The Crowd Was Right: Stock Opinions on Social Media Actually Predict Returns

For most of financial history, investment opinion was produced by a licensed priesthood. Sell-side analysts at investment banks, with CFA charters, company access, and the institutional apparatus behind them. If you wanted a view on a stock, you read their research.

Then the internet happened, and anyone with a keyboard could publish an investment thesis. The professional reaction was predictable and dismissive: this is noise, it is amateurs talking to amateurs, it is a stock promotion scheme with a comment section.

Hailiang Chen, Prabuddha De, Yu Jeffrey Hu and Byoung-Hyoun Hwang went and read the amateurs. Then they checked what happened to the stocks.

The problem: is crowd opinion signal or noise?

There is a real theory on both sides of this, which is what makes it a good question.

The case for noise. Retail investors are famously prone to bias. They chase performance, they overtrade, they buy attention-grabbing stocks. Aggregating a lot of biased opinions does not average out the bias, it just gives you a very confident wrong answer. And an open publishing platform invites manipulation: people who own a stock have every incentive to write glowingly about it.

The case for signal. This is the wisdom-of-crowds argument, and it has real pedigree. If many people each have a small, independent piece of information, and their individual errors are random, then aggregating them cancels the errors and concentrates the information. A crowd guessing the weight of an ox is famously better than most individual experts. And unlike sell-side analysts, internet commentators have no investment banking relationship to protect, no fear of losing management access, and no institutional pressure to stay bullish. They can say a company is bad, which the sell side structurally struggles to do.

So which is it? This is an empirical question, and it had never been properly answered.

The key idea via analogy: read the crowd, then check the scoreboard

The design is refreshingly direct.

Take a major social media platform where individual investors publish articles and commentary about specific stocks. Gather the text. Run textual analysis over it to measure the tone of each article: how positive or negative is the author about the stock? Aggregate that tone across all the articles about a given company in a given period.

Now you have a number: the crowd's view of this stock. And you have a clean, objective scoreboard to check it against: what the stock actually did next, and whether the company beat or missed earnings.

The findings were clear.

The crowd's opinions predicted future stock returns. More positive aggregate tone was followed by higher returns. The relationship was not noise. The crowd was, on average, right.

The crowd's opinions predicted earnings surprises. This is the stronger and more interesting result. It is one thing to predict returns, which could be a story about sentiment temporarily pushing prices around. It is quite another to predict whether a company will beat or miss its earnings, because earnings are a hard fact reported by the company, and no amount of collective enthusiasm can move them.

That second finding is what elevates the paper. If the crowd merely moved prices, you would conclude they were causing sentiment-driven mispricing, the kind of thing Tetlock found in newspaper tone. But predicting earnings surprises means the crowd knows something. They are aggregating genuine information about the company's business, not just generating hype.

Where does that information come from? The plausible story is that the crowd contains people with real, dispersed, local knowledge. Customers who noticed the stores are busier. Employees who know a product launch is going well. Suppliers who saw an order increase. Enthusiasts who have read every filing more carefully than the two overworked analysts covering the stock. None of them alone has an edge. Collectively, they have a picture.

The authors also examined the comments on the articles, not just the articles themselves, and found the crowd's discussion added information beyond the original author's view. The crowd was not just a broadcaster. It was a filter and a corrective.

Why it mattered

• It legitimised social media as a financial data source. Before this paper, "we scrape investor forums" was something you said quietly. After it, in the Review of Financial Studies, it was a documented, peer-reviewed source of predictive information. An entire industry of social sentiment data providers grew in this direction.
• The earnings-surprise result is the killer finding. Predicting returns is ambiguous, it can always be a sentiment story. Predicting the actual earnings number is not ambiguous. It means the information content is real, and it forces you to take the crowd seriously rather than dismissing it as noise.
• It formalised the wisdom-of-crowds mechanism in markets. The theory was old. This was a clean, large-scale, financial demonstration that aggregating dispersed amateur opinion produces genuine information, in exactly the domain where you would most expect the professionals to dominate.
• It was an unflattering comparison for the sell side. The implicit contrast with professional analysts, who are conflicted, herd toward consensus and rarely say sell, is not subtle.
• It anticipated everything that followed. The rise of retail communities as a genuine market force, culminating in episodes where coordinated retail attention moved large stocks, was foreshadowed by the finding that these communities carry real information and real influence.

The honest limitations

• The sample period is a specific, benign one. The platform studied was a curated environment with editorial standards and a community of relatively serious contributors. It is not a general result about "the internet." The retail investing conversation has since moved to far less curated venues, and the wisdom-of-crowds logic depends on independent, honest opinions, which is exactly what a hype-driven forum does not have.
• Independence is the fragile assumption. The crowd is only wise if the individual errors are independent. When everyone reads the same viral post and repeats it, errors are correlated, and the crowd stops aggregating information and starts amplifying a single view. Modern social platforms are engineered to produce exactly that correlation.
• Manipulation is a persistent hazard. Anyone with a position has an incentive to post about it. Paid promotion of small-cap stocks through investor forums is a well-documented practice. The platform studied had some safeguards; many do not.
• Predicting is not profiting. The paper documents predictive power. It does not establish a strategy that survives transaction costs, and the effects are strongest in exactly the smaller stocks where trading is most expensive.
• Publication decays the edge. Once this was known, sentiment data providers commercialised it, and quant funds began consuming social data at scale. The specific edge documented here is not sitting on the table.

The one-line takeaway

Chen, De, Hu and Hwang read the amateur stock articles on a social media platform and found that the aggregated crowd view predicted not only future returns but actual earnings surprises, which means the crowd was not generating hype but genuinely aggregating dispersed information, and which turned social media from a punchline into a data source.