Liquidation, insurance funds and auto-deleveraging

How perpetual venues allocate losses: liquidation, cascades, insurance funds and auto-deleveraging rules, with October 10-11, 2025 as a case where key figures still disagree.

Lesson 13 of 15Advancedabout 35 minAll lessons

You’ll learn

  • Trace a failed position through liquidation, backstop, insurance fund and auto-deleveraging
  • Calculate a bankruptcy price and the shortfall when a position closes past it
  • Model a liquidation cascade as a feedback loop, and state when the loop amplifies a shock
  • Compare auto-deleveraging rules across venues, and compute queue and pro-rata allocations
  • Assess conflicting accounts of October 10-11, 2025 by asking what each figure measures and who reported it
Real story

October 10-11, 2025

On October 10, 2025 (UTC), crypto derivatives venues processed the largest liquidation event on record in figures compiled by CoinGlass, with reported totals near $19 billion within 24 hours. On Hyperliquid, about 35,000 auto-deleveraging executions closed positions held by nearly 20,000 wallets, in what was reported as the mechanism’s first activation there in about two years. Traders who believed they were hedged found profitable short positions closed by rule, and USDe, a dollar-tracking token, traded in the mid-$0.60s on Binance and stayed much nearer $1 elsewhere. A year later, the figures, the causes and the fairness of the outcome all remain disputed. This lesson builds the mechanism first, and then returns to the case.

Source: <a href='https://arxiv.org/abs/2512.01112'>Chitra (2025, rev. 2026)</a>; <a href='https://www.fticonsulting.com/insights/articles/crypto-crash-october-2025-leverage-met-liquidity'>FTI Consulting (Dec. 24, 2025)</a>; <a href='https://www.coingecko.com/learn/october-10-crypto-crash-explained'>CoinGecko</a>; <a href='https://www.gncrypto.news/news/adl-mechanism-crypto-exchanges-liquidation/'>GN Crypto</a>

The sequence when an account fails

A derivatives venue promises that winners get paid. Perpetual contracts are zero-sum before fees: one side’s gain is the other side’s loss. So when a loser cannot pay, the shortfall has to land somewhere. A venue’s rules on this point are its real risk policy. Most venues use a sequence of the following form, though each differs in detail.

StageWhat happensWho bears the cost
1. Margin breachEquity falls below maintenance margin, measured at the mark priceNobody yet
2. Market liquidationThe venue sends orders to the book to close some or all of the positionThe trader, including any penalty
3. BackstopA vault or insurance fund takes over what the book could not absorbThe vault or fund, if the close is below water
4. Insurance fundReserves cover losses beyond the account’s equityThe fund, built from fees and penalties
5. Auto-deleveragingProfitable opposing positions are closed by rule to cover the restWinning traders

A generalized sequence. Order and details differ by venue; see the rules table later in this lesson.

Definition

Bankruptcy price: The price at which an account’s equity is exactly zero. For a single position: P0 − C/q for a long, P0 + C/|q| for a short. Binance defines it as the price at which the liquidated user’s margin balance reaches zero.

Definition

Shortfall (bad debt): The loss remaining once an account’s equity is exhausted: for a long closed at Px, D = max(0, q(Pbk − Px)).

Definition

Auto-deleveraging (ADL): A rule-based closure of profitable positions, against a position that cannot be closed otherwise, used to prevent bad debt from falling on the venue. It is a form of loss socialization.

Example: from a liquidation trigger to a shortfall

  1. Example position: 10x long, $10,000 notional at $100 entry, so q = 100, with $1,000 of collateral. The bankruptcy price is 100 − 1,000/100 = $90.
  2. Suppose a maintenance fraction of 1.25%, which sets the liquidation trigger at (100 − 10) / 0.9875 = $91.14.
  3. The book is thin, so the closing orders fill at an average of $88.
  4. The position’s loss = 100 × (100 − 88) = $1,200. Equity = 1,000 − 1,200 = −$200.
  5. The shortfall of $200 is 20% of the collateral and 2% of the notional.
The gap between the trigger and the bankruptcy price is the only buffer. A move through it, or a close beyond it, creates a loss that someone else must absorb.

Hyperliquid’s documentation shows how specific these rules can be. It first sends market orders to the book, for the whole position, or only 20% for positions above 100,000 USDC, with a 30-second cooldown afterward. If equity falls below two thirds of maintenance margin and the book has not cleared the position, the Liquidator Vault, a component of HLP, can take it over, and maintenance margin is not returned. Binance’s Futures Insurance Fund takes over the portion of a liquidated position that cannot be closed normally, and it bears the loss or keeps the profit from those positions.

dYdX v4 prices liquidation orders with a limit called the Fillable Price, set from the oracle price and the account’s remaining margin ratio, and the insurance fund takes the profit or loss. The default maximum penalty is 1.5%, adjustable by governance. Its documents describe a boundary case: if an account’s value drops below zero, deleveraging is applied immediately and the insurance fund is not used.

Cascades as a feedback mechanism

A single liquidation is a sale or purchase forced by a price move. If that order moves the price further, it can trigger the next liquidation. The mechanism is a loop.

A simple linear model isolates the structure. Suppose each 1% move triggers β dollars of forced orders, and each dollar of forced orders moves the price by λ percent. Then a first move of s percent induces a second move of βλs, a third of (βλ)2s, and so on.

Total move = s + s(βλ) + s(βλ)2 + … = s / (1 − βλ) when βλ < 1

s is the initial shock, β the forced notional triggered per 1% of move, and λ the price impact per dollar of forced notional. If βλ reaches 1 or more, this linear model has no finite solution, meaning the loop stops only when positions or the model’s linearity run out.

Example: the same shock in deeper and thinner markets

  1. Example inputs: s = 1%, and β = $200 million of forced orders per 1% move.
  2. Deep market: λ = 0.0015% per $1 million. βλ = 0.30, so the total move = 1 / 0.70 = 1.43%.
  3. Normal market: λ = 0.003% per $1 million. βλ = 0.60, so the total move = 1 / 0.40 = 2.5%.
  4. Thin market: λ = 0.0045% per $1 million. βλ = 0.90, so the total move = 1 / 0.10 = 10%.
A threefold fall in depth turned a 1.43% outcome into a 10% outcome from the same shock. The example shows why depth and leverage cannot be studied apart.

This toy model is not an estimate, and it leaves out much of what happens. β changes along the path, because liquidations cluster at particular levels. λ changes too, because market makers often widen spreads or leave when volatility rises. Real outcomes also depend on arbitrage across venues, and on collateral whose value falls in the same move. Brunnermeier and Pedersen (2009) provide the classic theoretical account in a broader setting. In their model, traders supply market liquidity according to their funding, funding depends on market liquidity, and under certain conditions margins become destabilizing and the two reinforce each other in liquidity spirals. Schmeling, Schrimpf and Todorov suggest the mix of trend-seeking leverage, scarce arbitrage capital and high leverage may help explain frequent severe crypto crashes. That is a suggested mechanism, not a measured causal share.

Insurance funds

An insurance fund sits between bad debt and socialization. In the notation of Chitra (2025), if Dt is the shortfall and IFt the fund balance, the residual that must be socialized is Rt = max(0, Dt − IFt). ADL fires when Rt is positive. The fund typically grows from a part of revenue from fees and liquidation penalties. Chitra cites mid-2025 estimates of about $300 million at OKX and about $150 million at Bybit.

Three limits apply. First, funds may be segmented. Binance states that coin-margined contracts sharing a collateral asset share one fund, and it notes those funds are smaller, making ADL more likely. Second, a fund can be drawn on faster than it was built. Press reports citing Binance’s data (Bitcoin Sistemi, October 11, 2025) put its shared BTC, ETH and BNB USDT-margined fund at about $1.23 billion falling to about $1.04 billion during the crash, with $188 million deployed. Third, a large headline balance does not show coverage for every contract, because a balance held for other contracts or collateral may not be available to the one in trouble.

A vault with depositors, such as Hyperliquid’s HLP, mixes these roles. It earns from market making and liquidations and shares profit and loss with depositors. Hyperliquid warned in March 2025 that “HLP is not a risk-free strategy”, as The Block reported. CoinGecko’s July 2026 review, drawing on on-chain analysts and press reports, estimates HLP earned $40 million to $41.5 million in the October 2025 weekend, about a 10% return in under 48 hours. A backstop can therefore profit from a crash. This is not a contradiction, because a backstop takes on forced positions that others would not, and it has also lost money in other episodes such as March 2025. It does raise the question of incentives, addressed below.

Auto-deleveraging rules across venues

ADL has a surprisingly short history as a formal rule. Chitra’s review states that, to the author’s knowledge, Huobi introduced it in 2015, that BitMEX implemented a ranking of profit times leverage by 2016, and that Binance adopted this formula in 2019. The paper also estimates that well over 95% of perpetual volume is on venues using the ranking model. The closest traditional analogue is a clearing house default waterfall, which can include variation-margin gains haircutting. A key difference, as Chitra notes, is that only clearing members bear those losses in that setting, and the process has manual stages, while ADL can reach any winning trader automatically.

VenueTriggerWho is selectedClosing price
BinanceInsurance fund cannot take over a bankrupt positionQueue: profitable positions with PNL% × effective leverage ranked highest firstBankruptcy price of the liquidated order, which may lie outside the current market range. No trading fee for the affected trader
HyperliquidAccount value or isolated position value becomes negativeQueue of the opposite side, sorted by (mark ÷ entry) × (notional ÷ account value)The previous mark price
dYdX v4Account value below zeroDocumented as randomly chosen offsetting positionsBankruptcy price of the bankrupt account
GMXPending profit-to-pool ratio exceeds the market’s threshold, using raw oracle pricesProfitable positions partly or fully closed. The page read for this lesson does not state a selection ruleNot specified on the page read
BitMEX (closed Sept. 2026)Not verified for this lessonRanked by profit times leverage, as described by ChitraIts documentation could not be opened for this lesson
Drift, ParadexNot verified for this lessonPro-rata mechanisms, as described by ChitraVenue-specific. Primary documents were not checked for this lesson

Sources: venue documentation read November 2026, and Chitra (2026 revision) for BitMEX, Drift and Paradex.

The closing price matters as much as the selection. A trader closed at the bankruptcy price of the failed account, per the Binance and dYdX rules, may receive far less than the market price. In dYdX’s documented example, a long is closed at $1.20 while the market stands at $2. Hyperliquid instead closes at the previous mark price, so the affected trader keeps more of the gain as marked. Comparing outcomes across the rules takes care, because each one allocates a deficit at different prices. Hyperliquid also states an invariant: a user with no open positions will not absorb any platform losses through socialization.

Example: ranking under two queue rules

  1. Binance rule, score = (unrealized profit ÷ notional) × (notional ÷ (balance + unrealized profit)).
  2. Trader A: profit $1,000, notional $10,000, balance $1,000. Score = 0.10 × (10,000 / 2,000) = 0.50.
  3. Trader B: profit $2,000, notional $50,000, balance $50,000. Score = 0.04 × (50,000 / 52,000) = 0.038.
  4. Trader C: profit $300, notional $5,000, balance $250. Score = 0.06 × (5,000 / 550) = 0.545.
  5. Queue: C first, then A, then B.
  6. Hyperliquid rule, for three different example traders, index = (mark ÷ entry) × (notional ÷ account value). Entry $100 and mark $110, notional $11,000, account $2,000 gives 1.10 × 5.5 = 6.05. Mark $140, notional $14,000, account $14,000 gives 1.40. Mark $105, notional $10,500, account $700 gives 1.05 × 15 = 15.75.
Leverage dominates these scores. A small high-leverage trader goes ahead of a much larger profit held at low leverage.

Example: queue compared with pro-rata

  1. Example: four traders hold unrealized profits of $400, $300, $200 and $100, and the shortfall is $120.
  2. If the $400 trader ranks first, a queue takes all $120 from that trader, or 30% of their profit, and the rest lose nothing.
  3. A pro-rata rule applies a uniform haircut h = D ÷ U = 120 / 1,000 = 12%. The payments are $48, $36, $24 and $12.
  4. The queue’s cost is concentrated, while the pro-rata cost is dispersed.
A live queue may also remove more profit than the deficit requires. That excess is what the debate below tries to measure.

ADL has a second-order effect that its accounting hides. A trader’s short position may be half of a hedge, with the long half held in spot, on another venue, or as a staked token. When ADL closes the short, the account is suddenly long with no hedge at the worst moment. FTI Consulting reports that some well-hedged shorts were closed in this way on October 10. CoinDesk Data says ADL also closed positions held by market makers hedging spot exposure, and that the forced unwind likely led them to sell spot to limit losses. That second claim is an inference, not an observed fact. The mechanism is sound, though: a rule meant to stop loss from spreading can itself transmit risk across markets.

The October 10-11, 2025 case: sequence

Time (UTC, Oct. 10)EventReported by
14:57US tariff announcement on Chinese importsAmberdata timeline, cited by CoinGecko
About 20:00 to 20:20BTC decline accelerated (CoinDesk Data dates the sell-off from 20:00)CoinDesk Data; Amberdata via CoinGecko
20:50Binance places the period of market turmoil from 20:50 to 22:00. Amberdata gives $6.93 billion liquidated over the 40 minutes from this timeBinance statement; Amberdata via CoinGecko
21:15$3.21 billion liquidated in 60 secondsAmberdata via CoinGecko
21:16 to 21:27Window of Hyperliquid ADL fills, with $2.1 billion closedChitra (2026), public-data reconstruction
21:20 to 21:21Lowest market prices, according to BinanceBinance statement
21:36 to 22:16Depeg window for USDe, BNSOL and wBETH on Binance, as used for compensation. AMBCrypto reports Binance’s later review as 21:36 to 22:15The Block; AMBCrypto

Chitra’s introduction also states that at least 40 ADL events began at “10:21:00 UTC”. The paper’s data section and all other accounts place the activity after 21:00, so that time seems to be an error in the paper. Treat that one reference with caution.

The case: where figures disagree

QuantityFigures reportedSource and basis
Total liquidations$19.141 billion; $19.27 billion; about $19.3 billion; “about $19 billion to $20 billion”ChainCatcher, The Daily Hodl and KuCoin, citing CoinGlass, with different snapshot times. CryptoSlate gives the range. CoinDesk Data calls it about $19 billion in notional positions unwound
Could the total be higher?CoinGlass later suggested it may be much higher, with an estimate of $30 to $40 billionBitcoinist, reporting CoinGlass, which cited late disclosure by some platforms and feedback from market makers. Coin Metrics documents that Binance’s feed shows at most one liquidation per market per second
Accounts and side1,621,284 accounts; $16.686 billion of longs and $2.455 billion of shorts, so about 87% longChainCatcher, snapshot of CoinGlass-style data; CoinGecko gives over 1.6 million with no source stated
By venueHyperliquid more than $10 billion to $10.3 billion, Bybit $4.6 billion, Binance about $2.4 billionCryptoSlate; GN Crypto, citing CoinGlass. Venue reporting conventions differ
Open interestDown 27.5%, nearly $60 billion, to $158 billion; or about $65 billionCoinDesk Data; Bitwise portfolio manager Jonathan Man via CryptoSlate. Definitions differ
BTC low$122,574 to $104,782 (−14.5%); about $126,000 to about $104,000CoinGecko; GN Crypto. The start points differ, since one is the day’s high and the other nearer the earlier peak
Hyperliquid ADL$2.1 billion closed in 12 minutes; 34,983 executions, 19,337 wallets, 162 tickersChitra (2026), from the on-chain fill log
Hyperliquid losses$1.23 billion of trader equity lost across about 6,300 wallets; $100.1 million realized deficit in the ADL windowCoinmarketman and GN Crypto (no primary source given); Chitra (2026)
USDe on BinanceBelow $0.66; mid-$0.60s; $0.65The Block; FTI Consulting; CoinGecko
Binance compensationAbout $283 million at first; $283 to $328 million in later accountBinance statement (Oct. 2025); CoinGecko
Binance fund drawdown$188 million, with balances of $1.23 to $1.04 billion or $1.22 to $1.03 billionBitcoin Sistemi, citing Binance data. Binance’s own post-mortem could not be opened for this lesson

Each row states a figure and its origin. Secondary sources may repeat one another, so several reports of one number are not independent confirmation.

The headline number measures notional closed, not money lost. A $19 billion figure is the size of positions force-closed. Trader losses, venue deficits and the sum moved by ADL are each different, and much smaller, quantities. On Hyperliquid in this window the realized deficit Chitra measures is about 5% of the $2.1 billion closed by ADL.

The case: competing accounts

Open question

What caused the crash and how much did Binance’s depeg matter?

Macro and leverage account. Binance attributes the event to a macro risk-off shock colliding with heavy leverage and thin liquidity. It says the market’s low came at 21:20 to 21:21, before severe depegging after 21:36, and a January 2026 press report of its review says about 75% of industry liquidations had already occurred by 21:36. FTI Consulting reports that top-of-book BTC depth fell by more than 90% on key venues, and that perpetual funding had risen to nearly 30% annualized by October 6. Venue-specific account. Critics, as CoinGecko summarizes, emphasize Binance’s use of internal order books to price collateral and a USDe yield-loop structure, including claims by OKX’s chief executive that are disputed. USDe’s decline was specific to Binance. Binance paid compensation to users whose collateral in USDe, BNSOL or wBETH was liquidated in that window. Ethena’s chief executive objected to calling it a USDe depeg at all, as The Block reported. What each side can and cannot show. Timing can show that the depeg did not start the first leg. It cannot by itself show that the depeg did not worsen later closures, or that its effects did not spread. Binance’s account comes from an interested party, and its account-level data are private. Critics’ accounts often lack data of their own.

Open question

Did Hyperliquid’s ADL take too much from winners?

Chitra’s paper proves a trilemma, namely that no ADL policy can simultaneously guarantee exchange solvency, revenue and fairness to traders. Applying the framework to Hyperliquid, its numbers changed materially. The December 2025 version said Hyperliquid’s queue overused ADL by about 28 times relative to the paper’s optimal policy, with roughly $653 million of unnecessary haircuts, and press stories repeated that. The February 2026 revision, with “corrections from public feedback” and a corrected empirical analysis, reports excess profit lost of $45.0 million to $51.7 million. It keeps $653.6 million only as the corresponding amount of positions closed, measured in equity dollars, and attributes the gap to winners’ overcollateralization of about 6.66 times. Jeff Yan of Hyperliquid responded that ADL does not transfer profit to HLP, treats HLP like any user, and that its closures netted users hundreds of millions by closing profitable shorts at favorable prices. He also accused the author of misrepresentation, and other supporters alleged bias over investments in competitors, as Protos reported. The author leads a risk-analytics firm, Gauntlet. Both sides have an interest. The most defensible statements are narrow: a queue can remove more profit than the needed amount, and any estimate depends on a counterfactual reconstruction of a market that did not happen. The paper also states that Binance appeared to use ADL far more heavily, but it also concedes that Binance’s account-level data are not public.

Other failures that day were operational. dYdX’s chain halted for about eight hours, which the protocol attributed to a misordered code process, and after it resumed, some liquidations were processed at stale oracle prices. Its governance considered paying up to $462,000 from the insurance fund, as Cointelegraph reported on October 27, 2025. Binance said some platform modules had brief failures after 21:18, while its core engines ran normally, and it attributed extreme prints on some altcoins partly to old limit orders that were filled in thin books. Chitra also notes that open interest fell by nearly 50% on Hyperliquid after the event as of late November 2025, while Lighter and Binance recovered. The author calls that a coarse measure and reports others attributing it to ADL policy. Falling open interest has several possible causes, so that attribution is unproven.

Reading such an event honestly

Good practice for an analyst is mostly bookkeeping. State the unit (notional or money lost), the time window, and whether a figure covers one venue or many. Prefer on-chain records where they exist, and describe the gap where they do not. Check whether reports are independent or copy one source. Separate a sequence from a cause. And note revisions: the most repeated number in this case’s fairness debate was changed by its own author months later.

Common mistakes

  • Reading $19 billion as $19 billion of losses. It measures closed notional.
  • Comparing venues’ liquidation totals as if reported alike. Coin Metrics documents different reporting conventions, and Binance’s feed shows only one per market per second.
  • Assuming ADL only hits risky traders. A hedged position is one leg of a position whose other leg the rule cannot see.
  • Treating an insurance fund’s total as available for every contract. Funds can be segmented by collateral or by contract.
  • Citing a working paper’s early figure after it has been revised.
  • Inferring cause from timing alone. A later event cannot have caused an earlier one, but an earlier event does not rule out later amplification.
  • Assuming a backstop’s profit proves the mechanism was rigged. Taking on risk at a discount normally earns a profit, and fairness needs a separate argument.

Problem set

  1. An example 20x long has $20,000 notional at a $100 entry, with a 2.5% maintenance fraction. Find the collateral, bankruptcy price and liquidation trigger. If the fill is at $93, find the shortfall.
    Show solution
    q = 200 and C = 20,000/20 = $1,000. Bankruptcy price = 100 − 1,000/200 = $95. Liquidation trigger = 95/0.975 = $97.44. Loss at $93 = 200 × 7 = $1,400, so equity = −$400, and the shortfall is $400.
  2. In the cascade model with s = 2% and β = $150 million per 1%, find the total move for λ = 0.002% and λ = 0.006% per $1 million. Then find the λ at which the model has no finite solution.
    Show solution
    With λ = 0.002%: βλ = 0.30, total = 2/0.70 = 2.86%. With λ = 0.006%: βλ = 0.90, total = 2/0.10 = 20%. The model diverges at βλ = 1, so λ = 1/150 = 0.00667% per $1 million.
  3. A shortfall of $5 million occurs. An insurance fund holds $2 million. Winners hold $30 million in unrealized profit. Find the residual, and the pro-rata haircut. What does a queue do instead?
    Show solution
    R = max(0, 5 − 2) = $3 million. Pro-rata haircut = 3/30 = 10% of each winner’s profit. A queue closes top-ranked positions in order until $3 million is covered, so some traders may lose much more than 10%, while others lose nothing.
  4. Compute Binance’s score for an example trader with profit $5,000, notional $100,000 and balance $20,000. What happens to the score if the trader adds $25,000 to the balance?
    Show solution
    PNL% = 5%. Effective leverage = 100,000 / 25,000 = 4, so score = 0.20. After adding funds, leverage = 100,000 / 50,000 = 2, so score = 0.10. The rule rewards a lower leverage with a later queue position, without changing profit.
  5. Explain why the following is a weak inference: “Hyperliquid’s open interest fell by nearly half after October 10, so its ADL policy drove traders away.” List two alternative explanations and one test.
    Show solution
    A fall in open interest has several causes. Alternatives include a general reduction in risk after a crash, a change in incentive programs or in market maker capacity at a venue, or a rotation to venues offering other terms. One test compares wallets that were closed by ADL with similar wallets that were not, and tracks each group’s later activity. That design still has selection problems, since ADL selects on leverage and profit.
  6. Report A says $19 billion was liquidated. Report B says the realized deficit in Hyperliquid’s ADL window was about $100 million. Report C says trader equity fell by $1.23 billion on Hyperliquid. Can all three be true? Say what each measures.
    Show solution
    Yes. A is market-wide notional closed in about 24 hours. C is the equity lost by Hyperliquid accounts, as one analytics firm reports it, which is lost money and not notional, and is a figure with no primary source given. B is the deficit in a 12-minute window on one venue, which is the part of losses that exceeded the failed accounts’ own equity. They differ in unit, scope and window, so they do not conflict.

References

  1. Chitra, T. (2025; v3 February 2026). Autodeleveraging: Impossibilities and Optimization. arXiv:2512.01112. Source
  2. Brunnermeier, M., and Pedersen, L. (2009). Market Liquidity and Funding Liquidity. Review of Financial Studies 22(6), 2201-2238. Source
  3. Schmeling, M., Schrimpf, A., and Todorov, K. (2023). Crypto carry. BIS Working Paper 1087. Source
  4. Hyperliquid. Liquidations and auto-deleveraging documentation (read November 2026). Source
  5. Binance. What is auto-deleveraging and how does it work? (read November 2026). Source
  6. Binance. Binance Futures Liquidation Protocols (read November 2026). Source
  7. dYdX. Liquidations and contract loss mechanism documentation (read November 2026). Source
  8. GMX. Liquidations and auto-deleveraging documentation (read November 2026). Source
  9. Binance statement on market volatility (October 2025), as republished by Deep News via Tiger Brokers. Source
  10. FTI Consulting (December 24, 2025). The Crypto Crash of October 2025: When Leverage Met Liquidity. Source
  11. CoinDesk Data (October 28, 2025). How CCIX Performed During Crypto’s Largest Liquidation Event. Source
  12. CoinGecko. What Is October 10th? Crypto’s 10/10 Mass Market Liquidation Event. Secondary compilation including Amberdata’s timeline. Source
  13. The Block (October 12, 2025). Binance pays $283 million in compensation following Friday’s depegs. Source
  14. Protos (December 10, 2025). Outdated algorithm caused $650M excess losses on Hyperliquid, report. Reports the figure later corrected, and Hyperliquid’s response. Source
  15. Bitcoinist. Bitcoin News: Here’s How Much Was Liquidated In The Crypto Market In 2025. Reports CoinGlass’s $30 to $40 billion estimate. Source
  16. Coin Metrics. Market Liquidations documentation. Source

Exercise

  1. Open the listing and compare the abstracts of v2 and v3. Write down which number changed, by how much, and what the author says caused the change.
  2. In v3, find the section on the October 10 dataset. Note the window, the deficit and the production haircut, and check that the overshoot equals the difference of the two published totals.
  3. Read the Hyperliquid ADL documentation, then build your own 5-trader example and rank it with the published index.
  4. Find two press reports of the $19 billion figure. Note each report’s source, time window and whether it says notional or losses.
  5. Write one paragraph stating what is known about this event and what remains uncertain, without assigning blame.
Open Chitra (2026), arXiv →

Check yourself

1. A position closes past its bankruptcy price. Who bears the shortfall first on a venue with a fund?

The fund or backstop generally absorbs a shortfall first. ADL reaches winning traders only if that fails.

2. In the linear cascade model, what is the effect of a fall in market depth?

Less depth means more price impact per dollar. Because the multiplier is 1/(1 − βλ), amplification rises sharply as βλ nears 1.

3. Under the Binance ranking, who sits first in the ADL queue?

The score multiplies profit percentage by effective leverage, so a small highly leveraged winner can sit ahead of a much larger low-leverage one.

4. Why can ADL harm a trader who was hedged?

The rule sees one position on one venue. Closing that position leaves the rest of the hedge open.

5. A headline reads “$19 billion lost.” What is the most accurate correction?

Liquidation totals generally measure the size of positions closed. Losses, deficits and ADL amounts are different quantities.

6. Binance says most liquidations happened before its depeg window began. What does that finding establish?

Order of events can rule out a cause for earlier effects. It cannot rule out amplification of later ones, and the data come from an interested party.

7. A 2025 paper’s headline figure changed sharply in its 2026 revision. What should an analyst do?

Revisions are part of the record. Quoting a corrected figure without noting the change, or the old one as settled, misleads readers.

8. Why does a backstop vault’s profit during a crash not itself prove unfair treatment?

Profit is consistent with compensation for risk. It still prompts questions on incentives, but it isn’t sufficient evidence by itself.
← 12. Leverage and margin 14. Funding rates and open interest →

Describes past on-chain activity and how our data is built. Not investment advice. Method details: data methodology. Live panels: tracked whale wallets, updated 2026-11-05 15:41 UTC.