Should Retail Investors Listen to Social Media Analysts? Evidence from Text-Implied Beliefs

Abstract

Social media is increasingly affecting financial markets, with important implications for market efficiency. This paper uses machine learning to infer beliefs of nonprofessional social media investment analysts (SMAs) from opinions expressed about individual stocks on social media. On average, SMA beliefs are informative about future abnormal returns and earnings surprise. However, there exists substantial heterogeneity in SMAs' ability to form beliefs that generate value. A small fraction, 10%, of SMAs form beliefs that yield a sizeable abnormal return of 56 bps over a 5-day window, while the remaining 90% generate only 6 bps over the same horizon. SMA characteristics such as specialization, skin in the game, effort, popularity, and disagreement matter for belief formation skill. When forming beliefs, SMAs herd on the consensus; herding is less pronounced in bad times and when the consensus is optimistic, but more pronounced when the consensus is correct ex-post. SMAs also extrapolate from past returns but are less bullish about lottery-type stocks.

Keywords: Nonprofessional analysts, Belief formation, Investor skill, Market efficiency, Herding, Extrapolation, Machine learning, Natural language processing

JEL Classification: G11, G12, G14