Why stock sentiment depends on who is posting

A stream of posts about a company can look like a crowd taking its temperature. But a thousand posts need not represent a thousand independent, well-informed views. An economics preprint asks whether English-language stock sentiment becomes less informative when the people who know a company best are underrepresented in the discussion.

Social-media sentiment turns posts into numbers that investors can compare with market returns. The difficulty is knowing what those numbers contain: useful information, repeated commentary, passing enthusiasm or automated output. A new preprint by Piyush Akimitsu and Ulrich Hounyo examines that problem by comparing U.S.-headquartered firms with foreign-headquartered firms whose stocks also trade on U.S. exchanges.

Their proposed explanation is a mismatch between who posts and who knows the company. Investors in a foreign firm's home market may understand its management, products and regulatory environment particularly well, yet contribute relatively little to its English-language Twitter discussion. The study finds patterns consistent with that explanation, without observing individual contributors' locations or knowledge.

What the researchers measured

The authors combine Context Analytics' daily Twitter sentiment metrics with stock-market data from CRSP and company fundamentals from Compustat. Their primary analysis covers 2018–2024, avoiding a vendor sampling change in 2017. The main structural estimation uses 96,228 firm-quarter observations across 5,583 firms.

Rather than simply asking whether posts sound positive or negative, they measure how much daily sentiment per tweet fluctuates within a quarter. They call this sentiment noise. It is an observable proxy interpreted through their model—not a direct check of whether individual posts are accurate.

They then examine the following quarter's losses, defined as the sum of negative daily returns after subtracting the risk-free rate. This is not the same as a stock's net quarterly return: a volatile stock can accumulate both large daily losses and large daily gains.

A relationship that changes across borders

Across 27 usable quarters, higher sentiment noise was associated with deeper subsequent losses in roughly 78 percent of quarters for domestic stocks, compared with 48 percent for foreign stocks. The difference was statistically significant in the authors' block-bootstrap test, which resamples consecutive groups of quarters.

Those percentages describe the sign of a statistical relationship across quarters. They are not a trading strategy's success rate or the probability that an individual stock will fall.

The authors also compare stocks within three groups formed using a model-based proxy for informed trading. The domestic–foreign gap in sign consistency is significant only in the lowest-ranked group, not throughout all three. The result therefore does not establish a uniform reliability gap across every market segment.

What repeated posting can—and cannot—reveal

To interpret these patterns, the paper models three contributor types: specialists who post repeatedly with company information; occasional, attention-driven posters whose messages mix information with a temporary fad; and automated accounts that post repeatedly without information about the firm's underlying value.

The vendor provides aggregate counts, not account identities. A key measure is the number of distinct posting accounts divided by the number of tweets. A lower ratio indicates more concentrated posting, not greater disagreement about the stock.

Under the model's assumptions, foreign-stock feeds have higher estimated repeat-author shares and a larger fad component. Their median signal-quality index is also lower in each ranking group, by about 17, 16 and 7 percent respectively. That index combines the number of days with scored posts and the variability of daily sentiment per tweet.

Repeated posting alone does not identify expertise or automation. The data cannot separate specialist accounts from automated accounts. Moreover, translating concentration into author shares requires an assumed posting rate; the headline estimates fix it at 1.5 tweets per repeat author per day. The estimated direction of the domestic–foreign differences survives alternative rates considered by the authors.

The proposed price-correction mechanism did not hold up

The model allows optimistic sentiment errors to inflate prices, with short-selling constraints preventing a full correction until the following quarter. But the estimated fraction of overpricing that survives reaches its zero lower bound. Losses also do not respond more strongly to noise after more optimistic quarters, as that mechanism predicts.

Instead, the loss–noise association becomes statistically insignificant after controlling for realized return volatility. This is consistent with a volatility explanation: noisier feeds accompany larger market swings, rather than demonstrating that sentiment pushes down average future returns.

Where the evidence stops

The structural interpretation has important assumptions. It uses a common conversion factor to translate the signal-quality index into information precision, and its baseline attributes nonfundamental variance to fads by setting automated-content variance to zero. A formal test rejects the model's full set of restrictions, with liquidity measures accounting for most of its weighted misfit.

The domestic–foreign contrast is also strongest for American Depositary Receipts, whose U.S. prices are linked to home-market shares. The mean-slope difference for directly listed foreign shares has the same direction but is statistically imprecise.

The November 2026 arXiv preprint offers evidence that feed composition matters when interpreting sentiment—not proof of a causal mechanism or a ready-made investment rule. Its findings concern English-language Twitter and U.S.-listed stocks; they do not establish how local-language platforms would compare.