Does the long/short ratio predict price? We tested 2.98 million observations

The most repeated piece of positioning analysis in crypto goes like this: most traders are long, therefore the market is about to fall. It is stated as though it were arithmetic. Almost nobody checks it, largely because the exchange API only serves 30 days of history and the number is awkward to test properly. We have been recording long/short data continuously across hundreds of perpetuals, so we tested it. The rule does not merely fail in our sample. It points the wrong way.

What we measured

We took every long/short account-ratio observation in our archive between 19 May and 13 August 2026: 2,984,190 readings across 513 perpetual contracts. Each observation was joined to the five-minute close at the same timestamp, and we measured what the coin did 1, 4 and 24 hours later.

No symbol filtering. No removing the losers. No choosing a favourable start date. The method is deliberately boring, because the point of the exercise is that the result should survive someone else repeating it.

One number matters before any other: over this window the average coin returned −0.507% in the 24 hours following a randomly chosen moment. The market drifted down. Every figure below has to be read against that baseline, not against zero. This is the single most common error in studies of this kind, and it is how a losing rule gets published as a winner.

Result 1: the raw ratio explains almost nothing

Grouped by the ratio as a trader reads it on a screenshot, forward 24-hour returns look like this.

Average return over the following 24 hours, by long/short ratio at the moment of observation.

The spread is real but small, and it is not monotonic: the worst forward returns sit in the middle of the range, not at the extremes. If crowding drove reversals, the two ends of this table would be the interesting rows. They are the boring ones. Coins with a ratio above 3.00, where the crowd is heavily long and the contrarian rule screams sell, did better than the average coin.

There is a structural reason the raw number is weak. Crypto retail carries a long bias almost permanently, and each coin has its own resting level. A ratio of 1.8 is unremarkable on one contract and an extreme on another. Comparing every coin against the same fixed thresholds mostly measures which coins are popular, not who is crowded.

Result 2: measured per coin, the rule inverts

So we ranked each coin against its own history and looked only at its own extremes: the top 5% of readings for that specific contract, and the bottom 5%. To stop overlapping five-minute windows from inflating the sample, every figure here is collapsed to one observation per coin per day.

Average return over the following 24 hours. Coin-days, not raw observations.

Read those two rows against the baseline. When the crowd is at its most long on a given coin, that coin went on to outperform the average by about six tenths of a percentage point. When the crowd is at its most short, the coin went on to underperform by nearly two percentage points.

The folklore says a heavily shorted coin is a squeeze waiting to happen. In 1,951 coin-days of maximum short crowding, the average outcome was the worst in the entire dataset.

Does it hold up, or is one month carrying it?

This is the question that kills most published findings, so we split the same collapsed data by month.

* May covers 19–31 only, when our recording began. Difference column: long-extreme minus baseline.

Three of four months agree, and the one that disagrees is the 13-day partial month at the start of the archive. That is better than most retail rules manage, and it is still not enough. Four months is four observations of market regime, no matter how many data points sit inside them. The whole window also leans bearish, and an effect measured entirely inside one regime has not been tested against the other one.

Why we do not trade this, and neither should you

The long-side effect is smaller than its costs. Six tenths of a percentage point over 24 hours, before fees, spread and funding. A round trip in perpetuals eats most of that. An edge that only exists in a spreadsheet is not an edge.

The causal arrow probably runs the other way. Traders do not short a coin and then it falls. A coin falls, and traders pile into shorts on the way down. If that is what is happening, the bottom-5% row is not positioning predicting price, it is positioning describing price that has already moved, and the −2.37% is momentum wearing a costume. Our data cannot separate those two stories, and we are not going to pretend otherwise.

An average is not a trade. These are averages of forward returns with no stop, no target and no position sizing. Turning an average into a strategy requires answering when you exit and what you risk, and that is where most apparent edges die.

What positioning data is actually good for

Discarding the entry rule does not mean discarding the data. Three uses survive this test.

As a measure of fuel, not direction. A one-sided book is stored energy. It does not tell you when the move comes, it tells you how violent it will be when it does. That belongs next to open interest and the liquidation stream, not on its own.

As a sizing input. If you are on the crowded side of a lopsided book, you are standing in the path of any forced unwind. That is not a reason to skip the trade. It is a reason to take it smaller.

As a change, not a level. The level is contaminated by the structural long bias. The direction of the last few hours says the crowd is arriving or leaving right now, which is a different and more useful statement.

For the mechanics of the three different ratios and why they disagree with each other, we wrote that up separately in what the long/short ratio really measures.

Check it yourself

The exchange API serves 30 days of positioning history and no paid tier extends it, which is why studies like this are rare. Our archive exists because we started recording early and never stopped. The per-coin history is published free in the Data Terminal, so any claim above can be checked against the same source, including by someone hoping to prove us wrong.

And if a rule built on positioning looks compelling to you, run it against a random-entry control before risking money on it. Ours is free and takes ten seconds: Reality Check. When we ran twenty-four widely marketed strategies through it, none of them beat random entry after costs. This one would not have either.