of Accounting Studies 9, 5-34. Holthausen, R., Larcker, D., Sloan, R., 1995. Annual bonus schemes and the manipulation of earnings

Journal of Accounting and Economics 19, 29 74 Hribar, P., Nichols, C., 2007. The use of unsigned earnings quality measures in tests of earnings

management. Journal of Accounting Research 45, 1017-1053. Hutton, A., A. Marcus, H. Tehranian, 2009. Opaque financial reports, R2, and crash risk. Journal of

Financial Economics 94, 67–86. Irvine, P., and Pontiff, J., 2005. Idiosyncratic return volatility, cash flows and product market

Competition, Review of Financial Studies 22, 1149–1177. Jones, C., and Romer, P., 2010. The new Kaldor facts: Ideas, institutions, population, and human capital,

American Economic Journal: Macroeconomics 2, 224-245. Jones, J., 1991. Earnings management during import relief investigations. Journal of Accounting

Research 29, 193-228. Kothari, S., 2001. Capital markets research in accounting. Journal of Accounting and Economics 31, 105

–231. Kothari, S., Leone, A., Wasley, C., 2005. Performance matched discretionary accrual measures. Journal

of Accounting and Economics 39, 163–197. Liu, L., Zhang, 2008. Momentum Profits, Factor Pricing, and Macroeconomic Risk, Review of Financial

Studies 21, 2417-2448. McNichols, 2002. Discussion of “The quality of accruals and earnings: the role of accrual estimation

errors. The Accounting Review 77, 61-69. Milgrom, P., Roberts, J., 1995. Complementarities and fit strategy, structure, and organizational change in

manufacturing, Journal of Accounting and Economics 19, 179-208.

Mitchell, M., Lehn, K., 1990. “Do Bad Bidders Become Good Targets?” Journal of Political Economy 98, 372-398.

Porter, M., 1979. How competitive forces shape strategy. Harvard Business Review 57 (2), 137-156. Prahalad, C., Hamel, G., 1990. The core competence of the corporation. Harvard Business Review 68, 79-

91. PwC 15th Annual Global CEO Survey 2012, http://www.pwc.com/us/en/issues/ceo-agenda/assets/15th-

annual-global-ceo-survey-us-findings.pdf

43

Ramanna, K., Watts, R., 2012. “Evidence on the use of unverifiable estimates in required goodwill impairment,” Review of Accounting Studies 17, 749–780.

Savor, P., 2012. Stock returns after major price shocks: The impact of information, Journal of Financial

Economics 106, 635–659. Schmalensee, R., 1989. Inter-Industry Studies of Structure and Performance, R. Schmalensee and R.

Willig eds. Handbook of Industrial Organization 2 (Elsevier) 951- 1099. Schumpeter, J., 1942. Capitalism, Socialism and Democracy. Harper, New York. Subramanyam, K., 1996. The pricing of discretionary accruals. Journal of Accounting and Economics 22,

249–281. Sutton, J., 2007. Market Structure: Theory and Evidence, M. Armstrong and R. Porter eds. Handbook of

Industrial Organization 3 (Elsevier) 2301-2367. Teece, T., 2010. Business models, business strategy and innovation. Long Range Planning 43, 172-194. Journal of Accounting and Economics 21 (1996) 279-295

Teets, W. and Wasley, C., 1996. Estimating earnings response coefficients: Pooled versus firm-specific models, Journal of Accounting and Economics, 21, 279-295.

Vorhies, J., 2005. The new importance of materiality. Journal of Accountancy (May). Wysocki, P. 2008. Assessing earnings and accruals quality: U.S. and international evidence. Working

paper, MIT Sloan School of Management. Zhang, Y. and Zhuang, Z., 2012. Modeling fundamental performance to improve the estimation of

discretionary accruals. Working paper, The Chinese University of Hong Kong.

44

‐60

‐50

‐40

‐30

‐20

‐10

0

10

20

30

1985 1990 1995 2000 2005 2010 2015

Airlines’ Operating Cycle (Days)

Alaska

US Air

American

Delta

Southwest

United

Frontier

American

Delta

Figure 1 Variance of Operating Cycle in the Airline Industry Figure 1 presents a time-series plot of the time-series and cross-sectional variation in the operating cycles of several major airline companies.

45

0

5000

10000

15000

20000

25000

Frequency

Absolute Percentage

MaxMUAR

Figure 2 Frequency histogram of large stock return shocks Figure 2 presents a histogram of sample realizations (using the full sample of 125,956 observations) of the magnitude of the largest monthly abnormal stock return in each firm-year (MaxMUAR).

46

0

20

40

60

80

100

120

140

160

0.000

0.025

0.050

0.075

0.100

0.125

0.150

0.175

0.200

0.225

0.250

0.275

0.300

0.325

0.350

0.375

0.400

0.425

0.450

0.475

0.500

0.525

0.550

0.575

0.600

0.625

0.650

0.675

0.700

More

Frequency

R2

Industry‐year Jones Model R2 Frequency

0

5

10

15

20

25

30

35

40

45

Frequency

R2

Industry‐year Nonlinear model R2 Frequency

0

10

20

30

40

50

60

70

Frequency

R2

Industry‐year Model with ROA Performance Control R2 Frequency

Figure 3 Histogram of R2 from estimation of the cross-sectional Jones model by industry-year Figure 3 presents frequency histograms of the estimated R2s from the 878 industry-year estimations within our sample.

47

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

% Firms with Shocks/Large Abnormal Accruals

% OpShock

% RetShock

% LUAA

% LUAA_NL

% LUAA_PF

Figure 4 Time-series plots of the proportion of sample firms in each year that experience shocks and large unsigned abnormal accruals Figure 4 presents time-series plots of the proportion of the "full sample" firms in each year (i.e., 125,956 observations) that experience large operational shocks (i.e., OpShock = 1), large return shocks (i.e., RetShock = 1), large unsigned abnormal accruals from the original Jones model (i.e., LUAA = 1), large unsigned abnormal accruals from the nonlinear Jones model (i.e., LUAA_NL = 1), and large unsigned abnormal accruals from a Jones model that includes ROA (i.e., LUAA_PF = 1). All variables are further defined in Appendix B.

48

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.00

0.10

0.20

0.30

0.40

0.50

0.60

1 2 3 4 5 6 7 8 9

Model Number (see Panel A)

All Observations ‐ Nonlinear model

PcntUAA

AvgUAA

0.000

0.005

0.010

0.015

0.020

0.025

0.030

0.035

0.040

0.045

0.050

0.00

0.05

0.10

0.15

0.20

0.25

0.30

1 2 3 4 5 6 7 8 9


Initial Uncontaminated Obs ‐ Nonlinear model

PcntUAA

AvgUAA

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.00

0.10

0.20

0.30

0.40

0.50

0.60

1 2 3 4 5 6 7 8 9


All Observations ‐model with ROA performance control

PcntUAA

AvgUAA

0.000

0.005

0.010

0.015

0.020

0.025

0.030

0.035

0.040

0.045

0.050

0.00

0.05

0.10

0.15

0.20

0.25

0.30

1 2 3 4 5 6 7 8 9


Initial Uncontaminated Obs ‐model with ROA performance control

PcntUAA

AvgUAA

Figure 5 Graphical presentation of Table 5, Panel B results

49

0.00

0.02

0.04

0.06

0.08

0.10

0.12

0.14

0.16

0.18

0.20

1 2 3 4 5 6 7 8 9


Original Jones

Mean IY R‐sq.

Median IY R‐sq

0.44

0.46

0.48

0.50

0.52

0.54

0.56

0.58

0.60

0.62

0.64

1 2 3 4 5 6 7 8 9


Nonlinear model

Mean IY R‐sq

Median IY R‐sq

0.20

0.22

0.24

0.26

0.28

0.30

0.32

0.34

0.36

0.38

0.40

1 2 3 4 5 6 7 8 9


Model with ROA performance control

Mean IY R‐sq

Median IY R‐sq

Figure 5, continued

50

0.00%

10.00%

20.00%

30.00%

40.00%

50.00%

60.00%

70.00%

80.00%

90.00%

100.00%

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Rejection Rate Frequency

Correlation between partition variable and MaxMUAR

NoControls

Controls

0.00%

10.00%

20.00%

30.00%

40.00%

50.00%

60.00%

70.00%

80.00%

90.00%

100.00%

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Rejection Rate Frequency

Correlation between partition variable and OtherMaxMUAR

NoControls

Controls

Figure 6 Rejection rate frequencies (Predictions 5 and 6) Fig. 6 presents the frequency of rejecting the null hypothesis of no relation between absolute discretionary Jones model accruals and a partitioning variable that is correlated with business model shocks. In Panel A, the correlation is with a firm's own shock during year t (MaxMUAR). In Panel B, the correlation is with shocks to other industry firms during year t (OtherMaxMUAR). The "diamond" lines are from univariate tests at the 5% level, and the "square" lines are from multivariate tests with the following controls: LnSize, MktToBook, Leverage, σCFO, and σRev. Results are based on 250 trials using a sample size of 1,000 randomly drawn observations. Panel A: Own-firm shocks (MaxMUAR) (Prediction 5)

Panel B: Other firm shocks in the industry (OtherMaxMUAR) (Prediction 6)

51

Table 1 Descriptive statistics Table 1 presents descriptive statistics for key variables used in our Jones model analyses. |AbAccrual| is the unsigned abnormal accrual from estimating the original Jones model by industry-year (i.e., the absolute value of the residual from Eq. 4). |AbAccrual_NL| is the unsigned abnormal accrual from estimating the nonlinear accrual model by industry-year (i.e., the absolute value of the residual from Eq. 5). |AbAccrual_PF| is the unsigned abnormal accrual from estimating the Jones model with inclusion of ROA by industry-year (i.e., the absolute value of the residual from Eq. 6). All variables are further defined in Appendix B.

N Mean Std P1 P25 Median P75 P99

TotalAccruals 90,822 -0.070 0.141 -0.581 -0.112 -0.054 -0.010 0.291

SalesChange 90,822 0.108 0.349 -0.724 -0.022 0.057 0.197 1.357

PPE 90,822 0.573 0.516 0.000 0.211 0.453 0.825 2.079

ROA 90,822 -0.036 0.239 -1.059 -0.050 0.029 0.074 0.283

CF 90,822 0.036 0.189 -0.753 -0.001 0.069 0.129 0.345

ABNRET 90,822 0.008 0.688 -0.980 -0.389 -0.102 0.214 3.025

MktToBook 90,615 2.780 4.081 -8.317 1.055 1.811 3.225 26.204

Leverage 90,664 0.507 0.260 0.044 0.307 0.502 0.673 1.342

LnSize 90,657 5.261 2.320 0.377 3.553 5.176 6.881 10.780

σCFO 65,050 0.076 0.076 0.006 0.028 0.050 0.091 0.376

σREV 65,982 0.174 0.158 0.006 0.065 0.124 0.227 0.733

|AbAccrual| 90,822 0.080 0.102 0.001 0.021 0.049 0.099 0.497

|AbAccrual_PF| 90,822 0.070 0.084 0.001 0.019 0.044 0.089 0.399

|AbAccrual_NL| 90,822 0.068 0.091 0.001 0.017 0.040 0.083 0.451

52

Table 2 Correlation matrix Table 2 presents Pearson (Spearman) correlations above (below) the diagonal among key variables we use in our analyses. Correlations significant at the p < 0.05 level are reported in bold. Variables are defined in Appendix B. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14)

TotalAccruals (1) 0.141 -0.138 -0.049 0.080 -0.036 -0.127 0.085 -0.155 -0.048 -0.375 -0.190 -0.378 -0.116

SalesChange (2) 0.161 0.045 0.091 0.205 0.111 -0.036 0.088 -0.003 0.133 0.076 0.105 0.050 0.021

PPE (3) -0.186 0.081 0.200 0.045 -0.038 0.090 0.086 -0.153 -0.114 -0.076 -0.084 -0.078 -0.097

CF (4) -0.303 0.189 0.281 0.113 -0.088 -0.050 0.293 -0.508 -0.112 -0.285 -0.294 -0.307 -0.231

ABNRET (5) 0.118 0.264 0.079 0.239 0.219 -0.071 0.142 0.003 0.016 0.017 0.057 -0.005 0.061

MktToBook (6) 0.020 0.263 0.009 0.160 0.304 -0.069 0.166 0.171 0.041 0.092 0.133 0.118 0.034

Leverage (7) -0.085 -0.054 0.126 -0.085 -0.053 -0.102 0.046 -0.083 0.043 0.048 -0.012 0.011 -0.035

LnSize (8) 0.056 0.145 0.120 0.358 0.246 0.380 0.073 -0.364 -0.308 -0.217 -0.202 -0.205 -0.395

σCFO (9) -0.085 0.000 -0.203 -0.250 -0.096 0.061 -0.205 -0.452 0.375 0.412 0.468 0.402 0.317

σREV (10) -0.044 0.120 -0.090 -0.094 -0.054 0.015 -0.029 -0.331 0.483 0.226 0.217 0.207 0.243

|AbAccrual| (11) 0.010 0.025 -0.135 -0.241 -0.061 0.035 -0.049 -0.249 0.373 0.234 0.795 0.855 0.221

|AbAccrual_PF| (12) -0.065 0.060 -0.133 -0.154 -0.032 0.080 -0.081 -0.241 0.425 0.235 0.678 0.691 0.216

|AbAccrual_NL| (13) -0.047 0.004 -0.137 -0.221 -0.126 0.070 -0.104 -0.242 0.369 0.228 0.633 0.480 0.232

LUAR -0.113 -0.019 -0.124 -0.246 -0.073 -0.054 -0.067 -0.391 0.395 0.285 0.245 0.244 0.270

53

Table 3 Operational shocks and large unsigned abnormal returns Table 3 presents results from estimating the probit specification of Eq (3). OpShocki,t is an indicator that equals one if firm i has an observed operational shock in year t, and equals zero otherwise. RetShocki,t is an indicator that equals one if firm i has a large monthly stock return shock in year t, and equals zero otherwise. All variables are further defined in Appendix B. Robust t-statistics clustered by both firm and year are reported in parentheses. *, **, and *** indicate significance (two-sided) at the 10%, 5% and 1% levels, respectively. Dep. Var.: LargeMergAcq IndChange LargeDiscOps LargeRestruc LargeSpecItem OpShock Column: (1) (2) (3) (4) (5) (6) Intercept -7.520*** -3.881*** -4.122*** -6.382*** -3.124*** -2.554*** (-29.968) (-35.992) (-26.029) (-23.499) (-48.554) (-49.583) RetShockt 0.561** 0.562*** 0.488*** 1.238*** 0.977*** 0.837*** (2.102) (6.872) (5.683) (6.749) (17.729) (18.557) RetShockt-1 0.435** 0.239*** 0.263*** 0.980*** 0.583*** 0.485*** (2.281) (3.176) (3.548) (3.330) (7.408) (9.416) RetShockt-2 -0.025 0.089 0.200*** 0.452** 0.213*** 0.188*** (-0.127) (0.942) (2.849) (2.130) (6.179) (7.283) N 125,964 125,964 125,964 125,964 125,964 125,964 Pseudo-R2 0.010 0.014 0.014 0.066 0.062 0.051

54

Table 4 Business model shocks and large abnormal accrual estimates (Predictions 1 and 2) Table 4 presents results from estimating Eqs. (7), (8), and (9). LUAA is an indicator that equals one if firm i's year t unsigned original Jones model abnormal accrual is greater than 5% of total assets, and equals zero otherwise. LUAA_NL is an indicator that equals one if firm i's year t unsigned nonlinear Jones model abnormal accrual is greater than 5% of total assets, and equals zero otherwise. LUAA_PF is an indicator that equals one if firm i's year t performance Jones model abnormal accrual is greater than 5% of total assets, and equals zero otherwise. OpShocki,t is an indicator that equals one if firm i has an observed operational shock in year t, and equals zero otherwise. RetShocki,t is an indicator that equals one if firm i has a large monthly stock return shock in year t, and equals zero otherwise. MaxMUAR is firm i's maximum monthly unsigned absolute return during year t. OtherMaxMUA is the average MaxMUAR across all other firms in firm i's industry during year t. All variables are further defined in Appendix B. R2 refers to adjusted- R2 (pseudo- R2) in columns (1)-(3) (columns 4-6). Robust t-statistics clustered by both firm and year are reported in parentheses. *, **, and *** indicate significance (two-sided) at the 10%, 5% and 1% levels, respectively. Panel A: Operational Shocks reflected in Financial Statements (Prediction 1) Model: OLS OLS OLS Logit Logit Logit Dep. Var.: UAA UAA_NL UAA_PF LUAA LUAA_NL LUAA_PF Column: (1) (2) (3) (4) (5) (6) Intercept 0.025*** 0.019*** 0.023*** -0.985*** -1.376*** -1.258*** (19.248) (19.413) (28.768) (-18.168) (-29.866) (-32.802) OpShockt 0.046*** 0.045*** 0.026*** 0.725*** 0.849*** 0.606*** (30.481) (34.962) (25.684) (41.616) (30.514) (21.242) OpShockt-1 0.003** 0.005*** 0.001 0.071*** 0.113*** 0.033 (2.073) (3.544) (1.466) (2.920) (3.166) (1.568) OpShockt-2 0.001 0.003** -0.002 0.027 0.057** -0.014 (0.641) (2.358) (-1.399) (1.040) (2.073) (-0.553) σCFO 0.435*** 0.386*** 0.449*** 6.859*** 7.376*** 9.311*** (24.818) (21.966) (33.528) (21.733) (32.863) (30.591) σREV 0.048*** 0.034*** 0.024*** 1.350*** 1.205*** 1.037*** (10.330) (8.039) (5.092) (13.001) (15.109) (10.830) N 64,193 64,193 64,193 64,193 64,193 64,193 R2 0.212 0.209 0.239 0.076 0.092 0.095

55

Table 4, continued Panel B: Market-based shocks (Prediction 1) Model: OLS OLS OLS Logit Logit Logit Dep. Var.: UAA UAA_NL UAA_PF LUAA LUAA_NL LUAA_PF Column: (1) (2) (3) (4) (5) (6) Intercept 0.021*** 0.015*** 0.019*** -1.189*** -1.657*** -1.488*** (13.708) (11.486) (23.096) (-18.769) (-30.443) (-38.312) RetShockt 0.015*** 0.016*** 0.009*** 0.395*** 0.556*** 0.382*** (10.185) (10.664) (15.797) (11.316) (11.988) (18.327) RetShockt-1 0.006*** 0.006*** 0.004*** 0.208*** 0.216*** 0.217*** (6.326) (6.813) (7.170) (5.938) (5.797) (7.963) RetShockt-2 0.004*** 0.005*** 0.003*** 0.182*** 0.215*** 0.184*** (2.790) (3.806) (3.007) (5.015) (5.548) (6.095) σCFO 0.438*** 0.387*** 0.441*** 5.823*** 6.211*** 8.035*** (23.333) (20.903) (31.255) (19.192) (26.108) (26.829) σREV 0.040*** 0.025*** 0.017*** 1.019*** 0.829*** 0.719*** (7.517) (5.177) (3.565) (10.481) (10.938) (8.173) N 64,599 64,599 64,599 64,599 64,599 64,599 R2 0.184 0.178 0.227 0.075 0.093 0.098

56

Table 4, continued Panel C: Shocks of other industry firms (Prediction 2) Dep Var: UAA UAA_NL UAA_PF Column: (1) (2) (3) Intercept 0.003 -0.006 0.008*** (0.959) (-1.641) (2.696) MaxMUARt 0.021*** 0.022*** 0.013*** (7.292) (7.440) (5.077) MaxMUARt-1 0.008*** 0.010*** 0.005*** (4.422) (5.134) (3.003) MaxMUARt-2 0.007*** 0.008*** 0.001 (2.864) (5.168) (0.848) OtherMaxMUARt 0.075*** 0.069*** 0.046*** (5.761) (5.965) (7.478) OtherMaxMUARt-1 -0.013 -0.005 -0.002 (-0.616) (-0.258) (-0.161) OtherMaxMUARt-2 0.003 0.014 0.003 (0.139) (0.761) (0.188) σCFO 0.417*** 0.361*** 0.434*** (23.899) (19.468) (30.649) σREV 0.040*** 0.024*** 0.018*** (7.222) (5.116) (3.744) N 64,599 64,599 64,599 Adj.-R2 0.193 0.189 0.232

57

Table 5 Business model shock sample contamination and abnormal accruals (Prediction 3) Table 5 presents results from iterative estimation of accrual models to show the effect of inclusion of "contaminated" observations in the cross-sectional industry-year estimation on the incidence of large unsigned absolute accruals and industry-year estimation R2s. We begin by estimating the accrual models with only uncontaminated observations (i.e., observations with RetShockt = 0 and RetShockt-1 = 0 and RetShockt-2 = 0), then successively add 12.5% of the contaminated sample observations (i.e., observations with RetShockt = 1 or RetShockt-

1 = 1 or RetShockt-2 = 1) and re-estimate the models. Panel A presents the sample characteristics of each estimation iteration. Panel B reports the median industry-year R2 from each estimation, mean unsigned abnormal accruals (UAA), and the percentage of sample observations with a large unsigned abnormal accrual (LUAA = 1). All variables are defined in Appendix B. Panel A: Sample compositions across contamination iterations

Number of Observations

% of Contaminated Uncontaminated Contaminated

Iteration Obs. Included (1) (2)

1 00.0% 10,844 0

2 12.5% 12,722 6,447

3 25.0% 13,937 14,426

4 37.5% 14,867 23,833

5 50.0% 15,502 33,634

6 62.5% 15,938 43,659

7 75.0% 16,400 53,753

8 87.5% 16,701 63,831

9 100.0% 16,917 73,905

58

Table 5, continued Panel B: Estimation results from sample iterations (Prediction 3)

All Observations

Initial 10,844 Uncontaminated Obs.

Column: (1) (2) (3) (4) (5)

Panel B1: Jones Model

Iteration % LUAA Avg. UAA % LUAA Avg. UAA Med. I-Y R2

1 0.1936 0.0331 0.1936 0.0331 0.1403

2 0.3413 0.0562 0.2278 0.0364 0.1152

3 0.3968 0.0647 0.2378 0.0373 0.1049

4 0.4340 0.0705 0.2464 0.0380 0.0991

5 0.4567 0.0740 0.2542 0.0386 0.0940

6 0.4706 0.0762 0.2585 0.0389 0.0948

7 0.4810 0.0783 0.2635 0.0394 0.0913

8 0.4873 0.0794 0.2646 0.0396 0.0853

9 0.4927 0.0805 0.2678 0.0399 0.0836

Panel B2: Nonlinear Accrual Model

Iteration % LUAA_NL Avg. UAA_NL % LUAA_NL Avg. UAA_NL Med. I-Y R2

1 0.1117 0.0233 0.1117 0.0233 0.6016

2 0.2659 0.0434 0.1629 0.0285 0.5583

3 0.3291 0.0535 0.1718 0.0296 0.5274

4 0.3668 0.0591 0.1810 0.0304 0.4844

5 0.3853 0.0625 0.1848 0.0311 0.4731

6 0.3996 0.0646 0.1904 0.0313 0.4680

7 0.4109 0.0665 0.1968 0.0316 0.4607

8 0.4154 0.0675 0.1969 0.0318 0.4630

9 0.4201 0.0684 0.1942 0.0318 0.4545

Panel B3: Model with ROA Performance Control

Iteration % LUAA_PF Avg. UAA_PF % LUAA_PF Avg. UAA_PF Med. I-Y R2

1 0.1872 0.0320 0.1872 0.0320 0.2466

2 0.3211 0.0498 0.2086 0.0342 0.3095

3 0.3714 0.0574 0.2124 0.0349 0.3406

4 0.4024 0.0617 0.2124 0.0350 0.3371

5 0.4201 0.0644 0.2151 0.0353 0.3334

6 0.4318 0.0663 0.2193 0.0355 0.3380

7 0.4413 0.0678 0.2200 0.0356 0.3425

8 0.4464 0.0689 0.2190 0.0357 0.3497

9 0.4515 0.0697 0.2195 0.0358 0.3464

59

Table 6 Business model shocks and increases in operating cycle volatility (Prediction 4) Table 6 presents results from estimating Eq. (10). OpCycVolUPt is an indicator that equals one if firm i's 4-quarter operating cycle volatility increased from year t-1 to t, and equals zero otherwise. To calculate operating cycle volatility in year t, we first calculate the operating cycle for each quarter in year t as "days sales outstanding" + "days inventory outstanding" - "days payables outstanding." We then take the standard deviation of these four quarterly operating cycles as year t's operating cycle volatility. If this standard deviation increases from year t-1 to t, the dependent variable = 1. OpShocki,t is an indicator that equals one if firm i has an observed operational shock in year t, and equals zero otherwise. RetShocki,t is an indicator that equals one if firm i has a large monthly stock return shock in year t, and equals zero otherwise. All variables are further defined in the Appendix. Robust t-statistics clustered by both firm and year are reported in parentheses. *, **, and *** indicate significance (two-sided) at the 10%, 5% and 1% levels, respectively.

Dep. Var.: OpCycVolUPt OpCycVolUPt Column: (1) (2) Intercept -0.070** -0.059 (-1.97) (-1.42) OpShockt 0.230*** (6.26) RetShockt 0.047* (1.66) N 47,056 47,056 R2 0.001 0.0001

60

Table 7 Unsigned abnormal accruals and cost of debt Table 7 presents results from estimating Eqs. (11) and (12). Spread is the loan interest rate in excess of LIBOR. UAA is the absolute value of firm i's year t abnormal accrual, estimated as the residual from the estimation of the cross-sectional version of the original Jones model by industry-year. ContamRet is an indicator that equals 1 if firm i has a return shock (RetShock) in either year t, t-1, or t-2, and equals 0 otherwise. ContamOS an indicator that equals 1 if firm i has an observed operational shock (OpShock) in either year t, t-1, or t-2, and equals 0 otherwise. All variables are further defined in Appendix B. Robust t-statistics clustered by firm are reported in parentheses. *, **, and *** indicate significance (two-sided) at the 10%, 5% and 1% levels, respectively. Dep. Var.: Spread Spread Spread Contam: ContamRet ContamOS Column: (1) (2) (3) Intercept 497.049*** 452.044*** 484.130*** (28.707) (26.098) (28.463) Contam 38.601*** 15.454*** (11.626) (5.021) UAA 80.577*** 25.176 30.486 (4.834) (0.660) (1.061) Contam*UAA 48.069 45.391 (1.151) (1.334) Leverage 84.024*** 78.758*** 80.769*** (11.295) (10.796) (10.919) LnAssets -8.023*** -5.716*** -7.972*** (-6.334) (-4.692) (-6.461) Tangibility -11.998* -9.751 -11.635* (-1.930) (-1.616) (-1.909) CurrRatio -2.706** -3.661*** -2.823** (-2.251) (-3.083) (-2.372) MktToBook -1.887*** -1.775*** -1.924*** (-5.499) (-5.266) (-5.674) DefaultRisk 167.856*** 162.351*** 164.284*** (12.877) (12.709) (12.782) LnFacility -20.067*** -19.828*** -19.600*** (-16.872) (-17.293) (-16.818) LnMaturity 5.783*** 6.632*** 5.941*** (3.927) (4.563) (4.100) Secured 89.018*** 84.172*** 88.322*** (34.296) (32.949) (34.186) HO: UAA + Contam*UAA = 0 73.245*** 75.877*** χ2 Statistic 17.69 15.04 Prob. > χ2 0.0000 0.0001

N 23,738 23,738 23,738 Adj.-R2 0.423 0.435 0.427

61

Table 8 Specification tests of signed abnormal accruals Table 8 reports the proportion of 250 random samples of 200 firms each where the null hypothesis of zero abnormal accruals is rejected at the 5% level using a one-tailed t-test (the assocated actual numbers of random trials out of the 250 where the null is rejected is presented in brackets). The samples are drawn at random within each sample quartile based on several different variables: Market-to-book, Size, and Sales Growth. Figures in bold signify rejection rates that are significantly different from the 5% significance level of the test (two-tailed), where proportions above (below) 0.05 indicate that tests are biased against (in favor) of accepting the null hypothesis of zero abnormal accruals. The first five columns ('Full Sample') use the full sample in the Jones model estimation and random sampling pools. The second five columns ('Subsample') includes only those firm-years where LUARt = 0 (i.e., only retains those firm-years without a contemporaneous shock) from both the initial Jones model estimation and the random sampling pool. * indicates that the proportion in the no-shock subsample analysis is different from the corresponding proportion in the full sample analysis at the 5% level based on a two-sample Wald test.

Column (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)

Full Sample Subsample With No Year t Shocks

Sample Quartile All 1 2 3 4 All 1 2 3 4

N 90,822 22,705 22,705 22,705 22,705 Sum 1-4 33,027 8,256 8,256 8,256 8,256 Sum 1-4

Panel A: Market-to-Book

Ha : AbAccrual > 0 0.060 0.008 0.420 0.468 0.008 0.068 0.124* 0.132* 0.048* 0.008

[2] [105] [117] [2] [226] [31] [33] [12] [2] [78]

Ha : AbAccrual < 0 0.044 0.336 0.000 0.000 0.168 0.036 0.020* 0.004 0.032* 0.208

[84] [0] [0] [42] [126] [5] [1] [8] [52] [66]

Panel B: Size

Ha : AbAccrual > 0 0.060 0.000 0.088 0.320 0.564 0.068 0.000 0.004* 0.036* 0.040*

[0] [22] [80] [141] [243] [0] [1] [9] [10] [20]

Ha : AbAccrual < 0 0.044 0.468 0.024 0.004 0.000 0.036 0.024* 0.072* 0.092* 0.040*

[117] [6] [1] [0] [124] [6] [18] [23] [10] [57]

Panel C: Sales Growth

Ha : AbAccrual > 0 0.060 0.004 0.444 0.592 0.008 0.068 0.016 0.120* 0.144* 0.028

[1] [111] [148] [2] [262] [4] [30] [36] [7] [77]

Ha : AbAccrual < 0 0.044 0.248 0.000 0.000 0.124 0.036 0.044* 0.024* 0.004 0.088

[62] [0] [0] [31] [93] [11] [6] [1] [22] [40]

Business Model Shocks and Abnormal Accrual Models Model Shocks and Abnormal Accrual Models1 Edward Owens Joanna Shuang Wu Jerold Zimmerman Simon School of Business, University of Rochester,

Documents