Does sentiment lead prices?

News sentiment

News tone tracks the price move it describes. It does not tell us where price goes next.

Each trading day at 3:30 PM ET we rank the nine stocks by their sentiment reading and compare that ranking with each stock's return against QQQ. The same-day match is strong and steady in both halves of the period (best: Level vs own 60 days, t +6.9). The next-day and next-week match is flat: the largest lead anywhere, daily or intraday, is t -1.6, which is noise. Put simply, stories are written about moves that already happened.

Bars show the average rank correlation per day across nine stocks (−1 to +1). Labels give the t-statistic; we call a result clear at |t| ≥ 3 with both halves agreeing, weak at |t| ≥ 2.

Sorted into fifths by "level vs own 60 days"

The most negative fifth fell -47 bp against QQQ that day and the most positive rose +86 bp. The next day shows no ordering: +16 bp, +6 bp, +19 bp, -3 bp, +25 bp from most negative to most positive.

Inside the day (15-minute readings, next hour)

Nothing leads the next hour. The readings also fail to follow the past hour, because historical stories are only counted an hour after publication, so a story about the last hour's move usually lands in the next hour.

Events

Analyst upgrades keep working after the open. Earnings beats and positive news bursts fade.

Each dot is the average move against QQQ in the event's own direction, from the first 15-minute bar after the event was public; for after-hours news that means the next open, so the overnight gap is excluded. Lines are 95% intervals. A dot above zero means price kept going the way the event pointed; below zero means it reversed.

Scheduled macro days (CPI, jobs, Fed): QQQ moved 0.92% on average across 57 macro days and 0.85% on 289 other days (t +0.6). They were not noticeably bigger days.

Prediction markets

The odds form clean expected-value lines. Day to day they do not lead the market. Sharp moves in some lines are worth testing on new data.

Each Kalshi question is a ladder of contracts ("CPI above 0.3%", "above 0.4%"…). Their prices form a probability distribution, so each day gives one expected value. The line follows the next unresolved event and its daily change is scored against its own last 60 changes (z). Snapshots are taken before 3 PM ET; returns start at the 4 PM close. On days when expected inflation, unemployment or recession odds fell sharply, QQQ tended to rise the next day. Each of these groups holds 13 to 24 days and we looked at 24 groups, so one or two would clear the bar by chance alone. Treat this as a hypothesis for out-of-sample testing, not a result.

Tweets

Not testable yet.

Tweet tone is scored with a locally run language model, and recording has only just started. Unlike news, tweets cannot be reliably backfilled, so the same tests need a few months of live history. A sizeable share of stock-tagged posts are templated promotion (group invitations, "follow this trader"), so filtering them out comes before any measurement.

Takeaways

• Sentiment measured this way is a good description of what just happened and a poor forecast of what happens next. Using it as a buy or sell signal on its own would mostly mean chasing moves that are already in the price.
• It is more useful as context. Knowing that a move came with a burst of positive coverage, and that such moves tended to give a little back, is information a price chart alone does not carry.
• The event patterns (upgrade drift, earnings-beat fade) are consistent with well-known effects, but on a few dozen cases each they need fresh data before anyone should trust them.
• Most published sentiment "signals" are tested in ways that quietly use information from after the fact. Being strict about timing removed almost all of the apparent edge here, which is the result worth taking away.