The Lab

How we tested

• Minute-by-minute spot, futures and mark prices (the mark is the price the exchange uses to liquidate), for Bitcoin and Ether. Solana, XRP, BNB, Dogecoin and Cardano were run as a check.
• The crash days: COVID (12 March 2020), 19 May 2021, June 2022, the FTX week (November 2022), 5 August 2024 and, in focus, 10 October 2025. Plus every day Bitcoin or Ether moved 15% from the day before: 42 periods from 2020 to 2026, each from the day before to the day after.
• The book: the same coins long and short; reset each day, if still alive.
• Leverage on both sides:
  • spot: paid in full, or 2, 3 or 5 times with the rest borrowed;
  • cash cushion: none, 5%, 10%, 20% or 50% of the value.
• The rule for one account (Bybit's): it is liquidated when what it must keep reaches its counted value. The counted value is the coins at the collateral rate, less the loan, plus the cash and the short's profit or loss. What it must keep is the short's maintenance margin (0.33% for Bitcoin and Ether) and 5% of the loan.
• Bybit's collateral rates, first tier: Bitcoin and Ether 98%; Solana and XRP 95%; BNB, Dogecoin and Cardano 90%. They step down for large holdings: Ether to 95% above 2,700 coins and 80% above 22,000; the others to 60–70%. We also ran 100% down to 50% for every coin.
• Calm months: what each set-up earned in the 45 months with no crash, after fees and a loan rate of 10% a year (5% and 20% also run).

The maths, step by step

1. One account. Hold 1 Bitcoin at 100,000, paid for, and short 1 Bitcoin. The price jumps to 110,000. The coins count as 110,000 × 98% = 107,800. The short has lost 10,000. Counted value: 97,800. The account must keep 0.33% × 110,000 = 363. It's nowhere near liquidation.

2. Separate accounts. The same short, isolated with 5% margin (5,000). At 110,000 it has lost 10,000, double its margin: it was liquidated on the way up, at about +4.7%. The coins gained 10,000 in another account that couldn't help.

3. Borrowing, and a cut in the collateral rate. Pay 33,333 and borrow 66,667 to hold 1 Bitcoin (3x on spot). If the exchange cuts the collateral rate to 50%, the coins count as 50,000: less than the 66,667 loan. The account is liquidated without the price moving at all.

4. What borrowing earns. The paid-for book earned about 5.9% a year on its money in calm months. At 3x, the loan of two thirds costs 10% a year, 6.7% of the value, more than the funding brings in. So it loses about 2% a year.

What we found

1. One account survived every crash. Separate accounts mostly didn't.

• At Bybit's 98% rate, all 20 set-ups in one account were never liquidated, on Bitcoin or Ether, in any of the 42 periods. That covers every spot leverage and cushion, up to 5 times the money.
• In separate accounts, only 1 of 20 survived: no loan and a 50% cushion.
• On 10 October 2025:
  • Bitcoin books in separate accounts with 5x on spot were liquidated at 21:17 UTC: the spot loan account was sold off in the crash, losing 9% to 25% of the money, while the short's gain sat in another account.
  • The same books in one account fell at most 3.5% at the worst minute, and ended the day within 1%.

2. The real danger is the collateral rate. Of the 20 set-ups:

3. The gaps on 10 October 2025.

• Around 21:20 UTC the futures mark fell below spot by 18.7% (Solana), 11.6% (XRP) and 8.6% (Dogecoin). Bitcoin's gap was only 0.8%, Ether's 2.9%.
• A mark below spot helps this book: the short gains more than the coins lose.
• The dangerous gap is the mark above spot. That happened in Dogecoin's squeeze on 29 January 2021 (+10.6%).

4. Smaller coins are less safe. At their own Bybit rates (90–95%), books in one account with 2x or more on spot were liquidated on Dogecoin (January and February 2021), Solana (the FTX week, a 20.6% gap), BNB (February 2021) and Cardano (March 2025). With the coins paid for, none was, on any coin.

5. Borrowing for the spot side doesn't pay. A year on the money, calm months, at a 10% loan rate:

6. Two checks.

• Real liquidation records are public for only one stretch: Binance's coin-margined futures, June 2023 to October 2024. On 5 August 2024 the minutes with the most liquidations (01:10, 06:08 and 06:24 UTC) matched the day's sharpest minute drops. So minute prices do see the bursts. None of that history is free for 10 October 2025.
• A real case: a third-party post-mortem by EarnPark describes a hedged book of about 5.5x in a Binance cross-margin account. It was liquidated on 23 August 2026 for about −30%. Their stated cause: their own model counted the collateral at full value while the exchange counted it at less.

Update, 16 October 2026: a test on a demo account

Added after this page was published. The findings above are unchanged. This was an exchange demo account, with no real money.

We set up the book this page describes, in one cross-margin account: Bitcoin, Ether and Solana, a third each, bought partly with borrowed money, each meant to be hedged by an equal short. It was set to borrow far more than the 5x this page tested.

What happened (UTC):

• 07:19. Bitcoin and Ether were bought, about 2.9 times the account's money each. Solana's buy was refused for lack of balance. All three shorts were refused. At the futures' 10x leverage, each needed a tenth of its size as margin, and the loan had used most of the room.
• That left the account long about 5.7 times its money, with no hedge. A fall of about 12% in Bitcoin and Ether would have liquidated it.
• About 09:40. The shorts' leverage was raised to 100x, so each needed only a hundredth of its size as margin. Bitcoin and Ether took it. Solana was refused at 50x: the exchange's leverage cap falls as a position grows, and Solana's size put it in a tier capped at 40x.
• 09:42. The Bitcoin and Ether shorts filled, about 3.2 times the money each. The book was hedged again. The exchange's own estimate put liquidation about 20% above the price.
• 09:43. Our software's loan-cost check found the loan cost 3.93% a year, while the shorts earned 0.90% a year in funding. It cut the book back to no borrowing, selling coins and closing shorts in pairs. Nothing was liquidated.

What it adds to this page:

• Once both sides were on, one account did what this page says. The hedge held and the margin was ample. But the full hedged book lasted under two minutes, so this tells us nothing new about crashes.
• The two sides have to open together. For over two hours the account held the coins without the shorts: a plain bet on the price. The cause was in our own software's order of steps, and it has been fixed.
• The leverage cap depends on size. This page used the lowest tier only. A large short can't use the top leverage.
• At this funding, borrowing loses. The loan cost more than four times what the shorts earned, as step 4 warned.

Update, 16 October 2026: a correction and a clarification

Added after this page was published. The text above is left as it was.

• "At today's loan rates it loses money" went too far. The study assumed a 10% loan. Today's rate is 3.93%, closer to the 5% case it also ran. At 5%, in calm months, borrowing earned a little more than paying in full: Bitcoin +6.9% a year at 2x, +7.8% at 3x and +9.7% at 5x, against +5.9%; Ether +7.4%, +8.6% and +11.0%, against +6.2%. It loses when funding falls under the loan's rate, as it had on the day of the demo test above.
• The advice stands for a different reason. Borrowing adds the one risk that can liquidate the account without the price moving (step 3), for a gain of a few points a year at best.
• "Highest safe leverage" in the collateral table is the whole account's: the book's value over the money put in, cash cushion included. So 4x is 5 times on the coins with a 5% cushion (1 ÷ (1/5 + 5%) = 4), and 1.4x is 2 times with a 20% cushion (1 ÷ (1/2 + 20%) = 1.43).

Update, 27 October 2026: where the idea comes from

Added after this page was published. The text above is left as it was.

• The carry trade is older than crypto: borrow where money is cheap and hold where it pays. In crypto the pay is funding. Its history and maths are on our theory page "Funding carry: getting paid to hold a hedged coin".
• Margining a whole book at once comes from the futures exchanges. The Chicago Mercantile Exchange's SPAN system, from 1988, sets one margin for a whole portfolio's risk, not for each position alone. Bybit's unified account works on the same idea.
  • For example: 1 Bitcoin held, and 1 sold short on futures, at 60,000. The price jumps 10%. The coin gains 6,000 and the short loses 6,000.
  • In one account the two cancel, and nothing happens.
  • In separate accounts, a short opened at 10x has 6,000 of margin. The 6,000 loss uses all of it, so the short is liquidated, though the coin gained the same 6,000 in the other account.
• The catch. One account only helps while the exchange counts the coins at nearly full value. Bybit can cut its collateral rates at any time, and a crash is when that is most likely. The real case on this page shows how a book can fail even when the hedge is right.
• Theirs and ours. The carry trade, the perpetual future (Shiller), portfolio margining (the CME) and Bybit's rules are theirs, credited below. The crash replay, the 5% loan margin and the comparison of one account with separate ones are ours.

Credits

• The carry trade is long-standing market practice, not ours. It is measured in crypto by:
• Maik Schmeling, Andreas Schrimpf and Karamfil Todorov, "Crypto Carry", BIS Working Paper 1087, Bank for International Settlements, 2023 (revised 2025).
• Robert J. Shiller, "Measuring Asset Values for Cash Settlement in Derivative Markets: Hedonic Repeated Measures Indices and Perpetual Futures", Journal of Finance, 1993: the perpetual future, proposed. Crypto exchanges built it, starting with BitMEX in 2016.
• Bybit's published collateral rates, unified-account liquidation rule and maintenance margins, as read on 15 October 2026.
• EarnPark, "Post-mortem: the 23 August 2026 liquidation and what we changed", published on its own site, as the real case. It is EarnPark's own account of events.
• Chicago Mercantile Exchange, SPAN (Standard Portfolio Analysis of Risk), 1988: one margin for a whole portfolio.
• The crash replay and the 5% loan margin are ours.
• Data: Binance's public minute prices (spot, futures and mark) and its liquidation records.
• PerpGremlin and Claude, The carry trade in one account: does it survive the crashes?, TerminalGremlin, 2026: the tests, the code and the words.