Does crypto bounce after a liquidation cascade? 1,955 flushes on 5-minute bars
Published by vyx
A little, on average, and only the biggest flushes bounce by enough to matter. After a liquidation flush, a 5-minute drop of 3 standard deviations or more on one of the biggest forced-selling bursts of the last 10 hours, price was higher 30 minutes later 57.5% and 60.0% of the time on two separate windows, against 49%–50% from a random moment, by +11.4 bp and +16.6 bp on average: barely more than the ~10 bp a market-order round trip costs. The biggest flushes bounced +49.1 bp and +36.4 bp; drops as big without the record burst, only +15.4 bp and +5.7 bp.
Open the Liquidation Flush desk: the crypto markets flushed in the last 5 minutes on a record forced-selling burst, the big ones first, live on 5-minute bars.
The question
When a market drops hard and longs get liquidated, traders split two ways. One camp buys the flush: forced sellers are done, the drop overshot, the bounce is coming. The other waits: a cascade feeds itself, and the first liquidations set off the next. Most answers to which is right are a chart someone remembers. We wanted a number, measured on every flush across the most-traded crypto markets, beside what a random moment gives.
Reading liquidations from the tape
Hyperliquid publishes no market-wide liquidation feed, only each account's own. So a flush has to be read from the trades. vyx scores every bar's forced flow: the biggest burst of same-side trading inside any 2 seconds of the bar, against the market's own normal rate. A liquidation engine selling into the book leaves exactly that print. This is the rule we tested, on the 40 most-traded crypto perpetuals on 5-minute bars:
- The bar fell 3 standard deviations or more of the market's own normal 5-minute move, measured over its last 100 bars.
- Its forced-flow burst ranks in the top 5% of the market's last 120 bars (10 hours), and the bar's net aggressive flow was selling.
- One flush per market per hour. The bounce is measured from the flush bar's close: long, as someone buying the flush would be.
- Compared fairly: against a random moment in the same market at the same time of day on another day, and against sharp drops that had no record burst.
- Open interest is the check on the proxy: it fell 0.5% or more on the bar itself on 37.0% of flushes, against 19.7% of all sharp drops. The burst finds bars where positions were closed about twice as often as price alone does, but on most flushes open interest gave no such sign.
Share of events after which price was higher 30 minutes on, beside chance: a random moment in the same market at the same time of day on another day, 400 times over. Sep 1 – Oct 22, 2026 is the window the rules were looked at on; Jul 28 – Aug 31, 2026 came before it and no test had seen it.
The numbers
| Price higher 30 minutes later | Chance (mean) | Chance (95% range) | |
|---|---|---|---|
| Sharp drop (3σ) · Sep 1 – Oct 22 | 56% | 49% | 47%–50% |
| Sharp drop (3σ) · Jul 28 – Aug 31 | 57% | 50% | 48%–53% |
| Liquidation flush · Sep 1 – Oct 22 | 58% | 49% | 46%–51% |
| Liquidation flush · Jul 28 – Aug 31 | 60% | 50% | 46%–54% |
| Flush, open interest fell · Sep 1 – Oct 22 | 60% | 49% | 45%–53% |
| Flush, open interest fell · Jul 28 – Aug 31 | 64% | 50% | 44%–57% |
| 5σ drop, no record burst · Sep 1 – Oct 22 | 52% | 49% | 44%–54% |
| 5σ drop, no record burst · Jul 28 – Aug 31 | 61% | 52% | 45%–59% |
| Big liquidation flush · Sep 1 – Oct 22 | 67% | 49% | 39%–58% |
| Big liquidation flush · Jul 28 – Aug 31 | 63% | 50% | 39%–61% |
Flushes did bounce more often than chance: price was higher 30 minutes later after 57.5% of 1,258 flushes on the first window and 60.0% of 697 on the unseen one, against 48.7% and 49.8% from a random moment. But so did any sharp 5-minute drop: 55.8% and 56.9%. Part of the bounce is simply what crypto does after a sharp 5-minute move down, burst or no burst. It is also a lean, not a rule. Measured as a race, half the drop back against half the drop again, flushes won 52.4% and 55.0% of the time, where chance gives 46.9%–52.1% and 46.9%–54.1%.
Basis points (0.01%) from the event bar's close, long, averaged over every event. A market-order round trip costs about 10 bp. The record burst is the biggest of the market's last 10 hours, the one a big flush requires.
The numbers
| Sep 1 – Oct 22, 2026 (rules looked at) | Jul 28 – Aug 31, 2026 (unseen) | |
|---|---|---|
| A random moment (chance) | −0.7 bp | +2.1 bp |
| Sharp drop, no record burst | +9.1 bp | +6.0 bp |
| Liquidation flush | +11 bp | +17 bp |
| Flush, open interest fell | +17 bp | +27 bp |
| 5σ drop, no record burst | +15 bp | +5.7 bp |
| Big liquidation flush | +49 bp | +36 bp |
The average flush bounced +11.4 bp (±3.6) and +16.6 bp (±4.1) over 30 minutes, a few basis points more than sharp drops without the burst (+9.1 bp and +6.0 bp), and barely more than the ~10 bp a market order costs to get in and out. Where the burst matters is at the extreme. Bars that fell 5 standard deviations or more on the biggest burst of the last 10 hours bounced +49.1 bp (±12.1) and +36.4 bp (±24.7), higher 30 minutes later 66.7% and 62.9% of the time. Drops just as big without that record burst bounced +15.4 bp and +5.7 bp. One reading: a big drop with no forced selling behind it is more often the market repricing, while one on a record burst is more often forced sellers overshooting. Flushes where open interest fell with the price, positions closed rather than sold, sat in between: +16.9 bp and +26.5 bp.
Basis points from the flush bar's close, long, averaged over every flush. A random moment drifted +3.8 bp and +16.9 bp over 4 hours on the two windows: the unseen window was a rising market.
The numbers
| Sep 1 – Oct 22, 2026 (rules looked at) | Jul 28 – Aug 31, 2026 (unseen) | |
|---|---|---|
| Held 15 minutes | +6.7 bp | +11 bp |
| Held 30 minutes | +11 bp | +17 bp |
| Held 1 hour | +13 bp | +18 bp |
| Held 2 hours | +8.2 bp | +31 bp |
| Held 4 hours | +16 bp | +32 bp |
Against chance, the bounce is a first-half-hour thing. Net of what a random moment drifted, a flush was ahead +12.1 bp and +14.5 bp after 30 minutes and +11.8 bp and +14.7 bp after 4 hours: holding longer added drift and a lot of noise, not more bounce; after 4 hours the error bars (±7.9 and ±8.9) are as wide as the bounce. Flushes come in bursts too: a market-wide flush can light dozens of coins inside 15 minutes. Counting each such episode once, the 30-minute bounce was +20.9 bp and +16.3 bp, so the average is not one bad afternoon.
Every idea we tested, both windows
| Idea | Events | Reversed in 30 min | Chance | Average over 30 min | Chance | Verdict |
|---|---|---|---|---|---|---|
| Sharp drop: 3σ or more in 5 minutes (the baseline) | 3,354 / 2,012 | 55.8% / 56.9% | 48.8% / 50.4% | +8.1 bp / +8.8 bp | −0.6 bp / +3.1 bp | Above chance on both windows |
| Liquidation flush: sharp drop, top-5% burst, sellers aggressing | 1,258 / 697 | 57.5% / 60.0% | 48.7% / 49.8% | +11.4 bp / +16.6 bp | −0.7 bp / +2.1 bp | Above chance on both windows |
| Big flush: 5σ or more on the biggest burst of 10 hours | 138 / 62 | 66.7% / 62.9% | 48.5% / 50.0% | +49.1 bp / +36.4 bp | −1.7 bp / +0.8 bp | Above chance on both windows |
| Control: 5σ drop without the record burst | 402 / 241 | 52.2% / 61.0% | 48.6% / 51.7% | +15.4 bp / +5.7 bp | −1.3 bp / +5.3 bp | Mixed: 2 of 4 reads above chance |
| Flush with open interest down 0.5% or more | 497 / 262 | 60.0% / 63.7% | 49.1% / 50.4% | +16.9 bp / +26.5 bp | −0.1 bp / +3.1 bp | Above chance on both windows |
| Flush with longs paying above the base funding rate | 117 / 59 | 61.5% / 50.8% | 48.7% / 50.4% | +19.3 bp / +9.9 bp | +0.3 bp / +4.7 bp | Mixed: 1 of 4 reads above chance |
| Flush while Bitcoin also fell 2σ | 638 / 252 | 56.1% / 60.3% | 48.6% / 50.8% | +11.2 bp / +19.9 bp | −0.7 bp / +4.0 bp | Above chance on both windows |
| Flush while Bitcoin did not | 664 / 461 | 58.7% / 60.1% | 49.3% / 49.5% | +13.3 bp / +15.3 bp | 0.0 bp / +0.7 bp | Above chance on both windows |
| Sharp rise (3σ), faded | 3,572 / 2,493 | 52.5% / 53.5% | 50.2% / 49.3% | −5.5 bp / −2.8 bp | −0.2 bp / −2.2 bp | Mixed: 2 of 4 reads above chance |
| Squeeze (the mirror of a flush), faded | 909 / 629 | 52.8% / 51.7% | 49.8% / 49.3% | −4.9 bp / −4.5 bp | −0.7 bp / −2.1 bp | No better than chance |
Each cell reads Sep 1 – Oct 22 / Jul 28 – Aug 31. Reversed means price was higher 30 minutes later after a drop, lower after a rise; the average is in the reversal's direction, in basis points. Chance is a random moment in the same market at the same time of day. Above chance means both numbers sat above chance's 95% band.
What did not help
Three ideas that sound right did not hold up. Crowded longs, a flush while longs were paying above Hyperliquid's base funding rate, bounced +19.3 bp and +9.9 bp over 30 minutes on only 117 and 59 events, inside the wide band chance gives so few events, and the race went the wrong way on both windows. Whether the whole market flushed with it made no reliable difference: flushes while Bitcoin also fell 2 standard deviations bounced +11.2 bp and +19.9 bp, flushes on their own +13.3 bp and +15.3 bp, and which was bigger flipped between the windows. And the mirror does not work: fading a squeeze, a sharp rise on a record buying burst, made −4.9 bp and −4.5 bp over 30 minutes. Squeezes did not reverse; flushes did, a little.
GRAM around Oct 22, 16:50 UTC, when it fell 0.48% in 5 minutes, 3.3 standard deviations of its normal move, on the biggest burst of its last 10 hours. Shaded: the bar the desk lit. Picked as the latest flush, not the best one.
The numbers
| Hour | GRAM, % from the flush bar's close | The desk lit it |
|---|---|---|
| -30 | +1.235% | |
| -25 | +1.258% | |
| -20 | +1.037% | |
| -15 | +1.159% | |
| -10 | +1.03% | |
| -5 | +0.48% | |
| 0 | 0% | yes |
| +5 | +0.267% | |
| +10 | +0.16% | |
| +15 | +0.389% | |
| +20 | +0.48% | |
| +25 | +0.9% | |
| +30 | +0.831% | |
| +35 | +0.846% | |
| +40 | +0.991% | |
| +45 | +0.892% | |
| +50 | +0.808% | |
| +55 | +0.747% | |
| +60 | +0.831% | |
| +65 | +0.694% | |
| +70 | +0.465% | |
| +75 | +0.61% | |
| +80 | +0.328% | |
| +85 | +0.519% | |
| +90 | +0.48% | |
| +95 | +0.496% | |
| +100 | +0.595% | |
| +105 | +0.587% | |
| +110 | +0.648% | |
| +115 | +0.618% | |
| +120 | +0.564% | |
| +125 | +0.519% | |
| +130 | +0.496% | |
| +135 | +0.389% | |
| +140 | +0.404% | |
| +145 | +0.572% | |
| +150 | +0.618% | |
| +155 | +0.61% | |
| +160 | +0.496% | |
| +165 | +0.45% | |
| +170 | +0.412% | |
| +175 | +0.419% | |
| +180 | +0.724% | |
| +185 | +0.564% | |
| +190 | +0.343% | |
| +195 | +0.343% | |
| +200 | +0.374% | |
| +205 | +0.351% | |
| +210 | +0.412% | |
| +215 | +0.259% | |
| +220 | +0.221% | |
| +225 | +0.206% | |
| +230 | +0.275% | |
| +235 | +0.084% | |
| +240 | +0.069% |
This one bounced. Thirty minutes after the flush GRAM was up 0.83%, an hour after it up 0.83% and 4 hours after it up 0.07%. That is the shape of the average too: most of the bounce early, and any single flush can go either way.
r = close / prior(close, 1) - 1; r / stdev(r, 100) <= -3 pctRank(cascadeScore, 120) >= 0.95 change(cvdClose, 1) < 0
Line one: the bar's move in the market's own standard deviations of 5-minute moves, over its last 100 bars, 3 or more down. Line two: the bar's forced-flow burst ranks in the top 5% of its last 120 bars. Line three: net aggressive flow on the bar was selling. The desk adds two more screens: the big flush (5 standard deviations, the biggest burst of the 120 bars) and the flush with open interest down 0.5% or more.
These are events measured close to close, not trades. In a cascade the book is thin and fills are worse than the ~10 bp round trip quoted; a limit order costs about 3 bp instead but may not fill on the bar that bounces. The flush is read from the tape, not from liquidation records. Nothing here says to buy anything, and the desk is a watchlist of where a flush just happened, not a signal that it will bounce.
Why buying the flush feels like it works
The bounce is real, and the big ones are memorable: a coin that drops 2% in a minute and gives back half of it is the chart everyone screenshots. What the average says is quieter. A plain flush leans toward a bounce by about as much as it costs to trade, and the clear edge sits in the rarest, most violent flushes, the ones on the biggest forced-selling burst of the last 10 hours. The Liquidation Flush desk lights those first, as they happen, so you can see which kind you are looking at before you decide.
How this was done
The same method as every study here — how vyx tests a signal.
- Venue and data
- Hyperliquid perpetuals, 5-minute candles from vyx’s own stored history, with each bar’s trade-by-trade forced-flow score.
- Markets
- The 40 most-traded Hyperliquid crypto perpetuals when the study ran: AAVE, ADA, AERO, ARB, AVAX, BNB, BTC, CASHCAT, CRV, DOGE, DOT, ENA, ETH, FARTCOIN, GRAM, HYPE, INIT, INJ, kPEPE, LINK, LIT, LTC, MON, NEAR, ONDO, PONS, PUMP, SOL, STRK, SUI, TAO, TIA, UNI, VVV, WLD, XMR, XPL, XRP, ZEC, ZRO.
- Windows
- Sep 1 – Oct 22, 2026: the window the rules were looked at on. Jul 28 – Aug 31, 2026: the window before it, which no test had looked at. Both fixed, so a re-run grades the same candles.
- A flush
- A 5-minute bar down 3 standard deviations or more of the market’s last 100 bars, whose biggest 2-second same-side burst ranks in the top 5% of its last 120 bars, with net aggressive selling; at most one per market per hour. 1,258 on the first window in 575 episodes, 697 on the second in 396.
- What counts
- The move from the flush bar’s close, long, at 15 minutes to 4 hours; the share of flushes with a higher price 30 minutes later; and a race from the close: half the bar’s drop back before half the drop again, within 4 hours. Error bars are one standard error across events.
- Chance
- For every event, a random moment in the same market at the same time of day on another day in the window, with the same holding time and, for the race, the same distance in that moment’s own standard deviations; 400 draws, the band is where 95% land.
- Costs
- A market-order round trip of about 10 bp (Hyperliquid’s base taker fee of 4.5 bp a side plus a little slippage); about 3 bp with limit orders on both sides.
- Ideas tried
- 9: eight written down before any outcome was computed (the sharp drop, the flush, the big flush, the flush with crowded longs, with open interest falling, with and without Bitcoin, and the two mirrors for rises), plus one control added after the first results, the big drop without the record burst.
- Reproduce
- The study re-runs on request from the same fixed windows: every number on this page is regenerated by the same run, and the desk linked above runs the very screens that were measured.
What it does not show
- The flush is a proxy read from the trade tape, not liquidation records: Hyperliquid publishes none market-wide. Only 37.0% of flushes on the first window saw open interest fall 0.5% or more on the bar, so most flushes carry no sign in open interest that positions were closed: some may be ordinary aggressive selling rather than liquidations.
- Survivorship: the markets are today’s 40 most traded. A coin that faded or was delisted during the windows is not in the data.
- 87 days on one venue. The big-flush result rests on 138 and 62 events; on the unseen window its error bar (±24.7 bp) is wide, and counted once per episode its bounce there was +27.5 bp.
- Close-to-close events, not fills. Slippage in a cascade is worse than the quoted round trip, and the bar's close is the earliest anyone could act on it.
- 9 ideas were tried. With that many looks one can pass by luck; the ones reported as above chance held on both windows.
Questions
Does crypto bounce after a liquidation cascade?
A little, on average. On Hyperliquid's most-traded crypto perps, price was higher 30 minutes after a liquidation flush 57.5% and 60.0% of the time on two windows, against 49%–50% from a random moment, by +11.4 bp and +16.6 bp: about what a market-order round trip costs.
What is a liquidation cascade?
A chain of forced selling: a drop pushes leveraged longs past their liquidation price, the exchange closes them by selling into the book, and that selling pushes price into the next set of liquidations. On the tape it shows as a burst of same-side selling far bigger than the market’s normal flow, often with open interest falling.
Is buying the liquidation flush profitable?
Not reliably for an ordinary flush, in this study: the average 30-minute bounce of +11.4 bp and +16.6 bp is close to a 10 bp round trip, before the worse fills a cascade brings. The biggest flushes bounced +49.1 bp and +36.4 bp, but they are rare. This is a study of events, not advice.
How can you see liquidations on Hyperliquid?
Hyperliquid publishes each account’s own liquidations, not a market-wide feed. vyx reads them from the trades instead: the biggest 2-second burst of same-side trading in each bar against the market’s normal rate, alongside open interest falling on the same bar.
Does a falling open interest make a liquidation bounce more likely?
A little, here. Flushes where open interest fell 0.5% or more on the bar bounced +16.9 bp and +26.5 bp over 30 minutes, against +11.4 bp and +16.6 bp for flushes as a whole.
Do short squeezes reverse?
Not in this study. Fading a squeeze, a sharp 5-minute rise on a record buying burst, made −4.9 bp and −4.5 bp over 30 minutes on the two windows: no better than chance. Flushes leaned toward a bounce; squeezes did not lean toward a reversal.
Watch it on today’s board
Open the Liquidation Flush desk: the crypto markets flushed in the last 5 minutes on a record forced-selling burst, the big ones first, live on 5-minute bars.
Open the deskFree, no install. Or grade your own signal the same way — how vyx tests a signal.
Get new research by email
One email when a new study lands or a study is re-run on fresh data. Nothing else, and one click to leave.
Your address is used for this and nothing else. Privacy