Do crowded longs precede a sell-off?

The claim is familiar: when more than 70% of accounts are long, a sell-off is coming. We tested a precise version of that claim against recorded history. The ordinary 2% decline rate changed little; the larger-move thresholds need a more cautious reading because the episodes and comparison minutes differ.

What we measured

We registered the claim before measuring it. The library entry says: the global long account fraction at 0.70 or above, the sharpest countable version of “the crowd is crowded long.” The threshold was calibrated on how often the state occurs, before looking at any outcome. On BTC, 0.80 and 0.75 literally never happened in this window; 0.65 held 27% of the time, which is not what anyone means by crowded; 0.70 held 3.7% of the time, a rare state on BTC. So 0.70 it was.

Worth saying: the companion claim, “smart money is positioned against retail, so follow the top traders,” was staged for the same treatment and failed the calibration gate. Its usual threshold turned out to be true a quarter of the time in this record. That threshold did not meet the planned rarity criterion, so we did not measure its outcomes. Writing the question down first cuts both ways.

We ran the crowded-long condition over every recorded market-minute across 862 Binance USDT-M crypto and TradFi perpetuals from 2026-05-15 to 2026-07-26: 47,426,721 eligible market-minutes, 677 markets. It became freshly true, deduplicated to the hour, 8,637 times.

One honest wrinkle before the results. Universe-wide, the crowded state covers about 28% of eligible minutes, because many small perpetuals live crowded long more or less permanently. The 3.7% rarity is a BTC fact, not a universe fact. The episodes we count are the fresh crossings into the state, and most of them come from markets where the crowd leans long by habit. Keep that composition in mind when you read the comparison.

What followed

Of the 8,637 episodes, 7,351 had a complete 24 hours of record after them. Thirty minutes in, the tape had no opinion at all: 623 were down 1% or more, 630 were up 1% or more. Over the full day:

• 2,302 of 7,351 (31.3%) fell 2% or more. Across every eligible minute where the condition was false, that rate is 30.6%.
• 1,947 (26.5%) rose 2% or more, against a background rate of 25.2%.
• 1,065 (14.5%) fell 5% or more, against a background rate of 11.8%. That is the one rung where the folklore shows up: roughly a fifth more often than ordinary. The matching up-rung barely moved: 11.1% against 10.7%.
• And the ending the story promises most loudly is the one that got rarer: 0.5% of episodes closed 20% or more down, against a background rate of 0.8%. Same at 15%: 1.3% against 1.5%.

Read that back against the screenshot thesis. At the moment the crowd tipped past 70% long, the next day was not a flush. It was, to within about a point at the ±2% rungs, an ordinary day. What actually changed was the shape of the bad outcomes: mid-size flushes, the 5% to 10% kind, were more frequent in the recorded sample, while the catastrophic kind came less often than the market’s own background. Crowding, in this record, was not a sell signal. It was a mild marker for tail shape.

Before you trade this, read this

This comparison does not establish a directional edge. The observed 31.3% decline rate was close to the 30.6% comparison rate. These frequencies alone cannot explain a trading strategy’s profitability.

The two sides of the comparison are different kinds of thing. The 14.5% counts deduplicated episodes; the 11.8% counts every eligible minute where the condition was false, and neighbouring minutes share almost all of their forward path. Our engine ships that caveat with the result and it applies here. Treat the gap as descriptive of this sample, not as an isolated effect of crowding.

Composition is doing work. Crowded-long episodes concentrate in markets whose crowds lean long by habit, and those markets move differently from BTC. Part of any gap between episode rates and the all-market baseline is just which markets get crowded, not what crowding does. The pinned report carries per-market counts, so this is checkable rather than arguable.

Coverage grew mid-window. Positioning data covers a small set of markets before 2026-06-05 and a much larger one after, so most episodes date from the last seven weeks of the window. The engine only counts minutes where the reading exists, and absence is never treated as zero, but a ten-week label on what is mostly a seven-week measurement deserves this sentence.

Ten weeks, one regime, one of 27. This claim was swept alongside 26 others this week, so some spread between them is expected by chance alone. Our own sweep labels this one a null at the ±2% rung. We are publishing the count, not a verdict.

These are counts of what happened in the scanned markets and dates, not a promise it happens again. Nothing here is trading advice.

Check us

The condition, the window, the 8,637 episodes and every distribution above are pinned in the linked report under q#5e498394. The background rates come from the cohort comparison of the same document, same key, same dataset revision, run against every eligible minute where the condition was false. Both re-run to the same bytes on any account, on any day.

The claim went into our library before its first measurement, with the threshold fixed on prevalence alone. Its twin failed that gate and was sent back for a tighter threshold rather than measured anyway. That ordering, question first, count second, is the entire reason this blog exists: the argument is here, the receipt is one click away.

A useful next test would compare similarly volatile markets and separately report their results. Define those groups before looking for a stronger effect. See base rates and selection bias for the comparison principles.