Print
E-MailShimon Kogan is Associate Professor of Finance at IDC Herzliya; Tobias J. Moskowitz is the Dean Takahashi ’80 B.A., ’83 M.P.P.M. Professor of Finance at the Yale School of Management; and Marina Niessner is Vice President at AQR Capital Management. This post is based on their recent paper.
An increasing number of professional and retail investors obtain information about financial markets from knowledge sharing platforms. For example, a 2015 study by Greenwich Associates found that 48% of institutional investors use social media to “read timely news.” While crowd-sourced outlets can lower the cost of information acquisition and speed its dissemination, they also provide a venue for interested parties to spread fake information in an attempt to manipulate the markets. In this paper, we employ a methodology developed by linguistic psychologists to identify a large set of fake articles on financial knowledge sharing platforms. We document their prevalence on these platforms, and examine the effect of fake news on volume, volatility, and prices. We further document a broader spill-over effect of fake articles on all news posted on knowledge sharing platforms. Using a clean natural experiment, we show that after investors are made aware of the presence of fake news on knowledge sharing platforms, the effect that fake and non-fake articles have on the trading volume and volatility goes down substantially.
Specifically, in order to identify fake articles, we turn to a linguistic algorithm developed by James Pennebaker and his colleagues (Pennebaker, et al., 2015) to identify deception in spoken and written language. The methodology is summarized by the authenticity measure in LIWC software. While the exact formula for the authenticity score is proprietary, Pennebaker (2011) describes which linguistic traits are associated with honesty. In particular, 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.
However, since the methodology was developed using personal essays and criminal writing, and linguistic methods can be very subject-specific, we first need to verify that the algorithm works in a financial setting. In order to do so, we turn to a scandal that erupted in early 2014, when a contributor to Seeking Alpha, Rick Pearson, was approached by a PR firm to write promotional articles on Seeking Alpha, without disclosing the payments. Rick went undercover and turned over his findings to the SEC, which then filed several lawsuits against implicated authors and companies. We obtain a unique dataset of for-sure fake paid-for articles from Rick and the SEC investigation. We then use this clean, albeit small, set of for-sure fake promotional articles to validate the linguistic authenticity measure. The algorithm produces statistically significantly different authenticity scores for for-sure promotional and non-promotional articles that were written by the same authors. Thus, a unique and critical advantage of our study is that we can use the for-sure fake articles to validate the methodology, and to calibrate the authenticity score into a probability of fake news. We then apply this measure to a much larger set of news articles published on Seeking Alpha and Motley Fool between 2005 and 2015, and identify probabilistically fake and non-fake articles. By our measure, the prevalence of fake news on these platforms is not insignificant and varies meaningfully through time: we classify 2.8% of articles as fake, with the frequency peaking in 2008 at 4.8%. We then use these articles to study the effects of fake news on non-fake news and on financial markets, more broadly.
First, we examine the direct impact of fake news on trading activity. We find that abnormal trading volume, which is a gauge of investor activity, rises on the days that articles appear on these platforms. Furthermore, looking specifically at the small SEC sample of for-sure fake articles, we find an even larger trading response to fake news relative to non-fake articles published at the same time on the same platform. This is likely driven by fake articles often being more sensational and diffusing more quickly across consumers (Vosoughi, Roy, and Aral (2018)). Turning to the broader set of probabilistically fake articles, we find similar results. The direct effect on trading is stronger for smaller firms with higher retail ownership and for articles with greater circulation (measured by number of clicks and readers of each article), lending credence to these platforms influencing investor behavior.
We then use a setting where readers are explicitly made aware of the presence of fake news on knowledge sharing platforms, to examine whether there is a spill-over effect on how much investors trade on regular, presumably non-fake, news. While fake articles on these platforms have appeared for a long time, in early 2014, news outlets published stories highlighting the promotional schemes on Seeking Alpha and other platforms, including an announcement of an SEC investigation. Using the announcement as an exogenous shock to readers’ awareness of the presence of fake articles on these platforms, we find a marked decrease in the probability of false content appearing on these platforms as well as a decrease in reaction by investors to all news, even legitimate news. These findings are the first to document some of the effects of fake news empirically and are consistent with theoretical models of fake news.
We then turn to pricing effects to see if fake news move prices in a distortive way. Using the small sample of for-sure fake articles from the SEC lawsuit, we find that the fake promotional articles are able to pump up the stock price for small companies, which subsequently gets fully reversed over the course of a year. Mid-size firms, however, experience a permanent negative price impact when fake articles are written about the firm. Looking at the broader set of probabilistically fake articles, we first find that the incidence of fake content is higher for small firms and very low for large firms. We similarly find strong temporary positive price effects for smaller firms, that then fully reverse and turn negative, immediate negative returns for mid-size firms and no price impact for large firms. These results mirror those from the SEC sample and suggest our methodology for detecting fake news is valid.
The complete paper is available for download here.
