Fake News in Financial Markets - idc.ac.il · for fake news to inﬁltrate users’ information sets quickly and extensively is becoming a major concern. One prominent example includes

Fake News in Financial Markets

Shimon Kogan∗

MIT Sloan School of ManagementInterdisciplinary Center Herzliya

Tobias J. MoskowitzYale University

School of ManagementNBERAQR

Marina NiessnerYale University

School of Management

March 2018

Preliminary and incomplete. Do not cite without permission.

∗ We thank Tony Cookson, Diego Garcia, James Pennebacker, Eric So, Denis Sosyura as well as confer-ence and seminar participants at UCLA (Anderson), Rice University (Jones), University of Miami BusinessSchool, ASU Sonoran Winter Finance Conference, 3rd Annual News & Finance Conference, University ofColorado at Boulder, Northwestern University (Kellogg), FSU SunTrust Beach Conference, MIT Sloan, YaleSOM, Catolica-Lisbon, University of Kentucky Finance Conference, FEB, for their helpful comments andsuggestions. We also thank Elli Hoffmann and Keren Ben Zvi for providing and helping us with data.Contact emails: [email protected], [email protected], and [email protected]

Fake News in Financial Markets 2

Abstract

While social media platforms, blogs, and other unmonitored media outlets are becominga main source of news for many, they also offer scope for providing misleading or falseinformation. We use two unique datasets and a linguistic algorithm developed to detectdeception in expression, to examine the impact of fake news in financial markets. Thefirst dataset is a set of paid-for articles obtained from an SEC investigation that areknown to be false, that allow us to validate the linguistic algorithm. The second datasetapplies the linguistic algorithm to quantify the probability of an article being fake ona much larger set of articles. We find a strong temporary price impact and subsequentreversals from the fake news articles for small firms, permanent negative price impactfor mid-size firms, and no impact for large firms. In addition, for small and mid-sizefirms we find that around the release of fake articles, managers are more likely to issuepress releases, file 8-K forms with the SEC, and buy stock in their own firm, hintingthat such firms are possibly engaging in stock price manipulation. No such patternsare found for large firms.


1. Introduction

The potential role of new shared information platforms for information transmission is

receiving ever increasing attention. Social media platforms, blogs, and other unmonitored

media outlets are quickly becoming a main source of news for many people.1 While such

platforms can enhance the speed with which information is disseminated and lower the

cost of obtaining information, they also offer scope for providing biased, misleading, or

false information. One prominent example is the proliferation of “fake news," defined as

hoaxes, frauds, or deceptions designed to mislead consumers of information.2 As Alcott

and Gentzkow (2017) argue, social media platforms enable content to be disseminated with

no significant third party filtering or monitoring, allowing false information to be spread

quickly through a vast social network. Vosoughi, Roy, and Aral (2018) find that fake news

diffuses faster, deeper, and more broadly than actual news, in part because the fake news is

often more extreme and exaggerated and designed to increase diffusion. Hence, the potential

for fake news to infiltrate users’ information sets quickly and extensively is becoming a

major concern. One prominent example includes the potential influence fake news had on

the 2016 U.S. Presidential election (Alcott and Gentzkow (2017)). Another example is a

recent study by ReviewMeta (2016) that fake reviews on Amazon are misleading consumers

toward various products (often paid for by the producers of the products) and becoming an

increasing problem. The issue of fake content has become important enough that Amazon,

Google, Facebook, Wikipedia, and many others are currently investing heavily to curb its

dissemination.

The economics of fake news is an interesting area of study that is very recently receiv-

ing attention. Alcott and Gentzkow (2017) for instance, argue that fake news can arise in

equilibrium where it is cheaper to provide than precise signals when consumers cannot cost-

lessly infer accuracy. Extending the model of Gentzkow and Shapiro (2006) on media bias,1According to a survey from the Pew Research Center (Gottfried and Shearer (2016)), 62% of American

adults get news from a social media site.2Facebook defines fake news as “hoaxes shared by spammers" for personal or monetary reasons.


they argue that fake news may generate utility for some consumers (who may enjoy fake

stories or derive utility from content slanted toward their priors or biases as in Mullainathan

and Shleifer (2005)), and where news is deemed higher quality if closer to the consumer’s

priors. When verifying accuracy is costly, this will lead to biased reporting in equilibrium

(Gentzkow, Shapiro, and Stone (2006)), of which fake news can be one aspect. Alcott and

Gentzkow (2017) argue that fake news may maximize short-run gains over long-term rep-

utation. Consumption of fake news occurs in equilibrium because agents cannot verify the

truth costlessly and the news is tailored to match the agent’s priors. Aymanns, Foerster, and

Georg (2017) provide an equilibrium model of an adversary using fake news to target agents

with a biased private signal. The model shows that transmission of fake news is more suc-

cessful when the adversary knows the agent’s signal and information network. Conversely,

when agents are aware of the adversary they discount news appropriately and make fake

content less effective.

There is ongoing debate over the potential costs and importance of fake news. False

content can impose private and public costs by making it more difficult for consumers to

infer the truth. Alcott and Gentzkow (2017) point to the cost of consumers having less-

accurate beliefs as the result of fake news, which could not only make them worse off but also

possibly reduce positive social externalities from these media platforms. As a byproduct,

consumers may also become more skeptical of legitimate news producers if they are hard

to distinguish from fake news producers. Finally, these distortions could cause resource

misallocation, such as misleading or misdirecting consumers toward worse products in the

Amazon example or diverting votes in elections. These potential costs must be balanced

against any welfare benefits from consumers deriving utility from fake news. However, there

is very little evidence documenting these costs, their significance and where and when they

arise.

In this paper, we examine the impact of fake news in financial markets – a setting

where we can readily assess the costs of such news and where arguably there is little benefit


to consumers of such news. Fake financial news, unlike political or social news, should not

provide any utility from an entertainment or bias perspective as in Mullainathan and Shleifer

(2005). The goal is to accurately assess the state of the world for financial gain. On the other

hand, the costs of fake content here are clear in that if fake news causes less accuracy or

erroneous financial decisions, we can directly measure the consequences and financial costs

of those actions to the individual investor, as well as perhaps in aggregate from a resource

misallocation perspective. For example, if fake financial news moves prices, then this would

provide direct evidence of resource misallocation.

Of course, if financial markets are informationally efficient (Fama (1970)), where the cost

of information is close to zero, then fake news should not matter at all and would have no

bearing on prices. Hence, finding a significant price response to fake news suggests three

things: 1) that markets may not be fully efficient, 2) that the cost of information (at least

for that security) may be significant,3 and 3) that fake news may be even more important in

other settings, where competition for information and the ability to exploit it is less fierce.

Our study, therefore, provides a unique test of market efficiency that circumvents the joint

hypothesis problem – fake news under any equilibrium model of pricing should have no effect.

Our experiment is the flip side of the classic event study test (Fama, Fisher, Jensen, and

Roll (1969)): we conduct a “fake-event" study where the price response should be precisely

zero and can measure the magnitude and duration of any price response. Second, while

the marginal cost of information is at the heart of determining how informationally efficient

financial markets are (Grossman and Stiglitz (1980)), attempts to estimate such costs are

empirically elusive. Our study may perhaps provide a glimpse into the cost and speed of

price discovery across firms by examining the impact of fake news events. Finally, given the

competition for information and the ability to trade on that information in financial markets,

if we find significant effects from fake news here, it stands to reason that its impact may3The cost of information can be both a direct cost of gathering, processing, and analyzing information,

as well as the indirect costs of misperceiving or misreacting to information stemming from psychologicalor behavioral biases (e.g., a mental processing cost that can include psychological barriers to interpretinginformation).


be even greater in settings where competition and arbitrage activity are more limited (e.g.,

consumer retail, political elections). Thus, the effects of fake news on financial markets may

provide a lower bound in its potential influence more broadly.

Our study begins with a unique dataset of fake news articles. Rather than try to estimate

information directly and its impact on prices, which is a product of both the cost of informa-

tion and its interpretation and value placed by investors, we examine the impact of fake or

false information on prices, where the price response should be zero under any equilibrium

asset pricing model. Our sample consists of two datasets to estimate the impact of fake news.

The first is a unique dataset of paid-for articles obtained from an industry “whistle-blower”,

Rick Pearson, a regular contributor on Seeking Alpha, a crowd-sourced content service for

financial markets. Mr. Pearson went undercover to investigate other authors on the site and

uncovered fake paid-for articles now being investigated by the SEC. The sample size is small,

but the identity of fake news is clean – 171 articles by 12 authors covering 46 companies.

We compare these to all other articles written by the same authors (334 in total) that were

published on the same platform that may not have been fake – covering 171 stocks in total.

While the first sample represents our cleanest experiment, where there is no ambiguity in

identifying fake articles, it is a small and narrow sample, which may make it more difficult

to draw general conclusions. To address this issue, we also use a second dataset of hand-

collected articles that were published on two (and eventually three) of the most prominent

financial crowd-sourced platforms: Seeking Alpha and Motley Fool covering 203,545 articles

from 2005 to 2015 for Seeking Alpha, and 147,916 articles from 2009 to 2014 for Motley Fool.

Using a linguistic algorithm that scores the authenticity of an author’s text, together with

the first dataset with taged fake news, we probabilistically identify “fake” news. This creates

a second and a much larger set of false news events. Importantly, we use our first and smaller

dataset of known fake articles to validate the algorithm’s ability to identify fake news stories.

Having an unambiguous sample of fake articles from the undercover sting operation by the

SEC is a key advantage, because without it the authenticity algorithm cannot be validated.


Echoing the importance of this statement are the challenges Google, Facebook, and Twitter

are currently facing trying to identify fake news on their own platforms. For example, Google

and Facebook are currently using human editors to evaluate content in the hopes of training

an algorithm to identify false content systematically (Leong (2017)). Absent a set of known

and identifiable fake articles, such endeavors have yielded little success. For the same reason,

an investor at the time of the fake article’s publication could not have constructed such an

algorithm either, since the fake articles from our dataset were not yet known or identified.

Using the set of identifiable fake articles from the SEC to train our algorithm and cross-

validate it, we find our type II error to be very low – less than 5% of articles are identified

as false positives. Hence, our method for identifying fake news in the second dataset is quite

conservative, where we classify only 2.8% of articles as being fake in our sample, with the

frequency peaking in 2008 at 4.8%, but where we have high confidence that these articles

contain false content. Hence, our method is designed to minimize type II errors at the

expense of increasing type I errors, where we are likely missing many other fake articles.

Using the samples of truly fake articles and probabilistically fake articles, we investigate

whether they impact the market and test various theories of fake news production. We

find that the incidence of fake articles is slightly higher for small firms, and that the price

response in markets is much larger for small and mid-size firms and negligible (precisely zero)

for large firms. Small firm prices rise on fake news, which is predominantly positive, by 8%

upon its release, which subsequently gets fully reversed over the course of a year. Hence, for

small firms the market appears fooled by these articles initially, overpricing small firms with

fake news by 8% on average, but then eventually corrects the mispricng. For mid-size firms,

the price impact is negative immediately, and there is a permanent 4% discount associated

with fake articles written about the firm, suggesting that mid-size firms having fake articles

is a bad signal about the firm. For large firms, there is no price impact – initially or long-

term – from fake articles written about it. These results suggest that the market is efficient

with respect to large firms and appears inefficient for small firms, consistent with intuition


suggesting that the cost of information (direct and indirect/psychological) is larger for small

firms. Where the cost of information is lowest and competition is highest, we see no impact

from fake news, and we show that production of fake news in equilibrium is consistent with

these results, as paid-for fake content is done by small firms, but not large firms.

To better understand these results, it is useful to consider what motivates the production

of fake news about firms. One motivation for the fake articles, which is related to how Rick

Pearson went undercover and why the SEC is involved, is that the firms themselves may

be orchestrating a promotional pump-and-dump campaign to manipulate the stock price.

Another possibility, of course, is that rogue authors wish to create a false narrative about a

firm for their own intentions, having no direct connection to the firm itself. To investigate

the first possibility we look at a set of firm actions the firm may be pursuing at the time of

the fake articles’ release. For example, if these articles are part of paid campaigns by firms

orchestrated by a public relations agency, then other actions taken by the firms around these

events are likely to be present. We find that the fake news articles are often accompanied

with press-releases by the firm among small firms, but not among large cap firms. We also

find that 8-k filings are more likely to accompany the news of small firms, but not large firms.

We further find that insider trading in the direction of the fake news (to take advantage of

the price impact) is also more likely for small and mid-cap firms, but not large stocks. These

results are consistent with a deliberate campaign by the firm to manipulate the stock price

and take advantage of the price impact among small firms. For large firms, however, we

do not find the same patterns, consistent with fake news about large firms being driven by

authors outside of or unassociated with the firm.

We explore what characteristics of firms and articles are associated with the propensity

of fake news as well as the magnitude of temporary price impact from the fake news, to test

theories of fake news and to better understand the variation in information environments

across firms and articles. We find that for small and mid-cap firms, high past volatility

and volume are associated with higher propensity of fake news, consistent with more retail


investor attention (see Barber and Odean (2007)). This finding could be consistent with fake

news production catering toward less sophisticated investors who may derive psychological

utility from fake news correlated with their priors (Alcott and Gentkzkow (2017)). Likewise,

article email circulation, a proxy for the popularity of the firm among retail investors, is

positively related to fake news, consistent with fake news being more effective with a broader

network (Aymanns, Foerster, and Georg (2017)).4 In addition, other proxies for attention

and strength of network, such as larger analyst coverage, larger number of stock tweets

mentioning the firm, more media coverage, and more readers’ comments, are all associated

with higher propensity of fake news. We also examine other actions taken by the firm and

its insiders prior to the fake news event date, such as share purchases, insider sales, and

initiation of press releases. These actions are more prevalent and coincide with the fake

news for small and mid-cap firms, but not for large firms, consistent with smaller firms

engaging in short-term profit maximizing schemes from fake news.

We also find that these variables are associated with the magnitude of the positive tem-

porary price impact found for small firms, and the permanent negative price impact found

for mid-cap firms. The price impact for large firms is non-existent and does not vary with

any of these variables. Specifically, for small firms, we find that the price reaction is stronger

for firms with higher turnover and volatility, less media coverage, more retail ownership,

and more frequently talked about on StockTweets. These characteristics predict both the

magnitude of the initial price rise as well as the size of the subsequent reversal. We also find

that firms whose articles usually get a lot of comments have a quicker return reaction than

firms with fewer comments. Similarly, if the author of the article has many followers, or if

many readers subscribe to that firm’s articles, the return reaction is much stronger for those

firms. Finally, the amount of trading by insiders and the number of press releases issued

by the firm affect the differential return reaction, where firms whose management engages

in insider buying and press releases around the time of the article have a stronger return4We have data on email circulation for Seeking Alpha articles only.


reaction.

For mid-size firms, these same characteristics also predict the magnitude of the price re-

action. However, the same characteristics associated with small firms having a more positive

return reaction are associated with mid-size firm’s more negative permanent price reaction.

These results suggest that both small and mid-cap firms may be engaging in stock manipula-

tion strategies to pump up the share price and take advantage of the higher price by issuing

shares or buying shares before the run-up. In the case of small cap stocks, the market ap-

pears to be fooled by this scheme, causing a temporary price increase that subsequently gets

reversed. For mid-cap stocks, however, one interpretation is that the market does not get

fooled because markets are more efficient for these larger stocks (e.g., the cost of information

for these stocks is lower). Instead, the market identifies the news as fake and reacts immedi-

ately to the fake news in a negative way by permanently discounting the firms’s share price.

This evidence is consistent with the reputational story of Alcott and Gentzkow (2017) where

fake news producers may sacrifice longer-term reputational capital in lieu of short-term cap-

ital gain. In the case of mid-cap stocks, the market deciphers the fake news and punishes

the firm with a permanent price discount.5

For large cap firms, they neither seem to be initiating or taking actions in conjunction

with the fake news articles and there appears to be no market price reaction of any kind

to fake articles written about large firms. These results are consistent with the market

being efficient with respect to fake news about large firms and as a result, large firms not

attempting to engage in a futile effort to manipulate the share price. Rather, it appears that

the fake articles about large firms are being written by rogue authors unaffiliated with the

firms in an effort to increase their own utility.

Our study provides evidence on the prevalence and effect of fake articles from crowd-5Of course, it’s also possible that mid-cap firms pumping scheme actually works if the returns would have

been even worse had they not initiated the fake articles. Hence, another interpretation is that the mid-capfirms fool the market, too, but only do so when other bad news about the firm is present. This narrative isless consistent with the data, however, since we find no evidence of other bad news associated with mid-sizefirms around the time of the articles.


sourced financial platforms that continue to grow and gain attention. How important are

these platforms and what impact are they having, and specifically, how pervasive and im-

portant are fake articles on these platforms? While these social news networks may simply

be a side-show for financial markets, our results indicate that there is significant price im-

pact from these fake articles. The results are consistent with other findings suggesting that

crowd-sourced services can impact markets (Hu, Chen, De, and Hwang (2014), who show

that information content and sentiment from these services predicts returns) and suggests

that if fake content can impact financial markets in the U.S., it is likely to have an even

greater influence in settings where the cost of information is higher and the ability to exploit

it is more limited, such as on-line consumer retail reviews, crowd-funding, elections, news

cites, etc.).

The rest of the paper is organized as follows. Section 2 details our sample on fake news

articles and presents our methodology for identifying fake news. Section 4 examines the

market’s response to fake news, including both temporary and permanent price responses,

as well as the drivers of differential market responses across different types of firms. Section

5 then examines whether managers are more likely to issue press releases and file 8-K forms

around publications of fake articles, and also whether they engage in insider trading. Section

6 concludes the paper.

2. Data and Identifying Fake News

We describe the data we obtain on fake articles, our algorithm and its validation, and

provide an example of a specific fake article and its consequences.

2.1. Obtaining Fake Articles

The popularity of financial crowd-sourced platforms such as Seeking Alpha, Motley Fool,

TheStreet, etc., has grown exponentially over the last fifteen years. For example, Seeking

Alpha went from having two million unique monthly visitors in 2011 to over nine million

by 2014. While this innovation has allowed for unprecedented levels of ‘democratization’


of financial information production, some concerns have been raised about these platforms

being susceptible to pump-and-dump schemes, as they are frequented by retail investors, and

since many authors on these platforms use pseudonyms instead of writing under their real

name.6 Whereas it is theoretically legal for an author to talk up or down a stock that she

is either long or short, she has to disclose any positions she has in the stock in a disclaimer

that accompanies the article. While many authors add such disclaimers to their articles,

that is not something the platforms actually verify. What is illegal, according to Section

17b of the securities code, is to fail to disclose any direct or indirect compensation that the

author received from the company, a broker dealer, or from an underwriter7.

Even though it does not come as a surprise that such promotions and pump-and-dump

schemes exist, they can be hard to identify. Even harder is to prove that the authors of

such articles received payment for writing them. Our analysis starts out with a unique

dataset of paid-for articles obtained from an industry “insider," Rick Pearson. Rick, who is

a regular contributor to Seeking Alpha, was approached by a PR firm that helps promote

stocks. The PR firm asked him to write articles for a fee without disclosing the payment.

Rick went undercover to investigate other authors and has uncovered many fake, paid-for

articles where the authors did not disclose their compensation. These fake articles were

subsequently taken down by the platforms where they originally appeared and the firms are

now being investigated by the SEC. The SEC had since filed two lawsuits: first in 2014

and subsequently in 2017 against authors of fake articles and the PR firms who were paying

the authors to generate those articles.8 Rick has kindly shared with us the articles that he

has determined to be fake, providing us with 111 fake articles by 12 authors covering 46

companies. We also were able to obtain a second set of for-sure fake articles. During the

investigation, the SEC lawyers were able to identify further articles that were paid for by6Even though the platforms claim that they always know the true identity of the author, in case that

information is subpoenaed by the SEC.7In June 2012, Seeking Alpha announced it would no longer permit publication of articles for which

compensation had been paid.8http://securities.stanford.edu/filings-documents/1051/GBI00_01/20141031_r01c_14CV00367.pdf and

https://www.sec.gov/litigation/complaints/2017/comp23802-lidingo.pdf


the stock promotion firms.9. We contacted Seeking Alpha, and they kindly shared 147 of

those articles with us. Of those, we were able to match 60 with CRSP dataset (the rest of

the articles were about firms that traded over the counter). So our final dataset of for-sure

fake articles consists of 171 articles written by 20 authors about 47 firms.

We furthermore were able to obtain all other articles written by the same authors (334

in total) that were published on Seeking Alpha, many of which presumably not paid-for

fake articles, as a baseline comparison for the same authors. These other articles were often

written about large firms (171 stocks in total), which as we will show, are less likely to engage

in this sort of stock promotion. Furthermore, authors need to establish a reputation writing

non-fake articles before they can write (and get away with) pump-and-dump type articles.

2.2. Example of a Fake Article

To illustrate the process and the impact of fake articles, we provide an example of Galena

Biopharma Inc., one of the companies that hired a PR firm to order paid-for fake articles

about its stock. Several very positive articles about Galena appeared on Seeking Alpha

and other platforms from 2012 to 2014, which coincided with a substantial run-up and a

subsequent drop in Galena’s stock price. Figure 1 shows the price of Galena’s stock from

2012 to 2015 in light blue, and the appearances of fake articles in red. Over that time period

six identified fake articles appeared about Galena, with four of them published towards the

end of 2013 and early 2014. The four fake articles were all written by the same author, John

Mylant. John Mylant had been an active contributor to Seeking Alpha since 2009, even

though since this incident all of his articles have been taken down by Seeking Alpha. As the

graph shows, Galena’s stock price started to increase drastically, when several fake articles

were published, more than tripling in 4 months, before it plummeted back down, once the

fake articles stopped.

One natural question that follows is why the companies pay for these fake articles. In9The full list can be found here: https://ftalphaville-cdn.ft.com/wp-

content/uploads/2017/04/10231526/Stock-promoters.pdf


Figure 1 in dark blue we plot all instances of insider trading between 2012 and 2015 (an

indicator variable for whether any insider buys/sells were reported to the SEC through

Form 4). The graph shows that insiders executed trades after the stock price almost tripled

and right before the stock price crashed again. The SEC brought charges against Galena and

its former CEO Mark Ahn “regarding the commissioning of internet publications by outside

fake firms.” Mr. Ahn was fired in August 2014 over the controversy, and in December 2016,

the SEC, Galena, and Mr. Ahn reached a settlement. The example of a fake article about

Galena is shown in Appendix A, and the 8-K form documenting the settlement in presented

in Appendix B. Interestingly, if one were to search for this fake article now, Seeking Alpha

just displays a message saying “This author’s articles have been removed from Seeking Alpha

due to a Terms of Use violation.”

2.3. Further Identifying Fake Articles – LIWC and the Authenticity Score

While the unique data of fake articles is illustrative of the phenomenon, the sample size

is small and it is difficult to draw more general conclusions based on it. The goal of our

paper is to estimate the prevalence and the effects of these fake articles on financial markets.

In order to do so, we manually download all articles that were published on two of the more

prominent financial crowd-sourced platforms: Seeking Alpha and Motley Fool. For Seeking

Alpha we obtained 203,545 articles dating from 2005 to 2015 and for Motley Fool we have

147,916 articles dating from 2009 to 2014.

To understand how pervasive the phenomenon is in general and what effect fake articles

have on financial markets we develop an objective and scalable measure that captures the au-

thenticity of the article. To that end, we use LIWC2015 (Pennebaker et al. (2015)). LIWC is

a linguistic tool that focuses on individuals’ writing or speech style, rather than content, and

thus appears to be uniquely adept at measuring individuals’ cognitive and emotional states

across domains. Specific to authenticity, Newman et al. (2003), use an experimental setting

to develop an authenticity score based on expression style components. While the exact for-

mula for the authenticity score is proprietary, James Pennebaker describes which linguistic


traits are associated with honesty in his book "The Secret Life of Pronouns" Pennebaker

(2011). In particular he finds that truth-tellers tend to use, for example, more self-reference

words and communicate through longer sentences compared to liars. Intuitively, when peo-

ple lie, they tend to distance themselves from the story by using fewer “I" or “me"-words.

Furthermore, people use fewer insight words such as realize, understand, and think, and in-

clude less information about time and space. On the other hand, liars tend to use more

discrepancy verbs, like could, that assert that an event might have occurred, but possibly

did not. James Pennebaker and co-authors then use a combination of these linguistic traits

to generate the authenticity measure in LIWC.

2.4. Validation

Given that the LIWC authenticity score was not developed in the context of financial

media, one may be skeptical about its ability to distinguish fake from non-fake articles in

Finance. After all, financial blogs and articles tend to point to facts, trends, and figures,

which are different from narratives. To address this concern, we start out by validating

the LIWC authenticity score using the small sample of 171 fake articles and 334 non-fake

articles, all written by the same set of authors. That is, we compare the LIWC authenticity

score, which is normalized between 0 and 100, for the two samples. The difference in the

LIWC authenticity score across the two samples is both economically and statistically large.

Relative to an average authenticity score of 33 for non-fake articles, fake articles had a much

lower average score of 19 (statistically significant at 1% level). The density plots in Figure

2, Panel A illustrate how different the two distributions are. It is important to note that we

control for any differences the authors’ writing style may have on the authenticity score, as

the sample consists of both fake and non-fake articles written by the same set of authors.

To provide more specific examples, in Panel B, we provide the density plots of authenticity

scores for two specific authors: John Mylant, and Equity Options Guru. We can see that

while some of the non-fake articles also had a low authenticity score, most of the fake articles

had a very low authenticity score.


While the exact composition of the authenticity score is proprietary, several language

characteristics are associated with being more or less authentic (described in Pennebaker

(2011)). In Table 1 we provide a summary of those characteristics for the for-sure fake and

non-fake articles (written by the same authors). For-sure Fake articles are articles that

have been shared with us by Rick Pearson and that were subpoenaed by the SEC. Non-fake

articles, are articles that were written by the same authors, but about larger firms, that are

unlikely to be fake. We display the number of articles in each category as well as the mean of

the Authenticity measure from LIWC. From the table, we see that the for-sure fake articles’

authenticity score is about half the size of non-fake articles. We also report the means of

several other variables provided by LIWC to help better understand the authenticity score. In

particular we display the means of the average of the 1st person singular measure (examples:

I, me, mine), Insight measure (examples: think, know), Relativity measure (examples: area,

bend, exit), Time measure (examples: end, until, season), Discrepancy measure (examples:

should, would), and the average number of words per sentence. According to research by

James Pennebacker and co-authors, when people lie they tend to use fewer self-referencing

words, fewer words per sentence, fewer insight and relativity words, and more discrepancy

verbs. The results in the table line up well with those findings: fake articles’ self-referencing

score is about half of non-fake articles, and fake articles have a lower insight, and relativity

scores, and higher discrepancy score. It’s important to note that the for-sure fake articles

that we obtain from Rick Pearson and from Seeking Alpha (that have been subpoenaed by

the SEC) are crucial to being able to use LIWC to identify fake articles in Finance, as they

provide an out-of-sample test for a methodology that was developed outside of finance.

2.5. Probability of Being Fake

The above validation demonstrates that the LIWC authenticity score has the ability to

distinguish between fake and non-fake articles. At the same time, it is not clear how to

interpret the cardinal nature of the score – what does a 14 point difference in authenticity

score mean? Ideally, we would measure the probability of an article being fake, but the


LIWC authenticity score is not a probability. To provide a more direct interpretation of the

results and their economic meaning, we develop a mapping of the authenticity score into the

probability space. Starting with the validation sample and applying Bayes rule to the overall

sample of Seeking Alpha articles, we create a function that maps the authenticity score into

a conditional probability of an article being fake.

Specifically, let S be the authenticity score and F (T ) denote a fake (true) article. The key

to our ability to tranform the ordinal authenticity score into a cardinal probability measure

is the fact that in the validation sample, we know which articles are F and which ones are

T . Thus, we can compute Prob(S|F ) and Prob(S|T ). From Bayes rule, we know that:

Prob(F |S) = Prob(S|F )Prob(F )Prob(S|F )Prob(F ) + Prob(S|T )Prob(T )

.

If we integrate Prob(F |S) over the empirical distribution of scores, we get Prob(F ). The

issue, of course, is that Prob(F ) is also an input in the calculation. The solution to the

fixed point problem can be found assuming that Prob(F ) in the sample is representative of

Prob(F ) in the overall population.

We apply this approach to the entire sample of Seeking Alpha articles published between

2005 and 2015, over 203,000 in total, covering over 7,700 firms. There are a number of

findings that arise. To start with, we observe the resulting mapping of LIWC authenticity

scores (S) into the conditional probability of being fake (Prob(F |S)). As Figure 3 depicts,

the relation between the two is highly non-linear. As the figure shows, an authenticity score

of 31 – the average for the non-fake articles – corresponds to a conditional probability of

being fake of close to zero, while an authenticity score of 17 – the average for the fake articles

– corresponds to a significant probability of being fake of 3.6%.

This has two important implications. The first is that using the LIWC authenticity score

is not equivalent to using the probability. The second is that the sharp increase in probability

in the very low authenticity range suggests that articles can be well classified into fake and


non-fake ones. Put differently, using a probability cutoff can be an efficient way of separating

articles into various types.

Another validation exercise that we perform is looking at articles by a Motley Fool author

Seth Jayson. Seth has been working for Motley Full time since 2004, and has written over

31,000 articles. Given that he works for Motley Fool, he is unlikely to write fake articles.

Mostly because promotional firms are probably less likely to approach him. Using our

method, from the articles that we classify, we classify 18,361 as non-fake and only 2 of his

articles as fake. This suggests that our algorithm works quite well, given that the number

of his articles that are fake are most likely close to zero.

Next, we use these results to answer a key question: how pervasive are fake articles on

financial crowd-sourced platforms? We find that the unconditional probability of a Seeking

Alpha article being fake is 2.8%, peaking at 4.8% in 2008 and dropping to 1.6% in 2013.

We next examine how accurate our method is at identifying fake news. We take the 505

articles (171 for-sure fake and 334 non-fake articles written by the same authors), generate

an authenticity score for them, and calculate their probability of being fake. We then use the

cutoff of Prob(Fake) > 20% to classify articles as being fake10. Our algorithm classifies 18

out of the 505 articles as being fake. Out of those 18 articles 17 are actual fake articles. This

suggests that our Type II error rate is very low - we have very few false positives, and our

method is very conservative. In other words, while we most likely miss some fake articles,

the ones our algorithm identifies as fake are highly likely to be truly fake articles.

We classify articles with Prob(Fake) < 1% as being non-fake. Our algorithm identifies

165 articles (out of 505) are being non-fake. Of those 14 are actually fake articles, and the

rest are not. Therefore, our Type I error is about 5%, which is quite low. So when we

look at articles that our algorithm identifies as being “non-fake," most of them happen to be

non-fake. We exclude articles with 1% ≤ Prob(Fake) ≤ 20% from our analysis, as for those

articles Type I and Type II errors will be large, and would make our analysis noisy.10Our results are not sensitive to the specific cutoffs


Table 2 shows, for different types of articles on Seeking Alpha and Motley Fool, summary

statistics of various LIWC textual measures, the probabilities of being fake, and firm charac-

teristics of the covered firms. For-sure Fake Articles are the articles that have been shared

with us by Rick Pearson, who went undercover to expose authors who were being paid to

write fake articles for companies, without disclosing the payments, and also articles that were

subpoenaed by the SEC and shared with us by Seeking Alpha. Seeking Alpha Articles and

Motley Fool Articles are regular articles that we downloaded from Seeking Alpha and Motley

Fool. Of those articles, Fake articles are articles whose probability of being fake (according

to our measure) is higher than 20%, Non Fake articles are articles with probability of being

fake less than 1%, and the rest are classified as Other.

In Panel A, we display the number of articles in each category as well as the mean of

the Authenticity measure that we use to construct the probabilities of being fake. In all

instances, the authenticity score is much lower for fake articles, than for non-fake articles in

both Seeking Alpha and Motley Fool datasets. The differences are statistically significant.

The authenticity scores for Fake Seeking Alpha and Motley Fool articles are especially low,

which is by construction.

We also report the means of several other variables provided by LIWC to help better un-

derstand the authenticity score. In particular we display the means of the average of the 1st

person singular measure (examples: I, me, mine), Insight measure (examples: think, know),

Relativity measure (examples: area, bend, exit), Time measure (examples: end, until, sea-

son), Discrepancy measure (examples: should, would), and the average number of words

per sentence. While LIWC doesn’t disclose details about their authenticity measure, Pen-

nebaker (2011) suggests that these variables are important linguistic differentiators between

honest and deceptive reports. Fake articles, as identified by our algorithm, have a much

lower fraction of 1st person singular then the non-fake articles, suggesting that the authors

try to distance themselves from the article. The fake articles also have a lower insight and


relativity scores relative to non-fake articles, suggesting that the authors seem to draw fewer

insights in the articles, and also reference time and space less.

In Panel B, we display the average probability of being fake for each of the article cate-

gories. For Seeking Alpha and Motley Fool, the difference in magnitudes of the probability

of being fake are by construction.

In Panel C, we display the average fraction of retail investors, the average number of

analysts covering the firm, and the average firm size (in Millions of dollars) for the respective

article groups. For-sure fake articles tend to cover firms with a higher fraction of retail

investors, whereas the fake versus non-fake articles seem to target firms with similar fraction

of retail investors. Similarly, for-sure fake articles tend to concentrate on smaller firms with

low analyst coverage, which is not the case for fake and non-fake articles.

Finally, we examine whether fake articles tend to cluster in specific industries. In par-

ticular, in Table 3, we separate articles by the 12 Fama-French industries that the firms

the articles are about belong to. Some interesting patterns emerge. For articles provided

to us by Rick Pearson, and articles that were subpoenaed by the SEC and shared with us

by Seeking Alpha, 81% of for-sure fake articles are about firms in the Healthcare industry.

This finding is not too surprising as these articles came from authors who were hired by

two PR firms, and concentrated on the healthcare industry. For the non-fake articles, the

majority of firms belong to Business Equipment, Healthcare, Finance , and Manufacturing

industries. The distribution of Fake versus Non-Fake articles on Seeking Alpha and Motley

Fool is similar to the Non-Fake articles in column (2), with majority of them coming from

Business Equipment, Finance, and Healthcare.

2.6. Other Datasets

We investigate the motivation behind these fake articles, where one hypothesis is that

these campaigns are ordered by firms and orchestrated by a PR agency. To test this hypoth-

esis, we obtain a dataset of press releases from RavenPack from 2001 to 2015 and collect

message volume for a given firm from a Twitter-like platform called StockTwits. If firms are


coordinating these articles, they may issue press releases simultaneously to provide material

for the fake articles and may also start Twitter campaigns to reinforce the messages in the

fake articles.

We obtain stock price data from CRSP and firms’ financial information from COMPU-

STAT. We also obtain data on insider trades from Form 4 from Thomson Reuters to see if

insiders are trading around these events. Finally, we obtain dates of SEOs from the SDC

Platinum database to look at firm equity issuance around these events.

3. 2014 SEC Lawsuit and Impact on the Market

There are several ways fake news can be costly to society. There are potential costs

directly to individuals who believe the fake news and act upon them. However, it can also be

costly to society if people stop placing as much trust in news that are actually non-fake. While

fake news are not unique to financial markets, in other areas it is unclear how to estimate

the potential spill-over effect of presence of fake news on people’s trust in non-fake news. In

our paper, we have a unique setting, where the presence of fake news on knowledge sharing

platforms became suddenly very salient through a set of exposé articles, and eventually a

lawsuit by the SEC, about companies hiring PR firms to place fake articles on websites like

Seeking Alpha. Furthermore, in financial markets, there are outcome measures we can use

to gage whether people’s trust in real news has been affected by the scandal. In this section

we examine whether the increased salience of the presence of fake news had an effect on how

much investors reacted to non-fake news. We start by documenting the pump-and-dump

scheme and it’s unraveling, and then examine how that has affected the impact of Seeking

Alpha and Motley Fool articles on abnormal trading volume.

3.1. Case Study: Galena Biopharma Inc.

Galena Biopharma Inc. was the first company that was prosecuted by the SEC for

manipulating it’s stock price on knowledge sharing platforms like Seeking Alpha and Motley

Fool. On 31 October, 2014 the SEC filed a lawsuit in the United States District Court (Case


3:14-cv-00558-SI)11 on behalf of all persons who obtained Gala’s common stock between

August 6, 2013 and May 14, 2014.

The timeline of the events (summarized from the lawsuit document) is presented in Figure

1. The figure depicts the stock price of Galena from April 2013 - May 2014, as well as the

events that lead to the lawsuit. According to the lawsuit, Galena worked with PR companies

Lidingo and DreamTeam to publish a series of promotional articles on third-party websites,

like Seeking Alpha, that Galena ordered. The articles didn’t disclose the payments that the

authors received, and in some cases claiming not to have received any payment. The lawsuit

documents at least twelve of these promotional articles. As can be seen in Figure 1, the

scheme worked and Galena’s price rose from about $2 to $7.48 between the summer of 2013

and January of 2014. In the meantime Galena insiders used the price-rise to help the firm

and themselves. On September 18, 2013 in an SEO Galena sold 17,500,000 units of stock for

net proceeds to Galena of $32.6 million. Then, on November 22, 2013, Galena held a board

meeting and granted stock options to executives and directors with a strike price of $3.88.

Starting January 17, 2014, after the stock price reached it’s highest level since 2010, seven

Galena insiders sold most of their stock within less than a month, for a combined total for

more than $16 million.

As the news of these insider sales broke, the stock price started to decline. Furthermore,

in February and early March, 2014, several investigative journalists, including Matt Gravitt,

Adam Feuerstein, and Richard Pearson, started publishing exposé articles on Seeking Alpha

documenting the fraud. While these journalists were uncovering events that were linked

directly to Galena, in February and March, 2014, several articles appeared on Barron’s and

in Fortune, discussing Seeking Alpha’s anonymous contributors policy and the fact that

Seeking Alpha, Wall St. Cheat Sheet, and other websites it had to pull over 100 articles that

have been used in stock promotion schemes. Finally, on March 17, 2014 Galena revealed in a

10-K filing that it was the target of an SEC investigation over the promotion. By this point11http://securities.stanford.edu/filings-documents/1051/GBI00_01/20141031_r01c_14CV00367.pdf


it’s stock prices has dropped back down to almost $2.

It is important to know that while Galena is a small firm, it had still visibility. In July

2013, before the promotion started, it’s market cap was approximately $350 million, and it

was followed by analysts at Cantor Fitzgerald, JMP Securities, Oppenheimer & Co, among

others. Furthermore, according to the SEC lawsuit, more than a hundred market makers

made market in the company’s stock.

3.2. Articles and Impact on Abnormal Volume

To examine the impact that articles on knowledge sharing platforms have on financial

market, we examine abnormal trading volume following the publication of articles before

and after February and March, 2014. We concentrate on volume and not on other measures

of financial markets, because we are interested in whether investors who read these articles

’trust’ the information in these articles. One way to proxy for whether investors trust that

information is to examine whether they act on the information provided in these articles,

which would be directly reflected in trading volume. Whereas, examining the articles’ effect

on volatility, for example, would get more at whether there is information in the articles

(which we examine later), but not necessarily whether and how many investors act on this

information.

First, we examine whether an article has any effect on trading volume. We define ab-

normal trading volume, as V ol(t)/V ol(t − 251, t − 1). We then sum the abnormal volume

over days t = 0, t + 1, and t + 2, and take the natural log of the sum. We regress this

abnormal volume on whether there was an article on a given day. We also include firm and

year-month fixed effects. We only examine firms that had at least one article published on

Seeking Alpha or Motley Fool about them. The results are presented in Table 4, Panel A. In

the first column, we see that an article on Seeking Alpha or Motley Fool is associated with

a 12% increase in trading volume on the three days following the article. Next, we examine

separately whether the article was written about a small, medium, or a large firm. Small

firms are defined as firms in the bottom 10th percentile of NYSE firms, mid-size firms are


defined as firms in the 20th-90th percentile of NYSE firms, and Large firms are defined as

the top 10th percentile of the NYSE firms. We can see that the effect of articles on abnormal

trading volume for small firms is six times larger than for large firms. This is consistent with

an idea, that small firms tend to have more retail investors and a more opaque information

environment.

Next, we only concentrate on days with articles that we either classify as fake or as

non-fake. We examine what author and company characteristics affect the impact of fake,

relative to non-fake articles. The results are presented in Table 4, Panel B. In the first

column we show that a fake article is associated with a abnormal volume increase that is

3.4% higher than non-fake articles. In the next two columns we examine whether the effect

of fake articles on trading volume differs by firm size or the fraction of retail ownership. We

find that authors’ influence and firm size don’t seem to affect the differential effect of fake

articles on trading volume relative to non-fake articles. However, fake articles have a much

stronger effect on trading volume, relative to non-fake articles for firms with higher retail

ownership.

So far we have shown that Seeking Alpha and Motley Fool articles have a strong effect

on investors’ attention and affect their trading behavior. Next, we concentrate on the time

period around the scandal that broke in February and March, 2014 as an exogenous shock to

people’s awareness of the presence of fake new in knowledge sharing platforms. In particular,

we concentrate on 6 months prior and 6 months after the scandal, to examine whether the

prevalence of fake news, and the effect of news on abnormal trading volume has changed.

The analysis is presented in Table 5. We define the ‘before’ period as August 2013 to January

2014, and the ‘post’ period as April - September, 2014. We exclude February and March,

2014, as that is when the exposé articles were published. First, in Panel A, we examine

whether the propensity of fake news has gone down after the scandal and whether the effect

varies by firm size. In the first column, for all firms, the prevalence of fake news before

and after the scandal was similar. In the second column, we examine whether the author’s


past impact played a difference in the decision to post fake news. Author Impact is the

average idiosyncratic volatility over days [t, t+2] observed for the author after the release of

all prior articles. In column two of Panel A, we we see that before the scandal, authors who

were more impactful were more likely to write fake news, whereas after the scandal, there

was no more difference. Which probably suggests that the successful authors decided that

the cost to publishing fake news was much higher than they previously thought. Next, we

separate our analysis by small, midsize, and large firms. Interestingly, and fairly intuitively,

the scandal had the biggest impact on the prevalence of fake articles in small firms, relative

to medium and large firms. And for large firms the effect was even slightly positive. This

suggests that if companies were doing any promotional articles, they started doing it much

less, for the fear of being caught by the SEC, and it’s mostly small firms that engage in this

behavior. Also the difference that author’s prior impact had on propensity to publish fake

news, was biggest for small firms.

In Panel B, we perform the same analysis, but separate firms by retail ownership (within

size deciles) and by industry classification. There doesn’t seem to be a big difference between

high and low retail investor base, yet there is a difference depending on the industry. As

Table 3 demonstrated, the majority of companies that got caught in the scandal were in the

Healthcare industry. Therefore, it might have been more salient for authors and investors

in that industry. The results confirm that intuition. The drop in the propensity of fake

articles was much larger for Healthcare firms than for non-healthcare firms. Also, the effect

of author impact on the propensity of fake articles was stronger for firms in the healthcare

industry.

Next, in Panels C and D of Table 5, we examine whether articles’ impact on abnormal

trading volume has been affected by the scandal. We have a similar setup as in Panels A

and B, expect we use abnormal trading volume over days t, t+1, and t+2 as the dependent

variable. The first two columns of Panel C, examine whether the impact of non-fake and

fake articles has been affected by the scandal. First, the results show that the effect of a


non-fake article was 15% lower after the scandal than before. Furthermore, can see a result

we already saw in Table 4, that before the scandal the effect of fake articles was much larger

than non-fake articles. The results also suggest that that difference went down after the

scandal, probably driven by the fact that people seem to trust all news less. In the second

column of Panel C, we see that before the scandal articles by more impactful authors led to

higher abnormal trading volume, but that differences has decreased after the scandal. The

rest of the panel separates the analysis by whether the firm was small, medium, or large.

Similar to the results in Panel A, the decrease in the effect of non-fake articles on abnormal

trading volume was largest for small firms, and so was the decrease in the effect of fake

articles and of articles by impactful authors.

In Panel D, we separate the firms by retail ownership (within size deciles) and by industry

classification. We find that for non-fake news the decrease in impact was stronger for firms

with fewer retail investors, yet the effect of fake articles and of high-impact authors before

the scandal and the decrease after the scandal were much larger for firms with high retail

ownership. The effects on abnormal volume were similar for firms in Healthcare and in

non-healthcare industries.

4. Impact on Returns

Using the sample of for-sure fake articles as well as the set of probabilistically fake articles

from the broader sample, we investigate the market’s response to fake articles.

4.1. Return Reaction

First we examine the return reaction to the for-sure fake articles that were provided to

us by Rick Pearson or that were subpoenaed by the SEC. We separate firms by size (small

and mid-size) and examine their return response to the release of for-sure fake articles. We

classify a firm as small if its market cap is in the bottom 10th percentile of NYSE stocks

and as medium if it is in the 20th-90th percentile by market cap among NYSE stocks. The

cumulative abnormal returns, measured as equal-weighted 4-factor residuals, are constructed


starting the day after the article was published until 251 trading days after the article was

published. We generate equal-weighted Mom, SMB, HML, and Mkt factors, and estimate

betas for a given stock i for day t using the window t−252 to t−1. We then use those betas

to calculate the residual (abnormal) cumulative returns for stock i for days t+ 1 to t+ 251.

Figure 4 plots the cumulative abnormal returns for the for-sure fake articles for small

(blue line) and mid-size (red line) firms. Out of the 111 articles in our sample, 69 are about

small firms, 35 are about mid-size firms, and 7 are about large firms. Returns for small firms

increase after the article is published, reaching as much as 20%, cumulatively, after about

60 days, before giving up all the gain, and ending up with a 10% loss towards the end of the

year. The permanent price impact of −10% for small firms indicates that once the market

figures out the news is fake, investors view this as a bad signal about the firm. Interestingly,

for mid-size firms, there is no gain followed by reversal – the price starts dropping after the

fake article comes out and continues to decrease throughout the year. These results suggest

that for both small and mid-size firms, the fact that management is trying to prop up the

stock price with for-sure fake articles is a signal for deteriorating underlying performance.

However, either due to larger limits to arbitrage or to less sophisticated investor base, the

for-sure fake articles are successful at temporarily propping up the stock price of small firms,

but not of mid-size firms.

The articles we obtained from Rick Pearson and the SEC’s subpoenaed articles are a

small sample. We next examine the market response to articles that we classify as fake using

LIWC. Since our analysis is at the firm-day level, we need to define whether a firm had a

for-sure fake article on a given day. In order to do so, we calculate the average authenticity

score among all articles for a given firm/day and define that a company had fake articles on

a given day if the average authenticity score translates into the probability of being fake of

20% or more, and that a firm had no fake articles if the average authenticity score translates

into the probability of being fake of less than 1%.

Figure 5 plots the difference between abnormal cumulative returns following days with


fake articles, relative to days with no fake articles separately for small, mid-size, and large

firms in our sample (that have at least one fake article). As the blue line shows, the returns

for small firms increase for 6 months by about 8% following a fake article, relative to a non-

fake article, and then revert back to their original level. Whereas the returns for mid-size

firms (red line) start dropping almost immediately, and come to a steady state of -5% after

about 10 months. It is important to note that small firms experience temporary positive

returns following fake articles, whereas mid-size firms see a decrease in returns, which is very

similar to return patters following for-sure fake articles that Rick Pearson shared with us

(shown in Figure 5), which helps to corroborate the patterns. For large firms (green line) the

returns appear to first go up and then decrease, but the magnitudes are quite small – 50 to

100 basis points and are not statistically different from zero. Suggesting that the markets for

those firms are quite efficient, and also that the articles that we identify as fake are probably

one-off rogue investors, rather than those companies launching fake-article campaigns. The

for-sure fake articles that we obtained from Rick Pearson only included a few firms in this

size category.

Next, we examine whether the patters in cumulative abnormal returns for different-sized

firms we observe in Figure 5 are statistically significant. In order to do so, we estimate the

following model:

Reti,(t+1,t+T ) = α + βFakei,t + εi,t

where Reti,(t+1,t+T ) are cumulative abnormal 4-factor returns for firm i, from 1 day after the

fake article was published until T days, where T is either 51, 101, 151, 201, or 251. The

results are presented in Table 6. As we saw in Figure 5, for small firms, the returns in the

first 6 to 7 months following fake articles, are more positive than following non-fake articles,

and the difference is statistically significant. This gain disappears after about 10 months

and is basically 0 after a year. For mid-size firms, the returns start decreasing following the

publication of fake articles, relative to days with non-fake articles, and continue to decrease

for about 10 months, before coming to a steady state at around -4%. Finally, for large


firms, the difference is very small (60 basis points after 3 months) and is barely statistically

significant.

4.2. Fake articles and firm fundamentals

While we saw in the last section that the presence of fake articles is usually bad news,

especially for mid-size firms, it still remains somewhat unclear whether it’s due to investors’

over/under reaction or whether it is indeed a sign of poor firm performance. Therefore, next,

we examine whether the presence of fake articles during the quarter are associated with

worsening fundamental performance. We measure fundamental performance as SUE, which

is defined as the seasonally-adjusted change in earnings scaled by the standard deviation of

seasonally-adjusted change over the prior eight quarters. We regress SUE on a fake article

dummy equal to 1 there was at least one fake article in the 90 days leading up to earnings

announcements, and 0 otherwise. We define an article as being fake if the probability of the

article being fake is > 20%. We only include firms in this analysis that had at least one fake

article in our sample. We include firm and year-month fixed-effects.

The results are presented in Table 7. We can see that for all firms, the presence of at

least one fake article during the quarter, is associated with a 0.125 lower SUE disclosed at

the end of the quarter. This is about a decrease in 1/10th standard deviation. In the next

three columns we examine the effect separately for small, midsize, and large firms. Small

firms are defined as firms in the bottom 10th percentile of NYSE firms, mid-size firms are

defined as firms in the 20th-90th percentile of NYSE firms, and Large firms are defined as

the top 10th percentile of the NYSE firms. For large firms, there is no effect. For small and

medium firms the effect is negative, and it’s very statistically significant for medium firms.

This result is consistent with the effect of fake articles on abnormal returns we presented in

the last section.


5. Firm Disclosures and Insider Trading

So far we have provided evidence that fake articles can influence asset prices. Next we

examine whether firms take any actions to facilitate the promotion of the articles, and what

incentives managers have to try to pump up their stock price. In particular, we examine

whether companies are more likely to issue press releases around the time of fake articles

such as to give the authors of the fake articles some material to write about. We further

look at whether there is indirect evidence to suggest that managers engage in what is called

"pump-and-dump" schemes, where one acquires shares at a low price, then inflates the price

through fake articles, and then sells the stock. While we cannot observe trades for regular

investors, this is something we can observe for managers, as they have to report their insider

trades to the SEC using Form 4.

5.1. Firm Disclosures

When authors write fake articles about the firms, it would make the articles much more

believable if the firm issued a press-release or submitted a material information filling with

the SEC at the same time. The press release and/or an SEC filing can also provide some

material facts for the author to write about and the author can then potentially exaggerate

what those facts mean for the future of the firm. In order to examine this question we regress

whether there were fake articles in a given week, on whether there was a press-release or

an SEC (8-K filling) in the prior week, the week of and a week after the fake articles are

published.12 We perform our analysis separately for small, mid-size, and for large firms.

We find that small firms are substantially more likely to have fake articles written about

them in the week before, the week of, and the after they issue a press release or file an 8-K

form with the SEC. The coefficients become insignificant if we go out further than those

weeks. Mid-size firms have an increased probability of having fake articles written about

them in the week of the press release or an 8-K filing, and the is no effect for large firms.12An 8K-filling is a form a firm has to file with the SEC if a material event has occurred at the firm within

the past 5 business days.


These results are consistent with the anecdotal evidence that companies often issue press

releases to provide some material for the fake articles. Large firms are unlikely and do not

seem to engage in this behavior. These results are consistent with a deliberate campaign by

the firm to manipulate the stock price and take advantage of the price impact.

5.2. Insider Trading

First, at the monthly level, we regress an indicator variable for whether a firm had

predominantly fake articles in a given month on whether the firm was a net insider buyer or

a net insider seller in the previous month and in the contemporaneous month. A firm is a net

buyer (seller) if insiders bought more shares, in dollar value, than they sold in a given month

(sold more shares than they bought). We define an indicator variable, Fake Article, to be

1 if the probability of being fake associated with the average authenticity score for articles

written about the firm in the given month is great than 20%. We perform our analysis

separately for small, mid-size, and for large firms. We do not tabulate the results for brevity

but we find that for small and mid-size firms, insider buying and issuing of press releases is

strongly associated with the prevalence of fake articles in the same month. Whereas there

is no association between insider trading and fake articles in the previous and the following

months.

The above analysis shows that insiders buy stock and issue press releases in the same

month as fake articles come out. Next, we zoom in on the weeks around fake articles and

examine the timing in more detail. We run similar regressions as we did at the monthly level,

except now everything is defined at the weekly level. Therefore, we regress an indicator

variable for whether a firm had predominantly fake articles in a given week, on whether

insiders were net buyers or net sellers in the week before, the week of, and the week after the

fake article came out. Net Buyer (Net Seller) is an indicator for whether insiders bought

more shares, in dollar value, than they sold in a given week (sold more shares than they

bought). We define a dummy variable for whether a firm had predominantly fake articles in

a given week as 1 if the probability of being fake associated with the average authenticity


score for articles written about the firm in the given week is great than 20%. We perform

our analysis separately for small, mid-size and for large firms.

The results are presented in Table 9. For small firms and mid-size firms, insiders do not

trade in the weeks leading up to fake articles, trade a little bit in the week before the fake

article, and then start actively buying the week of and the week after the fake articles come

out. These findings are consistent with Figure 1, where insiders start buying Galena’s stock

around/after fake articles come out. We do not find a similar result for large firms.

5.3. Insider Trading and Returns

In the last section we have documented that, on average, for small and midsize firms,

insiders tend to buy stock around the time that fake articles come out. Next, we examine

whether the impact of fake articles on returns is the differs whether insiders purchase stock

versus when they don’t. To do that, we separate articles by whether the firm was a net

insider buyer (insiders bought more stock than they sold) in the two weeks leading up to

the article being published. We concentrate on small firms, and examine fake and non-fake

articles separately. The results are presented in Figure 6. In Panel A, we examine the effect

of fake articles on returns, and separate them into articles that followed insider buying versus

the ones that didn’t. The figure clearly shows that when fake articles are published following

heavy insider buying, it usually happens after a stretch of falling stock prices, and it seems

to temporarily prop up the stock price. Whereas, we do not see a similar pattern for fake

articles that did not follow insider purchases. Especially, the difference in returns leading

up to the fake articles, suggests that the articles that are accompanied by insider purchases

might be more likely to be orchestrated by the firm, and thus come at a very specific time,

and might be accompanied by other promotional attempts. Whereas, fake articles that are

not accompanied by insider purchases might be one off attempts by individuals to manipulate

the stock price, and thus seem to be less successful.

One explanation could be that the above documented patterns have nothing to do with

fake articles, but are just the market’s response to insider buying. Therefore, in Panel B,


we perform a similar analysis, but we concentrate on non-fake articles. As the graph for

the non-fake articles with and without insider buying shows there is not a big difference in

returns. Suggesting, that it’s not the insider buying that caused the differences in returns

in Panel A, but the interaction of insider buying with fake articles. In Panel B, there is a

very distinct jump in returns leading up to the article. This is most likely caused by insider

purchases, as these are the two weeks over which we measure insider buying. It is quite

curious that we do not seem to have a similar increase in returns for insider purchases prior

to fake articles in Panel A.

6. Conclusion

Examining the impact of false information on prices using our unique datasets of fake

articles, our novel test of market efficiency finds that markets respond to erroneous infor-

mation in small stocks, possibly leading to potential price manipulation. The “non-event"

studies we conduct find strong temporary price impact and subsequent reversals from fake

news for small firms that coincide with insider trading and firm press releases, that predict

the magnitude of the price reaction and reversal. We find similar results for mid-size firms

except there is no temporary price increase and only a permanent price decrease associated

with fake articles, especially when coordinated with insider trades and press releases by the

firm. Large cap stocks exhibit none of these patterns nor any price impact from the fake

articles.

The evidence suggests that markets are efficient with respect to fake news for large and

possibly mid-cap firms, but is inefficient for small cap stocks, consistent with information

costs being greater for smaller firms. Small firms therefore engage in possible price manip-

ulation that temporarily props up the share price and eventually reverses over the course of

the year. Mid-size firms seem to engage in similar behavior, but the market is not fooled and

applies an immediate permanent price discount on those firms. Large firms do not engage

in this behavior, consistent with its share prices being immune from fake news and the cost


of information low enough in large firms that prices remain efficient.

Further research seeks to understand why firms may or may not engage in price manip-

ulation, how these social media platforms can amplify or hinder the price discovery process,

and potentially provide a way to measure the cost of information across firms.


References

Grossman, S. J. and J. E. Stiglitz (1980). On the impossibility of informationally efficient

markets. The American economic review 70 (3), 393–408.

Newman, M. L., J. W. Pennebaker, D. S. Berry, and J. M. Richards (2003). Lying words: Pre-

dicting deception from linguistic styles. Personality and social psychology bulletin 29 (5),

665–675.

Pennebaker, J., R. Booth, R. Boyd, and M. Francis (2015). Linguistic inquiry and word

count: Liwc 2015 operator’s manual. retrieved april 28, 2016.

Pennebaker, J. W. (2011). The secret life of pronouns: what our words say about us.

Chapter 6. New York: Bloomsbury Press.


APPENDIX

Appendix A: An example of a for-sure fake article about Galena Biopharma Inc.

Appendix B: 8-K form documenting the settlement between the SEC, Galena, and Mr. Ahn.


Figure 1. Example of a Pump-and-Dump Scheme: Galena Biopharma Inc.

This figure depicts the stock price of Galena Biopharma Inc. from April 2013 - May 2014, as well asoccurrences of fake articles being published on Seeking Alpha, instances of SEO and stock options beinggranted to senior executives, as well as instances of insider trading and exposé articles about the promotionalarticles. This information was obtained from the SEC Lawsuit filed against Galena on 31 October, 2014 inthe United States District Court (Case 3:14-cv-00558-SI)13.According to the lawsuit, the fake articles were published on August 6 and 22, 2013, September 26 and 30,2013, November 12, 13, and 22, 2013, December 4, 10, 16, 2013, January 15, 2015, and February 5, 2014.While this was happening, Galena sold on September 18, 2013 in an SEO 17,500,000 units of stock for netproceeds to Galena of $32.6 million. On November 22, 2013, Galena held a board meeting and granted stockoptions to executives and directors with a strike price of $3.88. The CEO received 600,000 options, the CMOand COO 300,000 options, the CAO 150,000 options and each of the six directors received 200,000 options.Galena has historically awarded options either at the end of December of in early January.During the board meeting on January 16, 2014, where the board reviewed the preliminary 2013 earningswhich have not been made public yet, the CEO declared that insiders could trade the company’s stockimmediately. Between January 17 and February 12, 2014 insiders sold over $16 million of their stock.On January 24 and 27, 2014 attention has been drawn to the large insider trades. Then on February 1,13, 14 and on March 13, 2014 articles started to appear on Seeking Alpha and TheStreet, documenting thepromotional scheme. Finally on March 17, 2014, Galena disclosed in it’s 10-K form an SEC probe.

$0

$1

$2

$3

$4

$5

$6

$7

$8

Apr-13 May-13 Jun-13 Jul-13 Aug-13 Sep-13 Oct-13 Nov-13 Dec-13 Jan-14 Feb-14 Mar-14 Apr-14 May-14

Stoc

k Pr

ice

Price Fake Articles SEO Stock Options Insider Selling Exposé Articles


Figure 2. Authenticity Scores

This figure depicts the distribution of authenticity scores for fake and non-fake articles. In PanelA, we plot authenticity scores for all the articles in our validation sample of 171 fake and 334 non-fakearticles. In Panel B, we plot authenticity scores for two authors in our validation sample with the most articles

Panel A:0

510

1520

25

0 20 40 60 80 0 20 40 60 80

Non Fake Fake

Perc

ent

Authenticity ScoreGraphs by fake

Panel B:

010

2030

4050

6070

0 20 40 60 80 0 20 40 60 80

Non Fake Fake

Perc

ent


John Mylant

010

2030

4050

6070

8090

100

0 20 40 60 80 0 20 40 60 80

Non Fake Fake

Perc

ent


Equity Options Guru


Figure 3. Authenticity score and the probability of being fake

This figure depicts the relationship between LIWC authenticity scores (S) and the conditional probabilityof being fake (Prob(F |S)).


Figure 4. Abnormal Returns for For-Sure Fake Articles

The figure depicts cumulative abnormal returns (measured as equal-weighted 4-factor residuals) for for-surefake articles that were provided to us by Rick Pearson and that were subpoenaed by the SEC. The cumulativereturns are measured starting with the day after the article was published until the 251 trading days after thearticle was published. For the time period before the article was published we measure cumulative returnsstarting with the day -120 and ending on the day before the article publication. Small firms are definedas firms in the bottom 10th percentile of NYSE firms, medium firms are defined as firms in the 20th-90thpercentile of NYSE firms.

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

-120 -90 -60 -30 0 30 60 90 120 150 180 210 240Small Firms Medium Firms


Figure 5. Abnormal Returns for Fake versus Non-Fake Articles

The figure depicts the difference in cumulative abnormal returns (measured as equal-weighted 4-factor resid-uals) between days with fake articles and days with non-fake articles separately for small, mid-size, and largefirms in our sample. We designate a given day t for company i to have a fake article, if the probability ofbeing fake, associated with the average authenticity score for all articles about firm i on day t, is greaterthan 20%. Similarly, we designate a day t for company i as not having any fake articles, if the probability ofbeing fake, associated with the average authenticity score for all articles about firm i on day t, is less than1%. The cumulative returns are measured starting with the day after the article was published until the 251trading days after the article was published. For the time period before the article was published we measurecumulative returns starting with the day -120 and ending on the day before the article publication. Smallfirms are defined as firms in the bottom 10th percentile of NYSE firms, medium firms are defined as firmsin the 20th-90th percentile of NYSE firms, and large firms are defined as firms above the 90th percentile ofNYSE firms.

-0.08

-0.06

-0.04

-0.02

0

0.02

0.04

0.06

0.08

-120 -90 -60 -30 0 30 60 90 120 150 180 210 240Small Firms Medium Firms Large Firms


Figure 6. Abnormal Returns and Insider Trading

The two figures below depict for small firms the cumulative abnormal returns (measured as equal-weighted4-factor residuals) for days with articles where the firm was a net insider buyer in the two weeks leading upto the article versus days with articles where the firm was not a net insider buyer. A firm is a net insiderbuyer if the dollar value of stock bought by insiders is larger than the value of stock sold by insiders. PanelA shows the results for days with fake articles and Panel B shows the results for days with non-fake articles.We designate a given day t for company i to have a fake article, if the probability of being fake, associatedwith the average authenticity score for all articles about firm i on day t, is greater than 20%. Similarly, wedesignate a day t for company i as having a non-fake article, if the probability of being fake, associated withthe average authenticity score for all articles about firm i on day t, is less than 1%. The cumulative returnsare measured starting with the day after the article was published until the 251 trading days after the articlewas published. For the time period before the article was published we measure cumulative returns startingwith the day -120 and ending on the day before the article publication. Small firms are defined as firms inthe bottom 10th percentile of NYSE firms.

Panel A: Fake Articles with/without Insider Buying for Small Firms

-0.1

-0.05

0

0.05

0.1

0.15

0.2

0.25

-120 -90 -60 -30 0 30 60 90 120 150 180 210 240No Insider Buying Insider Buying

Panel B: Non-Fake Articles with/without Insider Buying for Small Firms

-0.15

-0.1

-0.05

0

0.05

0.1

0.15

0.2

0.25

-120 -90 -60 -30 0 30 60 90 120 150 180 210 240No Insider Buying Insider Buying


Table 1. LIWC and Fake Articles

This table presents the summary statistics for various LIWC textual measures for for-sure fake articles thatwere shared with us by Rick Pearson, articles that were subpoenaed by the SEC that Seeking Alpha sharedwith us, and non-fake articles that were written by the same authors. For-sure Fake articles are articlesthat have been shared with us by Rick Pearson and that were subpoenaed by the SEC. Non-fake articles,are articles that were written by the same authors, but about larger firms, that are unlikely to be fake.We display the number of articles in each category as well as the mean of the Authenticity measure fromLIWC. We also report the means of several other variables provided by LIWC to help better understandthe authenticity score. In particular we display the means of the average of the 1st person singular measure(examples: I, me, mine), Insight measure (examples: think, know), Relativity measure (examples: area,bend, exit), Time measure (examples: end, until, season), Discrepancy measure (examples: should, would),and the average number of words per sentence.

For-sure Fake Non-Fake

Number of articles 171 334Authentic 19.08 32.77

(10.66) (15.73)1st pers singular 0.42 0.76

(0.43) (0.67)Words per sentence 57.54 65.23

(33.30) (31.73)Insight 1.52 1.67

(0.59) (0.71)Relativity 12.92 15.11

(2.04) (2.56)Time 4.97 5.35

(1.26) (1.45)Discrepancy 1.40 1.05

(0.73) (0.45)Clout 58.25 52.31

(9.78) (8.63)

FakeNew

sin

FinancialM

arkets42

Table 2. Summary Statistics

This table presents the summary statistics for various LIWC textual measures and firm characteristics of the covered firms, for different types ofarticles on Seeking Alpha and Motley Fool. For-sure Fake Articles are articles that have been shared with us by Rick Pearson, or that were subpoenaedby the SEC and shared with us by Seeking Alpha. Seeking Alpha Articles and Motley Fool Articles are regular articles that we downloaded fromSeeking Alpha and Motley Fool. Of those articles, Fake articles are articles whose probability of being fake was higher than 20%, Non Fake articlesare articles with probability of being fake less than 1%, and the rest are classified as Other, which are not used in our main analysis.In Panel A, we display the number of articles in each category as well as the mean of the Authenticity measure that we use to construct theprobabilities of being fake. We also report the means of several other variables provided by LIWCvariables provided by LIWC to help betterunderstand the authenticity score. In particular we display the means of the average of the 1st person singular measure (examples: I, me, mine),Insight measure (examples: think, know), Relativity measure (examples: area, bend, exit), Time measure (examples: end, until, season), Discrepancymeasure (examples: should, would), and the average number of words per sentence. In Panel B, we display the average probability of being fake,for each of the article categories. In Panel C, for the firms that are covered in the respective article groups, we provide the average fraction of retailinvestors, the average number of analysts covering the firm, and the average firm size (in Millions of dollars).

Rick Pearson & SEC Seeking Alpha Motley FoolFor-sure Fake Non Fake Fake Non Fake Other Fake Non Fake Other

Panel A: LIWC variablesNumber of articles 171 334 3,933 116,289 83,323 1,368 78,943 67,605Authentic 19.09 32.79 5.44 50.71 22.51 5.71 46.75 21.961st pers singular 0.42 0.76 0.25 0.98 0.54 0.20 0.53 0.23Words per sentence 57.55 65.23 23.89 21.76 22.18 31.23 19.28 19.39Insight 1.52 1.67 1.43 1.75 1.63 1.62 2.08 1.84Relativity 12.92 15.11 9.90 17.37 13.53 9.20 16.57 13.29Time 4.97 5.35 3.40 6.34 4.68 3.34 6.54 5.23Discrepancy 1.41 1.05 1.40 1.12 1.22 0.76 1.08 1.11

Panel B: Probability of being FakeProb(Fake) 0.08 0.02 0.45 0.01 0.03 0.42 0.01 0.03

Panel C: Firm characteristicsPercent of retail investors 76.66% 50.15% 42.32% 42.46% 44.96% 40.88% 36.78% 38.99%Numer of Analysts 6.96 16.76 16.83 18.33 16.67 23.21 19.84 20.34Firm Size ($Mil) 7.36 58.43 44.12 51.72 45.17 101.97 70.58 80.40

FakeNew

sin

FinancialM

arkets43

Table 3. Fake Articles and Industries

This table presents the distribution of articles by Fama-French 12 industries, for different types of articles on Seeking Alpha and Motley Fool. For-sureFake Articles are articles that have been shared with us by Rick Pearson, or that were subpoenaed by the SEC and shared with us by Seeking Alpha.Seeking Alpha Articles and Motley Fool Articles are regular articles that we downloaded from Seeking Alpha and Motley Fool. Of those articles, Fakearticles are articles whose probability of being fake was higher than 20%, Non Fake articles are articles with probability of being fake less than 1%,and the rest are classified as Other, which are not used in our main analysis.

Rick Pearson & SEC Seeking Alpha Motley FoolIndustry For-sure Fake Non-Fake Fake Non-Fake Others Fake Non-Fake Others

Consumer NonDurables - 2.45% 2.57% 5.19% 4.53% 5.67% 5.19% 5.19%Consumer Durables - 4.49% 3.13% 3.52% 3.37% 6.66% 5.04% 4.04%Manufacturing 2.30% 12.65% 4.55% 7.26% 5.82% 8.05% 9.98% 8.09%Energy - 8.16% 4.9% 6.52% 6.17% 5.26% 5.66% 6.68%Chemicals 1.15% 1.22% 1.46% 1.79% 1.78% 1.97% 2.44% 2.34%Business Equipment 4.60% 27.35% 28.13% 23.66% 25.91% 26.87% 26.22% 25.39%Telecom - 2.86% 6.39% 4.77% 4.72% 4.35% 3.61% 3.87%Utilities - - 1.11% 0.99% 1.46% 1.23% 1.66% 2.1%Shops - 2.86% 6.84% 12.19% 9.21% 13.72% 13.69% 11.62%Healthcare 81.61% 17.14% 10.63% 5.38% 9.6% 7.81% 7.92% 10.4%Finance - 13.06% 22.2% 16.67% 16.42% 10.85% 6.49% 8.9%Other 10.34% 7.76% 8.09% 12.06% 11.03% 7.56% 12.11% 11.38%


Table 4. Article Impact on Abnormal Volume

The tables examines how investor attention (proxied for by log of abnormal volume on days t = 0, t+1, andt + 2), responds to fake and non-fake articles as a function of article and author characteristics, abnormalvolatility, firm size, and fraction of retail investors. Panel A shows results for all firms that have ever hadan article written about them on Seeking Alpha or Motley Fool. Panel B concentrates on days with articles.Abnormal volume is defined as V ol(t)/AvgV ol(t− 250, t− 1), summed over days t = 0, t+1, and t+2, andthen we take the natural log of the sum. Fake Article is a dummy equal to 1 if the probability of an articlebeing fake is > 20%, 0 if the probability of an article being fake is < 1%, and missing otherwise. AuthorImpact is the average idiosyncratic volatility over days [t, t+2] observed for the author after the release ofall prior articles. Small firms are defined as firms in the bottom 10th percentile of NYSE firms, mid-sizefirms are defined as firms in the 20th-90th percentile of NYSE firms, and Large firms are defined as the top10th percentile of the NYSE firms. We include both firm and year-month fixed effects when firm size orpercentage of retail investors are not used as explanatory variables.

Panel ADependent variable Ln([t, t+ 2] day abnormal volume)

Firm Size

All Small Medium Large

Article dummy 0.121*** 0.651*** 0.183*** 0.106***(67.73) (21.8) (78.96) (64.78)

Observations 13,400,000 716,815 11,200,000 1,563,820R2 0.6 0.46 0.58 0.64Firm FE X X X XYear-Month FE X X X X

Panel B

Dependent variable Ln([t, t+ 2] day abnormal volume)

Fake Article 0.034*** 0.041 0.159**(3.17) (0.59) (2.42)

Author Impact 0.077***(49.52)

Fake Article × Author Impact 0.002(0.24)

ME percentile -0.022***(-256.18)

Fake Article × ME percentile -0.001*(-1.69)

Percentile retail investors 0.454***(42.48)

Fake Article × Percentile retail investors 0.284***(3.70)

Observations 190,152 185,146 173,384R2 0.783 0.786 0.310Firm FE X X XYear-Month FE X X X

FakeNew

sin

FinancialM

arkets45

Table 5. 2014 SEC Lawsuit Event Study

The table examines how fake article intensity and firm-level abnormal volume after the release of articles (days [t, t + 2]), change around the 2014exposé articles and SEC lawsuit. The exposé articles and the lawsuit occurred in February and March of 2014. We study the 6-month time periodsbefore and after February and March of 2014. Abnormal volume is defined as V ol(t)/AvgV ol(t− 250, t− 1), summed over days t = 0, t+1, and t+2,and then we take the natural log of the sum. Post Event is defined as the time period after February and March, 2014. Fake Article is a dummy equalto 1 if the probability of an article being fake is > 20%, 0 if the probability of an article being fake is < 1%, and missing otherwise. Author Impact isthe average idiosyncratic volatility over days [t, t+2] observed for the author after the release of all prior articles. Small firms are defined as firms inthe bottom 10th percentile of NYSE firms, mid-size firms are defined as firms in the 20th-90th percentile of NYSE firms, and Large firms are definedas the top 10th percentile of the NYSE firms. Retail Ownership is divided into high and low, depending on whether the fraction of retail ownerswas above or below the median fraction of retail ownership in the firm’s size decile on August, 2013. Healthcare industry membership is according toFama-French’s 12 industry classification. We include firm fixed-effects.

Panel A

Dependent Variable Fake Article

Firm Size

All Firm Small Medium Large

Post event 0.001 -0.040*** -0.047*** -0.248*** -0.001 -0.031** 0.004*** -0.002(0.52) (-4.75) (-6.06) (-7.13) (-0.30) (-2.11) (2.71) (-0.18)

Author Impact 0.008*** 0.025*** 0.007*** 0.003**(8.80) (7.25) (4.19) (2.14)

Post event × Author Impact -0.005*** -0.027*** -0.004** -0.001(-4.90) (-5.91) (-2.05) (-0.57)

Observations 63,918 62,283 2,441 2,309 22,175 21,538 30,458 29,847R-squared 0.266 0.266 0.540 0.536 0.127 0.132 0.029 0.030

Panel B

Dependent Variable Fake Article

Retail Ownership Industry Classification

Low Retail High Retail Healthcare Non-Healthcare

Post event 0.003 -0.026** -0.000 -0.075*** -0.020** -0.177*** 0.002 -0.028***(1.62) (-2.02) (-0.14) (-5.88) (-2.19) (-3.47) (1.23) (-3.23)

Author Impact 0.006*** 0.009*** 0.033*** 0.006***(3.87) (6.44) (6.09) (6.81)

Post event × Author Impact -0.003** -0.009*** -0.020*** -0.004***(-2.24) (-5.92) (-3.19) (-3.47)


FakeNew

sin

FinancialM

arkets46

Panel C

Dependent Variable Log Abnormal Volume [t, t+2]

Firm Size

All Firm Small Medium Large

Post event -0.152*** -0.422*** -0.615*** -1.547*** -0.154*** -0.095 -0.119*** -0.115**(-21.92) (-11.10) (-12.08) (-9.15) (-12.78) (-1.39) (-14.98) (-2.39)

Fake Article 0.220*** 0.621*** 0.057 -0.180*(3.58) (3.46) (0.56) (-1.82)

Post event × Fake Article -0.193*** -0.770** -0.041 0.192*(-2.91) (-2.41) (-0.37) (1.85)

Author Impact 0.090*** 0.144*** 0.078*** 0.049***(21.59) (7.80) (10.16) (9.69)

Post event × Author Impact -0.033*** -0.116*** 0.005 -0.001(-7.28) (-5.21) (0.63) (-0.25)


Panel D

Dependent Variable Log Abnormal Volume [t, t+2]

Retail Ownership Industry Classification

Low Retail High Retail Healthcare Non-Healthcare

Post event -0.212*** -0.201*** -0.101*** -0.693*** -0.111*** -0.717*** -0.153*** -0.394***(-22.88) (-3.95) (-8.30) (-11.23) (-3.12) (-4.97) (-21.78) (-9.99)

Fake Article -0.143 0.613*** 0.426*** 0.164**(-1.59) (6.44) (3.17) (2.33)

Post event × Fake Article 0.180* -0.558*** -0.127 -0.158**(1.90) (-5.21) (-0.76) (-2.11)

Author Impact 0.071*** 0.109*** 0.095*** 0.089***(12.68) (16.23) (5.89) (20.51)

Post event × Author Impact -0.001 -0.070*** -0.061*** -0.030***(-0.09) (-9.49) (-3.45) (-6.43)



Table 6. Return Window Regressions – Unconditional

The table reports results from regressing 4-factor cumulative abnormal returns Ret1,51, Ret1,101, Ret1,151,Ret1,201, Ret1,251 on a dummy variable for whether an article was fake. Small firms are defined as firms inthe bottom 10th percentile of NYSE firms, mid-size firms are defined as firms in the 20th-90th percentile ofNYSE firms, and Large firms are defined as the top 10th percentile of the NYSE firms.

Ret1,51 Ret1,101 Ret1,151 Ret1,201 Ret1,251

Small Firms

Fake Article 0.034 0.063*** 0.055* 0.027 0.017(1.61) (2.66) (1.85) (0.77) (0.45)

Constant -0.022*** -0.045*** -0.064*** -0.078*** -0.086***(-8.81) (-12.62) (-13.48) (-13.20) (-12.25)

Observations 11,622 11,622 11,622 11,622 11,622R2 0.000 0.000 0.000 0.000 0.000

Medium Firms

Fake Article -0.017*** -0.020** -0.028** -0.045*** -0.038**(-3.04) (-2.45) (-2.50) (-3.51) (-2.50)

Constant -0.006*** -0.012*** -0.017*** -0.025*** -0.031***(-7.76) (-11.45) (-12.21) (-15.07) (-16.05)

Observations 68,087 68,087 68,087 68,087 68,087R2 0.000 0.000 0.000 0.000 0.000

Large Firms

Fake Article 0.006* 0.004 0 -0.007 -0.011(1.71) (0.90) (0.06) (-0.90) (-1.33)

Constant 0.001** 0 -0.003*** -0.004*** -0.005***(2.26) (-0.23) (-2.77) (-3.96) (-4.02)

Observations 47,908 47,908 47,908 47,908 47,908R2 0.000 0.000 0.000 0.000 0.000


Table 7. Fake Articles and Fundamental Performance

This table examines whether the presence of fake articles during a quarter is associated with deterioratingfundamental performance. We measure fundamental performance as SUE, which is defined as the seasonally-adjusted change in earnings scaled by the standard deviation of seasonally-adjusted change over the prioreight quarters. Fake Article is a dummy equal to 1 there was at least one fake article in the 90 days leadingup to earnings announcements, and 0 otherwise. We define an article as being fake if the probability ofthe article being fake is > 20%. We only include firms in this analysis that had at least one fake article inour sample. Small firms are defined as firms in the bottom 10th percentile of NYSE firms, mid-size firmsare defined as firms in the 20th-90th percentile of NYSE firms, and Large firms are defined as the top 10thpercentile of the NYSE firms. We include firm and year-month fixed-effects.

SUE

All Firms Small Medium Large

Fake Article -0.125*** -0.200 -0.201*** -0.001(-3.43) (-1.25) (-4.17) (-0.01)

Observations 82,335 14,190 58,659 7,348R-squared 0.117 0.117 0.133 0.146Firm FEs X X X XYear-Month FEs X X X X

FakeNew

sin

FinancialM

arkets49

Table 8. Fake News and Firm Announcements (Weekly Level)

In this table, we examine whether there are more likely to be fake articles in the weeks around and contemporaneous with insider trading. At theweekly level, we regress a dummy variable for whether a firm had predominantly fake articles in a given week (w = 0) on whether the firm was anet buyer or net a seller in the previous week (w-1), the contemporaneous week (w=0), and the following week (w=1), and a dummy variable forwhether the firm issued a press release in weeks w-1, w=0, or w+1. Net Buyer (Net Seller) is an indicator for whether insiders bought more sharesin dollar value than they sold in a given week (sold more shares than they bought). We define a dummy variable (Fake Article) for whether a firmhad predominantly fake articles in a given week as 1 if the probability of being fake associated with the average authenticity score for articles writtenabout the firm in the given week is great than 20%. PR is an indicator variable for whether the firm issues at least one press release in a given week.We perform our analysis separately for small, mid-size, and large firms. Small firms are defined as firms in the bottom 10th percentile of NYSE firms,mid-size firms are defined as firms in the 20th-90th percentile of NYSE firms, and large firms are defined as firms above the 90th percentile of NYSEfirms. Standard errors are double-clustered at the year-month and firm level.

Fake Article

Small Firms Mid-size Firms Large Firms

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Press Release (week-1) 0.0011** 0.0002 -0.0010*(2.07) (0.85) (-1.75)

8K filing (week-1) 0.0013** 0.0001 0.0003(2.50) (0.29) (0.49)

Press Release (week=0) 0.0026*** 0.0020*** 0.0008(3.78) (6.36) (0.76)

8K filing (week=0) 0.0018*** 0.0017*** 0.0002(2.92) (4.57) (1.07)

Press Release (week+1) 0.0002 0.0002 0.0001(0.36) (0.71) (0.18)

8K filing (week+1) 0.0014** 0.0004 0.0006(2.49) (1.51) (0.88)

Observations 137,560 137,998 137,719 406,508 407,379 406,593 86,956 87,104 86,946R-squared 0.010 0.011 0.011 0.007 0.008 0.007 0.013 0.013 0.013Firm FE X X X X X X X X XYear-month FE X X X X X X X X X

FakeNew

sin

FinancialM

arkets50

Table 9. Insider Trading and Fake News (Weekly Level)

In this table, we examine whether there are more likely to be fake articles in the weeks around and contemporaneous with insider trading. At theweekly level, we regress a dummy variable for whether a firm had predominantly fake articles in a given week (w = 0) on whether the firm was anet buyer or net a seller in the previous week (w-1), the contemporaneous week (w=0), and the following week (w=1), and a dummy variable forwhether the firm issued a press release in weeks w-1, w=0, or w+1. Net Buyer (Net Seller) is an indicator for whether insiders bought more sharesin dollar value than they sold in a given week (sold more shares than they bought). We define a dummy variable (Fake Article) for whether a firmhad predominantly fake articles in a given week as 1 if the probability of being fake associated with the average authenticity score for articles writtenabout the firm in the given week is great than 20%. PR is an indicator variable for whether the firm issues at least one press release in a given week.We perform our analysis separately for small, mid-size, and large firms. Small firms are defined as firms in the bottom 10th percentile of NYSE firms,mid-size firms are defined as firms in the 20th-90th percentile of NYSE firms, and large firms are defined as firms above the 90th percentile of NYSEfirms. Standard errors are double-clustered at the year-month and firm level.

Fake Article

Small Firms Mid-size Firms Large Firms

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Buyer (week-1) 0.0025* 0.0020* -0.0062(1.92) (1.81) (-1.41)

Seller (week-1) 0.0008 0.0004 -0.0008(0.72) (0.90) (-0.91)

Buyer (week=0) 0.0051*** 0.0040*** -0.005(3.03) (4.15) (-0.62)

Seller (week=0) 0.0017 0.0004 -0.0013(1.56) (1.05) (-1.50)

Buyer (week+1) 0.0058*** 0.0022* -0.0013(3.27) (1.94) (-0.55)

Seller (week+1) 0.0010 0.0006 0.0005(1.04) (1.64) (0.53)

Observations 137,593 137,998 137,721 406,575 407,379 406,595 86,959 87,104 86,946R-squared 0.010 0.011 0.011 0.007 0.007 0.007 0.013 0.013 0.013Firm FE X X X X X X X X XYear-month FE X X X X X X X X X

Documents

Fake News in Financial Markets - idc.ac.il · for fake news to inﬁltrate users’ information sets quickly and extensively is becoming a major concern. One prominent example includes