LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
Emoji sentiment can add information about later stock returns after text sentiment is already in the regression. That is the result of a talk to a San Diego student trading club, given to a crowd of over 70. The method does not hand the emojis a dictionary. Surrounding words score them, year by year.
Why it matters
Language moves. "To the moon" and a rocket emoji mean a rise to a crypto audience now. A decade or more ago they need not have meant that in a market. Off-the-shelf sentiment tools handle emojis and do not track that drift. The research questions are whether token meanings move, whether a meaning can be measured at a date, whether that measurement works for emojis, and whether emojis then add something about cross-sectional returns that ordinary text does not. The lecture's warning about the word "forecast" is that the output is not a prediction of a path. It is an association in a regression.
Most of the job is the data. About 95 percent of the effort is collection and cleaning, and more than 98 percent of questions end inconclusive. The universe and the price vendor have to be chosen on purpose. Cashtags were kept and bare company names were dropped, so a piece of fruit did not become its own model. A named-entity pass might add signal and would itself need a defense.
The idea in plain words
Score each post with a general sentiment model aimed at words whose polarity is relatively stable: amazing, great, good, bad, terrible. Average those scores across posts that contain a given emoji, and that average is the emoji's score at that date. A central-limit argument is why the talk expects the average to look roughly normal. A rocket in 2000, next to posts about billionaires and flights, comes out near zero. The same emoji in 2018 is scored again from that year's posts. Nothing in the procedure types "rocket means up."
Applied to the collection, the scores came out in the intuitive direction anyway. A rising chart, green, a rocket, and a happy face scored positive. A falling chart, red, and an angry face scored negative. Those labels were read off the data. Each firm-level document can then receive a score. The regression puts cumulative abnormal returns on the left and, on the right, the emoji score plus controls the literature already uses: past returns, size, market-to-book, time fixed effects, news sentiment, and filing sentiment, with standard errors that allow heteroskedasticity and the autocorrelation in the panel.
How to use it
The main result is that the emoji term stays statistically significant, and the talk calls the size economically meaningful. A one-standard-deviation move in the emoji score lines up with an extra annualized return above 5 percent. The study does not find that this gap reverses over the following weeks or months, which matters because a gap that vanishes immediately is a different claim. Splitting accounts, the effect is driven more by accounts opened in 2020 or later than by older ones.
Whether that association is a trade is a longer list. Institutional pipes can pay for the data and the latency. A personal account, the talk says, generally cannot fund the same pipe, and the alpha can fade. The questions are the clock, a day or a week, turnover, costs, and whether the book can be known by 9:20 a.m. or only proxied. Stability is the test, and adding capital should, if the mispricing is real, shrink it. The short answer in the room was that related strategies have been run at firms. The long answer was that it depends. Reproducing a public result on a small parse is the suggested start. The talk will not play the live book for you.
What usually goes wrong
Reading "over 5 percent" as a personal strategy. Treating an implied betting price, 80 or 90 percent, as "nine times out of ten," which the questions turned into a separate point: the useful object is how hard the market moves if the priced outcome fails. Tuning hyperparameters until one sample looks tradable, then discovering the next sample does not. And calling a year of empty results wasted. The talk's line is that the experience was the work.
Questions people ask
How was an emoji scored without a label?
Posts were scored with a general sentiment model. The emoji inherited the average score of the posts that contained it, separately in each period, such as 2000 and 2018.
What survived the controls?
The emoji coefficient, in a regression for cumulative abnormal returns that already included standard return predictors, news, and filings. One standard deviation mapped to more than 5 percent annualized, without a reversal in the following weeks or months in that study.
Who drove the result?
Accounts created in 2020 or later, more than accounts created before 2020. The talk treats that split as a reason to keep asking questions of the same tape, not as a finished trading rule.