Eight Minutes on October 10

Under the rule Bybit publishes, a bet on bitcoin rising made with 100× leverage is closed once the exchange's reference price falls 0.5% from the entry price, and one made with 10× leverage after a 9.5% fall.

That rule starts with leverage: putting down only part of a position's value as a deposit, called margin, so that at 100× leverage a trader posts 1% of the position's value and at 10× leverage 10%. The exchange does not wait until that deposit is used up, because its rule keeps back a small slice of the position's value, and it closes the position by force as soon as the margin left falls to that slice. That forced closing is called a liquidation, and on Bybit it is triggered by the exchange's reference price, a price the exchange calculates rather than the last trade.^[3]

On Bybit's help page, the slice for a bitcoin position worth up to 2 million USDT, a token tied to the US dollar, is 0.5% of the position's value, an example rate that we cannot confirm for every position on October 10.^[3] Take the simplest case, a single bet on a rise backed only by its own margin, with no fees, funding or added margin: the distance from the entry price to liquidation is the margin posted minus that slice. At 100× a fall of 0.5% in the reference price therefore closes the position, and at 10× a fall of 9.5%, so the trader loses the position before the price has moved the full amount of the margin posted.

When a trader lets the whole account's balance back every position instead, the leverage chosen for one bet no longer sets how far the price can move before it is closed, so the figure shows how the rule works, not where any trader stood. The figure starts from a hypothetical entry price because Bybit, whose rule it uses, does not publish where its traders' liquidation prices sit. Whichever backs a position, its own margin or the whole account, closing it means trading, and the forced trade can go to the order book, the exchange's public list of offers to buy and sell, or be handled another way.

A forced closing of a bet on a rise is a sale, and when it is sold into the order book it can push the price further down, toward other positions' closing points.

Which way a forced trade pushes depends on the side of the bet: closing a long position, a bet on a rise, means selling, and closing a short position, a bet on a fall, means buying. Where the trade goes is set by each exchange's rules. On Hyperliquid, for example, a liquidation is first sent to the order book as orders to trade at once at the best prices waiting there. The rules also provide a second route off the order book, called the backstop: the exchange's own liquidator account takes the position over, or the position is closed directly against traders on the other side.^[4]

The first route is the one that can move the price. A forced sale sent to the order book fills against the best offers to buy that are waiting there, so it can push the price lower, and a lower price can bring other positions to their own liquidation points, whose forced sales can push it lower again. That loop is what a liquidation cascade means. In a 2009 paper, the economists Markus Brunnermeier and Lasse Heje Pedersen described a "loss spiral" in which losses on a position end up "forcing speculators to sell more, causing a further price drop, and so on."^[5] In 2022 the US Financial Stability Oversight Council warned of the same loop in crypto markets: "price declines could expose leveraged positions, potentially leading to deleveraging or liquidation and causing further price declines."^[6]

Neither source is evidence about October 10, since the paper models professional traders who use borrowed money rather than crypto exchanges, and the council warned in general terms. The price cannot show how much of a fall came from forced closings either, because a falling price records every trade alike, whether the seller was forced, chose to sell, or met fewer buyers as buyers stepped back from the book. What can be checked is the setup: a chain like this needs many leveraged positions to be waiting, so there should be signs of them beforehand, and two public numbers are commonly watched for them.

Two public numbers describe these contracts: open interest counts how many are open, and funding shows whether holders of long positions are paying more than the fixed rate built into the exchange's formula to keep them.

The first is open interest, the number of contracts open at a given moment. Every contract has a holder on each side, one long and one short, so open interest counts positions without saying which side is crowded; it also says nothing about how much margin stands behind them, which means it is not a measure of leverage. The second is funding, a payment between the two sides that both exchanges settle every eight hours; each payment is called a settlement. When funding is positive, longs pay shorts, and when it is negative, shorts pay longs.^[3][7]

Its size depends on the premium, the contract's distance from an index of bitcoin's price that the exchange publishes. While the contract stays within a narrow range of the index, both exchanges' formulas fix funding at 0.01% every eight hours, a fixed interest part that longs pay even on a quiet day; this essay calls that the resting level. Funding goes above it only when the contract trades far enough above the index, and then longs pay extra, as happens when many traders buy with leverage and lift the contract. Funding is therefore the price of leveraged buying rather than its amount, and it does not say whether a fall is coming. Because open interest counts both sides, crowding on one side shows up in funding, not in the count. Even so, if the chain on October 10 needed a build-up of leveraged longs, both numbers should have given an unusual reading in the weeks before, and on the exchanges where they can still be checked, neither did.

Before October 10, 2025, funding on Bybit and Binance never once rose above its resting level, and open positions on Bybit's bitcoin contract ended the sixteen days about 5% higher.

We looked at the sixteen days before the crash, from midnight on September 24 to midnight on October 10, on Bybit and Binance, the two exchanges in our composite.^[2] Each exchange settled funding 49 times in that window, and on neither was it above the resting level even once: on Bybit it sat at the resting level 13 times and below it 36 times, and on Binance it was at the resting level 9 times and below it 40 times. Below the resting level, longs pay less than the fixed interest part, and when funding turns negative, shorts pay longs instead. Under both exchanges' formulas, funding falls below that level only when the contract trades at a discount to its index large enough to move funding, the opposite of a premium; so on these two exchanges' bitcoin contracts, longs paid no premium for leveraged buying in the weeks before the fall.

Open interest can be checked on only one of the two exchanges, because Binance's public open-interest history reaches back just thirty days, while Bybit's can still be downloaded for those weeks. At midnight on October 10, Bybit's bitcoin contract had about 35,300 bitcoin of open contracts, counting one side of each, and the rise in the heading is the net change over the sixteen days, after a dip in late September. A change like that means little on its own, so the next section sets it against the same count, on the same exchange, before the other crash.

None of this means there was no leverage. The leverage was there, as always; what the usual gauges did not show is an unusual build-up. A build-up could have formed in smaller coins, on other exchanges, in dollar terms or in accounts where one balance backs every position, and these public series cannot rule that out.

Before bitcoin's crash on May 19, 2021, funding was above its resting level in most settlements and open positions on Bybit ended higher, so a crash can follow raised readings as well as calm ones.

The other time was May 19, 2021. On the same two-exchange composite, bitcoin fell 29.9% in just under twelve hours, the steepest fall between two closing prices no more than twelve hours apart that day.^[2] The IMF judged afterwards that "The sharp declines during May were likely exacerbated by high use of leverage …, which led to automatic liquidations of margin and futures positions by exchanges."^[8] That is an official assessment, worded as likely, of what made the fall worse, and it rests on the IMF's own data rather than on the gauges used here.

Over the same sixteen days before that crash, funding was above its resting level in 42 of Bybit's 49 settlements and in 45 of Binance's, so in most settlements longs were paying shorts more than the fixed interest part, which is what crowding on the long side looks like in funding. Separately, open positions on Bybit's bitcoin contract ended those sixteen days more than a fifth higher. In October 2025 funding showed no crowding on the long side, because it never rose above its resting level on either exchange; separately, open positions on Bybit ended the sixteen days about 5% higher. The two counts come from the same exchange over windows of the same length, which is why they can stand side by side.

Two cases cannot say how often a reading like 2021's is followed by a crash, or how often it fades with no crash at all. What they do show is that a crash can arrive from either state: a quiet gauge is not an all-clear, and a loud one is not a forecast. So the gauges we can still check gave no unusual reading before October 10, and even a loud one, as 2021 shows, would not have said when. What the price and the exchanges recorded in the eight minutes is another matter, and the chart from the opening section can now be read again with the rule in mind.

By 21:20, open contracts on Bybit's bitcoin market were 22% fewer than just before the post, and the count cannot show whether they were closed by force or by choice.

With the rule in mind, the steep section of the opening chart can be read again. Every liquidation price that the falling reference price crossed in those eight minutes could add a forced sale to an order book, and each such sale could push the price toward the next liquidation price, which is the loop described earlier. A fall that gathers speed after a slow start is therefore what the rules allow for. The price alone still cannot show how many of the sales in those minutes were forced, so the chart shows that the rules could have been at work, not that they were.

Bybit's open interest, recorded every five minutes, shows what happened to positions over the same stretch. By 21:20 the number of open contracts on Bybit's bitcoin market was 22% lower than just before the post. Part of that drop came during the slide, but most of it came in the ten minutes that contain the steep fall, and afterwards the count kept falling for about another fifty minutes, then leveled off. A fall in open interest means contracts were closed, but it cannot say whether by force or by choice, because a trader who closes a position and an exchange that liquidates one change the count in the same way. These numbers therefore show positions closing while the price fell, not which caused which. Nor do they explain why the fall ended: the stock of open positions shrank and the fall came to an end, but the data cannot show that one ended the other. The clearest record of what was being forced shut in those minutes comes from elsewhere, from an exchange that writes every trade to a public blockchain.

On Hyperliquid, an exchange that records every trade on a public blockchain, at least $2.1 billion of traders' positions, at face value, were force-closed between 21:15 and 21:27 UTC.

Hyperliquid's public data record every trade and label each liquidation with the route it took, so forced closings there can be counted trade by trade.^[4] Between 21:15 and 21:27, at least $2.1 billion of traders' positions were force-closed by liquidation there. That window is not one we chose: it is the window covered by an independent reconstruction of the exchange's public trade records, and it starts while the steep fall was under way and runs past its end. The amount is face value: the size of the positions, not the money their owners lost. The difference matters because a leveraged position is many times larger than the margin behind it, so the two kinds of number can be far apart. Traders' positions are those in the accounts of the exchange's users, not in its own accounts, and the amount covers one exchange, not the market as a whole.

That is a floor, not a total, because it combines two sources that each see only part of the picture. The first is the reconstruction itself, published on GitHub under the name ConejoCapital, and it contains only the liquidations that took the backstop route, so on its own it would miss every liquidation that went to the order book.^[9] The second is our own check: for a sample of accounts we looked up their trades in the exchange's public data, found liquidations through the order book that the reconstruction leaves out, and added them. For the accounts we checked, every row of the reconstruction matched a trade in the exchange's records; for the other rows that match has not been tested. The check therefore raises our confidence in the rows the reconstruction contains without telling us how much it leaves out in total. Forced closings can leave an exchange by two routes, and the records show which one each took.

On Hyperliquid both routes were in use: some forced closings went to the order book, and about $1.8 billion went through the exchange's backstop instead.

The exchange's records label each liquidation with its route, and both routes were in use. Some liquidations were closed by orders in the public order book, the route described earlier, the one that can push the price further. The rest went the other way: rather than being sold into the book, the positions were taken over by the exchange's liquidator accounts, or closed directly by auto-deleveraging, which matches the position against traders holding the opposite side, at the exchange's reference price, whether those traders want it or not.^[4] That second route happens off the order book, so at the moment of the takeover or the match these positions did not meet the offers waiting in the book. Both routes sit inside the floor of at least $2.1 billion: about $1.8 billion went through the backstop, the only route the reconstruction records, and the rest is the order-book liquidations our own check found in the sampled accounts, only part of what went through the book.

The size of the whole order-book part is the number a reader most wants, because that is the route that can push the price, and it is the one number this essay cannot give. A sample can show that the route was used, but it cannot be scaled up to the whole exchange, because the sample was drawn from the accounts in the reconstruction, and accounts liquidated only through the order book are not in it. Scaling it up would mean assuming that the accounts outside the reconstruction behaved like the ones inside it, and nothing in the records supports that assumption. Nothing caps the total from above either, which is why the essay makes no guess at it. So we do not know how large the order-book part was, and therefore not how much it could have added to the fall: the forced closings that went through the book could have pushed the price further, as the loop allows, while the ones that went through the backstop were not sold into the book when the backstop handled them. Even on the exchange that publishes every trade, then, we could only put a floor under the total, and where an exchange does not publish every trade, the total has no exact answer either.

Neither the lowest price nor the total liquidated on October 10 is a single exact number: exchanges traded at different lows, and the best-known total is, by its publisher's own account, an undercount.

Even the depth of the fall has no single number. Each exchange has its own order book and runs its own liquidations, so in a fast fall the lowest trade is not the same everywhere: on October 10 the lowest bitcoin trades on major spot markets, where bitcoin itself changes hands, and on futures markets ranged from about $101,000 to $107,000.^[2] Within our own composite the gap was clearest in ether: at 21:20 the two exchanges in our composite had lowest ether trades 2.6% apart. Our composite's low is the volume-weighted average of the two, not a trade on either, and we found no primary source that explains why the lows differed.

The total liquidated is harder still to pin down. By Coinglass's count, liquidations around the event "exceeded $19 billion" at face value, with no exact window given, and the data provider's own estimate is that "the actual nominal liquidation scale likely approached $30–40 billion".^[10] Another provider, Amberdata, reports $9.89 billion on its own dataset over about fourteen hours from 20:50.^[11] Our own floor covers only Hyperliquid, and only its window. These figures measure different things over different windows, none is a measured total, and they cannot be added or set against each other. One reason public totals can fall short: Binance's public stream reports "only the latest one liquidation order within 1000ms" for each contract, so other liquidations on that contract in the same second do not appear in it.^[12]

Each exchange's published rules turn a price drop into forced closings, and a fall like this one can come whether the usual gauges beforehand read calm, as in 2025, or raised, as in 2021.

So the answer to the opening question comes in two parts, one about the mechanism and one about what anyone could have seen. For the mechanism, the forced closings come from each exchange's published rules: a position is closed once the reference price crosses a line set by its margin, and closing a long is a sale. On October 10 those closings happened in minutes that overlapped the steepest part of the fall, and on Hyperliquid both routes were in use, so the rules were at work while the price fell fastest. How far the forced closings fed on each other cannot be read from the price, because the price records forced and chosen sales alike.

The second part concerns foresight. On the exchanges where the usual gauges can still be checked, they gave no unusual reading before October 10, 2025, while before May 19, 2021 they read raised, and both readings were followed by a crash. That is two cases, with open interest from one exchange and funding from two, and the size of October 10 is a floor on one exchange and an estimate everywhere else. Within those limits the answer is this: the rules that turn a fall into forced closings are published and work the same way whatever the gauges read beforehand, and the two cases show that a fall like the one on October 10 can come after calm readings, as in 2025, or after raised ones, as in 2021.