What happened after crypto liquidation cascades? 300 cases
We measured Hyperliquid's 300 largest liquidation episodes: price 15 minutes to a day later, longs vs shorts, before fees, and the limits.
In the 300 largest liquidation episodes we measured on Hyperliquid, from June 2025 to September 2026, price was higher one hour later in 58.33% of them. The middle one-hour move was +0.39% before fees. After forced selling of longs the middle move was +0.81%; after forced buying of shorts it was −0.59%; and a day later the overall split was close to even, with 48.00% higher.
These are past cases from one exchange's records, not a forecast, and not proof of a trade that made money after costs.
What is a crypto liquidation cascade?
A liquidation happens when an exchange closes a trade because the trader's money backing it ran out. A cascade means many liquidations in a short time.
A long is a position that gains when price rises. Closing a long by force means selling. A short gains when price falls. Closing a short by force means buying.
We used Hyperliquid perps: futures contracts that never expire.
A bar is a price record for a fixed span of time. We flagged a five-minute bar as a cascade bar when the forced closures on one side were at least $5,000 and above the 95th percentile of that side's previous 288 five-minute bars (the past 24 hours). In plain words, that side's liquidations were near the top of the last day's range. We skipped bars where both sides qualified.
Why do we count episodes instead of single bars?
A big crash rarely fits in five minutes. One bad day can hold hundreds of cascade bars across many coins, and counting each bar as a case lets that day fill the sample.
So we merged cascade bars into episodes. Any two cascade bars, on any coin, that started less than one hour apart joined the same episode, and chains of these joined too. From each episode we kept one bar: the one with the most dollars liquidated. Every price move starts at the close of that bar, on that bar's coin.
Here is the clearest example. On October 10, 2025, one episode ran from a PUMP bar at 14:15 UTC to a last bar starting at 03:30 UTC the next day. It held 868 cascade bars. Its largest bar was PUMP at 21:15 UTC, a forced-long bar with about $106 million liquidated. One hour after that bar closed, PUMP was +77.92% higher. That is the biggest one-hour rise among the 300 episodes, and it counts once, not 868 times.
Merging has a cost too. 196 of the 300 episodes had both longs and shorts forced out at different moments. The side label below is the side of the largest bar only, not of every bar in the episode.
What did prices do after the largest cascades?
The median is the middle result when you sort all the moves from lowest to highest. "Share up" counts cases where the price was strictly higher; flat counts as not up.
All 300 episodes had a price at every time step, so each row uses the same cases. Moves are before fees.
| Time after the bar's close | All 300: median (share up) | 210 forced-long: median (share up) | 90 forced-short: median (share up) |
|---|---|---|---|
| 15 minutes | +0.19% (58.33%) | +0.41% (65.71%) | −0.23% (41.11%) |
| 1 hour | +0.39% (58.33%) | +0.81% (66.19%) | −0.59% (40.00%) |
| 4 hours | +0.44% (56.33%) | +0.99% (64.76%) | −0.83% (36.67%) |
| 24 hours | −0.38% (48.00%) | +0.20% (52.38%) | −1.65% (37.78%) |
In this sample, price more often moved against the forced closures for the first few hours. After forced selling, the middle result was a rise. After forced buying, the middle result was a fall. By 24 hours, the forced-long group was close to a coin flip, with 52.38% up.
How much did cases disagree?
The middle result hides a wide spread. The 10th and 90th percentiles are cutoffs that leave about a tenth of cases below the lower number and a tenth above the upper one.
- Forced-long episodes, one hour: −1.91% to +3.72%.
- Forced-short episodes, one hour: −4.26% to +1.46%.
- All 300 episodes, 24 hours: −12.65% to +11.44%.
The extremes go both ways. The largest episode in the sample was XPL on August 26, 2025, a forced-short bar with about $136 million liquidated. One hour after it closed, XPL was −55.23% lower, and a day later it was still −48.94% lower. The PUMP episode above went the other way by a similar size. And a big forced-long case did not always bounce: POPCAT on November 12, 2025, with about $35 million of longs closed in one bar, was −28.52% lower an hour later.
Even in the forced-long group, about a third of the one-hour moves were not up. Our guide to why past crypto moves are not a forecast explains why a past share is not a promised chance next time, and when past crypto cases disagree shows how to read a spread like this.
What happens when we include trading fees?
We assumed a 0.05% fee to open and 0.05% to close, 0.10 percentage points in total. This is a taker fee, paid when an order trades right away at the price on offer.
- An example long after the 210 forced-long episodes: the one-hour median fell from +0.81% to +0.71%, with 65.24% of cases above zero.
- An example short after the 90 forced-short episodes: the one-hour median was +0.49% after fees, with 58.89% above zero.
These examples leave out the gap between buy and sell prices, worse fills on big orders, funding payments between perp traders, and delay. They assume a trade exactly at the bar's close. They are not returns anyone earned.
How we measured this
We used Hyperliquid's five-minute cascade bars from June 2025 to September 2026, across the perp coins in the archive. Liquidation records for April and May 2025 were missing or partial, so the first possible event is June 2, 2025. We dropped any bar whose previous 24 hours, its own five minutes, or its next 24 hours had a missing minute or no real trade at a price point we needed, which left 30,828 bars. Those merged into 2,514 episodes, and we kept the 300 with the largest single bar; the smallest of them had $878,798 liquidated in that bar. The 300 fall on 184 different UTC days and cover 35 coins. Each price move compares the coin's price at the end of a later minute with its price at the bar's close, using real trades in both minutes; we filled no missing prices.
Limits: we picked the largest episodes after the fact, so this is not a rule you could have followed live. Our records can only show liquidations the exchange's trade feed marked, and the strict data checks favor busy coins. Four-hour and 24-hour windows from different episodes can overlap, so the 300 cases are not 300 independent tests.
What can you take from this into your own research?
If you study events like these yourself, merge events that are close in time before you count them, and measure once per episode from a fixed starting bar. Without that, one crash day can count hundreds of times. Then split by side and look at the 10th and 90th percentiles, not just the middle result.
Try this in Comparand
With Comparand connected to your AI app, you can ask about Hyperliquid:
- "Show the cascades on October 10, 2025, with what followed each event." The daily tool (hl_cascade_events with after=1) returns that archive day's cascade events and what the price did after each one.
- "Find PUMP cascades from October 1 to October 31, 2025, with later outcomes." The search tool (hl_cascade_search) covers up to 548 archive days, one row per event. You can filter by coin, direction or minimum dollars liquidated. It returns no averages.
- "Show the largest liquidation minutes on October 10, 2025." The ranking tool (hl_liquidations_top) returns the largest single minutes in a span, for all coins or ones you list, and for the buy or sell side.
- "Show PUMP's hourly liquidation totals for October 2025." The totals tool (hl_liquidations_series) returns one coin's totals per complete UTC hour, or per day.
These tools return single events and totals; merging them into episodes and working out medians is a step you do yourself. A study of a large hourly price drop picks events by price change instead, which is a different question.
You can try 3 coins free, see the data tool docs, or check pricing.
The short version
- Across the 300 largest Hyperliquid liquidation episodes, price was higher one hour later in 58.33% of cases, with a middle move of +0.39% before fees.
- After forced long selling, the middle one-hour move was +0.81%; after forced short buying, it was −0.59%.
- A day later the overall split was close to even (48.00% up), and the spread ran from −12.65% to +11.44% at the 10th and 90th percentiles.
- We merged nearby events into episodes so one crash day counts once; the cases are still chosen after the fact and not independent.
- Past cases, not a forecast, and not a trading edge after real costs.
Questions people ask
Does crypto always bounce after a liquidation cascade?
No. In the 300 largest Hyperliquid liquidation episodes we measured (June 2025 to September 2026), price was higher one hour later in 58.33% of them. A day later, only 48.00% were higher.
What is the difference between long and short liquidations?
A long gains when price rises, and closing it by force means selling. A short gains when price falls, and closing it by force means buying. The label tells you which positions were closed, not which way price went next.
What happened after long liquidations compared with short liquidations?
In our 300 Hyperliquid episodes, the 210 led by forced long closures had a middle one-hour move of +0.81%. The 90 led by forced short closures had a middle one-hour move of −0.59%. Both groups had many cases that went the other way.
Do these results prove a profitable trading rule?
No. We picked the largest episodes after they happened, the numbers are before fees and other costs, and episodes close in time can still share the same market move. They describe past cases, not a forecast.