Employees Social Media

THE ACCOUNTING REVIEW American Accounting Association
Vol. 95, No. 2 DOI: 10.2308/accr-52519

March 2020
pp. 199–226
What Do Employees Know? Evidence from a Social Media

Platform
Kelly Huang
Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

Florida International University
Meng Li
Stanimir Markov
The University of Texas at Dallas
ABSTRACT: We use employee predictions of their companies’ six-month business outlook from Glassdoor.com to
assess the information content of employee social media disclosures. We find that average employee outlook is
incrementally informative in predicting future operating performance. Its information content is greater when the
disclosures are aggregated from a larger, more diverse, more knowledgeable employee base, consistent with the
wisdom of crowds phenomenon. Average outlook predicts bad news events more strongly than good news events,
suggesting that employee social media disclosures are relatively more important as a source of bad news.
Consistent with the organizational theory, we find systematic differences in the quantity and nature of the information
in employee disclosures when the disclosures are grouped based on employee attributes and job responsibilities.
Finally, average outlook predicts future returns of firms that attract less attention by analysts and investors,
suggesting that investors in these firms use outlook inefficiently.
Keywords: employee knowledge; job function; social media; wisdom of the crowd; stock returns.
I. INTRODUCTION
C
ompany employees possess valuable information that can be used not only in internal budgeting and planning, but also
in external disclosure and investor relations.1 Knowing how easily employees can share information through social
media, many companies prohibit the disclosure of material non-public information, such as financial and operational
predictions. Enforcing these bans is difficult because such disclosures have proliferated on social media and are often posted
anonymously. Furthermore, the bans do not explicitly prohibit the disclosure of information that is non-material by itself, but
material in the context of other information. This suggests that employee social media disclosures could be a source of valuable
information in capital markets.
Exploring the informativeness of employee social media disclosures is important for several reasons. First, existing
disclosure models generally assume that the quantity, quality, and timing of corporate disclosures are decided by top
management (Skinner 1994; Bamber, Jiang, and Wang 2010). But if valuable information is flowing via social media, from
rank-and-file employees to capital markets, then these models should be modified. Second, the conventional view of the
intermediary role of social media is that it facilitates information sharing among investors (Chen, De, Hu, and Hwang 2014;
Jame, Johnston, Markov, and Wolfe 2016). A distinct and still-unexamined role of social media is that it can increase the flow
We thank Qiang Cheng (editor), three anonymous reviewers, Ilia Dichev, Artur Hugon, Jay Jung, Art Kraft, Ningzhong Li, Xi Li, Greg Miller, James R.
Moon (discussant), Naim Bugra Ozel, Peter Pope, Gil Sadka, Ane Tamayo, Sorabh Tomar, Senyo Tse, Tara Vakil, Jieying Zhang, and seminar participants
at Cass Business School, The London School of Economics and Political Science, Southern Methodist University, Southwestern University of Finance and
Economics, Texas A&M University, University of Michigan, The University of Texas at Dallas, and the 2018 AAA Annual Meeting for their comments
and suggestions. We also thank Glassdoor, Inc. for sharing their data with us.
Editor’s note: Accepted by Qiang Cheng.
Submitted: September 2017
Accepted: June 2019
Published Online: August 2019
1
Meetings with technical staff and mid-level management are generally viewed as ‘‘best investor relations practices’’ and are highly sought after by
financial analysts and portfolio managers (Rossi 2010).
199
200 Huang, Li, and Markov
of information from companies to investors by helping employees share information beyond their immediate circle of friends
and relatives, a phenomenon that corporate managers may find difficult to monitor or curtail.
The primary challenge in exploring the informativeness of employee social media disclosures is the difficulty in observing
the disclosures themselves. Investment-oriented social media sites (e.g., Seeking Alpha, Estimize, and Motley Fool) do not
provide, and in many cases do not gather, employment information. We overcome this challenge by focusing on
Glassdoor.com, a popular job site that also distributes employee-provided company ratings and directional predictions of
company performance (i.e., business outlook over the next six months). We investigate whether employee-provided predictions
of a company’s future six-month performance (hereafter, outlook) can predict future performance, and also consider the nature
of the information embedded in outlook.2
Our sample includes 572,262 individual employee predictions—the employees chose between ‘‘get better,’’ ‘‘stay the
same,’’ and ‘‘get worse,’’ which we code as þ1, 0, and 1, respectively—for 2,270 unique companies from May 2012 to

December 2016. Together, these companies account for approximately 61 percent of the market capitalization and 63 percent of
aggregate earnings in the Compustat universe.
Our independent variable is the average of predictions issued by company employees during the quarter (Outlook).3 Our
primary measure of future operating performance is the average return on assets over the next two quarters. Controlling for
standard predictors of operating performance, competing information sources such as analyst and management forecast news,
insider net purchase, employee stock option exercises, and overall ratings on employers from Glassdoor,4 we find that a one-
standard-deviation increase in Outlook is associated with an increase in future ROA that is equivalent to 4.6 percent of its
standard deviation. Furthermore, the predictive ability of Outlook is greater when it aggregates a larger number of predictions,
when these predictions are made by employees who have diverse information (proxied for by dispersed job locations and
diverse job descriptions), and when employees are generally more knowledgeable about firm performance (indicated by greater
use of stock options in compensation).
Next, we explore whether the amount of information embedded in employee outlook varies systematically across
employees. Drawing on basic organizational theory, we expect that outlook has greater predictive value when it is obtained
from employees who work full time, have longer tenure (three or more years), have higher education (master’s or higher), hold
management positions, or work in the same state as their corporate headquarters. We augment our regression model by
including the average of predictions issued by employees who have a particular attribute. We find that full-time employees,
employees with longer tenure, and employees with higher education possess more information.
Of even greater interest is the extent to which the nature of the information embedded in outlook varies across employees.
Given the large average size of modern firms and the bounds placed on employees’ capabilities for information acquisition and
processing, employees are likely to be better informed about the activities in their own functional area than about activities in
other functional areas (Radner 1992). We use self-reported job descriptions to group employees into four functional areas
(Sales, Production, Research and Development [R&D], and Supply Chains), then test whether employees in different areas gain
superior information about different components of future operating performance (sales growth for Sales, cost of goods sold for
Production, R&D expenditures for R&D, and inventory turnover for Supply Chains). We find that outlook by employees from
a specific functional area is incrementally useful in predicting the component of firm performance most closely related to that
function.
We also hypothesize that outlook plays a greater role in conveying bad news than good news. Employees have stronger
incentives to acquire information about unfavorable future outcomes than about favorable future outcomes because their
payoffs, like debtholders’ payoffs, are less sensitive to firm performance in good states than to firm performance in bad states
(Chen, Kacperczyk, and Ortiz-Molina 2012). In addition, the incentives of employees to incorporate unfavorable information in
anonymous social media disclosures are not weakened by career concerns, known to weaken the incentives of corporate
managers and financial analysts to incorporate unfavorable information in non-anonymous public disclosures (Hong and Kubik
2003; Ke and Yu 2006; Kothari, Shu, and Wysocki 2009). We find evidence consistent with our asymmetric information
content hypothesis. When we partition the sample into (1) positive and non-positive Outlook, or (2) Outlook above and below
its sample median, we find that Outlook has significantly greater predictive value when it is non-positive or below its sample
median. We also find that Outlook predicts credit rating downgrades more strongly than credit rating upgrades, and predicts
dividend decreases more strongly than dividend increases.
2
We focus on employee outlook because its information content can be easily assessed by relating it to future performance. In contrast, employee
company ratings measure employee satisfaction, so their information content is more difficult to ascertain.
3
We include predictions provided by both current employees and former employees who left the company within the past two years. We present a
detailed discussion of this choice in Section III.
4
We control for changes in overall employer ratings from Glassdoor because Green, Huang, Wen, and Zhou (2019) show that employer rating changes
predict future stock returns.
The Accounting Review

Volume 95, Number 2, 2020
What Do Employees Know? Evidence from a Social Media Platform 201
All of this leads to the question of whether investors are taking full advantage of the information in employee outlook. In
general, investors have incentives to use all available information in forecasting future earnings, but limited cognitive resources
may lead to unawareness of Glassdoor disclosures or inability to discern their investment value, especially among stocks with
small capitalization and limited analyst following, where pricing is relatively less efficient. We find that among stocks with
below-the-median market capitalization and below-the-median analyst following, an equal-weighted hedge portfolio, formed
monthly by going long (short) in stocks in the top (bottom) decile of employee outlook, yields monthly alphas of 58 and 73
basis points, respectively. These results are robust to controlling for employer ratings, which Green et al. (2019) have shown to
be a predictor of future returns.
Our study fits within a growing empirical literature that investigates whether rank-and-file employees are, on average,
informed about overall firm performance. Early studies that use data on employee stock option exercises and stock purchases
find mixed evidence (Core and Guay 2001; Huddart and Lang 2003; Babenko and Sen 2016). More recently, Hales, Moon, and

Swenson (2018) and Green et al. (2019) analyze data from Glassdoor.com and find that employee disclosures are useful in
predicting operating and stock performance, respectively. We complement and extend these studies in two ways. First, we
strengthen the case that employee disclosures have predictive ability by including comprehensive controls for competing
information sources. Second, and more importantly, we present evidence that speaks to the amount and the nature of the
information possessed by employees. Specifically, we show that the amount of information held by an employee depends on
her employment status, experience, and education, and that the nature of her information depends on the nature of her job.
Consistent with employees having stronger incentives to acquire information on bad news events, we also demonstrate that
average outlook is a stronger predictor of bad news than good news.
Because our findings are consistent with the fundamental idea that information and knowledge are widely dispersed in
large organizations (Coase 1937; Hayek 1945), they should be of interest to researchers in organizational economics who seek
to understand how information is collected, processed, and acted upon (Gibbons and Roberts 2012). Large-sample evidence on
the distribution of information in companies is scarce due to the difficulty of observing what employees know, and is largely
limited to Li, Minnis, Nagar, and Rajan’s (2014) findings on the distribution of knowledge within a small group of top
managers who participate in conference calls. In contrast, this paper sheds light on how information is distributed among rank-
and-file employees.
Our study also contributes to the nascent literature that explores the significance of social media in capital markets. One
stream of research examines how investment opinions and earnings predictions on crowd-sourced investment sites, such as
Seeking Alpha and Estimize, reveal new information to capital markets (Chen et al. 2014; Jame et al. 2016). These studies,
although useful in assessing the effects of crowdsourcing, cannot identify which predictions are submitted by company
employees. In contrast, we focus on a non-investment social media platform, Glassdoor.com, where company employees are
the main source of new information. Our finding that investors inefficiently use employee predictions on Glassdoor
complements prior results that show analysts’ and investors’ inefficient use of investment research on Estimize and Seeking
Alpha (Chen et al. 2014; Jame et al. 2016) and of customers’ product reviews on Amazon (Huang 2018).
The rest of the paper is organized as follows. In Section II, we review the literature and develop our hypotheses. In Section
III, we discuss our sample and provide descriptive statistics. In Section IV, we present the results of our baseline analyses and
the cross-sectional tests. In Section V, we report stock return analyses. We conclude in Section VI.
II. BACKGROUND AND HYPOTHESIS DEVELOPMENT
Employees as an Information Source

In modern corporations, employees create substantial value by building client relationships and inventing new products
(McGregor 1960; Zingales 2000). As direct participants in firm operations, employees acquire valuable information on their
firms’ future prospects. For instance, sales associates who interact directly with customers obtain immediate information about
product demands; assembly line workers acquire firsthand information about product quality; and scientists and engineers learn
directly whether a new product can be developed successfully or launched to the market on time. Recognizing the value of this
information, many large corporations, including Ford Motor Company, Best Buy, and Hewlett-Packard, operate internal
prediction markets to extract information on product demand, project completion, and R&D from their employees (Wolfers and
Zitzewitz 2004; Dvorak 2008; Cowgill and Zitzewitz 2015).5 More broadly, a key objective of organizational design is to
ensure that information from throughout the organization is collected and used in centralized decision making (Brickley, Smith,
and Zimmerman 2009).
5
Internal prediction markets are markets where employee participants trade virtual securities based on their private information on the event outcome.
The price of the virtual securities corresponds to the aggregated beliefs on the probability of the event outcome by all the traders.

The employees’ dual role as drivers of firm value and sources of value-relevant information creates demand among
investors to know what employees know. To meet this demand, firms offer institutional investors opportunities for face-to-face
interactions with corporate insiders during analyst/investor days and corporate site visits (Cheng, Du, X. Wang, and Y. Wang
2016, 2019; Kirk and Markov 2016). According to a recent survey, investors place a premium on meetings with technical staff
and mid-level management (Rossi 2010). However, companies typically schedule these meetings infrequently and make only a
small subset of firm employees available, thereby limiting the transfer of information and knowledge to investors.
Prior empirical literature finds mixed evidence on whether rank-and-file employees have valuable information. While Core
and Guay (2001) find no evidence that option exercises by non-executives reflect private information, Huddart and Lang (2003)
find that lower employee stock option exercise is followed by higher stock returns. More recently, Babenko and Sen (2016)
document that employees’ aggregate purchases of company stock are predictive of future stock returns. In a closely related
study, Hales et al. (2018) find that employee outlook is useful in predicting key income statement items such as future sales and

gross margin, transitory items such as restructuring charges, earnings surprises, and management forecast revisions.
Employee Social Media Disclosures

Employees have always shared information with friends, relatives, and acquaintances, shaping the investment decisions of
those around them. According to Shiller and Pound (1989, 55), 24 percent of the individual investors who have recently
purchased company stock identify someone who worked at or is involved with the company as what first drew their attention to
it; this is only slightly below the 33 percent of the respondents who identify their broker as what first drew their attention to the
company. Moreover, cases of insider trading, where company employees act as tippers, are not uncommon.
Social media make it easier for employees to share information with others. With social media, employee disclosures are
no longer limited to an immediate circle of friends; they can be made anonymously and replicated many times to reach a
much broader audience. Unsurprisingly, many large companies have formal social media policies that prohibit disclosure of
any non-public financial or operational information.6 However, because social media disclosures do not always reveal
contributor affiliation or the nature of the underlying information, these prohibitions are difficult to enforce. Nor do the
prohibitions explicitly prohibit the disclosure of information that is non-material by itself, but material in the context of other
information.
Surprisingly little is known about the prevalence and significance of employee social media disclosures. One reason for
this is that investment-oriented social media sites (e.g., Seeking Alpha, Estimize, and Motley Fool) do not provide, and in many
cases do not gather, employment information. Another is the researchers’ tendency to focus on investors: social media is
commonly perceived as facilitating investors’ information sharing and reducing their reliance on gatekeepers, such as sell-side
analysts and firm management, for information. This view is somewhat incomplete in that it overlooks social media
participation by employees.
Hypotheses
As argued above, company employees acquire direct and unique information on the job. Recognizing the value of this
information, many large corporations operate internal prediction markets to extract the information (e.g., Wolfers and Zitzewitz
2004). Also, institutional investors appreciate the importance of employees’ information by placing a premium on meetings
with non-executive employees (Rossi 2010). Therefore, employees possess valuable information that can be shared easily,
broadly, and anonymously via social media. Our identification strategy is to use employee directional predictions of company
performance over the next six months available on a popular job site, Glassdoor.com.
Launched in 2008, Glassdoor.com is a website where current and former employees post reviews about employers, sharing
information on areas such as compensation and benefits and job interviews. Up to December 2016, more than 4.6 million
individual reviews on 34,114 U.S. companies had been posted. Since May 2012, as part of the company reviews on the site,
employees have been asked to rate their companies’ six-month business outlook by choosing one of three options: ‘‘get better,’’
‘‘stay the same,’’ or ‘‘get worse.’’7
Several features of Glassdoor suggest that its reviews reasonably accurately represent employee views. First, the reviews
are anonymous, allowing employees to express their views without fear of employer retaliation. Second, under Glassdoor’s
‘‘give to get’’ policy, effective since 2007, employees gain access to the most valuable information about employers only if they
6
For Best Buy Social Media Policy, please refer to: https://forums.bestbuy.com/t5/Welcome-News/Best-Buy-Social-Media-Policy/td-p/20492. For Ford
Social Media Policy, please refer to: https://corporate.ford.com/microsites/sustainability-report-2015-16/doc/sr15-ford-social-media-guidelines.pdf
7
In addition to providing business outlook, employees provide an overall employer rating (from one to five stars), as well as optional ratings of senior
management, career opportunities, compensation and benefits, work/life balance, and culture and values. Employees can also choose to approve or
disapprove of their company CEO and recommend or not recommend their employer to a friend.

themselves provide reviews.8 Marinescu, Klein, Chamberlain, and Smart (2018) suggest that the ‘‘give to get’’ policy
encourages reviews from individuals who hold moderate views and who would otherwise tend not to contribute, alleviating the
so-called polarization bias. Third, Glassdoor works to identify and remove employee reviews that appear to have been
incentivized or coerced by employers. The site lets users flag problematic content, which taps into the collective knowledge of
its user base; it investigates all flagged reviews, often by contacting employers; and it removes content determined to have been
affected by gaming or abuse.9 In sum, these Glassdoor features suggest that employees can be truthful in their reviews, and that
if any gaming or abuse by employers does occur, there are policies in place to curtail it.10
If employees truthfully disclose their information via outlook, then outlook would be informative about firm future
performance. One caveat is that on Glassdoor, predicting company performance is a secondary task that employees complete
after reviewing the company as an employer. As a result, employee outlook on Glassdoor may be less accurate than predictions
that employees contribute to investment sites such as Seeking Alpha or Estimize, where forecasting performance is the primary

task. At the same time, Glassdoor likely has a greater volume of employee predictions, as employees are more likely to provide
outlook on Glassdoor for accessing jobs data than to provide earnings forecasts on investment sites. The aggregation of more
employee predictions can enhance the information content of employee disclosure on Glassdoor. Therefore, our first hypothesis
is as follows:
H1: Employee outlook is informative about firm future performance.
There are reasons, however, why employee outlook may have no significant information value. First, employees may
unintentionally bias their outlook if their assessment of future performance is unduly affected by their personal experience with
the firm, i.e., if those with positive (negative) experiences predict improved (worsened) firm performance. Also, employees may
intentionally bias their outlook upward to preserve their jobs and curry favor with upper management,11 attract high-quality
coworkers, or increase the stock price (especially if they hold employer stock). Second, given that an employee’s job is only a
small component of the company’s business, an employee may overvalue information about her part of the company, making
her outlook less useful. Unless a sufficiently large number of employees contribute outlook on Glassdoor, the information
content from these contributions could be negligible. Finally, information contained in outlook may be subsumed by
information contained in employee stock trades and management disclosures (Huddart and Lang 2003; Hirst, Koonce, and
Venkataraman 2008; Babenko and Sen 2016).
We also derive and test a series of hypotheses about how the information content of employee outlook varies across firms
and employees. In a large organization, information is widely distributed across many employees who are located in different
geographic locations and who engage in different job activities (Hayek 1945). The wisdom of crowds theory suggests that
information aggregated from a large group of people has greater predictive ability than information from any individual
member (Surowiecki 2004). Furthermore, recent research suggests that the benefit from combining individual predictions
increases with the unique additional information contained in each prediction (Jame et al. 2016; Adebambo, Bliss, and Kumar
2016). Therefore, we expect employee outlook to be more informative if it is aggregated from a greater number of reviewers,
from more diverse reviewers (proxied for by the diversity of reviewer geographic locations and job activities), and from more
knowledgeable reviewers. Our second hypothesis is as follows:
H2: Employee outlook is more informative when it aggregates the opinions of a larger, more diverse, and more
knowledgeable reviewer base.
Our next prediction relates the information content of employee outlook to employee attributes. We expect that full-time
employees and employees with longer tenure, by virtue of their greater engagement with the firm, acquire more information
about the firm and, thus, provide more informative outlook than do other employees. We also expect that employees who have
more education and who work in management provide more informative outlook due to their greater ability to gather and
process information and their involvement in higher-level decision making. Finally, employees working in the corporate
headquarters states are likely to provide more informative outlook due to their greater access to top management and better
knowledge of corporate-level strategies and initiatives. Our third hypothesis, thus, is as follows:
8
Employees may also derive utility from sharing their views or from participating in sweepstakes. Under the ‘‘contribute and win’’ policy, Glassdoor
gives small rewards to randomly chosen reviewers.
9
Glassdoor uses proprietary technology and human moderators to monitor content. However, our private communications with a Glassdoor
representative reveal that most cases of fraudulent reviews are discovered by users.
10
Anecdotal evidence suggests that job seekers use Glassdoor reviews to screen out bad employers, and employers use Glassdoor reviews to design HR
policies and make acquisition decisions (Demers 2014; Fuhrmans 2017). This reassures us that the reviews generally represent employee views.
11
A recent Wall Street Journal article reports that some companies take measures to encourage their employees to post positive reviews on Glassdoor
(Winkler and Fuller 2019).

H3: Full-time employees and employees with longer tenure, with higher education, with management positions, and
located in headquarters states provide more informative employee outlook than other employees.
Our fourth prediction is that the nature of the information possessed by employees varies across functional areas. Because
the scope of activities in a modern firm is exceedingly large and employees’ capabilities for information acquisition and
processing are limited (Radner 1992), employees in one functional area are better informed about their own activities than
about the activities of employees in other functional areas. Also, activities in different functional areas affect different
components of firm performance. For example, activities in the Sales area affect sales growth the most, and activities in the
Production area affect the cost of goods sold in particular. This leads to our fourth hypothesis:
H4: Employee outlook provided by employees from a specific functional area is more informative about the component of
firm performance closely related to that function.

Our final prediction is that employees play a greater role in conveying bad news than good news. Employees have stronger
incentives to acquire information about unfavorable future outcomes than favorable future outcomes because their payoffs, like
debtholders’ payoffs, are less sensitive to firm performance in good states than to firm performance in bad states (Chen et al.
2012).12 In addition, since employee predictions on Glassdoor.com are anonymous, employees are less likely to avoid bad news
disclosures as top managers and financial analysts are known to do (Hong and Kubik 2003; Ke and Yu 2006; Kothari et al. 2009).
Therefore, employee outlook could be particularly important in revealing bad news. Thus, our final hypothesis is as follows:
H5: Employee outlook is more informative about future bad news than good news.
III. DATA AND DESCRIPTIVE STATISTICS
Data
We obtain approximately one million employee reviews for 6,467 public firms from May 2012 to December 2016 directly
from Glassdoor. Merging with the Compustat universe (using both ticker symbols and company names) reduces the sample to
719,280 reviews for 4,307 unique firms (43,075 firm-quarters). Some 395,300 reviews are by current employees, and 323,980 are
by former employees. At departure, former employees’ knowledge equals current employees’ knowledge, but the former
employees’ knowledge gradually becomes obsolete as circumstances change and as access to company insiders is curtailed. In
Appendix B, we show that current employees have more valuable information than former employees who departed more than two
years ago, but not more than former employees who departed within the past two years. We, therefore, exclude reviews by former
employees who departed more than two years ago, leading to 651,938 reviews from 4,229 unique firms (41,835 firm-quarters).13
Finally, we drop firm-quarters with fewer than three reviews or with insufficient Compustat control variables or missing stock return
information from CRSP, reducing the sample to 572,262 reviews for 21,763 firm-quarters from 2,270 unique public firms.14
Descriptive Statistics
Table 1 reports summary statistics of Glassdoor.com reviews for our sample firms. Panel A reports our sample distribution
by year. There are, on average, 26 reviews for each firm-quarter. Consistent with Glassdoor’s self-described rapid growth, the
average number of reviews per firm-quarter increases from 14 in 2012 to 31 in 2016, and the number of public firms that are
covered also increases from 755 in 2012 to 1,883 in 2016. Panel B reports our sample distribution by Fama-French 12
industries. The industries with the greatest number of employee reviews per firm-quarter are Wholesale/Retail,
Telecommunications, and Finance. The Wholesale/Retail sector’s average of 53 reviews per firm-quarter far exceeds other
industries, consistent with the large labor base in this sector. Panel C compares industry composition in our sample with that in
the Compustat universe during our sample period. We find no large difference between these two samples, although the
Business Equipment and Wholesale/Retail sectors are overrepresented in our sample relative to the Compustat universe (21.9
percent versus 14.6 percent for Business Equipment and 11.5 percent versus 6.5 percent for Wholesale/Retail).
We first code the individual employee outlooks ‘‘get better,’’ ‘‘stay the same,’’ and ‘‘get worse’’ as 1, 0, and 1,
respectively, then average all individual outlooks in a firm-quarter to construct our variable of interest, Outlook. We winsorize
12
Although some companies award stocks and options to rank-and-file employees, Bova, Kolev, Thomas, and Zhang (2015) find that about two-thirds of
their sample firms have zero employee stockholding.
13
Our results hold throughout the paper if we only include reviews by current employees.
14
We use three reviews as the cutoff to strike a balance between reducing noise in data and increasing sample firm coverage. Our results are generally
unchanged when we require each firm-quarter to have at least four, five, or six reviews.

TABLE 1
Summary Statistics on Glassdoor.com Reviews for Our Sample Firms
Panel A: Yearly Distribution

Mean Median
Reviews per Reviews per
Reviews Firm-Quarters Firms Firm-Quarter Firm-Quarter
2012 20,542 1,432 755 14 7
2013 68,132 3,585 1,262 19 8
2014 114,402 4,881 1,629 23 9

2015 185,976 5,954 1,872 31 11
2016 183,210 5,911 1,883 31 11
Full Sample 572,262 21,763 2,270 26 10
Panel B: Industry Distribution

Mean Median
Reviews per Reviews per
Reviews Firm-Quarters Firms Firm-Quarter Firm-Quarter
Consumer Durables 21,078 1,116 101 19 10
Consumer Non-Durables 7,512 527 58 14 7
Manufacturing 28,943 1,744 200 17 8
Energy 8,786 619 83 14 7
Chemicals 8,612 507 57 17 8
Business Equipment 114,996 4,852 529 24 9
Telecommunications 28,160 676 69 42 9
Utilities 4,174 457 51 9 6
Wholesale/Retail 178,630 3,376 279 53 15
Healthcare 24,950 1,632 204 15 8
Finance 84,047 3,273 361 26 10
Other 62,374 2,984 308 21 11
Panel C: Comparison of Sample Industry Composition to Compustat

Our Sample Compustat
Unique Sample Unique Sample
Firms Percentage Firms Percentage
Consumer Durables 101 4.18% 414 3.65%
Consumer Non-Durables 58 2.40% 212 1.87%
Manufacturing 200 8.27% 751 6.62%
Energy 83 3.43% 786 6.93%
Chemicals 57 2.36% 202 1.78%
Business Equipment 529 21.87% 1,653 14.57%
Telecommunications 69 2.85% 271 2.39%
Utilities 51 2.11% 325 2.87%
Wholesale and Retail 279 11.53% 739 6.52%
Healthcare 204 8.43% 1,411 12.44%
Finance 361 14.92% 2,211 19.49%
Other 308 12.73% 2,367 20.87%
This table reports summary statistics on Glassdoor.com reviews. Panel A presents the sample distribution by year. Panel B presents the sample distribution
by Fama-French 12 industries. Panel C compares the industry composition of our sample with that of Compustat during the sample period.

TABLE 2
Descriptive Statistics
Variable n Mean STD 25th 50th 75th
Outlook 21,763 0.1648 0.4132 0.1034 0.2000 0.4570
AvgROA(tþ1,tþ2) 21,763 0.0076 0.0259 0.0017 0.0093 0.0196
ROA 21,763 0.0075 0.0282 0.0015 0.0094 0.0201
LogSale 21,763 6.5208 1.6291 5.3874 6.4823 7.6474
MTB 21,763 3.7451 7.4573 1.4559 2.5330 4.4247
LEV 21,763 0.2385 0.2113 0.0586 0.2034 0.3570
Age 21,763 23.3860 20.2131 9.0000 18.0000 31.0000

RET 21,763 0.0263 0.1624 0.0656 0.0267 0.1189
STDRET 21,763 0.0204 0.0104 0.0133 0.0176 0.0244
TACC 21,763 0.0160 0.0324 0.0280 0.0128 0.0008
Acquisition 21,763 0.0987 0.2983 0.0000 0.0000 0.0000
AFNews 21,763 0.0020 0.0155 0.0015 0.0004 0.0034
MFNews 21,763 0.0004 0.0059 0.0000 0.0000 0.0000
InsiderNetBuy 21,763 0.1201 3.0735 0.0335 0.0000 0.0655
OptionExercise 21,763 0.0150 0.1799 0.0671 0.0000 0.0871
Crowd_Negative 12,172 0.0339 0.0184 0.0217 0.0317 0.0437
Crowd_Positive 12,172 0.0402 0.0166 0.0300 0.0391 0.0492
ChgRating 21,763 0.0222 0.7101 0.3333 0.0083 0.3593
MVE 21,763 14.7458 30.8433 1.2217 3.7324 12.7015
AnalystFollowing 21,763 11.7007 8.0126 5.0000 10.0000 17.0000
InstitutionalOwnership 21,763 0.5842 0.2812 0.4705 0.6670 0.7801
This table describes the sample used in our main regression analyses.
Detailed variable definitions are provided in Appendix A.
Variable Definitions:
Outlook ¼ the arithmetic average of employees’ assessments of firm business outlook made in the current fiscal quarter. Employees make such assessments
by choosing ‘‘get better,’’ ‘‘stay the same,’’ or ‘‘get worse,’’ which we code as 1, 0, and 1, respectively;
AvgROA(tþ1,tþ2) ¼ the average return on assets for the next two quarters;
ROA ¼ return on assets, measured as income before extraordinary items divided by total assets;
LogSale ¼ the natural logarithm of firm sales;
MTB ¼ market-to-book ratio;
LEV ¼ leverage ratio;
Age ¼ firm age;
RET ¼ firm stock return over the fiscal quarter;
STDRET ¼ the standard deviation of daily stock returns over the fiscal quarter;
TACC ¼ total accruals;
Acquisition ¼ an indicator variable that equals 1 if the acquisition item is greater than 5 percent of the firm’s total assets, and 0 otherwise;
AFNews ¼ analyst earnings forecast news;
MFNews ¼ management earnings forecast news;
InsiderNetBuy ¼ abnormal net purchase made by insiders in the quarter;
OptionExercise ¼ the abnormal stock option exercised by firm employees in the most recent fiscal year;
Crowd_Negative ¼ the average proportion of negative words in all single-ticker articles published on Seeking Alpha about the firm over the current fiscal
quarter;
Crowd_Positive ¼ the average proportion of positive words in all single-ticker articles published on Seeking Alpha about the firm over the current fiscal
quarter;
ChgRating ¼ the average overall employer rating in the current quarter minus the average rating in the previous quarter;
MVE ¼ market value of equity;
AnalystFollowing ¼ the number of analysts following the firm; and
InstitutionalOwnership ¼ the percentage of ownership held by institutional investors.
all continuous variables except Outlook at 1 percent and at 99 percent to reduce the impact of outliers, and report key sample
statistics in Table 2.
We observe in Table 2 that the mean (median) Outlook is 0.1648 (0.2), indicating that, on average, employees expect
performance to improve.15 Outlook varies substantially, with a standard deviation of 0.4132. Our sample firms are large (mean
15
Our sample firms indeed experience an improvement in performance, as indicated by a positive change in ROA.

sales revenue of 2.4 billion), mature (mean age of 23 years), well capitalized (mean market-to-book ratio of 3.7), and profitable
(mean ROA of 0.008). Also, the average sample firm has market value of 14.7 billion, is followed by 12 analysts, and has 58
percent of its shares owned by institutional investors.
IV. EMPIRICAL RESULTS
Predicting Future Operating Performance with Employee Outlook

We test whether employee outlook predicts firm future performance (H1) by estimating the ordinary least squares (OLS)
regression model:
X X
AvgROAi;tþ1;tþ2 ¼ b0 þ b1 Outlooki;t þ b2 Controlsi;t þ IndustryFE þ TimeFE þ ei;tþ1;tþ2 ; ð1Þ

where AvgROA is the average return on assets, measured over the next two quarters to approximately match the six-month
horizon of Outlook, the variable of interest,16 and where Controls is a vector of 16 variables, which include standard measures
of current performance (e.g., ROA and stock return) and firm characteristics (e.g., size), as well as measures of information
flows from competing sources: analyst forecast news (AFNews), management forecast news (MFNews), insider trades
(InsiderNetBuy), and employee stock option exercises (OptionExercise).17 We also include positive and negative crowd
sentiment from Seeking Alpha (Crowd_Negative and Crowd_Positive); this lets us control for general crowd opinions and
allows us to isolate the employees’ unique inside information from the public information that employees hold as a subgroup of
general crowd.18 We control for changes in employee overall rating from Glassdoor (ChgRating) because this variable is
positively correlated with Outlook, and Green et al. (2019) document that it predicts future stock returns. See Appendix A for
detailed definitions of all variables.
Table 3 reports the results from the estimation of three specifications: (1) control variables except general crowd opinions
included, (2) all control variables included, and (3) all control variables included and missing observations of AFNews,
MFNews, InsiderNetBuy, and OptionExercise dropped.19 The coefficient estimates on Outlook are consistently positive and
statistically significant at the 1 percent level. In terms of economic significance, based on specification (1), a one-standard-
deviation increase in Outlook is associated with an increase of AvgROA by 0.0012, about 4.6 percent of its standard deviation.
Three of the four measures of competing information sources in Table 3—analyst forecast news (AFNews), management
forecast news (MFNews), and net insider purchases (InsiderNetBuy)—are statistically significant in predicting future operating
performance. In terms of economic magnitude, a one-standard-deviation increase in AFNews, MFNews, and InsiderNetBuy is
associated with a 0.0039, 0.0010, and 0.0003 increase in AvgROA, respectively; these marginal effects are comparable to that of
Outlook.
In addition, Crowd_Positive and ChgRating do not load significantly in Table 3. The coefficient on Crowd_Negative is
significantly negative, and a one-standard-deviation increase in Crowd_Negative is associated with an increase of AvgROA by
0.0005. We conclude that Outlook has significant incremental predictive value relative to other major information sources.
Wisdom of Crowds Analysis

H2 is that the information content of Outlook increases with the size and diversity of the reviewer base and with
employees’ knowledge of firm performance. We suggest that diversity in the reviewers’ information sets is inherently related to
diversity in their job activities and locations. We propose that greater use of stock options in compensation indicates more
knowledgeable employees for two non-mutually exclusive reasons: (1) firms use employee stock options relatively more when
employees are more critical value drivers, and (2) employees who have more stock options have greater incentives to gather
information.
We augment Equation (1) by including an interaction term between Outlook and NumReviewer, HI_ReviewerState, HI_
ReviewerJob, and EmpStockOption separately. NumReviewer is the total number of reviewers in a firm-quarter. HI_ReviewerState
is the Herfindahl index of reviewer job locations, and HI_ReviewerJob is the Herfindahl index of reviewer job titles, with lower
values of the indices indicating greater diversity.20 EmpStockOption is the number of rank-and-file employee stock options,
16
Our results hold when we deflate earnings with market value of equity and when we define future operating performance as the change in ROA.
17
In a hand-collected small sample (3,811 firm-quarters reported in Column (3) of Table 3), we control for employee stock purchases. Untabulated results
show that the coefficient on outlook remains positive and significant, and the coefficient on employee stock purchase is positive, but insignificant.
18
We do not control for the two variables from Seeking Alpha for the rest of our tests because including them significantly reduces our sample size.
19
In specifications (1) and (2), missing observations of AFNews, MFNews, InsiderNetBuy, and OptionExercise are coded as zero.
20
The Herfindahl index of reviewer job locations ( job titles) is the sum of the squares of the number of reviewers in each state (with the same job title),
divided by the square of the total number of reviewers across all states ( job titles).

TABLE 3
Predicting Future Operating Performance
Dependent Variable: AvgROA(tþ1,tþ2)
(1) (2) (3)
Outlook 0.0028*** 0.0033*** 0.0040***
(5.93) (4.94) (3.48)
ROA 0.5738*** 0.5593*** 0.6246***
(32.98) (26.59) (12.27)
LogSale 0.0008*** 0.0005** 0.0000

(4.88) (2.45) (0.03)
MTB 0.0000 0.0000 0.0001
(0.96) (0.38) (1.05)
LEV 0.0025** 0.0024 0.0014
(2.14) (1.62) (0.48)
LogAge 0.0008*** 0.0007** 0.0003
(3.33) (2.21) (0.48)
RET 0.0095*** 0.0091*** 0.0125***
(8.17) (6.05) (3.60)
STDRET 0.4000*** 0.3632*** 0.3376***
(14.98) (10.67) (4.25)
TACC 0.0744*** 0.0667*** 0.0546*
(6.68) (4.44) (1.66)
Acquisition 0.0007 0.0026*** 0.0022*
(1.49) (3.76) (1.70)
AFNews 0.2497*** 0.2873*** 0.3895***
(8.11) (6.98) (3.36)
MFNews 0.1708*** 0.1619*** 0.092**
(5.35) (3.66) (2.47)
InsiderNetBuy 0.0001*** 0.0001 0.0002*
(3.17) (1.51) (1.77)
OptionExercise 0.0009 0.0006 0.0006
(1.23) (0.60) (0.36)
Crowd_Negative 0.0282** 0.0056
(2.58) (0.30)
Crowd_Positive 0.0085 0.0088
(0.83) (0.52)
ChgRating 0.0001 0.0000 0.0004
(0.61) (0.07) (0.60)
Industry/Time FE Yes Yes Yes
Observations 21,763 12,172 3,811
Adjusted R2 0.509 0.504 0.451
*, **, *** Significance at the 10 percent, 5 percent, and 1 percent levels, respectively.
This table presents results from OLS regressions of future return on assets on employee outlook. In specifications (1) and (2), missing values of AFNews,
MFNews, InsiderNetBuy, and OptionExercise are coded as zero; in specification (3), observations with missing values of the four variables are dropped. Fama-
French 48 industry fixed effects and quarter-end time fixed effects are included. t-statistics are reported in parentheses. Standard errors are clustered by firm.
Detailed variable definitions are in Appendix A.
calculated as the total number of stock options minus the number of stock options owned by top executives, scaled by the number
of shares outstanding.21 Each variable is converted to quintile ranks and transformed into values of 0, 0.25, 0.5, 0.75, and 1.
We interact Outlook with three measures of complexity—the number of geographic segments (NumGeoSeg), the number
of business segments (NumBusSeg), and earnings volatility (STDROA)—to alleviate the concern that firms with more complex
21
Due to the mechanical relation between the number of employees and the number of rank-and-file employee stock options, we first sort observations into
quintiles based on the number of employees; then, within each quintile, we sort observations into quintiles based on the number of rank-and-file employee
stock options; finally, we pool observations with the same employee stock options quintile rank across different number-of-employees quintile rank.

operations have a larger, more diverse, and more knowledgeable reviewer base, as well as more informative disclosures in
general due to greater information asymmetry.
Results, as tabulated in Table 4, are consistent with our predictions. In terms of economic importance, when Outlook is
equal to its sample mean, the difference in future ROA of firms in the top and bottom quintiles of number of reviewers,
geographic diversity, job title diversity, and employee stock options is 1.6 percent, 1.8 percent, 1.4 percent, and 2.5 percent of
the standard deviation of ROA, respectively.
Reviewer Attributes Analysis

H3 states that employees who are full-time, have longer job tenure, have more education, are at or above management
level, and work in headquarters states provide more informative outlook than other employees. To test whether employees with

a particular attribute have more information than other employees, we calculate the average of outlook provided by, e.g., full-
time employees in a firm-quarter, Outlook_FullTime, and augment Equation (1) by including Outlook_FullTime as a predictor.
Since Equation (1) includes Outlook (the average of full-time and part-time employees), we predict that the coefficient on
Outlook_FullTime is zero when full-time and part-time employees are equally informed, and positive when full-time employees
are better informed. In the same vein, Outlook_LongTenure is the average outlook of reviewers with three or more years of firm
experience; Outlook_HighEducation is the average outlook of reviewers with a master’s or doctoral degree; Outlook_Manager
is the average outlook of reviewers holding management positions;22 and Outlook_Headquarters is the average outlook of
reviewers located in the same state as their firm’s headquarters.
Panel A of Table 5 provides descriptive statistics about the percentage of predictions in a firm-quarter made by employees
with a particular attribute.23 We observe that, on average, full-time employees account for 84 percent of all predictions in a
firm-quarter. The corresponding figures for employees who have longer tenure, have higher education, work in management,
and are in headquarters states are 52 percent, 21 percent, 26 percent, and 48 percent, respectively.
Columns (1)–(5) in Panel B of Table 5 report the results for our five attributes. To alleviate multicollinearity concerns, we
exclude firm-quarter observations in which 100 percent of the reviewers possess the same attribute; e.g., when all reviewers are
full-time employees, or when all are part-time employees. Because principal component analysis reveals that a single factor
(Outlook_Factor) adequately explains the variation in our five attribute-based Outlook measures,24 Column (6) tabulates results
concerning its information content. We find that Outlook_FullTime, Outlook_LongTenure, Outlook_HighEducation, and
Outlook_Factor are significant predictors of future operating performance. The coefficients on Outlook_Manager and Outlook_
Headquarters are positive, as expected, but insignificant. In terms of economic significance, a one-standard-deviation increase
in Outlook_FullTime, Outlook_LongTenure, Outlook_HighEducation, and Outlook_Factor increases AvgROA by approxi-
mately 4.1 percent, 2.2 percent, 1.6 percent, and 5.6 percent of its standard deviation, respectively. As a reference, the
corresponding marginal effects of Outlook are 5.1 percent, 4.3 percent, 5.7 percent, and 4.0 percent. We conclude that the
information content of employee outlook varies systematically in relation to employee attributes.
Job Function Analysis

H4 states that employees from a specific functional area have superior information about the component of firm
performance related to that function. We assign employees to one of four major functional areas—Sales, Production, R&D, and
Supply Chain—based on self-reported job descriptions, and explore whether the nature of the information embedded in outlook
varies across functional areas.25 We expect employees in Sales to know more about future sales growth (AvgSG); employees in
Production to be better informed about future cost of goods sold (AvgCOGS); employees in R&D to be more knowledgeable
about future R&D expenditure (AvgR&D); and employees in Supply Chain to know more about future operating efficiency,
such as inventory turnover (AvgInvTurn).
To test H4, we calculate the average value of outlook made by employees in each functional area—Outlook_Sale,
Outlook_Production, Outlook_R&D, and Outlook_SupChain—and we augment Equation (1) by including, one at a time, each
individual function-specific outlook and replacing the dependent variable, overall operating performance, with a corresponding
component of operating performance: AvgSG, AvgCOGS, AvgR&D, and AvgInvTurn.
22
An employee holds a management position if her job title includes the following keywords: ‘‘manager,’’ ‘‘director,’’ ‘‘executive,’’ ‘‘leader,’’ ‘‘president,’’
‘‘officer,’’ ‘‘chief,’’ or ‘‘management.’’
23
We drop observations with missing attribute information.
24
In the principal component analysis, only the first factor identified has an eigenvalue greater than 1.
25
Note that only a subset of reviewers are categorized in these four functional areas. Other reviewers either do not disclose their job titles, or provide
generic job titles such as ‘‘analyst’’ or ‘‘manager.’’

TABLE 4
Wisdom of Crowds Analysis
(1) (2) (3) (4)
Outlook 0.0011 0.0038*** 0.0035*** 0.0027**
(1.12) (3.15) (2.69) (2.04)
NumReviewer 0.0004
(0.58)

Outlook NumReviewer 0.0025**
(2.01)
HI_ReviewerState 0.0006
(0.90)
Outlook HI_ReviewerState 0.0029**
(2.28)
HI_ReviewerJob 0.0001
(0.19)
Outlook HI_ReviewerJob 0.0021*
(1.67)
EmpStockOption 0.0023***
(2.96)
Outlook EmpStockOption 0.0031**
(2.11)
NumBusSeg 0.0001 0.0001 0.0001 0.0002
(0.58) (0.64) (0.71) (1.52)
NumGeoSeg 0.0001 0.0001 0.0001 0.0000
(1.03) (1.11) (1.31) (0.06)
STDROA 0.0415*** 0.0442*** 0.0451*** 0.0171
(2.81) (2.93) (2.92) (1.01)
Outlook NumBusSeg 0.0000 0.0001 0.0001 0.0002
(0.10) (0.22) (0.20) (0.73)
Outlook NumGeoSeg 0.0003** 0.0004** 0.0002* 0.0001
(2.27) (2.49) (1.67) (0.34)
Outlook STDROA 0.0001 0.0045 0.019 0.081**
(0.00) (0.15) (0.63) (2.47)
Other Controls Yes Yes Yes Yes
Industry and Time FE Yes Yes Yes Yes
No. of Observations 21,256 21,133 20,575 14,700
Adjusted R2 0.487 0.487 0.480 0.417
This table examines whether the predictive value of outlook varies with the number of firm reviewers, dispersion of reviewer locations, diversity of
reviewer job titles, and number of employee stock options. All other control variables are the same as in Column (1) of Table 3. Fama-French 48 industry
fixed effects and quarter-end time fixed effects are included. t-statistics are reported in parentheses. Standard errors are clustered by firm.
NumReviewer ¼ the number of reviewers who make outlook predictions on Glassdoor during the fiscal quarter;
HI_ReviewerState ¼ the Herfindahl index of states in which reviewers are located;
HI_ReviewerJob ¼ the Herfindahl index of reviewer job titles;
EmpStockOption ¼ the number of stock options owned by rank-and-file employees. We rank these variables into quintiles and transform them into values
of 0, 0.25, 0.5, 0.75, and 1;
NumBusSeg ¼ the number of business segments;
NumGeoSeg ¼ the number of geographic segments; and
STDROA ¼ the standard deviation of firm ROA over the past 12 quarters.

TABLE 5
Reviewer Attribute Analysis
Panel A: Descriptive Statistics on Reviewer Attributes

Mean STD 25th 50th 75th
FullTime 83.80% 20.64% 75.00% 90.91% 100.00%
LongTenure 51.57% 26.41% 33.33% 50.00% 66.67%
HighEducation 20.73% 26.50% 0.00% 11.11% 33.33%
Manager 26.25% 24.13% 0.00% 24.00% 40.00%

Headquarters 47.79% 34.48% 17.07% 46.15% 75.00%
Panel B: Regression Analysis

(1) (2) (3) (4) (5) (6)
Outlook 0.0029*** 0.0027*** 0.0037*** 0.0039*** 0.0030*** 0.0024
(2.89) (3.81) (4.70) (5.48) (4.11) (1.25)
Outlook_FullTime 0.0019***
(2.67)
Outlook_LongTenure 0.0010**
(2.38)
Outlook_HighEducation 0.0006*
(1.85)
Outlook_Manager 0.0001
(0.34)
Outlook_Headquarters 0.0005
(1.46)
Outlook_Factor 0.0006**
(2.04)
Controls Yes Yes Yes Yes Yes Yes
Industry/Time FE Yes Yes Yes Yes Yes Yes
Observations 12,573 18,049 10,331 15,120 13,912 5,450
Adjusted R2 0.435 0.460 0.478 0.456 0.456 0.464
This table examines whether the predictive value of outlook varies with reviewer attributes. We focus on whether an employee works full time (FullTime),
has an employment length of three or more years (LongTenure), holds a master’s degree or higher (HighEducation), holds a manager or above job position
(Manager), and works in the corporate headquarters state (Headquarters). Panel A reports the distribution of the percentage of reviewers with the above
attributes in a firm-quarter. Panel B tests whether outlook provided by reviewers with certain attributes is incremental to outlook provided by all reviewers
in predicting future return on assets. All control variables are the same as in Column (1) of Table 3. Fama-French 48 industry fixed effects and quarter-end
time fixed effects are included. t-statistics are reported in parentheses. Standard errors are clustered by firm.
Outlook_FullTime ¼ the average outlook of reviewers who are full-time employees in a firm-quarter;
Outlook_LongTenure ¼ the average outlook of reviewers who have worked for the company for at least three years;
Outlook_HighEducation ¼ the average outlook of reviewers with a master’s degree or higher;
Outlook_Manager ¼ the average outlook of reviewers with manager or above job positions;
Outlook_Headquarters ¼ the average outlook of reviewers who work in the corporate headquarters states; and
Outlook_Factor ¼ the first factor from factor analysis of Outlook_FullTime, Outlook_LongTenure, Outlook_HighEducation, Outlook_Manager, and
Outlook_Headquarters.

TABLE 6
Job Function Analysis
Panel A: Descriptive Statistics on Reviewer Job Functions

Mean STD 25th 50th 75th
Employee_Sale 10.83% 16.66% 0.00% 0.00% 16.67%
Employee_Production 2.26% 5.77% 0.00% 0.00% 0.85%
Employee_R&D 7.77% 13.49% 0.00% 0.00% 11.76%
Employee_SupplyChain 1.26% 4.58% 0.00% 0.00% 0.00%

Panel B: Job Function Analysis across Different Aspects of Operating Performance
Dependent Variable
AvgSG AvgCOGS AvgR&D AvgInvTurn
(1) (2) (3) (4)
Outlook 0.0650*** 0.0613*** 0.0268*** 0.1518
(14.38) (4.52) (5.09) (1.04)
Outlook_Sale 0.0027*
(1.65)
Outlook_Production 0.0119***
(3.46)
Outlook_R&D 0.0028**
(2.01)
Outlook_SupChain 0.0806**
(2.01)
Controls Yes Yes Yes Yes
Industry/Time FE Yes Yes Yes Yes
Observations 10,653 4,481 8,909 2,668
Adjusted R2 0.281 0.568 0.482 0.867
This table examines whether the predictive value of outlook depends on reviewers’ job function. We assign employees to the following functional areas:
Sales (Employee_Sale), Production (Employee_Production), R&D (Employee_R&D), and Supply Chain (Employee_SupplyChain). Panel A reports the
distribution of the percentage of reviewers in each functional area in a firm-quarter. Panel B tests whether outlook is more predictive of a specific aspect of
firm future performance when it is provided by employees whose job function is more closely related to that performance aspect. Outlook_Sale, Outlook_
Production, Outlook_R&D, and Outlook_SupChain represent the average outlook provided by reviewers in the Sales, Production, R&D, and Supply Chain
functional areas, respectively. AvgSG, AvgCOGS, AvgR&D, and AvgInvTurn are the average sales growth, the average cost of goods sold to sales ratio, the
average R&D expenditure to sales ratio, and the average inventory turnover ratio for the next two quarters, respectively. All control variables are the same
as in Column (1) of Table 3. Fama-French 48 industry fixed effects and quarter-end time fixed effects are included. t-statistics are reported in parentheses.
Standard errors are clustered by firm.
Panel A of Table 6 describes the percentage of individual employee outlook, in a firm-quarter, that each functional area
accounts for. On average, 11 percent of employee outlook in a firm-quarter is by employees from Sales, 2 percent by
Production, 8 percent by R&D, and 1 percent by Supply Chain.26
Panel B of Table 6 presents results from the test of H4. To alleviate multicollinearity concerns, we exclude firm-quarter
observations in which 100 percent of the reviewers possess the same attribute: e.g., when all reviewers are from the Sales area,
or when all are from non-Sales areas.27 As predicted, higher outlook by Sales, R&D, and Supply Chain reviewers predicts
higher sales growth, greater future R&D investment, and higher future inventory turnover ratio, respectively; higher outlook by
26
Because few employees are assigned to Production and Supply Chain, and because the overlap in activities conducted in these two areas is potentially
substantial, we form a single area: Production or Supply Chain. We find that outlook by employees in this area is a stronger predictor of future cost of
goods sold and future inventory turnover ratio.
27
In untabulated tests, we find that our results in Tables 5 and 6 hold when we exclude firm-quarters in which 95 percent or more of the reviewers possess
the same attribute. We note that none of the ordinary least squares (OLS) regressions in Tables 5 and 6 have a variance inflation factor greater than 10.

Production reviewers predicts lower cost of goods sold. In terms of economic significance, a one-standard-deviation increase in
Outlook_Sale, Outlook_R&D, and Outlook_SupChain and a one-standard deviation decrease in Outlook_Production increase
AvgSG, AvgR&D, AvgInvTurn, and AvgCOGS by 1.7, 2.2, 1.8, and 3.9 percent of their sample standard deviations,
respectively. As a reference, the corresponding marginal effects of Outlook are 23.1 percent, 13.3 percent, 1.7 percent, and 10.5
percent.
To assess the concern that these results are driven by chance, we conduct placebo tests. Specifically, we test whether
outlook made by employees from a specific functional area has greater predictive power for performance components not
closely related to that function. To do this, we estimate 12 regressions and match each functional area to three ‘‘wrong’’
performance components. We find insignificant coefficients in all 12 regressions, alleviating the concern that our Table 6
findings are due to chance.

Asymmetric Information Content Analysis
H5 states that employee outlook is more informative about future bad news than good news. To test H5, we first partition
the sample based on the favorability of Outlook: Positive versus Non-Positive and Above Median versus At or Below Median.
We then compare coefficients across subsamples. As expected, in Table 7, Panel A, unfavorable Outlook is the stronger
predictor of future performance: the coefficients on Outlook in the Non-Positive and At or Below Median subsamples are
0.0044 and 0.0048, significantly larger than 0.0017 and 0.0014, the corresponding coefficients in the Positive and Above
Median subsamples.
We then use probit regressions to test whether Outlook is a stronger predictor of bad news events—credit rating
downgrades and dividend decreases—than of good news events, credit rating upgrades, and dividend increases. The event
indicators are defined as follows: CR_Up (CR_Down) equals 1 if the firm experiences a credit rating upgrade (downgrade) over
the next six months, and 0 otherwise; DIV_Incr (DIV_Decr) equals 1 if the firm experiences a dividend increase (decrease) over
the next six months, and 0 otherwise. We include all control variables that appear in Equation (1), as well as the number of
shares sold short in the quarter (ShortSelling), since short selling is a known predictor of bad news events (e.g., Kecskes, Mansi,
and Zhang 2013).
The results, as tabulated in Panel B of Table 7, are consistent with our predictions. Outlook predicts credit rating
downgrades, but not credit rating upgrades. While Outlook does predict both dividend decreases and dividend increases, the
magnitude of the coefficient on Outlook is significantly larger in the dividend decrease prediction regression than in the
dividend increase prediction regression (0.3029 versus 0.1086). In an untabulated analysis, we find that Outlook is a strong
predictor of extreme bad events such as lawsuits, the receipt of adverse going concern opinions, and filings for bankruptcy.
Positive Bias as an Alternative Explanation

Although positive bias works against finding results in our baseline analysis (H1), it works in favor of finding results in our
tests of asymmetric information content (H5). For example, a sample split based on the level of Outlook, favorable versus
unfavorable, is also a split based on the error with which true outlook is measured, large versus small. To address the concern
that the asymmetric information content results are solely driven by positive bias, we examine whether our findings hold when
we only consider predictions by (1) employees who have minimal incentives to be optimistic, and (2) employees who display a
weaker tendency for optimism in their reviews. Accordingly, Outlook_NoCurrentEmp includes predictions by former
employees whose rank-and-file employee stock options are below the sample median, as these employees are likely to have a
smaller ownership stake and, therefore, fewer incentives to be optimistic; and Outlook_NoTopReviews includes predictions by
employees who did not give their employer the highest rating in all eight other review categories.28
Panel A of Table 8 compares the distributions of Outlook_NoCurrentEmp, Outlook_NoTopReviews, and Outlook. The
comparisons reveal that the two new measures are less optimistic than Outlook. For example, the mean (median) of Outlook_
NoCurrentEmp is 0.048 (0) while the mean (median) of Outlook is 0.165 (0.2). This validates our approach to ex ante
identifying less optimistic outlook. Column (1) of Panel B tabulates the coefficient differences that test the asymmetry
hypothesis in Table 7; Columns (2) and (3) tabulate the same differences when Outlook_NoCurrentEmp and Outlook_
NoTopReviews are the predictive variables. We continue to find that employee outlook is a stronger predictor of bad news than
good news. Specifically, all of the coefficient differences in Columns (2) and (3) are statistically significant and not much lower
than those in Column (1): across the eight new specifications, the average percentage decline is 23 percent. We conclude that an
upward bias affects, but does not fully explain, our findings of asymmetric information content.
28
These eight review categories include the overall rating of the company; ratings of senior management, career opportunities, compensation and benefits,
work/life balance, and culture and values; approval of the company CEO; and whether to recommend the employer to a friend.

TABLE 7
Test of Asymmetric Information Content of Outlook
Panel A: Predicting Future Operating Performance: Sample Split on Outlook

Above At or Below
Positive Non-Positive Median Median
(1) (2) (3) (4)
Outlook 0.0017** 0.0044*** 0.0014 0.0048***

(2.29) (3.84) (1.48) (5.27)
ROA 0.6389*** 0.4919*** 0.6577*** 0.4997***
(34.88) (20.82) (34.36) (23.11)
LogSale 0.0006*** 0.001*** 0.0007*** 0.0007***
(3.40) (4.11) (3.96) (3.48)
MTB 0.0000 0.0001*** 0.0000 0.0001***
(0.03) (3.15) (0.05) (2.84)
LEV 0.0018 0.0032* 0.0014 0.0034**
(1.36) (1.93) (0.98) (2.32)
LogAge 0.001*** 0.0002 0.001*** 0.0003
(3.85) (0.71) (3.30) (1.21)
RET 0.0087*** 0.0106*** 0.0081*** 0.0107***
(6.33) (6.01) (5.48) (6.72)
STDRET 0.3274*** 0.4538*** 0.3214*** 0.4506***
(9.32) (12.68) (8.68) (13.73)
TACC 0.058*** 0.0924*** 0.0581*** 0.0853***
(4.42) (7.01) (4.33) (6.77)
Acquisition 0.0015*** 0.0007 0.0018*** 0.0005
(2.87) (0.94) (3.20) (0.79)
AFNews 0.4267*** 0.1385*** 0.4611*** 0.161***
(7.94) (5.42) (8.46) (6.11)
MFNews 0.1314*** 0.1917*** 0.1694*** 0.1526***
(3.00) (3.99) (2.65) (3.62)
InsiderNetBuy 0.0001** 0.0001* 0.0001 0.0002***
(2.47) (1.71) (1.38) (2.71)
OptionExercise 0.0006 0.0015 0.0008 0.0011
(0.68) (1.18) (0.88) (0.99)
ChgRating 0.0004* 0.0002 0.0004* 0.0002
(1.78) (0.59) (1.80) (0.50)
Diff. in jCoeff.j Outlook (2) (1) 0.0027** Outlook (4) (3) 0.0034***
Observations 13,462 8,301 10,519 11,244
Adjusted R2 0.575 0.420 0.603 0.433
Panel B: Predicting Changes in Credit Ratings and Changes in Dividends

Dependent Variable
CR_Up CR_Down DIV_Incr DIV_Decr
(1) (2) (3) (4)
Outlook 0.0671 0.4074*** 0.1086** 0.3029***
(1.09) (6.50) (2.31) (3.77)
ROA 1.9095* 4.2763*** 8.4297*** 2.502
(1.69) (4.21) (7.80) (1.56)
LogSale 0.1371*** 0.2126*** 0.1627*** 0.1716***
(6.93) (10.56) (9.84) (6.35)
(continued on next page)

TABLE 7 (continued)
Dependent Variable
CR_Up CR_Down DIV_Incr DIV_Decr
(1) (2) (3) (4)
MTB 0.0002 0.0045 0.0001 0.0010
(0.09) (1.48) (0.04) (0.20)
LEV 0.8657*** 0.9025*** 0.0319 0.3802**
(6.99) (7.18) (0.26) (1.94)
LogAge 0.0276 0.1478*** 0.1281*** 0.117***
(0.90) (4.39) (4.78) (3.25)

RET 0.5825*** 0.8177*** 0.0199 0.2745
(4.87) (6.76) (0.21) (1.82)
STDRET 0.5289 10.5147*** 43.7682*** 11.7773***
(0.16) (3.89) (12.16) (2.89)
TACC 0.8304 0.2048 1.5626** 1.5224
(1.37) (0.26) (2.55) (1.74)
Acquisition 0.1083 0.1418 0.0708 0.0555
(1.30) (1.55) (1.36) (0.53)
AFNews 2.4287* 4.9628*** 0.8013 0.5537
(1.71) (4.86) (0.54) (0.29)
MFNews 9.426** 7.5028** 4.2485 12.8113***
(2.44) (2.46) (1.52) (3.57)
InsiderNetBuy 0.0100 0.0111** 0.0016 0.0046
(1.58) (2.07) (0.43) (1.02)
OptionExercise 0.4720*** 0.1326 0.2112*** 0.0656
(3.46) (0.87) (3.37) (0.55)
ChgRating 0.0086 0.0615** 0.0341** 0.06
(0.33) (2.22) (2.12) (2.52)
ShortSelling 0.0071 0.0155*** 0.0116** 0.0103
(1.40) (3.13) (2.22) (1.26)
Diff. in jCoeff.j Outlook (2) (1) 0.3403*** Outlook (4) (3) 0.1943*
Observations 21,182 21,028 20,494 19,161
Adjusted R2 0.114 0.198 0.194 0.116
This table examines the asymmetric information content of Outlook. In Panel A, we predict future operating performance in subsamples of positive and
non-positive Outlook, and above-median and at or below the sample median Outlook. In Panel B, we use probit regressions to predict credit rating
upgrades, credit rating downgrades, dividend increases, and dividend decreases. In each case, the dependent variable is coded as 1 when the firm
experiences the event in the next six months, and 0 otherwise. We include Fama-French 48 industry fixed effects and quarter-end time fixed effects, which
results in different numbers of observations being dropped in different specifications. We report t-statistics and z-statistics in parentheses in Panels A and
B, and Chi-square and z-statistics from tests that compare coefficients on Outlook across select specifications in Panels A and B. All test statistics and
significance levels are calculated based on standard errors clustered by firm.
We provide detailed variable definitions in Appendix A.
V. RETURN PREDICTION ANALYSES

Our findings that employee social media disclosures are incrementally informative about future operating performance
raise the natural question of whether and to what extent investors use this information. While investors have strong incentives
to use all available information, they may fail to do so due to limited attention and information-processing capacity, resulting in
predictable market returns (Hong and Stein 1999; Hirshleifer and Teoh 2003; Peng and Xiong 2006).29
We first test whether Outlook predicts future earnings announcement stock returns by estimating the following model:
X X
CARi;tþ1;tþ2 ¼ b0 þ b1 Outlooki;t þ b2 Controlsi;t þ IndustryFE þ TimeFE þ ei;tþ1;tþ2 ; ð2Þ
29
Untabulated results show that Outlook can predict future analyst forecast error for the next two quarters, and that this inefficiency only exists for stocks
with low analyst coverage or small market capitalization.

TABLE 8
Test of Asymmetric Information Content of Outlook
Addressing the Positive Bias Explanation
Panel A: Distribution of Alternative Measures of Outlook

Variable n Mean STD 25th 50th 75th
Outlook 21,763 0.1648 0.4132 0.1034 0.2000 0.4570
Outlook_NoCurrentEmp 10,494 0.0481 0.5444 0.4286 0.0000 0.3333
Outlook_NoTopReviews 21,763 0.1436 0.4139 0.1250 0.1667 0.4286

Panel B: Difference in the Absolute Value of the Slope Coefficients on Outlook
No No
Current Top
Table 7 Employees Reviews
(1) (2) (3)
jNon-Positive Outlookj jPositive Outlookj 0.0027** 0.0020* 0.0025*
(4.17) (2.97) (3.72)
jAt or Below Median Outlookj jAbove Median Outlookj 0.0034*** 0.0020* 0.0034***
(7.30) (2.97) (7.21)
jCredit Rating Downgradej jCredit Rating Upgradej 0.3403*** 0.1235* 0.3107***
(3.91) (1.71) (3.84)
jDividend Decreasej jDividend Increasej 0.1943* 0.1407* 0.1684*
(1.94) (1.89) (1.81)
This table addresses the concern that positive bias in employee outlook explains our Table 7 results. We construct alternative measures of outlook by
excluding reviews that are more likely to be positively biased (i.e., reviews made by current employees and reviews with top ratings in all review
categories). Outlook is the same variable as reported in Table 7. Outlook_NoCurrentEmp includes outlook by former employees who left the company
within the past two years at firms with low (i.e., below-median) rank-and-file employee stock options. Outlook_NoTopReviews is outlook computed based
on a subsample that excludes reviews with top ratings in all review categories. Panel A reports distribution of outlook constructed based on alternative
samples. Panel B reports the differences in the magnitudes of coefficients on outlook in predicting bad news versus good news, by estimating the same
regressions from Table 7 and using Outlook, Outlook_NoCurrentEmp, and Outlook_NoTopReviews, respectively. Chi-square statistics are reported in
parentheses when dependent variables are continuous variables (i.e., future ROA), and z-statistics are reported in parentheses when dependent variables are
indicator variables (i.e., future events). All test statistics and significance levels are calculated based on standard errors clustered by firm.
where CAR is the average of three-day buy-and-hold market-adjusted stock returns surrounding the announcement of earnings
for quarters tþ1 and tþ2, and Controlsi includes all control variables from Equation (1), as well as several possible predictors for
stock returns from prior studies: lagged forecast errors (LaggedAFE), analyst forecast dispersion (AFDispersion), share volume
turnover (Turnover), and advertising (Advertising) and R&D (R&D) expenditures (Chan, Jegadeesh, and Lakonishok 1996;
Huang 2018). We estimate Equation (2) on the full sample, on a subsample of stocks with analyst coverage below the sample
median, and on a subsample of stocks with market capitalization below the sample median. We expect stronger return
predictability for smaller companies and companies with lower analyst coverage because their shares are generally priced less
efficiently (Hong, Lim, and Stein 2000), and because these companies offer investors fewer opportunities to access company
employees.30
Table 9 reports the results. We find that Outlook predicts earnings announcement returns only in the subsamples of small
stocks and stocks with low analyst coverage; in terms of economic significance, a one-standard-deviation increase in Outlook is
associated with a 25 and 19 basis points increase in three-day earnings announcement returns.
We next use a calendar-time hedge portfolio approach to examine the investment value of outlook. Following Huang
(2018), we form hedge portfolios at the end of each month; that is, we sort firms into deciles based on the average outlook in a
30
Valuable information available on Glassdoor in the form of Outlook is likely to overlap, at least partially, with information obtained directly from
employees at corporate access events such as analyst/investor days (Kirk and Markov 2016), which are more prevalent among larger, more heavily
followed firms. Indeed, we find that 10 percent of the firms with market cap above the sample median hold analyst/investor days each year, whereas
only 3 percent of the firms with market cap below the sample median hold analyst/investor days each year (untabulated for brevity). We thank Marcus
Kirk for providing data on analyst/investor days.

TABLE 9
Predicting Future Stock Returns
Earnings Announcement Analysis
Dependent Variable: AvgCAR(1,1)
Low Small
Full Analyst Market
Sample Coverage Cap
(1) (2) (3)
Outlook 0.0009 0.0058*** 0.0045**

(0.84) (3.54) (2.49)
ROA 0.0378* 0.0153 0.0258
(1.84) (0.55) (0.93)
LogSale 0.0000 0.0000 0.0013
(0.09) (0.04) (1.41)
MTB 0.0001 0.0002 0.0001
(1.34) (1.64) (0.58)
LEV 0.0029 0.003 0.0018
(1.18) (0.79) (0.45)
LogAge 0.0015*** 0.0029*** 0.0027***
(2.84) (3.62) (2.93)
RET 0.0003 0.0100** 0.0078*
(0.09) (2.55) (1.96)
STDRET 0.1110* 0.1874** 0.168*
(1.75) (2.15) (1.82)
TACC 0.0138 0.0118 0.0117
(1.08) (0.61) (0.60)
Acquisition 0.0024* 0.0023 0.0009
(1.80) (1.07) (0.41)
AFNews 0.0238** 0.0182* 0.0119
(2.02) (1.69) (0.92)
MFNews 0.016 0.0084 0.0257**
(1.26) (0.52) (2.06)
InsiderNetBuy 0.0000 0.0000 0.0002
(0.36) (0.29) (0.89)
OptionExercise 0.0008 0.0082** 0.0058
(0.37) (2.34) (1.47)
ChgRating 0.0003 0.0008 0.0007
(0.62) (1.16) (0.98)
ShortSelling 0.0001 0.0002 0.0001
(1.26) (1.14) (0.85)
LaggedAFE 0.0195 0.0146 0.0089
(1.52) (1.17) (0.61)
AFDispersion 0.0111 0.0073 0.0063
(1.55) (0.97) (0.67)
Turnover 0.0003 0.0004 0.0003
(0.69) (0.67) (0.35)
Advertising 0.0158 0.0267 0.0227
(1.16) (1.37) (1.26)
R&D 0.0736* 0.1212** 0.0839
(1.93) (2.23) (1.55)
Industry/Time FE Yes Yes Yes
Observations 15,014 7,386 7,551
Adjusted R2 0.012 0.023 0.017

TABLE 9 (continued)
This table presents results from OLS regressions of earnings announcement returns on employee outlook in three samples: the full sample, a subsample
with analyst coverage below the sample median, and a subsample with market capitalization below the sample median. The dependent variable
AvgCAR(1,1) is the average of the three-day market-adjusted stock returns centered on the earnings announcement days of quarter tþ1 and tþ2 earnings.
Outlook is the average outlook posted by employees in quarter t. Fama-French 48 industry fixed effects and quarter-end time fixed effects are included. t-
statistics are reported in parentheses. Standard errors are clustered by firm.

calendar month and track the performance of the decile portfolios over the following three months.31 We construct spread
portfolios that buy stocks in the top decile of Outlook and sell stocks in the bottom decile, and rebalance the portfolios on a
monthly basis by adding stocks that enter the top or the bottom decile and dropping stocks that have reached their holding
periods. We compute equal-weighted and review frequency-weighted returns for each calendar month and regress the monthly
time-series of returns minus the risk-free rate on the three Fama-French and momentum factors (Fama and French 1993;
Carhart 1997).32
Panel A of Table 10 reports the alphas from the monthly calendar-time portfolio regressions. The long-short equal-
weighted portfolio delivers a significant four-factor alpha of 35 basis points per month in the full sample, while the review
frequency-weighted portfolio delivers a positive, but insignificant, alpha of 32 basis points. The alphas generated in the two
subsamples are consistently higher, ranging from 58 basis points (low analyst coverage, equal-weighted returns) to 87 basis
points (low analyst coverage, review frequency-weighted returns). These results suggest that Outlook can be used to predict
risk-adjusted returns for small firms and firms with low analyst coverage.
To explicitly control for other return predictors, we estimate regressions of three-month future market-adjusted returns on
current-month outlook and all Equation (2) control variables.33 We replace each variable with its decile ranks, transformed to
range from 0 to 1, so that the coefficient on each variable represents the incremental hedge return of buying the top decile stocks
and selling the bottom decile stocks.
Panel B of Table 10 reports the results from estimating Fama-McBeth regressions. The coefficient on Outlook is
insignificant in the full sample, but significantly positive in the two subsamples. Similar to the earnings announcement return
results, the insignificant results (untabulated) of hedge portfolio returns for large stocks and stocks with high analyst coverage
could be due to more frequent analyst/investor days in these firms. Of the control variables, only OptionExercise loads
significantly in all three specifications.34 Other known predictors of stock returns, such as past stock returns (RET), accruals
(TACC), insider trades (InsiderNetBuy), and short selling (ShortSelling), do not load significantly in our regressions. We
conclude that employee outlook is a unique, important predictor of future returns for stocks with small market capitalization or
low analyst coverage.
VI. CONCLUSION
In this study, we use employee business outlook on Glassdoor.com to assess whether employees’ social media disclosures
are a distinct source of information about operating performance. We also explore the nature of the information embedded in
these disclosures, and test whether investors use them efficiently.
Controlling for competing sources of information, such as management and analyst forecast news, insider net purchase,
and employee stock option exercises, we find that average employee outlook predicts future operating performance, and that
31
Because stock prices are likely to quickly incorporate value-relevant information, we form portfolios monthly to more precisely capture the return
predictability. The return predictability of outlook becomes weaker when we form portfolios quarterly, consistent with stock markets incorporating the
information embedded in outlook over time. We choose a three-month holding period because return predictability weakens after three months
(untabulated for brevity).
32
The choice to weight stocks based on the number of reviews is motivated by our findings that the information content of outlook is increasing in the
number of reviews.
33
Controls for competing information (AFNews, MFNews, InsiderNetBuy, ChgRating, and ShortSelling) are also measured over the current month.
34
The negative coefficient on OptionExercise is consistent with Huddart and Lang (2003), who show that future stock returns are lower when
option exercise is higher. However, we do not find a negative relation between OptionExercise and firm future operating performance. These
results are consistent with Babenko and Sen (2016), who suggest that the negative returns following high option exercise are likely caused by
stock dilution and who caution against using employee stock option exercises to infer employees’ private information about firm future
performance.

TABLE 10
Predicting Future Stock Returns
Hedge Portfolio Analysis
Panel A: Alphas from Fama-French Four-Factor Model

Low Small
Full Analyst Market
Sample Coverage Cap
(1) (2) (3)
Equal-Weighted 0.0035* 0.0058** 0.0073***

(1.82) (2.09) (3.17)
Review Frequency-Weighted 0.0032 0.0087* 0.0078*
(1.38) (1.69) (1.75)
Panel B: Coefficients from Fama-MacBeth Regressions

Dependent Variable: AbnRET
Low Small
Full Analyst Market
Sample Coverage Cap
(1) (2) (3)
Outlook 0.0028 0.0155** 0.0208**
(0.56) (2.06) (2.26)
ROA 0.0021 0.0179 0.0021
(0.20) (1.60) (0.13)
LogSale 0.0042 0.0052 0.0139
(0.57) (0.53) (1.41)
MTB 0.0176* 0.0033 0.0172
(1.92) (0.28) (1.21)
LEV 0.0085 0.0139 0.0100
(1.57) (0.94) (1.31)
Age 0.0042 0.0102 0.001
(0.64) (1.23) (0.10)
RET 0.0061 0.0136 0.0042
(0.66) (1.40) (0.32)
STDRET 0.0056 0.0217 0.0118
(0.43) (1.44) (0.92)
TACC 0.0051 0.0115* 0.0084
(1.30) (1.78) (1.33)
Acquisition 0.0097 0.0038 0.0127
(1.56) (0.66) (1.23)
AFNews 0.0168*** 0.0082 0.0227**
(3.80) (1.25) (2.64)
MFNews 0.0052 0.0098 0.0137*
(1.40) (1.21) (2.00)
InsiderNetBuy 0.001 0.0018 0.0004
(0.66) (0.48) (0.10)
OptionExercise 0.0101*** 0.0116* 0.0232***
(3.11) (1.90) (3.390)
ChgRating 0.0021 0.0051 0.0112*
(0.89) (1.32) (1.84)
ShortSelling 0.0024 0.0054 0.0125
(0.34) (0.49) (1.07)
LaggedAFE 0.007 0.0293*** 0.0145**
(1.11) (4.68) (2.02)

TABLE 10 (continued)
Dependent Variable: AbnRET
Low Small
Full Analyst Market
Sample Coverage Cap
(1) (2) (3)
AFDispersion 0.0016 0.0018 0.01
(0.23) (0.19) (0.83)
Turnover 0.0076 0.0122 0.0155
(0.89) (1.05) (1.45)

Advertising 0.0024 0.0162 0.0058
(0.29) (1.35) (0.57)
R&D 0.0136 0.0133 0.0297**
(1.41) (1.04) (2.28)
Average Observations 650 286 311
Average Adjusted R2 0.079 0.062 0.067
This table reports return analysis results based on our full sample, a subsample with analyst coverage below the sample median, and a subsample with
market capitalization below the sample median. At the end of each month, we sort sample stocks into decile portfolios based on outlook (i.e., mean
employee assessment over the current month). We then track the performance of these portfolios over the following three months and estimate the hedge
returns of buying the top decile portfolio and selling the bottom decile portfolio. Panel A reports monthly alphas for the hedge portfolio estimated based on
the Fama-French four-factor model. The raw portfolio returns are calculated at the end of each calendar month by equal weighting and review frequency
weighting (i.e., weighting by the number of outlook reviews in the month). Panel B reports the coefficient estimates from Fama-MacBeth regressions of
stock returns on outlook and control variables. We rank each continuous independent variable into deciles and transform it to range from 0 to 1 so that the
regression coefficient represents the incremental return for the hedge portfolio formed on the corresponding return predictor. AbnRET is the three-month
buy-and-hold market-adjusted return following the current month. Outlook, AFNews, MFNews, InsiderNetBuy, ChgRating, and ShortSelling are measured
over the current month. Other controls are measured based on the most recent information as of the end of the current month. Numbers in parentheses are
Newey-West adjusted t-statistics.
this predictive ability is greater when average outlook aggregates disclosures by a larger, more diverse, and more
knowledgeable employee base. In addition, we find that average outlook is a stronger predictor of bad performance than of
good performance, consistent with the idea that because employees hold debtor-like claims on firm value, they have stronger
incentives to anticipate bad performance. Drawing on basic organizational theory, we document systematic differences in the
nature of the information possessed by employees in different functional areas. Specifically, employees who perform different
job functions have superior information about different aspects of operating performance. Finally, we find that investors do not
fully incorporate the information in employee outlook for firms that attract less market attention, resulting in predictable market
returns.
Our findings should be of interest to several audiences: (1) investors who want detailed and disaggregated information and
access to mid-level management and technical staff, but who are not efficiently using the information in employee social media
disclosures; (2) corporate managers who want to control employee social media disclosures and manage the flow of information
to capital markets; and (3) researchers in organizational economics who need empirical measures of how information is
collected, processed, and acted upon in companies (Gibbons and Roberts 2012).
REFERENCES
Adebambo, B., B. Bliss, and A. Kumar. 2016. Geography, diversity, and accuracy of crowdsourced earnings forecasts. Working paper,
University of San Diego.
Babenko, I., and R. Sen. 2016. Do nonexecutive employees have valuable information? Evidence from employee stock purchase plans.
Management Science 62 (7): 1878–1898. https://doi.org/10.1287/mnsc.2015.2226
Bamber, L., J. Jiang, and I. Wang. 2010. What’s my style? The influence of top managers on voluntary corporate financial disclosure. The
Accounting Review 85 (4): 1131–1162. https://doi.org/10.2308/accr.2010.85.4.1131
Bova, F., K. Kolev, J. Thomas, and X. Zhang. 2015. Non-executive employee ownership and corporate risk. The Accounting Review 90
(1): 115–145. https://doi.org/10.2308/accr-50860
Brickley, J., C. Smith, and J. Zimmerman. 2009. Managerial Economics and Organizational Architecture. New York, NY: McGraw-Hill.

Carhart, M. 1997. On persistence in mutual fund performance. Journal of Finance 52 (1): 57–82. https://doi.org/10.1111/j.1540-6261.
1997.tb03808.x
Chan, L., N. Jegadeesh, and J. Lakonishok. 1996. Momentum strategies. Journal of Finance 51 (5): 1681–1713. https://doi.org/10.1111/j.
1540-6261.1996.tb05222.x
Chen, H., M. Kacperczyk, and H. Ortiz-Molina. 2012. Do nonfinancial stakeholders affect the pricing of risky debt? Evidence from
unionized workers. Review of Finance 16 (2): 347–383. https://doi.org/10.1093/rof/rfq028
Chen, H., P. De, Y. Hu, and B. Hwang. 2014. Wisdom of crowds: The value of stock opinions transmitted through social media. Review
of Financial Studies 27 (5): 1367–1403. https://doi.org/10.1093/rfs/hhu001
Cheng, Q., F. Du, X. Wang, and Y. Wang. 2016. Seeing is believing: Analysts’ corporate site visits. Review of Accounting Studies 21 (4):
1245–1286. https://doi.org/10.1007/s11142-016-9368-9
Cheng, Q., F. Du, X. Wang, and Y. Wang. 2019. Do corporate site visits impact stock prices? Contemporary Accounting Research 36 (1):

359–388. https://doi.org/10.1111/1911-3846.12417
Coase, R. 1937. The nature of the firm. Economica 4 (16): 386–405. https://doi.org/10.1111/j.1468-0335.1937.tb00002.x
Core, J., and W. Guay. 2001. Stock option plans for non-executive employees. Journal of Financial Economics 61 (2): 253–287. https://
doi.org/10.1016/S0304-405X(01)00062-9
Cowgill, B., and E. Zitzewitz. 2015. Corporate prediction markets: Evidence from Google, Ford, and Firm X. Review of Economic Studies
82 (4): 1309–1341. https://doi.org/10.1093/restud/rdv014
Demers, J. 2014. How negative online company reviews can impact your business and recruiting. Forbes (September 9).
Dvorak, P. 2008. Best Buy taps ‘‘prediction market.’’ Wall Street Journal (September 16). Available at: https://www.wsj.com/articles/
SB122152452811139909
Fama, E., and K. French. 1993. Common risk factors in the returns on stocks and bonds. Journal of Financial Economics 33 (1): 3–56.
https://doi.org/10.1016/0304-405X(93)90023-5
Fuhrmans, V. 2017. Who’s reading employees’ online reviews? Their CEOs. Wall Street Journal (July 11). Available at: https://www.
wsj.com/articles/whos-reading-employees-online-reviews-their-ceos-1499774400
Gibbons, R., and J. Roberts. 2012. The Handbook of Organizational Economics. Princeton, NJ: The Princeton University Press.
Green, C., R. Huang, Q. Wen, and D. Zhou. 2019. Crowdsourced employer reviews and stock returns. Journal of Financial Economics
134 (1): 236–251. https://doi.org/10.1016/j.jfineco.2019.03.012
Hales, J., J. Moon, and L. Swenson. 2018. A new era of voluntary disclosure? Empirical evidence on how employee postings on social
media relate to future corporate disclosures. Accounting, Organizations and Society 68-69: 88–108. https://doi.org/10.1016/j.aos.
2018.04.004
Hayek, F. 1945. The use of knowledge in society. American Economic Review 35 (4): 519–530.
Hirshleifer, D., and S. H. Teoh. 2003. Limited attention, information disclosure, and financial reporting. Journal of Accounting and
Economics 36 (1/3): 337–386. https://doi.org/10.1016/j.jacceco.2003.10.002
Hirst, E., L. Koonce, and S. Venkataraman. 2008. Management earnings forecasts: A review and framework. Accounting Horizons 22 (3):
315–338. https://doi.org/10.2308/acch.2008.22.3.315
Hong, H., and J. D. Kubik. 2003. Analyzing the analysts: Career concerns and biased earnings forecasts. Journal of Finance 58 (1): 313–
351. https://doi.org/10.1111/1540-6261.00526
Hong, H., and J. Stein. 1999. A unified theory of underreaction, momentum trading, and overreaction in asset markets. Journal of Finance
54 (6): 2143–2184. https://doi.org/10.1111/0022-1082.00184
Hong, H., T. Lim, and J. Stein. 2000. Bad news travels slowly: Size, analyst coverage, and the profitability of momentum strategies.
Journal of Finance 55 (1): 265–295. https://doi.org/10.1111/0022-1082.00206
Huang, J. 2018. The customer knows best: The investment value of consumer opinions. Journal of Financial Economics 128 (1): 164–
182. https://doi.org/10.1016/j.jfineco.2018.02.001
Huddart, S., and M. Lang. 2003. Information distribution within firms: Evidence from stock option exercises. Journal of Accounting and
Economics 34 (1/3): 3–31. https://doi.org/10.1016/S0165-4101(02)00071-X
Jame, R., R. Johnston, S. Markov, and M. Wolfe. 2016. The value of crowdsourced earnings forecasts. Journal of Accounting Research
54 (4): 1077–1110. https://doi.org/10.1111/1475-679X.12121
Ke, B., and Y. Yu. 2006. The effect of issuing biased earnings forecasts on analysts’ access to management and survival. Journal of
Accounting Research 44 (5): 965–999. https://doi.org/10.1111/j.1475-679X.2006.00221.x
Kecskes, A., S. Mansi, and A. Zhang. 2013. Are short sellers informed? Evidence from the bond market. The Accounting Review 88 (2):
611–639. https://doi.org/10.2308/accr-50313
Kirk, M., and S. Markov. 2016. Come on over: Analyst/investor days as a disclosure medium. The Accounting Review 91 (6): 1725–1750.
https://doi.org/10.2308/accr-51418
Kothari, S. P., S. Shu, and P. Wysocki. 2009. Do managers withhold bad news? Journal of Accounting Research 47 (1): 241–276. https://
doi.org/10.1111/j.1475-679X.2008.00318.x
Li, F., M. Minnis, V. Nagar, and M. Rajan. 2014. Knowledge, compensation, and firm value: An empirical analysis of firm
communication. Journal of Accounting and Economics 58 (1): 96–116. https://doi.org/10.1016/j.jacceco.2014.06.003

Marinescu, I., N. Klein, A. Chamberlain, and M. Smart. 2018. Incentives can reduce bias in online reviews. Working paper, National
Bureau of Economic Research.
McGregor, D. 1960. The Human Side of Enterprise. New York, NY: McGraw-Hill.
Peng, L., and W. Xiong. 2006. Investor attention, overconfidence and category learning. Journal of Financial Economics 80 (3): 563–
602. https://doi.org/10.1016/j.jfineco.2005.05.003
Radner, R. 1992. Hierarchy: The economics of managing. Journal of Economic Literature 30 (3): 1382–1415.
Rossi, T. 2010. Have a great day! Investor Relations Update (August): 14–15.
Shiller, R., and J. Pound. 1989. Survey evidence on diffusion of interest and information among investors. Journal of Economic Behavior
and Organization 12 (1): 47–66. https://doi.org/10.1016/0167-2681(89)90076-0
Skinner, D. 1994. Why firms voluntarily disclose bad news. Journal of Accounting Research 32 (1): 38–60. https://doi.org/10.2307/
2491386
Surowiecki, J. 2004. The Wisdom of Crowds: Why the Many Are Smarter Than the Few and How Collective Wisdom Shapes Business,

Economies, Societies and Nations. New York, NY: Doubleday Anchor.
Winkler, R., and A. Fuller. 2019. How companies secretly boost their Glassdoor ratings. Wall Street Journal (January 22). Available at:
https://www.wsj.com/articles/companies-manipulate-glassdoor-by-inflating-rankings-and-pressuring-employees-11548171977
Wolfers, J., and E. Zitzewitz. 2004. Prediction markets. Journal of Economic Perspectives 18 (2): 107–126. https://doi.org/10.1257/
0895330041371321
Zingales, L. 2000. In search of new foundations. Journal of Finance 55 (4): 1623–1653. https://doi.org/10.1111/0022-1082.00262

APPENDIX A
Variable Definitions
Variable Definition
Outlook The average assessment of business outlook over the current fiscal quarter made by current employees and
former employees who left the company within the past two years. We code ‘‘get better’’ as 1, ‘‘stay the
same’’ as 0, and ‘‘get worse’’ as 1. Data source: Glassdoor
AvgROAtþ1,tþ2 Average return on assets for the next two quarters. Return on assets is measured as income before
extraordinary items (IBQ) divided by total assets at the beginning of the quarter (ATQ). Data source:
Compustat
ROA Return on assets for the current quarter. Data source: Compustat

LogSale The natural logarithm of sales (SALEQ). Data source: Compustat
MTB Market-to-book ratio, measured as market value of equity (CSHOQ 3 PRCCQ) divided by the book value of
equity (CEQQ). Data source: Compustat
LEV Leverage ratio, calculated as long-term debt divided by total assets (DLTTQ/ATQ). Data source: Compustat
LogAge The natural logarithm of the number of years the firm has had stock returns in CRSP. Data source: CRSP
RET Cumulative stock return over the fiscal quarter. Data source: CRSP
STDRET The standard deviation of daily stock returns over the fiscal quarter. Data source: CRSP
TACC Total accruals, measured as the difference between earnings (IBQ) and operating cash flows (OANCF
XIDOC), scaled by beginning total assets (IBQ). Data source: Compustat
Acquisition An indicator variable that equals 1 if the acquisition item (AQCY) is greater than 5 percent of the firm’s total
assets (ATQ), and 0 otherwise. Data source: Compustat
AFNews The average of one-quarter-ahead and two-quarter-ahead consensus EPS forecasts minus the most recent
quarterly EPS actual, scaled by stock price at the end of the current quarter. The consensus is the average of
the most recent individual forecasts in the quarter. Data source: I/B/E/S
MFNews The most recent annual EPS forecast for the current year minus the sum of the four most recent quarterly EPS
actuals, scaled by stock price at the end of the current quarter. We use annual management earnings
forecasts because quarterly management forecasts, unlike analyst forecasts, are not consistently available for
most of our sample firms. Data source: I/B/E/S
InsiderNetBuy Abnormal net purchase made by all the insiders in the current fiscal quarter, measured as the net purchase
((purchase sale)/shares outstanding) in the current quarter minus the net purchase in the previous quarter.
Data source: Thomson Reuters
OptionExercise Abnormal stock option exercised by all employees in the most recent fiscal year, measured as the difference
between the option exercised (option exercised/(exercised þ exercisable)) in the two most recent years. Data
source: Compustat
Crowd_Negative The average proportion of negative words over the total number of words in all single-ticker articles published
on Seeking Alpha about the firm over the current fiscal quarter. Data source: Seeking Alpha
Crowd_Positive The average proportion of positive words over the total number of words in all single-ticker articles published
on Seeking Alpha about the firm over the current fiscal quarter. Data source: Seeking Alpha
ChgRating The average overall employer rating in the current quarter minus the average rating in the previous quarter.
Data source: Glassdoor
MVE The market value of equity in billions of dollars. Data source: Compustat
AnalystFollowing The number of analysts following the firm. Data source: I/B/E/S
InstitutionalOwnership The percentage of ownership held by institutional investors. Data source: Thomson Reuters
ShortSelling The average number of shares short sold in the current quarter scaled by the number of shares outstanding.
Data source: Compustat
NumReviewer The number of firm reviewers over the fiscal quarter, ranked into quintiles 0 through 4 and scaled by 4. Data
source: Glassdoor
HI_ReviewerState The sum of the squares of number of reviewers in each state, divided by the square of total number of
reviewers across all states, and then ranked into quintiles 0 through 4 and scaled by 4. Data source:
Glassdoor
HI_ReviewerJob The sum of the squares of number of reviewers with the same job title, divided by the square of total number
of reviewers across all job titles, and then ranked into quintiles 0 through 4 and scaled by 4. Data source:
Glassdoor

APPENDIX A (continued)
Variable Definition
EmpStockOption Stock options owned by rank-and-file employees, measured as total employee stock options minus stock
options owned by top executives scaled by the number of shares outstanding. To control for the
confounding effect of the number of firm employees, we sort firm-quarters into quintiles based on the
number of employees, and within each quintile, we sort firm-quarters into quintiles based on the number of
rank-and-file employee stock options. We then pool firm-quarters with the same employee stock options
quintile rank across different number-of-employee quintile rank together. Data source: Compustat and
Execucomp
NumBusSeg The number of business segments in the firm. Data source: Compustat
NumGeoSeg The number of geographic segments in the firm. Data source: Compustat

STDROA The standard deviation of firm ROA over the past 12 quarters. Data source: Compustat
Outlook_FullTime The average outlook of reviewers who are full-time employees. Data source: Glassdoor
Outlook_LongTenure The average outlook of reviewers who have worked for the company for at least three years. Data source:
Glassdoor
Outlook_HighEducation The average outlook of reviewers who have master’s or higher degrees. Data source: Glassdoor
Outlook_Manager The average outlook of reviewers who have manager or above job positions. Data source: Glassdoor
Outlook_Headquarters The average outlook of all reviewers who work in the headquarters states. Data source: Glassdoor
Outlook_Factor The first factor from factor analysis of Outlook_FullTime, Outlook_LongTenure, Outlook_HighEducation,
Outlook_Manager, and Outlook_Headquarters.
Outlook_Sale The average outlook of reviewers in the Sales function. A reviewer is regarded to be in Sales function if her
job title includes ‘‘sale,’’ ‘‘customer,’’ ‘‘client,’’ ‘‘retail representative,’’ or ‘‘store manager.’’ Data source:
Glassdoor
Outlook_Production The average outlook of reviewers in the Production function. A reviewer is regarded to be in Production
function if her job title includes ‘‘operation,’’ ‘‘product,’’ ‘‘manufacture,’’ ‘‘front end engineer,’’ or ‘‘assembly
line worker.’’ Data source: Glassdoor
Outlook_R&D The average outlook of reviewers in the R&D function. A reviewer is regarded to be in R&D function if her
job title includes ‘‘scientist,’’ ‘‘research,’’ ‘‘R&D,’’ or ‘‘engineer’’ (but not ‘‘front end engineer’’). Data
source: Glassdoor
Outlook_SupChain The average outlook of reviewers in the Supply Chain function. A reviewer is regarded to be in Supply Chain
function if her job title includes ‘‘supply chain,’’ ‘‘purchase,’’ ‘‘distribute,’’ ‘‘logistics manager,’’
‘‘merchandiser,’’ or ‘‘stock clerk.’’ Data source: Glassdoor
AvgSG Average sales growth for the next two quarters. Sales growth is measured as sales (SALEQ) in quarter t minus
sales in quarter t4, scaled by sales in quarter t4. Data source: Compustat
AvgCOGS Average cost of goods sold (COGSQ) to sales (SALEQ) ratio for the next two quarters. Data source:
Compustat
AvgR&D Average R&D expenditure (XRDQ) to sales (SALEQ) ratio for the next two quarters. Data source: Compustat
AvgInvTurn Average inventory turnover ratio for the next two quarters. Inventory turnover ratio is measured as cost of
goods sold (COGSQ) over average inventory balance (INVTQ) for the quarter. Data source: Compustat
CR_Up An indicator variable equal to 1 when a firm experiences a credit rating upgrade over the next two quarters.
Data source: Compustat S&P
CR_Down An indicator variable equal to 1 when a firm experiences a credit rating downgrade over the next two quarters.
Data source: Compustat
DIV_Incr An indicator variable that equals 1 when a firm experiences a dividend increase over the next two quarters,
and 0 otherwise. Data source: Compustat
DIV_Decr An indicator variable that equals 1 when a firm experiences a dividend decrease over the next two quarters,
and 0 otherwise. Data source: Compustat
AvgCAR(-1,1) The average of the three-day market-adjusted stock returns centered on the earnings announcement days of
quarter tþ1 and tþ2 earnings. Data source: CRSP
LaggedAFE Analyst forecast error for quarter t, measured as the actual earnings minus the consensus forecast as of the first
month of quarter t, scaled by stock price at the end of quarter t. Data source: I/B/E/S
AFDispersion Analyst forecast dispersion for quarter t, measured as the standard deviation of earnings forecasts made as of
the first month of quarter t, scaled by stock price at the end of quarter t. Data source: I/B/E/S
Turnover Share turnover, measured as the daily average of trading volume scaled by the number of shares outstanding
for the current fiscal quarter. Data source: CRSP
Advertising Advertising expense, measured as the adverting expense for the most recent year, scaled by total assets. Data
source: Compustat

APPENDIX A (continued)
Variable Definition
R&D Research and development expense, measured as the R&D expenditure for the current quarter, scaled by total
assets. Data source: Compustat
AbnRET The three-month buy-and-hold market-adjusted return following the current month. Data source: CRSP
APPENDIX B

Use of Former Employees in Measuring Outlook
In this Appendix, we justify our decision to include reviews by recently departed employees in constructing Outlook.
We suggest that employees depart from a company with a stock of valuable inside information and often continue to obtain
inside information from their former colleagues. Time, however, depreciates this initial stock of information (as memory fades
and information becomes less timely) and reduces the flow of new information (as former colleagues communicate less and
depart the company). This process likely unfolds over months, if not years, for two reasons. First, information about employee
morale, quality of management, and the profitability of long-term investments is likely to retain its predictive value for some
time. Second, there are significant career benefits to maintaining relationships and exchanging information with employees at
other companies. Anecdotally, employees who leave a company (or university) for a job in the same city/industry stay in touch
with former colleagues for years.
Because not all current employees provide outlook, our inclusion of former employees should yield a more accurate and
representative measure of current employees’ information in the aggregate. It should also reduce bias, since former employees
have weaker incentives to be optimistic than current employees do (see discussion in Section IV).
We obtain information on the last year of employment from Glassdoor and investigate how the predictive ability of former
employees’ outlook decays in time. Specifically, we partition former employees into five groups: those who departed in the
year of the review and those who departed one year, two years, three years, and four or more years before the year of the
review, and regress future operating performance on current employees’ outlook and outlook by former employees with the
same departure year. In Table 11, we find that the incremental information content of former employees’ outlook declines
monotonically from 0.0023 to 0.0014, consistent with the hypothesis that former employees possess inside information. This
information content does not disappear, however, not even for employees who left four or more years ago, consistent with
employees having some public information. Finally, we find that the predictive ability of current employees is stronger than that
of former employees who left four or more years ago and that of employees who left three years ago, but not that of more
recently departed employees. In our analyses, therefore, we construct Outlook by averaging outlook provided by current
employees and former employees who departed no more than two years prior to the year of the review, and include
comprehensive controls for publicly available information pertinent to forecasting future operating performance.

TABLE 11
Predicting Future Operating Performance with Former Employees’ Outlook
(1) (2) (3) (4) (5)
Outlook_Current 0.0030*** 0.0028*** 0.0026*** 0.0032*** 0.0030***
(5.10) (4.99) (4.64) (4.51) (5.29)
Outlook_Former0 0.0023***
(4.47)

(4.57)
(4.21)
(4.61)
(4.25)
Controls Yes Yes Yes Yes Yes
Industry/Time FE Yes Yes Yes Yes Yes
Observations 17,279 15,636 10,668 7,897 10,274
Adjusted R2 0.497 0.498 0.452 0.446 0.414
Difference between 0.0007 0.0006 0.0009 0.0017** 0.0016***
coefficients on the two variables (v2 ¼ 2.50) (v2 ¼ 2.33) (v2 ¼ 2.42) (v2 ¼ 4.75) (v2 ¼ 6.63)
This table reports results from OLS regressions of future return on assets on the average outlook made by current employees (Outlook_Current), and by
employees who left the company in the year of the review (Outlook_Former0), in the year prior to the review (Outlook_Former1), two years prior to the
review (Outlook_Former2), three years prior to the review (Outlook_Former_Recent3), and four or more years prior to the review (Outlook_Former4 ).
Fama-French 48 industry fixed effects and quarter-end time fixed effects are included. t-statistics are reported in parentheses. Standard errors are clustered
by firm.
Control variables are the same as in Column (2) of Table 3, and detailed variable definitions are in Appendix A.

Copyright of Accounting Review is the property of American Accounting Association and its
content may not be copied or emailed to multiple sites or posted to a listserv without the
copyright holder's express written permission. However, users may print, download, or email
articles for individual use.

Employees Social Media

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Employees Social Media

Uploaded by

Copyright:

Available Formats

THE ACCOUNTING REVIEW American Accounting Association

Vol. 95, No. 2 DOI: 10.2308/accr-52519

What Do Employees Know? Evidence from a Social Media

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

The Accounting Review

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

II. BACKGROUND AND HYPOTHESIS DEVELOPMENT

Employees as an Information Source

The Accounting Review

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

Employee Social Media Disclosures

The Accounting Review

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

The Accounting Review

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

III. DATA AND DESCRIPTIVE STATISTICS

The Accounting Review

Panel A: Yearly Distribution

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

Panel B: Industry Distribution

Panel C: Comparison of Sample Industry Composition to Compustat

The Accounting Review

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

The Accounting Review

IV. EMPIRICAL RESULTS

Predicting Future Operating Performance with Employee Outlook

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

Wisdom of Crowds Analysis

The Accounting Review

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

The Accounting Review

Reviewer Attributes Analysis

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

Job Function Analysis

The Accounting Review

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

The Accounting Review

Panel A: Descriptive Statistics on Reviewer Attributes

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

Panel B: Regression Analysis

The Accounting Review

Panel A: Descriptive Statistics on Reviewer Job Functions

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

The Accounting Review

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

Positive Bias as an Alternative Explanation

The Accounting Review

Panel A: Predicting Future Operating Performance: Sample Split on Outlook

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

Panel B: Predicting Changes in Credit Ratings and Changes in Dividends

The Accounting Review

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

V. RETURN PREDICTION ANALYSES

The Accounting Review

Panel A: Distribution of Alternative Measures of Outlook

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

The Accounting Review

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

The Accounting Review

Downloaded from http://meridian.allenpress.com/accounting-review/article-pdf/95/2/199/2435129/accr-52519.pdf by EBSCO, Sarah Scotti on 07 May 2020

The Accounting Review