Table of Contents

Lecture 1........................................................................................................................................................................1
Tutorial McVay 2006.....................................................................................................................................................6
Tutorial 1 Amiram 2015.................................................................................................................................................8
Tutorial 1..................................................................................................................................................................... 10
Lecture 2 Nichols, Wahlen 2004..................................................................................................................................16
Lecture 2 Nichols, Wahlen 2023..................................................................................................................................18
Lecture 2...................................................................................................................................................................... 25
Tutorial 2 Dechow (1994).............................................................................................................................................33
Tutorial 2 Leung, Veenman (2018)..............................................................................................................................38
Tutorial 2..................................................................................................................................................................... 42
Lecture 3 Sloan (1996).................................................................................................................................................51
Lecture 3...................................................................................................................................................................... 58
Tutorial 3 Bamber (2010).............................................................................................................................................72
Tutorial 3 Jung, Naughton, Tahoun, Wang (2018)........................................................................................................77
Tutorial 3......................................................................................................................................................................83
Lecture 4 Frankel, Jennings, Lee (2016)........................................................................................................................91
Lecture 4.................................................................................................................................................................... 101
Tutorial 4 Brown, Crowley, and Elliott (2020).............................................................................................................113
Tutorial 4 Frankel, Jennings, Lee (2022).....................................................................................................................124
Tutorial 4....................................................................................................................................................................131
Lecture 5 Basu (1997)................................................................................................................................................141
Lecture 5.................................................................................................................................................................... 158
Tutorial 5 Banker, Basu, and Byzalov (2017)...............................................................................................................170
Tutorial 5 Glaum, Landsman, and Wyrwa (2018)......................................................................................................176
Tutorial 6

. .183

Lecture 1

Bloom’s Taxonomy of Learning Domains

Two general approaches:
• Case-based approach: “Case studies focus on particular organizations, events, or
phenomena, and scrutinize the activities and experiences of those involved, as well as
the context in which these activities and experiences occur” 1
This is called the Qualitative research approach
• Big-Data approach: “an approach to research methodology that seeks to generalize
from specific data through the arms-length testing of hypotheses as to stable causal
laws.” 2
This is called the Quantitative research approach

The Objective of Financial Accounting Research (FAR):

FAR= Positivistic approach to research
• quantitative hypothesis-testing research
• Positive research: investigate elements of accounting practice that are
generalizable (global) rather than a unique case that occurs in a single
instance (local)

The Role of Theory:

Theory:
1. A group of ideas that aims to explain something
2. Existing knowledge (the academic literature)
• Something that informs, and at the same time, emerges out of data collection and
analysis

Key ingredient of Theory: Concepts

• Concept:
• an abstract idea generalized from particular instances
• A theoretical construct: a label we give to things that have common features
• Examples of concepts: Justice, Freedom, Love, Gravity, Quality,
• This course: Accrual Quality, Earnings Persistence, Earnings Management, Conservatism, Limited Attention.
The key take away that we can compare accounting reports
After the lecture you need to know about errors and accruals

Concept: Earnings management

• Earnings management = (What?) managers’ intentional over- or understatement of earnings to
achieve a particular reporting objective and thereby influence users’ perceptions of
company performance
(How?) • This can be achieved through:
1. Managing accruals: aggressive recognition of revenues, deferring or capitalizing costs,
understatement of liability reserves (e.g., warranty), etc.
2. Managing real activities: cutting R&D, using discounts to boost end-of-year sales, etc.
Why? – to be a good accountant and detect earnings management

In this paper, we focus on :

• Accruals-based earnings management is not necessarily fraud when it is done within

the discretion allowed under accounting rules
• Dechow and Dichev (2002, p.38): Accruals are temporary adjustments that shift the recognition of
cash flow over time
• Accrual accounting relies on estimates and assumptions made by managers

The company is unhealthy when there is fraud

• Under Armour was accused of aggressively recognizing revenues to meet quarterly

targets
• This leads to an increase in accruals and earnings but does not change the company’s cash flows

• This suggests that we can examine accruals to identify the possibility of earnings
management
• An accrual that results from revenue recognition (increase in receivables) should lead to a future cash
inflow
• But earnings management leads to errors in these accruals because the cash inflow is not realized
• But accruals can also contain errors because the future is uncertain (in other words,
there could be unintentional errors)
• This is the focus of Dechow and Dichev (2002)
We have to figure out what amount of error is exactable.

Measuring the quality of accruals and earnings

• Paper objective: present a new measure for the quality of accruals and earnings

• Why?
• We know from Dechow (1994): the “role of accruals is to shift or adjust the recognition of cash flows over time so that
the adjusted numbers (earnings) better measure firm performance” (Dechow and Dichev 2002, p. 35)
• But as we know from Sloan (1996), accruals can vary in the extent to which they make earnings a better measure of
performance, because “accruals require assumptions and estimates of future cash flows” => you can make an error in
assumptions (Dechow and Dichev 2002, p. 35) => greater uncertainty about the future likely reduces the quality of the
accrual estimates
• Dechow and Dichev (2002, p. 35) argue that “the quality of accruals and earnings is decreasing in
the magnitude of estimation error in accruals”
• Derive an empirical measure of this concept of accrual quality

• Example of errors in accruals (Dechow and Dichev 2002, p. 36):

“For example, recording a receivable accelerates the recognition of a future cash flow in earnings, and
matches the timing of the accounting recognition with the timing of the economic benefits from the sale.
However, accruals are frequently based on assumptions and estimates that, if wrong, must be corrected
in future accruals and earnings.”
“For example, if the net proceeds from a receivable are less than the original estimate, then the
subsequent entry records both the cash collected and the correction of the estimation error. We argue
that estimation errors and their subsequent corrections are noise that reduces the beneficial role of
accruals.”

When the system is fully cash-based, then just count cash.

Accounts receivable – booked profit, but did not receive it yet
Inventory – bought, and sold very quickly
Warranty – we paid for it, but we don’t know if we will use it

Why do accrual estimation errors occur?

• Simply because the future is uncertain – the accuracy of estimates in accruals depends on characteristics of the company,
such as the complexity of transactions, or the predictability of the information environment
• The earnings management literature suggests that management intent affects the incidence and magnitude of accrual
estimation errors
• Dechow and Dichev (2002, p. 36) focus on errors that are both intentional and unintentional:
“We argue that even in the absence of intentional earnings management, accrual quality will be systematically related to
firm and industry characteristics”
“For our purposes, we do not attempt to disentangle ‘‘intentional’’ estimation errors from unintentional errors because both
imply low-quality accruals and earnings”
• Focus is on (variation in) the quality of financial reporting more generally, not just low quality caused by
earnings management

In line with these examples, Dechow and Dichev (2002) build an empirical model to quantify accrual
estimation errors
• Focus on (short-term) working capital accruals because cash flow realizations related to working capital
typically occur within one year

Predictions
Regression residuals
Regression is the change in y based on change in x, everything else is residual => what this paper measures

Results

As predicted, current changes in working capital are negatively related to current cash flow from operations, and positively
related to past and future cash flow from operations.
Correlated with next week’s paper, the bigger the error = the lower the persistence.

Key learning objectives from Dechow and Dichev (2002)

• Accruals help to make earnings a better measure of performance (than cash flows), but the use of
accruals come at a cost: accrual estimation errors reduce the quality and persistence of earnings
• Dechow and Dichev (2002) introduce a method to quantify the quality of accruals and earnings
• Based on the extent to which accruals do not map into realizations of cash flows
• Using the residuals from a regression of working capital accruals on cash flows in different periods
• Accrual estimation errors result from both intentional and unintentional errors in the estimates and
assumptions reflected in accruals
• Several company characteristics are predictably related to the measure of accrual quality
• For example: firms with more volatile operations face more uncertainty => greater likelihood of estimation errors
Correlate error with incentives

Conclusion:

Our investigation of the interrelations between accrual quality, level of accruals, and earnings persistence also suggests a
reconciliation of the findings of Dechow (1994) and Sloan (1996). Dechow (1994) finds that accruals improve earnings'
ability to measure performance relative to cash flows. Sloan (1996) finds that the accrual portion is less persistent than the
cash flow portion of earnings, which suggests that firms with high levels of accruals have low quality of earnings. Our
reconciliation is based on the observation that a high level of accruals signifies both earnings that are a greater improvement
over underlying cash flows, and low-quality earnings. The reason is that accruals are largest when the underlying cash
flows have the most timing and mismatching problems, so more accruals signify greater improvement over the underlying
cash flows. However, this benefit comes at the cost of incurring estimation errors, and there will be a positive correlation
between levels of accruals and the magnitude of these estimation errors. Thus, everything else equal, large accruals signify
low quality of earnings and less persistent earning.

Tutorial McVay 2006

Earnings Management Using Classification Shifting: An Examination of Core Earnings and Special Items

I focus on the shifting of expenses between core expenses (cost of goods sold and selling, general, and administrative
expenses) and special items. There are many possible misclassifications to special items. Large charges, such as those
related to restructurings or mergers, offer a great deal of latitude and camouflage. For example, managers can classify
normal severance charges as charges resulting from the restructuring or merger. A manager might also allocate a greater
percentage of legal costs or other administrative expenses than were actually related to the restructuring or merger to the
‘‘special’’ fees. Classification shifting is not, however, limited to these large charges. Many unusual charges might contain
misclassified core expenses. Consider Y2K expenses, which might contain the salaries of permanent information
technology personnel, or a litigation gain or loss might contain day-to-day legal fees.

HYPOTHESES

Finally, as noted above, the shifting of expenses between core expenses and special items is viable; managers have
subjectivity over the classification of expenses, and the shifting is not expected to raise red flags for outside monitors.

H1: Managers classify core expenses as special.

In particular, I expect unexpected core earnings to be increasing in special items in year t.

Thus, the benefits to classification shifting are presumably greater for managers who can use their discretion to meet the
analyst forecast, especially in high-growth firms. This leads to my second hypothesis:

H2: Managers classify more core expenses as special in periods when the net benefits to classification shifting are expected
to be greater.

In particular, I expect classification shifting to be more pervasive when the shifting allows managers to meet the consensus
analyst forecast, especially in high-growth firms.

MEASURING CLASSIFICATION SHIFTING

I develop a methodology to measure classification shifting. I expect the core earnings of special-item firms to be overstated
in the year the special item is recognized. I model the level of core earnings and anticipate unexpected core earnings
(reported core earnings less predicted core earnings) in year t to be increased with special items in year t if managers are
classification shifting. As discussed above, an alternative explanation for this association is that core earnings are
unexpectedly high due to the immediate benefits of the restructuring charge or some other real economic event. In order to
distinguish between real economic changes and the opportunistic behavior of managers, I examine whether the
improvement associated with special items in year t reverses in year t+1.

Results:
TEST DESIGN AND RESULTS

H1 supported: Referring to the results for Equation (3b), presented in the lower half of Table 6, as predicted, special items
in year t are negatively associated with the unexpected change in core earnings in year t
+1. Overall, the results are consistent with managers classifying some core expenses as special in the year a special item is
recognized.

only special items that are susceptible to classification shifting experience unusually high core earnings in year t. This result
is consistent with real economic improvements as a result of special items, after controlling for the effects of classification
shifting.

H2 supported: The results are consistent with managers’ classification shifting to a greater degree when doing so allows
them to meet the analyst forecast.

There is some evidence that investors do not fully disentangle classification shifting and are negatively surprised when
previously shifted expenses recur as core expenses.

Real Economic Actions as an Alternative Explanation

I address this alternative explanation in two ways. First, I estimate Equations (3a) and (3b) by performance quintile (not
tabulated). If my findings merely reflect discretionary spending cuts in poorly performing firms, then I would expect to
only find results in the lowest quintiles of performance. However, evidence of classification shifting is present throughout
the performance quintiles, mitigating the likelihood that the effect is purely performance-driven. Second, I control directly
for changes in research and development expenditures to proxy for discretionary spending. Results are not sensitive to the
inclusion of this control variable. Thus, it does not appear that discretionary spending cuts are driving the relationship
between core earnings and special items.

Conclusion:

Overall, the evidence of classification shifting is compelling: (1) unexpected core earn- ings are increasing with special
items in year t, but this improvement reverses in the fol- lowing period; (2) the unexpected improvement only reverses if
there are no special items present in year t1, otherwise managers appear to classification shift again in year t1, thereby
maintaining the inflated core earnings; (3) these results hold only for those special items that, upon inspection, appear to be
amenable to classification shifting; (4) these results are stronger for firms that just met the analyst forecast, and stronger yet
for growth firms that just met the analyst forecast; and (5) there is some evidence that classification shifting is associated
with negative returns in the subsequent year, suggesting that investors are negatively surprised when expenses that were
previously excluded from core earnings recur.

Tutorial 1 Amiram 2015

Financial statement errors: evidence

from the distributional properties of financial statement numbers

Prior accounting literature outlines the limitations of current measures of financial statement errors, such as their correlation
with underlying firm characteristics and their reliance on time-series, cross-sectional, or forward-looking data, to name a
few

We construct a parsimonious, firm-year measure to assess the level of error in financial statements that overcomes some of
the concerns surrounding existing measures.

Literature in mathematics, statistics, and economics suggests that examining the distribution of the first or leading digits
(e.g., the leading digit of the number 217.95 is 2) of the numbers contained in a dataset allows users to assess the level of
error within the underlying data. The theoretical foundation of prior research using this method is based, implicitly or
explicitly, on the theorem proved by Hill (1995), which states that if distributions are selected at random and random
samples of varying magnitudes are then taken from each of these distributions, the leading digits of the combined mixture
distribution will converge to the logarithmic or Benford distribution, otherwise known as Benford’s Law. Specifically,
Benford’s Law states that the first digits of all numbers in an empirical dataset will appear with decreasing frequency (that
is, 1 will appear as the first digit 30.1 % of the time, 2 will appear 17.6 % of the time, and so forth).

We construct a measure, the Financial Statement Divergence Score (FSD Score for short) based on the mean absolute
deviation statistic as applied to the distribution of the leading digits of the numbers in annual financial statement data. The
FSD Score allows us to compare the empirical distribution of the leading digits of the numbers in a firm’s annual financial
statements to that of the theoretical or expected distribution defined by Benford’s Law

We contribute to this literature by implementing a measure that overcomes many of these limitations. First, the FSD Score
does not require time-series or cross-sectional data to estimate and does not model the error as a residual from a prediction
model. Second, based on its theoretical derivation, the measure is unlikely to have an ex ante relation with underlying firm
characteristics or business models since those characteristics or models do not theoretically cause firms to have financial
statement items that start with 1, 2, or any other digit. Third, the measure does not require forward-looking information.
Fourth, the measure does not require returns or price information.

We employ two statistics when measuring conformity to Benford’s Law—the Kolmogorov–Smirnov (KS) statistic and the
Mean Absolute Deviation (MAD) statistic.

Numerical analysis

We show under certain assumptions that this is the case with accounting data using a stylized numerical model. We also
show that, if the accounting estimates of the true cash flow realizations are without error, the distribution of the accounting
estimates (the financial statements) will follow Benford’s Law exactly. we summarize the results from the numerical
analysis here. In sum, we show that, under certain parameters, the FSD Score is increasing with the size of the error.
However, not all errors create deviations from Benford’s Law; the error needs to be applied in different rates to different
items in the distribution.

Simulation analysis

To further demonstrate how errors could alter conformity to Benford’s Law, we run a simple simulation that involves
changing the value of a single line item in a firm’s income statement and calculate how that change affected the financial
statements overall. As a result of the sales manipulation, a firm likely needs to adjust cost of goods sold and tax expense
accordingly. Therefore, we add three journal entries to the original numbers:
We find that firms with high FSD Scores tend to be smaller, younger, more volatile, and growing.

Results:

In examining the accrual quality measures, firms with higher FSD Scores tend to have more working capital accruals, more
discretionary accruals and higher values of the Dechow–Dichev measure, and are more likely to be a manipulator according
to Beneish’s M-Score. The result pertaining to the F_Score suggests that firms with higher FSD Scores are less likely to be
accused by the SEC of making material misstatements, which we explore in detail in Sect. 6. Finally, inspecting the
earnings quality measures, firms with higher FSD Scores tend to have less persistent earnings and are more likely to have a
loss.

Firms that misstate their financial results by manipulating select accounts may report numbers with first digits that are not,
in expectation, driven from the same interactions of random distributions that create conformity to Benford’s Law. Given
the nature of double-entry bookkeeping, this lack of consistency should trickle through several of the financial statement
line items.

We conjecture that current earnings will exhibit less persistence for firms with greater divergence from Benford’s Law.

Our prediction results collectively suggest that AAER firms are prosecuted for making material misstatements only once
they run out of room to manipulate their numbers, forcing them to report numbers that more accurately reflect their
underlying business activities and more closely reflect the theoretical distribution posited by Benford’s Law. In addition,
consistent with critics’ views that the SEC should ramp up its efforts to detect accounting fraud, these results provide
evidence that firms may be able to evade detection of financial statements errors, but their manipulations will still leave
traces in the distributional properties of their financial statements in the form of deviations from Benford’s Law.

Conclusion

Building on a method used in a variety of disciplines, we propose that firm stakeholders may find a firm-year measure of
financial reporting errors to be a useful tool to augment existing techniques to assess accounting data quality. Our measure,
the FSD Score, relies on the divergence from Benford’s Law, which states that the first digits of all numbers in a dataset
containing numbers of varying magnitude will follow a particular theoretical and mathematically derived distribution where
the leading digits 1 through 9 appear with decreasing frequency. This measure has significant advantages over alternative
measures of accounting quality currently used in the literature. For example, it does not require time-series, cross-sectional,
or forward-looking information, is available for essentially every firm with accounting information, and is uncorrelated ex
ante with firms’ operating performance and business models.

After providing intuition for the theory behind the measure, we use numerical methods to demonstrate that financial
statements without error are distributed according to Benford’s Law. We then provide several scenarios to demonstrate the
types of financial statement errors that are likely to create divergence from the law. For example, overestimating revenue,
underestimating expenses, meet-or-beat behavior, or a combination of these are likely to introduce deviation from the law.
To corroborate the results from the numerical analysis, we provide a simple simulation to demonstrate that when
accounting numbers are manipulated, there is a high likelihood of an increase in the divergence from the law.

Next, to establish whether the law applies to actual financial statement data, we show that at the aggregate level, financial
statement numbers conform to Benford’s Law in all industries and years. When assessing the conformity of individual firm-
years, we find that roughly 86 % of firm-years conform to the law as well. In examining the financial statements
individually, we find that the income statement has the greatest divergence from Benford’s Law. In examining the financial
statements by account types, we find that equity and liability (in contrast to asset) accounts, as well as income (in contrast
to expense) accounts, have the greatest divergence from the law.

Turning to firm characteristics, we find that firms that diverge from Benford’s Law tend to be smaller, younger, more
volatile, and growing. To shed light on the types of firm behavior associated with the FSD Score, we find that proxies for
accruals-based earnings management and earnings manipulation are related to divergence from Benford’s Law. However,
multivariate empirical analysis indicates that the FSD Score is incremental to these proxies. In addition, firms reporting
losses have weaker conformity to the law, and firms that report just above the zero earnings threshold have weaker
conformity than firms reporting just below zero.

We conclude by examining the relation between divergence from Benford’s Law and several ex post measures of earning
management. Our findings suggest that when restatements occur, the restated numbers are significantly closer to Benford’s
Law relative to the misstated numbers. Furthermore, as firms’ financial statements diverge from the law, their earnings
persistence decreases. The negative relations between the FSD Score and these ex post measures of earning management
support our claim that there exists a relation between the level of divergence from Benford’s Law and the informational
quality of reported financial results. Finally, we provide evidence that the FSD Score may serve investors, auditors,
regulators, and researchers by providing a leading indicator of material misstatements as identified by SEC AAERs.

Tutorial 1

Shifting expenses down the income statement to manage perceptions of “core” performance => Classification shifting

McVay (2006): “Earnings Management Using Classification Shifting: An Examination of Core Earnings and Special
Items”
What is core earnings and special items
Introduction:
1. In her study, McVay (2006) studies a phenomenon called “classification shifting” as an earnings
management tool for companies. Explain what she means with classification shifting.
• Earnings management: managers’ intentional over- or understatement of earnings to achieve a
particular reporting objective and thereby influence users’ perceptions of company performance.
Reporting objectives – what can be wrong – picture persistent with economic reality
• McVay’s (2006) definition (p. 501): “the misrepresentation or masking of true economic performance”
• Misrepresentation can be achieved through:
1. Managing accruals: aggressive recognition of revenues, deferring or capitalizing costs, understatement of
liability reserves (e.g., warranty), etc.
2. Managing real activities: cutting R&D, using discounts to boost end-of-year sales, etc.
Accounting performance is correct and economically it is wrong
• McVay (2006) introduces a third method:
3. Classification shifting: the deliberate misclassification of items within the income statement
• New information: This involves shifting expenses down, or revenues up, the income statement to “present a picture that is
not consistent with economic reality” (p. 501-502)
• Key distinction from 1. and 2.: this does not change bottom-line net income (“GAAP earnings”)

Classification shifting example:

income from operations => what investors are interested in

Take other business expenses to discontinued operations, which would increase income from operations.
Why is it bad? => changes alternative definitions of earnings
2. As she explains, classification shifting does not change “bottom-line” earnings. Why, then, can we still
label this activity as an earnings management tool?
Because you want to influence the perception of investors regarding core earnings.

• Why an earnings management tool?

• No effect on the bottom line, but classification shifting does change alternative definitions of earnings
• Analysts and investors often focus on a company’s “core earnings”; they rely on the presented subtotal Income
from operations AND they often ignore transitory expenses such as impairments and restructuring charges
• Classification shifting falls under McVay’s (2006) definition of earnings management because it leads
to a misrepresentation of an earnings subtotal
• McVay (2006) focuses on the U.S. setting and the misclassification of “core expenses” to “special
items” such as impairments and restructuring; this increases “core earnings”
• Core expenses = costs of goods sold (COGS) + SG&A
• Core earnings = revenues – COGS – SG&A => higher lead to higher share prices
Core earnings can be increased by selling more or misreporting => classification shifting

Another example:
Why would they do it? => Managers know that investors just hyperfocus on core earnings, so they manipulate data in their
favor. The objective is to overprice the company => compensations are tied to share price, share price is tied to company
value.
3. Think back about the discussion on investors’ processing of accounting information. Explain why and
how your understanding of investors’ processing of information can explain the existence of earnings
management in general, and classification shifting in particular.

Can we explain the existence of earnings management in general, and of classification shifting in
particular?
• From week 4, we know that the average investor has limited attention and processes accounting
information incompletely
Limited attention – if the investor sees the financial statement, they cannot focus on all the items, they just focus on
earnings.
• Example from Sloan (1996): on average, investors do not fully understand that the cash flow component of
earnings is more persistent than the accrual component; they “fixate” on the earnings number
• In general, investors’ fixation on earnings may explain the existence of earnings management, because
managers know that not all investors will adjust high earnings for the existence of large income--
increasing accruals (for example: aggressive revenue recognition or understatements of allowances)

4. The activity of classification shifting partly relates to companies’ presentation of non-GAAP measures
of earnings performance (see also week 2). Explain how we can make this link between classification
shifting and non-GAAP reporting.

• Why is the discussion on classification shifting related to non-GAAP reporting?

• managers, analysts, and investors often focus on adjusted measures of earnings performance
• Adjusted (non-GAAP) earnings: earnings that ignore expenses that are not predictive of future performance
• Examples of adjustments: stock-based compensation + transitory expenses like impairments
• McVay’s (2006) focus on core earnings is very similar to the focus on non-GAAP measures that exclude
transitory expenses such as impairments
• P506: “Special items tend to be excluded from core earnings by both managers (e.g., Lougee and Marquardt
2004) and analysts (e.g., Philbrick and Ricks 1991)”

More investors focus on adjusted earnings, by putting more items to special items => and more investors focus on core
earnings.

Understanding of the model: How?

Hypothesis:

• If users only focus on GAAP earnings, classification shifting would be “pointless”

• However, “individual components of the income statement are meant to be informative to financial
statement users, facilitating analysis by grouping items with similar characteristics (FASB Accounting
Concept No. 5). In general, the closer a line item is to sales, the more [persistent] this item tends to be
[...]. Furthermore, investors appear to recognize this distinction and weight individual line items within
the income statement differently” (p. 502)
• Consistent with the Nichols and Wahlen (2004) framework, investors value components of earnings differently, because
these components have different persistence
• Because net income does not change, the shifting of expenses in the income statement is less likely to
receive critical inspections by auditors and regulators

Certain line items are more important than others. We need to prove it.

• Focus on the allocation of expenses between (a) core expenses (COGS and SG&A expenses) and
(b) special items
• “As an anecdotal example of classification shifting, the SEC determined that Borden, Inc., classified $192
million of marketing expenses as part of a restructuring charge when it should have been included in selling,
general, and administrative expenses (Hwang 1994)” (McVay 2006, p. 502)
We need to find a method that proves that it is not one time thing.
• H1: Managers classify core expenses as special
• H2: Managers classify more expenses as special in periods when the net benefits to classification
shifting are expected to be greater
o Focus on companies meeting analyst expectations of core earnings, because meeting these
expectations is associated with high benefits (stock price increase)
*Meeting analyst expectations is an important reporting objective because investors are asking analysts whether they should
invest in that company or not. Investors look at analysts’ expectations and make estimations of how much they will earn
from dividends.
We need to show that certain line expenses are recorded as special items.

5. In simple words, explain how McVay (2006) empirically identifies the use of classification shifting by
companies.

• Panel dataset of U.S. firm years 1989-2003

• Empirical predictions following hypothesis:
• Classification shifting means that more expenses are misclassified as impairments or restructurings (special
items): this increases both core earnings and special items
 H1: positive relation between a measure of core earnings and special items
Why do we expect it to be positive? => the more special items you book, the higher core earnings.
• Recall that the Jones-model splits accruals into “nondiscretionary” (expected) and “discretionary”
(unexpected) accruals to identify accruals-based earnings management (see Stubben 2008 paper)
• McVay (2006) splits core earnings into expected and unexpected components

Regression is telling us that core earnings is too high and is incorrect. We can tell that it is too high, by checking accounts
receivable and cash, so model what the number should be by looking at sales, and accounts receivable… (more sales 20%
increase=>more core earnings cannot be 80% increase). The regression model predicts what that number should be.
We can explain 75% of core earnings.

Interpretation: if the firm reports a lot of special items => the firm is using earnings management.
The relation is negative next year because you already reported it the year before and you need to reverse it.
Next year you have less special items. A lot of core earnings this year will lead to less special items next year.
Moderation effect:

Reporting the exact number that was expected.

You need a little bit more, so you manage a little bit, so the auditor cannot notice.

Summary and conclusion:

• Managers do not only manage the earnings but also manage the classification of expenses within the
income statement
• Shift expenses from core expenses (cost of goods sold and SG&A) to special items
• No effect on bottom-line net income number
• Therefore less likely to receive critical inspection by auditors and regulators
• But this does overstate “core earnings”

• Important insights, because the market (analysts, the media, and investors) often focus on adjusted
measures of “core” earnings: also known as “pro forma” or “non-GAAP” earnings, which ignore
several expenses such as special items that are not persistent
• Given investors’ limited attention and the media’s and analysts’ focus on adjusted earnings measures,
unsophisticated investors might be misled by this classification shifting practice
• McVay (2006, p. 528): “There are many additional settings of classification shifting that future research
might investigate. For example, managers might shift core expenses to expense classifications other [...]
items, such as R&D, which is valued differently than other core operating expenses (e.g., Lev and
Sougiannis 1996; Aboody and Lev 2000)”

What is earnings management and error?

Why – auditors focus more at accrual earnings management, rather than classification shifting, so you can mislead investors
and analysts, without worrying auditors.

Limitation: what if you miss a variable in the regression model?

Amiram et al. (2015): “Financial statement errors: evidence from the distributional properties of
financial statement numbers”

• High quality financial reporting matters for well-functioning capital markets

• P1541: “Accurate financial reports enable efficient resource allocation and efficient contracting (Bushman and
Smith 2003). Therefore, assessing the errors in financial statements is an important task for investors,
analysts, auditors, regulators, and researchers.”
• Prior research creates estimates of reporting quality or likelihood financial statement errors using:
• Accrual or earnings quality (see Dechow and Dichev 2002)
• Discretionary accruals or discretionary revenues (see Stubben 2008)
o But these measures have major limitations, as Amiram et al. (2015) explain
• Objective of this paper: create and evaluate a new measure of the level of error in financial statements

What: Can we create a new measure that does not limit to what previous regression models do. => make a 4 th method to
measure earnings management.

6. The paper by Amiram et al. (2015) is an example of an academic study that uses “big data” and “data
analytics” techniques to quantify the possibility of material errors in financial statements, an issue
that is of high relevance to auditors. Their analyses are based on knowledge of the statistical behavior
of numbers, also known as “Benford’s Law.” Explain in simple words what Benford’s Law means and
how it may be applied to the data from financial statements.

This paper uses machine data analytics to find earnings management.

The question is does it work to find financial errors.
If you collect all the numbers in fin. Reports, some numbers should be mentioned more often than others, for example, 1
compared to 5.
The deviation from above mention distribution indicates an error.
The model is proved for other areas, but can we do it for accounting?

We know from our understanding of accrual accounting that financial statement line items represent
estimates of the realizations of companies’ cash flows
• Amiram et al. (2015, p. 1549-50): “Since accounting data are a series of estimations of the true cash flow
realizations of the underlying items (for example, cash flows from sales, cash flows from payments to
employees, etc.), the resulting distribution of the mixture of these cash flow realizations may [...] follow
Benford’s Law.”
• Here, error in the estimates of the true cash flows could be driven by both unintentional (prediction
errors) and intentional (manipulation) errors (as in Dechow and Dichev 2002)
• P1541: “Methods based on the law have been used to detect errors in published scientific studies,
questionable election data in Iran, suspicious macroeconomic data, internal accounts receivables data,
and misreported tax returns.”

Example:

We are taking absolute value, because we care whether it aligns with expectations, rather than whether it is positive or
negative.
7. Use Appendix 1 from Amiram et al. (2015) and the balance sheet of Google’s parent company
Alphabet Inc. for 2018, to compute the empirical distribution of numbers for Alphabet and the KS
and MAD statistics.

There are no material errors in fin. statement of google.

Why is it important:
Occurrence of restatements
Blue is bad reporting, because of deviations.

How: we calculate the deviation and look at the Benford’s Law

Higher error found by the fsd score(deviations) is confirmed by other measures.

After restatements, fsd score is lower.

It is good because you can review big data sets, and it is quick, and it aligns with other measures.

Summary:
• Investors, analysts, auditors, regulators, and researchers need measures to quantify the likelihood of
error and manipulation in financial statements
• This paper shows that we can use a statistical law (Benford’s Law) to create such as measure
• Based on the frequency of observing the first digit in the balance sheet, income statement, and cash flow
statement combined
• Greater deviation from the frequencies predicted by Benford’s Law are indicative of a higher potential for
material errors in the financial statements
• The measure has construct validity because it is correlated with other well-known measures of financial
reporting quality
• Example of a big data analytics approach to detecting financial statement fraud

Lecture 2 Nichols, Wahlen 2004

How Do Earnings Numbers Relate to Stock Returns? A Review of Classic Accounting Research with Updated
Evidence

In this article, we summarize the theory and evidence on how accounting earnings information relates to firms' stock
returns, particularly for the benefit of students, practitioners, and others who may not yet have been exposed to this
literature. In addition, we present new empirical evidence on the relation between earnings and returns by replicating and
extending three classic studies using data from 1988 through 2002.

Earnings (or more precisely, accounting net income) represents the "bottom-line" accounting measure of firm performance.
A firm's earnings number is an accrual accounting measure of the firm's profit or loss from business activities and events
during a quarter or annual period. A firm's earnings number represents an accounting measure of the change in the value of
the firm to common equity shareholders during a period (apart from the effects of direct transactions with shareholders,
such as paying dividends or issuing shares).

A firm's stock return, which equals the change in the firm's market value over a period of time plus any dividends paid,
represents the capital market's measure of the firm's "bottom line" performance over a period of time. How do these
"bottom lines" relate? How do accounting earnings numbers relate to stock returns?

We organize this review using the three theoretical links between earnings and share prices developed by Beaver (1998).
These "three links" are:
1. current period earnings provide information to predict future periods' earnings, which
2. provide information to develop expectations about dividends in future periods, which
3. provide information to determine share value, which represents the present value of expected future dividends.
THE RELATION BETWEEN EARNINGS AND STOCK RETURNS
The Three Links Relating Earnings to Stock Returns

Link 1 in the three-links framework assumes that a current period earnings number provides two important elements of
information useful for developing dividend expectations: information about current period wealth creation and information
about future earnings.

Link 2 in the three-links framework assumes that current and future earnings represent wealth created by the firm that will
ultimately be distributed to equity shareholders through dividends. Thus, current earnings and forecasts of future earnings
indicate future dividend-paying ability, which shareholders can use to develop expectations of future dividends.

Link 3 therefore represents the classical approach to equity valuation, which views share value as the present value of the
future dividends the shareholder expects to receive over the remaining life of the firm. Current period earnings numbers
(and related financial reports) provide shareholders with information to develop expectations for those future earnings,
which aid in developing expectations of future dividends, which ultimately form the basis for share value.

The Importance of Earnings Persistence and Unexpected Earnings

Earnings persistence refers to the likelihood a firm's earnings level will recur in future periods, an essential element of link
1.For example, if a firm generates one-for-one persistent earnings, then shareholders should expect the firm to generate the
same levels of earnings in future years. The brief analogy in the previous section of the $1,000 savings account firm with
100 percent dividend payout illustrates the case of one-for-one persistent earnings — the firm will generate $50 of earnings
this period and $50 of earnings in all future periods.

On the other hand, a firm might experience unusually high or low earnings in a given period because a component of
current period earnings has low or even zero persistence (i.e., transitory earnings); shareholders would not expect such
levels of earnings to recur. For example, if a firm recognizes a one-time gain or loss this period, that gain or loss will not
likely persist in future periods.

The three-links framework, therefore, provides a useful structure for analyzing the valuation implications of earnings
information. Under this framework, share value reflects the present value of expected future dividends, which are
determined by current and expected future earnings. When firms announce earnings that unexpectedly differ from the
market's expectations, share prices generally react to the "earnings news." Generally speaking, if earnings beat expectations,
share prices increase, and likewise, if earnings fall short of expectations, share prices fall. By how much? That depends on
many factors, but an important factor is the persistence of the unexpected earnings a firm announces an unexpected change
in earnings that is not likely to persist, then share prices will likely change by the amount of the one-time earnings change.
On the other hand, when the firm announces an unexpected change in earnings that will likely persist in the future, share
prices will generally move up or down by a larger amount due to the link between current and future earnings —
persistence. Thus, when a firm announces earnings that differ from expectations (unexpected earnings), the three-links
framework provides a set of steps one can follow to analyze the implications of an unexpected change in earnings for future
earnings (persistence), future dividends, and share value.

Isolating the Effect of Unexpected Earnings on Stock Returns

The market's use of information to price shares is a complex and dynamic process. As a result, the association between
unexpected earnings and stock returns depends on numerous factors, each of which is difficult to precisely specify in
empirical tests. Four such factors that researchers consider are:
 earnings information (central variable of interest),
In examining the association between unexpected earnings and stock returns, we test the fundamental question of whether
earnings numbers reflect information that the capital markets believe is relevant and reliable. Despite the power of accrual
accounting and GAAP, earnings numbers might not provide useful information to the capital markets. Accounting earnings
determined under GAAP may: ( 1) provide an incomplete measure of firm performance in a given period (e.g., a firm may
have developed a promising new product this period, but GAAP will not permit the firm to recognize any revenues until the
firm generates sales of that product), or ( 2) measure performance with a conservative bias (e.g., expensing R&D).
Moreover, capital markets participants may be wary that some firms under certain conditions may attempt to manage
reported earnings numbers to a point where they are not reliable indicators of economic performance. Indeed, recent
popularity of the notion that "cash is king" may reflect (in part) mistrust of the relevance and reliability of earnings
information. If factors such as these destroy the usefulness of accounting earnings, then we should observe no association
between accounting earnings and stock returns.
 earnings expectations,
= new information communicated by earnings (the unexpected earnings). To isolate the new information in earnings, we
subtract the earnings that the (presumably efficient) market expected to occur this period. Researchers and capital market
participants often use consensus analysts' earnings forecasts as proxies for earnings expectations in the capital market.
Alternately, if analysts' forecasts are not available, then prior-period earnings (or prior-period earnings compounded with an
expected growth rate) often serve as an estimate of expected earnings.
 market efficiency with respect to earnings information,
Refers to the scope of the information that prices reflect and the degree to which capital market prices react quickly and
completely (e.g., without bias, not consistently under- or over-reacting) to new value-relevant information, such as
unexpected changes in earnings. Market efficiency does not assume that the capital markets are omniscient — prices reflect
only the information known to the market. Nor does market efficiency assume prices are prescient — surprises happen in
our world of uncertainty. Market efficiency is not an absolute — it is not that prices either are or are not efficient. Instead,
market efficiency is a matter of degree, which describes how much information prices reflect and how quickly prices react
and reach new equilibrium levels. A highly efficient market with respect to accounting earnings numbers would react
quickly and completely when new earnings-related information becomes available.
 asset pricing.
Asset pricing models in financial economics predict that risky securities, such as shares of stock, should provide a sufficient
rate of return to compensate the securities holders for forgoing consumption and bearing risk. As one example, the capital
asset pricing model (CAPM) predicts that equity securities will earn the risk-free rate of return plus a risk premium that
depends on the nondiversifiable risk of that security (i.e., beta) multiplied by the average risk premium per unit of risk in
the economy. Thus, stocks should earn an expected rate of return to compensate investors for forgoing consumption and
bearing risk, separate from the change in share value that is attributable to unexpected news about the firm's earnings. Thus,
we control for the expected rate of return on each stock in order to isolate and detect the firm-specific incremental return
associated with the firm's unexpected earnings.

We can never be certain that accounting information causes stock market price reactions. Instead, we rely on carefully
constructed statistical and econometric tests to isolate and examine the association (the correlation) between unexpected
earnings and stock returns.

Lecture 2 Nichols, Wahlen 2023

The Essential Role of Accounting Information in the Capital Markets: Updating Seminal Research Results with
Current Evidence

In this paper, we use recent data to examine whether seminal findings from prior studies in the following three major areas
of research still hold: the earnings-returns relation, earnings management, and market efficiency. We also introduce some
new findings that were not in the original studies.

The relation between financial accounting information and stock prices reveals insights into the economic relevance of
financial reporting as a source of firm-specific information that has important implications for how the capital markets
allocate capital and value shares.

Since the turn of this century, a number of fundamental changes have occurred in the capital market's information
environment, including the following:
 Substantial developments in U.S. GAAP and International Financial Reporting Standards (IFRS);
 Innovations in technology that accelerate the dissemination of earnings information and reduce the costs of
acquiring and processing that information;
 Enhanced capital market regulations, including the Sarbanes-Oxley Act of 2002;
 Changes in firms' business models and global activities, including increasing expenditures to internally
develop intangible assets (such as research and development to create intellectual property, advertising to
 create brand names, and others) that are expensed immediately and are not capitalized on balance sheets or
amortized over useful lives in income statements;
 Significant expansion in information released with earnings announcements, including a greater frequency of
management guidance for future earnings and more frequent SEC filings (e.g., Form 10-K) on the same day;
 More frequent reporting and use of non-GAAP performance measures;
 Leaps forward in the speed and efficiency of share trading activities; and
 Substantial increases in trading activities that are not directly linked to firm-specific accounting information,
such as algorithmic trading and exchange-traded funds.
THE RELATION BETWEEN BUSINESS ACTIVITIES AND SHARE PRICES
How Do Business Activities Relate to Share Prices?
The extent to which market prices reflect the implications of accounting information depends on three steps:
Step 1: Financial Reporting
Each period, firms engage in a vast array of different transactions, events, commercial arrangements, and business
activities. Financial statement preparers follow professionally established accounting principles and standards (such as U.S.
GAAP and IFRS) to measure, summarize, and report information about firms' business activities in balance sheets, income
statements, and statements of cash flows. Reporting financial statements makes firms' private information about their
business activities available to external stakeholders.
Step 2: Analysis and Valuation
Investors and analysts use financial statements and a wide array of other information to evaluate firms' profitability, growth,
and risk. Using this information, they develop expectations about future earnings, cash flows, and dividends. In
fundamental analysis, analysts and investors use various valuation approaches, like the discounted cash flows model and
the residual income model ([40]), for estimating share values based on expected future earnings, cash flows, and dividends.
Step 3: Share Trading and Pricing
Investors trade on their information and their share value estimates, buying or selling shares when they believe share prices
are below or above fundamental values. These trading activities may cause share prices in the equity capital markets to
change. When share prices in the market deviate from fundamental values, these share trading activities should move prices
closer to fundamental values.

When these three steps function effectively, share prices should reflect the value-relevant information that financial
statements provide about firms' fundamental business activities. However, these three steps do not always function
perfectly.

Slippage can occur in the financial reporting process in step 1 when balance sheet and income statement amounts rely on
historical costs that no longer reflect current values (e.g., outdated historical costs for plant and equipment). In addition,
some firms rely heavily on research and development to develop valuable intellectual property or engage in customer
service activities and advertising to establish a valuable brand name.

Slippage in step 2, the analysis and valuation process, can arise from poor forecasts and expectations, poor valuation
methods, or both. Companies differ in the richness of information they report to investors for forecasting these future
business activities, which can create slippage in analysts' and investors' forecasts and value estimates. Also, to the extent
that analysts' and investors' forecasts of future performance are influenced by firms' reports of non-GAAP performance
measures that overstate (or understate) financial performance and growth relative to the GAAP–based performance
measures, it may also cause slippage in step 2. Further slippage can arise when analysts and investors do not conduct
careful analysis and valuation and instead rely on simplistic valuation heuristics (e.g., price-earnings ratios).

Step 3 assumes share trading is informed by analysts' and investors' share value estimates based on accounting information.
However, trading often occurs for many other reasons. For example, slippage in step 3 can arise from liquidity trading (e.g.,
selling shares to meet cash needs), noise trading (e.g., trading on rumors), market frictions (e.g., wide bid-ask spreads or
short-sale restrictions), and market sentiment (e.g., bubbles and crashes), which can lead to temporary departures of share
prices from fundamental values.

How Do We Test the Relation between Earnings and Stock Returns?

The theory explaining the relation between earnings information and stock prices makes three general predictions ([ 7]).[ 9]
First, the theory predicts that earnings numbers (or, more broadly, financial statements) provide information to equity
shareholders about current period profitability. Second, the theory predicts that current period accounting information helps
analysts and investors develop expectations about the firm's future earnings, cash flows, and dividends. Third, the theory
predicts that share price equals the present value of expected future cash flows into the firm, which should equal expected
future dividend payments to shareholders.

To test these predictions with empirical data, researchers examine the associations between changes in earnings numbers
and changes in stock prices by using following factors:


Earnings information (central variable of interest),
 Earnings expectations and unexpected earnings,
 Earnings persistence,
 Market efficiency, and
 Risk-adjusted stock returns.

EVIDENCE LINKING EARNINGS AND STOCK RETURNS

To What Extent Do Earnings Changes Relate to Stock Returns?
 These results are striking, suggesting that merely the sign of the change in earnings is associated with an average difference
of 25.4 percent in abnormal annual stock returns.[17] In comparison, [ 4] reported a difference of 16.8 percent based on
differences in the sign of the change in earnings per share over their 1957–1965 sample period.

Our results suggest that earnings changes are economically important pieces of information in the capital markets because
simply knowing the sign of the earnings change would allow investors to capture 33.8 percent (25.4 percent/75.1 percent)
of the total information impounded in price on average during the year.

How Do Changes in Other Firm Performance Measures Relate to Stock Returns?

The results in Table 2, Panel B and Figure 2 indicate that, for the 12-month period from month –11 through month 0, firms
with positive annual changes in cash flows from operations experience average abnormal returns of 9.3 percent, whereas
firms with negative changes experience average abnormal returns of –8.5 percent. These results suggest the sign of the
changes in cash flows from operations is very informative (an average returns spread of 17.7 percent), but the sign of the
changes in earnings contains more value-relevant information (an average returns spread of 25.4 percent). This occurs
because, in part, changes in cash flows from operations are more ambiguous measures of the firm's wealth creation and
profitability in a given period.

The results in Table 2, Panel B and Figure 2 indicate that firms with positive changes in EBITDA experience average
annual abnormal returns of 9.2 percent, whereas firms with negative changes experience returns of −6.2 percent (an average
returns difference of 15.4 percent). Together, the results from these comparative analyses suggest that measures of
performance based on cash flows or EBITDA are informative and capture value-relevant information, but they are not
nearly as informative as accounting earnings because they exclude important value-relevant information measured in
bottom-line earnings numbers.

The evidence suggests that earnings numbers contain important information that relates to changes in the market's share
prices, encompassing steps 1 through 3. The evidence also suggests that earnings reflect more value-relevant information
than other common measures of firm performance, such as changes in cash flows from operations or EBITDA (step 1).
Finally, this evidence reveals the potential value of investing based on accurate forecasts of the sign and magnitude of
future earnings changes.[25] This partly explains why so many investors and analysts devote so much time and energy to
forecasting earnings (step 2). In spite of the limitations of U.S. GAAP and IFRS and the potential for earnings management
(as discussed later), accounting earnings numbers appear to provide useful information about firm value to the capital
markets.
Do Earnings Numbers Convey New Information to the Market?

Capital market participants obtain information from a variety of sources, namely, analysts' research reports, articles in the
financial press, company press releases and conference calls, to name a few. In addition, over the past decade, firms have
substantially expanded the information released with earnings announcements, including bundling earnings guidance with
earnings announcements and filing Form 10-K on the earnings announcement day.

We examine daily returns surrounding the day firms announce quarterly earnings, from four days before (day −4), the day
of the announcement (day 0), and through five days after (day +5). We sort stocks into ten portfolios based on unexpected
earnings for that quarter. We sharpen our measure of the new information in the earnings announcement by calculating
unexpected earnings as earnings per share minus the analysts' consensus earnings per share forecast, divided by price per
share 60 trading days before the announcement. We measure the consensus analysts' earnings per share forecast as the
median forecast issued by analysts between days −120 and −60 relative to the earnings announcement. We group firms each
quarter into 10 portfolios, ranging from the lowest decile to the highest decile of unexpected earnings that quarter (the worst
earnings news to the best earnings news).

We report the results in Table 3 and Figure 4. During day 0 through the end of day +1, Figure 4 reveals sharp increases in
abnormal returns for stocks announcing good earnings news and sharp decreases for stocks announcing bad news. In
particular, by the end of day +1, prices for the best earnings news stocks (decile 10) rose by an average of 4.0 percent,
whereas prices for the worst earnings news stocks (decile 1) fell by an average of 3.8 percent. The difference in returns
through day +1 equals 7.7 percent and is strongly statistically significant. Of this 7.7 percent difference in reaction, 7.0
percent occurs on days 0 and +1. To put this in context, consider that the average market capitalization of the firms in this
sample is $6,432.89 million. An average firm of that size in the top earnings news decile would experience a $257.3 million
increase in market value by day +1, whereas an average firm of that size in the bottom decile of earnings news would
experience a $244.4 million decline in market value by day +1, which is a difference of $501.7 million.

The evidence in this section has two important implications. First, the results suggest that earnings numbers convey new
information (step 1) that has important consequences for share prices. Second, the stock price consequences of new
earnings information reveal that market participants process (step 2) and trade (step 3) on that information quickly—stock
prices react quickly and incorporate most of the new information by the end of day +1.

To What Extent Do Prices Lead Earnings?

To what extent does the additional two-thirds of the information impounded in share values (steps 2 and 3) predict future
earnings? To address this question, we formed portfolios each year using a nested sort procedure. We first sorted firms into
quintiles by current year earnings changes (scaled by total assets). Next, within each earnings change quintile, we sorted
firms into deciles by contemporaneous abnormal returns. Finally, we measured average one-year-ahead earnings changes
(scaled by total assets) for each abnormal returns decile portfolio. This procedure allows current period returns to vary
while controlling for current period earnings changes.
Controlling for current period earnings changes, the results in Table 4 and Figure 4 show that current period stock returns
predict earnings changes in year +1. On average, the firms in the highest abnormal returns decile in the current year
generated earnings (as a percent of total assets) that were 1.1 percent higher in year +1 than those in the current year. Firms
in the lowest abnormal returns decile in the current year generated earnings (as a percent of total assets) that were −1.0
percent lower in year +1 than those in the current year. Given that the median rate of return on assets (roughly speaking,
earnings divided by total assets) equals 3.5 percent for firms in our sample, these results suggest that current period stock
returns predict significant differences in return on assets one year ahead. Stock returns during the current year not only
reflect current year earnings news but also predict changes in earnings next year.

EVIDENCE ON SLIPPAGE IN THE THREE STEPS LINKING FIRMS' BUSINESS ACTIVITIES TO SHARE
PRICE...
Do Managers Influence the Earnings Reporting Process?

Managers may be concerned about meeting or beating three natural earnings benchmarks. In accounting research, analysts'
earnings forecasts are common proxies for the market's expectations for firms' earnings. This benchmark is salient to
managers because it reflects the level of earnings that analysts expect the firm to generate to justify the current share price.
Other prominent benchmarks include reporting positive earnings, indicating the company turned a profit for the common
shareholders, as well as meeting or beating last year's earnings, indicating that the company's performance improved.

To examine how reported earnings numbers relate to analysts' earnings forecasts as a benchmark, we examine the
distribution of analysts' earnings forecast errors around zero, replicating [22]. We measure forecast error by subtracting the
analysts' consensus quarterly earnings forecast from the reported quarterly earnings per share. We then divide the analysts'
forecast error by share price 60 days before the earnings announcement. (This is the same unexpected earnings variable we
used in our earnings announcement tests reported in Table 3.) Next, we place the observations into bins of narrow width
(for example, 0.5 percent increments) based on the value of the scaled earnings forecast error. Finally, we count the number
of observations in each bin and plot the frequency in Figure 5. The darkly shaded areas above and below the x-axis report
the number of observations in each bin that are above or below the number of observations in the corresponding bin on the
opposite side of the distribution. Notice the striking amount of darkly shaded areas just above zero and just below zero on
the x-axis—many more firms than expected report earnings that just beat analysts' forecasts, and many fewer firms than
expected report earnings just below analysts' forecasts. These statistics provide circumstantial evidence to suggest that firms
manage earnings to meet or beat analysts' earnings forecasts.
Although the evidence supports earnings management around these benchmarks, the results are substantially weaker than
those reported in the original studies. As [17], [18] note, our sample period is characterized by greater degrees of regulatory
attention focused on earnings management and greater consequences to managers who are caught cooking the books. One
possible explanation (which we do not test) for our results is that changes in the regulatory and legal environment have
diminished (but not eliminated) managers' propensity to engage in earnings management behavior around these
benchmarks. Also, another possible explanation noted earlier is that greater numbers of firms are increasingly reporting
non-GAAP performance measures, so perhaps managers are less concerned about meeting or beating GAAP-based earnings
benchmarks.

How Efficiently Does the Market Impound Earnings News into Share Prices?

To address this question, we use the same 10 earnings news portfolios each quarter (grouped from best decile to worst
decile earnings news) from our earnings announcement tests in Figure 3 and Table 3. As soon as a firm announces earnings
for a quarter, we add that stock to a portfolio based on the earnings news.[31] Following Bernard and Thomas, we cumulate
daily returns over 60 trading days before and 60 trading days after the earnings announcement. Daily returns over the
preannouncement period (from day −60 to −1) reflect the arrival of information prior to the earnings announcement (day 0)
that aids investors in predicting the sign and magnitude of the earnings news (step 2).
Immediately after the day of the earnings announcement, we begin cumulating returns from zero again for each of our 10
portfolios each quarter. This is equivalent to forming 10 portfolios each quarter based on the earnings news announced the
day before (again, best news decile to worst news decile). If the market is fully efficient with respect to quarterly earnings
information, then share prices should quickly and completely impound the earnings information at the announcement date,
and we should not observe any consistently significant abnormal returns in the postannouncement period.
We summarize each decile portfolio's stock return performance by averaging the abnormal returns for the firms in each
decile each quarter. We then average each decile's return across the 72 quarters in our sample (quarter 1 (Q1) 2002 through
Q4 2019). Our sample for this analysis includes 150,540 firm-quarter observations, described in Table 1. We present the
results in Table 5 and Figure 6.
Overall, the results in Tables 4 and 5 suggest that capital markets are highly efficient with respect to earnings news, but not
yet fully efficient, particularly when small-cap and mid-cap firms announce extremely good or extremely poor earnings
news. This phenomenon has become known as "post-earnings-announcement drift," and it remains one of the most puzzling
anomalies in tests of the degree of capital market efficiency with respect to earnings information.[33] However, contrasting
our results with those in [13] indicates that preannouncement returns have increased and post-earnings-announcement
returns have diminished in recent periods. The evidence points to the possibility that, in a highly efficient market with
substantial incentives to exploit this anomaly, share prices have become more efficient with respect to quarterly earnings
news. However, the evidence suggests that share prices for small-market-capital stocks still react with some delay to
extreme quarterly unexpected earnings, reflecting slippage in step 2 and/or step 3.
CONCLUSION
We use data from 2002 to 2019 to contribute updated empirical evidence on the extent to which some of the key results still
hold. Despite the many changes occurring in financial reporting, the capital markets, and securities regulation, our evidence
shows that changes in earnings still remain strongly associated with stock returns and are associated with roughly one-third
of the information impounded in share prices each year. In addition, we show that earnings convey more information than
cash flows from operations and non-GAAP performance measures such as EBITDA. Furthermore, we demonstrate that
market prices react to earnings news very quickly in days around earnings announcements. We also find that stock returns
continue to convey a lot of information about one-year-ahead earnings, beyond the information in current period earnings.
We also contribute updated evidence that indicates managers still intervene in the earnings reporting process to manage
earnings upward to just meet or beat earnings benchmarks, but this behavior has diminished considerably since the original
studies, perhaps due to more disciplined auditing and corporate governance, as well as tighter securities regulation.
Moreover, we find that market prices have become more efficient in pricing quarterly earnings news, but prices are not yet
completely efficient, particularly for mid-cap and small-cap stocks. Our evidence shows that the associations between
earnings and stock returns remain robust and explain (in part) why investors, managers, boards of directors, analysts, the
financial press, auditors, securities regulators, and others continue to place so much importance on accounting information.

From intro:

Our first set of results provides evidence on steps 1 through 3 by showing a significant relation between the sign of the
change in annual earnings and annual stock returns over the same period, updating [ 4]. Firms that reported earnings
increases experienced stock returns that were on average 12.3 percent higher than the returns of similar size firms during
the same period of time. In contrast, firms that reported earnings decreases experienced stock returns that were on average
13.1 percent lower than similar size firms' returns. These results suggest that even information as simple as the sign of the
change in earnings relates to an average difference of 25.4 percent in annual stock returns.
We benchmark these abnormal returns by sorting our sample firms into two portfolios each year, namely, one including the
stocks that experienced positive abnormal returns over the next year and the other including the stocks that experienced
negative abnormal returns over the next year. With perfect foresight of the sign of one-year-ahead abnormal returns, the
positive returns portfolio experienced average abnormal returns of 43.2 percent per year, whereas the negative returns
portfolio experienced average abnormal returns of −31.9 percent, which is a spread of 75.1 percent. This benchmark
indicates that, on average, changes in earnings are associated with roughly one-third (33.8 percent = 25.4 percent/75.1
percent) of the information impounded in share prices during the year.
Recent years have seen an increasing number of firms reporting non-GAAP performance measures like EBITDA (earnings
before interest, taxes, depreciation, and amortization) and pro forma earnings. To what extent are the capital markets
substituting these performance measures for earnings? We find that the capital markets draw substantially more value-
relevant information from changes in annual earnings than either changes in EBITDA or changes in annual cash flows from
operations, highlighting the importance of the accrual accounting information in earnings (step 1: financial reporting).
Given the technological innovations in information dissemination, the expansion of large amounts of information released
with earnings announcements, and the increasing use of trading mechanisms that are not linked to firm-specific earnings
information, we look more closely at the earnings-returns relation to examine when the capital market reacts to earnings
news. Like [10], we find that share prices react quickly to quarterly earnings news announcements. Over the ten-day period
surrounding quarterly earnings announcements, stock prices for the firms announcing the largest earnings increases rose on
average 4.7 percent, whereas stock prices for the firms announcing the largest earnings decreases fell by an average of 4.1
percent. Of this 8.8 percent difference in announcement period returns, 7.0 percent occurs on the announcement day and the
following day. These results show that quarterly earnings announcements still convey new information (step 1) and reveal
that market participants process (step 2) and trade (step 3) on that information quickly.
Taking a step further, does the information impounded in share prices relate to earnings beyond the current year? If the
current period change in earnings is associated with roughly one-third of the information impounded in share prices, does
the additional two-thirds of the information impounded in share values (steps 2 and 3) predict future earnings? To address
this question, we formed portfolios each year by using a nested sort procedure that allows current period returns to vary
while holding current period earnings changes constant. Controlling for current period changes in earnings, the results
indicate that stocks with the highest abnormal returns in the current year are more likely to experience earnings increases in
year +1, whereas stocks with the lowest abnormal returns in the current year are more likely to experience earnings declines
in year +1. The difference in year +1 earnings changes amounts to 2.0 percent of total assets. This difference is
economically significant because the median firm in our sample generates earnings equal to 3.5 percent of total assets.
Thus, the changes in stock prices during the current year reflect not only current year earnings news, but also portend
changes in next year's earnings.
The empirical results to this point indicate that the three steps mapping business activities to financial reports to share
values and capital market prices continue to operate well, despite major changes in the market's information environment.
But these steps do not completely explain share prices or stock returns. What are potential sources of slippage in each of the
three steps?
Because earnings information is associated with important capital market consequences, a significant stream of accounting
research asks the following: do firm managers influence the financial reporting process (step 1) to report higher earnings in
order to meet or beat earnings expectations?[ 5] Prior studies find unusual patterns in the distributions of earnings numbers
around earnings benchmarks, including meeting or beating analysts' forecasts, avoiding reporting losses, and avoiding
reporting earnings declines (e.g., [19]; [22]). These results imply that firm managers influence the financial reporting
process to manage earnings up, especially if earnings numbers would otherwise fall short of benchmarks. However, it is
possible that in the post-Sarbanes-Oxley 2002 era, with more rigorous auditing and corporate governance, tighter securities
regulation (e.g., SAB 99), and heightened awareness among executives of their responsibility for reported earnings
numbers, the frequency with which managers intervene in the financial reporting process may have diminished.[ 6] Also, as
noted earlier, given the increased reporting of non-GAAP performance measures, perhaps managers are less concerned
about meeting or beating GAAP-based earnings expectations. Consequently, it is not clear whether the prior evidence on
earnings management to meet or beat benchmarks persists in recent years.
To re-examine earnings management, we extend the earnings distribution tests using reported quarterly earnings numbers
and analysts' consensus earnings forecasts (as proxies for the market's expectations) as benchmarks. Like the prior studies,
our evidence indicates that more firms than expected report earnings that meet or beat analysts' forecasts, and fewer firms
than expected report earnings that are just below analysts' forecasts. However, our evidence is weaker than the findings in
the original studies. The spikes in earnings distributions just above and just below the benchmarks are still observable, but
are smaller than they were in the early studies. This evidence suggests that managers continue to intervene in the financial
reporting process (slippage in step 1) to manage earnings upward to meet or beat analysts' earnings forecasts, but the
frequency of this behavior has diminished.
In our final set of analyses, we revisit the seminal findings of [13], [14], who documented that stock prices continue to drift
after earnings announcements, suggesting the capital markets were not perfectly efficient in reacting quickly to earnings
surprises (slippage in steps 2 and 3). The decades since the [13], [14] study period have seen innovations in technology that
accelerate the dissemination of earnings information and reduce the costs of acquiring and processing that information, as
well as faster and less costly share trading activities. In addition, the findings of [13], [14] are well known, so sophisticated
market participants should be aware of the potential to earn excess returns following earnings announcements. Have these
developments and the incentives to earn post-earnings-announcement returns enhanced the degree of market efficiency with
respect to earnings information?
During the 60 trading days leading up to quarterly earnings announcements, the firms in the top decile of earnings increases
each quarter experienced average abnormal returns of 7.3 percent, whereas the firms in the bottom decile of earnings
decreases experienced negative abnormal returns of −6.2 percent, which is a difference that averaged 13.5 percent. These
preannouncement returns are more pronounced than those observed in Bernard and Thomas (roughly 10.4 percent). These
results reveal that share prices anticipate earnings news weeks in advance of the actual announcement (steps 2 and 3).
However, during the 60-day period after these earnings announcements, stock prices for the firms that announced the
largest earnings increases continued to rise by an average of 2.1 percent, whereas stock prices for the firms that announced
the largest earnings decreases continued to fall by an average of 0.9 percent. The average 3.0 percent spread we observe in
abnormal returns extends the seminal findings on the post-earnings-announcement drift anomaly documented by [13], [14],
but is considerably smaller than the 4.2 percent average post-earnings-announcement drift returns they observed. Taken
together, our results suggest that the capital markets are not yet completely efficient (slippage in steps 2 and 3) with respect
to pricing quarterly earnings information, although the markets have become more efficient, perhaps due to incentives to
exploit this anomaly and advances in the efficiency of share trading.
Lecture 2
• Conceptual foundations underlying the role of financial reporting and stock prices
• Nichols and Wahlen (2023) three-step framework mapping business activities into share prices?
• Nichols and Wahlen (2004) theoretical framework as part of this three-step framework: why and how do
earnings relate to stock returns?
• Link between IASB conceptual framework and the conceptual foundation of financial accounting research
papers
• Importance of understanding earnings persistence for valuation impact of earnings news
• Basic results from Nichols and Wahlen (2023)
• Accounting earnings are “value relevant”
• Earnings reflect new information, but are not particularly timely
• Managers continue to intervene in the financial reporting process to manage earnings upward to meet or beat
analysts’ earnings forecasts, but the frequency of this behavior has diminished.
• Stock prices react quickly to news, but subsequent “drift” suggests this reaction is not complete

Nichols and Wahlen (2023)

Introduction
• Nichols Wahlen (2023, p.105): “The relation between financial accounting information and stock prices
reveals insights into the economic relevance of financial reporting as a source of firm-specific
information”
• This research demonstrates:
• the associations between accounting income numbers, share trading activity, and changes in share
prices;
• and the complex and dynamic relations between accounting information and stock prices.
1. Nichols and Wahlen (2023) present a three-step framework that maps firms’ business activities
into share prices:
1. Financial Reporting
2. Analysis and valuation
3. Share trading and pricing

• The financial reporting process provides useful information about firms’ financial
position and performance, which impacts investors’ and analysts’ expectations about
future earnings and cash flows and, therefore, share values. Changes in share value
estimates trigger share trading, which in turn causes share prices to change in the
capital markets.
The key question is when does accounting information impact expectations?

• This paper:
• Why does this work? The 3-Steps Framework
• When does this not work? Slippage in (one of) the three steps

We want to understand step 1, what do these journal entries represent.

We want to use financial reporting to do financial analysis in order to understand how much we will earn in the future (step
2).

If we see that the share price does not align with expectations, we buy it or sell it (step 3).

When is the link between firm-specific information and share prices weakened?
Slippage in the 3-step framework

Slippage in Step 1: The Financial Reporting Process

• Conservatism: outdated historical cost/asymmetric conservative bias
• Internally Developed Intangible Assets: large off-balance sheet assets and equity
• GAAP provides an incomplete or biased measure of firm performance
• e.g. revenue recognition for new products until all performance obligations are met

Accounting is not easy. There are accounting standards, but the chices that you are making within those standards will
change financial reports. If you use historical data, it does not represent the actual value, and it cannot be used in steps 2
and 3 (conservatism).

Revenue recognition – delayed/deferred revenue. We cannot show that in our revenue recognized and we cannot use this
information in steps and 3.

In addition R&D and Advertisement.

Slippage in Step 2: The Analysis and Valuation Process

• Poor forecasts and/or valuation methods
• Companies differ in the richness of the information they provide
• Non-GAAP measures affecting forecasts
• Analysts and investors do not conduct careful analysis and valuation and instead rely on simplistic
valuation heuristics (e.g., price-earnings ratios)
Investor attention.

Slippage in Step 3: The Share Trading and Pricing Process

• Trading occurring for reasons other than based on accounting information:
• Liquidity trading
• Noise trading
• Market frictions
• Market Sentiment
• Algorithmic trading
• Exchange-traded funds
Computer problems

Theory

• Prior research provides many insights that helps us understand the relevance of financial reporting
• The theory predicts that new accounting information (e.g., unexpected earnings) that triggers a change in
investors’ expectations for future cash flows and dividends should trigger a change in the share value of
the firm.
• Nichols and Wahlen (2023, p.111): This requires an understanding of:
1. Earnings Information
2. Earnings Expectations and unexpected earnings
3. Earnings Persistence
4. Market efficiency
5. Risk-adjusted stock returns

The earning information now is not accurate, because there may be R&D, deferred revenue etc..

Earning is a change in a book value.

Stock change is looked at by investors over the long period of time.

• This theory is better summarized in Nichols and Whalen (2004), who present a framework to link
earnings and stock returns
• Three links form the foundation of the relation between earnings and returns (Nichols and Wahlen, p.
264; Nichols and Wahlen, p.110):
1. Current period earnings provides information to predict future periods’ earnings, which ...
2. ... provide information to develop expectations about dividends in future periods, which ...
3. ... provide information to determine share value, which represents the present value of expected future dividends

We want to determine the share price (sell) and dividend (receive) in a long term

By how much should it increase – expected rate of return.

We care about making expectations about future dividends.

Not every firm pays out dividends, so we also look at free cash flow models.

Cash flows can be expressed in future earnings.

The framework

That is why we are interested in current earnings, as it reflect future dividends, and it is interesting to investors.

This framework explains why accounting is interesting for investors.

• The theory of link #1 assumes that earnings, and financial reporting

more broadly, provides information to investors about current, and expected future, profitability
1. Current period earnings summarizes information about the wealth created by the company during the
period for shareholders
2. Current period earnings provides information useful for prediction
• Nichols and Whalen (2004, p. 266): “firms depend on financial reporting to convey credible information
about their ability to generate future wealth for equity shareholders and other stakeholders”
• Note that the IASB recognizes the importance of “predictive value”

Predicting is a key of accounting numbers.

• The theory of link #2 assumes that current and future earnings areinformative about dividend distributions
• From practice and academic insights, we know this assumption is valid:
dividends are typically linked to earnings
• Even if firms do not pay dividends, the firm could pay higher dividends it was more profitable
• The theory of link #3 goes back to our valuation example and assumes that share prices equal the
present value of expected future dividends to the shareholders
• Simply put, the framework explains that the information in earnings should relate to changes in stock
price, because earnings are informative about the future dividends/cash flows/earnings the company
is expected to generate

Unexpected earnings

• Nichols and Wahlen (2023, p.111): New information is important because it can cause
expectations to change, which can trigger changes in share prices.
• To isolate the new information in earnings, we subtract the earnings that the market
expected: unexpected earnings
• Example:
• Analysts and investors expect a firm will announce $3.00 EPS (earnings per share) this period.
• If the firm announces that EPS are $4.00, then the new information is that the firm generated
unexpected EPS of $1.00 this period. The firm was more profitable than expected.
• Will this unexpected $1.00 of EPS cause investors and analysts to revise expectations for future
earnings? That depends on earnings persistence.
How do we know if the information is new: => Unexpected earnings
If this earning number comes out, will it change the share price?

Unexpected Earnings Persistence

Trying to understand if investors find information useful.
We have an expectation and we wonder if our expectation will happen or something changes => look at unexpected
earnings/ new information

We bought shares at 30

We expect a 3$ per share return.

• Nichols and Wahlen (2004, p. 268): “Earnings persistence refers

to the likelihood a firm’s earnings level will recur in future periods,
an essential element of link 1”
• Example: Savings account of $100: with a constant interest rate of 1%, earnings on a
savings account are fully persistent: every year you receive the same $1
• Prediction: more persistent earnings should have a stronger impact on stock prices
• Firm with 10% discount rate and persistent earnings of €1 per year (fully paid out as dividends)

The present value of all future revenue is 11 because no unexpected earnings

Compensating for holding money at the company and not a bank.

 *34

If something new happens just for one time

If something new happens not just for one time

We need to figure out how the market uses financial information => looking for unexpected earnings and whether they are
persistent or not.

• The theoretical framework helps us understand why changes in earnings

and changes in stock market valuations should be related
• Explains why stock prices react to news in earnings announcements
• Understanding the persistence of earnings further helps us determine how strongly the stock price
should respond
• For example: the reduction in earnings for Philips was substantial, but the stock price went down by less than
2%; but why?
• The major driver of the negative earnings change was an impairment charge
• Impairments are one-time, transitory, events
• Thus: impact on value should be limited

Nichols and Wahlen (2004)

• Nichols and Wahlen (2004): The association between earnings and returns is complex and dynamic, depending among
others on:
1. Earnings information (“the central variable of interest”)
2. Earnings expectations (“new information communicated by earnings (the unexpected earnings) ”)
3. Market efficiency (“the scope of the information that prices reflect and the degree to which capital market prices react
quickly and completely”)
4. Asset pricing (“expected rate of return”)
• Nichols and Wahlen, 2004, p.269: “We can never be certain that accounting information causes
stock market price reactions”. Because it is complex and dynamic
• Nichols and Wahlen, 2004, p.269: “We rely on carefully constructed statistical and econometric
tests to isolate and examine the association (the correlation) between unexpected earnings and stock
returns”.

• Analyst expectations: see weeks 1 2, 3

• One important reason for the limited drop in Philips’ stock price was that the impairment was already
announced by the company, and (partly) reflected stock price
• To better understand the stock market reaction to earnings, we should have a precise measure of
unexpected earnings
• Options:
1. Same quarter of last year (this is what the press referred to with Philips)
Used by Nichols and Wahlen (2023)
2. A comprehensive prediction model based on multiple signals
3. Forecasts of analysts made on or before the day of the earnings announcements
 This is what many researchers use because it includes recent news (such as the impairment
announcement of Philips) easiest

What if we know at month -10 we make an investment, how at year 0 does the share price behave?
The share price tends to go the same way as earnings.
If you can predict the decline in earnings, you can sell shares and save your money.
It you do the expectations of earnings good, you can make money as an investor.

Firms that have a loss=> share price goes down

Firms that have a gain => share price goes up

Earnings are very relevant to the share price. Up – up, worn - down

Loss is not so persistent. So gain leads to a higher change. Investors do not react to a decrease in price as much, as an
increase.

You need to look at other sources because when the company sends a report it is too late to react to the share price change.

Accruals is the best predictor of share prices compared to EBITDA and Cash flows.

When it is a long window, how do we know if the change in a share price depends on the firm value or other external
factors => look at shorter window.
How efficient is a market in doing following expectations The top firm does the best of all. The lowest one has the most
negative surprise. The positive surprise is more rapid because people don’t think about it, as they do it with negative
surprises.

The effect is smaller for firms that meet expectations, rather than the ones that did not meet.

Firms have a higher earnings management when they want to meet the benchmark.

Do managers influence the earning reporting process?

• If the underlying business performance is falling a little short of the benchmark, some
managers might resort to managing earnings upward to meet or beat the benchmark
• To examine how reported earnings numbers relate to analysts’ earnings forecasts as a
benchmark, Nichols and Wahlen (2023) examine the distribution of analysts’ earnings
forecast errors around zero
• Number of observations just above zero and just below zero: many more firms than
expected report earnings that just beat analysts’ forecasts, and many fewer firms than
expected report earnings just below analysts’ forecasts.
Very few firms have negative expectations.

Summary of results:

• The relation between news in earnings and changes in stock price is robust and still exists today
• Accounting earnings are “value relevant”
• Results from tests of the role of earnings persistence highlight the importance of the three-link
framework in evaluating stock price responses to earnings
• Earnings reflect new information, but are not very “timely” a lot of the change in the share price happened before the
announcement, so you lose the possible gains
• Stock prices react quickly to news, but subsequent “drift” suggests this reaction is not complete. The market is not
completely efficient
• Managers continue to intervene in the financial reporting process to manage earnings upward to meet or
beat analysts’ earnings forecasts, but the frequency of this behavior has diminished

Exam questions:
• How to empirically evaluate the usefulness of financial reporting
If it is useful, you should exoect more change in share price, when the gains are positive.
• Importance of earnings as a summary measure of firm performance
If you invest based on earnings, you can expect the return.
• Role of earnings “persistence” and three links between earnings and returns
Depending on that you will include in for one tear or a continuity.
• Research design, and measures of the “news” in earnings

We expect significant association between what is reported and what is in the market, and are there unexpected events and
are there one time or will they repeat.
Earnings and returns have a positive association. But it takes until you can report earnings. So it is not timely to look at the
reports. Plus there is also earnings management. So look also at other sources.

Tutorial 2 Dechow (1994)

Accounting earnings and cash flows as measures of firm performance: The role of accounting accruals

This paper investigates circumstances under which accruals are predicted to improve earnings’ ability to measure firm
performance, as reflected in stock returns.

Earnings are the summary measure of firm performance produced under the accrual basis of accounting. Earnings are
important since they are used as a summary measure of firm performance by a wide range of users.

The view adopted in this paper is that the primary role of accruals is to overcome problems with measuring firm
performance when firms are in con- tinuous operation. Information asymmetries between management and other
contracting parties create a demand for an internally generated measure of firm performance to be reported over finite
intervals. This measure can be used to contract and recontract as well as to evaluate and reward management. The success
of a firm depends ultimately, on its ability to generate cash receipts in excess of disbursements. Therefore, one performance
measure that could be used is net cash receipts (realized cash flows). However, over finite intervals, reporting realized cash
flows is not necessarily informative. This is because realized cash flows have timing and matching problems that cause
them to be a ‘noisy’ measure of firm performance. To mitigate these problems, generally accepted accounting principles
have evolved to enhance performance measurement by using accruals to alter the timing of cash flows recognition in
earnings.

Two important accounting principles that guide the production of earnings are the revenue recognition principle and the
matching principle. The revenue recognition principle requires revenues to be recognized when a firm has performed all, or
a substantial portion, of services to be provided and cash receipt is reasonably certain. The matching principle requires cash
outlays associated directly with revenues to be expensed in the period in which the firm recognizes the revenue. By having
such principles, the accrual process is hy- pothesized to mitigate timing and matching problems inherent in cash flows so
that earnings more closely reflects firm performance.

‘Information about enterprise earnings and its components measured by accrual accounting generally provides a better
indication of enterprise performance than does information about current cash receipts and payments.’
However, the use of accruals introduces a new set of problems.’ Management typically have some discretion over the
recognition of accruals. This discretion can be used by management to signal their private information or to oppor-
tunistically manipulate earnings.

However, to the extent that manage- ment use their discretion to opportunistically manipulate accruals, earnings will
become a less reliable measure of firm performance and cash flows could be preferable.

The accrual process is therefore the result of a trade-off between relevance and reliability

In this paper, the empirical tests use stock price performance as the bench- mark against which to compare realized cash
flows and earnings. Stock prices are viewed as encompassing the information in realized cash flows and earnings
concerning firm performance.

The purpose of this paper is to examine why earnings is the most frequently used summary measure of firm performance.

Testable predictions

The existence of information asymmetries between the firm’s managers and outside parties contracting with the firm
creates a demand for a summary measure of firm performance. This measure can be used to evaluate manage- ment and as
a source of information to investors and creditors on the firm’s cash generating ability. The problem faced by contracting
parties is that although management is the most informed party to report on the firm’s performance, they are also evaluated
and rewarded based on the firm’s performance. There- fore, in the absence of objective procedures to determine
performance, external parties have difficulty assessing the reliability of signals produced by manage- ment.

On the one hand, contracting parties could demand that managers report realized cash flows. These can be objectively
measured but are influenced by the timing of cash receipts and disbursements. On the other hand, management could
attempt to determine the firm’s expected future cash flows. This, however, would provide management with so much
reporting flexibility that any signal produced would be difficult to verify and would result in an unreliable measure of firm
performance.

The accrual process can be viewed as trading off these two problems when producing earnings.

Measurement interval predictions

This study compares the ability of earnings relative to net cash flows and cash from operations to reflect firm performance.
Net cash flows will fluctuate with cash inflows and outflows associated with the firm’s investment and financing activities
as well as the firm’s operating activities. Net cash flows have no accrual adjustments and are hypothesized to suffer
severely from timing and matching problems (see Appendix 1). Cash from operations reflects the net cash flows generated
by the firm’s operating activities. This measure includes accruals that are ‘long-term’ in nature (i.e., do not reverse within
one year) and mitigate timing and matching problems associated with the firm’s investment and financing activities.
However, cash from operations exclude accruals associated with changes in firms’ working capital requirements. Earnings
contain accruals that mitigate the timing and matching problems associated with firms’ operating, investment, and financing
cash flows. Therefore, earnings are predicted, on average, to be a more useful measure of firm performance than either cash
flow measure.

Hypothesis 1. There is a stronger contemporaneous association between stock returns and earnings than between stock
returns and realized cashjows over short measurement intervals.

A short measurement interval is defined as one quarter or one year. Consider increasing the time interval over which
performance is measured (e.g., four years). Over longer intervals, cash flows will suffer from fewer timing and matching
problems and so the importance of accruals diminishes. Therefore, over longer intervals, earnings and realized cash flows
are expected to converge as measures of firm performance (assuming clean surplus).

Hypothesis 2. The contemporaneous association of stock returns with realized cash Jlows improves relative to the
contemporaneous association of stock returns with earnings as the measurement interval is increased.

Cross-sectional predictions

Thus, earnings will differ from realized cash flows in each period to the extent that (i) credit sales are excluded from
realized cash flows and (ii) realized cash flows include collections from the previous period’s credit sales.

Eq. (2) reveals that the magnitude of the difference between revenues and cash flows for any period will be greater (i) the
larger cp(i.e., the proportion of sales on credit) and (ii) the larger the magnitude of the change in revenues (AS,).

The accrual process is most important for firms that have had large changes in the net balance of their noncash accounts.
If cash collection is reasonably certain, then the actual timing of the cash collection is not relevant for reporting purposes. If
accruals reduce timing and matching problems in cash flows, then earnings are expected to reflect relatively more value-
relevant events when earnings and cash flows differ by the greatest magnitude.

!When aggregate accruals are large in magnitude (either positive or negative) earnings will more closely reflect firm
performance than realized cash flows.

However, when firms undertake new investment and financing activities or experience large changes in their working
capital requirements (when cpdS, is large in absolute magnitude), realized cash flows are expected to be a relatively poor
measure of firm performance. Under such circumstances, realized cash flows suffer from timing and matching problems
and are less able to reflect firm performance. Accruals are expected to reduce these problems in earnings.

Hypothesis 3. The larger the absolute magnitude of aggregate accruals made by a jirm, the lower the contemporaneous
association between stock returns and realized cash$ows relative to the association between stock returns and earnings.

Firms with longer operating cycles are expected to have larger working capital requirements for a given level of operating
activity. Therefore, in firms with longer operating cycle, a given change in the level of operating activity (AS,) is expected
to translate into a larger change in the required level of working capital (cpdS,). Thus, the length of the operating cycle is
predicted to be an underlying determinant of the volatility of working capital. Cash from opera- tions excludes accruals
relating to the firm’s operating activities. Hence the ability of cash from operations to measure firm performance is
expected to decline as the length of the operating cycle increases.

Hypothesis 4. The longer a jirm’s operating cycle, the more variable the jirm’s working capital requirements and the lower
the contemporaneous association between stock returns and realized cash$ows.

Stock returns as u benchmark measure of firm performance

This paper assumes that stock markets are efficient in the sense that stock prices unbiasedly reflect all publicly available
information concerning firms’ expected future cash flows. Therefore, stock price performance is used as a benchmark to
assess whether earnings or realized cash flows better summarize this information.

Sample description and variable measurement

Variable defenitions

Empirical results

Measurement interval predictions

Hypothesis 1 is examined by performing three pooled regres- sions: (1) stock returns on earnings, (2) stock returns on cash
from operations, and (3) stock returns on net cash flows.

These results support Hypothesis 1.

Consistent with Hypothesis 2, there is a relative increase in the explanatory power of cash flows over longer measurement
interval.r3

In summary, the results demonstrate that accruals improve the association of earnings with contemporaneous stock returns.
Table 2 provides evidence consis- tent with cash flows suffering from matching problems since cash flows exhibit larger
temporary components than earnings. Table 3 indicates that cash flows have a relatively lower association with stock
returns than do earnings. To- gether, these results suggest that cash flows suffer from greater timing and matching problems
than earnings. This indicates that the negative correlation observed between cash flows and accruals in Table 2 is not due
solely to management ‘arbitrarily’ smoothing earnings. The results suggest that accruals are performing a useful role in
mitigating timing and matching problems in cash flows, so that earnings better summarizes firm performance.

Cross-sectional predictions

Hypothesis 3 predicts that, when aggregate accruals are large in absolute value, net cash flows will be a relatively poor
measure of firm performance. Table 5 provides results of a test to determine whether the ability of net cash flows to reflect
firm performance declines as the absolute value of aggregate accruals [abs(AA)] increases.

The results presented in Table 5 support Hypothesis 3. They demonstrate that cash flows are not a poor measure of firm
performance per se. In steady-state firms, where the magnitude of accruals is small and cash flows and earnings are most
similar, cash flows are a relatively useful measure of firm performance. However, when the magnitude of accruals
increases, indicating that the firm has large changes in its operating, investment, and financing activities, cash flows suffer
more severely from timing and matching problems.
Hypothesis 4 predicts that the length of the operating cycle is an underlying determinant of the variability of working
capital and hence cross-sectional variation in cash from operation’s association with stock returns. To investigate this
hypothesis, two proxies for the length of the operating cycle are calculated for the annual interval

Thus, in industries where the operating or trade cycle is long, working capital requirements also tend to be more volatile.
This is consistent with the length of the operating cycle being an economic determinant of the volatility of working capital
requirements. Cash from operations excludes short-term working capital accruals, therefore cash from operations are
predicted to be a poor measure of firm performance in industries with long operating cycles.

Evaluation of accrual components

They suggest that operating accruals, as a group, are important for mitigating timing and matching problems in cash from
operations.

These results are consistent with long-term operating accruals playing a less important role than working capital accruals in
mitigating timing and matching problems in cash from operations.

Conclusion

This paper hypothesizes that one role of accounting accruals is to provide a measure of short-term performance that more
closely reflects expected cash flows than do realized cash flows. The results are consistent with this prediction. First, over
short measurement intervals earnings are more strongly associated with stock returns than are realized cash flows. In
addition, the ability of realized cash flows to measure firm performance improves relative to earnings as the measurement
interval is lengthened. Second, earnings have a higher association with stock returns than do realized cash flows in firms
experiencing large changes in their working capital requirements and their investment and financing activities. Under these
conditions, realized cash flows have more severe timing and matching problems and are less able to reflect firm
performance.

This paper also predicts that although accruals improve earnings’ association with stock returns, certain accruals are less
likely to mitigate timing and matching problems in realized cash flows. Evidence is presented indicating that long-term
operating accruals play a less important role in this respect. In addition, the inclusion of special items in earnings is shown
to reduce earnings’ association with stock returns over short intervals. However, this does not imply that special items
should be excluded from earnings from continuing operations.
If it is desirable that management are held accountable for such charges, then it is important to include special items in
earnings.

Tutorial 2 Leung, Veenman (2018)

Non-GAAP Earnings Disclosure in Loss Firms

This study examines the incremental information in loss firms’ non-GAAP earnings disclosures relative to GAAP earnings.
Using a large sample ob- tained through textual analysis and hand-collection, we posit and find that loss firms’ non-GAAP
earnings exclusions offset the low informativeness of GAAP losses for forecasting and valuation.

Prior research indicates that losses tend to be less persistent than profits, because of shareholders’ option to liquidate the
firm. As a result, the aggregation of items with varying persistence can reduce the informativeness of GAAP losses for
forecasting and valuation.

Our main tests focus on the ability of GAAP versus non-GAAP earnings measures to predict future operating cash flows
and earnings. For a bench- mark sample of firm-quarters without non-GAAP earnings (GAAP-only loss firms), we first find
that GAAP earnings are significantly positively associ- ated with future performance. For firms with GAAP and non-
GAAP losses (non-GAAP loss firms), however, we find that the predictive coefficient on GAAP earnings is less than half
of the coefficient obtained for GAAP-only loss firms. This result is even stronger for loss converters, where we find GAAP
earnings to be unrelated to future performance. Non-GAAP earn- ings are strongly predictive in both sets of GAAP loss
firms, while non-GAAP exclusions are not predictive of future performance. Consistent with the ar- gument that the
aggregation of items with different persistence can reduce the informativeness of GAAP earnings, we conclude from these
results that loss-firm managers use non-GAAP disclosures to offset the low predictive ability of aggregate GAAP earnings.
Background and Predictions

LOSS FIRMS, INFORMATION UNCERTAINTY, AND DEMAND FOR DISCLOSURES

A substantial proportion of U.S. firms report negative earnings (Hayn [1995], Givoly and Hayn [2000], Klein and
Marquardt [2006]). Besides the practical challenge of implementing valuations, losses are associated with increased
uncertainty about future earnings. This uncertainty partly stems from investors’ ability to liquidate (parts of) a firm that
operates poorly

They also find that losses of firms not engaged in R&D are valued consistent with the abandonment option, while losses of
R&D firms are valued as if R&D creates value. These results suggest there is substantial variation in the persistence and
valuation implications of losses.

Assuming stationarity in the variance of earnings over time and all else being equal, Dichev and Tang [2009] show
analytically that lower earn- ings persistence can translate into less predictable, and more uncertain, fu- ture earnings.

Consistent with the notion that losses are associated with in- creased uncertainty because of their lower persistence, Brown
[2001] finds that analysts’ forecast errors are larger for loss firms, while Liu and Natara- jan [2012] document that loss
firms are associated with greater dispersion in forecasts.

This leads to investors request more supplementary disclosure.

NON-GAAP EARNINGS DISCLOSURE AND INCOME STATEMENT DISAGGREGATION

By excluding GAAP income statement line items that are less informa- tive about future firm performance, managers’
disclosures of non-GAAP measures can be more informative to investors than aggregate GAAP earn- ings.

Prior studies also find that managers are more likely to report or emphasize non-GAAP earnings when GAAP earnings are
less in- formative, and that non-GAAP earnings measures are, on average, more strongly associated with stock returns than
GAAP earnings

Non-GAAP measure allow investors to separate cash and noncash components, recurring and nonrecurring components in
order to have a better picture of the firm performance.

NON-GAAP EARNINGS DISCLOSURE IN LOSS FIRMS

Because of the elevated demand for supplemental disclosures and the benefits to firms of providing credible voluntary
disclosures, we argue that loss firms provide a unique setting in which the informative (op- portunistic) considerations are
likely to be relatively more (less) important in explaining non-GAAP disclosures.

A potential argument against the prediction that non-GAAP earnings dis- closures are particularly informative for loss firms
is that the salience of the profit–loss benchmark (Degeorge, Patel, and Zeckhauser [1999]) can also trigger managers to
opportunistically exclude certain expenses from GAAP earnings and inflate outsiders’ perceptions of firm performance.
Managers may have increased opportunity to do so given the elevated uncertainty about future earnings for loss firms, and
many studies use the conversion of a GAAP loss into a non-GAAP profit as an indicator of aggressive non- GAAP
reporting

Finally, it is important to note that prior studies mostly conclude that non-GAAP earnings are informative because of
transitory item exclusions (Bradshaw and Sloan [2002], Bhattacharya et al. [2003]). In recent years, the earnings
components that managers typically exclude are recurring ex- penses such as stock-based compensation and the
amortization of intangi- bles (Whipple [2015]). Prior studies generally view recurring expense ex- clusions as more difficult
to justify than transitory item exclusions, and of- ten interpret recurring exclusions as “low-quality” exclusions indicative of
aggressive non-GAAP reporting

Data and Descriptive Statistics

Most importantly, we find strong differences in the role of the underlying GAAP earnings. Consistent with evidence in
related literature (e.g., Black, Christensen, Ciesielski, and Whipple [2017]), the negative coefficients on the cash flow and
accrual components of earnings for profit firms suggest that non-GAAP reporting is more likely when GAAP earnings are
lower.
Main Tests and Results

PREDICTIVE ABILITY OF GAAP VERSUS NON-GAAP EARNINGS FOR FUTURE FIRM PERFORMANCE

where Future performance captures firms’ operating cash flows or operating earnings summed over the subsequent four
fiscal quarters, scaled by total assets. In equation (2), we split GAAP earnings into non-GAAP earnings and exclusions:

GAAP earningsiq ≡ Non-GAAP earningsiq + Exclusionsiq .

Controls is a vector of com- mon control variables, that is, firm size, age, growth opportunities, and earnings volatility. In
addition, untabulated year-quarter and industry fixed effects are included in the estimation.

we find that GAAP losses are highly predictive of future cash flows.

Moving to non-GAAP loss firms, this predictive ability is strongly reduced

The difference with GAAP-only loss firms is again highly sig- nificant

These results suggest that the disclosure of non-GAAP earnings relates to the limited information in GAAP losses about
future cash flows

To examine the incremental information in the non-GAAP measures, we next estimate equation (2). Results in the fourth
and fifth regression columns reveal that non-GAAP earnings in both sets of loss firms are strongly predictive of fu- ture
cash flows: a $1 increase in non-GAAP earnings is associated with about $2.50 higher future cash flows. In contrast,
exclusions are not predictive of future cash flows. These results suggest that the items excluded in calcu- lating non-GAAP
earnings contribute to the low predictive ability of GAAP earnings, and that these items are so “noisy” for the set of loss
converters that they eliminate the predictive ability of GAAP earnings.

We formally assess the incremental information in non-GAAP over GAAP earnings by testing the significance of the
difference in coefficients on GAAP versus non-GAAP earnings in columns (2) versus (4) for non-GAAP loss firms, and
columns (3) versus (5) for loss converters. Tests based on stacked regressions reveal that the estimate of the incremental
information in non-GAAP earnings is highly significant, with the greatest incremental information among loss
converters.30

Overall, we find that non-GAAP earnings are highly predictive of future performance for loss firms and offset the low
predictive ability of GAAP losses. Inconsistent with evidence based on profit and loss firms combined (e.g., Doyle,
Lundholm, and Soliman [2003], Kolev, Marquardt, and McVay [2008]), we find that the items excluded from GAAP
earnings are largely uninformative about future firm performance. Hence, these results sug- gest that non-GAAP
disclosures are particularly informative for loss firms and help users disaggregate losses into components that are
differentially informative for forecasting and valuation. Moreover, the lack of predictive ability of the exclusions suggests
that non-GAAP disclosures are likely to be less strategic for loss firms than for profit firms.

TESTS BASED ON MATCHED SAMPLES AND ENDOGENEITY CONCERNS

we propensity-score match each non-GAAP observation with a GAAP-only ob- servation to ensure that any differences in
observable characteristics across the samples can be considered random. Effectively, this procedure helps us compare the
predictive ability of GAAP earnings for firms with sim- ilar levels of GAAP earnings, expenses such as stock-based
compensation, and other characteristics, but that differ because non-GAAP earnings are disclosed.

Effectively, this procedure helps us compare the predictive ability of GAAP earnings for firms with sim- ilar levels of
GAAP earnings, expenses such as stock-based compensation, and other characteristics, but that differ because non-GAAP
earnings are disclosed.

The results suggest that even when observations are matched based on the observable char- acteristics, the GAAP earnings
of non-GAAP-disclosing loss firms are sig- nificantly less predictive.

we conclude that it is not highly likely that an unob- served confounding variable is strong enough to overturn the results.

PREDICTIVE ABILITY OF NON-GAAP EARNINGS IN LOSS VERSUS PROFIT FIRMS

Our results presented thus far suggest that non-GAAP disclosures are particularly predictive of future performance and
potentially less strategic for loss firms than for profit firms.

This evidence is consistent with the non-GAAP disclosure more likely reflecting the limited information in GAAP for
losses than for profits

Con- sistent with the non-GAAP disclosure providing incremental information, the coefficients on profit firms’ non-GAAP
earnings are significantly larger than those on GAAP earnings (difference of 1.262 and 1.282, respectively). However, we
also find that the coefficients on the Exclusions variable are significant and nontrivial. This result suggests the exclusions
are persistent, which is in line with prior evidence on combined samples of profit and loss firms

To assess whether the incremental predictive ability of non-GAAP earn- ings and exclusions differ between loss and profit
firms, we pool the sam- ples of loss and profit firms with non-GAAP earnings and interact all vari- ables to test whether the
differences are significant. The results suggest that the incremental predictive ability of non-GAAP earnings for loss firms
is significantly greater than for profit firms.

Leung and Veenman (2018) find that adjusted earnings measures, from which managers exclude expenses such as stock-
based compensation and intangible amortization, have stronger correlations with future company performance than the
GAAP earnings of loss-making companies.

Additional Analyses

MARKET REACTION (ERC) TESTS

persistent losses are positively associated with firm values when R&D expenses contribute to the loss.

Among loss converters, most firms exclude stock-based compensation expense in the calculation of non-GAAP earnings
(see table 2). To the extent that stock-based compensation has positive consequences for firm value (e.g., Hanlon et al.
[2003]), this notion might explain the marginally significant coefficient on the exclusions surprises. Consistent with this
con- jecture, the results in table 7 also suggest that the coefficient on Exclusions surprise becomes more negative and
significant when we restrict the sample to observations for which stock-based compensation is excluded, and the coefficient
is insignificant for the remaining observations. These results, however, should be interpreted with caution, given the small
magnitudes of the coefficients and the limited number of observations in which loss converters do not exclude stock-based
compensation.

FUTURE PERFORMANCE LEVELS OF LOSS CONVERTERS VERSUS GAAP-ONLY LOSS FIRMS

Our results suggest that while non-GAAP disclosures are particularly in- formative for loss firms, loss converters are a
special subset of loss firms. By indicating that the GAAP loss is attributable to expenses that are not infor- mative for
forecasting and valuation, managers may signal that their firm differs from other loss firms for which the loss provides a
better indication of (weaker) future performance. This logic implies that the non-GAAP dis- closure should also indicate
that the set of loss-converting firms is expected, on average, to have stronger future performance and a greater likelihood of
loss reversal.

Table reports mean future performance levels for loss converters versus matched GAAP-only loss firms. For all three
measures, loss converters perform significantly better than their matched counter- parts.
These results are consistent with the informative nature of loss convert- ers’ non-GAAP disclosures and suggest these firms
significantly outperform other loss firms that have similar characteristics but do not disclose non- GAAP earnings.
Moreover, tests of the sensitivity of these results to potential hidden bias reveal that if the differences in future performance
outcomes are attributable to omitted variables, these variables would have relatively strong effects on both the disclosure of
non-GAAP profits and the future performance variables. Although we cannot definitely rule out that endo- geneity bias
affects our results, and the results should be interpreted with that caveat in mind, we conclude that endogeneity bias is
unlikely to have a major qualitative influence on our inferences.

MISPRICING TESTS

Although our results for loss converters suggest that non-GAAP earnings are highly predictive of future firm performance
and valued by investors, while GAAP earnings are not, it is still possible that investors attach valu- ations that are too high
to firms that convert GAAP losses into non-GAAP profits. This evidence would be consistent with concerns that this type
of non-GAAP disclosure inflates users’ perceptions of firm performance and value

Results are similar to those based on raw returns and suggest that loss converters are not overvalued.

An alternative and potentially more powerful way to examine mispricing is to test for predictable differences in returns
around future earnings an- nouncements.

These results indicate a reversal of returns and suggest an initial overweighting of the implications of non-GAAP earnings
for the future.40 On the other hand, the insignificant coefficients on the exclusion surprise variable suggest that investors do
not overestimate the implications of exclusions for future earnings. This result is consistent with the result that exclusions
are not predictive of future firm performance, consistent with their exclusion in the calculation of non-GAAP earnings.

Conclusions

Consistent with the argument that the aggregation of items with different persistence reduces the predictive ability of
GAAP earnings for future performance, we find that non-GAAP earnings disclosures are highly informative about the
nature and implications of GAAP losses. Specifically, for firm-quarters without non-GAAP earnings (GAAP-only loss
firms), we find that GAAP earnings are strongly predictive of future firm performance. However, for firms with GAAP and
non-GAAP losses, we find that GAAP earnings are less than half as predictive compared to the GAAP earnings of GAAP-
only loss firms. This result is even stronger for firms that exclude expenses to convert a GAAP loss into a non-GAAP profit
(loss converters), where we find GAAP earn- ings to be unrelated to future performance. For both sets of loss firms, non-
GAAP earnings are strongly associated with future performance.

Inconsistent with prior evidence on the strategic disclosure in combined samples of profit and loss firms, we do not find that
the non-GAAP exclusions of loss firms are predictive of future performance. In fact, we find that the items excluded from
GAAP earnings are so “noisy” for some firms (i.e., loss converters) that they completely eliminate the predictive ability of
GAAP earnings for future performance. Additional tests of ERCs provide similar insights. Moreover, we find that loss
converters have substantially better future performance than matched GAAP-only loss firms, and we do not find evidence
consistent with concerns that these firms are overvalued. These results suggest that managers effectively signal that their
firms’ losses are attributable to expenses that are less informative for forecasting and valuation. Using additional data on
non-GAAP earnings disclosures of profitable firms, we find that non-GAAP disclosures have significantly greater
incremental predictive ability and are less strategic for loss firms.

Tutorial 2

Topic: Capital Markets and Financial Reporting

• Pricing of accruals
• Alternative performance measures

From accounting perspective we don’t tell investors what strategy to use. But our concern is how to make market reliable.

Today we want to understand first part better and take away how accounting works for this process.

By comparing the measures we look at where and how accounting can help us.
What does performance mean? The usefulness of earnings. How does our wealth grow. Non-accounting measure.

Every paper has a story part (what, why) and empirical part (how)

What – if what accountants do, is it useful and how it happens.

How – statistics of interest – gives the answer whether in improves our ability use performance

Dechow (1994): “Accounting Earnings and Cash Flows as Measures of Firm Performance:
The Role of Accounting Accruals”
Comparison of accounting performance measures based on cash flows and accruals

Introduction
Nichols and Wahlen (2023) also foundthat the relation between earnings and returns was stronger for news in earnings than
for news in cash flows
• But why is this so?

Main takeaway from the picture – there are positive and negative changes in earnings. If we start one year before the
earnings are announced, we can predict how earnings are going to change and receive higher returns as an investors.
Why is that the case: there is the relation between earnings and returns. Earnings predict better than EBITDA. We want to
understand why.
• The stock market typically focuses on estimates of accrual-based earnings or revenues, instead of cash flows
• This also suggests that investors perceive performance measures based on accruals to be useful summary measures of
performance
• But again, why?
• Dechow (1994) answers this question using a careful explanation of what accounting accruals do, combined
with empirical evidence
The paper explains the usefulness of earnings.

• Dechow (1994, p. 3) “investigates circumstances under which accruals are predicted to improve earnings’ ability to
measure firm performance, as reflected in stock returns”
Earnings = Cash flows + Accruals (adjustments)
• Microsoft’s June 2019 operating section of the cash flow statement
• Cash from operations (CFO) = Net income – Operating accruals (OA)
Where OA consists of:
• Long-term: depreciation, stock-based compensation
• Short-term: changes in accounts receivables,
inventories, accounts payable, etc.
If you work at the store, you can count in the beginning of the year and at the end, the difference will be your profit-
performance. But usually accounting is more complicated, so there are adjustments-accruals (depreciation, account
receivable). In practice Net income is used more often than Net cash from operations.
• Earnings are important, because they are used by many stakeholders as a summary measure of firm performance
• Executive compensation plans  contracting role of earnings
• Debt covenants  contracting role of earnings
• IPO prospectuses  valuation role of earnings
• Investors and creditors  valuation role of earnings
• Key view in this paper (Dechow 1994, p. 4): “the primary role of accruals is to overcome problems with
measuring firm performance when firms are in continuous operation”, while “[i]nformation
asymmetries between management and other contracting parties create a demand for an internally
generated measure of firm performance to be reported over finite intervals.”
In practice Net income is used more often than Net cash from operations. Stakeholders choose Net cash from operations.
We need to measure performance measure over finite intervals.
If we know why these adjustments are useful, we understand the paper.
Theory
1. In her study, Dechow (1994) examines whether, and how, accrual accounting improves the ability of
earnings (or net income) to measure a company’s periodic performance. One of her primary
conclusions is that (p. 26) “accruals are performing a useful role in mitigating timing and matching
problems in cash flows, so that earnings better summarizes firm performance.” Explain what she
means with “timing” and “matching” problems in cash flows, and why these problems are fixed by
accrual accounting.

• Over “finite intervals” (e.g., a year or quarter), cash flows are noisy measures of firm performance
because of the existence of “timing” and “matching” problems
• Accounting principles (GAAP or IFRS) have evolved such that accruals alter the timing of the
recognition of cash flows in earnings to fix the problems
Revenue recognition principle:
• Recognize revenue when all (or at least most) of the services have been provided (i.e., the “performance
obligation” is met), not necessarily when cash is received
Matching principle:
• Recognize cash outlays as expenses in same period as in which revenues are recognized
Cash flow is noisy because of timing and matching problems. That is why earnings are better.
Revenue recognition is a solution to timing problem.
Timing and marching problems

Based on cash flow manager doesn’t do good, because of the timing and matching problems. Based on earnings, he does
good.

This is not a problem, if we would measure it in 2years.

This problem becomes bigger if operating cycle is longer than 2 years. According to accounting, it would be a lot of losses,
because of the long operating cycle.

It is not only about investments, but also for expenses.

The trade-off
• The trade-off: accruals are associated with discretion (see Dechow Dichev 2002 in week 1)
• Ideally, discretion in the recognition of accruals allows managers to signal their private information
about the firm’s cash-generating ability
• Example: capitalization instead of expensing of development costs (good signal) or recognition of a
goodwill impairment (bad signal)  this improves earnings as a performance measure
• But, discretion in the recognition of accruals also allows managers to manage earnings (up or down)
towards a desired level (“earnings management”)
• Example: pre-mature recognition of revenues or incorrect capitalization of expenses  this makes
earnings less reliable as a performance measure
• Conventions in accounting standard setting therefore limit the ability of management to use their
discretion to manage the recognition of revenues and expenses:
• Trade-off between relevance and reliability” (p. 5)

Research design X variable

2. To test this, she compares the ability of earnings and cash flow measures to measure performance using
statistical analyses. In simple words, explain how she empirically determines which performance
measure is better. What is the primary statistic of interest she uses to draw this conclusion?
Is earning better, even though it is less reliable. What is the difference between operating cash flows and net cash flow. Net cash flow
includes financing, investing. Operating cash flow only includes core cash flows.

Predictions (how - we look at the R^2 of regression above)

We are looking whether data confirms intuition.
• Do the benefits of accruals outweigh the drawbacks?
• If “yes”, given the existence of timing and matching problems over finite measurement intervals,
earnings should better relate to a benchmark measure of performance, than does a cash flow measure
H1: There is a stronger contemporaneous association between stock returns and earnings than between
stock returns and realized cash flows over short measurement intervals
H2: The contemporaneous association of stock returns with realized cash flows improves relative to
the contemporaneous association of stock returns with earnings as the measurement interval is
increased
• These are “alternative” hypotheses: the underlying “null” hypothesis predicts that these effects do not
exist, because of frictions like earnings management
If we do not find association, it means that people don’t trust accounting

• For firms in a “steady state”, cash flows are a relatively useful measure of firm performance because
cash flows have fewer timing and matching problems
• These are firms with relatively stable requirements for working capital, investments, and financing
 accruals are small in magnitude
• Instead, for firms in volatile environments, timing and matching problems are more severe, and are
more likely to reduce the ability of cash flows to measure firm performance
• These are firms with large changes in working capital, investment, and financing needs
 accruals are larger in magnitude
The more volatile – the more accurate.
H3: The larger the absolute magnitude of aggregate accruals made by a firm, the lower the
contemporaneous association between stock returns and realized cash flow relative to the association
between stock returns and earnings

• As the ship building example suggested, the operating cycle plays a role as well
• Longer operating cycle: larger and more volatile working capital needs for a particular operating activity
• In industries with relatively long operating cycles, cash flows will be relatively less useful in measuring firm
performance because timing and matching problems are greater
H4: The longer a firm’s operating cycle, the more variable the firm’s working capital requirements and
the lower the contemporaneous association between stock returns and realized cash flows
• The ability of earnings to measure firm performance should not be affected by the operating cycle as
much as the cash flows

Question 3: Dechow (1994) uses a company’s stock returns over the year as a benchmark measure of company performance
to assess which accounting measure better captures performance over the year. Do you agree with this choice to use stock
returns as the benchmark? Why, or why not?
How do we know which measure is the best. The author believes it is stock market.

Research design: Y variable

Dechow (1994) uses a company’s stock returns over the year as a benchmark measure of company
performance to assess which accounting measure better captures performance over the year. Do you
agree with this choice to use stock returns as the benchmark? Why, or why not?
• How to test whether earnings are a superior measure of firm performance?
• We need a benchmark for this evaluation
• Here, Dechow (1994) has chosen to use stock price performance as the benchmark
• Key assumption (p. 12): “paper assumes that stock markets are efficient in the sense that stock prices
unbiasedly reflect all publicly available information concerning firms’ expected future cash flows”

• Question: if we assume the price is right, why do we need an accounting measure of performance?
• We know that the price is not always right
• Dechow (1994) assumes a semi-strong efficient market for the average company
• Also, she relies on the evidence from many prior studies ( see Nichols and Wahlen 2004 that stock
prices appear to respond rapidly to new information in earnings
• Still: important to recognize that price reactions are not always complete (as the evidence on post-
earnings announcement drift (PEAD) suggest, see Nichols and Wahlen 2004)
• We discuss related evidence in week 3

If we have returns, why do we need accounting. New information is relevant, which is in line with Nichols and Wahlen
2004 results.

Results
The higher R^2, the better the measure. Investors prefer earnings than cash flows. Accruals are relevant.

In long run, cash flows become better, because there is less timing and matching problems.

H3 and H4 confirm above mentioned results by looking at firm specific data. When there are no timing, we do not need
accountants.

Operating cycles effect cash flows, but not earnings.

• Remember from Dechow and Dichev (2002):

• The results reveal that companies with low-quality accruals ...
• ... have longer operating cycles
• ... are smaller
• ... have more volatility in their revenues, cash flows, accruals, and earnings
• ... more frequently report losses
• ... and have larger absolute working-capital accruals
• So, didn’t Dechow and Dichev (2002) tell us that accrual quality is lower when the operating cycle was
longer and when companies have larger absolute working-capital accruals?
• Yes, and that is still the case: accruals still help because cash flows suffer from greater timing and matching
problems in these situations
• At the same time, however, the larger the accrual adjustments, the greater the potential for estimation errors
• This is what the data show us
The more accruals you have, the more estimates you have to make. The key moments, investor rely more on relevance,
rather than reliability.

Summary:

• Accruals improve the association of earnings with stock returns

• Table 2: negative relation accruals and cash flows suggests that accruals smooth out fluctuations in cash flows
• Table 3: cash flows have a weaker association with returns  evidence of the timing and matching problems
• Effect is weaker with longer performance-measurement windows
• Table 5: when the magnitude of accruals increases, suggesting large changes in operating, investing, and
financing activities, cash flows suffer more severely from timing and matching problems
• Key quotes from Dechow (1994):
• P26: “accruals are performing a useful role in mitigating timing and matching problems in cash flows, so that
earnings better summarizes firm performance”
• P27: “consistent with the hypothesis that accountants accrue revenues and match expenditures to revenues so
as to produce a performance measure (earnings) that better reflects firm performance than realized cash flows”
• P28: “suggests that manipulation of accruals is of second-order importance and the first-order effect of the
accrual process is to produce a summary measure that more closely reflects firm performance”
Leung and Veenman (2018): “Non-GAAP Earnings Disclosure in Loss Firms”
Presentation of adjusted earnings measures by companies reporting negative earnings

4. Leung and Veenman (2018) examine companies’ presentation of alternative earnings metrics, also
known as “non-GAAP” earnings measures. Explain why the Leung and Veenman (2018) study can
provide useful insights about the changes in relevance of accounting information over time.
Why would we use alternative performance indicators.

Non-GAAP reporting
• Why should we care about non-GAAP reporting?
• What does non-GAAP reporting tell us about the changing value relevance of accounting over time?
• Answer: the increasing use of non-GAAP measures of performance by managers and users suggests that
GAAP / IFRS earnings are perceived to have become less useful in evaluating company performance
• Note the quote from Lev (2019): “‘Why else would we voluntarily provide extensive non-GAAP information?’ asked me
recently a CEO. ‘Financial reports today are essentially a compliance [with GAAP rules] exercise, not really intended to
inform investors.’”
• Linsmeier (2016) quote from Leung and Veenman (2018, p. 1089): “EPS ... increasingly may not be serving allusers’
needs”
• Financial statement users demand additional information needed to evaluate company performance
The increase in non-GAAP reporting increased. This leads to a question why GAAP reporting is less useful. Where we use
GAAP, we need to follow rules. Which means that we have to use information, that is not always useful to predict future
returns. We want to know what makes GAAP not useful. GAAP is compliance with rules, and not for providing useful
information to investors.

Non-GAAP reporting by loss-making companies

• Lev (2019, p.734): “Nothing demonstrates better the consequent irrelevance of financial information
than the fact that almost half of U.S. firms reported losses in 2016–2017 (and 70% of high tech and
science-based firms were also ‘losers’), during a booming economic period.”
• P. 718 “For almost half of the high tech and science-based firms the ‘loss’ was caused by the expensing
of intangible investments.”
• From a user perspective, this is important because losses tend to be much less persistent than profits
(similar to earnings decreases, see Nichols and Wahlen 2004)
• That is, losses have different implications for performance evaluation than profits
We see that economy is doing okay, but a lot of firms are reporting loss. This was caused by investments in intangibles and
R&D.

Marked items show why there is a problem with GAAP reporting’s, because these items have to be amortized.

GAAP – loss, non-GAAP – profit

You can see that if you add stock profit compensation, the GAAP would show a loss. This would not be very useful.

Research Question
Prior Research:
• Non-GAAP earnings are incrementally informative about future performance
• Non-GAAP earnings are disclosed opportunistically to meet certain benchmarks
Prior Research: earnings are not informative for (GAAP) losses due to certain expenses (amortization)
• Non-GAAP earnings often exclude these expenses (“exclusions”)
• Exclusions are accruals that are not considered informative
Research Question: are non-GAAP earnings informative for loss firms?
• Distinction in GAAP losse (4 groups):
• GAAP Loss & non-GAAP Loss
• GAAP Loss & non-GAAP Profit [“loss converter”]
Trade-off non-GAAP relevant, but potentially not reliable. Are these specific non-GAAP measure informative for law
firms.

Research design

If it helps us we have evaluate future performance better and if it will increase stock price.

We want to look at the coefficients of non-GAAP earnings and Exclusions.

Data

Firms that disclose non-GAAP are divided on profit and loss.

• From Table 2-B, we also learn that most of these companies that report a non-GAAP earnings measure
ignore recurring expenses like stock-based compensation and the amortization of intangible assets
• Non-GAAP loss firms:
• 60.0% ignore stock-based compensation
• 36.8% ignore intangible amortization
• Loss converters:
• 72.0% ignore stock-based compensation
• 59.6% ignore intangible amortization
• For example, ignoring intangibles amortization is consistent with Warren Buffett (see lecture), who
argued that these expenses are not informative about company performance
Results
Dependent variable Future cash flows. Cash earning predict future earnings significantly. GAAP earning are not significant
for loss firms. Non-GAAP information is predictable.

For profit firms we don’t find this effect.

Implications of accrual accounting

5. Leung and Veenman (2018) find that adjusted earnings measures, from which managers exclude
expenses such as stock-based compensation and intangible amortization, have stronger correlations
with future company performance than the GAAP earnings of loss-making companies. They also find
that these adjusted measures are more highly correlated with stock market returns. Explain whether
you believe this result is consistent, or inconsistent, with what we have learned from Dechow (1994)
and Nichols and Wahlen (2004).
Not consistent with previous paper. Because accruals are not always better.
• Results suggest that for these loss-making companies, non-GAAP earnings have stronger relations with
future performance outcomes than GAAP earnings
• In other words: non-GAAP earnings have higher persistence than GAAP earnings
• Most of the non-GAAP measures ignore stock-based compensation and/or intangible amortization
• When ignored from non-GAAP measures, results indicate these expenses have little predictive value
• In the framework of Nichols and Wahlen (2004), less persistent earnings should trigger weaker stock
market reactions to earnings announcements
• Results from the ERC tests in this paper confirm this  Consistent with Nichols and Wahlen (2004)

• Important: stock-based compensation and/or intangible amortization are accruals!

• These are non-cash expenses that have a negative effect on earnings (i.e., negative accruals)
• GAAP earnings (which includes these accruals) has a weaker relation with stock market valuations and
future performance outcomes, than non-GAAP earnings (which ignores these accruals)
• This result is not consistent with Dechow (1994), because her arguments and results suggest that accruals
should improve the usefulness of earnings as a performance measure, and hence increase the association of
earnings with stock returns
• P1089: “This evidence also contributes to the literature on the role of accruals in forecasting cash flows
(Dechow [1994] ....), since it illustrates how certain accrual components reduce, rather than enhance,
the informativeness of GAAP earnings for forecasting and valuation.”

Summary

• Results for loss-making companies reveal that many companies disclose non-GAAP earnings measures
because their GAAP earnings are not value relevant
• On average, the non-GAAP earnings measure provide substantial incremental information above the
official GAAP earnings measures
• For these companies, many of the expenses excluded from non-GAAP earnings are not value relevant
• Most often, these are stock-based compensation and the amortization of acquired intangible assets
• These results highlight the importance of understanding the declining value relevance of earnings and
accounting information over time
• Non-GAAP disclosures provide information to standard setters on how to “fix” this issue
• And, inconsistent with Dechow (1994), accruals can also sometime decrease the value relevance of earnings

Lecture 3 Sloan (1996)

Do Stock Prices Fully Reflect Information in Accruals and Cash Flows About Future Earnings?

This paper investigates whether stock prices reflect information about future earnings contained in the accrual and cash
flow components of current earnings.

Texts on financial statement analysis frequently advocate examining the acc flow components of current earnings for the
purpose of predicting future earnings.

The results indicate that earnings performance attributable to the accrual component of earnings exhibits lower persistence
than earnings performance attributable to the cash flow component of earnings. The results also indicate that stock prices
act as if investors "fixate" on earnings, failing to distinguish fully between the different properties of the accrual and cash
flow components of earnings. Conse- quently, firms with relatively high (low) levels of accruals experience negative
(positive) future abnormal stock returns that are concentrated around future earnings announcements.

DEVELOPMENT OF HYPOTHESES

The importance of analyzing the accrual and cash components of current earnings in the assessment of future earnings is
frequently emphasized in texts on financial statement analysis.

The common theme underlying this reasoning is that the accrual and cash flow components
of current earnings have different implications for the assessment of future earnings. While both components contribute to
current earnings, current earnings performance is less likely to persist
if it is attributable primarily to the accrual component of earnings as opposed to the cash flow component. For example,
high earnings performance that is attributable to the cash flow component of earnings is more likely to persist than high
earnings performance that is attributa to the accrual component of earnings. This reasoning forms the basis for the first
testable hypothesis:

Hi: The persistence of current earnings performance is decreasing in the magnitude of t accrual component of earnings and
increasing in the magnitude of the cash flow component of earnings.

The remaining hypotheses concern the extent to which stock prices reflect the different properties of the accrual and cash
flow components of earnnings

A meaningful test of whether stock prices fully reflect available information requires the specification of an alternative
"naive" expectation model, against which to test the null of market efficiency. The naive model employed in this study is
that investors "fixate" on earnings and fail to distinguish between the accrual and cash flow components of current earnings.
This naive earnings expectation model is consistent with the functional fixation hypothesis, which has received empirical
support in capital markets, behavioral and experimental research (Hand 1990; Abdel-khalik and Keller 1979; Bloomfield
and Libby 1995). This model is not as restrictive as
the random walk model implicit in Ou and Penman (1989) and Bernard and Thomas (1990). Earnings expectations are
permitted to reflect the overall level of persistence in earnings performance, but are hypothesized not t able to the accrual
and cash flow components of earnings. The second hypothesis is then:

H2(i): The earnings expectations embedded in stock prices fail to reflect fully the higher earnings persistence attributable to
the cash flow component of earnings and the lower earnings persistence attributable to the accrual component of earnings.

This hypothesis makes predictions about both the direction and the magnitude of deviations in the expectations embedded
in stock prices from the actual relationships. Two extensions of this hypothesis provide corroborative evidence on the
extent to which stock price behavior deviates
from the rational expectations model. The first extension develops a trading strategy to exploit the naive earnings
expectations embedded in stock prices, providing insight into the economic significance of deviations from the rational
expectations model. If investors naively fixate on
earnings, then they will tend to overprice (underprice) stocks in which the accrual component is relatively high (low). This
occurs because the lower persistence of earnings performance attributable to the accrual component of earnings is not fully
anticipated. The mispricing will b corrected when future earnings are realized to be lower (higher) than expected, resulting
in predictable negative (positive) abnormal stock returns. A simple strategy that exploits this mispricing can therefore be
implemented as follows:4

H2(ii): A trading strategy taking a long position in the stock of firms reporting relatively low levels of accruals and a short
position in the stock of firms reporting relatively high levels of accruals generates positive abnormal stock returns.

The above extension of the stock price hypothesis concerns the sign and magnitude of abnormal stock returns resulting
from naive fixation on earnings. A second extension of the stock price hypothesis relates to the timing of the abnormal
stock returns resulting from naive fixation on earnings. If the abnormal stock returns represent a delayed response to
predictable changes in future earnings, then they should be concentrated around information events that reveal the
predictable earnings changes, such as future earnings announcements. Thus, the second extension
of the naive expectations model is that the predictable stock returns will be clustered around future earnings
announcements:

H2(iii): The abnormal stock returns predicted in H2(ii) are clustered around future earnin announcement dates.

Bernard and Thomas (1990) conduct a similar test in their examination of the post-earnings announcement drift. Consistent
with their naive-expectations model, they find that almost 40 percent of the drift is clustered around future earnings

announcements.

The computation of abnormal returns requires adjustment for the normal or expected return. Two alternative adjustment
procedures are employed in this study. First, size is a well- documented predictor of future returns, and prior research in this
area typically employs a size adjustment (Ou and Penman 1989; Bernard and Thomas 1990). In this study, size-adjusted
returns are computed by measuring the buy-hold return in excess of the buy-hold return on a value- weighted portfolio of
firms having similar market values. The size portfolios are formed by CRSP and are based on size deciles of NYSE and
AMEX firms. Membership in a particular portfolio is determined using the market value of equity at the beginning of the
calendar year in which the return cumulation period begins.

The second adjustment procedure estimates Jensen alphas at the portfolio level using the technique first suggested by
Ibbotson (1975). The procedure involves estimating the following time-series regression separately for each portfolio for
each of the three years in the evaluation period.

EMPIRICAL ANALYSIS

Tests of Hi

the relation between current earnings performance and future earnings performance can be expressed as:

Table 2 reports results from the estimation of

equation (4) to establish the average level of persistence in earnings performance
 Overall, the results in table 2 confirm previo evidence that accounting
rates of return are mean reverting, with an average persistence parameter, ax,, of approximately 0.8.

Table 3 provides parameter estimates for equation (5), which does not constrain the persistence coefficients on the accrual
and cash components of earnings to be equal. The first column of panel A provides the pooled regression results. The y,
coefficient is less than the y2 coefficient in 99 percen of the industries, and the null hypothesis of equality is rejected using
a sign. the results in table 3 therefore provide strong evidence in support of H1.

Tests of H2(i)
The likelihood test for market efficiency is 0.007 (marginal significance level = 0.933) and the null hypothesis of market
efficiency is not rejected. The results in panel B yield much the same conclusions. The persistence of earnings is somewhat
weaker using the decile rankings, but is still anticipated rationally in the pricing equation. These results can be interpreted
as indicating the absence of a post-earnings announcement drift in annual earnings. Stock prices correctly reflect the
implications of current annual earnings for future annual earnings.

uares in table 3. efficiency implies that the different implications of the accrual and cash flow components of
current earnings for future earnings should be reflected in stock prices. However, the results from
the stock return equation reveal that this is not the case. The coefficient on accruals, <,is 0.91 1,
while the coefficient on cash flows, y*, is 0.826. Thus, the coefficient on cash flows is smaller than its counterpart in
forecasting equation, while the coefficient on accruals is larger than its counterpart in the forecasting equation.

Stock prices do not appear to anticipate rationally the lower (higher) persistence of earnings performance attributable to the
accrual (cash flow) components of earnings. The likelihood ratio statistic is 180.91, rejecting the null hypothesis of market
efficiency.

Overall, the results in table 5 indicate that stock prices act as if investors fail to anticipate fully the lower (higher)
persistence of earnings performance attributable to the accrual (cash flow) component of earnings. The earnings
expectations embedded in stock prices consistently deviate from rational expectations in the direction predicted by naive
fixation on earnings. However, precise inferences as to whether the magnitudes of the deviations from rational expectations
are consistent with a naive fixation on earnings are sensitive to model specification.

Tests of H2(ii)
The remaining three columns in table 6 provide portfolio abnormal returns measured using Jensen alphas. The results are
generally consistent with those obtained using the size adjusted returns. The hedge portfolio return to a long position in the
lowest accrual portfolio and a short position in the highest accrual portfolio is 10.4% (t=4.42) in the first year, 4.8%
(t=2.41) for the second year and 3.8% (t= 1.62) for the third year. 17 The results in table 6 demonstrate the economic
significance of investors' apparent inability to distinguish correctly between the accrual and cash flow components of
earnings.

Figure 2 provides evidence on the stability of the abnormal

returns to the trading strategy. It plots the annual hedge portfolio return for each of the 30 fiscal years in the sample. The
returns used to produce the plot are size-adjusted returns from the first year subsequent to portfolio formation.
Consequently, the average of the 30 yearly returns corresponds to the hedge portfolio return of 10.4% reported in the first
column of results in table 6. The hedge portfolio return is positive in 28 of the 30 years examined, illustrating that the
relation is fairly stable over time. The only exceptions are 1966, when the return was -19.5%, and 1981, when the return
was -2.2%. The fact that the returns are positive in over 90 percent of 30 years examined helps rule out risk- based
explanations.

Panel C of table 7 regresses stock returns on the accrual component of earnings and a variety
of other variables that have been shown to predict future stock returns (Fama and French 1992). The objective of the
regressions is to demonstrate that the predictive ability of accruals is not subsumed by these other variables. The additional
variables considered are size (measured as the natural logarithm of the market value of equity), book-to-market (measured
as the log of the ratio
of the book value of equity to the market value of equity), historical beta (measured by estimating the market model on the
prior 60 monthly stock returns) and earnings-to-price (measured using the ratio of earnings per share to fiscal year end
stock price). The magnitude and statistical significance of the coefficients on accruals are very similar in panel C to those
reported for the univariate regressions in panel A. Thus, the ability of accruals to predict future returns is incremental to
these previously documented effects.

Tests of H2(iii)
The results from the announcement period tests are presented in table 8. The first column reports the total annual size-
adjusted return in the year following portfolio formation. The negative relation between accruals and abnormal stock
returns is clearly evident and the magnitudes of the abnormal returns are almost identical to those reported in table 6. The
next two columns of table 8 report the amounts of the total annual return that are attributable to the announcement and non-
announcement period respectively. The negative association between operating accruals and abnormal stock returns is
evident in both the announcement and non announcement periods. Focusing first on the hedge portfolio returns, the
announcement period return is 4.5%, while the non-announcement period return is 6.0%. Thus, over 40 percent of the
predictable stock returns are concentrated around the subsequent quarterly earnings announce- ments, even though the
announcement period contains less than five percent of the total trading days. Figure 3 plots the hedge portfolio return for
each of the 19 calendar years in the sample. The returns are positive in all 19 years, which is inconsistent with the returns
providing compensation for risk. These results are therefore consistent with a delayed price response to information in
accruals and cash flows about future earnings.

The individual portfolio results reveal additional information. The lowest accrual portfolio,
for which good earnings news is forecast, has an announcement period return of 4.5% and a non- announcement period
return of 0.9%. Thus, over 80 percent of the predictable stock returns are concentrated around subsequent earnings
announcements for this "good news" portfolio. The highest accrual portfolio, for which bad earnings news is forecast, has
an announcement period return of 0.0% and a non-announcement period return of -5.1%. Thus, essentially none of the
predictable stock returns are concentrated around the subsequent earnings announcements for this "bad news" portfolio. The
asymmetry of the results for good and bad news announcements is consistent with the evidence in Chambers and Penman
( 1984) and Skinner (1994) that bad news earnings announcements are more likely to be preempted. Evidence of this nature
is provided by the proportion of late reporters statistics in the final column of table 8, where 34.5% of the firm- years in the
highest accrual portfolio miss at least one quarterly earnings announcement compared to 28.5% for the lowest accrual
portfolio. These results are consistent with preemption of earnings announcements causing some of the predictable stock
returns to be realized in the non- announcement period.

CONCLUSIONS

This paper investigates whether stock prices reflect information about future earnings contained in the accrual and cash
flow components of current earnings. The persistence of earnings performance is shown to depend on the relative
magnitudes of the cash and accrual components of earnings. However, stock prices act as if investors fail to identify
correctly the different properties of these two components of earnings.

The stock price results are inconsistent with the traditional efficient market's view that stock prices fully reflect all publicly
available information. However, the finding that stock prices do not fully reflect all publicly available information does not
necessarily imply investor irrationality
or the existence of unexploited profit opportunities. The information acquisition costs and processing costs associated with
implementing the strategy outlined in this paper in real time are non-trivial. Moreover, the returns to exploiting the strategy
are potentially limited by price
pressure effects. Observing a historical trade price on the CRSP tapes does not imply that unlimited quantities of the stock
could have been traded at that price. Perhaps the results in th paper are simply evidence of a normal return to an active
investment strategy based on financial
statement analysis.
The results raise additional issues for future research. Of particular interest is the extent to which the lower persistence of
earnings performance attributable to the accrual component of earnings is due to earnings management. Dechow et al.
(1995) examine a sample of earnings manipulations subject to SEC enforcement actions and find that these earnings
manipulations are primarily attributable to accruals that reverse in the year following the earnings manipulations.
Thus, their evidence is consistent with earnings management contributing to the lower persistence
of the accrual component of earnings. Related issues include establishing whether earnings management is made with the
intent of temporarily manipulating stock prices and the motivations for any such stock price manipulations

Lecture 3

Talk about how well users understand and process accounting information.
Users are investors who invest in equity, debt investors, stakeholders. We focus primely on investors.
Focus on paper which talks about are markets effective. KNOWLEDGE CLIP – trading part.

Started week 1: we need to understand reporting – what is quality, errors.

Week 2 – capital market, users of market and how they understand reporting, alternative performance measures.
Week 3 – to what investors don’t understand the process correctly, pricing problems.

Start with businesses that are selling products and services.

Accountants are there to provide a high-quality report.
When we evaluate the report, we look at how the information is used.
And we can improve fin reports to be useful to people who use it.
If people do not use the report, it is not useful.
Share prices are useful.

Low quality – harder to predict future earnings, less earnings.

The lower the quality of earnings, persistence also becomes lower.
We care about it because predictive ability is one of the goals of accounting.

Key issue for this week: how investors process information (there are problems, they do not have enough expertise)
Look at the issue in persistence of accruals and cash flows.

Remember
Earnings matter as a performance measure (Nichols and Wahlen 2004; 2023)
-Accruals (through the revenue recognition principle and matching principle) help to make earnings a better measure of
performance than cash flows
- Greater predictive value for future performance
- Stronger relations with changes in stock price

-But earnings, and components of earnings, vary in their persistence(predictive value) and therefore vary in their value
relevance
- For example, the expenses excluded from the calculation of non-GAAP earnings were less persistent and less value
relevant (Leung and Veenman 2018)
-Sometimes, some accruals are not persistent or predictive of future performance

Users’ processing of accounting information

Recall that the IASB objective is that financial reporting provides information that is “useful to users”
-In making investment decisions, providing/settling loans with a company, or influencing management’s actions

But this assumes that these users fully understand how to interpret the accounting information
-For example: it assumes that users understand the difference between:
- A loss that results from a goodwill impairment, which is likely to be non-recurring (transitory)
- And a loss that results from intensified competition in a company’s core business, which likely has a persistent effect
on the company’s future earnings expectations
There are fin errors, line items. And all these changes have pricing implications. And if investors are not very
knowledgeable, it is not got enough for them.
They cannot read the statements and say whether they want to invest or not. If they do not understand it, they use shortcuts
to process information.

It also assumes that just providing the information to users is sufficient

But in practice we know this is not the case: the amount of information presented to users, as well as the way in which it is
communicated, can make a difference

• Quote from Hirshleifer and Teoh (2003, Journal of Accounting and Economics, p. 338):
“Firms and regulators care not just about the information made publicly available to investors, but the form in which it is
revealed. One issue of great concern to practitioners is whether information items should be recognized as part of earnings,
or merely disclosed as a footnote. Another is the prominence with which different kinds of information are displayed in
financial statements. There is also intense concern as to the form of disclosure, even when the information content of the
alternative formats is identical. Evidently regulators and commentators think that investors are imperfect processors of
publicly available information. Such concerns are reflected in the structure of accounting regulation, and in politically
charged debates over such issues as merger accounting, whether employee option compensation should be expensed, and to
what extent firms should be free to make pro forma disclosures that differ from GAAP definitions of earnings.”
We look how feasible the information is presented. We take the position of the regulator, and we know that investors are
stupid. So we try to fix it and find problems.

Illustration, in 2003 Apple had the option to either (1) recognize stock-based compensation as an expense in the income
statement, thereby reporting low net income, or (2) not recognize this expense in the income statement and only disclose the
what-if effect in the notes to the financial statements
Apple chose for option (2), but why? The same information is reported, but in different locations
Income statement (p. 57): Notes (p. 67):

We want to see if it clear to all users if some information is put on the income statement or notes.
The goal of this class is to discuss and identify:

Whether this location choice is likely to make a difference. In other words, to assess whether the market is not fully
efficient when processing accounting information

And if so, why this choice of where (and how) to report accounting information makes a difference
Users’ processing of accounting information

Sloan (1996) “Do Stock Prices Fully Reflect Information in Accruals and Cash Flows About Future Earnings?”

When we see information on the screen, we should not assume that it is correct. We should make our own research, so we
can earn money by better understanding information.

Introduction
Read the income statements carefully! Investors do not look at earnings components, they look at the total earnings at the
end and they are missing information.
Is there important information in earnings components. When is it important? => if it helps you predict future earnings,
because it related to future value of the company. If there is a difference in predictive value. The next question is do
investors understand it?
“Texts on financial statement analysis frequently advocate examining the accrual and cash flow components of current
earnings for the purpose of predicting future earnings. Indeed, some financial analysts argue that since investors tend to
"fixate" on reported earnings, analysis of this type can be used to detect mispriced securities.” (Sloan 1996, p. 289)
Recall: Earnings = Cash flow + Accruals
Research question #1: Do these accrual- and cash-flow components of earnings have different predictive value for future
earnings performance?
Research question #2: If the answer to question #1 is yes, do investors fully understand the difference in the nature of
information contained in cash flows and accruals?
This is different from Dechow (1994), because she looked at the comparison of cash flows vs. earnings
Sloan (1996) focuses on the comparison between cash flows and accruals
Although accruals help to make earnings (= cash flows + accruals) a better measure of performance, this does not mean that
accruals have greater predictive value than cash flows
In other words: accruals are not necessarily more persistent than cash flows

Last week we wanted to see whether earnings pr cash flow is better. CF leads to timing and matching problems. Even
though accruals(earnings) are better, does it mean they have better predictive value.

Theory and hypotheses

• From Dechow (1994) and other studies we know that accruals should make earnings a more informative (predictive)
performance measure for expectations of future performance, which implies:

Classical texts on financial statement analysis suggest that it is important to separately examine the accrual- and cash-flow
components of earnings, because some accruals are less likely to recur in future periods

We know that cash flows are less predictive value is less than earnings. So we look at which values earnings predict better.

Bernstein (1993) as quoted in Sloan (1996, p.291): “CFO (cash flow from operations), as a measure of performance, is less
subject to distortion than is the net income figure. This is so because the accrual system, which produces the income
number, relies on accruals, deferrals, allocations and valuations, all of which involve higher degrees of subjectivity than
what enters the determination of CFO. That is why analysts prefer to relate CFO to reported net income as a check on the
quality of that income. Some analysts believe that the higher the ratio of CFO to net income, the higher the quality of that
income. Put another way, a company with a high level of net income and a low cash flow may be using income recognition
or expense accrual criteria that are suspect.”
We hire accountants to check whether managers use accruals in the correct or wring way.

Even if we argue that managers use accruals in a good way (i.e., they do not use accruals for earnings management),
accruals are likely to be less persistent than cash flows because:

-Accruals rely on expectations of the future; uncertainty about

- For example: managers have to make estimates regarding expected future losses from uncollectible receivables or
make estimates on the rates of returns on pension plan assets
- In contrast: cash flow realizations are “hard” evidence – no uncertainty
-Many transitory events/transactions are recognized as accruals: asset impairments, restructuring charges with deferred cash
payments, etc.
We test if managers use accruals correctly. And we want to know if it helps investors process and understand information
better.
If there is subjectivity used in accruals, it will be less predictive than cash flows.

Based on this reasoning, Sloan (1996) predicts that the accrual-component of earnings is likely to be less persistent
compared to the cash-flow component of earnings
This can be true even given Dechow’s (1994) result that accruals make earnings a better performance measure

Two companies report the same earnings number. But in income statement company A reports more cash flows, but less
accruals. No uncertainty whether the customer will pay back.
Company B – a lot of accruals, so a lot of uncertainty whether the customer pays back.

H1: The persistence of current earnings performance is decreasing in the magnitude of the accrual component of earnings
and increasing in the magnitude of the cash flow component of earnings.
From a user perspective, this suggests that, looking at information from both the income statement and the cash flow
statement is important to better understand the sustainability of current performance

Question is why cash flows provide more predictability. Because with cash flows you do really need to estimate the amount
of money you will receive next year, because you received all the money this year, and there are no accruals.
People who are buying with credit (accruals), they are not very reliable. If economy become bad, they maybe will not be
able to pay you back.
SO how much can you use earnings statements to predict future earnings.
If investors do not understand the difference between these 2 companies, they cannot choose more profitable option –
option A. And can invetor actually benefit from investing in company A?
Now we need prove that these companies are not the same.

We know from Nichols and Wahlen (2004) that when earnings (components) have different persistence, they should also be
valued differently by investors
-But do investors in actually value these components differently?
-Do investors actually look at cash flow statement information, or do they just “fixate” on earnings (net income)?

Hence: “Do stock prices fully reflect information in accruals and cash flows about future earnings?”

Prior research has developed a “functional fixation hypothesis” (see references in Sloan 1996, p. 291)
-For example, Hand (1990, p. 740) states: “The traditional functional fixation view states that investors are always
unsophisticated and, therefore, fail to unscramble the true cash flow implications of accounting data.”
-In contrast: “the efficient market hypothesis states that investors are always sophisticated and very accurately
unscramble the true cash flow implications of accounting data.”

• Illustration of press coverage of Apple’s earnings (does not mention cash flow!):

https://

Under investor “fixation” on earnings, stock prices are expected to not fully reflect the higher persistence of the cash flow
versus accrual component of earnings:

H2(i): The earnings expectations embedded in stock prices fail to reflect fully the higher earnings persistence attributable to
the cash flow component of earnings and the lower earnings persistence attributable to the accrual component of earnings.
If reporting goes wrong, the analysis will be wrong , the forecast is wrong, the trading and value is wrong. We should see it
in stock returns after the announcement.

H1 is about earnings performance.

H2 is about earnings expectations, trading expectations, future returns.

If, on average, investors are indeed naive and fixate on reported earnings, hedge funds should be able to exploit this
mispricing with a trading strategy:
-Investors will overprice stocks for which accruals are relatively high, because they overestimate the persistence of
earnings
-Investors will underprice stocks for which cash flows are relatively high, because they underestimate the persistence
of earnings
As Sloan (1996, p. 292) explains: “mispricing will be corrected when future earnings are realized to be lower (higher) than
expected, resulting in predictable negative (positive) abnormal stock returns. A simple strategy that exploits this mispricing
can therefore be implemented as follows:”

H2(ii): A trading strategy taking a long position in the stock of firms reporting relatively low levels of accruals and a short
position in the stock of firms reporting relatively high levels of accruals generates positive abnormal stock returns.

Note:
-“Long” position means that we buy the stock in our portfolio
-“Short” position means that we sell stock that we do not have in our portfolio (the stock is borrowed, and subsequently
sold “short”)

Because the prediction is that this mispricing reflects investors’ delayed response to information about future earnings, we
should see that these abnormal stock returns are concentrated around future earnings announcements (i.e., when the
realizations of future earnings become known):

H2(iii): The abnormal stock returns predicted in H2(ii) are clustered around future earnings announcement dates.
Amsterdam Business School

Research design
How do we find prof for the story?
Panel dataset of > 40,000 firm-years: US companies 1962-1991
Important to understand earnings, because analysts focus on that.
Earnings is measured as “operating income after depreciation”
- Also known as “EBIT”: Earnings Before Interest and Taxes (not on income statements)
- Ignores special and extraordinary items, as well as non-operating income and expenses

Earnings = Cash flows + Accruals; but how should we empirically measure these variables?
- Key issue: for most of Sloan’s (1996) sample period, companies did not present a cash flow statement in their annual
report!
- Cash flows have to be computer indirectly using earnings (income statement) and accrual (balance sheet)
information: Cash flows = Earnings – Accruals

• Sloan (1996, p. 293): “the accrual component of earnings is computed using information from the balance sheet and income
statement”

1. Depreciation expense is a negative accrual: it lowers net income but does not affect cash flow
Research design
Accruals is the change income statement without cash component. Example – depreciation.

Also exclude s-t debt. Because it is not informative.

Assets go up – higher earnings, if all accruals are repaid.

Research design: scaling example

We want the results of a regression to capture an average representation of the population, and not be driven by just the
largest companies in the data

This is because large companies tend to have large (positive or negative) values of accounting variables, while small
companies tend to have small (positive or negative) values

For example, consider the relation between depreciation expense and earnings
Potential hypothesis are:
H1: Higher depreciation expense is associated with lower earnings  negative relation
H2 (null): Higher depreciation expense is not associated with earnings  zero relation
Descriptive statistics
10 portfolios created by sorting accruals from low (group 1) to high (group 10):

Large variation in accruals across companies:

In group 1, companies on average have negative accruals equal to 18% of their total assets
In group 10, companies on average have positive accruals equal to 15% of their total assets
A lot of different firms with negative accruals (investing), and positive accruals (when people owe you)

Descriptive
statistics

In terms of cash flows there are also differences. From week 2, the same conclusion, accruals and cash flows have negative
correlation.
We also see substantial variation in cash flows, but opposite to the variation accruals:
When accruals are low (group 1), cash flows are high
When accruals are high (group 10), cash flows are low
P295: “Consistent with prior research (Dechow 1994), there is evidence of a strong negative relation between accruals and
cash flows. The mean (median) value of cash flow falls from 0.22 (0.23) for the lowest accrual portfolio to 0.00 (0.00) for the
highest accrual portfolio.”
Descriptive statistics

• When accruals are split into its

We can see that the most important variation in accruals comes from the changes in (non-cash) current assets:
For example: changes in accounts receivables and changes in inventory

Results
To test H1 on the differential persistence of the accrual- and cash-flow components of earnings, Sloan (1996) first estimates a
baseline earnings persistence regression:

We want to show that different earnings components have different predictive value for the next year. Can we us ethe number
in this year income statement to predict a value in nest year income statement.
Here, provides an estimate of how persistent earnings are
Recall that persistence refers to the likelihood that current year earnings will recur in the future

Next, because
H1: There is a difference
We want to predict earnings at time T+1.
The idea behind the test, there should the accruals this year and next year
84% 1 euro this year will predict 84 cents next year. Not 1 year, because it is not fully persistent. It is difficult to be the best
firm, so the averaged value is used.
On average, one dollar/euro of earnings this year will recur as 84 cents next year

Result
Per industry, because outlier may differ per industry:
We can estimate the extent the firm will continue for many year=> 1 on 1 relation => fully persistent

.
We are interested if the value is the same for accrual is the same vale as cash flow.

Accrual 77%: 1 euro in accruals, we will ear 77 cents. So we will earn less in cash flows. Accruals are worse to have for
company.
One dollar/euro of accruals this year will recur as 77 cents next year; one dollar of cash flows will recur as 86 cents next year
 accruals are less persistent (more transitory) than cash flows
Differences in
 Differen
Result earnings
quic earnings
more
• Figure 1 shows how accruals mean-revert

High-accrual companies High-cash flow companies

If there are a lot of accruals this year, next year we expect to receive it back. We will not continue giving on credit. We will
first receive it back and then again give it on credit.
Cash flows stay the same for a longer time. So it is better to use them for forecasting.
Investors believe that those 2 dots are the same => they look only at earnings, and not components.

HOW? We care about earnings, to predict the future. Does the stock price reflect the difference in dots. Results: H2(i)
Note: details of the estimation approach (“Mishkin framework”) on pp. 302-303 can be ignored
To test whether investors understand this difference in persistence, Sloan (1996) uses a complicated estimation approach
This method quantifies the persistence of earnings that is expected by investors, as reflected in stock prices

We want to predict the earnings at t+1. If the market is efficient then both regressions should be the same.
Stock market is efficient in pricing the implications of annual earnings for future earnings

Result
• Tab
Acc
Cash

Investors expect 84%, the actual persistent is 84.1%

Accrual components – investors overestimate accruals. They underestimate the persistence of cash flows. Markets are not
efficient. Investors do noy understand that accruals are not the same as earnings.

For accruals, investors expect higher persistence (0.911) than the true persistence (0.765)
For cash flows, investors expect lower persistence (0.826) than the true persistence (0.855)
 Market efficiency is now rejected: investors do not fully understand the difference in persistence
 Consistent with investors “fixation” on earnings and not looking at the accrual and cash flow components

Results: H2(ii)
We buy companies that do not have accruals. If we have a lot of accruals, we
have a big negative surprise. If you but firms that have cash flows, you will get a
small positive surprise, So trade those. Trading strategy – accrual anomalies.
ALL done by estimating future returns. If this is done by using wrong
expectations, it should be confirmed after the announcement.
Can we make money by trading against investors’ misunderstanding of the
difference in persistence of accruals and cash flows?
Table 6 suggests yes: Supports H2(ii)
-On average, companies with low (negative) accruals have positive stock returns of 0.049 in the next year
-Companies with high (positive) accruals have a negative stock returns of -0.055 in the next year
-Thus, buying stocks of companies with low accruals, and selling (short) companies with high accruals, provides an average
“hedge” return of 0.104, or 10.4% per year

Results: H2(iii)
In support of H2(iii), a large part of the abnormal stock returns are realized
around future earnings announcements (next 4 quarters)
Hedge return for the full period (252 days): 11.2%
Hedge return around earnings announcements (4x3 days): 4.5%
Thus, about 40% of the returns are realized at the time when the true
persistence of earnings becomes publicly known

Sloan (1996): Summary and conclusions

For the user of financial statements, it is important to split earnings into its
accrual- and cash flow components
- The accrual- and cash flow components of earnings have different
predictive value for future performance
- Similar to the results for non-GAAP versus GAAP earnings we
saw in week 3
When earnings are high because of high accruals (low cash flows), this high level of earnings is less likely to recur in future
periods
- Because accruals rely on subjective estimations
- Investors do not understand this differential persistence: investors “fixate” on the earnings number, and do not look
beyond income statement information to assess whether high/low earnings are caused by high/low accruals or cash
flows

Why should we care about this, from an accounting perspective?

-Standard setters: information should be “useful to users”, but users should also be able to identify and understand the
information
-Companies: misunderstanding of performance may increase the costs of external financing
- Also: changes in reporting strategies can impact investor decision making (Jung et al. 2018 tutorial paper)
-Also: earnings management using accrual adjustments can work to mislead investors: investors do not examine the accrual
components of earnings as a “quality check” (see week 5)
-Auditors: if users place excessive weight on some components of the annual report, the verification of this information
becomes even more important

Lastly, Sloan (1996) asks whether the lower persistence of accruals is driven by earnings manipulation
Accruals might not only reverse because of their subjective estimation component, but also because of their use in earnings
management (more on this issue next week)

We want to understand earnings, we can create the higher quality of earnings and find the earnings management. This paper
does not explain why the market is not efficient and why investors are stupid. Because they can do only so much at the same
time.

Theories on users’ information processing

While Sloan’s (1996) evidence suggests that investors fixate on reported earnings numbers, his evidence does not tell us
much about why investors do so
- And why the stock market is not fully efficient in pricing the implications of accounting information

“Behavioral” theories from psychology can help us to better understand this phenomenon

In general, these theories suggest that decision makers typically have “limited attention” and are not able to completely
process all of the information that is presented to them
 - A decision can be informed by multiple pieces of information; in practice, however, attention should typically be
directed only to a subset of the information
- James (1980), as quoted in Hirshleifer et al. (2009, JF, p. 2289): “[Attention] is the taking possession by the mind in
clear and vivid form, of one out of what seem several simultaneously possible objects or trains of thought . . . It
implies withdrawal from some things in order to deal effectively with others.”

In the context of the processing of accounting information, the decision maker is an investor

In practice, investors receive many pieces of information, and should make decisions based on this information in a relatively
short period of time
- Example: annual reports contain many value-relevant pieces of information and are often 100+ pages long; at the
same time, investors often follow a portfolio of companies at the same time

Theories of limited attention suggest that it takes time and effort for an investor to fully understand the implications of each
piece of accounting information for (future) firm performance and valuation
- Because attention and time are scarce resources, sophisticated investors should prioritize their attention
- Other, unsophisticated, investors might not even have the time and resources to invest in their knowledge of
(accrual) accounting NOISE

Limited attention theories predict that investors’ neglect of information signals can lead to the mispricing of publicly
available accounting information (as we saw in Sloan 1996)

Hirshleifer and Teoh (2003, Journal of Accounting and Economics, p. 338):

“An immediate but far-reaching consequence of limited attention is that informationally equivalent disclosures can have
different effects on investor perceptions, depending on the form of presentation.”

What matters is how salient the information is presented to investors

- That is, information is more salient when it is more easy for the user to identify and process
Something is salient when the bottom line is clear to philosophers.

Theories on limited attention predict that investors are more likely to prioritize information that is presented in a more salient,
easy-to-process, manner

Leung and Veenman (2018, p. 1093): “Although non-GAAP disclosures do not necessarily provide investors with new
information, theories of limited investor attention and information processing costs (e.g., Bloomfield [2002], Hirshleifer and
Teoh [2003]) suggest that non-GAAP disclosures can be informative by making the differential predictive ability of earnings
components more salient.”

We have the non-GAAP because not all inventors have spread their attention to all the point.

“This [...] is important given prior evidence that less sophisticated investors are more likely to use nonGAAP earnings in their
decision making (Frederickson and Miller [2004], Elliott [2006], Allee et al. [2007], Bhattacharya et al. [2007]).”
Because these “less sophisticated” investors are more likely to suffer from limited attention

Both of this week’s tutorial papers (Bamber et al. 2010 and Jung et al. 2018) also refer to these theories of limited attention
and the Hirshleifer and Teoh (2003) paper to build up their hypotheses

Connection to Dechow Dichev 2002

We said earning persistence is a measure of earning quality. The more persistent – the higher quality. If you have high quality
(not much variability) => better persistence.
We talked whether users completely understands the earnings

Also if managers want to use it for their benefit or try to help investors understand it better.
Example of an exam question:

Question: Sloan (1996) argues that investors fixation on reported earnings is a possible explanation for the
results in his paper. Explain what Sloan means with investors “fixating” on earnings. Explain and motivate your
answer.

Answer: Investor fixation on earnings means that investors are not sophisticated enough to understand that
components of earnings (such as the cash-flow and accrual-components of earnings, or an impairment) have
different implications for performance measurement and valuation. When investors fixate their attention on
earnings, this means that they do not look beyond the net income number when evaluation company
performance. For example, they do not use additional information from the cash flow statement to better
understand how much of the earnings are backed by cash flows or accruals. As a result, they fail to anticipate
that the earnings may be more or less likely to recur in future periods.
Example of an exam question:
Question: Sloan (1996) empirically examines the persistence of earnings components, as represented in
Table 3 in Sloan (1996), and part of this table is presented below.

Explain based on the result of Sloan (1996) (a) which earnings component is most persistent, and (b) if
the result in table 3 of Sloan presented above is in conflict with the results of Dechow’s (1994) study on
accruals.
Example of an exam question:

Answer:
Tutorial 3 Bamber (2010)

Comprehensive Income: Who's Afraid of Performance Reporting?

Firms can report comprehensive income in either an income-statement like performance statement or the
statement of equity. Traditional theories of contract ing incentives cannot explain this reporting location choice
that only affects where comprehensive income data appear, because the contractible values of net income, other
comprehensive income items, and comprehensive income are exactly the same regardless of the location where
the firm reports comprehensive income. Drawing on theory, analysis of comment letters, and results of survey-
based and behavioral re search, we identify two factors?equity-based incentives and concerns over job security?
that help explain why most firms do not follow policymakers' preference to report comprehensive income in a
performance statement. Our empirical evidence on a broad cross-section of firms shows that managers with
stronger equity-based incentives and less job security are significantly less likely to use performance reporting.
Overall, our study suggests that even though the reporting location choice is inconsequential in a traditional
rational markets view, managers act as if they believe that comprehensive income reporting location matters.

INTRODUCTION

Financial reporting standards in the U.S. currently allow firms to report comprehensive income in either an
income-statement-like format as part of a statement of performance, or in the statement of equity (Financial
Accounting Standards Board [FASB] FAS No. 130). Firms report exactly the same information (i.e., the same
values of net income, other comprehensive income, and comprehensive income) either way. It is only the
reporting location that differs. In traditional models of financial markets, rational investors fully process
information regardless of its location, so it does not matter where firms report comprehensive income.

Standard-setters, however, believe the location of comprehensive income does matter. FAS No. 130 stated a
preference for an income-statement-like presentation—referred to as performance reporting—viewing this as the
more transparent presentation (FASB 1997, paragraph 67).2 Managers say developing a reputation for
transparent reporting is a key factor in their disclosure decisions (Graham et al. 2005, 54), so we would expect
managers who believe reporting location has little consequence to follow policymakers’ recommen- dations and
use performance reporting. Yet most firms report comprehensive income in the statement of equity, contrary to
policymakers’ preferences

We investigate determinants of managers’ comprehensive income reporting location choices. The only empirical
evidence on this choice is the Lee et al. (2006) investigation in the property-liability insurance industry. They
provide evidence that insurers who report comprehensive income in a statement of equity are more likely to
smooth earnings by cherry-picking realized gains and losses on available-for-sale (AFS) securities.

Outside the insurance industry, AFS securities are a smaller proportion of firm assets, which reduces
opportunities for cherry-picking. Yet firms outside the insurance industry are even less likely to use
performance reporting (Lee et al. 2006), so cherry-picking cannot be the primary reason non-insurers report
comprehensive income in the statement of equity.

many managers believed re- porting comprehensive income in a salient performance statement would lead
investors and other stakeholders to increase their assessments of the volatility of the firm’s performance.

Managers responding to the Graham et al. (2005) survey also say they believe higher perceived volatility of firm
performance could hurt the firm’s stock price and assessments of the manager’s own performance. We therefore
hypothesize that managers with more powerful equity-based incentives (whose wealth would suffer more
from lower stock prices) and managers with less job security (whose jobs are at greater risk if the firm
experiences unfavorable future performance) would be less likely to report comprehensive income in a
performance statement.

DEVELOPMENT OF HYPOTHESES

Traditional contract- ing incentives are often used to explain these kinds of choices that affect the values of
accounting numbers used in contacts. For example, debt covenants often incorporate rec- ognized accounting
amounts such as net income and debt-to-equity ratios, so incentives stemming from debt contracts help explain
accounting choices that affect recognized ac- counting numbers (e.g., Press and Weintrop 1990; DeFond and
Jiambalvo 1994).

But the choice we study is fundamentally different: the only difference is the location where these numbers are
reported. Therefore, we look beyond the contracting literature and draw on behavioral finance, comment letters,
and survey-based and behavioral research to identify a parsimonious set of factors that we expect to explain
most firms’ decisions not to follow policymakers’ preference for performance reporting.

We start by reviewing evidence that managers likely believed performance reporting would lead financial
statement users to perceive the firm’s performance as more volatile. Then we explain why managers would
expect an increase in the perceived volatility of firm performance to hurt the firm’s stock price and their own
performance evaluation. We use these discussions to develop our hy- potheses that managers who avoid
performance reporting have stronger equity-based in- centives and less job security.

Performance reporting = net income

Comprehensive income includes unrealized gains and losses on (1) AFS securities; (2) foreign currency
translations; (3) minimum pension obligations; and (4) certain hedging and derivative activities. These
unrealized gains and losses stem from uncontrollable and volatile market forces (e.g., stock market trends and
changes in currency exchange rates and interest rates) and are therefore transitory.

Thus, other comprehensive income items are generally volatile, transitory, and incomplete representations of the
firm’s unrealized gains and losses.
Their argument suggests that if comprehensive income is more salient, then: (1) users will put more weight on
comprehensive income as a performance measure than they would if it were less salient (Bloomfield’s [2002]
incomplete revelation hypothesis makes a similar prediction), and (2) users will fail to consider offsetting un-
realized gains and losses on other assets and liabilities that are not recognized in the current accounting model. 6
Both effects will lead users to believe the firm’s performance path is more volatile. 7

In sum, we expect that when managers were making their initial location decision, they expected
performance reporting would lead users to assess the volatility of the firm’s per- formance as higher. For
this maintained assumption to hold, all that is relevant is what managers believed at the time they made their
initial choice.

Evidence that Managers Would Expect an Increase in the Perceived Volatility of Firm Performance to Hurt
Stock Prices and Their Performance Evaluations

stakeholders perceive more volatile performance measures as indicating higher firm risk.

also show that financial statement users perceive uncontrollable items as increasing risk, so the relatively
uncontrollable nature of other comprehensive income items likely exacerbates users’ perceptions of firm risk.

GHR conclude that managers believe higher perceived risk hurts the firm’s stock price, with the interview
evidence leading to the persistent theme, ‘‘the market hates uncertainty’’

In addition to expecting the above adverse effects on equity investors, GHR’s managers also believe that more
volatile earnings paths lead to: (1) lower credit ratings and higher costs of debt; (2) greater customer and
supplier concern about the firm’s viability, and thus inferior terms of trade; and (3) lower growth projections by
investors, all of which nega- tively affect the firms’ stock price. Empirical research provides some support for
these additional concerns, by linking smoother earnings paths with: (1) better credit ratings and lower cost of
debt (e.g., Gu and Zhao 2006), and (2) better terms of trade with customers (e.g., Sommer 1996).

Hypotheses

Managers whose wealth is more sensitive to changes in the firm’s stock price (i.e., managers with more
powerful equity-based incentives) have more to lose from a lower stock price, and thus likely prefer reporting
methods that reduce the perceived volatility of firm performance (Goel and Thakor 2003). Our first hypothesis
(stated in the alterna- tive) is:

H1: The likelihood that a firm avoids reporting comprehensive income in a performance statement increases in
the power of the CEO’s equity-based incentives.

Managers who face greater risk of losing their jobs if they receive unfavorable per- formance evaluations—
including repercussions of poor stock price performance—have less job security. We hypothesize that managers
with less job security have more to lose from increased perceived volatility of firm performance. CEOs are
rightfully concerned about job security because dismissals increased significantly from 1971 to 1994 (Huson et
al. 2001) and further increased by 170 percent from 1995 to 2003 (Lucier et al. 2004). Our second hypothesis
(stated in the alternative) is:

H2: The likelihood that a firm avoids reporting comprehensive income in a performance statement decreases as
the CEO’s job security increases.

RESEARCH METHOD

Power of the CEO’s Equity-Based Incentives

We hypothesize that managers with more powerful equity-based incentives are more likely to avoid
performance reporting. We use Bergstresser and Philippon’s (2006) measure of the sensitivity of the CEO’s
stock and stock option holdings to changes in stock prices. Bergstresser and Philippon (2006) measure the dollar
change in the value of a CEO’s stock and stock option holdings arising from a one percentage point increase in
the firm’s stock price:

ONEPCT = the effect of a one percentage point increase in the firm’s stock price on the value of the firm’s
shares held by the CEO (i.e., 0.01 share price number of shares the CEO owns) plus the effect of a one
percentage point increase in the firm’s stock price on the value of the CEO’s options, calculated following Core
and Guay (2002).16

Bergstresser and Philippon (2006) then deflate ONEPCT to construct EQUITY INC, and argue that EQUITY
INC reflects the portion of the CEO’s total annual compensation from the firm stemming from a one percentage
point increase in the firm’s stock price:

Job Security

Model

RESULTS
The results support our hypothesis that firms whose CEOs have higher equity-based incentives are more likely
to avoid performance reporting (p 0.01), whether or not we control for industry. Table 4 also shows that CEOs
who enjoy greater job security are less likely to avoid performance reporting, consistent with our second
hypothesis (p 0.03). These inferences are robust to alternative specifications of equity incentives and job
security.24

Analysis of Firms Changing their Comprehensive Income Reporting Location

Table 5, Panel A reports the changes in our explanatory variables between the year the firm first reported
comprehensive income and the year the firm changed its reporting lo- cation. We denote these ‘‘change’’
variables with the prefix . For the 13 firms that changed away from performance reporting, EQUITY INC
increased, while J SECURITY decreased, on average. In stark contrast, for the seven firms that switched to
performance reporting, EQUITY INC decreased and J SECURITY increased. VOLATILITY increased more for
firms that changed to the statement of equity (p 0.08), but none of the other explan- atory variables changed
significantly.

Panel B of Table 5 reports the results of a logit analysis on the change firms, where the dependent variable
equals 1 for firms that switch from performance reporting to the tatement of equity location, otherwise 0. We
include only the change version of our two key variables, EQUITY INC and J SECURITY, because we have so
few degrees of freedom (adding a control for VOLATILITY does not affect our inferences). Consistent with the
univariate change analysis, the coefficient is positive on EQUITY INC (p 0.018) and negative on J SECURITY
(p 0.015). CEOs with increasingly powerful equity incentives and decreasing job security are more likely to
switch away from perform- ance reporting. The consistency of our changes and levels analyses suggests our
conclusions are not attributable to some unidentified firm-specific effect or correlated omitted variable. The
results of our change analysis further indicate that managers continue to act as if reporting location matters.
CONCLUSIONS

Our empirical results support our hypotheses that managers who avoid performance reporting are those with
more powerful equity-based incentives that will be devalued by a lower stock price and those with less job
security who have more to lose from a poor performance evaluation. The magnitudes of these effects suggest
managers’ unique personal equity-based incentives and job security concerns lead to meaningful differences in
report- ing behavior. Analysis of the small set of firms that change their comprehensive income reporting
location further supports our inferences: CEOs with increasingly powerful equity incentives and decreasing job
security are more likely to switch away from performance reporting. Our inferences also generalize to a random
sample of smaller firms not in the S&P 500.

Our evidence suggests that at the time managers made the initial comprehensive income reporting location
decision, they believed performance reporting would be more salient— consistent with views of the FASB and
IASB and evidence from laboratory studies (e.g., Hirst and Hopkins 1998; Maines and McDaniel 2000; Hunton
et al. 2006). Our change analysis suggests managers continue to believe reporting location matters. Our
evidence on who refuses to follow policymakers’ stated preferences on what seems to be an inconse- quential
reporting location choice is relevant as policymakers consider requiring all firms to use performance reporting
(IASB 2008; FASB 2008).

Our result that managers with less job security on average make reporting choices that reduce transparency is of
interest in its own right, and also helps fill the void that Graham et al. (2005) identify when they point out the
dearth of evidence on how managers’ career concerns affect their financial reporting choices. Our results also
extend prior research showing that equity-based compensation increases incentives for earnings management
(e.g., Cheng and Warfield 2005; Bergstresser and Philippon 2006) by showing that equity incentives affect
another transparency-related accounting choice: whether to disclose com- prehensive income in a more or less
salient location.

Finally, in comment letters on the initial FAS No. 130 proposal, managers expressed concern that performance
reporting would lead stakeholders to view firms’ performance as more volatile than warranted by the actual
economics. Our evidence suggests that the concerns expressed in the comment letters are real (as distinct from
excuses).

Tutorial 3 Jung, Naughton, Tahoun, Wang (2018)

Do Firms Strategically Disseminate? Evidence from Corporate Use of Social Media
We examine whether firms strategically disseminate information to the public. Strategic dissemination is
different from strategic disclosure, whereby firms voluntarily provide information to the public if the benefits
outweigh the costs

In contrast, strategic dissemination refers to firms choosing to use or not use certain channels of communication
to distribute both voluntary and mandatory information.

Either the two decisions are inseparable (e.g., quarterly disclosures must be filed with the SEC, and the vast
majority of firms always conduct quarterly conference calls to discuss the disclosures) or the dissemination
channel is not controlled by the firm (e.g., the business press makes editorial decisions about which firms to
cover)

Our investigation of strategic dissemination focuses on quarterly earnings announcements disseminated on the
social media platform Twitter by Standard & Poor's (S&P) 1500 firms. We focus on earnings announcements
because they are of first-order importance to investors and they are mandatory disclosures that do not need to be
disseminated on Twitter, which allows us to isolate the dissemination decision from the disclosure decision. We
focus on Twitter because our data reveal that it has surpassed other social media platforms in general corporate
adoption and for disseminating investor-related announcements, such as earnings announcements, which
provides power for our empirical tests.

Our primary focus is to examine if firms are strategic in their dissemination of news. Ex ante, it is unclear if
firms selectively disseminate some types of news and not others. If firms want to build or maintain a reputation
for transparency and trustworthiness, then they should have a consistent policy (e.g., always or never) for
disseminating corporate news. However, if firms only want to promote good news or explain bad news ([ 39] ;
[ 33] ), then they will be selective in their use of social media. In the setting of earnings announcements, where
news can be considered good or bad relative to analyst consensus forecasts, we hypothesize that firms are more
likely to disseminate good news over social media. While we defer a fuller discussion of the incentives for
strategic dissemination over social media to Section II, the intuition for our hypothesis is simple—firms (and
their managers) want to disseminate good news as widely as possible, but not publicize bad news any
more than is necessary. Bad earnings news is already disseminated through traditional channels (e.g., press
releases sent to newswires) and firms are not obligated to broaden the dissemination over social media.
Consistent with this expectation, we find that firms are less likely to disseminate quarterly earnings news
through Twitter when the news is bad and when the earnings miss is greater in magnitude, consistent with
strategic dissemination behavior. We further examine this behavior by testing whether the extent (or amount) of
strategic dissemination is associated with the direction of the news.

Thus, we measure the extent of dissemination by the number of tweets that a firm sends in a given quarter about
the same earnings announcements. We find some evidence, albeit weaker, that firms tend to send fewer earnings
announcement (EA) tweets when the news is bad, again supporting the notion that firms strategically
disseminate. Moreover, we examine earnings announcement “preview” and “rehash” tweets (discussed in
Section III) and find that firms tend to send fewer rehash tweets when the magnitude of earnings news is worse,
consistent with our main results using EA tweets.

While disclosure and dissemination are distinct, they can both be utilized to shape a firm's information
environment. Therefore, incentives that influence a firm's level of disclosure may also influence the level of
dissemination. We consider three different disclosure- and dissemination-related incentives—the firm's
overall litigation risk, the sophistication of its investor base, and the size of its social media audience. In
our cross-sectional analyses, we find evidence that firms' strategic dissemination, or the incentive to do so, is
greater for firms with a lower level of investor sophistication and firms with a larger social media audience. We
also find that strategic dissemination behavior is detectable in high litigation risk firms, but not low litigation
risk firms.

Next, we investigate how social media users respond to strategic dissemination. Twitter enables not only firms
to directly tweet information to their followers, but also enables the firms' followers to “retweet” the information
to their followers. We are interested in examining whether the frequency of retweeting a firm's earnings news
and the size of the audience that receives the retweet is associated with the direction of the earnings news. We
find that while firms exhibit strategic behavior in their dissemination of earnings news over Twitter, their
followers are not more or less likely to retweet good or bad news.
Finally, we examine the relation between Twitter dissemination and reactions from the traditional media and the
capital markets. We caution that any direction of causality is difficult to prove, as a firm's dissemination
decision could be a response to the media and market reactions or vice versa. Therefore, we conduct these
analyses with the intent to provide descriptive evidence on whether Twitter dissemination is associated with
media and market reactions. We find some evidence that the dissemination of bad earnings news over Twitter
by a firm, and the subsequent retweeting of the news by the firm's followers, are associated with more negative
news articles written about the firm at the time of the earnings announcement. In terms of market reactions, we
find that abnormal bid-ask spreads are smaller when a firm makes an EA tweet to a greater number of followers,
while the reduction in spreads is mitigated when there are more retweets of a firm's EA tweet to users who do
not directly follow the firm. These results are consistent with the notion that firm-initiated social media
dissemination may improve a firm's information environment, but user-initiated dialogues not controlled by the
firm may have a countervailing effect.

BACKGROUND AND HYPOTHESIS DEVELOPMENT

Varying levels of dissemination should be a consideration within a firm's disclosure strategy because broader
dissemination increases public awareness of the firm's disclosure and can increase investor recognition of the
firm itself, which, in theory, increases firm value

A number of recent empirical studies have provided evidence showing that how disclosures are disseminated
matters. For example, studies have shown that increased newswire dissemination affects stock prices ([ 34] ),
reduces information asymmetry ([ 10] ), and affects the price discovery process ([ 47] ).

Social media, and in particular, Twitter, provide a unique setting to examine whether firms strategically
disseminate. Conventionally, if a firm wanted to publicize investor-related information, such as an earnings
announcement, then it would send a press release to intermediaries, such as newswire services, equity research
databases, and brokerage firms ([ 20] ). Under this approach, a firm would not know if or when any of its
existing or prospective investors received the information. In contrast, a firm can use Twitter to: ( 1) directly
disseminate information to its followers without an intermediary, ( 2) control the timing of the dissemination,
( 3) send multiple repeated messages (or similar messages) over several days related to the same information
event, and ( 4) know its exact number of followers.

As mentioned earlier, it is unclear, ex ante, if firms selectively or strategically use social media to disseminate
some types of news and not others. First, if firms want to build or maintain a reputation for transparency and
trustworthiness, then they should have a corporate policy to either: ( 1) never use social media, ( 2) only use
social media for marketing (nonfinancial) purposes, or ( 3) use social media for financial news consistently,
regardless of whether the news appears good or bad. Second, it is also possible that firms may use social media
more to mitigate bad news, such as investor uncertainty created by large negative earnings surprises ([ 39] ), or
other crises such as product recalls ([ 33] ). As a third possibility, firms may be more likely to disseminate only
good news on social media. Incentives to promote good news relate to the previously discussed strategic
disclosure incentives to increase firm value.

n our setting of quarterly earnings announcements, in which good and bad earnings news can be based on
whether earnings met or missed analyst expectations, we expect incentives to emphasize positive performance to
dominate. As a result, we posit that managers with good earnings news will attempt to increase the breadth of
dissemination using social media, but that managers with bad earnings news will not. Evidence of such behavior
is consistent with strategic dissemination. We state our first hypothesis as follows:
H1a: Strategic dissemination is associated with the direction of the news; firms are more (less) likely to
disseminate good (bad) news over social media.

We also exploit the feature of Twitter that enables firms to send multiple repeated tweets (or similar tweets)
about the same quarter's earnings announcement to measure the extent (or amount) of strategic dissemination.
Similar to our first hypothesis, we expect the extent of strategic dissemination to be associated with the direction
of the news:
H1b: The extent of strategic dissemination is associated with the direction of the news; within the same quarter,
firms tend to send more good news (fewer bad news) tweets over social media.
 DETERMINANTS OF TWITTER USAGE

General Twitter Usage

TWi is an indicator variable set to 1 (0 otherwise) if firm i had a Twitter account anytime during the sample
period. To test potential determinants, we include variables related to a firm’s traditional media activity and
other firm characteristics. We capture a firm’s level of self- initiated disclosures using PRESS_RELEASES, the
log of one plus the number of corporate press releases issued by the firm and distributed via a news provider.
We capture the amount of attention a firm receives from the traditional media using MEDIA_NEWS, the log of
one plus the number of news articles written by traditional media organizations about the firm. For other firm
characteristics, we include variables that have been shown in the literature to be determinants of disclosure
through other communications channels such as conference calls (Frankel et al., 1999), corporate websites
(Ettredge et al., 2002), and conference presentations (Bushee et al., 2011). We include firm size, measured as
the log of total assets (SIZE), the market-to-book ratio (MTB), return-on-assets (ROA), yearly sales growth
(GROWTH), the debt- to-asset ratio (LEVERAGE), and the log of one plus the number of analysts who cover the
firm (ANALYSTS). We also include variables that have been used in prior papers examining social media
adoption (e.g., Lee et al., 2015), including advertising expense scaled by total sales (ADVERTISING), the
number of years since a firm’s founding (FIRMAGE), an indicator for whether a firm is headquartered in the
Silicon Valley region of Northern California (SILICON), and the age of the CEO (CEOAGE). Data for these
variables come from Compustat, I/B/E/S, ExecuComp, or a firm’s website, and they are measured as of the
10
latest quarter prior to a firm’s Twitter adoption, or for non-adopters, the last quarter in our sample period. We
include industry fixed effects and all variables are summarized in Appendix A.

The results of estimating Equation (1) are presented in Column (1) of Table 2 for the 1,422 firms (out of the
1,500 firms in the S&P 1500 index) for which we have requisite data. The positive coefficients on SIZE, MTB,
and ADVERTISING suggest that firms with Twitter accounts tend to be larger, more valuable, and spend more
on advertising expenses (e.g., because they are retail firms). In addition, these firms have lower leverage, higher
analyst coverage, and issue more press releases. However, the negative coefficient on MEDIA_NEWS indicates
that fewer articles are written about these firms in the traditional media. Last, firms that adopted Twitter tend to
have younger CEOs, but the age of the firm and whether it is located in Silicon Valley are not significant
factors. In the next subsection, we compare these results to our findings for the determinants of Twitter usage

for earnings news.

Twitter Usage for Earnings News

We next investigate the determinants of a firm’s choice to use Twitter to disseminate earnings news at least once
during our sample period. We run a firm-level, cross-sectional probit regression similar to Equation (1), with the
primary difference being that the dependent variable, TW_EAi, is an indicator variable set to 1 (0 otherwise) if
firm i used Twitter to disseminate earnings information at least once during the sample period (i.e., made at least
one EA tweet).11

The results of the full sample regression are provided in Column (2) of Table 2, and the results of the subsample
regression are in Column (3). The significantly positive coefficients for PRESS_RELEASES and SIZE in both
columns indicate that larger firms and firms that issue more press releases tend to disseminate earnings news
over Twitter. The significantly negative coefficients for MEDIA_NEWS in both columns indicate that firms
with fewer articles written about their earnings news from the traditional media are more likely to disseminate
earnings news over Twitter. The negative coefficient for CEOAGE in Column (2), but not Column (3), suggests
that across all firms, those with younger CEOs are more likely to adopt Twitter and use it for earnings news. But
conditional on having a Twitter account already, CEO age is not a significant factor. The negative coefficient for
FIRM_FOLLOWERS in Column (3) indicates that firms with fewer followers are more likely to use Twitter for
earnings information. This result is consistent with the summary statistics in Section III, which indicated that
firms in retail customer-facing industries (e.g., Retail) were less likely to use Twitter for earnings news, while
firms in industrial industries (e.g., Oil and Steel) were more likely to do so.

A number of differences exist between the choice to have a Twitter account, modeled in Table 2, Column (1),
and the choice to use Twitter to disseminate earnings information, modeled in Column (3). While firms with
greater analyst coverage are more likely to have a Twitter account, there is no difference in analyst coverage for
firms that use and do not use Twitter for earnings news. In addition, while firms with high levels of advertising
tend to have Twitter accounts, those with low levels of advertising tend to use Twitter to disseminate earnings.
This evidence suggests that firms that use Twitter for earnings and those that use Twitter in general are
fundamentally different, consistent with the earlier discussion in Section III.

Quarter-by-Quarter Dissemination of Earnings News Using Twitter

In this subsection, we formally test our first hypothesis (H1a) that strategic dissemination is associated with the
direction (good or bad) of quarterly earnings news. We model a firm’s choice to make an earnings
announcement (EA) tweet on a quarter-by-quarter basis using a panel regression with firm-quarter observations.

TW_EA_Qi,q is an indicator variable set to 1 (0 otherwise) if firm i disseminated at least one EA tweet (i.e.,
earnings-related tweet on firm i’s earnings announcement date) for fiscal quarter q. The independent variables of
interest are MISSESTi,q, an indicator variable set to 1 (0 otherwise) if firm i’s actual earnings per share (EPS) is
below the latest consensus mean analyst forecast for quarter q; the absolute earnings surprise
jEARNINGS_SURPRISEji,q, defined as the absolute value of the firm’s actual EPS minus the latest consensus
mean analyst forecast, scaled by stock price at the end of the quarter; and their interaction term MISSEST i,q
jEARNINGS_SURPRISEji,q. For ease of interpretation and to reduce multicollinearity, we demean the
continuous jEARNINGS_SURPRISEj variable when computing the interaction term. All other independent
variables are previously defined and we include quarter fixed effects, in addition to industry fixed effects. 14

Descriptive statistics of the variables used in the panel regression are provided in Table 3, Panel A. Continuous
variables are winsorized at the 1st and 99th percentiles. Of the 18,706 firm-quarters in our full sample, firms
made an EA tweet in 11.8 percent of the firm-quarters, and they missed analysts’ consensus expectations in 26.6
percent of the firm-quarters.

The results of the full-sample regression are provided in Column (1) of Table 3, Panel B, and the results of the
subsample regression are in Column (2). In both columns, the coefficients for MISSEST and the interaction
term MISSEST jEARNINGS_ SURPRISEj are significantly negative, indicating that firms that miss analyst
earnings expectations and miss by larger amounts are less likely to tweet earnings news over Twitter. 15 The
marginal effect of missing expectations (MISSEST 1⁄4 1) is a 1.4 percent decrease in the probability of sending
an EA tweet, which may appear nominal on an absolute basis, but it represents 12 percent of the unconditional
probability (11.8 percent) of a firm sending an EA tweet in a given quarter. These results support our first
hypothesis that the decision to disseminate earnings news over Twitter is related to the direction of earnings
news, which is consistent with strategic dissemination behavior by firms.

To provide further evidence on strategic dissemination in the social media setting, we test the second part of our
hypothesis (H1b) that the extent (or amount) of dissemination is associated with the direction of quarterly
earnings news. We replace the binary dependent variable in Equation (2) (TW_EA_Q) with a continuous
variable (TW_EA_NUM) that is the log of 1 plus the number of EA tweets that a firm made for fiscal quarter q.
The results of the full-sample ordinary least squares (OLS) regression are provided in Column (3) of Table 3,
Panel B, and the results of the subsample regression are in Column (4). The coefficient for MISSEST is negative
in Column (3) (significant at the 10 percent level), indicating that across all firms, the quarters in which a firm
missed analyst earnings expectations tend to have fewer EA tweets.16 The coefficients for the interaction terms
are negative, but not significant, suggesting that the magnitude of the earnings miss is not correlated with the
number of EA tweets. Overall, the results in Panel B provide some support for H1b, that the extent of strategic
dissemination each quarter is associated with the direction of earnings news. 17

Tutorial 3
This week we are saying that it is not the accounting that is not working, but the users who do not know how to
use information => misuse
Looking at step 2 of the framework => looking at how investors look at the information.
Slippage=> step 2, the information is correct, but the users do not understand it.
The more we can predict the future, the more we can earn.

Bamber et al. (2010): “Comprehensive Income: Who’s Afraid of Performance Reporting?”

 Why do managers choose to “hide” information in the equity statement, rather than present the information
more prominently in the income statement?

1. Bamber et al. (2010) study managers’ choice to report comprehensive income in a “performance
statement” close to the income statement, or separately in the statement of shareholders’ equity.
Explain what comprehensive income means and how it differs from net income as reported in an
income statement. Also give an example of an item or transaction that is part of comprehensive
income, but not part of net income.

Comprehensive income is a potential problem for the investors to analyze.

Why is it a problem?

Comprehensive income is a broader measure, that includes unrealized gain and losses.

We need a broader income measure, because it may provide better prediction value.

Why do we have income? => so we can measure performance on an annual basis

Comprehensive income =
• The company’s performance over the period measured as the total change in shareholders’ equity (net of any
transactions with shareholders, e.g., dividends, buybacks, or share issuances)
Why do we measure the total change in shareholders’ equity => so shareholders know how much dividends they
earn.

Comprehensive income (CI) differs from net income (NI) because not all items/events that affect shareholders’
equity are recognized as a gain or loss in the income statement
• These are designated as “other comprehensive income” (OCI) and “bypass” the income statement
Why do we not include unrealized gain/losses in net income => because net income are operational and happen
year to year.

Examples given by Bamber et al. (2010) for the (old) U.S. setting – unrealized gains/losses on:
• Available-for-sale (AFS) securities; Foreign currency translation adjustments; Minimum pension obligations;
and Certain hedging/derivative activities

Other example under IFRS:

• Unrealized gains/losses on (a) non-trading equity investments, or (b) debt investments “held-for-collection and
selling”

We have financial reporting to predict future earnings. We do not want to use comprehensive income in
predicting the future, because unrealized gains and losses are not persistent.

2. Explain why it makes sense for accounting standards to require comprehensive income to be reported
separately from net income (i.e., regardless of the location choice). Hint: in answering this question,
you may want to refer to the concept of “persistence” and the link between persistence and value as
discussed by Nichols and Wahlen (2004).

Something is transitory if it does not happen year to year.

As Bamber et al. (2010) explain, the financial accounting standards previously allowed firms to report
comprehensive income (CI = NI + OCI) either:
A. As supplemental information in the income-statement (“performance reporting”)
B. Or in the statement of changes in shareholders’ equity
 Same information in the annual report, but different location
Traditional efficient markets view: rational investors fully process information regardless of location
However, standards setters believe location does matter and state a preference for performance reporting,
viewing this as more transparent

It is the same information but in a different location.

In a perfect situation, it should not matter where you put it, because investors can inform themselves.

Option A (Unilever annual report 2018, p. 75)

• This is the “performance reporting” method where the company reports OCI in a “performance statement”
• Net income and (other) comprehensive income reported on the same page => very clear
Option B would mean: exclude the statement of comprehensive income from this page and only report (other)
comprehensive income in the equity statement => less feasible
3. Bamber et al. (2010) find that most companies “hide” comprehensive income in the equity statement
instead of reporting it more prominently in a performance statement, and that this choice relates
predictably to managers’ incentives. What does this choice reveal about managers’ perceptions of
how efficiently investors and other stakeholders process accounting information?

As Bamber et al. (2010) report, most managers (81%) choose option B – but why?
• What does this choice reveal about managers’ perceptions of how efficiently investors and other
stakeholders process accounting information?
Similar to standard setters, we can conclude that managers believe that location does matter
• That investors process the information differently when reported less transparently in the statement of
shareholders’ equity
• Evidence from comment letters by companies: “over 80 percent of the comment letters expressed
concern about financial statement users’ reactions to mandatory performance reporting” (p. 101)
The objective of the paper is to shed more light on the motives behind managers’ choice to report the same
information in a different, arguably less transparent, location
Why would firms chose to report in equity change statement.
Because the managers want to hide some information.

What is the concern?

1. Psychology => behavioral theory – invective perception
2. How managers think => what is more important?
We want to figure out what managers believe and what is their perception.
• Part 1 of the theory: managers believe that users will perceive that company performance is more volatile
when OCI reported in a salient “performance statement”
• Empirically, comprehensive income is significantly more volatile than net income
• Behavioral theory from Hirshleifer and Teoh (2003): users will attach more weight to information that is
reported more saliently (e.g., in performance reporting)
• But the persistence tests suggested that users should not attach more weight to (other) comprehensive income
because it is more transitory

Salience = very clear, everybody can see it.

The more you show it, the more people use it to make decisions. If the investors see the comprehensive income
more often, they will believe that the company is volatile. The more volatile the company, the less is expected
stock price.
However, in perfect world, investors should not put too much relevance to the comprehensive income, because
it has transitory components.

• Part 2 of the theory: survey evidence from Graham et al. (2005, Table 8) suggests that managers believe
that users will think there is more uncertainty – and therefore that the company is more risky – when
performance is perceived as being more volatile
• Consequence of higher uncertainty and perceptions of risk: lower stock price
Manager will not show comprehensive income, if he has a lot of share-based compensations, because investors
will think the company is volatile and stock price will drop.
If the manager has a high job security, he does not give a shit.

• Because lower stock prices reduce the value from equity-based compensation:
H1: The likelihood that a firm avoids reporting comprehensive income in a performance statement
increases in the power of the CEO’s equity-based incentives
• Because lower stock prices may also affect managers’ performance evaluations:
H2: The likelihood that a firm avoids reporting comprehensive income in a performance statement
decreases as the CEO’s job security increases

Research Design

• 440 U.S. firms’ (S&P500) reporting location choice during 1998-2001

• 85 (19%): reports comprehensive income in a performance statement
• 355 (81%): reports comprehensive income in the equity statement
• On average, comprehensive income is material for the average company in the sample (Table 1):

A. Measuring equity-based incentives: the sensitivity of a CEO’s wealth that results from a hypothetical
one-percent change in the stock price
B. Measuring CEO job security using corporate governance variables:
1. CEO-Chair duality
2. Percentage of outsiders on the board of directors
• Prior research: when (1) the CEO chairs the board and (2) there are fewer outsiders on the board, the
CEO’s job security is higher because s/he is less likely fired in case of poor (perceived) performance

A lot of equity incentive = avoid posting comprehensive income

A lot of job security = does not avoid.
4. Look up the results in Table 4 of Bamber et al. (2010). Explain why these results are consistent with
hypotheses H1 and H2. Make sure you also explain based on which statistics from Table 4 you draw
this conclusion.
Results

Green one shows that the volatility is high

• Change tests (Table 5): CEOs with increasingly powerful equity incentives and decreasing job security
are more likely to switch away from performance reporting
• Strengthens the earlier findings
• Suggests managers continue to act as if reporting location matters after the initial reporting decision
• Same results found when focusing on smaller firms that are not in the S&P 500 (Table 6)

Summary and conclusions

• Although reporting location should not matter in a rational market when the same amount of
information is provided, managers act as if location does matter
• Key issue is the perceived volatility of a firm’s performance, which affects (according to managers) stock
prices and performance evaluations
• Managers who avoid performance reporting have greater stock-based incentives and lower job security
• These findings help explain why so many companies (>80%) avoided performance reporting
• Findings are in line with Sloan (1996) and suggest that managers know that investors typically fixate
their attention only on certain items in the annual report
• And are consistent with behavioral theories that suggest that the salience of information presented by
companies affects financial statement users’ processing of the information
Even though the information is the same, manager hide comprehensive income (not show in the income
statement, but equity statement), because investors do not understand the volatility concept and fixate on
numbers.

2. Jung et al. (2018): “Do Firms Strategically Disseminate? Evidence from Corporate Use of SocialMedia”
 Does the manner in which companies communicate their earnings information to investors make adifference?
Evidence from companies’ use of Twitter

5. Jung et al. (2018) examine companies’ use of Twitter to “strategically disseminate” their quarterly
earnings announcements. Explain what they mean with companies’ strategic dissemination of
information and how it differs from companies’ strategic disclosure of information.

Companies communication with investors

How do managers reach their investors. Do managers use twitter t to promote their earnings announcements.
• What is “strategic dissemination” and how is it different from “strategic disclosure”?
• Dissemination refers to a company’s choice to use (or not use) a certain channel of communication to
distribute both voluntary and mandatory information
• This information is provided to investors anyway; the question is only how it reaches investors
• Strategic means that companies change their dissemination when managers perceive the benefits of doing so to
outweigh the costs
• Here: companies’ choice to use Twitter to further promote their earnings announcement news
• Instead, strategic disclosure refers to the choice of a company whether or not to voluntary provide
information to investors
• Again when the perceived benefits of doing so outweigh the costs

6. Although the choice of how to disseminate information does not speak to financial reporting per se, it
does speak to the manner in which company-specific financial information reaches its users, such as
investors. Explain why Jung et al. (2018) believe the choice of whether and how to disseminate
financial information, through social media such as Twitter, makes a difference in the way investors
process the information.

• Why should the choice of whether and how to disseminate financial information, through social media
such as Twitter, make a difference in the way investors process the information?
• Investors typically suffer from limited attention and information processing constraints: value relevant
information is more likely missed by investors when presented in a less salient format
• Even more likely when more of the investors are not “sophisticated”
• Theoretical arguments used by Jung et al. (2018):
1. Broader dissemination increases public awareness of company disclosures and increases investors’
recognition of the firm (Merton 1987)
2. P228: “firms can use Twitter to broaden dissemination and overcome a lack of investor attention (Hirshleifer
and Teoh 2003; ...) that can persist despite dissemination through traditional channels”
Not tweeting might cause investors to not know the earnings results, and not increase the share price as
supposed.

Hypothesis
• Following these theoretical arguments, Jung et al. (2018) predict that if managers are strategic:
H1a: Strategic dissemination is associated with the direction of the news; firms are more (less) likely to
disseminate good (bad) news over social media
• Underlying null hypothesis: if companies want to build or maintain a reputation for transparency and
trustworthiness, they should be consistent in their use of social media, regardless of the news
• And because companies can send multiple repeated tweets about the same earnings announcement:
H1b: The extent of strategic dissemination is associated with the direction of the news; within the same
quarter, firms tend to send more good news (fewer bad news) tweets over social media

Good news => tweeting, bad news= silence, not feasible.

Research design
7. In which table of Jung et al. (2018) can we find the test of hypothesis H1a? Explain based on which
statistics from this table we can draw conclusions regarding H1a.

Main results

They are less likely to tweet when they do not meet the expectations. The worth news get, the less probability,
that they will post it.
The likelihood of tweeting is the same if the news are good.

Variation in strategic dissemination

CEO with a lot of followers in social media is more likely to tweet, if he has audience that listens.

Summary and conclusions

• Consistent with the strategic dissemination hypothesis, companies appear to be strategic in their use of
Twitter to promote their earnings announcement news
• Less likely to tweet earnings news when news is bad
• Likelihood is further decreasing with the magnitude of the bad news
• Results suggest that managers believe that dissemination changes the way in which accounting information
reaches, and is processed by, investors
• Even though the earnings announcement information is publicly available anyway
• Consistent with managers believing in investors’ limited attention

In the first paper managers were strategic about the place of reporting.
In second article, the managers are strategic about making news public on social media.

Lecture 4 Frankel, Jennings, Lee (2016)

Using unstructured and qualitative disclosures to explain accruals

We examine the usefulness of support vector regressions (SVRs) in assessing the content of unstructured,
qualitative disclosures by relating MD&A-based SVR-accrual estimates (MD&A accruals) to actual accruals.

We apply support vector regressions (Manela and Moreira, 2014) to management’s discussion and analysis
(MD&A) in the 10-K by identifying words and word-pairs that explain firm-level accruals. We also use
support-vector regressions (SVR) to examine whether the MD&A contains information useful for predicting
future cash flows. To show the applicability of SVR to other contexts and provide an intuitive benchmark for the
MD&A results, we apply SVR to conference call transcripts. We offer a means to assess the explanatory power
of narrative disclosures.

In contrast, SVR automates identification of narrative patterns that occur in conjunction with firm
fundamentals.5 Therefore, researchers can use SVR even when they do not know what combination of words
managers will use to describe a firm fundamental (e.g., accrual levels).

We illustrate this flexibility by assessing the narrative content for both contemporaneous accruals and future
cash flows, using both the MD&A and conference call transcripts.

Statistical methods used in textual analysis

Accounting and finance researchers use many textual analysis techniques to understand the relation between the
text in firm disclosures and firm fundamentals. Early studies use indices (e.g., FOG index), pre-determined
dictionaries, word counts, and disclosure length to identify specific textual characteristics (e.g., readability,
tone), which are subsequently associated with firm fundamentals (e.g., future performance).

Other accounting and finance papers categorize firm disclosures into topics using statistical learning methods
(e.g., latent dirichlet allocation) and relate those topics to specific fundamentals. These learning algorithms
typically require researcher judgment in determining the number and categorization of the topics identified by
the learning algorithm and are often estimated using an unsupervised approach (i.e., without fitting directly to an
outcome variable)

Support vector regression (SVR) is a supervised statistical learning method that can automate the
identification of patterns in a narrative that occur in conjunction with firm fundamentals. SVR does not require
researcher intervention in the form of identifying pre-determined dictionaries, manually categorizing topics or
disclosures, or relying on researcher judgment to assess the narrative content of the disclosure in explaining firm
fundamentals. SVR allows the researcher to identify a dynamic “dictionary” of key words and weightings that
can vary as economic circumstances change.

As a result, SVR has two advantages over other statistical methods widely used in accounting research. First,
SVR can be applied to a variety of disclosures, languages, and contexts at a relatively low cost. Second, SVR
can be applied to circumstances where researchers have diffuse priors regarding the words associated with a
quantifiable fundamental.

Support-vector regressions and accruals

We use SVR to examine how specific words and short phrases in the MD&A explain accrual levels.

We construct a firm-year level dataset with counts of all one- and two-word phrases included in each firm’s
MD&A, replacing highly frequent words (i.e., “stop words”) such as “and” and “the” with an underscore
symbol and removing words containing digits. We scale the counts of all words and phrases by the total word
and phrase count of each MD&A. We also remove infrequent words and phrases by requiring each word and
phrase to be included in the MD&A of at least 10 firms in each year.

We use this dataset to explain accruals (Accrualsi,t) using the following SVR procedure, where Accrualsi,t
equals working capital accruals for firm i in year t scaled by total assets in year t1.8

We estimate the SVR procedure in two ways. First, we estimate the model to obtain coefficients by industry,
which we define using GICS codes, including training data for the years t5 to t1. Performing the analysis at the
industry level allows the estimation method to identify important words and phrases that are specific to firms in
each industry for explaining accruals. We then apply the estimated coefficients to the words and phrases in year
t to obtain an out-of-sample estimate of accruals for firm i in year t, which we label MDA Accruals – Industry i,t.

Second, we estimate the SVR model using training data for all firm-year observations in year t1 to obtain
coefficients on each word and phrase count. As in the industry estimation method described previously, we
apply the estimated coefficients to the word and phrase counts for each firm i in year t to obtain an out-of-
sample estimate of accruals for firm i in year t, which we label MDA Accruals – Year i,t. We calculate the
estimated accruals at the yearly level across all firms in the sample to identify words and phrases that are
associated with economy-wide factors that affect accrual generation. We note some influential words and
phrases that are useful in explaining accruals at the yearly level are “receivable,” “net profit,” “net loss,” and
“substantial doubt.”

After calculating the MDA Accruals – Industryi,t and MDA Accruals – Yeari,t variables, we calculate a
composite variable (MDA Accrualsi,t) by averaging the two estimates of accruals.

Explanatory power of MDA Accrualsi,t

Our first test of the narrative content of MDA Accrualsi,t regresses Accrualsi,t on MDA Accrualsi,t using
ordinary least squares (OLS) regression (see Eq. (1)).

Consistent with MDA Accrualsi,t having narrative content in explaining accruals, we find a positive and
significant coefficient (1% level) on MDA Accrualsi,t, which is equal to 0.544. The adjusted R2 of the model is
equal to 0.097, suggesting that approximately 9.7% of the variation in Accrualsi,t is explained by MDA
Accrualsi,t.12, 13

Incremental explanatory power of MDA Accrualsi,t relative to the modified Dechow and Dichev (2002) model

Second, we assess whether MDA Accrualsi,t has explanatory power beyond the current financial statement
numbers. The SEC emphasizes that “MD&A should not be a recitation of financial statements in narrative form
or an otherwise unin- formative series of technical responses to MD&A requirements” (SEC, 2002). We
augment Eq. (1) to include controls for factors that a well-informed financial statement user would know before
reading the MD&A. We proxy for these factors using the Dechow and Dichev (2002) model as modified by
McNichols (2002) and Ball and Shivakumar (2005).14 This model yields a higher adjusted R2 than the basic
Jones model and has economic support for the added complexity. We present the augmented model in Eq. (2).

We present the results of Eq. (2), including and excluding MDA Accrualsi,t, in Columns 2 and 3 of Table 4,
respectively. The adjusted R2 of the model excluding MDA Accrualsi,t (Column 2) is equal to 0.153, suggesting
that approximately 15.3% of the variation in accruals is explained by the modified Dechow and Dichev (2002)
model.16 The adjusted R2 of the model including MDA Accrualsi,t (Column 3), is equal to 0.191, representing
a 24.8% increase relative to the explanatory power of the modified Dechow and Dichev (2002) model alone. We
also find that the coefficient on the MDA Accrualsi,t variable equals 0.398 and is significant at the 1% level.
This evidence suggests that our method provides an estimate of accruals using the MD&A text that is
incremental to the fundamental determinants of accruals.

Consistent with the prior literature, we find a significantly positive (1% level) coefficient on the CFO i,t-1
variable and a significantly negative (1% level) coefficient on the CFOi,t variable. The interaction between
CFOi,t and the Neg CFOi,t variable is significantly positive at the 1% level, suggesting that accrued losses are
more likely in periods of negative cash flow. We also find a significantly positive (1% level) coefficient on the
ΔSalesi,t variable, suggesting that higher sales growth results in higher accruals. We find a significantly
negative (1% level) coefficient on the PPEi,t variable, suggesting that high tangible asset levels result in higher
depreciation accruals.

Incremental explanatory power of MDA Accrualsi,t relative to a dictionary method

Third, we create an alternative accrual estimate using counts of accrual-related terms from an accounting
dictionary and compare the explanatory power of this estimate to MDA Accruals i,t. This test has two purposes:
1) to understand whether SVR has any advantage over a an accounting dictionary method; and 2) to understand
whether the explanatory power of MDA Accrualsi,t is derived from phrases in the MD&A that directly describe
accrual levels.

We then estimate an out-of-sample value for accruals following a similar procedure to that which we use to
estimate MDA Accrualsi,t. Specifically, using observations from the prior year, we estimate an OLS regression
of accruals on the counts of these accounting-related terms. We then apply the estimated coefficients from this
estimation to their respective counts in year t. We label the estimation of accruals generated from this procedure
as Dict MDA Accrualsi,t.

We present the results of estimating Eq. (3), including and excluding MDA Accruals i,t, in Columns 4 and 5 of
Table 4, respectively. The adjusted R2 of the model excluding MDA Accrualsi,t (Column 4) is equal to 0.052.
The coefficient on the Dict MDA Accrualsi,t is equal to 0.304 and is significant at the 1% level. The adjusted
R2 of the model including MDA Accrualsi,t (Column 5) is equal to 0.106, representing a 104% increase from
the adjusted R2 in Column 4. The coefficient on MDA Accrualsi,t is equal to 0.476 and is significant at the 1%
level, and the coefficient on Dict MDA Accrualsi,t is equal to 0.157 and is significant at the 1% level. These
results suggest that the SVR estimate of accruals 1) is incremental to the dictionary estimate of accruals, and 2)
is not simply driven by direct discussion of accrual levels in the MD&A.

We perform three additional untabulated tests to explore the possibility that the explanatory power of the MDA
Accrualsi,t is driven by direct references to accrual levels. First, we examine whether the incremental
explanatory power of SVR results from direct references in the MD&A to one-time items. We generate an
alternative estimate of accruals by repeating the dictionary method described previously, replacing counts of
terms from the Oxford Reference Dictionary of Accounting with counts of the following items: Write-off *,
Special Item*, Impairment*, Restructur*, Going Concern*, Disposal*, and Loss*. We find that the adjusted R 2
of the model including only this estimate of accruals is equal to 0.066. When we further include the MDA
Accrualsi,t variable in the model the adjusted R2 increases to 0.120, representing an additional 81.8% increase.
Thus, the SVR method is not subsumed by an approach that estimates accruals using direct references in the
MD&A to one-time items.

Second, we use the 49 footnote headers identified by Chen and Li (2014), which correspond to the existence of
specific accounting activities, to assess whether SVR explains accruals simply because it identifies the existence
of accrual accounts. We regress MDA Accrualsi,t on 49 indicator variables, each set equal to 1 when the
corresponding accounting activity is mentioned in the MD&A. We then include the residual value of this
regression as the sole regressor in the accrual model and find an adjusted R 2 equal to 0.066, suggesting that
MDA Accrualsi,t includes information incremental to the mere existence of accounting activities.

Finally, we exclude from the MD&A all mentions of the terms in the Oxford Reference Dictionary of
Accounting and re- estimate SVR. We then include this alternative SVR accruals estimate as the sole regressor
in the accrual model and find an adjusted R2 equal to 0.065. This result corroborates the dictionary approach
presented in Table 4 and suggests that excluding accounting terms does not extinguish the explanatory power of
the SVR estimate. Overall, these dictionary-related tests suggest that direct accrual descriptions do not fully
drive the explanatory power of MDA Accrualsi,t. Moreover, the expla- natory power does not result from a few
unique situations or obvious accounting terms.

Incremental explanatory power of MDA Accrualsi,t relative to MD&A tone

Finally, we assess whether the MD&A’s tone can approximate the explanatory power of MDA Accruals i,t. The
tone (Tonei,t) of the MD&A is equal to the number of positive words less the number of negative words divided
by the total number of words in the MD&A for firm i in year t, using the Loughran and McDonald (2011) word
lists. We augment Eq. (1) with Tonei,t and present the augmented model in Eq

We present the results of Eq. (4), including and excluding MDA Accrualsi,t, in Columns 6 and 7 of Table 4,
respectively. The coefficient on Tonei,t is positive and significant, suggesting that more optimistic MD&As are
associated with positive accruals. However, the adjusted R2 of the model excluding MDA Accrualsi,t (Column
6) is equal to 0.017, suggesting that Tonei,t has little explanatory power for Accrualsi,t. The adjusted R2 of the
model including MDA Accrualsi,t (Column 7) is equal to 0.098. These results suggest that Tonei,t cannot
approximate the explanatory power of MDA Accrualsi,t.

Readability

We next examine whether MD&A accruals are less useful when the MD&A is less readable. Li (2008) suggests
that managers decrease the readability of the 10-K to obfuscate information about poor performance. If
managers do this and the automated technique cannot overcome these vagaries, we expect MD&A accruals to be
less useful in understanding accruals. We use Eq. (5) to examine whether support vector regressions are less
useful in explaining accruals when the MD&A is less readable. We perform these analyses including the
fundamental determinants of accruals because this model yields the greatest explanatory power in Table 4.

We define MD&A readability in three ways: Fog index (Li, 2008), word length (Li, 2008), and file size

Fog index as a measure of 10-K readability using the computational linguistics literature. The measure increases
as the 10-K becomes more complex or difficult to read. Li (2008) also suggests that longer documents are more
difficult to read. Loughran and McDonald (2014) suggest that the 10-K document size is an additional measure
of 10-K readability. They suggest that more complicated and less readable reports are larger. Bonsall et al.
(2015) suggest that 10-K file size is affected by content unrelated to the underlying text in the 10-K (e.g.,
HTML, XML, pdf, and jpeg file attachments). For this purpose, we restrict our file size measure to include only
the 10-K document and the associated EX-13 (i.e., annual report), if available. We then regress the file size
variable on an indicator variable equal to 1 if the document is HTML and 0 if the document is text-based. The
residual is our measure of 10-K file size. We perform three separate regressions—one for each readability
proxy. The Low Readabilityi,t variable equals one if the Fog index, word length, or file size is above the sample
median. We expect a negative coefficient on the interaction between MDA Accruals i,t and Low Readabilityi,t if
MDA Accrualsi,t is less useful in explaining accrual generation when the MD&A is less readable.

We report these results in Table 5. Column 1 includes the results when the Fog index is used to calculate the
Low Readabilityi,t variable, Column 2 includes the results when the MD&A word length is used to calculate the
Low Readabilityi,t variable, and Column 3 includes the results when the file size of the 10-K is used to calculate
the Low Readabilityi,t variable. Consistent with our expectations, we find a significantly negative (1%
level) coefficient on the interaction between the MDA Accrualsi,t and Low Readabilityi,t variables in
Column 2 and 3. This evidence is consistent with SVR capturing an information content concept similar to that
measured by prior readability proxies. These results reinforce the construct validity of prior measures as well as
the measure presented in this paper.

Persistence of MD&A accruals

We now turn our attention to understanding whether MDA Accrualsi,t is more or less persistent than those
accruals that are not explained by our method. Dechow et al. (2010) suggest that understanding the persistence
of accruals is a valuable input for equity valuation. We use these persistence tests to explore the nature of the
accruals identified in the SVR estimation.

We first examine the persistence of operating cash flows and total accruals to establish a benchmark in
evaluating the accruals identified by SVR. Eq. (4) describes our base model, which is similar to that presented in
Sloan (1996)

All variables are as previously defined with exception to ROAi,t þ 1, which equals net income for firm i in year
tþ1 scaled by total assets in year t. We also include year and industry fixed effects and cluster standard errors by
firm. We present the results of Eq. (6) in Column 1 of Table 6. The coefficient on CFOi,t equals 0.511 and is
significant at the 1% level. Consistent with prior literature, we find that the coefficient on Accruals i,t is
significantly less than the coefficient on CFOi,t at the 1% level, suggesting that cash flows are more
persistent than accruals. The coefficient on Accrualsi,t equals 0.380 and is significant at the 1% level.
To test whether the accruals explained using SVR are more or less persistent than those that are not explained
this way, we replace Accrualsi,t in Eq. (6) with MDA Accrualsi,t and with Non-MDA Accrualsi,t (Accrualsi,t
less MDA Accrualsi,t) (see Eq. (7)).

The results for this specification are included in Column 2 of Table 6. We find that the coefficient on MDA
Accrualsi,t equals 0.447, and the coefficient on Non-MDA Accrualsi,t equals 0.372, both of which are
significant at the 1% level. We also note that the coefficient on MDA Accruals i,t is significantly greater than the
coefficient on Non-MDA Accrualsi,t at the 1% level, suggesting that accruals explained using SVR are more
persistent than those that are not.

Flexibility of support vector regressions

MD& A future cash flows prediction

In Column 1 of Table 7, we regress CFOi,t þ 1 on MDA CFOi,t þ 1. We anticipate a coefficient of 1 on the

MDA CFOi,t þ 1 variable if MDA CFOi,t þ 1 predicts future operating cash flows without error. We find that
the coefficient on MDA CFOi,t þ 1 is equal to 0.852 and is statistically significant at the 1% level. We also find
that MDA CFOi,t þ 1 explains 45.8% of the variation in future operating cash flows, which we believe is
economically significant.

We then estimate the following model (Eq. (8)), similar to that in Barth et al. (2001), with and without MDA
CFOi,t þ 1 to assess the predictive ability of the MDA CFOi,t þ 1 variable after controlling for the known
determinants of future operating cash flows found in the financial statements.

The ΔRECi,t, ΔINVi,t, ΔAPi,t, ΔTAXi,t, and ΔOTHi,t variables are from the statement of cash flows and equal
changes in receivables, inventory, accounts payable, taxes payable, and other net assets and liabilities,
respectively, scaled by lagged total assets for firm i in year t. The Depi,t variable equals depreciation expense
scaled by lagged total assets for firm i in year t. Table 7 presents the results using Eq. (8). In Column 2, we
exclude the MDA CFOi,t þ 1 variable and find an adjusted R2 of 0.512. We include the MDA CFOi,t þ 1
variable in Column 3 and find that the adjusted R2 increases to 0.551, which corresponds to a 7.6% increase
relative to the Barth et al. (2001) model. The coefficient on MDA CFOi,t þ 1 equals 0.408 and is significant at
the 1% level. Overall, these results suggest that the MD&A contains useful information for predicting future
cash flows.
Conference calls analysis

We next apply support vector regressions to earnings conference calls to highlight the flexibility of SVR for
assessing alternative disclosures and to compare the narrative content of the MD&A to that of the conference
call.

We therefore test whether managers provide information during conference calls that is useful for understanding
current accruals.

We present the results in Column 1 of Table 8. The adjusted R2 is equal to 0.157, which is slightly lower than
that reported for the full sample in Table 4. In Column 2, we include the MDA Accruals i,t variable generated by
the SVR procedure using the MD&A. The adjusted R2 is equal to 0.177, an increase of 0.020, or approximately
12.8 percent, relative to the base model. In Column 3, we include the CC Accruals i,t variable generated by the
SVR procedure using the conference call transcripts. The adjusted R2 is equal to 0.176, an increase of 0.019, or
approximately 12.1 percent, relative to the base model, and is lower than that obtained from the MD&A. In
Column 4, we include both the MDA Accrualsi,t variable and the CC Accrualsi,t variable and find an adjusted
R2 of 0.184 representing an increase of 17.2%, relative to the base model. These results suggest that the
MD&A and conference calls are both useful in understanding current accruals. However, the conference
call and the MD&A contain incremental information for under- standing accruals relative to each other as
evidenced by the greatest increase in explanatory power when both accrual estimations are included in the
model.

Time-series analysis

Our method allows us to examine whether there is a systematic change in the information content of the MD&A
over time. The SEC issued new MD&A guidelines in 2003, encouraging firms to reduce boilerplate disclosure
and improve the narrative content of the MD&A. The Sarbanes-Oxley Act also added an additional layer of
certification requiring the CEO and CFO to certify that the financial statements and footnote disclosures fairly
represent the firm's financial conditions and results of operations.

We test the change in the narrative content of the MD&A in explaining accruals over time by estimating Eq. (1)
by year. The adjusted R2 provides an indication of the explanatory power of MDA Accrualsi,t each year. We
restrict our estimation to firms with at least 10 firm-years throughout the 20-year sample period, though our
results are robust to including all firm- years. Fig. 1 plots the incremental adjusted R 2 for each year. We observe
a discernable increase in the incremental adjusted R2 over time. To determine whether the increase in the
narrative content of the MD&A is statistically significant, we estimate a time-series regression (N 1⁄4 20) with
the incremental adjusted R2 as the dependent variable and a trend variable equal to 1 in 1994, equal to 2 in
1995, etc. The coefficient on the trend variable is equal to 0.0085 with a t-stat of 4.71, suggesting that the
narrative content of the MD&A has improved over time.

Robustness tests

Industry-specific regressions

We next provide several robustness tests in relation to our primary analyses. In Panel A of Table 9, we present
industry- specific regressions. GICS codes yield 56 separate industry regressions. In Column 1 we present the
results with MDA Accrualsi,t as the sole regressor. The average coefficient on MDA Accruals i,t is equal to
0.459 and is significant at the 1% level. The average (median) R2 for the industry-specific regressions is equal
to 0.078 (0.071), which is slightly lower than the adjusted R2 for the pooled regression in Table 4. This
evidence suggests that MDA Accrualsi,t explains a significant portion of the variation in Accrualsi,t. when
estimated at the industry level.

Similar to our previous results, we focus on the modified Dechow and Dichev (2002) model because it has the
most explanatory power in explaining accruals in Table 4. In Column 2, we present the results using the base
Dechow and Dichev (2002) model estimated at the industry level. We find that the average and median
coefficients from the Dechow and Dichev (2002) model are consistent with those in Column 2 of Table 4. The
average (median) R2 is equal to 0.243 (0.185). In Column 3, we find that the mean (median) coefficient from
the industry-specific regressions on the MDA Accrualsi,t variable equals 0.309 (0.346) and is significant at the
1% level. We find that the mean (median) R2 increases by 12.3% (24.9%) relative to the model reported in
Column 2. While these results are attenuated relative to the pooled regression results, they remain economically
significant.

Explanatory power of MD&A accruals relative to future cash flows

Second, we examine whether the ability of the MD&A to explain current accruals is solely driven by its
explanation of accruals that convert into cash flows in the following year as described by Dechow and Dichev
(2002). In untabulated results, we find that the MDA Accrualsi,t is positively correlated with CFOi,t þ 1 at the
1% level, providing preliminary evidence that MDA Accrualsi,t partially explains future cash flows. In addition,
we re-estimate Eq. (2), including cash flows for firm i in year tþ1 (CFOi,tþ1) and re-assess the explanatory
power of the MDA Accrualsi,t variable. Columns 1 and 2 of Panel B in Table 9 present the results. The adjusted
R2 of the base model in Column 1, excluding MDA Accrualsi,t, equals 0.200, which is higher than the 0.153
adjusted R2 of the base model reported in Column 2 of Table 4. Including the MDA Accrualsi,t variable in
Column 2 increases the adjusted R2 to 0.227, which is a 13.5% increase relative to the base model reported in
Column 1. The coefficient on the MDA Accrualsi,t variable equals 0.331 and is significant at the 1% level,
suggesting that the MD&A provides information useful for understanding accruals beyond their reversal
in the following year.

We note, however, that the percentage increase in the adjusted R2 of the model (13.5%) is lower than the
percentage increase in the adjusted R2 of the model when future cash flows are not included (24.8%). This
evidence suggests that the MD&A at least partially explains cash flows that are realized in the following
year.

Timeliness of MD& A accruals

Third, our objective is to assess the narrative content of the MD&A in explaining the accrual generation process,
regardless of the accrual's timeliness (i.e., whether the accrual persists from the prior period or is new to the
current period). Understanding whether MDA Accrualsi,t identifies accrual innovations can also provide insight
into the nature of the MD&A’s narrative. In particular, if accruals change little from year-to-year and the
description of accruals in the MD&A varies little, investors are unlikely to find timely insights by reading the
MD&A despite a high correlation between MD&A word counts and accrual levels.

To examine whether MDA Accrualsi,t identify timely accruals, we add the prior fiscal period accruals
(Accrualsi,t-1) to Eq. (2) as an additional independent variable, effectively converting our levels analysis to a
changes analysis of accruals. Because accruals are positively serially-correlated, the portion of firm-level
accruals explained by prior fiscal year accruals (Accrualsi, t-1) identifies the more persistent portion of accruals
from the prior period. In Column 3 of Panel B in Table 9, we re-estimate the empirical model from Eq. (2)
(excluding MDA Accrualsi,t) after including Accrualsi,t-1 in the model. Consistent with our expectations, we
find that the regression in Column 3 explains a much greater portion of accruals when compared to the portion
of accruals explained in Column 2 of Table 4. In fact, the adjusted R2 increases by approximately 22.2% after
including Accrualsi,t-1 in the model specification.

Column 4 presents the results after including MDA Accrualsi,t in the model. We note that the coefficient on
MDA Accrualsi,t continues to be significantly positive, suggesting that our method identifies a statistically
significant portion of timely accruals. As expected, the increase in adjusted R 2 from Column 3 to Column 4 is
approximately 6.4% and is much smaller than the percentage increase in adjusted R 2 reported in Table 4. Since
the percentage increase in adjusted R2 is much greater in Table 4 than in Panel B of Table 9, this evidence
suggests that our method is more effective at identifying and explaining the more persistent (i.e., not
timely) component of accruals. Nevertheless, our method explains a portion of timely accruals (i.e.,
changes to accruals).

Lecture 4
Today looking at the issue of using machine learning at predicting future earnings. The main topic is
understanding better financial accounting. Machine learnings is important to analyze qualitative data.
Week 1: understanding financial accounting by looking at the research, to develop critical thinking.
Reporting quality – is an extension of what you know about accounting, how and why we can measure quality.
Is it the right way to account for XYZ
Week 2: why are we spending so much time to analyze reporting information. Capital markets/ investors are
primary users of accounting information.
Sometimes accounting does not work, because some firms provide voluntarily disclosures.
Week 3: whether investors understand what is written in the annual reports. There are some problems regarding
processing information. Investors have a limited attention, they cannot comprehend a lot of items at the same
time, so they fixate on some numbers, and missing some information. This leads to mispricing of accruals.
Week 4: preparers can add textual components to the numbers to improve understandability.
There are advances in technology – machine learning – that can improve understandability of qualitative
disclosures.
Why qualitative disclosure are important?
Financial reporting exists because investors want be informed about company’s activity to make predictions
regarding share price.
Current numbers can inform about future numbers.
When we create standards, we want them to be useful to the investors. It is useful when it can predict stock
returns.
Earnings are useful because they relate to the changes in stock pricing, they are informative about future
dividends, returns.
Earnings is a concept that is equal accruals (adjustments) and cash returns.
Cash flows have timing and matching problems.

Recap: focus on investors as users of financial statement

The papers we cover in Financial Accounting Research focus on the user of financial reporting, in
line with the focus of standard setters, such as the IASB
In the IASB Conceptual Framework, the IASB explains that the objective of financial
reporting is to provide information that is “useful to users”
• And identifies investors as a primary user
The concept of usefulness is an important starting point to examine financial accounting questions
• The key question we ask is: “Is this information useful to investors?”
• We show this by showing that financial reporting information, such as earnings, is associated with stock
returns

Recap: predictive value

• Why should earnings relate to stock returns?
• Nichols and Wahlen (2004) present a framework to illustrate the three links that form the foundation of the
relation between earnings and returns (p. 264):
1. Current period earnings provides information to predict future periods’ earnings, which ...
2. ... provide information to develop expectations about dividends in future periods, which ...
3. ... provide information to determine share value, which represents the present value of expected future
dividends
• In summary, information in earnings should relate to changes in stock price, because earnings are informative
about the future dividends / cash flows / earnings the company is expected to generate
• Consistent with this framework, the IASB recognizes the importance of “predictive value”

Recap: we examine two related measures of firm performance

1. Earnings = net income = “bottom-line” measure of accounting performance
• Earnings is an “accrual accounting measure of the firm’s profit or loss from business activities and events
during a quarter or annual period”
• It is “an accounting measure of the change in the value of the firm to common equity shareholders during a
period”
2. Stock return = relative change in stock price of a firm ଵ ଵ ଴ ଴
• Stock return equals the change in the market value of a firm over a period of time plus any dividends that are
paid
• Represents the stock market’s estimate of the firm’s bottom-line performance over the period
• The more these two measures move together (in the same direction), the higher the usefulness

Recap: Theory
• Earnings are useful to investors, because accruals (through the revenue recognition principle and matching
principle) help to make earnings a better measure of performance measure for expectations of future
performance than cash flows, which suffer from timing and matching problems (Dechow 1994).
• Over “finite intervals” (e.g., a year or quarter), cash flows are noisy measures of firm performance
because of the existence of “timing” and “matching” problems
• Accounting principles (GAAP or IFRS) have evolved such that accruals alter the timing of the recognition
of cash flows in earnings to fix the problems
• Revenue recognition principle:
• Recognize revenue when all (or at least most) of the services have been provided (i.e., the “performance
obligation” is met), not necessarily when cash is received
• Matching principle:
• Recognize cash outlays as expenses in same period as in which revenues are recognized

Recap: Dechow (1994)

• For each measurement interval, earnings explain more of the variation in stock
returns than the cash flow measures
 Higher R^2
• The ability of cash flow to measure firm performance improves when the
measurement interval increases

Accounting adjustments make accounting more useful

Recap: accruals improve cash flows

Recap: issue with accruals based on Sloan (1996)

Does this mean that accruals are more predictive than cash flows. Accruals are beneficial, but they do not
always have a better predictive value.
Sloan (1996) focuses on the comparison of persistence between cash flows and accruals
Although accruals help to make earnings (= cash flows + accruals) a better measure of performance, this
does not mean that accruals have greater predictive value than cash flows
• In other words: accruals are not necessarily more persistent than cash flows
• Note: the higher the persistence, the easier it is for users to predict earnings

Recap: persistence of cash flows and accruals

The problem is that cash flows have a higher predictive value than accruals. 85% is recurring. Cash flows are
more recurring than accruals.
For the user of financial statements, it is important to split earnings into its accrual- and cash flow components
• The accrual- and cash flow components of earnings have different predictive value for future performance
When earnings are high because of high accruals (low cash flows), this high level of earnings is less likely to
recur in future periods
• Because accruals rely on subjective estimations
Also: investors do not understand this differential persistence: investors “fixate” on the earnings number, and do
not look beyond income statement information to assess whether high/low earnings are caused by high/low
accruals or cash flows
Investors fixate and they do not see the difference. It leads to predictive problems

Recap: the trade-off

• The trade-off: accruals are associated with discretion
• Ideally, discretion in the recognition of accruals allows managers to signal their private information about
the firm’s cash-generating ability
• But, discretion in the recognition of accruals also allows managers to manage earnings (up or down)
towards a desired level (“earnings management”)
• Conventions in accounting standard setting therefore limit the ability of management to use their discretion
to manage the recognition of revenues and expenses:
• Trade-off between relevance and reliability
Reliability – managers engaging in earning management.

Recap: accruals trade off relevance and reliability

• Even if we argue that managers use accruals in a good way (i.e., they do not use accruals for earnings
management), accruals are likely to be less persistent than cash flows because:
• Accruals rely on expectations of the future
• Uncertainty about the future will lead to accrual estimation errors
• In contrast: cash flow realizations are “hard” evidence – no uncertainty
• The accrual-component of earnings is likely to be less persistent compared to the cash-flow component
of earnings due to the existence of accrual estimation errors
• This can be true even given Dechow’s (1994) result that accruals make earnings a better performance measure

Recap: accrual estimation errors

Recap: Dechow and Dichev (2002)

The larger the error, the lower the persistence.

Recap: Dechow and Dichev (2002)

• Accruals help to make earnings a better measure of performance (than cash flows), but the use of accruals
comes at a cost: accrual estimation errors reduce the quality and persistence of earnings
• Accrual estimation errors result from both intentional and unintentional errors in the estimates and
assumptions reflects in accruals
• Several company characteristics are predictably related to the measure of accrual quality
• Confusing result? Dechow (1994) argues that accruals benefit firm performance as measured by stock
returns, but Sloan (1996) shows that accrual are less persistent than accruals, and that leads to lower returns?
Actuals give as better predictions, if the would not be estimation errors.

• Reconciliation of results of Dechow and Dichev (2002) with Dechow (1994)

• The results of Dechow and Dichev (2002) suggest that the result of Dechow (1994), that accruals improve
earnings as a measure of performance, only hold for high quality accruals
• Reconciliation of results of Dechow and Dichev (2002) with Sloan (1994)
• The results of Dechow and Dichev (2002) suggest that the result of Sloan (1996), that accruals are less
persistent than cash flows for the prediction of earnings, only hold for low quality accruals
• Investors who fixate on earnings do not “see” that a high magnitude of accruals in earnings also implies more
low quality accruals in earnings, leading to lower earnings next year, resulting in lower returns

On average accruals measure performance better, but only for high quality accruals.

Recap: one solution to the accruals issue: non-GAAP reporting

Solution – non-Gaap reporting. It helps investors to understate the data. Non-gaap has a better predictive value
than gaap.

Another solution to the accruals issue is qualitative disclosures

We want to understand the predictive value of earnings. By applying algorithm we can look at the extent of
error.
• Managers take Non-GAAP measures that take out non-recurring, transitory, items such as goodwill
impairments can be more useful, because the transitory items should have limited predictive value
• Another way managers can disclose information that allow users to better predict earnings is by
providing qualitative disclosure.
• Key Issue: how can we use machine learning to empirically test if qualitative disclosures allow users to
better predict future earnings?
Can we provide a better information by using qualitative disclosures?
Management discussion and analysis – provided with financial statement. Where management provides a story
of performance of the firm.

Machine Learning

Machine learning is a set of algorithm that helps us understand earnings and accruals.
• Machine Learning: the automated detection of meaningful patterns in data by using algorithms with
training data*
• Machine Learning (ML) is a sub-field of Artifical Intelligence (AI) which aims to implement intelligent
systems by learning from the data instead of using predefined rules
• ML methods are well known to be particularly useful for exploiting large volumes of data more
efficiently than traditional statistical methods
• ML has been particularly useful when dealing with unstructured data that are too complex or high-dimensional
for standard estimation methods, including language and image information
Why is it helpful in textual analysis?
Language processing – putting words in a story.
Quantify words – make themes.

• How can Machine Learning be used (in financial accounting research)?

• Why should we know more about this?
• Exploitation of the rich potential of various textual analyses
• Natural Language Processing (NLP)
• Quantification of qualitative and unstructured data (e.g. text and images) to create new measures for
existing and new research variables,
• Creation of better estimates and predictions
• In this course, the emphasis will be on using ML for textual analysis
• Conventional methods: “bag-of-words” approach, dictionaries, topic modelling
• Advanced methods: “word embeddings” and neural networks (not covered in this course)

We focus on machine learning, that involves different techniques.

We want to understand if this technology explains information better.

• Two methods:
• Supervised
• Machine learning style where the model is trained on inputs and labeled outputs to learn how to map
inputs to outputs
• Example: Support Vector Machines (SVM) (see Frankel et al. (2016); this lecture)
• Example: [Decision Trees] Random Forest (RF) (see Frankel et al. (2022); tutorial)
• Unsupervised (no label)
• Machine learning style where the model is trained only on unlabeled data, letting the model find the
structure in the data
Why data based on machine have a better predictive value.

• Example: Latent Dirichlet Allocation (LDA) (see Brown et al. (2021); tutorial) - supervised

the ocl
Training data based on beta 1,2,3. It is supervise because we are focusing on the working capital, Because if it
would be unsupervised, we would not know the outcome.
If you run supervised – linear model, you need to run a line. Having linear regression restricts you in how much
you can predict. We want to solve – why do we have a linear trend. So we ask the machine whither that line is
the better predictor. We also ask machine if it is still linear.

In Machine Learning, the full sample is split in a training sample and a testing ample (in the papers, the
validation is implicit). Then, a ML algorithm, such as Support Vector Machines (SVM), Latent Dirichlet
Allocation (LDA) or Random Forrest (RF), is used to estimate parameters. Finally, the parameters from the
training sample are used on out-of-sample prediction.

We predict from text what the number should be. But we are not constricted to the line.
Frankel (2016)
Using narratives to explain accruals
• Paper objective: present a measure of accruals based on the narrative disclosure in the Management
Discussion and Analysis (MD&A)
• Why?
• Frankel et al. (2016, p. 209) state that “Many papers study narrative content” to examine:
• Disclosure characteristics: uncertainty, tone, competition, sentiment
• Association firm fundamentals with measures of readability, similarity, deception,or length
• This paper expands on this prior literature, and uses narrative content to examine the ability of the MD&A to
explain accruals.
We do not look at the income statement, or balance sheet, we look at the narrative.
If you are the user what do you need to look at in the texts. When is information useful?
Quantitative data is not useful because of intentional or unintentional errors.
When we see certain words do they predict something regarding the company.

• Frankel et al. (2016, p.210): “MD&A is not necessary to derive the level of accruals; investors can
calculate accruals from changes in balance-sheet accounts or from net income and operating cash
flows. Our goal is to test whether the MD&A contains discussion of the accrual process.”
• Why is this?
• Public companies are required to file MD&A sections as an integral part of their 10-K filings
• The content of MD&A sections remains largely voluntary
• MD&A sections are reviewed by auditors only for consistency with the other parts of the annual
report

• The SEC intends that the MD&A provide investors with an opportunity to see the company through the
eyes of the management and has periodically provided guidance about the content of MD&A
disclosures
• In particular, the SEC has emphasized investors’ greater need for forward-looking disclosures than for
disclosures about past events, and it has guided companies to present any known trends, events,
commitments, plans, and uncertainties that are likely to materially affect company liquidity, capital
resources, or future operations. Voluntary projections of anticipated trends are also encouraged.
We want to predict future cash flows by looking at accruals and text. Not all investors can focus on data. So text
should help them understand it better.
MD&A is flexible to explain the data. So it is used as supportive information.
The problem with numbers in the back, they do not always predict the future.

Some firms want to report non-gaap, as they feel that not everything is disclosed there.

10-K
We have the numbers, but to understand it better you can use MD&A. Like forward looking information.

problems
They are able to use MD&A to express their concerns.
Corporate Sustainability Reporting and the MD&A
• Under the Non-Financial Reporting Directive (NFRD), adopted in 2014, companies must report on how
sustainability issues affect their performance, position, and development, and the impact of the company
on people and the environment.
• The required non-financial information can be presented in either the MD&A as part of the annual report
or in a separate report
• Firms had to report on Topics: activities on ‘environmental, social and employee matters, respect for
human rights, anti-corruption and bribery matters’.
• In Europe, the coming Corporate Sustainability Reporting Directive (CSRD) will mandate that the
MD&A include disclosures on how sustainability issues affect company performance, position, and
development and that these disclosures are aligned with the European Sustainability Reporting
Standards (ESRS) released the European Financial Reporting Advisory Group (EFRAG).
• The IFRS Sustainability Disclosures Standards (SDS) by the International Sustainability Standards
Board (ISSB) require the inclusion of sustainability-related financial disclosures in a company’s
MD&A. This is established by the IFRS Exposure Draft S1 General Requirements for Disclosure of
Sustainability-related Financial Information
• The European Sustainability Reporting Standards (ESRS) developed by the European Financial
Reporting Advisory Group (EFRAG) follow the IFRS definition of the MD&A as a complement to a
company’s financial statements. The ESRS are mandated by the Corporate Sustainability Reporting
Directive (CSRD), and will require that companies disclose material sustainability-related information
as part of their MD&A, for the same period as financial information, essentially placing sustainability
disclosures in general regulatory filings. The ESRS 1 General Requirements (Appendix G) also
specifies requirements for how sustainability information should be presented and identified as a single
section in the MD&A.

Firms are required to report something but they can present it either in MD&A or separate report, so now we
have CSRD. Now firms are mandated to report it, without flexibility of what it is about and where it is.

The question is how can we understand information in MD&A

Paper Frankel

Disclosure can be different from the numbers!

We first look at the content

• In their paper, Frankel et al. (2016) test:

1. The narrative content of accruals estimated from the MD&A (MDA Accruals) (convert it into numbers)
2. The incremental explanatory power of MDA Accruals relative to the modified Dechow and Dichev
(2002) model of accruals
3. Persistence of MD&A Accruals relative to Cash Flows (see Sloan (1996) )
4. Using SVR to predict future cash flows
• Note: dictionary methods will be covered in the tutorial paper of Frankel et al. (2022)
Can we predict current and future cash flows

Research Design:
• Support Vector Regression (SVR) is used to automatically identify patterns in a narrative that occur in
conjunction with firm fundamentals (supervised ML)
• SVR examines how specific words and short phrases in the MD&A explain reported accruals
• Frankel et al. (2016) first create a Word-Count of specific words/total words in the MD&A
• 296,329 unique words and phrases for 71,847 observation
• The number of observations is less than the number of unique words, so we cannot run OLS
regression
• The words created from the word-count then become independent variables in the SVR model
More words than observation, so regular model is not possible.
Observation – reports in every year.

We want to estimates MDS accruals. By taking the words and looking if they are useful.
Accruals t-1 W – estimates – it is suprvised here
MDA acrruals – our owm measure.
Accrual t – explain future accruals

ols
Words - x

svr
What is important and what is not
Results

The output of SVR is a coefficient

MDA accruals is significant – there is an association between what we picked up in the text and accruals.
Is MDA accruals is useful in combination with Dechow model.
It adds explanatory power.
Why do we care? Investors want to predict, text and numbers are useful to help with persistence and predicting
future cash flows

We wat to see if we can use MDA accruals to predict future earnings.

Sloan cash flows and accruals predict future earnings. Confirming with new data
MDA accruals are more persist then Sloan model.
We try to predict cash flows. Similar results.

Summary of results
• Unstructured qualitative disclosures have explanatory power for quantitative firm fundamentals
• Accruals estimated from the MD&A are associated with accruals from the financial statements
• Accruals estimated from the MD&A by machine learning algorithm Support Vector Machines (SVR)
• Results from tests of earnings persistence show that MD&A Accruals are very persistent for predicting
earnings
• The MD&A also contains information useful for predicting future cash flows
• Cash flow forecasts produced by similar SVR method as MD&A Accruals

Key learning objectives from Frankel et al. (2016)

• Unstructured and qualitative disclosures with narrative content can be very helpful to investors
• The narrative content of the MD&A may solve some of the issues with accruals from the financial
statements
• Machine learning is a helpful tool for investigating narrative content through textual analysis
• There are many machine learning algorithms, both supervised and unsupervised
• The choice of the ML method depends on the problem that needs to be solved

Tutorial 4 Brown, Crowley, and Elliott (2020)

What Are You Saying? Using topic to Detect Financial Misreporting

We use a machine learning technique to assess whether the thematic con- tent of financial statement disclosures
(labeled topic) is incrementally infor- mative in predicting intentional misreporting.

This study investigates whether a novel text-based measure of the thematic content of financial statement
disclosures (labeled as topic) is useful for detecting financial misreporting.

Detection models have long focused on quantitative financial statement and stock market variables as predictive
fac- tors (Beneish [1997], Brazel, Jones, and Zimbelman [2009], Dechow et al. [2011], Bao et al. [2020]). One
drawback of this approach is that financial misreporting can go undetected for multiple periods, because
misreport- ing firms often manipulate performance metrics and accounting transac- tions to blend in better with
their peers or the firm’s own past performance (Lewis [2013]). To address this weakness, recent studies analyze
the textual and linguistic features of management disclosures, finding that summary measures of these features
serve as useful warnings of misreporting

Despite the usefulness of communication style in revealing misreport- ing, the literature debates whether textual
and linguistic features ade- quately capture managers’ deliberate attempts to obfuscate or manipu- late financial
information (Bloomfield [2008], Bushee, Gow, and Taylor [2018]). Further, as Loughran and McDonald [2016]
highlight, commonly used textual measures do not reflect the context or meaning of manage- ment disclosures,
thereby limiting the inferences that can be drawn.

We tackle these issues by introducing a machine learning tool that simultane- ously detects and quantifies the
thematic content (topic) of annual report narratives. This approach departs from prior text-based research by
focus- ing on what is being disclosed by management rather than how.

Background and Research Questions

PREDICTING FINANCIAL MISREPORTING

recent re- search explores the predictive value of various language-based measures. The premise is that the
linguistic features of management disclosures re- veal communication patterns that foretell financial
misreporting. This lit- erature relies on two general approaches to uncover and analyze commu- nication
patterns in written disclosures

The first approach relies on predefined word categorizations (or dictio- naries) to investigate the link between
intentional misreporting and lan- guage tone as well as deception cues.

The second approach employs machine learning algorithms to discrim- inate between “bags of words” or textual
style markers that predict inten- tional misreporting. These markers include textual features, such as verbal
complexity, readability, and disclosure tone, as well as grammar and word choices.

LDA TOPIC MODELING

We use this approach to construct a firm-specific measure (topic) of the topics discussed in annual financial
statements in a given reporting year. This unique measure (defined as the normalized percentage of the annual
report attributed to each topic iden- tified by the algorithm) captures the extent to which a particular topic is
discussed within a given annual report filing.

As a result, our study seeks to provide new insights into the benefits of statistical topic analysis in assessing the
likelihood of misreporting.

We therefore ex- tend Hoberg and Lewis [2017] by employing a rolling-window estimation procedure that
accounts for the time-varying nature of the topics discussed by management. We use the same rolling-window
setup to predict misre- porting in real time using the topics identified in the window immediately preceding the
prediction period.

RESEARCH QUESTIONS

Our research questions explore the incremental value of thematic content in identifying intentional misreporting,
relative to the predictive value of quantitative financial and textual style characteristics.

Research Question 1 (RQ1): Do disclosure topics improve the detection of intentional financial misreporting,
relative to quantitative financial measures?

Research Question 2 (RQ2): Do disclosure topics improve the detection of intentional financial misreporting,
relative to aggregate textual style features?

EMPIRICAL MEASURES

Financial and Textual Style Measures.

We draw our quantitative fi- nancial statement and stock market variables from the Dechow et al. [2011] F-
score model (see model 3 in table 9 of their paper). These variables capture accrual quality, firm performance,
off-balance-sheet activities, and market pressures. We augment the F-score model with variables capturing firm
size, audit quality, and involvement in complex business transactions, namely, M&As and restructurings. These
additional variables capture char- acteristics that are correlated with reporting risks and the quality of firms’
external and internal monitoring mechanisms

LDA Topic Measure.

The LDA model is based on a few simple assumptions. It assumes a col- lection of K topics in a given document
and that the vocabulary of each topic is distributed following a Dirichlet distribution, β K ∼ Dirichlet(η). The
model further assumes that the topic proportions in each document d are drawn from a Dirichlet distribution θd
∼ Dirichlet(α). Given these assumptions, a specific number of topics to identify, and a few learning pa-
rameters, the LDA model categorizes the words in a given set of documents into well-defined topics. Because
the model uses Bayesian analysis, a word is allowed to be associated with multiple topics. This is a
distinguishing fea- ture of LDA, as words can have multiple meanings, especially in different contexts. In short,
LDA is a probabilistic process that condenses the vocab- ulary in a collection of documents into a dictionary of
topics and a set of topic weights.

We implement LDA using a dynamic time-series process, because annual report content is likely to vary over
time, due to macroeconomic or industry trends, changes in disclosure requirements, or managerial turnover

We follow Hoffman, Bach, and Blei [2010] and implement LDA using an online batch variant of the algorithm.
We draw the filings in each batch in random order to mitigate overweight- ing of early years in the online LDA
tool. We then set the algorithm to identify 31 topics in each five-year window.

Empirical Results

EVALUATION OF LDA TOPIC MEASURE

Because LDA is unsupervised, it is necessary to evaluate the algo- rithm’s effectiveness in capturing human
intuitions.

Interpretation and Labeling of LDA Topic Output.

Our first method evaluates the semantic meaning of the LDA output by labeling the topics and assessing the
extent to which they provide meaningful economic con- tent. This form of evaluation is qualitative

We derive the combined topics by matching topics across years based on the Pearson correlation of the word
weights within the topics.

To determine the underlying content of each combined topic, we gener- ate a list of the highest weighted phrases
and sentences associated with each topic.

Word Intrusion Tasks.

Our next evaluation method uses “word in- trusion” tasks to assess the semantic coherence of the unaggregated
topics derived by the algorithm across each of the rolling windows.

Chang et al. [2009] argue that the overall interpretability of LDA-derived topics can be evaluated by the extent
to which human subjects agree with the makeup of the topics. Using this logic, they develop a word intrusion
task in which hu- man subjects attempt to identify an unrelated or “intruder” word inserted into a list of words
that LDA selects as belonging to the same topic. If the set of words from the LDA model is coherent, then the
human subjects should easily identify the intruder word at a rate that is significantly higher than random chance.
Thus, a higher identification rate indicates higher interpretability of the LDA output.

The results from both procedures indicate iden- tification rates that are statistically higher than random chance
(25% or one out of four words) at the 1% level. The machine-based procedure cor- rectly identifies the intruder
word with accuracy rates ranging from 50% to 53%, while the human-subjects task produces an average
accuracy rate of 40%.

Taken together, our qualitative and quantitative evaluation methods sug- gest that the LDA algorithm provides a
valid set of semantically meaningful topics that are reasonably coherent and interpretable by human judges.

PREDICTIVE VALUE OF LDA TOPIC MEASURE

Empirical Methodology

We investigate our research questions by es- timating in-sample prediction models over rolling five-year
windows. (See panel A of figure 1.) We then conduct out-of-sample tests using the regres- sion estimates from
each five-year window to predict the likelihood of in- tentional misreporting in the year after the end of each
window.

Our first prediction model regresses misreport on vectors of the disaggre- gated topic proportions (topic) as

follows:

Similar to Dechow et al. [2011], we construct a

prediction score (p misreport) using the estimated coefficients from equa-

tion (1) and apply this scoring in our out-of-sample tests as follows:

We also introduce a comprehensive model that includes all three sets of prediction variables: topic, F-score, and
Style. We benchmark this model against a final model of F-score and Style to assess the incremental power of
topic over the full set of financial metrics and textual features. The comprehensive model is specified in
equation 3:

In-Sample Predictive Value of topic.

We first evaluate the in-sample performance of topic in detecting misreporting. For each sample, we esti- mate
annual in-sample regressions of equation (1) using the unaggregated, industry-normalized topic proportions.
For each prediction year, we present green (red) boxes if at least one subtopic for a given combined topic loads
as positive (negative) and signif- icant at the 10% level or greater, and all other subtopics are insignificant. We
code the boxes as gray (“Other”) if all subtopics for a given combined topic are insignificant, if multiple
subtopics are significant but with oppos- ing signs, or if all subtopics are dropped from the regression due to
multi- collinearity.
Predictive Value of topic Versus Financial Variables (RQ1).

Table 2 presents separate summary statistics of our financial variables for misre- ported and non-misreported
firm-years in each sample. We provide tests of differences in the means of each variable (clustered by firm)
between each set of firm-years. We find that only one of the financial variables be- haves similarly across
the three samples. Firms are more likely to issue securities during periods of misreporting in all three
samples.
Table 3 presents out-of-sample tests of the predictive role of topic and F-score. We reiterate that these tests are
conducted using the unaggregated topics (not the combined topics) discovered in each rolling window. Panels
A, C, and E present the pooled AUC statistics for the AAER, AA, and 10-K/A samples, respectively. The AUCs
across the three panels indicate that quantitative financial metrics (F-score) are signif- icant predictors of
misreporting, especially for events that trigger an enforcement action. The AUCs for the F-score model range
from 0.589 in the 10-K/A sample to a high of 0.708 in the AAER sample. All of the AUCs for the F-score
model are statistically greater than a random classification model (AUC = 0.500) at the 5% level and higher.
The AUCs for the topic model also exceed the 0.50 threshold in all three samples, indicating the ability of
thematic content to independently detect various forms of financial misreporting. The predictive value of
topic is markedly higher in the AAER sample, with a predictive gain of 18% over a random model, compared to
a 12% gain in the AA and 10-K/A samples (AUC of 0.680 in the AAER sample versus 0.616 for both
irregularity samples; statistically different at the 1% and 5% levels). This differential result suggests that
disclosure content varies strongly with instances of misreporting where there is high confidence of the
intent to mislead.

Collectively, the results in table 3 suggest that content-based information drawn from annual report narratives
improves the detection of misreport- ing beyond what can be achieved by financial metrics. Our evidence also
suggests that disclosure content and financial variables performs equally well in detecting misreporting, though
both predictors serve as comple- mentary warning signals.

The Predictive Value of topic Versus Textual Style

Table 4 presents separate summary statistics for our style characteristics. We find that many of the style features
shift inconsistently in misreporting years and, at times, contradict conventional views. For instance, misreported
filings in the AA and 10-K/A samples are more readable, relative to nonmisreported filings, whereas
misreported filings in the AAER sample are less readable. Such opposing evidence is not unique to our study, as
Purda and Skillicorn [2015] find contradictory evidence of more deceptive and negative language in filings
classified as “truthful.” These findings underscore the potential pitfalls in relying on basic textual measures to
identify financial misreporting.

Table 5 presents the pooled AUC statistics for out-of-sample tests of the predictive performance of topic,
relative to the performance of the textual style (Style) vector. Panels A and B present the test statistics for
AAERs; pan- els C–F present the results for the two irregularity samples. The benchmark- ing tests for the
AAER and AA samples indicate that topic by itself better predicts misreporting than the standalone Style model
(3.1% and 3.5% in- crease in AUC, in panels B and D, respectively; Wald p-values = 0.01 and 0.06). The topic
model, however, performs worse than the Style specification in the 10-K/A sample (panel F).

We observe that a joint model of topic and Style outperforms the individ- ual Style model by almost 6% in the
AAER sample (panel B, Wald p-value = 0.00). Both models perform at the same level in the AA and 10-K/A
sam- ples (panels D and F, respectively; Wald p-values = 0.13 and 0.27). Although the joint topic and Style
model dominates in the AAER sample, the basic topic model achieves an accuracy level in the AA sample that
ranks just as high as the joint model. Thus, the predictive value of topic over Style is quite strong when detecting
misreporting events that trigger an enforcement action or a financial restatement.

Joint Predictive Value of topic, Financial Variables, and Textual Style. We next examine the interplay between
all three sets of predictors: topic, F-Score, and Style
In table 6, we find that the three-vector model performs well in detecting misreporting out of sample. The AUCs
across the three samples are well above the 0.50 threshold, ranging from 0.635 in the AA sample (panel C) to as
high as 0.752 in the AAER sample (panel A). The benchmarking tests indicate that the addition of topic to the
F-score and Style model improves predictive accuracy by 3.2% for AAERs (Wald p-value = 0.00 in panel B)
and 2.8% for irregularity restatements (Wald p-value = 0.01 in panel D). The incremental value of topic is
insignificant in the 10-K/A sample when we compare the three-vector model to the joint F-score and Style
model. We also find that the three-vector model does not perform any better than the joint topic and F-score
model in the AAER and AA samples (Wald p-value = 0.13 and 0.80 in panels B and D) or the topic and Style
model in the 10-K/A sample (Wald p-value = 0.74 in panel F). This evidence corrobo- rates our previous results:
topic and F-Score are strong predictors of AAERs and restatements, whereas topic and Style provide robust
power for detecting broad misrepresentations and disclosure omissions.

THE ECONOMIC SIGNIFICANCE OF topic

Table 7 reports the percentage and number of filings that are correctly classified as misreported using multiple
models. Panel A reports results for the AAER sample; panels B and C report results for the AA and 10-K/A
sam- ples, respectively. For the AAER sample, we find that the basic F-score model correctly classifies 72.51%
of misreported filings at the 50th percentile cut- off (total count of 319 filings). The joint model of topic and F-
Score or, alternatively, the three-vector model performs better, flagging about 79% of misreported filings on
average. When we focus on high-risk prediction scores, we observe that the three-vector model captures the
most misreport- ing events at the 90th and 95th percentiles (32.90% and 22.44% or total counts of 150 and 96
misreported filings, respectively).

To quantify the economic value of topic, we note that the classifica- tion rate at the 95th percentile improves by
59% when topic is added to the benchmark F-score and Style model (accuracy rate of 22.44% versus
14.11%).25 In terms of raw numbers, we capture 26 additional misreporting firms at the 95th percentile when
topic is added to the benchmark model. This relative improvement is striking when we consider the low
frequency of AAER filings and the high costs associated with misreporting events that are not detected by
traditional prediction models (Beneish and Vorst [2019]).26 The classification rate at the 90th percentile also
increases by 36% when topic is added to the benchmark model (32.90% versus 24.28%).

This increase leads to 33 more filings being correctly classified at this cut- off. These findings corroborate our
inference that topic is incrementally valuable in detecting misreporting. The same interpretation holds when we
evaluate the detection rates at the 99th percentile based on the NDCG@k measure. Here, the inclusion of topic
in the three-vector model improves detection accuracy by 15%, compared to the benchmark F-score and Style
model (0.188 versus 0.163).

Our results for the AA sample (panel B) further demonstrate the eco- nomic value of our topic measure. We find
that topic increases detection accuracy by 50% at the 95th percentile cutoff when added to the bench- mark
model (classification accuracy of 9.68% versus 6.44%). We also ob- serve that topic by itself or joint models of
topic paired with F-score or Style perform just as well or higher than the three-vector model at the 90th and 95th
percentile cutoffs. Thus, there is little value in adding F-score or Style as a third predictor of high-risk
restatements, once topic is included in the model. In panel C, the three-vector model is most efficient at
detecting high-risk events in the 10-K/A sample. For instance, the inclusion of topic in the three-vector model
improves classification accuracy by roughly 6% at the 95th percentile cutoff when benchmarked against the
joint F-score and Style model (16.83% versus 15.94%). This improvement equates to an ad- ditional seven
filings being correctly classified as misreported. Collectively, the results in table 7 suggest that the incremental
predictive power of topic is economically significant and that its value is quite salient when detecting high-risk
reporting practices.

Tutorial 4 Frankel, Jennings, Lee (2022)

Disclosure Sentiment: Machine Learning vs. Dictionary

Methods
We compare the ability of dictionary-based and machine-learning methods to cap- ture disclosure sentiment at 10-K filing
and conference-call dates.

Our re- search follows Loughran and McDonald (2011) who develop a measure of disclosure sentiment based
on a business-context dictionary (LM dictionary) and pro- vide evidence of a stronger association with 10-K fil-
ing date returns than a measure based on the Harvard Psychosociological Dictionary (Harvard Dictionary).

Our re- search follows Loughran and McDonald (2011) who develop a measure of disclosure sentiment based
on a business-context dictionary (LM dictionary) and pro- vide evidence of a stronger association with 10-K fil-
ing date returns than a measure based on the Harvard Psychosociological Dictionary (Harvard Dictionary). 1
Although dictionary methods are easy to implement, they have flaws that machine-learning methods can
alleviate. For example, dictionaries do not reflect lan- guage change over time or across industries; do not al-
low for variation in word importance; can be costly to update; and are subject to researcher subjectivity, which
may lead to bias, incompleteness, or overfitting.

We provide evidence that machine-learning meth- ods yield an improvement over the LM sentiment mea- sure
in capturing returns at both 10-K filing dates and conference-call dates—an improvement similar to that of the
LM dictionary over the Harvard Dictionary. Moreover, the ability of dictionary-based methods to capture
sentiment at 10-K filing dates is time-period specific. Using the Loughran and McDonald (2011) sample period,
we confirm their results; but we find the LM and Harvard sentiment measures are unassociated with 10-K filing
returns when the sample peri- od is extended to 2019. In contrast, machine-learning methods consistently
explain 10-K filing and confer- ence call returns in both samples. We also compare multiple machine-learning
models and find that the random-forest regression-tree method explains 10-K and conference call returns better
than the support- vector regression and supervised-latent-Dirichlet al- location methods. Overall, our results
suggest that machine-learning methods offer a more powerful and reliable measure of disclosure sentiment than
dictionary-based methods.
One hurdle in this line of research is quantifying and sum- marizing qualitative disclosure content. Studies use
different techniques to extract disclosure signals such as readability, similarity, and sentiment.

We use the LM dictionary-based sentiment measures as a benchmark of comparison because of their wide use.2

Prior Literature

Manual classification is beneficial because it can be more precise (Li 2010). However, it also has its dis-
advantages. First, manual classification is time con- suming, leading to studies with small sample sizes, limited
scope, and reduced statistical power. Second, manual coding involves subjectivity, reducing its rep- licability
and limiting follow-up studies.

The accounting literature adopted alternative meth- ods to classify and extract information in disclosures to
improve replicability and applicability. These in- clude counting words based on dictionaries (e.g., tone,
competition)4 and using alternative computa- tional linguistic algorithms to measure specific text characteristics
(e.g., readability, cosine similarity)

Dictionary-based approaches also present drawbacks. First, dictionaries depend on context. Second, researcher
judgment in the creation of these methods can reduce their reliability

Prior literature has primarily used two dictionaries to create sentiment measures. The first is the Harvard
Psychosociological Dictionary, which identifies posi- tive and negative words that are used to calculate the
sentiment of a document or disclosures.

Because words can have many different meanings depending on the context, Loughran and McDonald
developed a sentiment dictionary specific to financial documents.

The first applications of automated machine learn- ing in the accounting literature rely upon manual text
classification to aid in the learning process

A natural progression in this literature is the imple- mentation of automated machine-learning methods, which
rely less on researcher judgment and instead al- low an algorithm to categorize a disclosure. Rather than
focusing on ill-defined constructs, such as “tone” or “readability,” automated methods identify topics or words
contained in a disclosure that can be used to predict. The most widely used automated machine- learning method
in the accounting and finance litera- ture is unsupervised latent Dirichlet allocation (LDA) (Blei et al. 2003). 6
Supervised latent Dirichlet allocation is an alternative topic modeling approach that creates topics that predict a
dependent variable (Blei and McAuliffe 2007). Alternative supervised methods (i.e., methods that are trained to
predict a dependent vari- able) include support vector regression and random forest regression trees.

Measures of Narrative Content

We compare the ability of dictionary-based and machine-learning measures to capture disclosure sen- timent of
both 10-K reports and earnings conference call transcripts. Although these two disclosures dis- cuss financial
results, they are different in two funda- mental ways. First, the conference call is timelier than the 10-K. Second,
conference calls are typically more spontaneous and dynamic than the 10-K.

Dictionary-Based Sentiment Measures

Following prior research, we use two dictionaries to measure disclosure sentiment or tone: the Harvard
Psychosociological Dictionary and the Loughran and McDonald (2011) business-context dictionary. We
calculate positive, negative, and net tone using the words included in each dictionary

Supervised Machine-Learning Senti- ment Measures

We create machine-learning measures of disclosure sentiment for both 10-K filings and conference calls using
three machine-learning models: support vector regression, supervised latent Dirichlet allocation, and random
forest regression trees. We use these models to map the counts of all one- and two-word phrases contained in
each disclosure

SVR allows the estimation of a unique weight for each one- and two-word phrase count that is included in the
disclosure (e.g., 10-K or conference call). SVR esti- mates weights on each one- and two-word phrase by
simultaneously minimizing both the coefficient vector magnitude and the prediction error. The simultaneous
minimization of the coefficient vector magnitude and the prediction error works to reduce overfitting.

sLDA categorizes the words and phrases of a dis- closure into a set of latent (i.e., unknown) topics that are
predictive of a dependent variable (Blei and McAuliffe, 2007). The algorithm assumes that all dis- closures
share the same set of topics; however, the mix of each topic varies by disclosure. The algorithm sorts the words
and phrases from the disclosures into topics based on the probability of words co-occurring within the
disclosures while simultaneously creating topics that are predictive of a dependent variable.

We also use RF to predict 10-K and conference call returns. The standard regression tree method uses an
iterative process called binary recursive partitioning, which creates a decision tree by recursively partition- ing
observations based on features (e.g., one- and two-word phrases) to predict a specific value or char- acteristic.
Each partition made by the algorithm is identified with a node (or branch), which is a binary classification of the
data using one of the data set’s features. At each node, the algorithm examines each of the remaining binary
splits of the data using the re- maining features and chooses the feature that mini- mizes the sum of squared
errors within each partition. The algorithm continues to partition the data using nodes until the number of
observations within each partition falls below a prespecified number (e.g., two or five observations) or when the
sum of the squared errors within the partition is equal to zero. When the process stops, the average value of the
response value at each terminal (i.e., final) node represents the pre- dicted value of the response given the
preceding bina- ry partitions of the data.

Sample and Main Results

10-K Sample—Dictionary-Based Senti- ment Measures

We next examine whether the sentiment measures based on the LM dictionary explain the two-day re- turn on
the 10-K filing date. Similar to Loughran and McDonald (2011), we use the following equation.

The results are re- ported in Panel A of Table 3. Consistent with Loughran and McDonald (2011), we find a
significantly negative (1% level) coefficient on LM 10-K NEG TONEi,t equal to −0.239 using the early sample
period (column (3)). We then extend the sample to include firm-year observa- tions between 1996 and 2019 and
fail to find a signifi- cant coefficient on LM 10-K NEG TONEi,t (column (6)). These results suggest that
negative tone measures based on the LM dictionary do not consistently relate to stock market reactions at
the 10-K filing date.
Although Loughran and McDonald (2011) focus on negative tone measures to capture sentiment, we also
examine whether positive and net tone measures ex- plain 10-K filing returns. Consistent with expectations, we
find a positive and significant (1% level) coefficient equal to 0.003 on LM 10-K TONEi,t in column (1) using the
early sample period. Similar to LM 10-K NEG TO- NEi,t, we cease to find a significant coefficient on LM 10-K
TONEi,t when the sample is extended to 2019 (col- umn (4)). These results reinforce the inference that the
LM dictionary does not consistently capture investor responses to the 10-K. Contrary to expectations, we
find a negative and significant (5% level) coefficient on LM 10-K POS TONEi,t when estimating the regression
using the early sample (column (2)) and an insignificant coef- ficient when extending the sample to 2019.

Given the insignificance of the sentiment measures based on the LM dictionary for the full sample, we next
examine whether sentiment measures based on the Har- vard Dictionary provide different implications. We re-
estimate Equation (1), replacing SENTIMENTi,t with the Harvard measures (HARV 10-K TONEi,t, HARV 10-K
POS TONEi,t, HARV 10-K NEG TONEi,t). We present the regression results using sentiment measures based on
the Harvard Dictionary in Panel B of Table 3.

We report redundant results from Panel A in Table 3 for measures based on the LM dictio- nary for ease in
comparison.

In the early sample (1996–2008), none of the coeffi- cients on the sentiment measures based on the Harvard
Dictionary are significant at conventional levels. In the full sample (1996–2019), none of the coefficients on the
Harvard measures have the expected sign at convention- al significance levels. Contrary to expectations, the
coeffi- cient on HARV 10-K TONEi,t is significantly negative (1% level) and the coefficient on HARV 10-K
NEG TONEi,t is significantly positive (1% level). Consistent with the gen- eral takeaway from Loughran and
McDonald (2011), these results suggest that the sentiment measures based on the Harvard Dictionary are
not well suited for measuring disclosure sentiment in a 10-K setting.

10-K Sample—Machine-Learning Senti- ment Measures

We re-estimate Equation (1) replacing SENTIMENTi,t with RF 10-K CAR[0,1]i,t, SVR 10-K CAR[0,1]i,t, sLDA
10-K CAR[0,1]i,t, and ML 10-K CAR[0,1]i,t. We present the results in Panel B of Table 3

We find that the coeffi- cients on RF 10-K CAR[0,1]i,t, sLDA 10-K CAR[0,1]i,t, and ML 10-K CAR[0,1]i,t are
positive and statistically significant at the 1% level in both the early (i.e., 1996–2008) and full sample (i.e.,
1996–2019). The coefficient on SVR 10-K CAR[0,1]i,t is statistically signifi- cant at the 10% level in the early
sample but insignifi- cant in the full sample. When inspecting the adjusted- R2 for the full sample period, we
note that RF 10-K CAR[0,1]i,t yields the highest adjusted-R2 (0.493%) across all dictionary-based and machine-
learning meas- ures included in Panel B of Table 3, including the com- bined machine-learning measure (ML
10-K CAR[0,1]- i,t).11 The adjusted-R2 for the model including RF 10-K CAR[0,1]i,t is statistically higher than
the adjusted-R2 for the model including LM 10-K TONEi,t using a Vuong likelihood ratio test (p-value equal to
0.059); however, the economic significance of the difference in explanatory power is not large. The relatively
small ad- justed-R2s in the 10-K setting are likely due to the weak stock market reaction to 10-K filings
documented in Li and Ramesh (2009). Thus, any improvement in mea- surement error for sentiment
measures is unlikely to yield economically large improvements in explanatory power in the 10-K market
reaction tests.

To put these results in perspective, in Figure 1 we graph the average two-day 10-K filing return (10-K
CAR[0,1]i,t) for quintiles based on HARV 10-K TONEi,t, LM 10-K TONEi,t, and RF 10-K CAR[0,1]i,t. These
sentiment meas- ures yield the highest adjusted-R2 when using the LM dic- tionary, Harvard Dictionary, and
machine-learning meth- ods. We expect to observe a monotonic increase in average returns as the quintiles for
each measure increase if these measures capture 10-K filing sentiment. A steeper line suggests that the
sentiment measure better captures 10-K filing returns. Both HARV 10-K TONEi,t and LM 10-K TONEi,t yield
relatively flat lines, whereas the line for RF 10-K CAR[0,1]i,t is steeper and monotonic, with the ex- ception of
the fifth quintile for which there is a slight de- crease in the average return from the fourth to fifth quin- tiles.
These results provide graphical evidence that RF 10-K CAR[0,1]i,t better captures 10-K filing returns
than the dictionary-based measures and thus is a superior measure of 10-K sentiment, although we
recognize the inherent limitation of market reaction tests in the 10-K setting.

Additional Analyses

Future Earnings Surprises and Future Cumulative Abnormal Returns

We therefore examine whether the dictionary-based sentiment measures and machine- learning measures predict
earnings surprises and stock market returns at the next earnings announce- ment date.

We define FUT EARN SURPi,t as the differ- ence between actual earnings per share reported in
Frankel, Jennings, and Lee: Disclosure Sentiment Management Science, 2022, vol. 68, no. 7, pp. 5514–5532, © 2021 INFORMS

quarter q + 1 less the mean analyst estimate of earn- ings per share available immediately before the quar- ter q
+ 1 earnings announcement date scaled by stock price at the end of quarter q. We define FUT EA CAR[0,1]i,t as
the cumulative market-adjusted return forfirmiondaytanddayt+1relativetotheearn- ings announcement date in
quarter q + 1. Using a sim- ilar process that we use to train the machine-learning models for concurrent returns
in our previous analy- sis, we retrain the RF models to predict both FUT EARN SURPi,t and FUT EA
CAR[0,1]i,t. We create sep- arate predictions for each of these variables using both the text of the 10-K filing
and the text of the con- ference call transcript. We label the predictions of the future earnings surprise as RF
FUT EARN SURPi,t and the predictions of the future earnings announcement cumulative abnormal return as RF
FUT EA CAR[0,1]i,t.

In Table 5, we report results of estimating models including FUT EARN SURPi,t as the dependent vari- able and
each disclosure content measure (LM TONE, HARV TONE, and RF FUT EARN SURP) as the prima- ry
independent variable of interest. We include all pri- or control variables in the models with the addition of
DISPERSION and REVISIONS, following Loughran and McDonald (2011). The sample size for this analy- sis
reduces to 53,749 observations with the inclusion of these additional control variables. In the first three
columns, we present results using the 10-K sample. We find the coefficient on the RF FUT EARN SURP
measure is positive and significant, which suggests that this measure predicts future earnings surprises.

In the last three columns, we present results using the confer- ence call sample. We find that the coefficient on
the RF FUT EARN SURP measure is positive and significant, but the coefficients on the LM CC TONE and
HARV CC TONE measures are insignificant. The adjusted-R2 of the model including RF FUT EARN SURP is
equal to 5.791%. Overall, these results suggest that only the machine-learning sentiment measures are
positively correlated with information that is unexpected at the subsequent earnings announcement date.

We first create equal-weighted FUT EA CAR[0,1]i,t portfolio returns for quintiles based on each 10-K dis-
closure measure (LM 10-K TONEi,t, HARV 10-K TONEi,t, and RF FUT EA CAR[0,1]i,t). To avoid look- ahead
bias, we assign observations to quintiles based on rankings relative to all other observations in the sample over
the prior 365 days. We expect the average return in each quintile to increase monotonically from the first to fifth
quintile for each disclosure sentiment measure. We report the results in columns (1)–(3) of Table 6. We graph
the average return for each quintile in Figure 3 to more easily detect monotonic increases in average returns
across quintiles. We do not find evi- dence of a monotonic increase in average returns from the first to fifth
quintiles of the LM 10-K TONEi,t and HARV 10-K TONEi,t measures.
When we subtract the average return in quintile one from the average return in quintile five, we find that the
change in portfolio re- turns for the LM 10-K TONEi,t and HARV 10-K TONEi,t measures are negative, which is
opposite expectations. These results suggest that neither the LM nor Harvard Dictionaries yield positive
hedge portfolio returns.

In contrast, consistent with expectations, we find that RF FUT EA CAR[0,1]i,t is monotonically increasing
from the first to fifth quintile in Figure 3.

We perform a similar analysis with the conference call sample in columns (4)–(6) of Table 6 and Figure 4. We
report and plot the average return for each quin- tile of LM CC TONEi,q, HARV CC TONEi,q, and RF FUT EA
CAR[0,1]i,q. Figure 4 provides graphical evi- dence that RF FUT EA CAR[0,1]i,q yields a monotonic
increase in average returns across quintiles and also represents the steepest line. The line for LM CC TO-
NEi,q is increasing nearly monotonically but is less steep, and the line for HARV CC TONEi,q is nearly flat or
somewhat decreasing.
Tutorial 4

Tutorial week 4
• Unstructured and qualitative disclosures: understanding financial reporting narratives using machine
learning
• Papers:
1. Brown et al. (2020): “What are You Saying? Using topic to Detect Financial Misreporting?”
 Can users of financial statements predict misreporting based on the reporting topics that are
disclosed in the narrative of the annual report (10-K)?

1. Brown et al. (2020) examine how the narrative of the 10-K can be used to detect financial
misreporting. They argue that (p. 238) “commonly used textual measures do not reflect the context or
meaning of management disclosures”. Explain what they mean with the “context” of management
disclosures.

Context
We want to detect misreporting. If we know that managers want to do earnings management, but then
managers know, because we are looking at the numbers. But then if he does not want to influence
numbers, he/she can influence with words.
• Qualitative disclosure is argued to improve financial misreporting detection models beyond quantitative
financial statement and stock market variables, because firms know how to make financial reporting
variables appear “normal”
• misreporting firms often manipulate performance metrics and accounting transactions to blend in better
with their peers or the firm’s own past performance
• However, it remains a question if commonly used measures of qualitative disclosure adequately capture
intentional misreporting
• For instance aggressive revenue recognition in the case of Under Armour in Week 1
Will there be a difference in data, if managers want to misreport.

Intentional suggest that misreporting happens with a specific objective, or context

• Just looking at how the revenue number is reported is not likely to capture the motivation of the manager
• Rather, one should look at what is said in conjunction with how revenue is reported

• In their investigation, the SEC stated that Under Armor was engaged in the practice of pulling forward
about $408 million in orders over six consecutive quarters, beginning in the third quarter of 2015. The
sales involved orders that customers had requested be shipped in future periods.
Numbers seem abnormally high and abnormally low after. To avoid it , when managers misreport they try
to make numbers to look normal all around. Normal compared to past performance or to similar firms.

Selective disclosure
Under Armour needed to pull forward sales orders to meet quarterly target. The company offered
incentives such as discounts or extended payment terms to wholesale customers.. In one case, in
September 2016, a company that was asked to commit to buying more goods replied: “We just brought a
bunch of your goods in early to help out your quarter...Now you want more...More..More..more..30%
[price discount] please.” Under Armour gave the company a 25% discount and an extra month to pay, the
SEC said in the order.
• The SEC claimed in their investigation that Under Armour failed to disclose that it was pulling forward
orders from future quarters, a practice that allowed it to meet Wall Street’s revenue estimates.
The question is can you do it or not. Yes you can, but you need to disclose it.

They lied about the interest in product part because they just offered discounts.
So you can say in forecast: we believe there is a high demand and we will be earning more.
You can figure it out by looking at the demand of other firms.

So in this paper we want to see how words reflect the misreporting of manager.
How do we do it?
Two approaches to analyze text
• To address the weakness of quantitative measures, recent research explores the predictive value of
various language-based measures.
• The premise is that the linguistic features of management disclosures reveal communication patterns that
foretell financial misreporting. This literature relies on two general approaches to uncover and analyze
communication patterns in written disclosures.
• Approach 1: predefined word categorizations, or dictionaries, to investigate the link between intentional
misreporting and language tone as well as deception cues:
• the relative frequency of self-reference words, causation words, positive emotion words, and future tense
verbs
• negative and uncertain language words
• deceptive language, such as emotion words and anxiety words
The problem is in what you put in the dictionary, you can miss some information

• The problem is that the meaning of a word depends on the context in which the word is used:
• For example, the word “bank” in the sentences “They obtained a loan from the bank” and “They met at
the river bank” is represented by different contexts
• Approach 2: use machine learning to discriminate between “bag-of-words”, or textual style markers
• These markers include textual features, such as verbal complexity, readability, and disclosure tone, as
well as grammar and word choices.
• Context reflects the meaning of the words
• Context refers in the paper of Brown et al. (2020) to a coherent set of semantically meaningful topics
We are trying to figure out if the story CEO is giving can help us detect misreporting by analyzing what he
said. Looking at the meaning

2. To examine context, they use a topic modelling approach. In simple words, explain how they
empirically determine the topic of the management disclosures. What is the primary outcome of the
machine learning methodology they use to determine topic?

Topic Modelling
Brown et al. (2020) generate a measure of topic using Latent Dirichlet Allocation (LDA)
• Bayesian topic modelling algorithm
LDA is for instance used by internet search engines to increase the probability that the word you type in
the search box is associated with a specific website
The LDA technique identifies the thematic structure of text using a Bayesian probabilistic model
• It assumes a collection of a specific number of topics (K topics), and that the language used to express
these topics follows a known distribution (Dirichlet distribution)
• Furthermore, the proportion of each topic in a document is drawn from this distribution.
• LDA uses the knowledge of this distribution to categorizes the words in the text into well-defined topics
It looks for structures in the text that refer to the theme.
Lda first reduce number of words, then looked at connection between words and combination of words. It
optimizes finding the topics.

Given assumptions, a specific number of topics to identify, and a few learning parameters, the LDA model
categorizes the words in a given set of documents into well-defined topics. Because the model uses
Bayesian analysis, a word is allowed to be associated with multiple topics.

LDA is a way to earn from the text, by using a bag of words and finding topic themes. But still you (not
machine) can give different labels on topics.
LDA is an unsupervised and unstructured probabilistic model that “learns” or discovers the latent thematic
structure of words within a corpus of documents.
• LDA is a “bag of words” algorithm that uses the distribution of words across documents to classify and
quantify themes without the need for predefined or researcher-determined word lists or topic categories.
• However, human judgment is necessary to interpret and label the topics inferred from the algorithm,
because the LDA output for a given topic consists only of word clusters and word probabilities (also
referred to as word weightings)
• In summary: LDA simultaneously detects and quantifies the thematic content (topic) of annual report
narratives.
• The primary output of LDA is a firm-specific measure of the topics discussed in annual financial
statements in a given reporting year (topic) , defined as the normalized percentage of the annual report
attributed to each topic identified by the algorithm, and captures the extent to which a particular topic is
discussed within a given
annual report filing.

LDA topics
What are we gonna do with this data? We want it to help us detect misreporting with it.
Using topic to detect financial misreporting
• In their paper, Brown et al. (2020):
1. evaluate semantic validity of the LDA output and the ability of topic to detect misreporting
2. Assess the usefulness of topic to detect misreporting compared to a comprehensive set of quantitative
financial and textual style variables >> this is the focus of this tutorial
3. The classification accuracy of their empirical models
• Note: Brown et al. (2020) examine both the topic and style of qualitative disclosures
We want to look at the usefulness of topics.
They take into consideration that people have different style of talking. So they want to correct for it. That
is why they analyze the style of qualitative disclosures.

Research Questions
RQ1: Do disclosure topics improve the detection of intentional financial misreporting, relative to
quantitative financial measures?
Does text helps detect misreporting relative to quantitative data.
RQ2: Do disclosure topics improve the detection of intentional financial misreporting, relative to aggregate
textual style features?

3. To evaluate the predictive value of topic, Brown et al. (2020) compare its predictive value to other
variables. Explain in you own words how topic differs from the other variables used to evaluate the
predictive value of topic. What is the primary statistic of interest they use to draw their conclusion?

Research design
Brown et al. (2020) use LDA to assess whether the thematic content of financial statement disclosures
(topic) is incrementally informative in predicting intentional misreporting, relative to the predictive value
of quantitative financial and textual style characteristics (p.246).
• How is intentional misreporting identified?
• AAER: This measure is also used in Amiram et al. (2015) in week 1
1. Accounting and Auditing Enforcement Releases that refer to SEC enforcement actions (for
instance the case of Under Armour – the exchange committee in US released the statement that the
company did something wrong, then everyone knows that that organization is going after you).
• Irregularities This measure is also used in Amiram et al. (2015) in week 1
2. Restatements from Audit Analytics (AA) database in WRDS: restatements categorized as
“fraudulent” or “nonclerical” or those associated with a regulatory investigation
3. Amended annual 10-K/A filings: the original 10-K filing was misreported (due to “fraud” or
“irregularity” or those associated with a regulatory investigation)
• p. 240: “There is significant overlap between our three data sources”
We want to identify which of these 3 measures detects misreporting better.

AAERs vs Irregularities
• AAERs measure with high confidence intentional misreporting, because the SEC targets firms when
there is strong evidence of intent to mislead.
They go after the firm only when they have strong evidence, because they do not have enough staff.
• The AAER data also outperform restatement samples in identifying infractions that meet the statutory
definition of securities fraud.
• However, one drawback is that many instances of misreporting are not pursued by the SEC, due to
resource constraints
• Another shortcoming is that cases pursued may reflect selection biases arising from the SEC’s evaluation
and investigation processes.
• Irregularities (restatements) mitigate these limitations, because the events are drawn from broader
sources. These samples could, however, introduce other selection or identification issues as the procedures
for gathering these data depend on how firms disclose and discuss misreporting events in filings.

Brown et al. (2020) examine if topic is incrementally informative to :

• F-Score: Financials from the Dechow et al (2011) model. This measure is also used in Amiram et al.
(2015) in week 1
• Style: various measures of textual complexity, language voice, and disclosure tone.

We want to see if text is better than style and F-score.

F-Score and Style
• F-Score: Dechow et al. (2011) conduct a comprehensive analysis of the quantitative financial and stock
characteristics of misreporting firms identified in AAERs. They find that AAER firms have low- quality
accruals, deteriorating financial and nonfinancial performance, and high stock valuations, relative to
fundamentals. AAER firms are also more likely to engage in aggressive off-balance-sheet and financing
activities during the misstatement period. Using these characteristics, Dechow et al. (2011)
develop a composite prediction measure (F-score)
• Brown et al. (2020) use the variables from the F-Score model as quantitative financial measures
(RQ1)
• Style: a comprehensive set of textual features include common proxies for readability, textual
complexity, and disclosure tone, deeper linguistic markers, such as language voice, lexical variety, and
disclosure emphasis (RQ2)

Research Design: standalone model

We train data to see if topics pick up misreporting.

Research design: Comprehensive model

We cannot use the whole data, because if we try to predict if the person will attend the lecture in week 4,
by using data through the weeks 1-4.
Research design: test statistic
• The test: search for the measure with the highest predictive ability, as measured by the area under the
receiver operating charicteristics (ROC), the AUC.
• The highest AUC reflects the measure that best predicts intentional misreporting
• AUC is comparable to the ଶ in in OLS regression
• The Area Under the ROC curve (AUC) is a widely used indicator of a model’s predictive ability
• The AUC values can range from 0.50 to 1 (1- is highest predictive value) and represent the probability
that a randomly chosen positive instance of misreporting will be ranked higher by the respective model,
compared to a randomly chosen negative instance
• Any reasonable detection model should have an AUC greater than 0.5 (i.e., the model should perform
better than a random classification model)

increase in misreporting from year to year,

with peak in 2003. After SOX it decreased.

Different topics and when they are linked to misreporting.

We see a predictive ability of 68%. The topic is useful. F-score had a 70% predictive ability. The
difference between then is not significant.
Topic has a higher predictive ability rather than the style - tone. The difference is significant.

Topic compared to the joint model doesnt have a significant influence

Summary and conclusions from Brown et al. (2020)
• Detecting intentional misreporting requires both quantitative and qualitative disclosures
• Textual analysis has limitations
• The focus on How firms disclose is inadequate; What is disclosed in also important
• Latent Dirichlet Allocation (LDA) is a machine learning algorithm that can be used to measure latent
topical content
• An empirical test of incremental predictive ability is required
• AUC is used as a benchmark to measure predictive ability
• Main test: logit regression on likelihood of intentional misreporting
• Topic does not appear to outperform quantitative disclosures (F-Score) on its own

2. Frankel et al. (2022): “Disclosure Sentiment: Machine-Learning vs. Dictionary Methods”

 Comparison of narrative disclosure measures of sentiment based on textual analysis using
dictionary methods and machine learning methods

4. Frankel et al. (2022) measure disclosure sentiment and compare a dictionary-based method to measure
sentiment to a machine-learning based method. Explain in your own words the difference between these
two methods to measure disclose sentiment.

Dictionary Method vs Machine Learning

• Dictionary Method
• counting words based on dictionaries
• Harvard Psychosociological Dictionary
• Loughran and McDonald Dictionary
• Machine Learning
• Adopting techniques in natural language processing from fields like computer science,
linguistics, and artificial intelligence
• Support Vector Regressions
• Latent Dirichlet Allocation
• Random Forrest

Dictionary Method

• Dictionary Method
• Dictionaries depend on context
• Research judgement on which words to include and exclude from the dictionary
• Narrowly applied to specific constructs
• Research Design
• POS (NEG) TONE: the sum of positive (negative) words/ total words in the disclosure

Machine Learning

• Machine Learning
• Reduce research subjectivity
• Broader scope of data is possible (e.g. pictures)
• Low costs
• May require manual text classification to aid in the learning process
• Same issues that lead to dictionary methods
• Automated methods
• May identify words with little economic intuition
• Research design:
• 3 Supervised learning methods: SVR, LDA (see before), and Randon Forrest (RF)

Random Forrest

The standard regression tree method uses an iterative process called binary recursive partitioning, which
creates a decision tree by recursively partitioning observations based on features (e.g., one- and two- word
phrases) to predict a specific value or characteristic.
• Random Forrest is an application of the regression tree method, combining many regression trees

Dictionary Method vs Machine Learning

• Dictionary Method
• counting words based on dictionaries
• Harvard Psychosociological Dictionary
• Loughran and McDonald Dictionary
• Machine Learning
• Adopting techniques in natural language processing from fields like computer science,
linguistics, and artificial intelligence
• Support Vector Regressions
• Latent Dirichlet Allocation
• Random Forrest

Summary and conclusions from Frankel et al. (2016)

• Textual analysis has improved over time
• From manual classification to automated classification
• Two main methods: Dictionary methods vs Machine Learning methods
• Both methods have benefits and drawbacks
• Empirical question which methods best classifies Sentiment
• Stock returns are used as a benchmark
• Main test: 10-K announcement stock return
• Machine learning seems to outperform the Dictionary methods

5. Frankel et al. (2022) use a company’s stock returns around the filing of the 10-K as a benchmark
measure to assess which method better captures sentiment. Do you agree with this choice to use stock
returns as the benchmark? Why, or why not?
Machine learning outperforms the dictionary.

Lecture 5 Basu (1997)

The conservatism principle and the asymmetric timeliness of earnings

I interpret conservatism as resulting in earnings reflecting 'bad news' more quickly than 'good news'. This
interpretation implies systematic differences between bad news and good news periods in the timeliness
and persistence of earnings. Using firms' stock returns to measure news, the contemporaneous sensitivity
of earnings to negative returns is two to six times that of earnings to positive returns. I also predict and find
that negative earnings changes are less persistent than positive earnings changes. Earnings response
coefficients (ERCs) are higher for positive earnings changes than for negative earnings changes, consistent
with this asymmetric persistence.
Fig. 1illustrates the predictions tested in this paper. Consider a firm receiving news that changes its
estimate of the productive life of a fixed asset. Panel A depicts the effects of the news on the reported book
value of the asset, and Panel B depicts the associated effect on the firm's earnings. If the new estimated life
is longer, the firm is economically better off, but under historical cost accounting no gain is recorded
currently. Instead, the depreciation charges that would have been taken in the current and future periods are
spread out over the new remaining life, resulting in lower depreciation charges in Panel A, and higher
income in Panel B, each year. 2 If the expected life decreases, a symmetric treatment would increase
depreciation charges over the entire shorter remaining life. In practice, however, the accountant records an
asset impairment (Panel A), which results in (often sharply) reduced current income, but no effect on
future income in Panel B.

In short, reported earnings responds more completely or quickly to bad news than good news. My first
prediction is that earnings is more timely or concurrently sensitive in reflecting publicly available
'bad news' than 'good news'. To test this prediction of asymmetric timeliness, l use negative and positive
unexpected annual stock returns to proxy for ~bad news' and 'good news', respectively.

My second prediction is that the concurrent earnings-return association is relatively stronger than
the concurrent cash flow-return association for publicly available 'bad news' compared to 'good
news'. In the example above, news about the expected life of the fixed asset has little or no effect on
current cash flows, but 'bad news' is reflected in current earnings. Therefore, current cash flows are
weakly associated with current returns for both bad and good news, whereas accruals are more
likely to recognize current 'bad news' and incorporate it in current earnings.

Asymmetric timeliness in news recognition is expected to manifest itself also as asymmetric persistence in
earnings. Since accountants typically report the capitalized value of bad news as losses, bad news earnings
is more timely but less persistent. In contrast, good news is reflected in earnings on a less timely basis, but
good news earnings tends to be more persistent. Good news earnings is less timely because accountants
require more verifiable information before they recognize good news. But good news earnings is more
persistent than bad news earnings because the capitalized value of the good news is only partially reflected
in current earnings, and after verification, is also reflected in sub- sequent earnings.

My third prediction is that unexpected earnings increases are more likely to be persistent while
unexpected earnings declines are more likely to be temporary. This asymmetric persistence prediction
can also be seen from the earnings time-series. Continuing with the fixed asset example in Panel B of Fig.
1, after good news, the firm's earnings increases currently, because of the reduced depreciation, and
remains at this higher level in future periods. Good news has a persistent effect on earnings. In contrast,
after bad news, the firm's current earnings is lower due to the impairment charge, but this decrease reverses
in the next period, and future years' earnings are unaffected. Hence, bad news has a temporary impact on
the earnings time-series.

My final prediction is that the abnormal return per dollar of unexpected earnings (the short window
earnings response coefficient or ERC) is smaller for 'bad earnings news' than 'good earnings news'.
In the fixed asset example, if the firm does not disclose that it has revised its estimate of the asset's life
until it announces earnings, the new depreciation numbers will surprise the market. The stock market's
reaction to the earnings release reflects the 'earnings news' about both current and future earnings. Hence,
the market reaction per unit of unexpected earnings is greater for more persistent news. In the example,
'bad earnings news' has no impact on future earnings, but 'good earnings news' is more persistent.

Empirical tests support all four predictions.

The conservatism principle

Alternative definitions of conservatism and the d(ff'erences between them

Accountants traditionally expressed conservatism by the rule "anticipate no profits but anticipate all
losses" (e.g. Bliss, 1924). I interpret this rule as denoting accountants' tendency to require a higher degree
of verification to recognize good news as gains than to recognize bad news as losses.

In contrast, some interpret conservatism more broadly as accountants' prefer- ence for accounting methods
that lead to lower reported values for shareholders' equity.

Conservatism no longer requires deferring recognition of income beyond the time that adequate evidence
of its existence becomes available or justifies recognizing losses before there is adequate evidence that
they have been incurred". This view of conserva- tism also appears inconsistent with accounting practice.

Financial accounting since the mid-1930s has emphasized the income state- ment, with a corresponding
emphasis on conservatism in the income statement. CAP (1939), (ARB 2) states, "conservatism in the
balance sheet is of dubious value if attained at the expense of conservatism in the income statement, which
is far more significant". Hence, I conduct tests on earnings, rather than on balance sheet values.4

Historical developments, and theories of the role of conservatism

Several costly contracting explanations have been advanced for the existence and pervasive influence of
conservatism (see Watts and Zimmerman, 1986; Ball, 1989; Basu, 1995; for expanded versions of the
following arguments). In a world of uncertainty regarding future profits, managers often possess valuable
private knowledge about firm operations and asset values. If managerial compensation is linked to reported
earnings, then managers have incentives to withhold from reported earnings any information that would
adversely affect their compensa- tion. Rational claimholders would reduce managerial compensation by
the expected effect of such malfeasance. The emergence of the conservatism principle and the preparation
of audited financial statements can be ascribed to manage- rial attempts to bond against exploiting their
asymmetrically informed position relative to other claimholders. Debtholders and other creditors also
demand timely information about "bad news' because the option value of their claims (Smith, 1979) is
more sensitive to a decline than an increase in firm value.5

Conservatism is thus argued to play an ex ante efficient role in contracting between the parties constituting
the firm. Phrased differently, if accounting were not regulated, contracting parties would voluntarily agree
that the accounting numbers used to partition cash flows amongst them should be determined
conservatively. Consistent with this argument, Leftwich (1983) reports that all departures from GAAP
specified in private debt covenants are conservative.

While contracting considerations appear to explain the origins of conserva- tism, tax, litigation, political
process and regulatory forces have also influenced the degree of conservatism in GAAP, particularly
during this century (Watts and Zimmerman, 1986: Basu, 1995). In the last ten years, the Financial Ac-
counting Standards Board (FASB) has mandated the recognition of formerly off-balance sheet liabilities
such as pensions, post-retirement health benefit obligations and environmental liabilities, along with their
associated expenses.

The FASB also issued standards for asset impairment recognition. These stan- dards have arguably
increased U.S. accounting conservatism in recent years. Both the costly contracting and regulatory
rationales can explain the continuing importance of conservatism in GAAP. Later in the paper, l explore
whether recent increases in accounting conservatism were caused by increases in auditor legal liability
exposure.

Data

The samples used for the tests consist of all firm-year observations from 1963 to 1990 with returns data on
the CRSP NYSE/AMEX Monthly files, and with the necessary accounting data on the COMPUSTAT
Annual Industrial and Research files. The use of the Research file reduces survivor bias. The require- ment
of data from both COMPUSTAT and CRSP causes a selection bias towards large firms, limiting the
generalizability of the results.

The asymmetric effect of conservatism on the properties of earnings

Conservatism and the sensitivity of earnings to returns

Because stock prices reflect information received from sources other than current earnings, stock prices
lead accounting earnings, by up to four years. Since accountants anticipate future losses but not future
profits, conservatism results in earnings being more timely and more sensitive concurrently to pub- licly
available 'bad news' than 'good news'. In the fixed asset example in Fig. 1, a longer estimated life (good
news) results in a small concurrent depreciation reduction, while a shorter estimated life (bad news) results
in a relatively large concurrent write down. Earnings is predicted to be more strongly associated with
concurrent negative unexpected returns, proxying for "bad news', than positive unexpected returns, which
proxy for 'good news'.

I regress annual earnings on current annual returns. Earnings, the dependent variable, contains more timely
information for 'bad news' firms, resulting in a higher predicted R2 for this sample. The slope coefficient,
[;¢,is predicted to be greater for the 'bad news' sample because earnings is predicted to be more sensitive to
contemporaneous unexpected returns. 7

Hypothesis l: The slope coefficient and R2 from a regression of annual earnings on annual unexpected
returns are higher for negative unexpected returns than for positive unexpected returns.
Fig. 2 displays the predicted relation between contemporaneous earnings and unexpected returns. The
slope coefficient for negative returns (in the second and third quadrants) is predicted to be higher than the
slope coefficient for positive returns (in the first quadrant). This is because unrealized losses (bad news)
are more likely to be recognized immediately under conservative accounting than unrealized gains (good
news). Some current unrealized gains will be recognized in future periods when they are realized. Because
news, by definition, is uncor- related through time, these current unrealized gains will be uncorrelated with
the recognition period news and returns. I predict that these realized gains reflecting previous good news
will result in a positive intercept in the "reverse' regression. In contrast, since most unrealized losses are
captured immediately, little bad news is postponed to future periods.
Panel A of Table 1 presents pooled cross-sectional regression results for price deflated earnings on inter-
announcement period returns. For the full sample, the adjusted R2 is 7.99%, which is consistent with prior
studies (Lev, 1989). The slope coefficient on returns is 0.113. The second regression in Panel A divides
Panel A of Table 1 presents pooled cross-sectional regression results for price deflated earnings on inter-
announcement period returns. For the full sample, the adjusted R2 is 7.99%, which is consistent with prior
studies (Lev, 1989). The slope coefficient on returns is 0.113. The second regression in Panel A divides
firm year observations into 'good news' and 'bad news' samples based on whether the return was greater
than or less than zero. Dummy variables capture the intercept and slope effects for the negative return
sample. The interactive slope coefficient,/~1, which measures the difference in sensitivity of earnings to
negative and positive returns is significant, and implies that earnings is about four and a half times
(4.66 = [-0.216 + 0.059]/0.059) as sensitive to negative returns as it is to positive returns. Adjusted
R^2sfrom separate regressions on the two samples indicate that the explanatory power of negative returns
(6.64%; 17,790 firm years) is greater than positive returns (2.09%; 25,531 firm years). These results are
consistent with earnings being more timely or concurrently sensitive in reporting publicly available 'bad
news' than 'good news'. The intercepts for both samples are positive and significant, consistent with the
current recognition of unrealized gains from previous periods that are uncor- related with current news.

Panel B of Table 1 presents results using market-adjusted variables to control for time-series non-
stationarity in the earnings and return processes that could affect the pooled cross-sectional standard errors.
Returns are adjusted by the CRSP equal-weighted index, and earnings-price ratios are adjusted by an
equal-weighted index of the sample EP ratios for that calendar year. The adjusted R2 for both regressions
are higher than in Panel A, which uses an identical sample, indicating a better fit. The slope coefficient on
negative market-adjusted returns is about six and a half times (6.45 = [0.256 + 0.047]/0.047) that on
positive market-adjusted returns. The explanatory power of negative market-adjusted returns, 10.00%,
exceeds that of positive market- adjusted returns, 1.07%. Both intercepts are significantly positive.

Panel C of Table 1 presents results using fiscal year returns, which exclude the market reaction to the
current year's earnings announcement. I use this speci- fication to isolate the impact of publicly available
'news' received from other sources before earnings is announced. Buy-and-hold annual returns are cumu-
lated to end at the fiscal year-end, in contrast to Panel A, which cumulated annual returns to end three
months later. The results are quite similar to those in Panels A and B.

Overall, Table 1 is consistent with my predictions under conservatism. Earn- ings is more timely in
reporting publicly available 'bad news' than 'good news', as measured by the difference in either the
adjusted R2s or slope coefficients. The intercepts are reliably positive, as expected if the recognition of
unrealized gains, but not unrealized losses, is postponed to future periods.

Comparisons of earnings and cash.flow to isolate the impact of conservatism

An analysis of conservatism provides insights into the nature of accounting accruals. Receipts and
disbursements of cash are typically considered objective evidence, absent fraud, that the underlying
transactions have been consum- mated, and cash transfers are usually recorded as they occur. If there is
substan- tial evidence of contractual performance before cash has been exchanged, accruals are recorded,
but some discretion exists in estimating the appropriate amounts. This results in earnings being more
timely than cash flow measures (Dechow, 1994).

The effects of conservatism on earnings and cash flow can be analyzed similarly. Accruals enable
accountants to recognize bad news about future cash flow on an asymmetrically timely basis. Unrealized
losses reduce current

earnings but do not impact current cash flow, while unrealized gains affect neither current earnings nor
current cash flow. Since earnings is the sum of cash flow and accruals, if unrealized losses but not
unrealized gains are recognized, then earnings is more conservative than cash flow. I predict that the
difference in the sensitivities of earnings and cash flow to current publicly available 'bad news' is greater
than the difference in their sensitivities to current publicly available 'good news'.

Note that earnings recognizes some 'good news' and "bad news' about assets equally quickly. For example,
both increases and decreases in gross accounts receivable are reflected quickly in earnings. Such accruals
should dampen the predicted effect, but not negate it.

Hypothesis 2: The increase in the timeliness of earnings over cash flow is greater for negative unexpected
returns than positive unexpected returns.
Hypothesis 2 is tested on a sample of 34,266 firm-year observations after outlier deletion from 1963 to
1990 for earnings before extraordinary items and discontinued operations (XE), cash flow from operations
(CFO), and cash flow from operating and investing activities (CFOI). The first two measures replicate
Rayburn (1986), and the third is calculated by subtracting the gross change in long-term assets (change in
non-current assets plus depreciation and amortiza- tion expense, a crude measure of investing cash flows)
from CFO. 9 To control for the possibility that extraordinary items and discontinued operations were the
sole cause for the results in Table 1, they are excluded from earnings in Table 2.

Panel A of Table 2 presents the results from regressions of XE, CFO and CFOI on returns and dummies
for negative returns to compare the asymmetry in their timeliness to news. The first three regressions show
that CFOI, CFO, and XE have successively stronger associations with returns as measured by the a d j
u s t e d RZs (0.73%, 4.55% and 9.31% respectively), consistent with Dechow (1994). This suggests that
the accruals that are successively added to CFOI to construct earnings either increase the timeliness and/or
reduce the noise in accounting numbers (Collins et al., 1994). The last three regressions show that while all
three measures have similar slopes on positive returns, the difference in the slope dummies and the
adjusted R2s from the regressions increase monotonically. This result is consistent with accruals
making earnings more timely in reporting 'bad news' but not 'good news'. The relative size of the
coefficients on negative and positive returns, (ill + fio)/flo, increases as we move from CFOI to CFO to XE
(1.65, 2.60 and 3.64). Panel B reports adjusted R2s from separate regressions on the two samples. Results
for CFOI show that the adjusted R 2 is similar for 'good news' (0.31%) and 'bad news' (0.24%).
However, for CFO and XE, the adjusted RZs are higher for 'bad news' (2.58% and 5.58%) than
'good news' (1.17% and 2.64%), and the increases in explanatory power as we move from CFOI to
XE are greater for 'bad news' than 'good news'. These results are consistent with conservatism being
reflected in accruals and not in cash flow.
Panel C reports the results of multivariate F-tests on equality of coefficients across regressions, which are
appropriate when different dependent variables are regressed on common independent variables (Mardia et
al., 1979).l° The tests on the slope coefficient for the difference in sensitivity to negative and positive
returns,/~1, indicate that the coefficient in the regression for CFOI is significantly smaller than that for
CFO, which in turn is significantly smaller than that for XE. The coefficient on positive returns is
significantly smaller for XE than CFO, suggesting that accrual adjustments make earnings less
sensitive than CFO to 'good news'. These results suggest that the greater timeliness of accruals for
'bad news' than 'good news' makes a large contribution to the increased timeliness of earnings
relative to cash flow.

Comparing Panel A of Table 1 with the XE regression in Panel A of Table 2, we see that excluding
extraordinary items reduces the conservatism in earnings, but does not eliminate it. A replication after also
excluding (pre-tax) 'special items' from earnings before extraordinary items resulted in a smaller but
statistically significant difference in the slope coefficients (0.139), but almost identical R2s.

The increased timeliness of earnings over cash flows for 'bad news' but not 'good news' is consistent with
accounting conservatism being reflected in accruals.

Conservatism and the pers&tence of accounting measures conditional on news

Conservatism results in the lower persistence of earnings in bad news periods relative to good news
periods. The underlying argument is that timeliness and persistence are different ways of viewing the same
phenomenon. More timeliness means that more current value relevant news is recognized
contemporaneously in earnings, leaving less current value relevant news to be recognized in future
earnings. More persistence means that less current value relevant news is reported in current
earnings, and more of it will be reported in future earnings.

Since earnings is predicted to be more timely for bad news, it is also predicted to be less persistent
for bad news.

The prediction on asymmetric persistence can also be derived by examining the effect of conservatism on
firms' earnings time-series. Conservatism implies that earnings anticipates future losses by concurrently
reporting an estimate of the expected future cash flow consequences of current bad news. Future periods'
reported earnings are protected from current bad news, so that next period's earnings will be close to what
it would have been if no bad news were received currently. From a time-series viewpoint, the bad news
reflected in current earnings will appear as a transitory shock or a one-time dip in the earnings process. In
contrast, the effects of a current positive shock will be spread over several future periods' earnings as
anticipated gains are realized. Thus, good news events are likely to appear as persistent shocks to the
earnings stream.

Hypothesis 3: Negative earnings changes have a greater tendency to reverse in the following period than
positive earnings changes.
Panel A of Table 3 reports that in cross-section, the price-deflated change in earnings tends to reverse in
the next period, with a significant slope coefficient of - 0.273.

When intercept and slope dummies for negative prior earnings changes are included in Panel B, a
significant and large difference in the reversion tendency for the two samples in the predicted direction is
observed, and the adjusted R^2 doubles. The slope coefficient on positive prior earnings changes, -
0.040, is not statistically significant, consistent with good news in earnings being permanent. The
slope on negative prior earnings changes, (- 0.692 = [- - 0.040 + - 0.652]), is significantly different
from zero, but not significantly different from - 1 , the theoretical limiting coefficient if negative
earnings changes contain only bad news that reverse completely in the next period. Finally, the adjusted
R2 from a regression on negative prior earnings changes (24.19%, 15,413 observations) is much
larger than that on positive prior earn-ings changes (0.68%, 20,981 observations). These results are
consistent with bad news earnings changes having a greater tendency to reverse in the next period
than good news earnings changes. 14 Earnings level-based (Panel C) and return- based (Panel D)
partitions produce weaker but qualitatively similar results.

To confirm that conservatism is reflected in accrual adjustments, the tests were replicated for XE, CFO,
and CFOI, respectively on a sample of 28,376 firm-year observations from 1964 to 1990 (reported in Basu,
1995). Excluding extraordinary items, which are unusual and infrequent by definition and thus unlikely to
recur, should make earnings changes reverse less. As expected, the slope coefficients on changes in XE
were all more positive than their counterparts in Table 3. Consistent with conservatism being reflected
in accruals, bad news cash flow changes reverse less than good news cash flow changes, although this
difference is not statistically significant.

Conservatism and the information content of earnings releases

The final hypothesis examines how conservatism affects the capital markets' reaction to 'earnings news',
the information that the markets learn when earnings is announced. The information content of earnings
releases is usually measured by the short window earnings response coefficient (ERC), which is the
abnormal return per unit of unexpected earnings at the earnings announcement.

Thus, firms with positive earnings changes are predicted to have higher ERCs than firms with
negative changes.

A short window abnormal return around the earnings announcement is used to isolate the market reaction
to news conveyed by the earnings announcement. Since the abnormal return captures the market's response
to the 'earnings news', I specify the abnormal return as the dependent variable and the change in earnings
as the independent variable for the regression test.

Hypothesis 4: In a regression of announcement period abnormal returns on earnings changes, the slope on
positive earnings changes is higher than on negative earnings changes.

The first three regressions in Table 4 show that the announcement period abnormal returns are weakly
associated with earnings changes. Intercept and slope dummy variables for the positive earnings change
sample are included in the last three regressions. The slope coefficient of announcement period
abnormal returns on positive earnings changes is significantly greater than the slope coefficient on
negative earnings changes. This result is consistent with the market recognizing that 'good earnings
news' is more persistent than 'bad earnings news' as a result of conservatism. The slope on 'bad
earnings news',/~o, is negative, but small and statistically insignificant, ranging from -0.000 to 0.014.15
Separate regressions, not reported, on the two samples for the three return intervals reveal that the adjusted
R^2s are higher for 'good earnings news' (0.93%, 0.57%, and 0.34%; 18,567 observations) than 'bad
earnings news' (0.00%, 0.02%, and -0.01%; 10,336 observations). Overall, the results are consistent with
the 'good earnings news' contributing most of the explanatory power in the full sample.

A potential alternative explanation for the lower ERCs and RZs for "bad earnings news' in Table 4 is that
earnings changes are worse proxies for 'bad earnings news' than 'good earnings news'.

Additional tests

I conduct some additional tests to increase confidence in my inferences. First, I test an implication of my
argument that earnings is more timely in reflecting 'bad news' than 'good news'. Second, I examine whether
changes in auditor legal liability exposure explain recent changes in conservatism.

Length of the measurement interval

Under "clean surplus' accounting, over the life of a firm, all the news in returns is also reported in
accounting earnings. Easton et al. (1992) argue that lengthen- ing the aggregation interval reduces the
relative lack of timeliness in earnings, thus improving the measured association with returns. Lev (1989),
Easton et al. (1992), Kothari and Sloan (1992), Dechow (1994) and others show that the association
between earnings and returns is stronger as the aggregation interval is lengthened. This suggests that
the longer the aggregation period, the higher the explanatory power of publicly available 'good news' for
earnings relative to 'bad news', and the less the difference in their slopes and RZs.

Table 5 presents results for a sample of 20,748 market-adjusted observations with variables cumulated
over four year intervals. The adjusted R^2s for both regressions are higher than in Panel B of Table 1, and
the slope coefficient for the full sample is about one and a half times that in Panel B of Table 1, consistent
with prior research. The difference between the slope coefficients for the 'good news' and 'bad news'
samples, 0.215, is less than in Panel B of Table 1, which was 0.256. Moreover, the difference in the
relative size of the slope coefficients drops from 6.45 to 2.57. Results from separate regressions on
the two samples show that the adjusted Rz for the 'bad news' sample (17.47%, 12,372 observations)
remains greater than that for the 'good news" sample (10.88%, 8,375 observations), but that the
relative and absolute difference have decreased. The results in Table 5 are consistent with a reduction in
the effect of accounting conservatism on earnings as the aggregation interval increases.

Variation in auditor legal liability exposure and conservatism over time

The legal liability exposure of auditors and managers for tardy disclosure of 'bad news' has increased
significantly over the last three decades (Kothari et al., 1989; Skinner, 1994). Conservatism reduces
auditors' liability exposure, and auditors are thus expected to have increased the asymmetric
timeliness of earnings in response to exogenous increases in their legal liability exposure.
Alternatively, it is possible that the courts enforce increased conservatism because contracting parties have

increased their demand for conservatism

Table 6 replicates the regression in Table 1 Panel B with interactive dummies for the different subperiods.
Market-adjusted variables are used to control for time-series non-stationarity. During the initial low auditor
liability regime of 1963 66, the coefficient on market-adjusted returns is small, and the difference in the
slope coefficients for "good' and 'bad' news, 0.009, is insignifi- cant. During the increased auditor liability
regime of 1967 75, the slope coefficient on 'good news' doubles to 0.072 ( = 0.034 + 0.038), while the
coeffic- ient on "bad news' increases six-fold from 0.043 (= 0.034 + 0.09) to 0.259 ( = 0.034 + 0.009 +
0.038 + 0.178). The increase in the difference in sensitivity to 'good" and 'bad' news, 0.178, is
statistically significant. The low auditor liability regime in 1976-82 coincides with a drop in the
sensitivity of earnings to both 'good news', 0.033, and "bad news', 0.222. The final high auditor liability
regime of 1983-90 sees no increase in sensitivity to 'good news', but a large and significant increase in the
sensitivity to 'bad news', of 0.214. The slope coefficient of 0.437 on 'bad news' for 1983-90 is nearly
thirteen times that on 'good news', 0.034. The high asymmetric timeliness of earnings after 1983 is
consistent with a rational response by auditors to being exposed greater legal liability after 1983.

Annual cross-sectional regressions for each year (reported in Basu, 1995) corroborate the results in Table
6. The annual slope coefficients on 'good' and 'bad' news are plotted in Fig. 3. From 1963 to 1966, a period
of low auditor liability, the coefficient on 'bad news' does not differ much from that on 'good news'. Over
the next 24 years, the coefficient on 'bad news' is greater than that on 'good news' in every year, and
significantly greater at the 1% level in 20 out of 24. The difference in the coefficients grows during the
high auditor liability period 1966 75, and then drops down to a lower level during the low auditor liability
period 1976 82. From 1983 onwards, a high auditor liability period, the coefficient on 'bad news' is always
at least five times as large as that on "good news', and fifteen times larger in 1989. Over these last eight
years, the coefficient on 'good news' is significant in only four years, whereas that on the dummy for 'bad
news' is always significant, and the absolute difference in the coefficients ranges from 0.33 to 0.51. Fig. 3
indicates that the results in Table 6 did not reflect unusual circumstances or outliers in a few years.

Alternative explanations

Shareholders' liquidation or abandonment option

Hayn (1995) argues that shareholders would prefer to liquidate a firm rather than bear predictable
losses, i.e., they have a put option on the firm. Hence, observed losses are those that were expected to be
temporary. This implies that reported losses should be less persistent than reported profits, and Hayn
(1995) predicts and shows that in a regression of annual returns on annual earnings, the slope coefficient
and R: are higher for profit firms (firms currently reporting profits) than loss firms (firms currently
reporting losses). This prediction is displayed in Fig. 4. A comparison of Fig. 4 with Fig. 2 shows that they
are reflections of each other along the 45~'line, because their axes are reversed. In other words, the slope
coefficient prediction from the abandonment option is the same as that under conservatism. However,
conservatism predicts a higher R2 for bad news or negative return firms, while the abandonment option
theory predicts a higher R 2 for good news or profit firms.
I replicated the test by Hayn (1995) using the sample in Table 1 (Panel A) and found that the earnings-
return R 2 was greater for profit firms than for the full sample. These results are consistent with hers.
However, when I used the market-adjusted sample in Panel B of Table 1, the R2 was greater for the full
sample than for profit firms only, contrary to the prediction by Hayn (1995). Chambers (1996) reports that
when sample firms are size-matched, loss firms have a higher R 2 than profit firms. Thus, Hayn's annual R
2 results (Hayn, 1995) appear sensitive to risk-adjustment, while the conservatism prediction is not
rejected under many different specifications. Also, conservatism predicts the positive regression intercepts
in Table 1, while the abandonment option theory has no prediction on the intercepts.

The abandonment option theory does not predict this time- series variation in the earnings-return
relation. In sum, the abandonment option theory does not explain all the conservatism results, although
some predictions and results are similar.

Delayed disclosure of bad news by opportunistic managers

McNichols (1988) argues that opportunistic managers have incentives to leak upcoming good news before
earnings is disclosed, and earnings announcements will thus reveal relatively more 'bad earnings news' to
the market. This argu- ment is consistent with viewing conservatism as a mechanism for managers to bond
against exploiting their asymmetrically informed position relative to other stakeholders, so that the
earnings report via auditing and conservatism often conveys the first tidings of bad news.

Although this theory implies that firms with 'bad earnings news' will have bigger market reactions
on average at the earnings announcement, it does not make a prediction about per unit reactions
(ERCs).

Managers'incentives to disclose bad news early

Skinner (1994) argues that managers release bad news early before earnings are announced to reduce their
investor litigation exposure. This parallels my argument that U.S. accounting has become more
conservative as a response to increased auditor liability exposure. Skinner (1994) finds that large negative
earnings surprises are preempted by voluntary corporate disclosures more often than other earnings
releases. Such early disclosures of bad news could lead to a lower ERC for negative earnings changes
when earnings are announced because the bad news was preempted.

Managers may also disclose bad news early to deter entry or competition in their firms' product markets
(Darrough and Stoughton, 1990; Wagenhofer, 1990) or to signal their 'quality' (Teoh and Hwang, 1991).
These incentives potentially reinforce those arising from their legal liability exposure.

Effect of mandated accounting changes

The large one-time transition effects that firms recognized when they adopted new accounting standards
for liabilities such as pensions and healthcare benefits could potentially confound my results on
conservatism. However, Balsam et al. (1995) show that firms adopting new FASB standards tend to report
income- increasing effects for prior years in the income statement (after extraordinary items); and income-
decreasing effects directly in retained earnings, thus shielding net income. This evidence suggests that new
mandated accounting standards do not account for the greater sensitivity of earnings to bad news than good
news.

Lecture 5

Slippage in step 1 of Nichols and Wahlen – conservatism. We want to understand which principles lead to
fin errors.
Why do we need fin information => because we need a performance measure that periodically tells us how
we are doing, earnings is one of the best measures

Recap
• “Earnings” (net income) is the key performance measure produced by the accounting system
• Nichols and Wahlen (2004): earnings are “value relevant”
• Why?
• Because earnings today relate to expected future earnings and dividends, and these future earnings and
dividends
affect the value of the company for an ordinary shareholder
• Dechow (1994): accrual accounting makes earnings more value relevant than cash flows
• Accruals fix the “timing” and “matching” problems in cash flows
• Other key concepts: earnings persistence, event study, earnings response coefficients (ERC)

Every asset is the net present value of future benefits. When we buy shares, the economic benefit is future
dividends
Remember

• Also recall that DCF models express value in terms of expected

future (free) cash flows:
• Accounting rules rely on similar present value techniques to determine the current “fair” value of an
asset based on expectations of future cash flows
• IASB’s definition of an asset: “A present economic resource controlled by the entity as a result of past
events [...] An economic resource is a right that has the potential to produce economic benefits”
• The “economic benefits” refer to the cash flows the asset is expected to generate in the future
• A change in the expected future cash flows changes the value of the asset
• We might have to recognize this change on the balance sheet if the change is material

As accountants we need to think about what if there is a change in our expectations.

The concept of conservatism is present in almost all standards. Especially in inventory, you do not look at
how much it costed you, you look at lower of cost and net realizable value. So, if the value goes down in
next year, we need to apply the rule of NRV, but if it goes up, in the book it stays the same. Also,
impairments. You need to apply the list of external and internal indicators of impairment. So, every year
for infinite intangibles you need to apply internal and external indicators of impairment.
So, if expectations change, this will be reflected in the share price.

Conservative accounting
• Conservative accounting practices refer to how accountants deal with such changes in asset values that
result from news about expected future cash flows
• Questions to be answered in this class:
• What is accounting conservatism?
• Why is accounting conservative?
• What types of conservative accounting can we distinguish?
• What are examples of conservative accounting?
• How can we empirically measure conservatism in accounting?
• First: let’s see why we should care about conservative accounting practices

Example #1
“By increasing goodwill and reducing tangible assets, GE avoided costs that could have
weighed on its earnings each quarter for years.”

“In April 2017, GE’s industrial businesses posted sharply negative quarterly cash flow, $1 billion below
internal expectations, raising concerns about its accounting and the health of the power business.”

They increased their balance sheet by 22mil, but then they took it out. When A buys b for 100B, but book
value was 50B, the difference is the goodwill. However, we assign 80 to PPE and 20 to goodwill, then we
pay amortization expense of 8. If we are conservative, then we assign 60 to PPA and pay amortization of
6 and goodwill of 40, and then next year we amortize it to conservative amount.
Example #2
“A U.S. investigation into the lack of write-downs at Exxon Mobil Corp. despite
the slump in oil-prices has brought to light the challenge of assessing impairments under U.S. Generally
Accepted Accounting Principles (GAAP).”
“Substantial distinctions between IFRS and GAAP” :
• Impairment trigger: discounted (IFRS) vs. undiscounted (GAAP) future cash flows
• Impairment reversals under IFRS (not for goodwill)
It is not only hard to choose when to impair, but also there is a difference between IFRS and GAAP,
because IFRS uses discounted cash flows, but GAAP uses undiscounted.
Example #3
• “Kraft Heinz said it booked a charge of $1.22 billion for the first six months of its fiscal year. The bulk of
the charge was related to the declining values of
brands and businesses, but $474 million of it reflected the company’s lower stock price.”

Stock markets warn you of what will happen in the books. One of the external triggers

Large increase in interest rates => the PV goes down

From the same WSJ article (Trentmann, Aug. 8, 2019)

• “H.J. Heinz Co. merged with Kraft Foods Group Inc. in 2015 in a deal worth about $49 billion. The
combined company in February wrote down more than $15 billion for some of its key Kraft and Oscar
Mayer brands amid waning consumer demand.”
• “A sustained decrease in a company’s stock price both in absolute terms and relative to its peers also
qualifies for impairment charges.”
• “Tests for impairments must conducted annually. In the interim, certain events—such as the discovery
of new value-depreciating issues with acquired assets—can trigger additional writedowns.”
• “Regulators and auditors are expected to pay more attention to write-downs related to stock price
declines should the U.S. economy continue to slow”
• “If the market does not share the company’s view [...] the company has to take a charge.”

Insights and questions from the examples

• Poor performance of (divisions of) companies triggers the recognition of losses
• Does good performance lead to the recognition of gains?
• Typically not!
Loss recognition is driven by accounting rules, but also partly by companies’ discretion in reporting
• Conservative accounting rules aim to prevent the overstatements of assets
• But loss recognition may also signal a correction of previous overstatements of assets
• Do companies recognize these losses in a sufficiently timely manner?
How can we identify such conservative accounting practices in research?
• Look at how timely the accounting is in recognizing losses triggered by indicators such as:
• Changes in consumer demand
• Changes in cash flows of company divisions
• Changes in stock price
Impairment is an accrual.
We want to provide a critical analysis of whether conservatism exists.
Can we generalize is and is it important.

Basu(1997)
If you look at accounting for a long time, the accounting is a principles of anticipating losses, but not
profits. He wants to reexamine conservatism, because accountants want to see that inventory will be sold,
but most importantly book the loss immediately.
Conservative accounting
• Bliss (1924, as quoted by Basu 1997 on p.7) defines conservatism in accounting as:
“anticipate no profits but anticipate all losses”
• Examples of conservative accounting:
• Inventory valuation: lower of cost or net realizable value (“market”)
• Impairments of tangible assets
• Impairments of intangible assets other than goodwill
• Impairments of goodwill
• Immediate expensing of internally generated intangible assets (e.g., R&D)

• Purpose of Basu’s (1997) paper: “re-examine the conservatism principle” (p. 4)

• Conservatism: “accountants’ tendency to require a higher degree of verification for recognizing good
news as gains than to recognize bad news as losses” in financial statements” (p. 7)
• This implies that earnings (net income) should recognize bad news more quickly than good news
• Example: “unrealized” losses often recognized earlier than unrealized gains
 “Asymmetric” recognition of news in financial reporting
Earnings give us asymmetric news. News is a change in expectations of future benefits. Indications of
news are external and internal sources from impairments
• What is news?
• Recall that with “news”, we refer to a change in the expectations of future cash flows to be generated by
a company, business unit, or asset

An example of a fixed asset – machine. Book value at t is 70000, we can use it for 10 years, so,
depreciation charge is 7000. Now, we get news in year 3 – bad ones ( we cannot sell all the products, so,
depreciation is reduced by 3 years – we have to impair, but deprecation is the same) or good ones (we can
use machine for 3 years longer – the depreciation charge drops).
Anticipate all losses, but not gains
 Some accruals are reversing quicklier than others, like these bad news.
When we purchase an asset we need to be prudent in assessing how much we can earn back.
Explanations for conservative accounting
• Some accounting standards rely on the principle of “prudence” (in Dutch: “voorzichtigheidsbeginsel”)
• In its revised Conceptual Framework, the IASB (2018) notes about prudence (article 2.16):
“Neutrality is supported by the exercise of prudence. Prudence is the exercise of caution when making
judgements under conditions of uncertainty. The exercise of prudence means that assets and income are
not overstated and liabilities and expenses are not understated. Equally, the exercise of prudence does not
allow for the understatement of assets or income or the overstatement of liabilities or expenses. Such
misstatements can lead to the overstatement or understatement of income or expenses in future periods.”
• For example: high uncertainty about the possible future realization of cash flows from a particular
asset (e.g., investment in research and development (R&D)) might preclude the accountant from
recognizing the asset in accordance with prudence

Why we do not show all R&D as asset, because we do not know if it is profitable.
Explanations for conservative accounting
• Interestingly, however, the IASB (2018) also notes in its Conceptual Framework in article 2.17:
“The exercise of prudence does not imply a need for asymmetry, for example, a systematic need for more
persuasive evidence to support the recognition of assets or income than the recognition of liabilities or
expenses. Such asymmetry is not a qualitative characteristic of useful financial information. Nevertheless,
particular Standards may contain asymmetric requirements if this is a consequence of decisions intended to
select the most relevant information that faithfully represents what it purports to represent.”
We look for the most relevant information. Some assets you need to book with conservatism, but for some
you use FV.
• Basu (1997, p. 10): “The FASB also issues standards for asset impairment recognition. These standards
have arguably increased U.S. accounting conservatism”

• In addition to the role accounting standards in explaining conservatism, academics have proposed that
contracting (executive compensation, where we need a performance measurement, so we can use accruals
for that. However, manager has a lot of private information, and we know that he/she has a discretion in
the change in prices. So he would not want to show bad news, but only good news. So conservatism is a
way of stopping managers to exploit their private information) considerations can also explain the
existence of conservatism
• Basu (1997, p. 9):
• “In a world of uncertainty [...], managers often possess valuable private information”
• “If managerial compensation is linked to reported earnings, then managers have incentives to” not
recognize bad news in earnings
• Conservatism “can be ascribed to managerial attempts to bond against exploiting [...] other claimholders”
• “Debtholders also demand timely information about ‘bad news’ because the option value of their claims
[...] is more sensitive to a decline than an increase in firm value”
Conservatism versus fair value accounting
• Although they are not perfect opposites, the intuition behind the difference between conservative
accounting and fair value accounting is:
• Conservative accounting: when news about future cash flows arrives, accountants typically only
recognize unrealized losses, not unrealized gains (asymmetric recognition of news)
• Fair value accounting: when news about future cash flows arrives, accountants recognize both unrealized
gains and unrealized losses (symmetric recognition of news)
Sometimes it becomes a problem, when we always have conservative numbers, so we apply FV. However,
the changes we show in comprehensive income.

• Earnings management and manipulation can lead to both an overstatement and understatement of assets
and net income, so sometimes has similar effects as conservative accounting:
• Big bath accounting: reduce the value of assets to allow for future overstatements of earnings
• Cookie-jar accounting (earnings “smoothing”): increase liability reserves to allow for future
overstatements of earnings
• Important distinction:
• Earnings management results from management’s discretion over financial reporting
• Conservatism is, for a large part, trigger by accounting rules
The whole idea is conservatism stops earnings management.

Conditional versus unconditional conservatism

• Basu (1997) focuses on what we know as “conditional conservatism (showing bad news early, but not the
good news)”
• This refers to the higher degree of verification that is needed to recognize good news as gains than to
recognize bad news as losses; i.e., this is conservatism that is conditional on the direction of news
• Or: “news-dependent” conservatism
• Alternative: “unconditional conservatism”
• Understatement of net assets that is news-independent (R&D - expense immediately)
• Examples (Beaver and Ryan 2005):
• Accelerated depreciation on assets
• Immediate expensing of internally generated intangible assets (e.g., R&D)
• Historical cost accounting for projects with positive NPV

Main hypothesis

Because accountants are more likely to recognize future losses than they recognize future profits, a
consequence of such conservatism is that earnings should be found to be more sensitive to publicly
available bad news than to good news
Sensitive the difference in change in variables
 If stock returns capture “news”, earnings should be more sensitive to negative stock returns than they are
to positive stock returns
• H1: The slope coefficient and R2 from a regression of annual earnings on annual unexpected returns are
higher for negative unexpected returns than for positive unexpected returns (null hyp: the slope coefficient
and R2 are not different)

Research design
We explain the variation in earnings conditional on returns. Earnings summarizes information in a timely
manner, and this information is reflected in a share process. How well we summaries it is reflected in R^2

We have seen it in tweeter.

The steep shows do we have more news.

The R^2 shows how much variation there is in a model
Main result

Beta 0 is the variation of returns on earnings. If returns increase by 1 unit, earnings increase by 11%
The we split it up in two regressions.
Beta 1 is positive. R^2 for bad news is 3 times higher than for good news.

If you have news in returns, earning will pick up 11%. The remained will be picked in future years. After
the split, when we pick up good news we pick up 5.9%, with bad news it is 27.5%

Replication of Basu (1997)

Now we fit the same regression separately for good and bad news.
 Remember!

Alpha is an incremental effect –

inerecept
Importance of interaction regressions
• Understanding how these interaction regressions work is essential for (financial) accounting research
• Tests of “moderating” effects also rely on interaction analyses
• Experimental research often relies on them
• Archival studies that use “natural experiments” and “shocks” use them as well
• The remainder of Basu’s tests also rely on interaction regressions
• In week 6, we will discuss difference-in-difference analyses, which also rely on interaction regressions

Basu’s other hypotheses: H2

• H2: The increase in the timeliness of earnings over cash flow is greater for negative unexpected returns
than positive unexpected returns
• Dechow (1994): earnings are better than cash flows in measuring performance, because accrual
accounting shifts the recognition of cash flows over time
• Accruals enable the timely recognition of bad news in earnings: unrealized losses reduce earnings but not
cash flows, while unrealized gains do not affect earnings and cash flows
• Particularly the case for accruals that reflect writeoffs and impairments

Basu’s other hypotheses: H3

• Recall: persistence refers to the likelihood that the current earnings level will recur in the future
• H3: Negative earnings changes have a greater tendency to reverse in the following period than positive
earnings changes
• “Timeliness and persistence are different ways of viewing the same phenomenon” (p. 19)
• Greater timeliness means: more value relevant news reflected in earnings in the current period, and less
to be recognized in future periods
• “Transitory” shock to earnings: e.g., an asset impairment is unlikely to recur next year
• Greater persistence means: less value relevant news is reflected in earnings today, but more to be
recognized in (and spread out over) future periods

Basu’s other hypotheses: H4

• H4: In a regression of announcement period abnormal returns on earnings changes, the slope on positive
earnings changes is higher than on negative earnings changes
• Recall from last week the link between market reactions to earnings and the underlying persistence of
earnings (Nichols and Wahlen 2004):
• More persistent earnings should trigger stronger market reactions
• In H3, the prediction was that negative earnings changes are less persistent than positive changes
 This should trigger weaker reactions

• H4: In a regression of announcement period abnormal returns on earnings changes, the slope on positive
earnings changes is higher than on negative earnings changes
• Important:
• This analysis looks like the reverse of the earnings-returns regression from H1: now we look at the
relation between (announcement) returns and earnings (changes)
• But it is very different!
• H1: association study: relation between long-window earnings and stock returns
• H4: event study: relation between short-window stock returns and news in earnings
 This is what Ball and Brown (1968) and Nichols and Wahlen (2004) used:
Event = earnings announcement
Important research after publication of Basu (1997)
• Watts (2003): provides conceptual discussion on the (contracting) reasons for conservative accounting
practices, and discusses the body of empirical literature at that time
• Many studies use the “Basu-measure”, or related measures, of accounting conservatism to study:
• Effects of cross-country differences in institutional settings
• Effects of international accounting standards
• Differences in the demands for conservative accounting in public versus private companies
• Effects of conservative accounting on the costs of external debt and equity financing
• Relation between conservative accounting and the value relevance of earnings
• Effects of conservative accounting on managers’ decisions (e.g., investments)

Summary of learning objectives

• “Understand how to read an academic paper and distill key attributes such as the research question,
contribution, main hypothesis, key concepts, measurements of the key concepts, and the primary
conclusions and implications of the study.” (see course outline)
• Accounting conservatism
• Definition, origins, and examples
• Differences between conservatism and fair value accounting
• Differences between conditional and unconditional conservatism
• How to empirically measure conservative accounting practices for research purposes
• Understanding how to conduct regressions with interactions, and how to interpret their results
We tried to figure out the issues of conservative accounting. In countries with more contractive, we see
more conservatism. As demand for conservatism is higher, the difference in effect of bad and good news in
higher.

Tutorial 5 Banker, Basu, and Byzalov (2017)

Implications of Impairment Decisions and Assets’ Cash- Flow Horizons for Conservatism Research

Accountants examine multiple indicators when assessing whether individual assets are impaired. Different
indicators predict cash flows over varying time horizons, and their importance varies with how far into the
future individual assets are expected to generate cash flows. We predict that earnings exhibits asymmetric
timeliness with respect to multiple indicators, including stock return, sales change, and operating cash flow
change, which differentially explain write-downs of current assets, long-lived tangible assets, and
indefinite-lived goodwill. We predict an interaction effect between indicators, such that the total impact of
several consistent indicators is greater than the sum of their individual impacts.

We study how accountants use data about future cash flows in their impairment decisions. Prior research
shows that bad news is recognized in earnings more quickly than good news, and infers the presence of
conditional conservatism from this finding

Earnings likely responds asymmetrically to multiple indicators. Different asset classes, such as inventory
and long-lived tangible assets, are tested for impairment separately. Therefore, accountants will use
indicators that best predict future cash flows for these individual asset classes

We identify sales change as an important new indicator in conservatism research. Impairment tests for
long-lived assets are based on operating cash flow forecasts. Sales is the fundamental driver of cash
inflows and outflows. Therefore, the sales forecast is a major input in the operating cash flow forecast.
Because sales forecasts are typically formed by projecting recent sales trends, current sales change is likely
to be an important determinant of accountants’ cash flow forecasts.

HYPOTHESIS DEVELOPMENT

Conditional conservatism is defined as the higher degree of verification used to recognize good news as
gains than to recognize bad news as losses. Conditional conservatism implies asymmetric timeliness of
earnings with respect to good versus bad news about future cash flows. Basu (1997) estimates asymmetric
timeliness using a piecewise- linear regression of net income on stock return, where positive (negative)
unexpected return is a proxy for good (bad) news. He finds that the slope coefficient and R2 are higher for
negative return than for positive return, i.e., net income reflects bad news more quickly than good news, as
predicted.

Asset write-downs are the most fundamental manifestation of conservatism.4 However, conservatism
research primarily focuses on earnings (rather than impairments) because of its central contracting role and
because of the limited quality, availability, and scope of impairment data in Compustat. Therefore, we use
the asset impairment guidance as a theoretical foundation to develop insights into conservatism for
earnings.
Impairment tests are conducted for asset groups ‘‘at the lowest level for which identifiable cash flows are
largely independent of the cash flows of other assets and liabilities’’ (SFAS 144, para. 10). Therefore,
indicators that help predict future cash flows for individual asset classes will be relevant for impairment. In
a semi-strong efficient market, these indicators are embedded in the stock price, which incorporates all
public information about firm value. Hence, researchers typically use stock return (rather than individual
indicators) as the main proxy for unrealized gains and losses. However, stock return reflects the change in
total discounted cash flows over an infinite horizon, whereas many impairment tests focus on asset cash
flows over a much shorter horizon. These shorter-term cash flows are likely better predicted by short-
term indicators than by stock return.

SFAS 144 requires a two-step impairment test for long-lived assets. This test can be triggered by indicators
of potential impairment, including a significant adverse change in the business climate (for example,
declining sales) or deterioration in current or projected future cash flows. The first step tests whether the
book value of the asset (or asset group) is less than the sum of undiscounted future cash flows from the
asset, where the cash flow estimates ‘‘shall consider all available evidence’’. Because operating cash flow
projections are normally based on sales forecasts less cost forecasts, the relevant evidence includes various
predictors of future sales and costs.

If the asset is deemed to be impaired based on this comparison, then in the second step, the asset is written
down to its fair value, estimated as the sum of discounted future cash flows (in the absence of quoted
market prices) or based on market prices (for assets traded in active markets). Indicators of a reduction in
expected future cash flows, such as a decrease in current sales or operating cash flow, increase both the
probability (in step 1) and the magnitude (in step 2) of the asset impairment. Therefore, conditional on
stock return, which combines gains for some assets and losses for others, firms will recognize
unrealized losses reflected in additional unfavorable indicators, but will not recognize unrealized
gains reflected in additional favorable indicators.

In summary, the impairment standards for both long-lived assets and goodwill suggest that accountants
will quickly recognize unrealized losses reflected in unfavorable short-term indicators such as decreasing
sales or operating cash flow. Therefore, after controlling for stock return, earnings is likely to exhibit
asymmetric loss recognition with respect to multiple additional indicators, including (but not limited to)
sales change and operating cash flow change.

We propose sales change as a major new indicator motivated by the standard cash flow forecasting
procedures used by both internal and external analysts. Because revenue is the primary driver of operating
cash inflows and outflows, future sales are predicted first. The sales projection in the standard time-series
methods, such as exponential smoothing or autoregressive integrated moving average (ARIMA) modeling,
is a function of the most recent sales change. The sales projection is next used to forecast the income
statement and balance sheet items, which, in turn, yield the cash flow forecast

We predict that sales change has an asymmetric effect on earnings incremental to both stock return and
operating cash flow change.Operating cash flow incorporates noise due to variation in working capital
during the normal course of operations. Many expenditures are matched poorly with expected revenues.
For example, because R&D and advertising expenditures are not capitalized as investments, they reduce
current operating cash flow (but likely increase expected future cash flows). This mismatching adds more
noise to operating cash flow. Because managers cannot adjust many resources on short notice, they adapt
to economic shocks slowly. Thus, operating cash flow change reflects only partial resource adjustment in
response to current shocks and is a noisy indicator of the impact of these shocks on future cash flows.
Sales change is not affected by these sources of matching-related noise. Further, it likely captures the
fundamental drivers of future cash flows, such as the overall performance of the firm’s business model,
providing significant incremental information for impairment tests:

H1: After controlling for the asymmetric effects of both stock return and operating cash flow change,
earnings exhibits asymmetric association with sales change.
 Compustat separates write-downs of goodwill (and other unamortized intangibles with indefinite future
lives) from write- downs of tangible assets (and amortized intangibles with a definite future life). Because
conservatism manifests most directly in asset write-downs, which are not confounded by other potential
asymmetries in earnings, the write-down data let us verify that the asymmetric effect of a given indicator
on earnings is attributable to conservatism.16 These data also let us test our time- horizon argument for
why accountants use different combinations of indicators for different asset classes. Future cash flows in
impairment tests are estimated for the remaining useful life of the asset or asset group (e.g., SFAS 144,
para. 18). Because tangible assets have a finite useful life, long-term data have limited relevance.
Goodwill, by contrast, has an indefinite future life. Further, because the value of goodwill only
incorporates cash flows that are not attributed to (finite-lived) assets in place, the long-term data are much
more important. Therefore, short-term indicators, such as sales change or operating cash flow, will be
relatively more relevant in assessing impairment of finite-lived tangible assets, whereas the long-term
indicator, stock return, will be relatively more informative for goodwill impairment: 17

H3a: The relative impact of short-term indicators, such as sales change or operating cash flow change, is
greater for tangible asset write-downs than for goodwill impairments.

H3b: The relative impact of stock return is greater for goodwill impairments than for tangible asset write-
downs.

DATA AND EMPIRICAL MODELS

Empirical Models
EMPIRICAL RESULTS

The estimates of Models (1)–(3) are presented in Table 3. Consistent with Basu (1997), in all models, the
asymmetric timeliness coefficient on DR 3 RET is positive and significant, i.e., bad news (negative RET)
is recognized in concurrent earnings more fully than good news (positive RET), indicating conditional
conservatism for stock return. Similar to Ball and Shivakumar (2005, 2006 ), we find a significant positive
coefficient on DC 3 DCF, which indicates asymmetric timeliness with respect to concurrent operating cash
flow change. When we add sales change, the coefficient on DC 3 DCF is reduced by 41 percent, from
0.823 in the Ball and Shivakumar (2006) model (Column (2)) to 0.482 in our full three-indicator model
(Column (3)). This suggests that sales change is an important correlated omitted variable in the Ball and
Shivakumar (2006) model. However, this smaller asymmetry coefficient on DC 3 DCF remains
significant, consistent with our argument that unrealized gains and losses are best modeled by a vector of
indicators.

As predicted in H1, in Table 3, after controlling for both stock return and operating cash flow change, as in
Ball and Shivakumar (2006), sales change has a significant asymmetric effect on earnings (the coefficient
on DS3DSALES is 0.174, t 1⁄4 8.74 in Column (3)).20 The sales indicator plays an important incremental
role in conditional conservatism. For example, while a $1 sales increase (good news) raises earnings by
just 5.9 cents, on average, a $1 sales decrease (bad news) reduces earnings by 23.3 cents, on average (1⁄4
0.059 þ 0.174), indicating much quicker recognition of bad news than good news for sales.
In Table 4, we examine additional proxies for unrealized gains and losses. First, to ensure that the results
are not driven by the accrual component of sales, we redefine the sales indicator as the change in cash
sales, DCSALES. We consider three additional indicators from Compustat: the level of current operating
cash flow, CF; the change in cash gross margin, DCGM; and the change in order backlog, DOB. We also
examine indicators from I/B/E/S that are based on analysts’ forecast revisions for future earnings per share
(DEPS_F), future operating cash flow per share (DCPS_F), and future sales (DSALES_F), espectively.
Analysts’ forecasts likely embed some of the firms’ forward-looking information that is not fully captured
by our Compustat-based indicators (e.g., Hutton 2005; Cotter, Tuna, and Wysocki 2006). To isolate the
news that arrived during the current fiscal year, we define these indicators as the change in analysts’
consensus forecast for year tþ1 that occurred from the beginning to the end of year t.22 In all cases, the
additional indicators have a significant asymmetric effect in the Basu (1997) and Ball and Shivakumar
(2006) models. Further, even when they are added as a fourth indicator in our full model, all but two (DOB
and DCPS_F) exhibit significant asymmetric loss recognition
To test H3, we estimate Models (1)–(3) for write-downs of tangible assets and goodwill (Table 6).27 For
both tangible assets and goodwill, we find asymmetric timeliness with respect to all three indicators (i.e.,
the coefficients on DR 3 RET, DC 3 DCF, and DS 3 DSALES are all positive and significant). In
untabulated tests, we also find significant asymmetric loss recognition for the additional indicators from
Table 4. Because conservatism flows through write-downs while other asymmetries in earnings do not,
these estimates confirm that the documented effects of multiple indicators on earnings are attributable to
conservatism.

The asymmetric timeliness coefficients for goodwill impairment in Table 6, Columns (4)–(6) differ from
the parallel long-lived asset write-down coefficients in Columns (1)–(3), as predicted by H3a and H3b. The
coefficient on DR 3 RET increases significantly by 100–146 percent, whereas the coefficients on DC 3
DCF and DS 3 DSALES decrease by 17–33 percent. Thus, the relative impact of stock return, a long-term
indicator that incorporates expected cash flows over an infinite horizon, is greater for impairment of
(indefinite-lived) goodwill, whereas cash flow change and sales change, our short-term indicators, play a
greater role in write-downs of (finite-lived) tangible assets. This evidence is consistent with our time-
horizon explanation for accountants’ use of multiple indicators.
Tutorial 5 Glaum, Landsman, and Wyrwa (2018)

Goodwill Impairment: The Effects of Public Enforcement and Monitoring by Institutional Investors

In this study, we investigate the determinants of goodwill impairment decisions for firms applying International
Financial Reporting Standards (IFRS) using an international dataset from 21 countries. We find that firms’
goodwill impairment decisions are associated with both stock return, our proxy for the decline in the economic
value of goodwill, and with proxies for managerial and firm-level incentives.

Conceptually, the impairment test for goodwill can be interpreted as an example of managerial discretion in
financial reporting. If applied neutrally, discretion allows management to convey private information and, thus,
make financial statements more informative. However, management may also use the discretion
opportunistically by delaying (or accelerating) goodwill impairments, or by ‘‘managing’’ the amounts of
goodwill impairment losses. Ultimately, it is an empirical question whether goodwill impairment losses reflect,
in a timely manner, declines in the economic values of firms’ goodwill balances, or whether they are unduly
influenced by managerial or firm-level incentives.

INSTITUTIONAL BACKGROUND AND RELATED LITERATURE

Goodwill arises when the price paid for a target company by an acquirer exceeds the fair value of the target’s
net assets that the acquirer recognizes in its consolidated financial statements.

conceptual problem, identified by the IASB itself, is that acquired goodwill that is allocated to a CGU is mixed
with unrecognized internally generated goodwill. Effectively, the internally generated goodwill provides a
‘‘cushion’’ against future goodwill impairment losses.5 Moreover, companies may be able to lower the
likelihood of future goodwill impairment losses by purposefully allocating goodwill to CGUs that have high
levels of unrecognized internally generated goodwill. The goodwill impairment tests defined in IAS 36 also give
companies wide scope for discretion. The estimation of fair values or values-in-use for the CGUs generally are
based on company-specific forward-looking information: business plans with expected future cash inflows and
outflows, long-term growth expectations, and discount factors reflecting the risks of the business units. By its
nature, such information is subjective and hard to verify and audit

Related Literature

In particular, the study’s findings suggest that overvaluation of firms’ stock induces managers to engage in
value-destroying acquisitions that ultimately lead to goodwill impairment.
To conclude, the available evidence on the determinants of goodwill impairment based on firms applying the
impairment- only approach of IAS 36 (rev. 2004) is limited. Most existing studies focus on the U.S. However,
the accounting for goodwill differs in some regards between U.S. GAAP and IFRS and, furthermore, financial
reporting is strongly influenced by incentives and by capital market supervision and enforcement. Hence, it is
unclear whether findings based on U.S. sample firms generalize to non-U.S. firms applying IFRS.11 More
importantly, because we analyze the incidence of goodwill impairment for a comprehensive sample of firms in
21 countries where firms apply IFRS, we can investigate not only the impact of firm-level characteristics on
goodwill impairment, but also whether goodwill impairment differs across countries, depending on the strength
of national enforcement of accounting and auditing. Because goodwill impairment is potentially highly relevant
to capital market participants, but also is subject to managerial discretion, it is an empirical matter how
managerial and capital market incentives interact with national enforcement mechanisms in affecting the
impairment decision. In addition, by also providing evidence regarding the role that institutional investors play
in countries with relatively high and low levels of public enforcement, our tests permit us to examine the extent
to which private enforcement can substitute for public enforcement.

RESEARCH DESIGN AND PREDICTIONS

Estimating Equation

To examine the determinants of firms’ goodwill impairment decisions, we estimate a multivariate logistical
regression where the dependent variable, IMPAIR, equals 1 if goodwill is impaired in a given firm-year, and 0
otherwise. Based on economic reasoning and following prior research

we use firms’ stock market return in year t, adjusted by the return of the respective MSCI country index in year
t, RETURN, as a market-based measure of economic performance. We expect that share prices generally reflect,
in a timely manner, information about companies’ abilities to generate cash flows. We, thus, interpret a negative
stock market performance as an indication that the company’s assets have lost some of their ability to generate
future cash flows and are, therefore, impaired. Hence, we expect a negative correlation between RETURN
and incidence of goodwill impairment.

We investigate the timeliness of companies’ goodwill impairment losses by including the stock return from the
prior year, adjusted by the prior-year return of the respective MSCI country index, RETURN_LAG, in the
estimating equation. If goodwill impairment is delayed (timely), then we expect the RETURN_LAG
coefficient to be negative (not different from zero).

We include several variables reflecting managerial and firm incentives as potential influences on the goodwill
impairment decision. The first, COMP, is the ratio of the CEO’s variable income to total income. Hence, we
predict a negative association between COMP and the incidence of goodwill impairment.

we include a second variable, CEOTURN, which is an indicator variable that equals 1 if there is a change in
CEO in the current year, and 0 otherwise.

Two additional variables, adapted from Riedl (2004), are intended to capture incentives of managers to absorb
impairment losses in periods with unusually high or low income before recognition of a goodwill impairment
charge. We classify a firm as having an unusually high income if its income is positive and the change in its
income in the current year is above the median among those firms with a positive change in income. For such
firms, the indicator variable SMOOTH equals 1; for all other firms, SMOOTH equals 0. We classify a firm-year
as a big bath year (when a firm experiences an unusually large loss, prior literature suggests that managers have
an incentive to accelerate recognition of losses, thereby taking a ‘‘big bath’’ ), i.e., BATH equals 1, if its income
is negative and if it experiences a negative change in income in the current year that is below the median among
those firms with a negative change in income; otherwise, BATH equals 0. We predict a positive association
with goodwill impairment incidence for both SMOOTH and BATH.

A further incentive-related variable is LEVERAGE, which is total liabilities divided by total assets before
goodwill impairment. Therefore, we expect a negative association between LEVERAGE and goodwill
impairment incidence.

How impairment decisions are influenced by managers’ incentives likely is affected by the firm’s governance
and monitoring system, although it is difficult to predict the sign of the association. We include in our goodwill
impairment incidence model four governance and monitoring variables used in prior research. The first, BIG 4,
is an indicator that equals 1 for firms that are audited by a Big 4 auditing firm in a given financial year, and 0
otherwise. The second and third variables, FREEFLOAT and INSTI_OWN, pertain to firms’ ownership
structures. FREEFLOAT is the percentage of equity shares freely available to the investing public.
INSTI_OWN is the percentage of equity shares held by institutional investors. The fourth,
ANALYST_FOLLOW, is the average number of analysts that follow a firm during the year. better monitoring
could be negatively associated with the incidence of impairment.

We also include as controls several variables that prior literature suggests can affect the impairment decision.
The first set comprises variables relating to goodwill. The first is the ratio of goodwill, before goodwill
impairment, to total assets, before goodwill impairment, GW/TA. Other things equal, the greater is the ratio,
the greater is the firm’s exposure to impairment

The second is the number of operating segments, SEGMENTS, which is a proxy for the number of CGUs

The third is YEARS_IMP, the number of consecutive years with goodwill impairment losses before the current
year.

The second set comprises four additional firm-level control variables included in prior studies. These include
country- and industry-adjusted return on assets before goodwill impairment, ROA; a size proxy, SIZE, log of
end-of- year total assets before goodwill impairment; the year-end ratio of market value of equity to book value
of equity, MV/BV; and the standard deviation of monthly equity return during the year, RISK.

Influence of Enforcement on Timeliness of Impairment and Incentives

These studies lead us to investigate whether the strength of countries’ accounting and auditing enforcement
systems affects the discipline with which goodwill impairment is applied. Although it is difficult to predict the
sign of the direct effect of the strength of enforcement on the incidence of goodwill impairment, we predict that
the strength of enforcement has a positive interactive effect with timeliness and a negative interactive effect with
incentives. Therefore, we instead divide our total sample into observations from relatively low and high
enforcement environments, and estimate Equation (1) separately for each group. We predict goodwill
impairment in high enforcement countries to be more timely, i.e., more strongly related to
contemporaneous stock returns and less strongly related to lagged returns, than in low enforcement
countries. Similarly, to the extent that a higher level of enforcement mitigates the influence of managerial and
firm-level incentives, we expect predicted relations between these variables and the incidence of goodwill
impairment to be smaller in magnitude.
DATA AND SAMPLE

Our sample comprises all firm-years with available data in Worldscope for exchange-listed firms in the
countries that require IFRS in consolidated financial statements from 2005 through 2011.

Dependent Variable: Incidence of Goodwill Impairment

Descriptive Statistics for Determinants of Goodwill Impairment

The statistics presented in Table 3, Panel A reveal that firm-years with impairment tend to have lower
contemporaneous and lagged stock returns, RETURN and RETURN_LAG; higher proportions of CEO
changes, CEOTURN; abnormally high changes in pre-impairment operating income, SMOOTH; higher
leverage, LEVERAGE; higher analyst following, ANALYST_ FOLLOW; and higher free-float,
FREEFLOAT. Among the control variables, firm-years with impairment also have more goodwill as a
fraction of total assets than firm-years without impairment, GW/TA; a higher number of segments,
SEGMENTS; a higher incidence of goodwill impairment in prior years, YEARS_IMP; lower accounting
rate of return, ROA; larger size, SIZE; lower equity market-to-book ratios, MV/BV; and higher risk, RISK

RESULTS

Primary Findings
Table 4 contains regression summary statistics for Equation (1) and presents estimated regression
coefficients, z-statistics, and the associated p-values.22 Columns 1, 2, and 3 contain the findings for the
full sample, nonfinancial firms, and financial firms. For the sake of parsimony, we focus our discussion on
findings relating to the full sample, noting relevant differences for financial firms. 23

The findings in Table 4 reveal that contemporaneous stock return, RETURN, has the predicted significant
negative relation with the log-likelihood of impairing goodwill (coefficient 1⁄4 0.320; z-statistic 1⁄4
3.876 ). Thus, firms with poorer (better) economic performance, as reflected by market return, are more
(less) likely to impair goodwill. However, the findings also reveal that lagged stock return,
RETURN_LAG, also has a significantly negative relation with the log-likelihood of impairing goodwill
(coefficient 1⁄4 0.168; z-statistic 1⁄4 2.003). This finding suggests that although firms with poorer
economic performance are more likely to impair goodwill, they do not do so entirely on a timely basis.
This could reflect both the economic incentives of managers to delay impairment and weaknesses in the
enforcement of accounting and auditing standards at the national level.

Two of the managerial and managerial/firm incentives variables in Table 4 have the predicted relations
with goodwill incidence and are significant. In particular, the CEOTURN coefficient is significantly
positive (coefficient 1⁄4 0.484; z-statistic 1⁄4 6.162), suggesting that reputational concerns play a role in the
impairment decision: new CEOs have incentives to ‘‘clean out’’ goodwill from acquisitions made by their
predecessors and, relatedly, established CEOs may be reluctant to write down goodwill from recent
acquisitions for which they are responsible. In addition, the coefficient for SMOOTH is significantly
positive (coefficient 1⁄4 0.723; z-statistic 1⁄4 9.617), suggesting that firms with unusually high income use
impairment as a means to smooth income.
Tutorial 6
Investors think differently about future accounts, why is there a difference between book value and market
value? That’s what we want to understand
On e of them is internally generated assets. Also error in forecasts by accountants or markets.

If we use only accounting info to determine whether to invest or not, we miss a lot of info. That means that
accounting is not relevant.
Sustainability tab in stock value websites
The alternative fin. Statement would look like this. This is making accounting more relevant.

1. Mittelbach-Hörmanseder et al. (2020): “The information content of corporate social responsibility

disclosure in Europe: an institutional perspective”
 Analyzing mandatory CSR reporting in Europe to examine the value-relevance of CSR disclosures

Background information
Since 2017, under the Non-Financial Reporting Directive (NFRD), adopted in 2014, companies must
report on how sustainability issues affect their performance, position, and development, and the impact of
the company on people and the environment.
• The required non-financial information can be presented in either the MD&A as part of the annual report
or in a separate report
• Firms had to report on Topics: activities on ‘environmental, social and employee matters, respect for
human rights, anti-corruption and bribery matters’.

• Mitelbach et al, p.312: The CSR directive states that companies are obliged to include in their annual
management report a non-financial statement on the impact of their ‘development, performance, position’
and activities on ‘environmental, social and employee matters, respect for human rights, anti-corruption
and bribery matters’.
• Mitelbach et al, p.313:However, the directive does not require companies to apply a specific,
standardized CSR framework or member states to demand that the CSR reports be verified by an assurance
auditor: companies may prepare their statements on the basis of any internationally recognized standard-
frameworks, such as the UN Global Compact, and it suffices for an auditor to confirm that they have
published the mandatory report.
a lot of freedom and choices when to report on sustainability. Not verifies be auditor.

In 2021, the European Parliament introduced a new Corporate Sustainability Reporting Directive (CSRD),
which would amend the existing reporting requirements of the NFRD. The proposal
• extends the scope to all large companies and all companies listed on regulated markets (except listed
micro-enterprises)
• requires the audit (assurance) of reported information
• introduces more detailed reporting requirements, and a requirement to report according to mandatory
EU sustainability reporting standards
• requires companies to digitally ‘tag’ the reported information, so it is machine readable and feeds into the
European single access point
now there is less freedom and audit. We want to see how it is gonna effect reporting
In the future it will be harder to do reporting. Because more info is required.
Information content – new information that investors trade on
Introduction
• Corporate Social Responsibility (CSR) Reporting: Companies publishing reports with the goal of
sharing their corporate social responsibility actions and results
• Since the early 2000s, CSR reporting has both gained importance on the company level and attracted
considerable levels from policy makers and standard setters
• In 2014, the European Parliament issued a CSR directive to incentivize the largest firms to invest in CSR,
thereby triggering change towards a sustainable economy, which came into effect in 2017 and covered
approximately 11,000 companies
• In April 2021, the European Commission extended the 2014 directive, extending the scope of
mandatory CSR reporting to all large companies and companies listed on regulated markets
• Starting October 2022, the information disclosed in CSR reports is required to be audited

They tell about how their business is effected by certain CSR topics.
Example from Facebook’s 2020 sustainability report:
• Facebook has outlined various CSR related issues based on their importance to shareholders and business
impact
• For example, Operational Waste is of low importance, but has mediate shareholder relevance
• Data Privacy and Consumer Trust are of high importance to both shareholders and Facebook itself

1. Mittelbach-Hörmanseder et al. (2020) study value-relevance of CSR disclosure and the role of the
institutional environment in this relation. Explain in your own words why the authors examine
mandatory CSR disclosure as an exogenous shock as opposed to studying voluntary CSR disclosure.
We want to see why CSR is an important setting to look at regarding the firms’ value.

• Mittelbach-Hörmanseder et al. (2020) examine the effect from the switch from voluntary to mandatory
CSR reporting in the EU as a result of the 2014 directive
• Focus on mandatory CSR reporting: voluntary CSR disclosure suffers from a dual-selection problem, as
firms can choose their CSR activities and disclosures based on cost-benefit consideration
• This is less of an issue when CSR disclosure is mandatory, as all firms, no matter their characteristics and
circumstances, must report on CSR activities
• Mittelbach-Hörmanseder et al. (2020) focus on the effect of CSR reporting on the value-relevance of
CSR disclosure
• In addition, the authors study on the effect of the institutional environment (e.g., CSR awareness, degree
of enforcement and strength of the legal environment) on the value-relevance of CSR disclosure
You can expect the if the firm has a good environmental score, they would show it voluntarily.
Because it is mandatory, you can see the effect better.

We can have 2 contradicting effects, whether it is gonna create value for other stakeholders on the cost of
shareholders, or it is gonna improve their operations and shareholders value.
Theory and hypotheses

Prior literature provides neither theoretical nor empirical results for the relationship between CSR
disclosure and firm value. Two contradicting views:
1. The shareholder expense view (Friedman, 1970): CSR activities have a negative effect on a firm’s value,
because these activities create value of stakeholders at the expense of shareholders.
2. The stakeholder maximization view (Deng et al., 2013): CSR activities can increase value, as focusing
on the interests of different stakeholders increases their support of the firm’s operations, leading to benefits
for shareholders too
• The effects of CSR disclosure on firm value are more difficult to examine than the effects of financial
disclosure:
1. CSR activities are more difficult to predict than financial activities
2. Financial disclosure is mandatory and well-regulated for all firms, whereas CSR disclosure is not (yet)
3. Users of disclosed CSR information are more diverse in their backgrounds and interests compared to the
users of financial information (p. 315)

2. Based on your reading of Mittelbach-Hörmanseder et al. (2020), briefly explain why the authors
examine four topic-specific CSR measures instead of one overall CSR score to measure the effect of
CSR disclosure.

Theory and hypotheses

According to Mittelbach-Hörmanseder et al. (2020), most prior research on the disclosure of CSR
activities fall prey to two major flaws:
1. Voluntary CSR disclosure suffers from the dual-selection problem (recall two slides back)
2. Most studies regard CSR as one whole concept and do not consider the distinct topics it is composed of
• Disregarding the different components of CSR disclosure ignores the fact that some activities may drive
the CSR disclosures whereas other activities are neglected.
• One overall Environmental, Social, and Governmental (ESG) score or Kinder, Leydenberg, Domini
(KLD) rating may be positively related to financial performance, but the real effect of CSR activities on
financial performance depends on the importance of each theme behind the considered company and
relevant cost (p. 316; Dorfleitner et al., 2015)
First, they take all CSR activities, then cteate topics that take into account the most popular activities.
Mittelbach-Hörmanseder et al. (2020) divide a firm’s total CSR activities into multiple topic-specific
CSR activities, based on the six factors addressed in the EU’s CSR directive:
1. Environmental matters
2. Social and employee matters
3. Respect for human rights
4. Anti-corruption and bribery matters
• Hypothesis H1: A company’s annual disclosure on ...
a) Environmental matters
b) Social and employee matters
c) Human rights matters
d) Anti-corruption and bribery matters
... provides information to capital market participants that is in addition to the financial information the
company publishes
3. Mittelbach-Hörmanseder et al. (2020) not only examine the value relevance of CSR disclosure, but
also the role of the institutional environment. The conflicting effects of various institutional factors are
reflected in hypotheses H2a and H2b, as well as H3a and H3b. Briefly explain why you believe the
authors specify non-directional hypotheses.

As the economic consequences of firm disclosure differ in the institutional environment, Christensen et
al. (2019) argue that a combination of various institutional factors influences the consequences of CSR
reporting
• Mittelbach-Hörmanseder et al. (2020) first examine the effects of CSR awareness and employee
protection across countries. Empirical evidence and theoretical predictions are mixed:
1. Countries with strong CSR awareness and high levels of employee protection have more credible
information concerning CSR disclosures, as stakeholders can more closely monitor a company’s CSR
activities.
• More value-relevant disclosures
2. However, stronger CSR regulation might weaken this relation because CSR disclosure is more implicit
• Less value-relevant disclosures
• Hypothesis H2: The level of ...
a) CSR awareness
b) Employee protection
... in a particular country has either a positive or negative impact on the value-relevance of CSR disclosure

The same holds for the degree of enforcement and the strength of the legal environment:
1. Strong institutional and legal environments can reinforce the alignment of shareholders and stakeholders
with managers, such that CSR disclosure and legal environment act as substitutes.
• More value-relevant disclosures
2. High enforcements and a strong legal environment provide incentivizes CSR reporting and more
credible disclosures.
• Less value-relevant disclosures
• Hypothesis H3: The ...
a) Degree of enforcement; and
b) Strength of the legal environment
... of a particular country have either a positive or negative impact on the value-relevance of CSR
disclosure

Research design

• Topic-specific CSR disclosure measures: textual analysis

• topic-specific vocabularies for each thematic group
• Similarity: how close the reports are with these vocabularies.
• Example: TopicHuman Rights
• Vocabulary: search for ‘human rights’
• Collect a 20-word window surrounding this word
• Examine similarity in annual report to the words in the 20-word window

Research Design
Topic T explains the price per share. T before mandatory rules and after.

We look at how R^2 increases depending on a setting where the company operates.

The fin prices explain price (1)

In 2 we see that non-financial topics also explain price.

4. In Table 3, Mittelbach-Hörmanseder et al. (2020) present the results of Hypothesis H1a to H1d.
Explain whether the results are consistent with these hypotheses and explain on which statistics you
draw this conclusion.
All 4 topics are important in explaining share price.

H2 and H3 looking at r^2

Mandatory => value relevance main effect, but we also look at the moderators.
Additional results

• Companies that publish separate CSR reports and firms with a better overall CSR performance
experience an increase in their stock price
• Inclusion of additional variables do not change main results
• Results remain robust when taking into account firm-specific characteristics
• Findings do not hold for the sub-sample of firms in environmentally sensitive industries
• Poor alignment between firms’ CSR disclosure and CSR performance reduces the value-relevance of
disclosed CSR activities
• Findings continue to hold for different institutional factors and combinations of factors

Summary and conlusions from Mittelbach-Hörmanseder et al. (2020)

• Topic specific CSR disclosures concerning environmental matters, employee matters, social matters,
human rights matters, and anti-corruption and bribery are value-relevant
• Indicating usefulness for capital market participants (p. 336)
• Country-level CSR awareness and employee protection generally have a significantly negative effect
on the explanatory power of all topic-specific disclosed CSR activities
• The degree of enforcement displays a significant relationship with social, environmental and
employee matters, whereas the strength of the legal environment provides additional explanatory
power concerning the human rights, anti-corruption and bribery matters
The main conclusion that CSR disclosures should be used by investors to buy or sell shares.

2. Chen et al. (2018): “The effect of mandatory CSR disclosure on firm profitability and social
externalities: Evidence from China”
 Mandatory reporting of Corporate Social Responsibility and its effect on firm performance
5. Chen et al. (2018) examine the mandatory CSR reporting on firm performance and predict that firms
experience a decrease in performance subsequent to the mandatory CSR disclosure. Briefly explain
the reasoning behind this hypothesis.
Looking at the different outcomes at fin outcome – profitability and non fin – social externalities.
Introduction
• Mittelbach et al. (2021) show that mandatory CSR disclosure affects value relevance, and that this effect
depends on the institutional environment
• Furthermore, the effect also depends on the CSR topic
• Given the theoretical and empirical ambiguity, Mittelbach et al. (2021) displayed high levels of
uncertainty in their hypotheses and research design.
• This paper: also examines whether the strengthening of disclosure regulation affects firm performance,
but tries to estimate the causal effect by improving the research design
• Chen et al. (2018) focus on only mandatory Environmental CSR disclosure in China
• One single topic
• One single institutional setting
• Difference-in-Difference research design
It looks only at one topic at single institution setting with a control sample.

• Mandatory disclosure should impact a firm’s activities because the increased transparency can make it
easier for governments and interest groups to pressure firms to engage in more CSR activities.
• Effect of firm performance (ROA/ROE):
• Decrease in firm performance, because if CSR had benefited the firm, it would already be in place.
No mandatory rules would be necessary
• Therefore, the mandatory disclosure forces firms to do things they normally would find too costly
• However, the effect on actual pollution is not clear
• External monitoring may be an incentive to reduce pollution
• However, firms may also focus on other dimensions of CSR instead
• Agency conflict: only do CSR projects that benefit the manager
We can firms to do good, if we see what they disclose.
• Chinese government issued a CSR mandate in 2006 requiring companies to undertake social
responsibility in the course of conducting business
• Mandate came in effect in December 2008
• In addition, multiple CSR actions were implemented by the Chinese government:
1. Issuance of additional CSR guidelines
2. Tying bank access to bank financing to a firm’s CSR performance
3. High polluting firms placed on ‘blacklist’
4. Publicizing CSR performance rankings
5. Granting CSR rewards to companies with strong CSR records
If the firms do better, we expect better environment coefficients.
• Chinese firms on the Shanghai Stock Exchange (SSE) and Shenzhen Stock Exchange (SZSE) were
required to report information regarding activities in the following areas (p.172):
1. Protection of the interests of shareholders and creditors
2. Protection of workers’ rights
3. Protection of suppliers, customers and consumers
4. Environmental protection and sustainable development
5. Public relations and social welfare services
• Research question 1: what is the effect of mandatory CSR disclosure resulting from the 2006
mandate on firm performance?
• Does it increase or decrease profitability? Or does it have no effect?
• Research question 2: what is the effect of mandatory CSR disclosure on city-level pollution?

Theory and hypotheses

• Prior studies conclude that disclosure of CSR activities lead to increased employee morale, better firm
reputation, and more harmonious growth
• However, these studies examine CSR disclosure in a voluntary setting
• As in Mittelbach-Hörmanseder et al. (2020), Chen et al. (2018) study the effects of mandatory CSR
disclosure to avoid the dual-selection problem
• When firms are required to disclose their CSR activities, they should feel pressure to commit to CSR
• As shareholders and other stakeholders can now closely monitor CSR activities, firms are incentivized to
invest in CSR activities
• Firms would have undertaken these activities voluntarily had these activities been profitable, Chen et al.
(2018) expect a negative relation between mandatory CSR disclosure and firm performance
• Hypothesis H1: Firms experience a decrease in performance subsequent to the mandatory CSR disclosure

To assess the environmental impact of the disclosure mandate, Chen et al. (2018) focus on
environmental pollution, as climate change is arguably the most important component of CSR
reporting
• Once firms disclose CSR activities, the Chinese government and other stakeholders may find it easier
to ‘shame’ these firms into reducing their pollutant emissions (Thaler and Sunstein, 2008).
• Mandated CSR disclosure is therefore expected to lower pollution
• However, pollution may potentially not decrease as a result of firms choosing to spend money on CSR
activities in other areas that provide private benefits rather than social benefits
• Hypothesis H2: Cities experience a decrease in pollution subsequent to the mandatory CSR disclosure

Research design
• Sample of 6,952 Chinese firms listed on either the Shanghai Stock Exchange (SSE) or Shenzhen Stock
Exchange (SZSE)
• 1,674 treatment firms, subject to mandatory CSR disclosure
• 5,278 benchmark (control) firms
• Chen et al (2018) use Return on Assets (ROA) and Return on Equity (ROE) as measures of firm
profitability:
1. ROA allows the authors to make causal inferences regarding firm performance independent of leverage
2. ROE is used to make causal inferences more comparable regarding both shareholder wealth as well as
stock prices
We focus on ROA

We wanna know beta 3

Hypotheses
• Because:
a) Regulatory setting are important (see Mittelbach et al. (2021), and Chen et al. (2018) have a strong
regulatory setting to examine the effects of mandatory disclosure, and
b) CSR is likely to not have only financial reporting and capital market effects, but also affect economic
externalities, such as actual pollution
... Chen et al. (2018) are able to measure causal effects of CSR Disclosure on firm performance
• Governments can apply more pressure on disclosure requirements
• Governments can actually intervene in economic activities to affect pollution levels

• H1 (directional): Firms experience a decrease in performance subsequent to the mandatory CSR

disclosure.
• H2 (directional): Cities experience a decrease in pollution subsequent to the mandatory CSR disclosure
• Two design choices improve the statistical tests of Chen et al. (2018) over Mittelbach et al. (2021):
• Difference-in-Difference model: allows for the counterfactual to be approximated
• Matching design: excludes other (unobservable) factors that may affect the outcomes
We look at the causal effect.
6. Create Libby boxes for the first hypothesis of Chen et al. (2018). You may ignore the control variables
in the framework.

7. Explain in which table(s) in Chen et al. (2018) we can find the results of hypotheses H1 and H2. Are
these results consistent with the hypotheses? On which test statistic(s) do you draw these
conclusions?

Negative effect, ROA is lower for firms in the mandatory setting, comparing to co
Additional results
• Firms subject to mandatory CSR disclosure experienced a significant decrease in sales revenue and
capital expenditures and a significant increase in operating expenses and impairment losses (Table 3,
Panel B, p. 176)
• Decrease in firm performance is mainly due to spending on environmental issues
• Cities firms with greater CSR spending experience a greater decrease in pollution
• Firms with greater government control spend more on CSR activities concerning staff protection and
public relations activities (Table 5, p. 181)
• Firms in the most polluting industries invest more in environmental protection (Table 5, p. 181)

• Mandatory disclosure of CSR activities leads to positive externalities, but firms bear the cost of lower
profitability

Summary and conclusions from Chen et al. (2018)

• The Mandatory CSR disclosure by Chinese firms leads to firms spending more on various CSR
activities
• This leads to firms experiencing a decrease in profitability
• Consistent with the idea that mandatory CSR reporting increases both social- and political pressure
regarding a firm’s CSR activities
• Cities with a relatively large number of treatment firms experience a decrease in pollution levels
subsequent to mandatory CSR disclosure
• Mandatory disclosure of CSR activities leads to positive externalities, but firms bear the cost of lower
profitability