Long/short ratio and open interest: do crowds predict?

This is the sequel to funding as a signal. Funding is what a crowded side pays; Binance also publishes who the crowd is — the share of accounts that are long against those that are short — and how much leverage is open. The idea traders repeat: when everyone is long, the market goes down to punish them, and when open interest balloons, a flush is near. Here is what six years of Binance perpetuals say about it.

Setup

• Data: Binance USDT-M perpetuals of ten large coins — BTC, ETH, BNB, XRP, ADA, DOGE, SOL, LINK, LTC and BCH — on 4h bars. Binance's positioning metrics reach back to 2020-09 for BTC but only to 2021-12 for the other nine, so each coin is studied from the first bar its metrics exist to 2026-10-01.
• The crowd: Binance's all-accounts long/short ratio (LSRATIO, kind ACCOUNT) — how many accounts hold a net long position for each account holding a net short one. It is the retail-positioning read, not the top-trader one. It was above 1 on 79.4% of BTC's bars and on 85.7–100% of each altcoin's, with medians from 1.48 (BTC) to 2.98 (DOGE): most accounts are long by default, so "crowded" has to be measured against each coin's own range.
• Three events, each in the top or bottom 10% of the coin's own readings over the half of the period it falls in: the ratio in its top decile (crowded long), in its bottom decile (crowded short), and open interest — contracts outstanding, not their dollar value — growing over 4 days (OICHG, 24 bars) in its top decile (leverage piling in).
• Statistics: Research Lab event scans: the mean return 1, 3 and 7 days (6, 18 and 42 bars) after the event against the coin's usual return over the same horizon, with a multiple-testing correction (q-value) and a rotation test that respects overlapping windows. Each coin's history is split into two halves — up to 2023-09-01 and after — scanned separately, so a finding has to show up twice. 10 coins × 2 halves × 3 events = 60 scans, 180 tests.
• Trading: two fixed fade rules on BTC, not tuned — short when the crowd is most long, long when it is most short, exit after 42 bars (7 days) — at 1x, isolated margin, 95% of the account, 0.05% fee and 0.05% slippage per side, funding charged and paid, 2020-09-01 → 2026-10-01.

The scans: a consistent tilt, not a signal

The 3-day return after each event, minus the coin's usual 3-day return. A tick marks a result that passed its own scan's correction (q below 0.10):

Crowded long is the one column with a clear direction: the next 3 days were weaker than usual on 18 of 20 coin-halves, by a median of -1.55%, and 18 of 20 were weaker at 1 and at 7 days too. On BTC's first half the effect was large — -1.84% over 3 days after a crowded-long reading against a usual +0.41%, q 0.017 — and the second half kept the sign at a third of the size without passing the correction (-0.79%, q 0.149). But the 20 cells are not 20 independent votes: the coins move together, so their crowded readings cluster in the same weeks. Per scan, 9 of 60 crowded-long tests passed, and none of them on both halves of the same coin.

Crowded short leaned the other way — better than usual on 14 of 20 at 3 days, median +0.89% — with 13 of 60 tests passing, 9 of them in 2023–2026 and almost all on DOGE, SOL and LINK. It is not stable: LINK's first half passed in the opposite direction at 7 days (-5.51%) before its second half passed at +7.64%. SOL is the only coin significant in both halves, and the scan flags its first half's own two halves as disagreeing.

Open-interest surges told nothing: better than usual on 9 of 20, worse on 11, and 1 of 60 tests passed. Read as numbers rather than events, neither the ratio nor the change in open interest predicted the next 1, 3 or 7 days in any of 120 tests.

Now the count that matters. 23 of the 180 tests passed the correction inside their own scan, where it covers only that scan's three horizons. Correct over all 180 together — the questions were asked of the same market, and that is how many were asked — and none passes: the strongest, LINK's crowded short in 2023–2026, has a p-value of 0.002 and a corrected q of 0.26.

Trading the fade

A trader cannot know in 2021 where the top decile of 2020–2023 will be, so the trading rules compare the ratio only with its own last 90 days (a trailing percentile rank over 540 bars). On BTC:

Backtest

Long/short ratio fade: short when the crowd is most long (90-day top 10%), hold 7 days

BTCUSDT · 4h · Futures · 2020-09-01 → 2026-10-01

Return

-66.62%

Max drawdown

72.80%

Sharpe

-0.35

Win rate

47.3%

Trades

110

Open the report →

Backtest

Long/short ratio fade: long when the crowd is most short (90-day bottom 10%), hold 7 days

BTCUSDT · 4h · Futures · 2020-09-01 → 2026-10-01

Return

+103.56%

Max drawdown

53.58%

Sharpe

0.53

Win rate

53.0%

Trades

100

Open the report →

Shorting the crowded long lost -66.62% over 110 trades, even with 1,705.39 USDT of funding received along the way. It lost in both halves. To see whether that is bad luck or no information, 2,000 sets of 110 random 7-day BTC shorts were simulated over the same window, with the same costs and funding: their median was -69.06%, and the rule landed at the 55th percentile — indistinguishable from shorting at random.

The scans and the backtest disagree because they ask different questions. The scan's "top 10%" is drawn over the whole half with hindsight. A dry run of the short with that hindsight threshold made +81.35% on BTC in 2020-09 → 2023-09 — and -7.32% in 2023-09 → 2026-10 with the second half's own hindsight threshold. The trailing version, the only one you could have traded, lost in both.

Buying the crowded short made +103.56% over 100 trades, after paying 2,284.53 USDT of funding. Against 2,000 sets of 100 random 7-day longs (median +5.81%) it sits at the 85th percentile — better than most, not unusual — and holding BTC over the same window made +615.67%.

Conclusion

• A crowded long book on Binance was followed by slightly weaker days on almost every large coin, 2020–2026, but no single result survived a correction over the 180 tests asked, and none repeated on both halves of a coin.
• The fade you could actually trade — short BTC when the ratio hits its 90-day top 10% — lost -66.62%, no better than random shorts; buying the crowded short made +103.56% against +615.67% for holding.
• Open-interest surges predicted nothing, as events or as numbers.

The method is the one from funding as a signal; the metrics are explained in the backtest metrics guide.