Post by miamianne67 on Aug 2, 2019 21:30:14 GMT
Post by miamianne67 on 4 hours ago
I am posting this for two reasons. I just received a class action settlement of $1.92 (whoo hoo) for my Galena stock and forgetting how and why I ever owned it, I went on a quest and turned up this fascinating abstract which may shed some light on last year's events, and upon practices that have been used to influence retain buyers for a very long time. Anyone with the stamina to get through this will see Biozone prominently mentioned, also. I believe (I am having a lot of difficulty navigating this abstract, that it may have been commissioned by someone or some entity named Garton (Investment Group or Dennis?) which I associate in some way with Icemandios having an opinion of Gartman ....I'm sure I will hear if it is negative. There are another 40 some pages which I believe to be charts etc. supporting the first 40+ pages. If you are really interested and cannot access it, let me know. This is quite dense reading, and surprising in that it wasn't written some years sooner before many retail investors were ....damaged. As Daddy Warbucks is fond of telling me after he has committed some household infraction: "Who ya gonna believe, ME or your lying' eyes?"
Fake News: Evidence from Financial Markets
Shimon Kogan∗
MIT Sloan School of Management Interdisciplinary Center Herzliya
Tobias J. Moskowitz
Yale School of Management NBER
AQR Capital Management
Marina Niessner
AQR Capital Management
March 2019
∗ We thank Tony Cookson, Diego Garcia, Gary Gorton, Bryan Kelly, Elizabeth Kempf, Bonnie Moskowitz, James Pennebacker, Kelly Shue, Eric So, Denis Sosyura, Sam Hartzmark, as well as conference and seminar participants at UCLA (Anderson), Rice University (Jones), University of Miami Business School, ASU Sono- ran Winter Finance Conference, 3rd Annual News & Finance Conference, University of Colorado at Boulder, Northwestern University (Kellogg), FSU SunTrust Beach Conference, MIT Sloan, Yale SOM, Catolica- Lisbon, University of Kentucky Finance Conference, FEB, 3rd Rome Junior Finance Conference, the 2018 WFA meetings, the U.S. Securities and Exchange Commission Division of Economic and Risk Analysis, the Red Rock Conference, Behavioral NBER, and Jackson Hole Finance Conference for their helpful comments and suggestions. We also thank Elli Hoffmann and Keren Ben Zvi for providing and helping organize the data. AQR Capital Management is a global investment management firm, which may or may not apply sim- ilar investment techniques or methods of analysis as described herein. The views expressed here are those of the authors and not necessarily those of AQR. Contact emails: skogan@mit.edu, tobias.moskowitz@yale.edu, and marina.niessner@aqr.com
Fake News: Evidence from Financial Markets 2
Abstract
We examine fake news in financial markets, a laboratory that offers an opportunity to quantify its direct and indirect effects. We study three experimental settings. The first is a unique dataset of unambiguous fake articles on financial news platforms prosecuted by the Securities and Exchange Commission. The second applies a linguistic algorithm to detect deception in expression on the universe of articles on these platforms, using the first sample as a validation and calibration set. The third is an event study exploiting the SEC investigation as a public shock to investor awareness of fake news. We find that fake news increases trading activity and price volatility relative to non-fake news for the equity securities of firms mentioned in the articles. Following public revelation of the existence of fake news, we find an immediate decrease in reaction to all news, including legitimate news, on these platforms, consistent with indirect spillover effects of fake news conjectured by theory. These findings are predominant among small firms with high retail ownership, and are stronger for more circulated articles. Our results are consistent with economic theory on media bias and its application to fake news.
Fake News: Evidence from Financial Markets 1
1. Introduction
False or misleading information can potentially impact social, political, and economic rela- tionships. A recent and prominent example is the increased attention “fake news” is receiving. Fake news is a form of disinformation, including hoaxes, frauds, or deceptions, designed to mislead consumers of news. The economics of fake news is an interesting and young area of study. What motivates fake news? What impact does it have? What are the welfare costs and benefits of monitoring it? What policy prescriptions should be considered?
Analysisoftheseissueshasprimarilybeentheoretical.1 Falsecontentcanimposeprivate
and public costs by making it more difficult for consumers to infer the truth, reduce positive
social externalities from shared-information platforms, increase skepticism and distrust of
legitimate news, and potentially cause resource misallocation. On the other hand, consumers
may derive utility from fake news (as entertainment or if slanted toward their biases as in
Mullainathan and Shleifer (2005)). Little empirical evidence on these issues exists, however,
due to lack of data, particularly the identification of fake content.2 With the explosion
of largely unmonitored shared information platforms, such as social media, blogs, and other
crowd-sourced content, the potential influence of fake and biased news is a growing concern.3
Indeed, a major challenge currently facing Amazon, Facebook, Twitter, and other crowd-
1Allcott and Gentzkow (2017) model fake news as an extension of Gentzkow and Shapiro (2005) and Gentzkow et al. (2015) on media bias, where fake news occurs in equilibrium when agents cannot costlessly verify the truth and the news matches the agent’s priors, with some debate over the relevance and conse- quences of fake news. Aymanns et al. (2017) provide an equilibrium model of an adversary using fake news to target agents with a biased private signal, where knowledge of the adversary causes agents to discount all news. Kshetri and Voas (2017) discuss the pervasiveness of fake news and its dissemination across news consumers.
2For example, Amazon, Google, Twitter, and Facebook are currently using human editors to evaluate content in the hopes of training an algorithm to identify false content systematically with limited success (Cullan-Jones (2016), Leong (2017), Leathern (2017)).
3According to a survey from the Pew Research Center (Gottfried and Shearer (2016)), 62% of American adults get news from a social media site. Allcott and Gentzkow (2017) argue that 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 et al. (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 in order to increase diffusion. Fake news has been suggested to have influenced the 2016 U.S. Presidential election (Allcott and Gentzkow (2017), Silverman (2016), Timberg (2016), Silverman and Alexander (2016)), and a study by ReviewMeta (2016) found that fake reviews on Amazon are misleading consumers toward various products (often paid for by producers).
Fake News: Evidence from Financial Markets 2
sourced content is the ability to detect fake content.
We provide some of the first empirical estimates of the direct and indirect impact of
fake news using three empirical settings. The first is a dataset of identified fake articles from a Securities and Exchange Commission (SEC) investigation into paid-for false articles on shared financial news networks. The sample is small, but the identity of fake news is clean – stemming from an undercover investigation by an industry whistle blower, Rick Pearson, resulting in 171 articles by 20 authors covering 47 companies knowingly providing false information about the stock. The data offer a singular look at identified fake content.
While the first setting provides known fake content, the sample is small and narrow. To broaden the analysis, and perhaps draw more general conclusions, we collect all articles from two prominent financial crowd-sourced websites – Seeking Alpha and Motley Fool – obtaining 203,545 articles from 2005 to 2015 for Seeking Alpha, and 147,916 articles from 2009 to 2014 for Motley Fool, covering 7,700 publicly traded firms. We then attempt to identify fake content within this broader set of articles using a linguistic algorithm (Pennebaker et al. (2015), Newman et al. (2003)) designed to detect deception in expression to assess the authenticity of each article. Using an “off-the-shelf" algorithm allows us to avoid in-sample overfitting, on the other hand we need to verify that the method works in a financial setting. Therefore, it is important that we use the first empirical setting of known fake articles from the SEC to validate the algorithm. It furthermore allows us to calibrate a model for the probability of fake news. Even with our identified set of fake articles, we show how difficult it is to detect fake content systematically. Our calibrated model classifies news as fake, non-fake, and ambiguous with the objective of minimizing classification error so that we can confidently identify fake and non-fake content. The algorithm has a type II error on the known fake articles of less than 1% (false positives) and a type I error on the non-fake articles by the same authors of less than 10% (false negatives). However, the low classification error comes at a cost because there are many articles that cannot be confidently classified, highlighting one of the major challenges and tradeoffs in quantifying fake news generally.
Fake News: Evidence from Financial Markets 3
Despite using a conservative measure, the prevalence of fake news we find in the broader sample is significant (2.8% of articles).
Our third empirical setting does not require identification of fake news at all. Rather, we exploit the public revelation of the SEC’s investigation on these platforms (that ultimately led to the dataset for our first experiment) as a shock to the market’s awareness of fake news. We first show that the market seemed largely unaware of fake content before the announcement, and then examine the market’s response to news before versus after the event. We use this setting to test another implication from theory that fake news imposes externalities on other news. We examine the shock from the public’s awareness of fake news on the market’s reaction to news in general, including legitimate news.
We first examine the direct impact of fake versus non-fake news on trading activity. Using the first empirical setting of known fake articles from the SEC, we find a larger trading response to fake news relative to non-fake articles published at the same time on the same platform. Abnormal trading volume rises by more than 50% over the three days following a fake article relative to a non-fake article. This effect is concentrated in the smallest ten percent of firms on public stock exchanges and is not significant among the largest firms. The effect is also bigger for stocks with higher retail investor ownership, where a ten percent increase in retail ownership results in a 7% increase in the trading volume response to fake news relative to non-fake news. The stronger impact on trading activity is likely driven by fake articles being more sensational and diffusing more quickly across consumers (Vosoughi, Roy, and Aral (2018)). Further corroborating that story, we find that fake articles generate more “clicks” and more “reads.”
Turning to the broader set of articles in the second experiment, where we estimate the probability of fake news, we find similar but more muted results. We similarly find that the direct effect on trading activity from fake news is stronger for smaller firms with higher retail ownership and for articles with greater circulation (measured by number of clicks and readers), lending credence to these platforms influencing investor behavior.
Fake News: Evidence from Financial Markets 4
Exploring the indirect effects of fake news on trading activity using the third empirical setting from the SEC announcement event, we find that trading volume drops significantly for any news article written on these platforms after the public became aware of fake content. Comparing trading volume before versus after the SEC announced investigation, trading volume drops by 5.2% for all news on these platforms following the information shock, with the drop being even larger for small firms with high retail ownership. We also find decreases in trading volume on the Motley Fool platform, which is a competitor platform that was not part of the SEC investigation, indicating that awareness of fake news caused a spillover effect in trading volume for other news platforms as well. These findings are consistent with theory (Allcott and Gentzkow (2017) and Aymanns et al. (2017)) arguing that awareness of fake news causes agents to discount all news. Looking at the the comments section to these articles using natural language processing (NLP), we find a significant increase in uses of the words “fake” and “fraud” after the SEC investigation came to light, consistent with participants being more concerned and aware of fake news after the event. Importantly, the frequency of these words in the comments has no predictive power to detect fake articles, indicating that readers had no ability to identify fake news, but were just more aware of its existence, consistent with their response to distrust all news.
Looking at news media more generally outside of these platforms, we obtain newspa- per articles about the firms from the Wall Street Journal and New York Times (obtained from Factiva) and examine whether there are any spillover effects from the SEC investiga- tion more broadly. Perhaps not surprisingly, and somewhat reassuringly, we do not find a commensurate decline in trading volume in response to news from non-social media sources such as the WSJ and NYT, suggesting that investors’ distrust of news from the social media platforms did not extend to more prominent types of media. The lack of a change in trading activity in response to newspaper articles also provides a useful falsification test that rules out any trend in response to news in general or unobservable effect on trading volume that just happened to coincide with the SEC event, potentially driving our results.
Fake News: Evidence from Financial Markets 5
We also examine pricing effects to see if fake news moves prices in a distortive way. If markets are informationally efficient, then despite the changes in trading volume, prices will be unaffected. In this case, fake news may distort attention and trading, but not firm values. Using the sample of known fake articles from the SEC, we find that the fake promotional articles increase idiosyncratic stock volatility by roughly 40 percent relative to non-fake articles over the three days after publication, with the effects concentrated among small firms with high retail ownership. Looking at the direction of price movement, the average fake article is positive and pushes up stock prices for the smallest 10% of companies on the NYSE by an average 8% over the next six months, which subsequently gets fully reversed over the course of a year and eventually becomes cumulatively negative at −2.5%. Looking at the broader set of articles where we estimate the probability of fake news, we find similar but much weaker patterns of temporary positive price effects for the smallest firms that get fully reversed and turn negative. For large firms, we find no price impact. Finding similar patterns lends further support to our algorithm for detecting fake content.4
The results are consistent with markets being less efficient for small firms, where the cost of information is higher and the average investor is smaller and perhaps less sophisticated.5 Consistent with this notion, paid-for fake content from the SEC investigation was exclusively engaged by small firms, and not by large firms, as expected in equilibrium. We also find that small firms are more likely to issue press releases and 8-K filings coinciding with the fake articles, consistent with a coordinated effort to influence the narrative of news about the firm, and find evidence of insiders positioning themselves to benefit from the price movement.6
Our results provide some of the first empirical estimates of the direct and indirect impact
4An investor at the time of the article’s publication could not have constructed or used a similar method- ology to detect the probability of false content since the fake articles from which we calibrate our model were not yet known or identified.
5The 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 psychological or behavioral biases. Allcott and Gentzkow (2017) argue information costs are necessary for fake news production.
6In fact, The reason Rick Pearson went undercover initially and why the SEC got involved was because many of these fake articles were tied to promotional pump-and-dump schemes to manipulate the stock price, orchestrated by the firms themselves.
Fake News: Evidence from Financial Markets 6
of fake news, which have implications for theory. The prevalence of fake content and its impact on trading activity is consistent with fake news being tailored to consumer’s priors as suggested by Allcott and Gentzkow (2017), and more broadly consistent with media bias (Gentzkow and Shapiro (2005) and Gentzkow et al. (2015)). The price patterns we find for small firms may also be consistent with fake news producers sacrificing longer-term reputational capital in lieu of short-term gains (Allcott and Gentzkow (2017)). The decline in trading activity to all news, including legitimate news, following the public’s awareness of fake news from the SEC investigation is also consistent with Aymanns et al. (2017) and Allcott and Gentzkow (2017), that fake news increases distrust of media in general.7 The results may also be related more generally to the economics of norms and institutions such as trust and social capital (Guiso et al. (2004), Guiso, Sapienza, and Zingales (Guiso et al.), Guiso et al. (2010), Sapienza and Zingales (2012)).
Finally, Our setting is financial markets, and specifically shared-information platforms on financial news and opinions. There are reasons to be both cautious and optimistic on what we can learn about the impact of fake news more broadly. One of the benefits of financial markets is that we can quantify the influence of fake news through prices and trading activity. On the other hand, these information platforms may have little influence on markets either because they are unimportant or due to markets already incorporating the information. A 2015 study by Greenwich Associates found that 48% of institutional investors use social media to “read timely news.” On the other hand, fake news should not matter if markets are informationally efficient (Fama (1970)), regardless of the equilibrium asset pricing model. In this sense, our setting offers a unique test of market efficiency that circumvents the joint hypothesis problem: we run the flip side of the classic event study (Fama et al. (1969)) by examining market responses to fake news events and find (at least for small stocks) that trading activity and prices are affected. Our results provide more
7See also “Trust in Social Media Falls – Raising Concerns for Marketers,” by Suzanne Vranica, Wall Street Journal, June 19, 2018, which discusses research by Edeleman, the world’s largest public relations firm, that found trust in social media has fallen world-wide and particularly in the U.S. over the last year.
Fake News: Evidence from Financial Markets 7
evidence on the growing impact of crowd-sourced information platforms (Hu, Chen, De, and Hwang (2014)). If fake news can impact U.S. equity markets, where there is competition for information and arbitrage activity exists, then it could have even greater influence in settings where information costs are high and the ability to correct misinformation is more limited, such as online consumer, political, and social shared-information networks.
The rest of the paper is organized as follows. Section 2 details our first two empirical settings: the sample of fake news articles obtained from the SEC and the broader set of all articles from the shared-information platforms that we apply a linguistic algorithm for assessing fake content. Section 3 analyzes the direct impact of fake news through investor trading activity and price impact. Section 4 describes the third empirical setting – an event study of the SEC’s investigation that provides a shock to public awareness of fake news – to measure the indirect effects of fake news on the market’s reaction to news generally. Section 5 considers the motivation of fake news. Section 6 concludes.
2. Identifying Fake News
We detail our first two empirical settings using articles from knowledge-sharing financial platforms. We first describe these platforms and the data we obtain. We then describe our first sample of fake articles from the SEC. Using this sample, we validate and calibrate a linguistic model for identifying probable fake content. We then apply this model to the broader sample of articles with unknown authenticity from the same media platforms that generates our second empirical setting.
2.1. Knowledge Sharing Platforms
Our sample of articles comes from the two largest financial crowd-sourced platforms: Seeking Alpha and Motley Fool. Seeking Alpha is an online news service provider for fi- nancial markets, whose content is provided by independent contributors. The company has had distribution partnerships for its content with MSN Money, CNBC, Yahoo! Finance, MarketWatch, NASDAQ and TheStreet. The Motley Fool is a multimedia financial-services
Fake News: Evidence from Financial Markets 8
company that provides financial advice for investors also through a shared-knowledge plat- form. We obtain the articles posted on these platforms, including their content, authorship, and in the case of Seeking Alpha, commentary from other users. Appendix A details how authors on these sites contribute and are compensated for their articles.
The popularity of these sites has grown considerably. Seeking Alpha grew from two mil- lion unique monthly visitors in 2011 to over nine million in 2014, generating 40 million visits per month. While these platforms allow for the ‘democratization’ of financial information production, concerns have been raised about their susceptibility to fraud, since they are virtually unregulated, frequented predominantly by retail investors, and authors on these platforms can use pseudonyms (though the platforms claim they know the true identity of each author, in case that information is subpoenaed by the SEC, which it was in the cases we examine below). Authors are allowed to talk up or down a stock that they are long or short, provided they disclose any positions they have in the stock in a disclaimer accompanying the article. Failure to disclose can have legal ramifications. 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 underwriter.8 We exploit a subset of fraudulent articles on these platforms, identified by an undercover investigation of paid-for articles of false content and eventual SEC prosecution for our first empirical setting.
2.2. “For-Sure” Fake Articles
We examine a unique dataset of articles whose authors were paid to write fake articles,
but where the authors illegally did not disclose payment. The articles are obtained from
an industry insider, Rick Pearson, who, as a regular contributor to Seeking Alpha, was
approached by a public relations firm to promote certain stocks by writing articles with false
information for a fee without disclosing the payment. Mr. Pearson instead went undercover
to investigate how rampant this practice was on these platforms and uncovered more than one
8In June 2012, Seeking Alpha announced it would no longer permit publication of articles for which compensation had been paid.
Fake News: Evidence from Financial Markets 9
hundred fake, paid-for articles by other authors who did not disclose their compensation. He turned the evidence over to the SEC, who investigated each of these cases. The fake articles were subsequently taken down by the platforms once the SEC informed them of the investigations. The SEC filed its first lawsuit pertaining to these articles on October 31, 2014, prosecuting the authors, promotion firms who paid them, and in some cases the companies and their executives who hired the promotion firms.9
Mr. Pearson kindly provided the articles to us that he determined to be fake: 111 fake articles by 12 authors covering 46 publicly traded companies. We also obtained a second set of known or, as we will refer to them, “for-sure” fake articles, which the SEC identified during the investigation containing paid-for fake content.10 We also contacted Seeking Alpha, and they kindly shared 147 of those articles with us. Among those articles, we match 60 to firms publicly traded on U.S. exchanges to obtain price and volume information from Center for Research in Security Prices (CRSP). The rest of the articles pertain to firms traded over the counter. Combining all of the data sources, our final sample of for-sure fake articles consists of 171 articles written by 20 different authors about 47 publicly traded firms.11
It is important to define what we mean by fake articles. In this smaller sample from
Rick Pearson and the SEC, the fake articles are those that were paid for by a promotional
firm and not disclosed, and many of the authors admitted that the articles were written to
deceive the market and manipulate the stock price. Consequently, these articles contained
false information that authors thought to be incorrect at the time. How false or wrong that
information turns out to be is difficult to assess. For example, an article could intend to
deceive by embellishing the prospects of the firm, but could turn out to be mostly correct
in that assessment ex post. In other instances, the deception may be grossly off. Hence, our
9Subsequent lawsuits were also filed on April 10, 2017 and September 26, 2018. See filing documents at: securities.stanford.edu/filings-documents/1051/GBI00_01/20141031_r01c_14CV00367.pdf; www.sec.gov/litigation/complaints/2017/comp23802-lidingo.pdf.
10The full list can be found here: ftalphaville-cdn.ft.com/wp-content/uploads/2017/04/10231526/Stock-promoters.pdf.
11While we gain 60 additional articles from the SEC, we only gain one additional firm. Most of the additional articles pertain to firms already covered by Rick Pearson, and hence simply give us more fake articles about the same firms, with only one new firm identified.
Fake News: Evidence from Financial Markets 10
fake articles are about intent to deceive and not necessarily about whether they are right or wrong ex post. We focus on authenticity and not accuracy. Articles may be fake and (mostly) right, as well as fake and (very) wrong. Some of our analysis on the language used in the articles and on their impact on stock prices may help distinguish between these two cases, where we conclude that most of the articles perpetuated false information.
To provide some insight into the content of these fake articles, we highlight a recent example from our sample that was prosecuted by the SEC in September 2018. One of the fake, paid-for articles in this case was a publication that appeared on Seeking Alpha on September 26, 2013 about the company Biozone. The article stated,
Biozone has developed a new method of drug delivery, QuSomes that provides improved efficacy, reduced side effects, and lower costs. This technology will allow Biozone to reformulate and sell certain FDA approved drugs at a reduced cost, which should help Biozone capture a large percentage of these drug markets.
From the SEC lawsuit filed in September 2018 in the District Court of New York City:
Keller misleadingly stated that Company A had a formulation ready for testing to be brought to the billion-dollar injectable drug market. Yet, as Keller knew, as of summer 2012, all R&D efforts had been shut down without the successful formulation of an injectable drug and Company A had ceased all efforts to develop this technology in mid-2012.
Keller refers to Brian Keller, the co-founder and Chief Scientific Officer of Biozone, who had paid for the promotion article. Many of the fake, paid-for articles from the SEC involve similar issues and often coincide with a public relations campaign orchestrated by the firm to artificially prop up the stock price, including press releases, filings, and corporate actions. We investigate these issues more below and attempt to control for these actions when assessing outcomes on trading activity and prices.
We compare the fake, paid-for articles to a set of non-fake articles by the same authors. Specifically, we obtain a sample of other articles written by the same 20 authors under the SEC’s investigation that were not paid for by a PR firm and having no evidence of being false, totaling 334 additional articles about 171 companies published on the same platform
Fake News: Evidence from Financial Markets 11
Seeking Alpha. We use this set of non paid-for articles by the same authors to provide a clean comparison to the fake articles to control for heterogeneity across authors. We refer to these non-paid for articles as “non-fake” following our definition above and make no statement about the accuracy of the articles themselves. We examine the impact of the fake versus non-fake articles on markets.
2.3. Assessing Authenticity in the Broader Set of Articles
Our second empirical setting uses the broader set of all articles written on these platforms and attempts to assess their probability of being fake. While our unique sample of fake articles from the SEC provides unambiguous fake news, the sample is small and therefore may raise external validity concerns that may make it difficult to draw general conclusions. To complement these data, we manually download all articles published on Seeking Alpha and Motley Fool, obtaining 203,545 articles from Seeking Alpha over the period 2005 to 2015 and 147,916 articles from Motley Fool from 2009 to 2014.
2.3.1 How do you tell if someone is lying?
The downside of this much larger dataset of 351,461 articles is that their authenticity is unknown. To assess authenticity, we develop a probability function for detecting fake content using an objective and scalable measure based on quantifiable research from linguis- tics. Specifically, we use a linguistic algorithm designed to detect deception in expression – the Linguistic Inquiry Word Count model (LIWC2015) from Pennebaker et al. (2015) – which focuses on individuals’ writing or speech style, and appears adept at measuring indi- viduals’ cognitive and emotional states across various domains. The LIWC model outputs the percentage of words that fall into one of more than 80 linguistic, psychological, and topical categories, one of which is the authenticity score that detects deception in expres- sion. While the exact formula for the authenticity score is proprietary, Pennebaker (2011) describes which linguistic traits are associated with honesty. In particular, truth-tellers tend to use more self-reference words and communicate through longer sentences compared to
Fake News: Evidence from Financial Markets 12
liars. When people lie, they tend to distance themselves from the story by using fewer “I” or “me”-words. Furthermore, liars use fewer insight words such as realize, understand, and think, and include less specific information about time and space. Liars also tend to use more discrepancy verbs, like could, that assert that an event might have occurred, but possibly did not. Newman et al. (2003) use an experimental setting to develop an authenticity score based on expression style components using similar techniques, and the Central Intelligence Agency and Federal Bureau of Investigation have used similar methods to assess authenticity in speech or writing.
For example, consider the two statements of former U.S. congressman Anthony Weiner before and after his admission in the “sexting” scandal.
Before admission:
We know for sure I didn’t send this photograph. [...] We don’t know where the photograph came from. We don’t know for sure what’s on it. [...] If it turns out there’s something larger going on here, we’ll take the requisite steps.
After admission:
I would like to make it clear that I have made terrible mistakes, that I have hurt the people I care about the most, and I am deeply sorry. I have not been honest with myself, my family, my constituents, my friends, my supporters and the media.
The use of “we” versus “I” and “my”, the discrepancy words “don’t know” and “if”, and the lack of insight words like “mistakes,” “clear,” and “hurt” are all more prevalent in his statements when he was lying to the public.
The algorithm uses a combination of these linguistic traits to generate the authenticity measure. A unique and critical advantage of our study is that we use the for-sure fake articles from Rick Pearson and the SEC to validate the linguistic algorithm and calibrate the authenticity score into a probability of fake news. One of the major challenges to studying this issue is the lack of known fake content. Since the LIWC authenticity score was not developed in the context of financial media, it is useful to assess its ability to distinguish fake from non-fake articles in our context. Financial blogs and articles tend to point to
Fake News: Evidence from Financial Markets 13
facts, trends, and figures, which may be decidedly different from narratives that were used to develop the linguistic algorithm. Our unique sample of 171 for-sure fake articles and 334 non-fake articles written by the same authors, provides a validation and test of the generalizability of the linguistic algorithm. We compare the LIWC authenticity score, which is normalized between 0 and 100, for the two samples and control for author fixed effects to capture heterogeneity in author style, content, or reputation, and any matching of authors to types of articles.
Panel A of Table 1 reports the difference in the LIWC authenticity scores for the fake and non-fake samples. Relative to an average authenticity score of 33 for non-fake articles, fake articles have a much lower average score of 19 (statistically significant at the 1% level). A plot of the distribution of authenticity scores for fake and non-fake articles in Figure 1, Panel A highlights the differences, controlling for author heterogeneity since we examine fake and non-fake articles for the same authors. Panel B of Figure 1 provides more specific examples for two authors: John Mylant and Equity Options Guru. The distribution of authenticity scores across fake and non-fake articles for each author are quite different. While some of the non-fake articles also have low authenticity scores, most of the fake articles have very low authenticity scores.
Panel A of Table 1 reports summary statistics on language characteristics associated with authenticity as described in Pennebaker (2011) for the for-sure fake and non-fake articles. We report the average use of 1st person singular (examples: I, me, mine), Insight (examples: think, know), Relativity (examples: area, bend, exit), Time (examples: end, until, season), Discrepancy (examples: should, would), and the average number of words per sentence. According to Pennebaker (2011) and Pennebaker et al. (2015), when people lie they also tend to use fewer words per sentence. Fake articles appear less authentic, having fewer self- referencing, lower insight, lower relativity, and higher discrepancy scores on average. These findings provide an out-of-sample test of the LIWC algorithm that validates it in a unique setting, which would not be possible without the sample of known fake articles from the
Fake News: Evidence from Financial Markets 14
SEC.
2.3.2 Probability of Being Fake
The sample of for-sure fake and non-fake articles also allows us to calibrate the au- thenticity scores into a probability of fake content. While the LIWC authenticity score is statistically different between fake and non-fake articles, it is not easy to interpret the car- dinal nature of the score – what does a 14 point difference in authenticity score mean? To provide a more direct interpretation of the results and their economic meaning, we develop a mapping of the authenticity score into probability space. Again, this exercise is only possible because we have a set of known fake articles from which to calibrate probabilities. Using the smaller sample of for-sure fake and non-fake articles, we map the authenticity score into the frequency of fake articles and apply Bayes rule to the broader sample of Seeking Alpha and Motley Fool articles. We use the known fake and non-fake articles to map authenticity scores into a conditional probability of being fake.
Specifically, let S be the authenticity score and F (T) denote a fake (true) article. We compute P rob(S|F ) and P rob(S|T ), where, crucial to this exercise, we use the smaller valida- tion sample, where we know which articles are F and T , in order to measure the probabilities. From Bayes rule,
P rob(F |S) = P rob(S|F )P rob(F ) . P rob(S|F )P rob(F ) + P rob(S|T )P rob(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 is found by solving the fixed point problem in which the observed P rob(F ) in the sample is representative of Prob(F) in the overall population.
Figure 2 plots the mapping of LIWC authenticity scores (S) into the conditional prob- ability of being fake (Prob(F|S)) for the entire sample of 203,545 Seeking Alpha articles
Fake News: Evidence from Financial Markets 15
published between 2005 and 2015. The relation between the LIWC authenticity score and the probability of being fake is highly nonlinear. Specifically, the sharp increase in probability in the very low authenticity range suggests that articles may be more efficiently and better classified into fake and non-fake using a probability cutoff. We use a cutoff of Prob(F) > 0.20 to classify articles as being fake and classify articles with Prob(F) < 0.01 as being non-fake, with the remaining articles (0.01 ≤ P rob(F ) ≤ 0.20) being classified as ambiguous or “other.”12 This cutoff implies an authenticity score that is even lower (about half) than the average authenticity score for the known fake articles in the SEC sample. Hence, this cutoff is conservative and designed to reduce type II errors.
We first examine how accurate our method is at identifying fake news from our small sample of 171 for-sure fake and 334 non-fake articles written by the same authors. We generate an authenticity score for each article, and calculate its probability of being fake. Our algorithm classifies 18 articles as being fake, of which 17 are actually fake, indicating that the Type II error rate is very low – one false positive. Our method is very conservative, however, since it misses a lot of fake articles – 154 fake articles are not classified as fake. Our algorithm identifies 165 articles as being non-fake. Of those, 17 are actually fake, implying a Type I error of about 10%, which is quite low considering our methodology is designed to minimize type II errors. However, our methodology is highly conservative in that 186 (334 - 165 + 17) non-fake articles are not classified. We exclude articles with 0.01 ≤ P rob(F ake) ≤ 0.20 from our analysis, since both Type I and Type II errors will be larger for these articles, which amounts to 322 (505-18-165) articles or about 64% of the sample not being classified.
These statistics highlight the challenges in identifying fake news. There is a tradeoff
in how confident we wish to be in our classification versus how many articles we wish to
classify. In our case, we choose to minimize the number of falsely identified fake articles
and falsely identified non-fake articles, so that when we look at their differential impact,
we are confident that we are measuring the difference between fake and authentic news.
12Our results are not sensitive to different cutoffs in the 0.10 to 0.30 probability range for fake, where 0.20 was chosen based on the steep increase in probability in Figure 2.
Fake News: Evidence from Financial Markets 16
This, however, means we are unable to classify many articles. If the goal is to capture as much fake content as possible, then we would choose a different threshold that would classify more articles but invite more classification error. This tradeoff highlights the limitations of the linguistic algorithm and echoes some of the challenges facing social media platforms in flagging fake content. For our purposes of identification, it is more important to minimize classification error at the expense of not being able to classify many articles.
Using our fake news probability model, calibrated to the sample of known, for-sure fake and non-fake articles, we apply our methodology to the broader sample of all articles. Table 1 Panel A shows summary statistics for the Fake, Non Fake, and Other articles identified by our algorithm. The number of articles in each category, the mean of the Authenticity measure that we use to construct the probabilities, and the components of that authenticity measure from the LIWC algorithm are reported. The difference in authenticity measures translates into large differences in the estimated probability of being fake from our calibrated function: the articles we identify as fake have an average 0.45 Prob(F) based on their authenticity score, while the average probability for articles we identify as non-fake is less than 0.01. Obviously, the articles were sorted based on the probabilities, but the magnitude of the difference is interesting and suggests substantial differences in authenticity scores between the two groups of articles, which Table 1 reports are 5.4 versus 50.7.
We also apply our methodology for identifying fake articles to another sample of articles from another crowd-sourced financial news platform – Motley Fool, where we have 147,916 articles from 2009 to 2014. Applying the LIWC algorithm, we obtain similar differences in authenticity scores and probabilities in classifying Motley Fool articles into Fake and Non- Fake. The unconditional probability of fake news on the Motley Fool sample is 2.7%, almost identical to the 2.8% we found for Seeking Alpha. Looking at the rest of the components of the authenticity score, the algorithm does a similar job on the Motley Fool sample.
Finally, as another validation exercise we analyze a particular set of articles written by a Motley Fool author, Seth Jayson, who has been working for Motley Fool full-time since 2004
Fake News: Evidence from Financial Markets 17
as a journalist, and has written over 31,000 articles. Mr. Jayson’s articles are a good test case because he works directly for Motley Fool and has for a long time. Hence, it is unlikely he has written fake articles on their platform and unlikely promotional firms would even approach him to do so. We, therefore, use his articles as a placebo test of our classification methodology. Using our methodology, we classify 18,361 of Mr. Jayson’s articles as reliably non-fake and only 2 of his articles as probabilistically fake (the rest being indeterminate). In other words, we classify 0.006% of his articles as fake, which is consistent with our prior that he would have essentially zero fake articles and suggests our conservative classification methodology works well at identification (though, again, the tradeoff is that many of Mr. Jayson’s articles cannot be classified by the methodology).
Finally, Panel C of Table 1 reports the average fraction of retail investors, the average number of analysts covering the firm, and the average firm size (in $US millions) for each article group. For-sure fake articles tend to cover small firms with a high fraction of retail in- vestors and low analyst coverage. The probabilistically determined fake and non-fake articles from the broader Seeking Alpha and Motley Fool samples exhibit more muted differences. Notably, the Motley Fool articles are written about significantly larger firms than Seeking Alpha, and the for-sure fake articles identified by Rick Pearson and the SEC are about tiny firms whose average market capitalization is only $7.4 million.
Table C1 in the Internet Appendix also examines whether fake articles tend to cluster in specific industries. We separate articles into one of the 12 Fama-French industries that the firms mentioned in the articles belong to. For the for-sure fake articles provided to us by Rick Pearson and the SEC, 81% are about firms in the Healthcare industry. This finding is not too surprising as these articles came from authors who were hired primarily by two PR firms that concentrated on the healthcare industry. For the non-fake articles, the majority of firms belong to Business Equipment, Healthcare, Finance, and Manufacturing industries. The industry composition of Fake and Non-Fake articles we identify on Seeking Alpha and Motley Fool using our algorithm is similar to the Non-Fake articles’ industry composition
Fake News: Evidence from Financial Markets 18
from the smaller sample of articles from the SEC.
3. Trading Activity and Price Impact
In this section, we examine the impact of fake news on trading activity and prices. Financial markets provide a useful setting to examine the impact of fake news, because they provide high frequency outcomes, such as trading volume and market prices.
3.1. Do Articles on the Platforms Have Impact?
Before looking at fake articles, we first address whether articles posted in general (fake or non-fake) on these social platforms have any influence on market participants or markets. We start by examining abnormal trading volume around the publication of all articles. We focus on abnormal volume because we are interested in whether investors “react” to these articles. Of course, it is also quite plausible that articles react to trading activity. To try to establish some causal interpretation, we only examine abnormal volume and look at future changes in it. Abnormal trading volume is trading volume relative to expected volume in the stock, which we proxy for using the recent average daily volume in the stock (defined below). We look at future changes in abnormal volume in the days following the article’s publication. A reverse causal story would need to imply that authors are writing articles in anticipation of future unexpected trading activity. If true, we would call that “news.” Nevertheless, we also control for lagged abnormal volume from the previous trading day and other news coming from the firm such as recent SEC filings and press releases about the firm. We also examine the articles’ effect on price volatility to address whether there is information in the articles that is not already impounded into market prices. Of course, quantities traded can vary significantly with no price movement (Fama (1970), Grossman and Stiglitz (1980)) or trade can be zero with substantial price movement (Milgrom and Stokey (1982)). Thus, we look at both quantities and prices separately.
Panel A of Table 2 examines the effect on abnormal trading volume from articles published
Fake News: Evidence from Financial Markets 19
140
on these sites. We define abnormal trading volume for stock i as V ol(i, t)/ 1 V ol(i, t−k),
T
k=20
which is the trading volume for stock i on day t relative to the average daily trading volume in stock i over the last 6 months (skipping a month).13 We sum abnormal volume over days t = 0, t+1, and t+2, where t = 0 is the date the article appears on the website and then regress the natural logarithm of abnormal volume on an indicator variable for whether there is any article on these sites about the firm on a given day, regardless of its authenticity. We also include year-month fixed effects in the regression. We examine only firms that had at least one article published on Seeking Alpha or Motley Fool over the sample period.
As the first column of Panel A of Table 2 shows, an article published on Seeking Alpha or Motley Fool is associated with an 86% increase in abnormal trading volume over the three days following publication. This result implies either that investors are trading in direct response to the articles or, more generally, are trading in response to whatever news is coming out that day that these articles may be discussing. Likely both. While the increase seems large, this first regression controls for no other variables, except time fixed effects. Moreover, as we show below, the bulk of the effect is concentrated in very small and illiquid firms, where trading volume changes (as a percentage) can be enormous, and where outliers can be in the thousands of percents.
As we will show, articles on knowledge sharing platforms in the SEC-prosecuted cases are often written following press releases or SEC filings. In the second column of Panel A, we control for whether there is an SEC filing (10-K, 10-Q, or 8-K) or a company-issued press release in the three days leading up to the article. Furthermore, to control for serial correlation in abnormal trading volume, we include lagged abnormal trading volume on day t − 1 as a regressor, which also captures other events we may be missing that could affect trading activity. After the controls, the effect on abnormal trading volume declines to 37%. To make sure these results are not all coming from the day the news is released, Table ?? in
the internet appendix reports the effect on trading volume separately for the same day and
13Results are identical defining abnormal volume relative to the last 30, 60, or 180 days.
Fake News: Evidence from Financial Markets 20
for one and two days after the article’s publication. Of the 37% rise in abnormal trading volume, 15.5% occurs on the day the article is published, 12.1% the following day, and 10.1% two days later. Figure 4 in the internet appendix plots the abnormal volume response for the next 20 trading days and finds that abnormal volume increases for about two weeks.
The next three columns of Panel A of Table 2 report results separately for small, medium, and large firms. Small firms are defined as smaller than the bottom 10th percentile of NYSE firms, mid-size firms fall in the 20th to 90th size percentile of NYSE firms, and large firms are in the top 10th size percentile of NYSE firms. The effect on abnormal trading volume declines strongly with firm size, with the effect six times larger for small firms than for large firms (80.9% increase versus 8.2% increase). This result is consistent with small firms having more retail investor trading and perhaps a more opaque information environment. In the last two columns, we separate firms into high and low retail ownership (above or below median retail ownership last month) and find that the effect on abnormal trading volume is twice as large for firms with high retail ownership.
The internet appendix reports some robustness tests of these results. First, Table C3 includes firm fixed effects to difference out any unobservable firm heterogeneity over the sample period – the results are nearly identical. Second, Table C4 controls for some outliers in percentage change in volume by winsorizing the most extreme five percent of abnormal volume observations. Since abnormal volume is the dependent variable, it is always question- able to winsorize, unless we think these extremes are data errors. The effects are obviously more muted from winsorizing but the patterns stay the same: a 33.3% increase in abnor- mal volume over the next three days that is larger for small firms (54.1%) and high retail ownership firms (41.7%). Our findings are not driven by a few extreme observations. The point estimate for the full sample is roughly the same, but the effect from winsorizing on the smallest decile of firms, where percentage volume changes are more extreme, reduces the effect from 80.9% to 54.1%. Given the similar patterns and the fact that we do not believe the extreme volume changes are errors, we do not winsorize observations.
Fake News: Evidence from Financial Markets 21
Panel B of Table 2 repeats the regressions in Panel A, replacing abnormal volume as the dependent variable with the idiosyncratic price volatility of the stock. We measure id- iosyncratic volatility as the square of the difference between the return of the stock and a matched-portfolio of stocks with similar size, book-to-market equity (value) and past 12- month returns (momentum) following the procedure of Daniel, Grinblatt, Titman, and Wer- mers (1997), which forms 125 equal-weighted portfolios based on 5 × 5 × 5 sorts of stocks using size, value, and momentum characteristics that are related to expected returns. The dependent variable is the sum of daily idiosyncratic volatility on the day the article is pub- lished plus the next two days. This analysis captures whether articles moved prices around the days they were published. We examine price volatility as opposed to returns because it is exceedingly difficult to sign the direction of the content of the articles.14 Hence, looking at volatility or the absolute value of returns captures whether prices moved significantly in relation to the articles published on that day. If the market has already incorporated the news, then the expected absolute return change should be zero. As Panel B reports, we find effects similar to those in Panel A that examine trading volume – daily price volatility or the absolute return of the stock rises following articles published on these platforms, even after controlling for recent SEC filings, firm press releases, and return volatility in the days leading up to the article. The effect is strongest for smaller firms with higher retail ownership. The magnitude of these effects is large but not unreasonable. Across all articles, the effect of a published article on idiosyncratic stock volatility is about 6.8% over the three days, which is roughly an additional 40% of the stock’s normal price movement on days when there is no news about the firm on these platforms. For the smallest firms, the effect is an order of magnitude larger, which is consistent with extreme price movement for the smallest stocks (Frazzini, Israel, and Moskowitz (2017)).
14Textual analysis used to derive sentiment (Antweiler and Frank (2005), Tetlock (2007), Das and Chen (2007), Jegadeesh and Wu (2013), Heston and Sinha (2017), Boudoukh et al. (2018)) is notoriously challeng- ing and noisy. In addition, price movements are deviations from expectations, so a “positive” article that is less positive than expected would predict a negative return. Not knowing expectations makes signing the price movement even more difficult.
Fake News: Evidence from Financial Markets 22
3.2. Impact of Fake Articles
Do fake articles have a differential impact on trading volume and volatility? Panel A of Table 3 reports results from the same regressions as Table 2, but includes a dummy variable for whether the article is “for-sure fake” from our SEC sample. As the first column of Panel A shows, the for-sure fake articles are associated with significant increases in abnormal trading volume, which is not too surprising since the SEC is more likely to target cases that had more impact. The magnitude of the impact of fake articles is another 50% increase in trading volume over the next three days relative to a non-fake article. This seems particularly large, butisreasonableaccordingtoexcerptsfromthemostrecentSEClawsuits.15 Thesecondand third columns of Panel A interact the fake article dummy with the market equity decile of the company the article is written about and the retail ownership percentage of the firm. The impact on trading volume is larger for smaller firms and firms with higher retail ownership, though the latter interaction is insignificant.
The first three columns of Panel B of Table 3 report results from the same regressions using idiosyncratic volatility as the dependent variable. Consistent with the abnormal trad- ing volume results, fake articles have an additional significant impact on price movements relative to non-fake articles. The magnitude is large, too, which is not terribly surprising as this sample is based off of the SEC investigations, which are chosen ex post (where the SEC is more likely to go after cases where the promotional articles had massive impact on prices).
3.3. Using Probabilistically Fake Articles
The last three columns of Panels A and B of Table 3 report results using a dummy variable
for whether the article is probabilistically fake using our calibrated probability function for
fake news. As Panel A shows, the coefficient on the LIWC Fake dummy is insignificant
15From Case 1:18-cv-08175 filed on September 7, 2018 in the U.S. District Court, Southern District of New York: “The market reacted strongly to the Company A promotion: the trading volume of Company A stock rose from approximately 1,100 shares on September 25, 2013 to over 4.5 million shares on September 27, 2013 and to more than 6 million shares on October 2, 2013.” And, in the same case about another firm, “The article did not disclose that the author had been paid by Company B – at Honig’s direction – to write the article. After the article was published on February 3, 2016, there was a 7000% increase from the previous day’s trading volume, and an intraday price increase of over 60%.”
Fake News: Evidence from Financial Markets 23
by itself (though of positive sign), but interacting it with size deciles and retail ownership delivers significant effects on trading volume for the smallest firms with the highest retail ownership. These results are consistent with those from the narrower set of for-sure fake articles. Panel B shows that the probabilistically fake articles have a more muted impact on stock volatility, producing the same signed coefficient we get from the for-sure fake sample, but where nothing is statistically significant. The LIWC algorithm appears to capture some of the fake content that may exist on the platforms, but is a noisy measure and illustrates the difficulty in identifying fake news.
3.4. More Evidence on Direct Impact
To further test a direct link between articles published on these platforms and trading activity, we obtain a proprietary supplemental dataset from Seeking Alpha on readership of articles. The data only covers calendar year 2017, but contains daily number of “clicks” (i.e., number of times a given article is uploaded) and the number of times the article is “read,” which is the instances in which a reader scrolled to the end of the article.16 In total, the dataset covers 25,596 articles about 3,118 publicly traded firms.
Table C5 in the internet appendix presents results from regressing abnormal trading volume following the release of the article on the readership circulation of the article over the first three days after the article is published. The table shows that future abnormal trading volume is positively related to the number of clicks and number of times the article is read by consumers. This evidence suggests that the articles are directly related to future abnormal trading activity in the stocks the articles discuss.
The last two columns report results from regressions of the readership circulation variables
on the fake article dummy to examine whether readership is affected by article authenticity.
We find that fake articles are clicked more heavily and read more heavily, consistent with
those articles also affecting trading volume more. Fake news seems to disseminate faster
16We obviously do not know if it was actually read, but scrolling through the article implies that some time was spent on it.
Fake News: Evidence from Financial Markets 24
and more widely and impacts trading activity more. These results are consistent with fake news being more sensational and more persuasive, catering to the biases and priors of their consumers, and propogating more diffusely through the network as suggested by Allcott and Gentzkow (2017) and Vosoughi, Roy, and Aral (2018).
4. A Shock to Investor Awareness of Fake News
Our third empirical setting examines a different aspect of fake news to test another theoretical implication. We use the announcement of the initial SEC investigation into the promotional articles that comprise our first sample as an exogenous shock to the public awareness of fake news and examine the market’s response to news before and after this shock. This exercise does not require being able to detect fake content.
4.1. Galena Biopharma Inc.
We begin with the case of Galena Biopharma Inc., which was the first prosecuted by the SEC for stock price manipulation on knowledge-sharing platforms. Galena encompasses the “event” which made the public aware of the existence of fake news. It also provides a microstudy of the direct impact these articles have on the stock’s trading activity and prices as well as some of the motivation behind fake articles.
On October 31, 2014 the SEC filed a lawsuit in the United States District Court on behalf of all persons who bought Galena’s common stock between August 6, 2013 and May 14, 2014.17 Figure 3 depicts the stock price of Galena from April 2013 to May 2014, as well as the events that led 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 paid for. The articles did not disclose the payments that the authors received, which violated the terms of Seeking Alpha and SEC regulation, and in some cases falsely claimed not to have received any payment. The lawsuit documents
at least twelve promotional articles of this type. Appendix B contains an example of one of
17(Case 3:14-cv-00558-SI): securities.stanford.edu/filings-documents/1051/GBI00_01/20141031_r01c_14CV00367.
Fake News: Evidence from Financial Markets 25
the fake articles written about Galena.18
Figure 3 shows that Galena’s share price rose from about $2 to $7.48 between the summer
of 2013 and January of 2014. The publications of the fake articles are highlighted on the graph by the green boxes and often coincide with a bump in stock price on that day and a steady increase in price several days after. The motivation behind the scheme seems to have been a pump-and-dump campaign, as Galena insiders took advantage of the price rise through corporate actions and their own personal trading. On September 18, 2013 Galena sold 17,500,000 units of stock in a seasoned equity offering for net proceeds of $32.6 million. On November 22, 2013, Galena held a board meeting and granted stock options to executives and directors with a strike price of $3.88. In January 2014, after the stock price reached its highest level since 2010, seven Galena insiders sold most of their stock in less than a month, for more than $16 million. These events are highlighted in Figure 3, where as news of insider sales broke, the stock price declined dramatically.
In February and early March 2014, several investigative journalists published exposé ar-
ticles documenting the fraud, including in Barron’s and Fortune. On March 17, 2014 Galena
revealed in a 10-K filing that it was the target of an SEC investigation over the promotion.
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. Appendix A reports the 8-K form documenting the settlement. By that point
Galena’s stock price had dropped to $2 a share.19
18This article and others like it that are part of the SEC investigation have been removed from Seeking Alpha. Searching for this fake article today, Seeking Alpha displays a message stating: “This author’s articles have been removed from Seeking Alpha due to a Terms of Use violation.”
19Interestingly, while Galena is a relatively small firm, it was not an obscure one. For example, in July 2013, before the promotion started, it had a market cap of 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 facilitated trading in the company’s stock.
Fake News: Evidence from Financial Markets 26
4.2. A Shock to Awareness of Fake News
The public revelation of the SEC’s investigation and subsequent media attention around it provides a unique shock to investor awareness of fake news. We exploit the timing of the announcement to test additional implications of fake news.
In addition to the direct costs of individuals believing and acting upon false content, fake news can be costly if it damages people’s trust in news generally and causes them to discount legitimate news (Allcott and Gentzkow (2017), Kshetri and Voas (2017), and Aymanns et al. (2017)). Our unique setting provides an opportunity to measure the potential spillover effects of fake news on people’s trust in news. Using the revelation of the SEC investigation, we examine whether investors behaved any differently before versus after the event, when the presence of fake news on knowledge sharing platforms suddenly became salient to many consumers on these platforms.
4.3. Spillover Effects from Fake News
We use the period from February to March 2014 as the event that provides a shock to people’s awareness of fake news. We examine the propensity of fake news and abnormal trading activity associated with articles six months prior to and six months after the event (August 2013 to January 2014 and April 2014 to September 2014, respectively).
Panel A of Table 4 first examines whether the propensity of fake news declines after the scandal. We regress a dummy variable of whether the article was probabilistically fake, on a dummy for 6 months after the SEC announcement event, controlling for SEC filings, firm press releases, and lagged abnormal volume in the days leading up to the article’s publication. In addition, we include the number of news articles about the firm from the NYT and WSJ, which we obtain from Factiva. The coefficient on the post-scandal period is indistinguishable from zero, indicating that the prevalence of fake news, or more precisely the authenticity score of the fake articles, is similar before and after the scandal. However, this average result masks substantial heterogeneity. The next three columns separately report results for small, midsize, and large firms (defined as the smallest 10%, middle 80%, and largest 10% of firms,
Fake News: Evidence from Financial Markets 27
respectively, based on NYSE market cap breakpoints). The prevalence of fake articles about small firms fell significantly by 1.2% following the scandal. These results are consistent with small companies, who engaged or were willing to engage in promotional articles before the scandal, ceasing or decreasing this activity after the SEC announcement.
Panel B of Table 4 examines the impact of published articles on abnormal trading volume before versus after the scandal. The first column of Panel B reports results from a regression of abnormal volume on an article indicator, the 6-month post event indicator, and their interaction. The positive coefficient on articles confirms our earlier result from Table 2 that articles are associated with larger trading volume in the three days after they are published. The negative interaction term with the post-event dummy shows, however, that the effect of articles on trading volume decreases significantly after the scandal. This result is consistent with investors becoming aware of fake content and muting their trading response to news on these platforms. In addition, the strong negative coefficient on the post-event dummy indicates that abnormal trading volume, in general, declines after the scandal. This result suggests that people responded less to news in general on these platforms, including legitimate news, after the scandal and is consistent with consumers having less trust of news once aware of the existence of fake news, as theory suggests (Alcott and Gentzkow (2017)). The economic magnitude of the effect is large – a 4.2% drop in trading volume associated with news articles after the scandal relative to before the event.
Figure 4 examines the daily abnormal trading volume response for four trading weeks after the article is published. We estimate the following model:
Log(AbVol)t =α+β1Article×PostEvent+β2Article+β3PostEvent+Controls+ε
and plot the coefficient β1 at the daily level, with 95% confidence error bars. The graph displays the average trading volume reaction to all articles, after the scandal, and shows significant trading decreases on the day the article is published, and for the next two trading
Fake News: Evidence from Financial Markets 28
weeks, before eventually returning to pre-scandal levels. This result suggests that investors’ reaction to articles on these platforms decreases after the scandal, and does not rebound with higher trading volume at a later date.
While the results in the first column of Panel B control for the level of SEC filings, press releases, other media (e.g., WSJ, NYT articles), and lagged abnormal trading volume in the days leading up to the article’s publication, in the second column, we also interact the frequency of SEC filings, firm press releases, other news media, and changes in abnormal trading volume with the post-scandal dummy. The interaction terms serve as falsification exercises or “placebo” tests of the market responding to news on these platforms and the shock of fake news awareness on the platforms. In particular, an alternative explanation for the decline in trading volume in response to news after the scandal is that there is less information content, less news, or less firm activity in the post-event period that happened by chance to coincide with the timing of the SEC announcement. Or, perhaps, the trading volume response to news generally declines over time and being confounded by the SEC event. In either case, the interactions between corporate filings, press releases, and other media news would be negative as well. As the table shows, however, the interaction terms are negligible and insignificant, and two out of three have the wrong sign to be consistent with this alternative story. The magnitudes of these interactions with the post-event dummy are trivially small – 0.2% increase in trading volume response to SEC filings, 0.3% decrease to press releases, and 0.6% increase to other news media – none of which are remotely statistically different from zero. We find no discernible difference in firm news or activity before versus after the scandal and no reliable difference between the trading volume response to WSJ or NYT articles before versus after the scandal. This despite the fact that SEC filings, press releases, and other media news from the WSJ and NYT by themselves have a significant impact on trading activity: SEC filings, press releases, and newspapers articles increase abnormal trading volume by 13%, 29%, and 12.5%, respectively. However, after the scandal we find no difference in response to these other sources of news. We only find a
Fake News: Evidence from Financial Markets 29
decreased response to articles published on the social platforms.
Hence, the drop in trading volume associated with published articles on these social
media platforms is likely a reduced response from investors to news specifically coming from these platforms, and not press releases, other public filings, or other media, and not any market trends in information production or lower trading activity. The evidence is most consistent with investors discounting all news on these platforms, even legitimate news, after the scandal due to increasing distrust of content from these platforms after the SEC revealed the existence of some fake articles. The magnitude of the drop in abnormal volume is even larger and more significant after accounting for the other activity post-scandal, decreasing volume by 7.5% per article after the event. These findings provide some of the first evidence on the indirect spillover effects of fake news on news in general, as conjectured by theory (Allcott and Gentzkow (2017)). As we will show below, consumers were largely unaware of and unable to detect fake news, consistent with their response to discount all news on the platforms following their awareness of fake content.
Columns 3 through 7 of Panel B report results separately for small, medium, and large firms, as well as for firms with high and low retail ownership. Consistent with our previous results, these effects are all much stronger for smaller firms, and firms with high retail ownership. Post scandal, the abnormal trading volume associated with articles published on these platforms drops by 35% for the smallest firms. Interestingly, even though few fake articles are written about large firms and none of the firms in the SEC probe were large firms, abnormal trading volume still declines by 11.7% for each published article about large firms that appeared on these platforms after the scandal, despite nearly all of these articles being authentic. This result provides further evidence of a spillover effect from fake news to other legitimate news content. Stocks with high retail ownership have a 30.3% drop in abnormal volume post-scandal compared to only a 9.5% drop for low retail ownership stocks. Since retail investors tend to dominate participation on these sites, this result provides a more direct link to these platforms influencing trading activity.
Fake News: Evidence from Financial Markets 30
4.4. Generalizing Spillover Effects
The spillover effect from the awareness of fake news to all news, including legitimate news, is interesting and consistent with theory (Allcott and Gentzkow (2017)). The result begs the question: How broadly does the awareness of fake news from the scandal affect investors’ response to news generally? Was the spillover response merely contained to similar articles on Seeking Alpha, where many of the promotional articles the SEC investigated resided, or did it impact news from other sources? We can first look at our previous results from the falsification tests. For SEC corporate filings, press releases, and other news media (namely, the WSJ and NYT), we do not find a reliable spillover effect from the public’s awareness of fake news from the SEC investigation on the market’s trading activity response to news from these other outlets. The insignificant interaction effects between filings, press releases, and other news media with the post-scandal period suggests that investors respond to these sources in the same manner after their awareness of fake news on the social platforms. Thus, the drop in trading response only for the social platform articles is consistent with investors discounting all news on the social platforms, but recognizing or believing press releases, the WSJ, and NYT, are less subject to fake news, or that the average investor who trades on press releases is different than the average investor who trades on news from these social platforms. For these reasons, we think the filings, press releases, and other news media interactions provide compelling falsification tests that support our main findings.
While the indirect effects of fake news on these platforms do not seem to spillover to press releases or the WSJ or NYT, we find that they do spillover to other similar media outlets that were not part of the scandal. Specifically, as a test to generalize the spillover effect from fake news on other news more generally, we examine the trading response after the scandal for articles on the Motley Fool platform only. Motley Fool was not part of the fake news scandal, and none of its articles were flagged for failing to disclose paid-for content as part of a promotional scheme or were investigated by the SEC. Hence, we examine whether the spillover effect from the scandal, contained largely on Seeking Alpha, also had
Fake News: Evidence from Financial Markets 31
an effect on the trading volume response to articles published on Motley Fool, a similar shared-knowledge platform that was not part of the investigation. The last column of Panel B of Table 4 reports the results and shows, interestingly, that abnormal trading volume declined significantly for Motley Fool articles after the scandal. The result points to the spillover effect from the scandal extending beyond the specific platform where the scandal occurred. The awareness of fake news seems to impact other related news sources – in this case a competitor shared-information platform where the scandal did not occur. But, as our previous analysis shows, it does not have an impact on very different news sources such as press releases or newspaper articles, such as the WSJ and NYT. These results make sense if investors simply discount all social news as a result of the scandal but think other news sources are more immune to false content, or if the set of investors who consume social news is simply different from those who consume other news sources.
Panel C of Table 4 reports the results from the same regressions as in Panel B, but uses idiosyncratic volatility as the dependent variable instead of abnormal trading volume. The results are consistent with the trading volume findings, where there is significantly reduced impact on price volatility from articles after the information shock from the scandal, especially for small firms with high retail ownership. We also find opposite-signed price movement for press releases, SEC filings, and other news media after the scandal, suggesting other trends or omitted variables in the market’s response to news in general are not driving the results. We also find that price movements on the Motley Fool articles only are consistent after the scandal with the reaction to articles on Seeking Alpha, providing further evidence of a spillover effect to other shared-information platforms as a result of fake news awareness. These findings are consistent with markets discounting news from these platforms after revelation of the existence of fake news.
4.5. More Direct Evidence of Spillover Effects
We provide some additional evidence that the decline in trading volume per published article, and spillover decline in volume for non-fake news after the scandal, is due to investors
Fake News: Evidence from Financial Markets 32
being made aware of fake news. Specifically, we examine the posted comments to the articles published on these sites in the six months before and after the scandal. In the comments section pertaining to each article, we add up the mention of the words “fake” or “fraud” and compute a variable Fake Words, which is a dummy equal to one if readers use these words. We then regress the frequency of Fake Words on a dummy for fake articles as well as a dummy for the six-month period after the scandal.
To test an alternative hypothesis, we also compute the frequency of the words “wrong” or “not right” from the comments section and create a dummy variable Wrong Words, which is equal to one if readers use these words in their comments. This variable helps distinguish between erroneous or inaccurate information from fraudulent or deceptive information. The distinction is subtle because it relies on intent. The comments section provides a glimpse of what consumers may be concerned about.
Panel A of Table 5 examines whether the appearance of Fake Words or Wrong Words is more prevalent for fake versus non-fake articles over the entire sample period. We regress the prevalence of Fake Words on the fake article dummy in the first column and find that the words “fake” or “fraud” are not used more frequently with fake articles. This null result suggests that participants on these platforms could not identify or differentiate between fake and non-fake articles. In our setting, participants on these platforms were deceived by these articles with no indication that consumers were skeptical or aware of fake content.
The second column of Panel A runs the same regression but uses Wrong Words as the dependent variable. Here, there is a strong negative association between fake articles and use of the words “wrong” or “not right” in the comments section. This result suggests that investors feel the fake articles are more accurate (less wrong) than the non-fake articles. Fake articles seem to be more convincing of their statements than the non-fake articles, which may be why they generate more trading volume (and may be why they are used in the promotional campaigns).
Panel B runs similar regressions using the Post Event dummy instead of the Fake Ar-
Fake News: Evidence from Financial Markets 33
ticle dummy, where the post-event dummy is the six-month time period after the scandal. Interestingly, after the scandal, the incidence of the words “fake” and “fraud” increased signif- icantly (t-statistic of 2.73), implying that participants on these platforms were indeed more concerned with or commented more about false content on these sites after the scandal. This evidence corroborates the decline in trading volume witnessed post-scandal for all articles and suggests general mistrust of news from these platforms. The use of “wrong” words is no more prevalent after versus before the scandal. Hence, after the SEC announced investiga- tion and subsequent exposé articles, participants on these platforms seemed more concerned with fake news rather then erroneous news.
Combining the results in Panels A and B of Table 5, the evidence paints a picture of investors and consumers of information on these platforms being largely unaware of fake news before the SEC investigation and then suddenly becoming aware after the scandal, but having no ability to differentiate or detect which articles are fake and not fake. As a consequence, we see a marked drop in investor trading volume to any articles published on these sites, regardless of their authenticity, creating a significant spillover effect from the revelation of the existence of fake news on legitimate news more generally.
To further examine the link between the articles and trading volume, we examine whether authors who have more followers, and have written more articles, have a bigger impact, as well as whether articles that receive more comments lead to greater trading volume. We also analyze whether the trading volume reaction is higher when the article is more quantitative in nature and/or references accounting data, where presumably it is less likely to be false since numbers, such as earnings, can be verified from other sources. We look at the fraction of the article text comprised of numbers, as well as the number of words that have “earn” as part of the word. We regress abnormal trading volume for stock i on the number of followers an author has, log number of comments an article received, the number of past articles the author has written, and the fraction of numbers that appear in the text as well as the fraction of mentions of “earn.” We also control for whether there was at least one SEC filing and one
Fake News: Evidence from Financial Markets 34
firm-issued press-release in the three trading days leading up to the publication date, and control for abnormal trading volume the day before the article’s publication. The results are presented in Panel A of Table 6. We find that articles by authors with more influence as well as articles that get more comments are associated with a larger impact on trading volume. Furthermore, articles that seem to be more quantitative, also have a bigger impact.
Next we examine whether these characteristics have an impact on how the SEC scandal affected the trading volume reaction to the article. In particular, we rerun the regressions from Panel A, examining the time period six months before and six months after the SEC scandal. The results are presented in Panel B of Table 6. Similar to earlier analysis, we find that the trading volume response to articles is lower in the post-period. However, the drop is not as large if the author has more followers and has written more articles in the past. This suggests that people’s trust in the articles decreases less for authors that have a better reputation. We further find that the decrease is not as large for articles that mention “earn” in the article, suggesting that articles that cover accounting-related and hard information are not discounted as much after the scandal.
5. What Motivates Fake News?
Finally, we investigate what might be motivating the fake articles on these platforms. Using the Galena case that launched the broader SEC investigation, we examine whether other cases have similar characteristics and motivations to better understand the existence and prevalence of fake news. This analysis serves several purposes: it may help us better quantify the economic impact of fake news, provides another test of the linguistic algorithm’s ability to detect fake content, and may help identify other fake news.
5.1. Firm Performance
We start by examining the price reaction to the other for-sure fake articles from the SEC to see if a similar pattern as Figure 3 for Galena exists for the other firms involved in the scandal. We conduct the flip-side of the classic event study in financial economics
Fake News: Evidence from Financial Markets 35
(Fama et al. (1969)) by examining the return response to false news. This exercise is a novel test of the informational efficiency of markets, where in a perfectly efficient market fake news should have no impact on prices, regardless of the underlying equilibrium asset pricing model. We separate firms by size into small and non-small (there are no large firms in this sample) and examine their return response to the release of for-sure fake articles, by plotting the cumulative abnormal returns, measured as the difference between the return of the stock and a matched-portfolio of similar stocks (one of 125 equal-weighted portfolios based on size, book-to-market equity, and momentum), for days t + 1 to t + 251 after a fake article appeared about the firm.
Figure 5 plots the difference between cumulative abnormal returns for the for-sure fake articles, relative to days with non-fake articles. Returns for small firms increase after the fake article is published (relative to non-fake articles), reaching as much as 13%, cumulatively, after about 60 days, before giving up all the gains, and ending with a cumulative negative 5% return after a year. This pattern matches that of Galena in Figure 3. The permanent price impact of −5% for small firms indicates either that once the market figures out the news is fake, investors view this as a bad signal about the firm or that the true price should have dropped by 5% initially, but the fake news temporarily propped up the price and delayed the decline. For non-small firms, the price starts dropping immediately after the fake article comes out and continues to decrease throughout the year. This result could be consistent with the market figuring out the news is fake immediately for larger firms, where the cost of information is lower, or that the returns would have been even worse had the firm not initiated the fake articles.
We next examine the market price response to articles that we classify as probabilistically fake using the linguistic algorithm on the larger universe of all articles on these platforms. Since our analysis is at the firm-day level, we define whether a firm had a fake article on a given day using the average probability of being fake of all articles written about the firm on that day. Figure 5 plots the difference between abnormal cumulative returns following days
Fake News: Evidence from Financial Markets 36
with (probabilistically) fake articles, relative to days with (probabilistically) non-fake articles, and plots the responses separately for small, mid-size, and large firms in our sample (that have at least one fake article). As the figure shows, among small firms, returns following fake articles relative to non-fake articles increase for 6 months by about 5% following publication, and then revert back to their original level. These patterns are remarkably similar to the return patterns we found for the for-sure fake articles from the smaller SEC sample, further supporting our algorithm to identify fake news. The magnitudes are, not surprisingly, much smaller here since, unlike the SEC example, we identify fake news with noise plus the SEC is likely to go after the most extreme cases, so there may be selection bias. For larger firms, we find nothing, which makes sense since the market is more efficient for larger firms who also are less likely to engage in promotional campaigns. The lack of results on larger firms is another useful falsification exercise. We formally test whether the patterns in cumulative abnormal returns for fake news articles about different-sized firms over different horizons are statistically significant in Table C6 in the internet appendix. We find statistically significant results for small firms and no impact for larger firms.20
5.2. Other Firm Actions
Fake news is designed to deceive for financial or personal gain, including perhaps the
utility of fooling people and/or influencing others. In our setting of financial markets, it seems
less likely that private utility benefits motivate fake news. The SEC investigation focused
on promotional articles as part of pump-and-dump schemes to defraud securities markets.
Our findings on the impact on abnormal trading and temporary prices are consistent with a
motivation to hire authors to write fake content to promote the stock. Consistent with this
motive, Table ?? shows that these firms are more likely to issue press releases and 8-K filings
20One question is whether the poor long-term returns to small firms that promote fake articles are due to investors’ over/under reaction or whether fake articles are a sign of poor fundamental firm performance. Table C7 in the internet appendix shows that the presence of fake articles is associated with worsening fundamental firm performance, as measured by surprise in unexpected earnings, the return on assets, and its recent quarterly change. These findings are consistent with a possible motivation for engaging in promotional campaigns for financially troubled small firms that include hiring fake articles to prop up the stock price.
Fake News: Evidence from Financial Markets 37
within the same week to accompany the fake articles, perhaps to give authors of the fake articles more material and credibility and to influence the narrative of the firm. We also find in Table C9 that insider trading coincide with the fake articles and is positioned to profit fromthepricemovementcausedbythepromotion.21 Whiletheseactionsarerampantamong the SEC-prosecuted sample, we also find similar evidence for our broader sample of articles where we probabilistically assess the occurrence of fake news using the linguistic algorithm. Consistent with our earlier results, we find these effects, too, to be predominantly contained among small firms. This provides additional support to the efficacy of our methodology to detect fake articles more broadly. Moreover, the evidence suggests that an improved method for detecting fake content may involve examining other actions taken by the firm in addition to textual analysis. As one example, when we combine the probability of fake articles with the dual presence of insider trading to benefit from stock promotion, we find sharper price impact patterns. In addition, fake articles published following insider purchases are preceded by very sharp drops in share price in the month before publication, whereas fake articles not associated with insider purchases have flat to lightly increasing returns before publication. However, even for firms with fake articles written about them that do not have insiders buying shares, there is still a small price increase that also turns negative after several months, suggesting the results are not just driven by insider trading. In addition, performing a similar analysis using only the non-fake articles, there is no difference in returns for non-fake articles with insider buying versus without insider buying. Hence, it is not insider buying per se that drives the returns. Rather, it is the combination of insider buying with fake articles that seems to matter most and is indicative of a comprehensive promotional campaign.
6. Conclusion
We study three empirical settings to assess the economic impact of fake news: a unique dataset of fake paid-for articles on financial media crowd-sourced platforms prosecuted by
21We obtain data on press releases from RavenPack from 2001 to 2015, 8-K disclosure filings from the SEC’s Edgar database, and insider trades from Form 4 from Thomson Reuters.
Fake News: Evidence from Financial Markets 38
the SEC, a broader set of articles on these platforms that we apply a linguistic algorithm to detect fake content, using the first sample of known fake articles to calibrate the algorithm, and the SEC’s announced investigation that provides a shock to the public’s awareness of fake news. We find that fake news increases abnormal trading volume and imposes temporary price impact on small firms. Following public revelation of the existence of fake news, we find a significant spillover effect to news generally, where investors react less to all news, even legitimate news on these platforms. These findings represent some of the first documented direct and indirect effects of fake news that are consistent with theory (Allcott and Gentzkow (2017), Aymanns et al. (2017), and Kshetri and Voas (2017)).
Our study provides evidence on the prevalence and effect of fake news from crowd-sourced information platforms that continue to grow and gain attention. Financial markets may provide a lower bound on the impact of disinformation in other settings, where information costs are higher and where the ability to take action to correct its distortions is more limited (e.g., online consumer retail, political news, elections, and social media). More broadly, our findings may have implications for news media generally (e.g., Gentzkow and Shapiro (2005) and Gentzkow et al. (2015)) and for trust and social capital (e.g., Guiso et al. (2004), Guiso, Sapienza, and Zingales (Guiso et al.), Guiso et al. (2010), and Sapienza and Zingales (2012)).
Fake News: Evidence from Financial Markets 39
References
Allcott, H. and M. Gentzkow (2017). Social media and fake news in the 2016 election. The Journal of Economic Perspectives 31(2), 211–235.
Antweiler, W. and M. Z. Frank (2005). Is all that talk just noise? the information content of internet stock message boards. The Journal of Finance 59(3), 1259–1294.
Aymanns, C., J. Foerster, and C. Georg (2017). Fake news in social networks. CoRR abs/1708.06233.
Boudoukh, J., R. Feldman, S. Kogan, and M. Richardson (2018). Information, trading, and volatility: Evidence from firm-specific news. The Review of Financial Studies, hhy083.
Das, S. R. and M. Y. Chen (2007). Yahoo! for amazon: Sentiment extraction from small talk on the web. Management Science 53 (9), 1375–1388.
Fama, E. F. (1970). Efficient capital markets: A review of theory and empirical work. The Journal of Finance 25(2), 383–417.
Fama, E. F., L. Fisher, M. C. Jensen, and R. Roll (1969). The adjustment of stock prices to new information. International Economic Review 10 (1), 1–21.
Gentzkow, M. and J. Shapiro (2005, October). Media bias and reputation. Working Paper 11664, National Bureau of Economic Research.
Gentzkow, M., J. Shapiro, and D. Stone (2015). Media bias in the marketplace: Theory. handbook of media economics, simon anderson, david stroemberg and joel waldfogel, eds.
Grossman, S. J. and J. E. Stiglitz (1980). On the impossibility of informationally efficient markets. The American Economic Review 70 (3), 393–408.
Guiso, L., P. Sapienza, and L. Zingales. Trusting the stock market. The Journal of Fi- nance 63(6), 2557–2600.
Fake News: Evidence from Financial Markets 40
Guiso, L., P. Sapienza, and L. Zingales (2004, June). The role of social capital in financial development. American Economic Review 94(3), 526–556.
Guiso, L., P. Sapienza, and L. Zingales (2010, March). Civic capital as the missing link. Working Paper 15845, National Bureau of Economic Research.
Heston, S. L. and N. R. Sinha (2017). News vs. sentiment: Predicting stock returns from news stories. Financial Analysts Journal 73 (3), 67–83.
Jegadeesh, N. and D. Wu (2013). Word power: A new approach for content analysis. Journal of Financial Economics 110(3), 712–729.
Kshetri, N. and J. Voas (2017, November). The economics of “fake news”. IT Profes- sional 19(6), 8–12.
Milgrom, P. and N. Stokey (1982). Information, trade and common knowledge. Journal of Economic Theory 26(1), 17–27.
Mullainathan, S. and A. Shleifer (2005, September). The market for news. American Eco- nomic Review 95(4), 1031–1053.
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.
Sapienza, P. and L. Zingales (2012). A trust crisis. International Review of Finance 12(2), 123–131.
Fake News: Evidence from Financial Markets 41
Tetlock, P. C. (2007). Giving content to investor sentiment: The role of media in the stock market. The Journal of Finance 62 (3), 1139–1168.
Vosoughi, S., D. Roy, and S. Aral (2018). The spread of true and false news online. Sci- ence 359(6380), 1146–1151.
I am posting this for two reasons. I just received a class action settlement of $1.92 (whoo hoo) for my Galena stock and forgetting how and why I ever owned it, I went on a quest and turned up this fascinating abstract which may shed some light on last year's events, and upon practices that have been used to influence retain buyers for a very long time. Anyone with the stamina to get through this will see Biozone prominently mentioned, also. I believe (I am having a lot of difficulty navigating this abstract, that it may have been commissioned by someone or some entity named Garton (Investment Group or Dennis?) which I associate in some way with Icemandios having an opinion of Gartman ....I'm sure I will hear if it is negative. There are another 40 some pages which I believe to be charts etc. supporting the first 40+ pages. If you are really interested and cannot access it, let me know. This is quite dense reading, and surprising in that it wasn't written some years sooner before many retail investors were ....damaged. As Daddy Warbucks is fond of telling me after he has committed some household infraction: "Who ya gonna believe, ME or your lying' eyes?"
Fake News: Evidence from Financial Markets
Shimon Kogan∗
MIT Sloan School of Management Interdisciplinary Center Herzliya
Tobias J. Moskowitz
Yale School of Management NBER
AQR Capital Management
Marina Niessner
AQR Capital Management
March 2019
∗ We thank Tony Cookson, Diego Garcia, Gary Gorton, Bryan Kelly, Elizabeth Kempf, Bonnie Moskowitz, James Pennebacker, Kelly Shue, Eric So, Denis Sosyura, Sam Hartzmark, as well as conference and seminar participants at UCLA (Anderson), Rice University (Jones), University of Miami Business School, ASU Sono- ran Winter Finance Conference, 3rd Annual News & Finance Conference, University of Colorado at Boulder, Northwestern University (Kellogg), FSU SunTrust Beach Conference, MIT Sloan, Yale SOM, Catolica- Lisbon, University of Kentucky Finance Conference, FEB, 3rd Rome Junior Finance Conference, the 2018 WFA meetings, the U.S. Securities and Exchange Commission Division of Economic and Risk Analysis, the Red Rock Conference, Behavioral NBER, and Jackson Hole Finance Conference for their helpful comments and suggestions. We also thank Elli Hoffmann and Keren Ben Zvi for providing and helping organize the data. AQR Capital Management is a global investment management firm, which may or may not apply sim- ilar investment techniques or methods of analysis as described herein. The views expressed here are those of the authors and not necessarily those of AQR. Contact emails: skogan@mit.edu, tobias.moskowitz@yale.edu, and marina.niessner@aqr.com
Fake News: Evidence from Financial Markets 2
Abstract
We examine fake news in financial markets, a laboratory that offers an opportunity to quantify its direct and indirect effects. We study three experimental settings. The first is a unique dataset of unambiguous fake articles on financial news platforms prosecuted by the Securities and Exchange Commission. The second applies a linguistic algorithm to detect deception in expression on the universe of articles on these platforms, using the first sample as a validation and calibration set. The third is an event study exploiting the SEC investigation as a public shock to investor awareness of fake news. We find that fake news increases trading activity and price volatility relative to non-fake news for the equity securities of firms mentioned in the articles. Following public revelation of the existence of fake news, we find an immediate decrease in reaction to all news, including legitimate news, on these platforms, consistent with indirect spillover effects of fake news conjectured by theory. These findings are predominant among small firms with high retail ownership, and are stronger for more circulated articles. Our results are consistent with economic theory on media bias and its application to fake news.
Fake News: Evidence from Financial Markets 1
1. Introduction
False or misleading information can potentially impact social, political, and economic rela- tionships. A recent and prominent example is the increased attention “fake news” is receiving. Fake news is a form of disinformation, including hoaxes, frauds, or deceptions, designed to mislead consumers of news. The economics of fake news is an interesting and young area of study. What motivates fake news? What impact does it have? What are the welfare costs and benefits of monitoring it? What policy prescriptions should be considered?
Analysisoftheseissueshasprimarilybeentheoretical.1 Falsecontentcanimposeprivate
and public costs by making it more difficult for consumers to infer the truth, reduce positive
social externalities from shared-information platforms, increase skepticism and distrust of
legitimate news, and potentially cause resource misallocation. On the other hand, consumers
may derive utility from fake news (as entertainment or if slanted toward their biases as in
Mullainathan and Shleifer (2005)). Little empirical evidence on these issues exists, however,
due to lack of data, particularly the identification of fake content.2 With the explosion
of largely unmonitored shared information platforms, such as social media, blogs, and other
crowd-sourced content, the potential influence of fake and biased news is a growing concern.3
Indeed, a major challenge currently facing Amazon, Facebook, Twitter, and other crowd-
1Allcott and Gentzkow (2017) model fake news as an extension of Gentzkow and Shapiro (2005) and Gentzkow et al. (2015) on media bias, where fake news occurs in equilibrium when agents cannot costlessly verify the truth and the news matches the agent’s priors, with some debate over the relevance and conse- quences of fake news. Aymanns et al. (2017) provide an equilibrium model of an adversary using fake news to target agents with a biased private signal, where knowledge of the adversary causes agents to discount all news. Kshetri and Voas (2017) discuss the pervasiveness of fake news and its dissemination across news consumers.
2For example, Amazon, Google, Twitter, and Facebook are currently using human editors to evaluate content in the hopes of training an algorithm to identify false content systematically with limited success (Cullan-Jones (2016), Leong (2017), Leathern (2017)).
3According to a survey from the Pew Research Center (Gottfried and Shearer (2016)), 62% of American adults get news from a social media site. Allcott and Gentzkow (2017) argue that 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 et al. (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 in order to increase diffusion. Fake news has been suggested to have influenced the 2016 U.S. Presidential election (Allcott and Gentzkow (2017), Silverman (2016), Timberg (2016), Silverman and Alexander (2016)), and a study by ReviewMeta (2016) found that fake reviews on Amazon are misleading consumers toward various products (often paid for by producers).
Fake News: Evidence from Financial Markets 2
sourced content is the ability to detect fake content.
We provide some of the first empirical estimates of the direct and indirect impact of
fake news using three empirical settings. The first is a dataset of identified fake articles from a Securities and Exchange Commission (SEC) investigation into paid-for false articles on shared financial news networks. The sample is small, but the identity of fake news is clean – stemming from an undercover investigation by an industry whistle blower, Rick Pearson, resulting in 171 articles by 20 authors covering 47 companies knowingly providing false information about the stock. The data offer a singular look at identified fake content.
While the first setting provides known fake content, the sample is small and narrow. To broaden the analysis, and perhaps draw more general conclusions, we collect all articles from two prominent financial crowd-sourced websites – Seeking Alpha and Motley Fool – obtaining 203,545 articles from 2005 to 2015 for Seeking Alpha, and 147,916 articles from 2009 to 2014 for Motley Fool, covering 7,700 publicly traded firms. We then attempt to identify fake content within this broader set of articles using a linguistic algorithm (Pennebaker et al. (2015), Newman et al. (2003)) designed to detect deception in expression to assess the authenticity of each article. Using an “off-the-shelf" algorithm allows us to avoid in-sample overfitting, on the other hand we need to verify that the method works in a financial setting. Therefore, it is important that we use the first empirical setting of known fake articles from the SEC to validate the algorithm. It furthermore allows us to calibrate a model for the probability of fake news. Even with our identified set of fake articles, we show how difficult it is to detect fake content systematically. Our calibrated model classifies news as fake, non-fake, and ambiguous with the objective of minimizing classification error so that we can confidently identify fake and non-fake content. The algorithm has a type II error on the known fake articles of less than 1% (false positives) and a type I error on the non-fake articles by the same authors of less than 10% (false negatives). However, the low classification error comes at a cost because there are many articles that cannot be confidently classified, highlighting one of the major challenges and tradeoffs in quantifying fake news generally.
Fake News: Evidence from Financial Markets 3
Despite using a conservative measure, the prevalence of fake news we find in the broader sample is significant (2.8% of articles).
Our third empirical setting does not require identification of fake news at all. Rather, we exploit the public revelation of the SEC’s investigation on these platforms (that ultimately led to the dataset for our first experiment) as a shock to the market’s awareness of fake news. We first show that the market seemed largely unaware of fake content before the announcement, and then examine the market’s response to news before versus after the event. We use this setting to test another implication from theory that fake news imposes externalities on other news. We examine the shock from the public’s awareness of fake news on the market’s reaction to news in general, including legitimate news.
We first examine the direct impact of fake versus non-fake news on trading activity. Using the first empirical setting of known fake articles from the SEC, we find a larger trading response to fake news relative to non-fake articles published at the same time on the same platform. Abnormal trading volume rises by more than 50% over the three days following a fake article relative to a non-fake article. This effect is concentrated in the smallest ten percent of firms on public stock exchanges and is not significant among the largest firms. The effect is also bigger for stocks with higher retail investor ownership, where a ten percent increase in retail ownership results in a 7% increase in the trading volume response to fake news relative to non-fake news. The stronger impact on trading activity is likely driven by fake articles being more sensational and diffusing more quickly across consumers (Vosoughi, Roy, and Aral (2018)). Further corroborating that story, we find that fake articles generate more “clicks” and more “reads.”
Turning to the broader set of articles in the second experiment, where we estimate the probability of fake news, we find similar but more muted results. We similarly find that the direct effect on trading activity from fake news is stronger for smaller firms with higher retail ownership and for articles with greater circulation (measured by number of clicks and readers), lending credence to these platforms influencing investor behavior.
Fake News: Evidence from Financial Markets 4
Exploring the indirect effects of fake news on trading activity using the third empirical setting from the SEC announcement event, we find that trading volume drops significantly for any news article written on these platforms after the public became aware of fake content. Comparing trading volume before versus after the SEC announced investigation, trading volume drops by 5.2% for all news on these platforms following the information shock, with the drop being even larger for small firms with high retail ownership. We also find decreases in trading volume on the Motley Fool platform, which is a competitor platform that was not part of the SEC investigation, indicating that awareness of fake news caused a spillover effect in trading volume for other news platforms as well. These findings are consistent with theory (Allcott and Gentzkow (2017) and Aymanns et al. (2017)) arguing that awareness of fake news causes agents to discount all news. Looking at the the comments section to these articles using natural language processing (NLP), we find a significant increase in uses of the words “fake” and “fraud” after the SEC investigation came to light, consistent with participants being more concerned and aware of fake news after the event. Importantly, the frequency of these words in the comments has no predictive power to detect fake articles, indicating that readers had no ability to identify fake news, but were just more aware of its existence, consistent with their response to distrust all news.
Looking at news media more generally outside of these platforms, we obtain newspa- per articles about the firms from the Wall Street Journal and New York Times (obtained from Factiva) and examine whether there are any spillover effects from the SEC investiga- tion more broadly. Perhaps not surprisingly, and somewhat reassuringly, we do not find a commensurate decline in trading volume in response to news from non-social media sources such as the WSJ and NYT, suggesting that investors’ distrust of news from the social media platforms did not extend to more prominent types of media. The lack of a change in trading activity in response to newspaper articles also provides a useful falsification test that rules out any trend in response to news in general or unobservable effect on trading volume that just happened to coincide with the SEC event, potentially driving our results.
Fake News: Evidence from Financial Markets 5
We also examine pricing effects to see if fake news moves prices in a distortive way. If markets are informationally efficient, then despite the changes in trading volume, prices will be unaffected. In this case, fake news may distort attention and trading, but not firm values. Using the sample of known fake articles from the SEC, we find that the fake promotional articles increase idiosyncratic stock volatility by roughly 40 percent relative to non-fake articles over the three days after publication, with the effects concentrated among small firms with high retail ownership. Looking at the direction of price movement, the average fake article is positive and pushes up stock prices for the smallest 10% of companies on the NYSE by an average 8% over the next six months, which subsequently gets fully reversed over the course of a year and eventually becomes cumulatively negative at −2.5%. Looking at the broader set of articles where we estimate the probability of fake news, we find similar but much weaker patterns of temporary positive price effects for the smallest firms that get fully reversed and turn negative. For large firms, we find no price impact. Finding similar patterns lends further support to our algorithm for detecting fake content.4
The results are consistent with markets being less efficient for small firms, where the cost of information is higher and the average investor is smaller and perhaps less sophisticated.5 Consistent with this notion, paid-for fake content from the SEC investigation was exclusively engaged by small firms, and not by large firms, as expected in equilibrium. We also find that small firms are more likely to issue press releases and 8-K filings coinciding with the fake articles, consistent with a coordinated effort to influence the narrative of news about the firm, and find evidence of insiders positioning themselves to benefit from the price movement.6
Our results provide some of the first empirical estimates of the direct and indirect impact
4An investor at the time of the article’s publication could not have constructed or used a similar method- ology to detect the probability of false content since the fake articles from which we calibrate our model were not yet known or identified.
5The 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 psychological or behavioral biases. Allcott and Gentzkow (2017) argue information costs are necessary for fake news production.
6In fact, The reason Rick Pearson went undercover initially and why the SEC got involved was because many of these fake articles were tied to promotional pump-and-dump schemes to manipulate the stock price, orchestrated by the firms themselves.
Fake News: Evidence from Financial Markets 6
of fake news, which have implications for theory. The prevalence of fake content and its impact on trading activity is consistent with fake news being tailored to consumer’s priors as suggested by Allcott and Gentzkow (2017), and more broadly consistent with media bias (Gentzkow and Shapiro (2005) and Gentzkow et al. (2015)). The price patterns we find for small firms may also be consistent with fake news producers sacrificing longer-term reputational capital in lieu of short-term gains (Allcott and Gentzkow (2017)). The decline in trading activity to all news, including legitimate news, following the public’s awareness of fake news from the SEC investigation is also consistent with Aymanns et al. (2017) and Allcott and Gentzkow (2017), that fake news increases distrust of media in general.7 The results may also be related more generally to the economics of norms and institutions such as trust and social capital (Guiso et al. (2004), Guiso, Sapienza, and Zingales (Guiso et al.), Guiso et al. (2010), Sapienza and Zingales (2012)).
Finally, Our setting is financial markets, and specifically shared-information platforms on financial news and opinions. There are reasons to be both cautious and optimistic on what we can learn about the impact of fake news more broadly. One of the benefits of financial markets is that we can quantify the influence of fake news through prices and trading activity. On the other hand, these information platforms may have little influence on markets either because they are unimportant or due to markets already incorporating the information. A 2015 study by Greenwich Associates found that 48% of institutional investors use social media to “read timely news.” On the other hand, fake news should not matter if markets are informationally efficient (Fama (1970)), regardless of the equilibrium asset pricing model. In this sense, our setting offers a unique test of market efficiency that circumvents the joint hypothesis problem: we run the flip side of the classic event study (Fama et al. (1969)) by examining market responses to fake news events and find (at least for small stocks) that trading activity and prices are affected. Our results provide more
7See also “Trust in Social Media Falls – Raising Concerns for Marketers,” by Suzanne Vranica, Wall Street Journal, June 19, 2018, which discusses research by Edeleman, the world’s largest public relations firm, that found trust in social media has fallen world-wide and particularly in the U.S. over the last year.
Fake News: Evidence from Financial Markets 7
evidence on the growing impact of crowd-sourced information platforms (Hu, Chen, De, and Hwang (2014)). If fake news can impact U.S. equity markets, where there is competition for information and arbitrage activity exists, then it could have even greater influence in settings where information costs are high and the ability to correct misinformation is more limited, such as online consumer, political, and social shared-information networks.
The rest of the paper is organized as follows. Section 2 details our first two empirical settings: the sample of fake news articles obtained from the SEC and the broader set of all articles from the shared-information platforms that we apply a linguistic algorithm for assessing fake content. Section 3 analyzes the direct impact of fake news through investor trading activity and price impact. Section 4 describes the third empirical setting – an event study of the SEC’s investigation that provides a shock to public awareness of fake news – to measure the indirect effects of fake news on the market’s reaction to news generally. Section 5 considers the motivation of fake news. Section 6 concludes.
2. Identifying Fake News
We detail our first two empirical settings using articles from knowledge-sharing financial platforms. We first describe these platforms and the data we obtain. We then describe our first sample of fake articles from the SEC. Using this sample, we validate and calibrate a linguistic model for identifying probable fake content. We then apply this model to the broader sample of articles with unknown authenticity from the same media platforms that generates our second empirical setting.
2.1. Knowledge Sharing Platforms
Our sample of articles comes from the two largest financial crowd-sourced platforms: Seeking Alpha and Motley Fool. Seeking Alpha is an online news service provider for fi- nancial markets, whose content is provided by independent contributors. The company has had distribution partnerships for its content with MSN Money, CNBC, Yahoo! Finance, MarketWatch, NASDAQ and TheStreet. The Motley Fool is a multimedia financial-services
Fake News: Evidence from Financial Markets 8
company that provides financial advice for investors also through a shared-knowledge plat- form. We obtain the articles posted on these platforms, including their content, authorship, and in the case of Seeking Alpha, commentary from other users. Appendix A details how authors on these sites contribute and are compensated for their articles.
The popularity of these sites has grown considerably. Seeking Alpha grew from two mil- lion unique monthly visitors in 2011 to over nine million in 2014, generating 40 million visits per month. While these platforms allow for the ‘democratization’ of financial information production, concerns have been raised about their susceptibility to fraud, since they are virtually unregulated, frequented predominantly by retail investors, and authors on these platforms can use pseudonyms (though the platforms claim they know the true identity of each author, in case that information is subpoenaed by the SEC, which it was in the cases we examine below). Authors are allowed to talk up or down a stock that they are long or short, provided they disclose any positions they have in the stock in a disclaimer accompanying the article. Failure to disclose can have legal ramifications. 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 underwriter.8 We exploit a subset of fraudulent articles on these platforms, identified by an undercover investigation of paid-for articles of false content and eventual SEC prosecution for our first empirical setting.
2.2. “For-Sure” Fake Articles
We examine a unique dataset of articles whose authors were paid to write fake articles,
but where the authors illegally did not disclose payment. The articles are obtained from
an industry insider, Rick Pearson, who, as a regular contributor to Seeking Alpha, was
approached by a public relations firm to promote certain stocks by writing articles with false
information for a fee without disclosing the payment. Mr. Pearson instead went undercover
to investigate how rampant this practice was on these platforms and uncovered more than one
8In June 2012, Seeking Alpha announced it would no longer permit publication of articles for which compensation had been paid.
Fake News: Evidence from Financial Markets 9
hundred fake, paid-for articles by other authors who did not disclose their compensation. He turned the evidence over to the SEC, who investigated each of these cases. The fake articles were subsequently taken down by the platforms once the SEC informed them of the investigations. The SEC filed its first lawsuit pertaining to these articles on October 31, 2014, prosecuting the authors, promotion firms who paid them, and in some cases the companies and their executives who hired the promotion firms.9
Mr. Pearson kindly provided the articles to us that he determined to be fake: 111 fake articles by 12 authors covering 46 publicly traded companies. We also obtained a second set of known or, as we will refer to them, “for-sure” fake articles, which the SEC identified during the investigation containing paid-for fake content.10 We also contacted Seeking Alpha, and they kindly shared 147 of those articles with us. Among those articles, we match 60 to firms publicly traded on U.S. exchanges to obtain price and volume information from Center for Research in Security Prices (CRSP). The rest of the articles pertain to firms traded over the counter. Combining all of the data sources, our final sample of for-sure fake articles consists of 171 articles written by 20 different authors about 47 publicly traded firms.11
It is important to define what we mean by fake articles. In this smaller sample from
Rick Pearson and the SEC, the fake articles are those that were paid for by a promotional
firm and not disclosed, and many of the authors admitted that the articles were written to
deceive the market and manipulate the stock price. Consequently, these articles contained
false information that authors thought to be incorrect at the time. How false or wrong that
information turns out to be is difficult to assess. For example, an article could intend to
deceive by embellishing the prospects of the firm, but could turn out to be mostly correct
in that assessment ex post. In other instances, the deception may be grossly off. Hence, our
9Subsequent lawsuits were also filed on April 10, 2017 and September 26, 2018. See filing documents at: securities.stanford.edu/filings-documents/1051/GBI00_01/20141031_r01c_14CV00367.pdf; www.sec.gov/litigation/complaints/2017/comp23802-lidingo.pdf.
10The full list can be found here: ftalphaville-cdn.ft.com/wp-content/uploads/2017/04/10231526/Stock-promoters.pdf.
11While we gain 60 additional articles from the SEC, we only gain one additional firm. Most of the additional articles pertain to firms already covered by Rick Pearson, and hence simply give us more fake articles about the same firms, with only one new firm identified.
Fake News: Evidence from Financial Markets 10
fake articles are about intent to deceive and not necessarily about whether they are right or wrong ex post. We focus on authenticity and not accuracy. Articles may be fake and (mostly) right, as well as fake and (very) wrong. Some of our analysis on the language used in the articles and on their impact on stock prices may help distinguish between these two cases, where we conclude that most of the articles perpetuated false information.
To provide some insight into the content of these fake articles, we highlight a recent example from our sample that was prosecuted by the SEC in September 2018. One of the fake, paid-for articles in this case was a publication that appeared on Seeking Alpha on September 26, 2013 about the company Biozone. The article stated,
Biozone has developed a new method of drug delivery, QuSomes that provides improved efficacy, reduced side effects, and lower costs. This technology will allow Biozone to reformulate and sell certain FDA approved drugs at a reduced cost, which should help Biozone capture a large percentage of these drug markets.
From the SEC lawsuit filed in September 2018 in the District Court of New York City:
Keller misleadingly stated that Company A had a formulation ready for testing to be brought to the billion-dollar injectable drug market. Yet, as Keller knew, as of summer 2012, all R&D efforts had been shut down without the successful formulation of an injectable drug and Company A had ceased all efforts to develop this technology in mid-2012.
Keller refers to Brian Keller, the co-founder and Chief Scientific Officer of Biozone, who had paid for the promotion article. Many of the fake, paid-for articles from the SEC involve similar issues and often coincide with a public relations campaign orchestrated by the firm to artificially prop up the stock price, including press releases, filings, and corporate actions. We investigate these issues more below and attempt to control for these actions when assessing outcomes on trading activity and prices.
We compare the fake, paid-for articles to a set of non-fake articles by the same authors. Specifically, we obtain a sample of other articles written by the same 20 authors under the SEC’s investigation that were not paid for by a PR firm and having no evidence of being false, totaling 334 additional articles about 171 companies published on the same platform
Fake News: Evidence from Financial Markets 11
Seeking Alpha. We use this set of non paid-for articles by the same authors to provide a clean comparison to the fake articles to control for heterogeneity across authors. We refer to these non-paid for articles as “non-fake” following our definition above and make no statement about the accuracy of the articles themselves. We examine the impact of the fake versus non-fake articles on markets.
2.3. Assessing Authenticity in the Broader Set of Articles
Our second empirical setting uses the broader set of all articles written on these platforms and attempts to assess their probability of being fake. While our unique sample of fake articles from the SEC provides unambiguous fake news, the sample is small and therefore may raise external validity concerns that may make it difficult to draw general conclusions. To complement these data, we manually download all articles published on Seeking Alpha and Motley Fool, obtaining 203,545 articles from Seeking Alpha over the period 2005 to 2015 and 147,916 articles from Motley Fool from 2009 to 2014.
2.3.1 How do you tell if someone is lying?
The downside of this much larger dataset of 351,461 articles is that their authenticity is unknown. To assess authenticity, we develop a probability function for detecting fake content using an objective and scalable measure based on quantifiable research from linguis- tics. Specifically, we use a linguistic algorithm designed to detect deception in expression – the Linguistic Inquiry Word Count model (LIWC2015) from Pennebaker et al. (2015) – which focuses on individuals’ writing or speech style, and appears adept at measuring indi- viduals’ cognitive and emotional states across various domains. The LIWC model outputs the percentage of words that fall into one of more than 80 linguistic, psychological, and topical categories, one of which is the authenticity score that detects deception in expres- sion. While the exact formula for the authenticity score is proprietary, Pennebaker (2011) describes which linguistic traits are associated with honesty. In particular, truth-tellers tend to use more self-reference words and communicate through longer sentences compared to
Fake News: Evidence from Financial Markets 12
liars. When people lie, they tend to distance themselves from the story by using fewer “I” or “me”-words. Furthermore, liars use fewer insight words such as realize, understand, and think, and include less specific information about time and space. Liars also tend to use more discrepancy verbs, like could, that assert that an event might have occurred, but possibly did not. Newman et al. (2003) use an experimental setting to develop an authenticity score based on expression style components using similar techniques, and the Central Intelligence Agency and Federal Bureau of Investigation have used similar methods to assess authenticity in speech or writing.
For example, consider the two statements of former U.S. congressman Anthony Weiner before and after his admission in the “sexting” scandal.
Before admission:
We know for sure I didn’t send this photograph. [...] We don’t know where the photograph came from. We don’t know for sure what’s on it. [...] If it turns out there’s something larger going on here, we’ll take the requisite steps.
After admission:
I would like to make it clear that I have made terrible mistakes, that I have hurt the people I care about the most, and I am deeply sorry. I have not been honest with myself, my family, my constituents, my friends, my supporters and the media.
The use of “we” versus “I” and “my”, the discrepancy words “don’t know” and “if”, and the lack of insight words like “mistakes,” “clear,” and “hurt” are all more prevalent in his statements when he was lying to the public.
The algorithm uses a combination of these linguistic traits to generate the authenticity measure. A unique and critical advantage of our study is that we use the for-sure fake articles from Rick Pearson and the SEC to validate the linguistic algorithm and calibrate the authenticity score into a probability of fake news. One of the major challenges to studying this issue is the lack of known fake content. Since the LIWC authenticity score was not developed in the context of financial media, it is useful to assess its ability to distinguish fake from non-fake articles in our context. Financial blogs and articles tend to point to
Fake News: Evidence from Financial Markets 13
facts, trends, and figures, which may be decidedly different from narratives that were used to develop the linguistic algorithm. Our unique sample of 171 for-sure fake articles and 334 non-fake articles written by the same authors, provides a validation and test of the generalizability of the linguistic algorithm. We compare the LIWC authenticity score, which is normalized between 0 and 100, for the two samples and control for author fixed effects to capture heterogeneity in author style, content, or reputation, and any matching of authors to types of articles.
Panel A of Table 1 reports the difference in the LIWC authenticity scores for the fake and non-fake samples. Relative to an average authenticity score of 33 for non-fake articles, fake articles have a much lower average score of 19 (statistically significant at the 1% level). A plot of the distribution of authenticity scores for fake and non-fake articles in Figure 1, Panel A highlights the differences, controlling for author heterogeneity since we examine fake and non-fake articles for the same authors. Panel B of Figure 1 provides more specific examples for two authors: John Mylant and Equity Options Guru. The distribution of authenticity scores across fake and non-fake articles for each author are quite different. While some of the non-fake articles also have low authenticity scores, most of the fake articles have very low authenticity scores.
Panel A of Table 1 reports summary statistics on language characteristics associated with authenticity as described in Pennebaker (2011) for the for-sure fake and non-fake articles. We report the average use of 1st person singular (examples: I, me, mine), Insight (examples: think, know), Relativity (examples: area, bend, exit), Time (examples: end, until, season), Discrepancy (examples: should, would), and the average number of words per sentence. According to Pennebaker (2011) and Pennebaker et al. (2015), when people lie they also tend to use fewer words per sentence. Fake articles appear less authentic, having fewer self- referencing, lower insight, lower relativity, and higher discrepancy scores on average. These findings provide an out-of-sample test of the LIWC algorithm that validates it in a unique setting, which would not be possible without the sample of known fake articles from the
Fake News: Evidence from Financial Markets 14
SEC.
2.3.2 Probability of Being Fake
The sample of for-sure fake and non-fake articles also allows us to calibrate the au- thenticity scores into a probability of fake content. While the LIWC authenticity score is statistically different between fake and non-fake articles, it is not easy to interpret the car- dinal nature of the score – what does a 14 point difference in authenticity score mean? To provide a more direct interpretation of the results and their economic meaning, we develop a mapping of the authenticity score into probability space. Again, this exercise is only possible because we have a set of known fake articles from which to calibrate probabilities. Using the smaller sample of for-sure fake and non-fake articles, we map the authenticity score into the frequency of fake articles and apply Bayes rule to the broader sample of Seeking Alpha and Motley Fool articles. We use the known fake and non-fake articles to map authenticity scores into a conditional probability of being fake.
Specifically, let S be the authenticity score and F (T) denote a fake (true) article. We compute P rob(S|F ) and P rob(S|T ), where, crucial to this exercise, we use the smaller valida- tion sample, where we know which articles are F and T , in order to measure the probabilities. From Bayes rule,
P rob(F |S) = P rob(S|F )P rob(F ) . P rob(S|F )P rob(F ) + P rob(S|T )P rob(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 is found by solving the fixed point problem in which the observed P rob(F ) in the sample is representative of Prob(F) in the overall population.
Figure 2 plots the mapping of LIWC authenticity scores (S) into the conditional prob- ability of being fake (Prob(F|S)) for the entire sample of 203,545 Seeking Alpha articles
Fake News: Evidence from Financial Markets 15
published between 2005 and 2015. The relation between the LIWC authenticity score and the probability of being fake is highly nonlinear. Specifically, the sharp increase in probability in the very low authenticity range suggests that articles may be more efficiently and better classified into fake and non-fake using a probability cutoff. We use a cutoff of Prob(F) > 0.20 to classify articles as being fake and classify articles with Prob(F) < 0.01 as being non-fake, with the remaining articles (0.01 ≤ P rob(F ) ≤ 0.20) being classified as ambiguous or “other.”12 This cutoff implies an authenticity score that is even lower (about half) than the average authenticity score for the known fake articles in the SEC sample. Hence, this cutoff is conservative and designed to reduce type II errors.
We first examine how accurate our method is at identifying fake news from our small sample of 171 for-sure fake and 334 non-fake articles written by the same authors. We generate an authenticity score for each article, and calculate its probability of being fake. Our algorithm classifies 18 articles as being fake, of which 17 are actually fake, indicating that the Type II error rate is very low – one false positive. Our method is very conservative, however, since it misses a lot of fake articles – 154 fake articles are not classified as fake. Our algorithm identifies 165 articles as being non-fake. Of those, 17 are actually fake, implying a Type I error of about 10%, which is quite low considering our methodology is designed to minimize type II errors. However, our methodology is highly conservative in that 186 (334 - 165 + 17) non-fake articles are not classified. We exclude articles with 0.01 ≤ P rob(F ake) ≤ 0.20 from our analysis, since both Type I and Type II errors will be larger for these articles, which amounts to 322 (505-18-165) articles or about 64% of the sample not being classified.
These statistics highlight the challenges in identifying fake news. There is a tradeoff
in how confident we wish to be in our classification versus how many articles we wish to
classify. In our case, we choose to minimize the number of falsely identified fake articles
and falsely identified non-fake articles, so that when we look at their differential impact,
we are confident that we are measuring the difference between fake and authentic news.
12Our results are not sensitive to different cutoffs in the 0.10 to 0.30 probability range for fake, where 0.20 was chosen based on the steep increase in probability in Figure 2.
Fake News: Evidence from Financial Markets 16
This, however, means we are unable to classify many articles. If the goal is to capture as much fake content as possible, then we would choose a different threshold that would classify more articles but invite more classification error. This tradeoff highlights the limitations of the linguistic algorithm and echoes some of the challenges facing social media platforms in flagging fake content. For our purposes of identification, it is more important to minimize classification error at the expense of not being able to classify many articles.
Using our fake news probability model, calibrated to the sample of known, for-sure fake and non-fake articles, we apply our methodology to the broader sample of all articles. Table 1 Panel A shows summary statistics for the Fake, Non Fake, and Other articles identified by our algorithm. The number of articles in each category, the mean of the Authenticity measure that we use to construct the probabilities, and the components of that authenticity measure from the LIWC algorithm are reported. The difference in authenticity measures translates into large differences in the estimated probability of being fake from our calibrated function: the articles we identify as fake have an average 0.45 Prob(F) based on their authenticity score, while the average probability for articles we identify as non-fake is less than 0.01. Obviously, the articles were sorted based on the probabilities, but the magnitude of the difference is interesting and suggests substantial differences in authenticity scores between the two groups of articles, which Table 1 reports are 5.4 versus 50.7.
We also apply our methodology for identifying fake articles to another sample of articles from another crowd-sourced financial news platform – Motley Fool, where we have 147,916 articles from 2009 to 2014. Applying the LIWC algorithm, we obtain similar differences in authenticity scores and probabilities in classifying Motley Fool articles into Fake and Non- Fake. The unconditional probability of fake news on the Motley Fool sample is 2.7%, almost identical to the 2.8% we found for Seeking Alpha. Looking at the rest of the components of the authenticity score, the algorithm does a similar job on the Motley Fool sample.
Finally, as another validation exercise we analyze a particular set of articles written by a Motley Fool author, Seth Jayson, who has been working for Motley Fool full-time since 2004
Fake News: Evidence from Financial Markets 17
as a journalist, and has written over 31,000 articles. Mr. Jayson’s articles are a good test case because he works directly for Motley Fool and has for a long time. Hence, it is unlikely he has written fake articles on their platform and unlikely promotional firms would even approach him to do so. We, therefore, use his articles as a placebo test of our classification methodology. Using our methodology, we classify 18,361 of Mr. Jayson’s articles as reliably non-fake and only 2 of his articles as probabilistically fake (the rest being indeterminate). In other words, we classify 0.006% of his articles as fake, which is consistent with our prior that he would have essentially zero fake articles and suggests our conservative classification methodology works well at identification (though, again, the tradeoff is that many of Mr. Jayson’s articles cannot be classified by the methodology).
Finally, Panel C of Table 1 reports the average fraction of retail investors, the average number of analysts covering the firm, and the average firm size (in $US millions) for each article group. For-sure fake articles tend to cover small firms with a high fraction of retail in- vestors and low analyst coverage. The probabilistically determined fake and non-fake articles from the broader Seeking Alpha and Motley Fool samples exhibit more muted differences. Notably, the Motley Fool articles are written about significantly larger firms than Seeking Alpha, and the for-sure fake articles identified by Rick Pearson and the SEC are about tiny firms whose average market capitalization is only $7.4 million.
Table C1 in the Internet Appendix also examines whether fake articles tend to cluster in specific industries. We separate articles into one of the 12 Fama-French industries that the firms mentioned in the articles belong to. For the for-sure fake articles provided to us by Rick Pearson and the SEC, 81% are about firms in the Healthcare industry. This finding is not too surprising as these articles came from authors who were hired primarily by two PR firms that concentrated on the healthcare industry. For the non-fake articles, the majority of firms belong to Business Equipment, Healthcare, Finance, and Manufacturing industries. The industry composition of Fake and Non-Fake articles we identify on Seeking Alpha and Motley Fool using our algorithm is similar to the Non-Fake articles’ industry composition
Fake News: Evidence from Financial Markets 18
from the smaller sample of articles from the SEC.
3. Trading Activity and Price Impact
In this section, we examine the impact of fake news on trading activity and prices. Financial markets provide a useful setting to examine the impact of fake news, because they provide high frequency outcomes, such as trading volume and market prices.
3.1. Do Articles on the Platforms Have Impact?
Before looking at fake articles, we first address whether articles posted in general (fake or non-fake) on these social platforms have any influence on market participants or markets. We start by examining abnormal trading volume around the publication of all articles. We focus on abnormal volume because we are interested in whether investors “react” to these articles. Of course, it is also quite plausible that articles react to trading activity. To try to establish some causal interpretation, we only examine abnormal volume and look at future changes in it. Abnormal trading volume is trading volume relative to expected volume in the stock, which we proxy for using the recent average daily volume in the stock (defined below). We look at future changes in abnormal volume in the days following the article’s publication. A reverse causal story would need to imply that authors are writing articles in anticipation of future unexpected trading activity. If true, we would call that “news.” Nevertheless, we also control for lagged abnormal volume from the previous trading day and other news coming from the firm such as recent SEC filings and press releases about the firm. We also examine the articles’ effect on price volatility to address whether there is information in the articles that is not already impounded into market prices. Of course, quantities traded can vary significantly with no price movement (Fama (1970), Grossman and Stiglitz (1980)) or trade can be zero with substantial price movement (Milgrom and Stokey (1982)). Thus, we look at both quantities and prices separately.
Panel A of Table 2 examines the effect on abnormal trading volume from articles published
Fake News: Evidence from Financial Markets 19
140
on these sites. We define abnormal trading volume for stock i as V ol(i, t)/ 1 V ol(i, t−k),
T
k=20
which is the trading volume for stock i on day t relative to the average daily trading volume in stock i over the last 6 months (skipping a month).13 We sum abnormal volume over days t = 0, t+1, and t+2, where t = 0 is the date the article appears on the website and then regress the natural logarithm of abnormal volume on an indicator variable for whether there is any article on these sites about the firm on a given day, regardless of its authenticity. We also include year-month fixed effects in the regression. We examine only firms that had at least one article published on Seeking Alpha or Motley Fool over the sample period.
As the first column of Panel A of Table 2 shows, an article published on Seeking Alpha or Motley Fool is associated with an 86% increase in abnormal trading volume over the three days following publication. This result implies either that investors are trading in direct response to the articles or, more generally, are trading in response to whatever news is coming out that day that these articles may be discussing. Likely both. While the increase seems large, this first regression controls for no other variables, except time fixed effects. Moreover, as we show below, the bulk of the effect is concentrated in very small and illiquid firms, where trading volume changes (as a percentage) can be enormous, and where outliers can be in the thousands of percents.
As we will show, articles on knowledge sharing platforms in the SEC-prosecuted cases are often written following press releases or SEC filings. In the second column of Panel A, we control for whether there is an SEC filing (10-K, 10-Q, or 8-K) or a company-issued press release in the three days leading up to the article. Furthermore, to control for serial correlation in abnormal trading volume, we include lagged abnormal trading volume on day t − 1 as a regressor, which also captures other events we may be missing that could affect trading activity. After the controls, the effect on abnormal trading volume declines to 37%. To make sure these results are not all coming from the day the news is released, Table ?? in
the internet appendix reports the effect on trading volume separately for the same day and
13Results are identical defining abnormal volume relative to the last 30, 60, or 180 days.
Fake News: Evidence from Financial Markets 20
for one and two days after the article’s publication. Of the 37% rise in abnormal trading volume, 15.5% occurs on the day the article is published, 12.1% the following day, and 10.1% two days later. Figure 4 in the internet appendix plots the abnormal volume response for the next 20 trading days and finds that abnormal volume increases for about two weeks.
The next three columns of Panel A of Table 2 report results separately for small, medium, and large firms. Small firms are defined as smaller than the bottom 10th percentile of NYSE firms, mid-size firms fall in the 20th to 90th size percentile of NYSE firms, and large firms are in the top 10th size percentile of NYSE firms. The effect on abnormal trading volume declines strongly with firm size, with the effect six times larger for small firms than for large firms (80.9% increase versus 8.2% increase). This result is consistent with small firms having more retail investor trading and perhaps a more opaque information environment. In the last two columns, we separate firms into high and low retail ownership (above or below median retail ownership last month) and find that the effect on abnormal trading volume is twice as large for firms with high retail ownership.
The internet appendix reports some robustness tests of these results. First, Table C3 includes firm fixed effects to difference out any unobservable firm heterogeneity over the sample period – the results are nearly identical. Second, Table C4 controls for some outliers in percentage change in volume by winsorizing the most extreme five percent of abnormal volume observations. Since abnormal volume is the dependent variable, it is always question- able to winsorize, unless we think these extremes are data errors. The effects are obviously more muted from winsorizing but the patterns stay the same: a 33.3% increase in abnor- mal volume over the next three days that is larger for small firms (54.1%) and high retail ownership firms (41.7%). Our findings are not driven by a few extreme observations. The point estimate for the full sample is roughly the same, but the effect from winsorizing on the smallest decile of firms, where percentage volume changes are more extreme, reduces the effect from 80.9% to 54.1%. Given the similar patterns and the fact that we do not believe the extreme volume changes are errors, we do not winsorize observations.
Fake News: Evidence from Financial Markets 21
Panel B of Table 2 repeats the regressions in Panel A, replacing abnormal volume as the dependent variable with the idiosyncratic price volatility of the stock. We measure id- iosyncratic volatility as the square of the difference between the return of the stock and a matched-portfolio of stocks with similar size, book-to-market equity (value) and past 12- month returns (momentum) following the procedure of Daniel, Grinblatt, Titman, and Wer- mers (1997), which forms 125 equal-weighted portfolios based on 5 × 5 × 5 sorts of stocks using size, value, and momentum characteristics that are related to expected returns. The dependent variable is the sum of daily idiosyncratic volatility on the day the article is pub- lished plus the next two days. This analysis captures whether articles moved prices around the days they were published. We examine price volatility as opposed to returns because it is exceedingly difficult to sign the direction of the content of the articles.14 Hence, looking at volatility or the absolute value of returns captures whether prices moved significantly in relation to the articles published on that day. If the market has already incorporated the news, then the expected absolute return change should be zero. As Panel B reports, we find effects similar to those in Panel A that examine trading volume – daily price volatility or the absolute return of the stock rises following articles published on these platforms, even after controlling for recent SEC filings, firm press releases, and return volatility in the days leading up to the article. The effect is strongest for smaller firms with higher retail ownership. The magnitude of these effects is large but not unreasonable. Across all articles, the effect of a published article on idiosyncratic stock volatility is about 6.8% over the three days, which is roughly an additional 40% of the stock’s normal price movement on days when there is no news about the firm on these platforms. For the smallest firms, the effect is an order of magnitude larger, which is consistent with extreme price movement for the smallest stocks (Frazzini, Israel, and Moskowitz (2017)).
14Textual analysis used to derive sentiment (Antweiler and Frank (2005), Tetlock (2007), Das and Chen (2007), Jegadeesh and Wu (2013), Heston and Sinha (2017), Boudoukh et al. (2018)) is notoriously challeng- ing and noisy. In addition, price movements are deviations from expectations, so a “positive” article that is less positive than expected would predict a negative return. Not knowing expectations makes signing the price movement even more difficult.
Fake News: Evidence from Financial Markets 22
3.2. Impact of Fake Articles
Do fake articles have a differential impact on trading volume and volatility? Panel A of Table 3 reports results from the same regressions as Table 2, but includes a dummy variable for whether the article is “for-sure fake” from our SEC sample. As the first column of Panel A shows, the for-sure fake articles are associated with significant increases in abnormal trading volume, which is not too surprising since the SEC is more likely to target cases that had more impact. The magnitude of the impact of fake articles is another 50% increase in trading volume over the next three days relative to a non-fake article. This seems particularly large, butisreasonableaccordingtoexcerptsfromthemostrecentSEClawsuits.15 Thesecondand third columns of Panel A interact the fake article dummy with the market equity decile of the company the article is written about and the retail ownership percentage of the firm. The impact on trading volume is larger for smaller firms and firms with higher retail ownership, though the latter interaction is insignificant.
The first three columns of Panel B of Table 3 report results from the same regressions using idiosyncratic volatility as the dependent variable. Consistent with the abnormal trad- ing volume results, fake articles have an additional significant impact on price movements relative to non-fake articles. The magnitude is large, too, which is not terribly surprising as this sample is based off of the SEC investigations, which are chosen ex post (where the SEC is more likely to go after cases where the promotional articles had massive impact on prices).
3.3. Using Probabilistically Fake Articles
The last three columns of Panels A and B of Table 3 report results using a dummy variable
for whether the article is probabilistically fake using our calibrated probability function for
fake news. As Panel A shows, the coefficient on the LIWC Fake dummy is insignificant
15From Case 1:18-cv-08175 filed on September 7, 2018 in the U.S. District Court, Southern District of New York: “The market reacted strongly to the Company A promotion: the trading volume of Company A stock rose from approximately 1,100 shares on September 25, 2013 to over 4.5 million shares on September 27, 2013 and to more than 6 million shares on October 2, 2013.” And, in the same case about another firm, “The article did not disclose that the author had been paid by Company B – at Honig’s direction – to write the article. After the article was published on February 3, 2016, there was a 7000% increase from the previous day’s trading volume, and an intraday price increase of over 60%.”
Fake News: Evidence from Financial Markets 23
by itself (though of positive sign), but interacting it with size deciles and retail ownership delivers significant effects on trading volume for the smallest firms with the highest retail ownership. These results are consistent with those from the narrower set of for-sure fake articles. Panel B shows that the probabilistically fake articles have a more muted impact on stock volatility, producing the same signed coefficient we get from the for-sure fake sample, but where nothing is statistically significant. The LIWC algorithm appears to capture some of the fake content that may exist on the platforms, but is a noisy measure and illustrates the difficulty in identifying fake news.
3.4. More Evidence on Direct Impact
To further test a direct link between articles published on these platforms and trading activity, we obtain a proprietary supplemental dataset from Seeking Alpha on readership of articles. The data only covers calendar year 2017, but contains daily number of “clicks” (i.e., number of times a given article is uploaded) and the number of times the article is “read,” which is the instances in which a reader scrolled to the end of the article.16 In total, the dataset covers 25,596 articles about 3,118 publicly traded firms.
Table C5 in the internet appendix presents results from regressing abnormal trading volume following the release of the article on the readership circulation of the article over the first three days after the article is published. The table shows that future abnormal trading volume is positively related to the number of clicks and number of times the article is read by consumers. This evidence suggests that the articles are directly related to future abnormal trading activity in the stocks the articles discuss.
The last two columns report results from regressions of the readership circulation variables
on the fake article dummy to examine whether readership is affected by article authenticity.
We find that fake articles are clicked more heavily and read more heavily, consistent with
those articles also affecting trading volume more. Fake news seems to disseminate faster
16We obviously do not know if it was actually read, but scrolling through the article implies that some time was spent on it.
Fake News: Evidence from Financial Markets 24
and more widely and impacts trading activity more. These results are consistent with fake news being more sensational and more persuasive, catering to the biases and priors of their consumers, and propogating more diffusely through the network as suggested by Allcott and Gentzkow (2017) and Vosoughi, Roy, and Aral (2018).
4. A Shock to Investor Awareness of Fake News
Our third empirical setting examines a different aspect of fake news to test another theoretical implication. We use the announcement of the initial SEC investigation into the promotional articles that comprise our first sample as an exogenous shock to the public awareness of fake news and examine the market’s response to news before and after this shock. This exercise does not require being able to detect fake content.
4.1. Galena Biopharma Inc.
We begin with the case of Galena Biopharma Inc., which was the first prosecuted by the SEC for stock price manipulation on knowledge-sharing platforms. Galena encompasses the “event” which made the public aware of the existence of fake news. It also provides a microstudy of the direct impact these articles have on the stock’s trading activity and prices as well as some of the motivation behind fake articles.
On October 31, 2014 the SEC filed a lawsuit in the United States District Court on behalf of all persons who bought Galena’s common stock between August 6, 2013 and May 14, 2014.17 Figure 3 depicts the stock price of Galena from April 2013 to May 2014, as well as the events that led 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 paid for. The articles did not disclose the payments that the authors received, which violated the terms of Seeking Alpha and SEC regulation, and in some cases falsely claimed not to have received any payment. The lawsuit documents
at least twelve promotional articles of this type. Appendix B contains an example of one of
17(Case 3:14-cv-00558-SI): securities.stanford.edu/filings-documents/1051/GBI00_01/20141031_r01c_14CV00367.
Fake News: Evidence from Financial Markets 25
the fake articles written about Galena.18
Figure 3 shows that Galena’s share price rose from about $2 to $7.48 between the summer
of 2013 and January of 2014. The publications of the fake articles are highlighted on the graph by the green boxes and often coincide with a bump in stock price on that day and a steady increase in price several days after. The motivation behind the scheme seems to have been a pump-and-dump campaign, as Galena insiders took advantage of the price rise through corporate actions and their own personal trading. On September 18, 2013 Galena sold 17,500,000 units of stock in a seasoned equity offering for net proceeds of $32.6 million. On November 22, 2013, Galena held a board meeting and granted stock options to executives and directors with a strike price of $3.88. In January 2014, after the stock price reached its highest level since 2010, seven Galena insiders sold most of their stock in less than a month, for more than $16 million. These events are highlighted in Figure 3, where as news of insider sales broke, the stock price declined dramatically.
In February and early March 2014, several investigative journalists published exposé ar-
ticles documenting the fraud, including in Barron’s and Fortune. On March 17, 2014 Galena
revealed in a 10-K filing that it was the target of an SEC investigation over the promotion.
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. Appendix A reports the 8-K form documenting the settlement. By that point
Galena’s stock price had dropped to $2 a share.19
18This article and others like it that are part of the SEC investigation have been removed from Seeking Alpha. Searching for this fake article today, Seeking Alpha displays a message stating: “This author’s articles have been removed from Seeking Alpha due to a Terms of Use violation.”
19Interestingly, while Galena is a relatively small firm, it was not an obscure one. For example, in July 2013, before the promotion started, it had a market cap of 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 facilitated trading in the company’s stock.
Fake News: Evidence from Financial Markets 26
4.2. A Shock to Awareness of Fake News
The public revelation of the SEC’s investigation and subsequent media attention around it provides a unique shock to investor awareness of fake news. We exploit the timing of the announcement to test additional implications of fake news.
In addition to the direct costs of individuals believing and acting upon false content, fake news can be costly if it damages people’s trust in news generally and causes them to discount legitimate news (Allcott and Gentzkow (2017), Kshetri and Voas (2017), and Aymanns et al. (2017)). Our unique setting provides an opportunity to measure the potential spillover effects of fake news on people’s trust in news. Using the revelation of the SEC investigation, we examine whether investors behaved any differently before versus after the event, when the presence of fake news on knowledge sharing platforms suddenly became salient to many consumers on these platforms.
4.3. Spillover Effects from Fake News
We use the period from February to March 2014 as the event that provides a shock to people’s awareness of fake news. We examine the propensity of fake news and abnormal trading activity associated with articles six months prior to and six months after the event (August 2013 to January 2014 and April 2014 to September 2014, respectively).
Panel A of Table 4 first examines whether the propensity of fake news declines after the scandal. We regress a dummy variable of whether the article was probabilistically fake, on a dummy for 6 months after the SEC announcement event, controlling for SEC filings, firm press releases, and lagged abnormal volume in the days leading up to the article’s publication. In addition, we include the number of news articles about the firm from the NYT and WSJ, which we obtain from Factiva. The coefficient on the post-scandal period is indistinguishable from zero, indicating that the prevalence of fake news, or more precisely the authenticity score of the fake articles, is similar before and after the scandal. However, this average result masks substantial heterogeneity. The next three columns separately report results for small, midsize, and large firms (defined as the smallest 10%, middle 80%, and largest 10% of firms,
Fake News: Evidence from Financial Markets 27
respectively, based on NYSE market cap breakpoints). The prevalence of fake articles about small firms fell significantly by 1.2% following the scandal. These results are consistent with small companies, who engaged or were willing to engage in promotional articles before the scandal, ceasing or decreasing this activity after the SEC announcement.
Panel B of Table 4 examines the impact of published articles on abnormal trading volume before versus after the scandal. The first column of Panel B reports results from a regression of abnormal volume on an article indicator, the 6-month post event indicator, and their interaction. The positive coefficient on articles confirms our earlier result from Table 2 that articles are associated with larger trading volume in the three days after they are published. The negative interaction term with the post-event dummy shows, however, that the effect of articles on trading volume decreases significantly after the scandal. This result is consistent with investors becoming aware of fake content and muting their trading response to news on these platforms. In addition, the strong negative coefficient on the post-event dummy indicates that abnormal trading volume, in general, declines after the scandal. This result suggests that people responded less to news in general on these platforms, including legitimate news, after the scandal and is consistent with consumers having less trust of news once aware of the existence of fake news, as theory suggests (Alcott and Gentzkow (2017)). The economic magnitude of the effect is large – a 4.2% drop in trading volume associated with news articles after the scandal relative to before the event.
Figure 4 examines the daily abnormal trading volume response for four trading weeks after the article is published. We estimate the following model:
Log(AbVol)t =α+β1Article×PostEvent+β2Article+β3PostEvent+Controls+ε
and plot the coefficient β1 at the daily level, with 95% confidence error bars. The graph displays the average trading volume reaction to all articles, after the scandal, and shows significant trading decreases on the day the article is published, and for the next two trading
Fake News: Evidence from Financial Markets 28
weeks, before eventually returning to pre-scandal levels. This result suggests that investors’ reaction to articles on these platforms decreases after the scandal, and does not rebound with higher trading volume at a later date.
While the results in the first column of Panel B control for the level of SEC filings, press releases, other media (e.g., WSJ, NYT articles), and lagged abnormal trading volume in the days leading up to the article’s publication, in the second column, we also interact the frequency of SEC filings, firm press releases, other news media, and changes in abnormal trading volume with the post-scandal dummy. The interaction terms serve as falsification exercises or “placebo” tests of the market responding to news on these platforms and the shock of fake news awareness on the platforms. In particular, an alternative explanation for the decline in trading volume in response to news after the scandal is that there is less information content, less news, or less firm activity in the post-event period that happened by chance to coincide with the timing of the SEC announcement. Or, perhaps, the trading volume response to news generally declines over time and being confounded by the SEC event. In either case, the interactions between corporate filings, press releases, and other media news would be negative as well. As the table shows, however, the interaction terms are negligible and insignificant, and two out of three have the wrong sign to be consistent with this alternative story. The magnitudes of these interactions with the post-event dummy are trivially small – 0.2% increase in trading volume response to SEC filings, 0.3% decrease to press releases, and 0.6% increase to other news media – none of which are remotely statistically different from zero. We find no discernible difference in firm news or activity before versus after the scandal and no reliable difference between the trading volume response to WSJ or NYT articles before versus after the scandal. This despite the fact that SEC filings, press releases, and other media news from the WSJ and NYT by themselves have a significant impact on trading activity: SEC filings, press releases, and newspapers articles increase abnormal trading volume by 13%, 29%, and 12.5%, respectively. However, after the scandal we find no difference in response to these other sources of news. We only find a
Fake News: Evidence from Financial Markets 29
decreased response to articles published on the social platforms.
Hence, the drop in trading volume associated with published articles on these social
media platforms is likely a reduced response from investors to news specifically coming from these platforms, and not press releases, other public filings, or other media, and not any market trends in information production or lower trading activity. The evidence is most consistent with investors discounting all news on these platforms, even legitimate news, after the scandal due to increasing distrust of content from these platforms after the SEC revealed the existence of some fake articles. The magnitude of the drop in abnormal volume is even larger and more significant after accounting for the other activity post-scandal, decreasing volume by 7.5% per article after the event. These findings provide some of the first evidence on the indirect spillover effects of fake news on news in general, as conjectured by theory (Allcott and Gentzkow (2017)). As we will show below, consumers were largely unaware of and unable to detect fake news, consistent with their response to discount all news on the platforms following their awareness of fake content.
Columns 3 through 7 of Panel B report results separately for small, medium, and large firms, as well as for firms with high and low retail ownership. Consistent with our previous results, these effects are all much stronger for smaller firms, and firms with high retail ownership. Post scandal, the abnormal trading volume associated with articles published on these platforms drops by 35% for the smallest firms. Interestingly, even though few fake articles are written about large firms and none of the firms in the SEC probe were large firms, abnormal trading volume still declines by 11.7% for each published article about large firms that appeared on these platforms after the scandal, despite nearly all of these articles being authentic. This result provides further evidence of a spillover effect from fake news to other legitimate news content. Stocks with high retail ownership have a 30.3% drop in abnormal volume post-scandal compared to only a 9.5% drop for low retail ownership stocks. Since retail investors tend to dominate participation on these sites, this result provides a more direct link to these platforms influencing trading activity.
Fake News: Evidence from Financial Markets 30
4.4. Generalizing Spillover Effects
The spillover effect from the awareness of fake news to all news, including legitimate news, is interesting and consistent with theory (Allcott and Gentzkow (2017)). The result begs the question: How broadly does the awareness of fake news from the scandal affect investors’ response to news generally? Was the spillover response merely contained to similar articles on Seeking Alpha, where many of the promotional articles the SEC investigated resided, or did it impact news from other sources? We can first look at our previous results from the falsification tests. For SEC corporate filings, press releases, and other news media (namely, the WSJ and NYT), we do not find a reliable spillover effect from the public’s awareness of fake news from the SEC investigation on the market’s trading activity response to news from these other outlets. The insignificant interaction effects between filings, press releases, and other news media with the post-scandal period suggests that investors respond to these sources in the same manner after their awareness of fake news on the social platforms. Thus, the drop in trading response only for the social platform articles is consistent with investors discounting all news on the social platforms, but recognizing or believing press releases, the WSJ, and NYT, are less subject to fake news, or that the average investor who trades on press releases is different than the average investor who trades on news from these social platforms. For these reasons, we think the filings, press releases, and other news media interactions provide compelling falsification tests that support our main findings.
While the indirect effects of fake news on these platforms do not seem to spillover to press releases or the WSJ or NYT, we find that they do spillover to other similar media outlets that were not part of the scandal. Specifically, as a test to generalize the spillover effect from fake news on other news more generally, we examine the trading response after the scandal for articles on the Motley Fool platform only. Motley Fool was not part of the fake news scandal, and none of its articles were flagged for failing to disclose paid-for content as part of a promotional scheme or were investigated by the SEC. Hence, we examine whether the spillover effect from the scandal, contained largely on Seeking Alpha, also had
Fake News: Evidence from Financial Markets 31
an effect on the trading volume response to articles published on Motley Fool, a similar shared-knowledge platform that was not part of the investigation. The last column of Panel B of Table 4 reports the results and shows, interestingly, that abnormal trading volume declined significantly for Motley Fool articles after the scandal. The result points to the spillover effect from the scandal extending beyond the specific platform where the scandal occurred. The awareness of fake news seems to impact other related news sources – in this case a competitor shared-information platform where the scandal did not occur. But, as our previous analysis shows, it does not have an impact on very different news sources such as press releases or newspaper articles, such as the WSJ and NYT. These results make sense if investors simply discount all social news as a result of the scandal but think other news sources are more immune to false content, or if the set of investors who consume social news is simply different from those who consume other news sources.
Panel C of Table 4 reports the results from the same regressions as in Panel B, but uses idiosyncratic volatility as the dependent variable instead of abnormal trading volume. The results are consistent with the trading volume findings, where there is significantly reduced impact on price volatility from articles after the information shock from the scandal, especially for small firms with high retail ownership. We also find opposite-signed price movement for press releases, SEC filings, and other news media after the scandal, suggesting other trends or omitted variables in the market’s response to news in general are not driving the results. We also find that price movements on the Motley Fool articles only are consistent after the scandal with the reaction to articles on Seeking Alpha, providing further evidence of a spillover effect to other shared-information platforms as a result of fake news awareness. These findings are consistent with markets discounting news from these platforms after revelation of the existence of fake news.
4.5. More Direct Evidence of Spillover Effects
We provide some additional evidence that the decline in trading volume per published article, and spillover decline in volume for non-fake news after the scandal, is due to investors
Fake News: Evidence from Financial Markets 32
being made aware of fake news. Specifically, we examine the posted comments to the articles published on these sites in the six months before and after the scandal. In the comments section pertaining to each article, we add up the mention of the words “fake” or “fraud” and compute a variable Fake Words, which is a dummy equal to one if readers use these words. We then regress the frequency of Fake Words on a dummy for fake articles as well as a dummy for the six-month period after the scandal.
To test an alternative hypothesis, we also compute the frequency of the words “wrong” or “not right” from the comments section and create a dummy variable Wrong Words, which is equal to one if readers use these words in their comments. This variable helps distinguish between erroneous or inaccurate information from fraudulent or deceptive information. The distinction is subtle because it relies on intent. The comments section provides a glimpse of what consumers may be concerned about.
Panel A of Table 5 examines whether the appearance of Fake Words or Wrong Words is more prevalent for fake versus non-fake articles over the entire sample period. We regress the prevalence of Fake Words on the fake article dummy in the first column and find that the words “fake” or “fraud” are not used more frequently with fake articles. This null result suggests that participants on these platforms could not identify or differentiate between fake and non-fake articles. In our setting, participants on these platforms were deceived by these articles with no indication that consumers were skeptical or aware of fake content.
The second column of Panel A runs the same regression but uses Wrong Words as the dependent variable. Here, there is a strong negative association between fake articles and use of the words “wrong” or “not right” in the comments section. This result suggests that investors feel the fake articles are more accurate (less wrong) than the non-fake articles. Fake articles seem to be more convincing of their statements than the non-fake articles, which may be why they generate more trading volume (and may be why they are used in the promotional campaigns).
Panel B runs similar regressions using the Post Event dummy instead of the Fake Ar-
Fake News: Evidence from Financial Markets 33
ticle dummy, where the post-event dummy is the six-month time period after the scandal. Interestingly, after the scandal, the incidence of the words “fake” and “fraud” increased signif- icantly (t-statistic of 2.73), implying that participants on these platforms were indeed more concerned with or commented more about false content on these sites after the scandal. This evidence corroborates the decline in trading volume witnessed post-scandal for all articles and suggests general mistrust of news from these platforms. The use of “wrong” words is no more prevalent after versus before the scandal. Hence, after the SEC announced investiga- tion and subsequent exposé articles, participants on these platforms seemed more concerned with fake news rather then erroneous news.
Combining the results in Panels A and B of Table 5, the evidence paints a picture of investors and consumers of information on these platforms being largely unaware of fake news before the SEC investigation and then suddenly becoming aware after the scandal, but having no ability to differentiate or detect which articles are fake and not fake. As a consequence, we see a marked drop in investor trading volume to any articles published on these sites, regardless of their authenticity, creating a significant spillover effect from the revelation of the existence of fake news on legitimate news more generally.
To further examine the link between the articles and trading volume, we examine whether authors who have more followers, and have written more articles, have a bigger impact, as well as whether articles that receive more comments lead to greater trading volume. We also analyze whether the trading volume reaction is higher when the article is more quantitative in nature and/or references accounting data, where presumably it is less likely to be false since numbers, such as earnings, can be verified from other sources. We look at the fraction of the article text comprised of numbers, as well as the number of words that have “earn” as part of the word. We regress abnormal trading volume for stock i on the number of followers an author has, log number of comments an article received, the number of past articles the author has written, and the fraction of numbers that appear in the text as well as the fraction of mentions of “earn.” We also control for whether there was at least one SEC filing and one
Fake News: Evidence from Financial Markets 34
firm-issued press-release in the three trading days leading up to the publication date, and control for abnormal trading volume the day before the article’s publication. The results are presented in Panel A of Table 6. We find that articles by authors with more influence as well as articles that get more comments are associated with a larger impact on trading volume. Furthermore, articles that seem to be more quantitative, also have a bigger impact.
Next we examine whether these characteristics have an impact on how the SEC scandal affected the trading volume reaction to the article. In particular, we rerun the regressions from Panel A, examining the time period six months before and six months after the SEC scandal. The results are presented in Panel B of Table 6. Similar to earlier analysis, we find that the trading volume response to articles is lower in the post-period. However, the drop is not as large if the author has more followers and has written more articles in the past. This suggests that people’s trust in the articles decreases less for authors that have a better reputation. We further find that the decrease is not as large for articles that mention “earn” in the article, suggesting that articles that cover accounting-related and hard information are not discounted as much after the scandal.
5. What Motivates Fake News?
Finally, we investigate what might be motivating the fake articles on these platforms. Using the Galena case that launched the broader SEC investigation, we examine whether other cases have similar characteristics and motivations to better understand the existence and prevalence of fake news. This analysis serves several purposes: it may help us better quantify the economic impact of fake news, provides another test of the linguistic algorithm’s ability to detect fake content, and may help identify other fake news.
5.1. Firm Performance
We start by examining the price reaction to the other for-sure fake articles from the SEC to see if a similar pattern as Figure 3 for Galena exists for the other firms involved in the scandal. We conduct the flip-side of the classic event study in financial economics
Fake News: Evidence from Financial Markets 35
(Fama et al. (1969)) by examining the return response to false news. This exercise is a novel test of the informational efficiency of markets, where in a perfectly efficient market fake news should have no impact on prices, regardless of the underlying equilibrium asset pricing model. We separate firms by size into small and non-small (there are no large firms in this sample) and examine their return response to the release of for-sure fake articles, by plotting the cumulative abnormal returns, measured as the difference between the return of the stock and a matched-portfolio of similar stocks (one of 125 equal-weighted portfolios based on size, book-to-market equity, and momentum), for days t + 1 to t + 251 after a fake article appeared about the firm.
Figure 5 plots the difference between cumulative abnormal returns for the for-sure fake articles, relative to days with non-fake articles. Returns for small firms increase after the fake article is published (relative to non-fake articles), reaching as much as 13%, cumulatively, after about 60 days, before giving up all the gains, and ending with a cumulative negative 5% return after a year. This pattern matches that of Galena in Figure 3. The permanent price impact of −5% for small firms indicates either that once the market figures out the news is fake, investors view this as a bad signal about the firm or that the true price should have dropped by 5% initially, but the fake news temporarily propped up the price and delayed the decline. For non-small firms, the price starts dropping immediately after the fake article comes out and continues to decrease throughout the year. This result could be consistent with the market figuring out the news is fake immediately for larger firms, where the cost of information is lower, or that the returns would have been even worse had the firm not initiated the fake articles.
We next examine the market price response to articles that we classify as probabilistically fake using the linguistic algorithm on the larger universe of all articles on these platforms. Since our analysis is at the firm-day level, we define whether a firm had a fake article on a given day using the average probability of being fake of all articles written about the firm on that day. Figure 5 plots the difference between abnormal cumulative returns following days
Fake News: Evidence from Financial Markets 36
with (probabilistically) fake articles, relative to days with (probabilistically) non-fake articles, and plots the responses separately for small, mid-size, and large firms in our sample (that have at least one fake article). As the figure shows, among small firms, returns following fake articles relative to non-fake articles increase for 6 months by about 5% following publication, and then revert back to their original level. These patterns are remarkably similar to the return patterns we found for the for-sure fake articles from the smaller SEC sample, further supporting our algorithm to identify fake news. The magnitudes are, not surprisingly, much smaller here since, unlike the SEC example, we identify fake news with noise plus the SEC is likely to go after the most extreme cases, so there may be selection bias. For larger firms, we find nothing, which makes sense since the market is more efficient for larger firms who also are less likely to engage in promotional campaigns. The lack of results on larger firms is another useful falsification exercise. We formally test whether the patterns in cumulative abnormal returns for fake news articles about different-sized firms over different horizons are statistically significant in Table C6 in the internet appendix. We find statistically significant results for small firms and no impact for larger firms.20
5.2. Other Firm Actions
Fake news is designed to deceive for financial or personal gain, including perhaps the
utility of fooling people and/or influencing others. In our setting of financial markets, it seems
less likely that private utility benefits motivate fake news. The SEC investigation focused
on promotional articles as part of pump-and-dump schemes to defraud securities markets.
Our findings on the impact on abnormal trading and temporary prices are consistent with a
motivation to hire authors to write fake content to promote the stock. Consistent with this
motive, Table ?? shows that these firms are more likely to issue press releases and 8-K filings
20One question is whether the poor long-term returns to small firms that promote fake articles are due to investors’ over/under reaction or whether fake articles are a sign of poor fundamental firm performance. Table C7 in the internet appendix shows that the presence of fake articles is associated with worsening fundamental firm performance, as measured by surprise in unexpected earnings, the return on assets, and its recent quarterly change. These findings are consistent with a possible motivation for engaging in promotional campaigns for financially troubled small firms that include hiring fake articles to prop up the stock price.
Fake News: Evidence from Financial Markets 37
within the same week to accompany the fake articles, perhaps to give authors of the fake articles more material and credibility and to influence the narrative of the firm. We also find in Table C9 that insider trading coincide with the fake articles and is positioned to profit fromthepricemovementcausedbythepromotion.21 Whiletheseactionsarerampantamong the SEC-prosecuted sample, we also find similar evidence for our broader sample of articles where we probabilistically assess the occurrence of fake news using the linguistic algorithm. Consistent with our earlier results, we find these effects, too, to be predominantly contained among small firms. This provides additional support to the efficacy of our methodology to detect fake articles more broadly. Moreover, the evidence suggests that an improved method for detecting fake content may involve examining other actions taken by the firm in addition to textual analysis. As one example, when we combine the probability of fake articles with the dual presence of insider trading to benefit from stock promotion, we find sharper price impact patterns. In addition, fake articles published following insider purchases are preceded by very sharp drops in share price in the month before publication, whereas fake articles not associated with insider purchases have flat to lightly increasing returns before publication. However, even for firms with fake articles written about them that do not have insiders buying shares, there is still a small price increase that also turns negative after several months, suggesting the results are not just driven by insider trading. In addition, performing a similar analysis using only the non-fake articles, there is no difference in returns for non-fake articles with insider buying versus without insider buying. Hence, it is not insider buying per se that drives the returns. Rather, it is the combination of insider buying with fake articles that seems to matter most and is indicative of a comprehensive promotional campaign.
6. Conclusion
We study three empirical settings to assess the economic impact of fake news: a unique dataset of fake paid-for articles on financial media crowd-sourced platforms prosecuted by
21We obtain data on press releases from RavenPack from 2001 to 2015, 8-K disclosure filings from the SEC’s Edgar database, and insider trades from Form 4 from Thomson Reuters.
Fake News: Evidence from Financial Markets 38
the SEC, a broader set of articles on these platforms that we apply a linguistic algorithm to detect fake content, using the first sample of known fake articles to calibrate the algorithm, and the SEC’s announced investigation that provides a shock to the public’s awareness of fake news. We find that fake news increases abnormal trading volume and imposes temporary price impact on small firms. Following public revelation of the existence of fake news, we find a significant spillover effect to news generally, where investors react less to all news, even legitimate news on these platforms. These findings represent some of the first documented direct and indirect effects of fake news that are consistent with theory (Allcott and Gentzkow (2017), Aymanns et al. (2017), and Kshetri and Voas (2017)).
Our study provides evidence on the prevalence and effect of fake news from crowd-sourced information platforms that continue to grow and gain attention. Financial markets may provide a lower bound on the impact of disinformation in other settings, where information costs are higher and where the ability to take action to correct its distortions is more limited (e.g., online consumer retail, political news, elections, and social media). More broadly, our findings may have implications for news media generally (e.g., Gentzkow and Shapiro (2005) and Gentzkow et al. (2015)) and for trust and social capital (e.g., Guiso et al. (2004), Guiso, Sapienza, and Zingales (Guiso et al.), Guiso et al. (2010), and Sapienza and Zingales (2012)).
Fake News: Evidence from Financial Markets 39
References
Allcott, H. and M. Gentzkow (2017). Social media and fake news in the 2016 election. The Journal of Economic Perspectives 31(2), 211–235.
Antweiler, W. and M. Z. Frank (2005). Is all that talk just noise? the information content of internet stock message boards. The Journal of Finance 59(3), 1259–1294.
Aymanns, C., J. Foerster, and C. Georg (2017). Fake news in social networks. CoRR abs/1708.06233.
Boudoukh, J., R. Feldman, S. Kogan, and M. Richardson (2018). Information, trading, and volatility: Evidence from firm-specific news. The Review of Financial Studies, hhy083.
Das, S. R. and M. Y. Chen (2007). Yahoo! for amazon: Sentiment extraction from small talk on the web. Management Science 53 (9), 1375–1388.
Fama, E. F. (1970). Efficient capital markets: A review of theory and empirical work. The Journal of Finance 25(2), 383–417.
Fama, E. F., L. Fisher, M. C. Jensen, and R. Roll (1969). The adjustment of stock prices to new information. International Economic Review 10 (1), 1–21.
Gentzkow, M. and J. Shapiro (2005, October). Media bias and reputation. Working Paper 11664, National Bureau of Economic Research.
Gentzkow, M., J. Shapiro, and D. Stone (2015). Media bias in the marketplace: Theory. handbook of media economics, simon anderson, david stroemberg and joel waldfogel, eds.
Grossman, S. J. and J. E. Stiglitz (1980). On the impossibility of informationally efficient markets. The American Economic Review 70 (3), 393–408.
Guiso, L., P. Sapienza, and L. Zingales. Trusting the stock market. The Journal of Fi- nance 63(6), 2557–2600.
Fake News: Evidence from Financial Markets 40
Guiso, L., P. Sapienza, and L. Zingales (2004, June). The role of social capital in financial development. American Economic Review 94(3), 526–556.
Guiso, L., P. Sapienza, and L. Zingales (2010, March). Civic capital as the missing link. Working Paper 15845, National Bureau of Economic Research.
Heston, S. L. and N. R. Sinha (2017). News vs. sentiment: Predicting stock returns from news stories. Financial Analysts Journal 73 (3), 67–83.
Jegadeesh, N. and D. Wu (2013). Word power: A new approach for content analysis. Journal of Financial Economics 110(3), 712–729.
Kshetri, N. and J. Voas (2017, November). The economics of “fake news”. IT Profes- sional 19(6), 8–12.
Milgrom, P. and N. Stokey (1982). Information, trade and common knowledge. Journal of Economic Theory 26(1), 17–27.
Mullainathan, S. and A. Shleifer (2005, September). The market for news. American Eco- nomic Review 95(4), 1031–1053.
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.
Sapienza, P. and L. Zingales (2012). A trust crisis. International Review of Finance 12(2), 123–131.
Fake News: Evidence from Financial Markets 41
Tetlock, P. C. (2007). Giving content to investor sentiment: The role of media in the stock market. The Journal of Finance 62 (3), 1139–1168.
Vosoughi, S., D. Roy, and S. Aral (2018). The spread of true and false news online. Sci- ence 359(6380), 1146–1151.