Research · test #11
Does buying a leverage flush work? Conditional
"Open interest just got wiped out - the leverage is flushed, nobody is left to sell, buy the dip." We wrote the rules down first and checked every flush on 522 Binance coins since 2024. It worked until mid-2025, then stopped - and the open interest part never added anything to a plain price drop.
In short: 5,893 flushes (open interest -10% and price -5% within four hours). Buying and holding 24 hours, after costs and funding: +1.32% on average in Jan 2024 - Jun 2025, -0.47% from Jul 2025. A 5% drop without the open interest condition did as well (p = 0.34). The median trade lost money: -0.98%.
The rules (fixed before the run)
- Data: 522 USDT-margined perpetuals on Binance; hourly closes, funding payments and open interest in coins (so a falling price does not shrink it by itself). 8,483,468 coin-hours from January 2024 to October 2026.
- Signal: over the last four hours the price fell 5% or more and open interest fell 10% or more. One signal per coin per 24 hours.
- Trade: buy at the close of that hour, sell 24 hours later. 0.15% costs and the funding paid during the hold are subtracted.
- Controls: 200,000 random hours on the same coins (luck), and every 5% drop without the open interest condition - to see whether the flush adds anything to a plain drop.
- Pass mark: positive average with t > 2 and beating random hours (p < 0.05) in both the first period (Jan 2024 - Jun 2025) and the second (Jul 2025 - Oct 2026), and beating plain drops. Passing only the first period counts as "conditional".
Result: worked, then stopped
| Period | Flushes | Coins | Average | Median | Winners | t | vs random hours, p |
|---|---|---|---|---|---|---|---|
| Jan 2024 - Jun 2025 | 2,452 | 346 | +1.32% | -0.05% | 50% | 5.89 | 0.0000 |
| Jul 2025 - Oct 2026 | 3,441 | 504 | -0.47% | -1.77% | 41% | -1.60 | 1.00 |
| Whole period | 5,893 | 515 | +0.27% | -0.98% | 45% | 1.40 | 0.0003 |
The first period passes every check. The second fails: the average turns negative and random hours do better 99.9% of the time. By the rule fixed in advance that is conditional - an effect that did not survive new data.
The open interest flush adds nothing
The popular story is about leverage: once open interest is wiped out, forced sellers are gone. If that were the reason, the flush should beat an ordinary sharp drop. It does not. Every 5% drop in four hours, with or without an open interest collapse, averaged +0.22% over 64,024 drops; the flush signal averaged +0.27%. A difference this size appears by chance 34% of the time. A deeper flush did not help either: with open interest down 20% or more the average was negative at every price threshold (table below).
By year
2024 and 2025 were positive on average, 2026 clearly negative (t -2.9). More than half of the second period's flushes (1,883 of 3,441) fell in 2026.
Why the average flatters the trade
- Most trades lose. Only 45% of flushes closed higher a day later; the median was -0.98%. A few large rebounds carry the average.
- Flushes come in crowds. 5,893 flushes fell on 905 different days - on a market-wide crash dozens of coins flush in the same hour. Averaging each day first and then across days gives -0.08% (t -0.27): the positive overall average rests on a handful of crash days.
- Holding three days averaged +1.90%, but the median was again negative (-1.33%).
- The tail is brutal. The worst trades lost 91%, 86% and 74% in a day - small coins that kept falling. With leverage, any of them ends the account.
- Mostly altcoins. Bitcoin and Ethereum flushed this hard only 3 and 3 times in almost three years.
Other thresholds (24 hours, whole period)
| Price drop in 4h | Open interest drop | Signals | Average | Any drop of that size |
|---|---|---|---|---|
| -3% | -5% | 24,967 | +0.08% | -0.02% |
| -3% | -10% | 7,284 | +0.18% | -0.02% |
| -3% | -20% | 1,068 | -0.65% | -0.02% |
| -5% | -5% | 16,352 | +0.17% | +0.22% |
| -5% | -10% | 5,893 | +0.27% | +0.22% |
| -5% | -20% | 958 | -0.68% | +0.22% |
| -10% | -5% | 5,567 | +1.33% | +1.08% |
| -10% | -10% | 2,964 | +1.02% | +1.08% |
| -10% | -20% | 680 | -0.55% | +1.08% |
Descriptive, not part of the verdict. Only the deepest price drops (10% in four hours) with a moderate open interest fall show a clearly positive average - and a 10% drop without the open interest condition does just as well.
What this test does not say
- Open interest falls when positions close for any reason - liquidations, stop losses or manual exits. We see the total, not the cause.
- Coins delisted before October 2026 are not in the data, which flatters buying dips.
- It does not test other ways to use a flush (waiting for a reclaim, scaling in, a stop loss). It tests the plain version: buy the flush, hold a day.
Data and method
Rules and code were committed before the run. Prices, funding and open interest: Binance public data. Open interest per coin and across exchanges right now: open interest and by exchange; what was actually liquidated: liquidations; how far a leveraged dip-buyer can fall: liquidation calculator. Method: methodology. All tests: research.
FAQ
Should you buy after a leverage flush?
Not on this evidence. Across 5,893 flushes, buying and holding a day averaged +1.32% in January 2024 - June 2025 but -0.47% from July 2025, after costs. The median trade lost in both periods.
Does the open interest drop add anything?
No. Buying every 5% drop in four hours without looking at open interest averaged +0.22%; the flush signal averaged +0.27%. The difference is well within chance (p = 0.34).
Why did it work in 2024 and not in 2026?
This test cannot say why. The day-after average was +0.95% in 2024, +0.93% in 2025 and -1.13% in 2026. A pattern that fades on new data is exactly what the out-of-sample check is for.
Is a drop in open interest the same as liquidations?
Not exactly. Open interest falls whenever positions close - liquidations, stop losses or traders closing by hand. A sharp drop together with a falling price usually means many longs were forced out, but the data cannot separate the causes.