fake news in financial markets - gsb.columbia.edu · •we have 171 for-sure fake and 334 non-fake...

Post on 27-May-2020

1 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Fake News in Financial MarketsShimon Kogan

(MIT Sloan School of Management, IDC)

Tobias J. Moskowitz

(Yale School of Management, NBER, AQR)

Marina Niessner

(Yale School of Management)

Detecting fake news is hard:

Detecting fake news is hard:

A study by Sam Wineburg et. al. (2016) from Stanford, assesses students’ ability to distinguish ads from articles, neutral sources from biased ones and fake accounts from real ones.

The authors described the results as “dismaying,” “bleak” and “threat to democracy.”

Does Fake News matter in Finance?

While the existence of fake new may be problematic in certain settings, it is

not clear it matters for financial markets

In a perfectly efficient financial securities market, where the cost of

information approaches zero, misinformation will have no impact

Does Fake News matter in Finance?

While the existence of fake new may be problematic in certain settings, it is

not clear it matters for financial markets

In a perfectly efficient financial securities market, where the cost of

information approaches zero, misinformation will have no impact

How prevalent are fake news on knowledge sharing platforms?

What effect do fake news have on financial markets?

Detecting Lying

How can we detect lying?

• Polygraph tests (about 60-65% accuracy)

• Could try linguistic methods. Do we speak differently when we lie?

Detecting Lying

How can we detect lying?

• Polygraph tests (about 60-65% accuracy)

• Could try linguistic methods. Do we speak differently when we lie?

It turns out the answer is yes:

• When people lie they use fewer self-references (I-words)

• They also tend to use more discrepancy verbs, like could, that assert that an event might have occurred, but possibly didn't

Detecting Lying

Our Paper

Use a linguistic measure of authenticity developed by James Pennebackerand co-authors (LIWC)

Obtain fake news articles that were published on knowledge sharing platforms, like Seeking Alpha, to test the measure in financial setting

Estimate the prevalence of fake news on financial information sharing platforms, and which authors

Examine the effects of fake news on returns and insider trading

Data

Knowledge Sharing Platforms

Seeking Alpha:

A moderated, crowd-sourced content service for financial markets

Was founded in 2004

Have 7 million unique visitors generating 40 million visits monthly (as of 2017)

Our data is 2005-2015, and contains 203,545 articles

Motley Fool:

Financial-services company that provides financial advice for investors through various stock, investing, and personal finance services

Was founded in July 1993

Alexa’s rank in the US is 631 (vs. 546 for Seeking Alpha)

Our data is 2009-2014 , and contains 147,916 articles

Knowledge Sharing Platforms

Such platforms can enhance the speed with which information is disseminated and lower the cost of obtaining information

Also offer scope for providing misleading or false information

Knowledge Sharing Platforms and Fake News

2014 Rick Pearson went undercover to investigate other authors and uncovered fake paid-for articles

He turned over the evidence to the SEC

SEC filed lawsuits against individuals, PR firms, and some companies on 31 October, 2014 and on 10 April, 2017

We obtained 171 articles by 20 authors about 47 firms

We further obtain 334 articles by the same authors that are non-fake

Linguistic Inquiry and Word Count Software

LIWC was developed by James Pennebaker and his colleagues

Outputs the percentage of words in a given text that fall into one or more of over 80 linguistic, psychological, and topical categories

From “The secret life of pronouns: what our words say about us”

From “The secret life of pronouns: what our words say about us”

Discrepancy words (examples: should, would)

Insight words (examples: think, know)

Relativity words (information about location and time)

Fake vs. non-fake articles by same authors

Fake news have: Much lower authenticity scores

Much fewer self-references

Shorter sentences

Lower insight score (examples: think, know)

Lower relativity score (information about location and time)

Higher discrepancy score (examples: should, would)

Higher clout score

Authenticity: For-sure Fake vs. Non-fake Articles

Identifying Fake Articles

in Overall Sample

Mapping into Probabilities

Generate the authenticity score for all articles published on Seeking Alpha and Motley Fool

Not clear how to interpret the cardinal nature of the authenticity score

Develop a mapping of the authenticity score into the probability space

Mapping into Probabilities

Authenticity score vs. conditional probability of being fake (Prob(F|S))

• Classify an article as fake if

P(Fake) > 20%

• Classify an article as non-fake

if P(Fake) < 1%

FakeNon-Fake

Other

Average probability of being fake:

Identifying Fake Articles

Type I and Type II errors

• We have 171 for-sure fake and 334 non-fake articles

• We use our algorithm to classify these articles into “fake articles” and “non-fake articles”

• Our algorithm identifies 18 articles as fake a 165 as non-fake

Our Algorithm For-sure Fake Non Fake Total

Fake 18

Non-Fake 165

Rick's and SEC articles

Type I and Type II errors

• We have 171 for-sure fake and 334 non-fake articles

• We use our algorithm to classify these articles into “fake articles” and “non-fake articles”

• Our algorithm identifies 18 articles as fake a 165 as non-fake

Our Algorithm For-sure Fake Non Fake Total

Fake 17 1 18

Non-Fake 165

Rick's and SEC articles

Our Algorithm For-sure Fake Non Fake Total

Fake 17 1 18

Non-Fake 14 151 165

Rick's and SEC articles

Type I and Type II errors

• We have 171 for-sure fake and 334 non-fake articles

• We use our algorithm to classify these articles into “fake articles” and “non-fake articles”

• Our algorithm identifies 18 articles as fake a 165 as non-fake

Impact of Fake News

on Financial Markets

We examine cumulative abnormal returns (measured as equal-weighted 4-factor residuals) for for-sure fake articles (171 articles)

Return Reaction: Rick’s and SEC articles

Small Firms Medium Firms

Return Reaction and Firm Size (all articles)

Small Firms Medium Firms

Large Firms

Insider Trading

and

Firm Actions

Fake Stories and insider trading: weekly

Fake Stories and insider trading: weekly

Conclusion

First paper in Finance to systematically identify fake articles on knowledge sharing platforms, and examine their prevalence

Document the effect of fake news on financial markets

Provide evidence of insider trading around the fake articles

Thank You!

Impact of Articles and Article Characteristics

Summary Statistics

top related