Auto-Deleveraging (ADL) Explained: When the Exchange Closes Your Winner

Most traders meet auto-deleveraging as a rumor: something bad that happens to people who win too hard. The mechanism is more specific than that, and the specifics change what you should do about it.

ADL is the last step of a chain that starts with a liquidation that failed. Crypto liquidations explained covers that chain up to the point where the position is handed to the risk engine, and who pays when a liquidation goes bad covers how different venue designs absorb the resulting deficit. This article is about the last absorber in the line, which on a centralized perpetual venue is you.

Why does ADL exist at all?

A perpetual contract is a closed accounting system. Every dollar of unrealized profit on one side is a dollar of unrealized loss on the other, and the venue is only the bookkeeper. That symmetry is the whole reason ADL exists, and it is covered in more depth in perpetual futures explained.

Liquidation is supposed to keep the books balanced. The risk engine takes over a failing position and closes it before the trader's equity reaches zero, so the loser's own margin pays the winner. When the market gaps faster than the engine can fill, the position closes past the bankruptcy price and the loser's margin is no longer enough. The books now show a profit credited to the winning side that nobody has funded.

The insurance fund is the first absorber. When it cannot or will not cover the shortfall, the venue has exactly two options left: let the shortfall stand and become insolvent, or take it back from the traders it was credited to. ADL is the second option, executed as an algorithm rather than a negotiation. Note that "cannot cover" does not always mean "empty" — several venues trigger ADL at a threshold well above zero, a point established in the companion article on loss allocation.

What actually happens to your position when ADL fires

The event is fast and there is no margin call, no warning, and no way to opt out. On the venues that publish the sequence, it runs like this.

Two things are worth being precise about, because they are the two most common misunderstandings.

ADL does not touch profit you have already realized. Money that has settled into your balance is not reachable. What ADL takes is the difference between what your open position was marked at and what it was settled at. It closes the position early and on unfavorable terms; it does not reverse history.

ADL does not always close the whole position. Bybit's documentation works an example in which a trader holding 5,500 contracts has 5,000 closed and retains 500. Whether you are fully or partially deleveraged depends on your rank and on how much the engine needs.

At what price is a deleveraged position closed?

This is the single most under-documented part of ADL in the trading press, and the venues genuinely disagree. The choice determines whether ADL costs you money or merely costs you a position.

What each venue publishes about ADL triggers, settlement price, and queue ranking

Venue	Documented trigger	Price the deleveraged position closes at	Queue ranking basis
Binance	Only if the futures insurance funds are unable to accept a bankrupt position	The bankruptcy price of the liquidated order	PnL percentage multiplied by effective leverage
Bybit	An 8-hour drawdown of the contract's insurance fund past a published trigger line	The bankruptcy price defined by the insurance fund	Leveraged return, calculated differently for isolated and cross margin
OKX	When the security fund falls significantly below a preset threshold	The mark price at the time of matching in normal circumstances, the bankruptcy price when the fund is near depletion	Leverage PnL percentage
Hyperliquid	If an account value or isolated position value becomes negative	The previous mark price	Mark over entry price, multiplied by notional over account value

Read the third column as a spectrum of how much of the deficit is pushed onto you. Closing at the bankruptcy price hands you the entire per-contract gap between that price and the market. Closing at the mark price hands you none of it: you lose the position and the future profit it would have earned, but not a cent of what was already marked to your account. Hyperliquid's previous mark price sits in the same family as the mark price, one tick behind.

The practical consequence is that "does this venue run ADL" is the wrong question. Most major perpetual venues run something like it, and the ones that do not name it usually have an equivalent step in their waterfall. The question is what the venue's ADL price is, because that is what converts a forced exit into a realized loss.

How the ADL queue is ranked, worked through with numbers

Every venue that publishes a formula uses the same shape:

Ranking (profitable position) = return on the position x effective leverage
Ranking (losing position)     = return on the position / effective leverage

Binance publishes exactly that, including the sign-dependent branch, and Delta Exchange publishes the same two lines. Bybit expresses the leverage term as a margin rate and OKX as "leverage PnL%", but the direction is identical everywhere: more profitable and more leveraged goes to the front.

The losing branch matters more than it looks. Bybit states plainly that opposing positions at a loss may still be selected, and that profitable ones simply get priority. Dividing a negative return by leverage produces a small negative number for a highly levered loser and a large negative number for an unlevered one, so even among losers the queue prefers the levered. If the profitable side of the book is exhausted, the engine keeps walking.

Effective leverage is where readers usually go wrong. It is not the "10x" on your order ticket. Binance defines it as position notional divided by account balance plus unrealized profit, which means it falls as your profit accrues: a winning position funds itself, and the ratio of notional to equity shrinks. Two traders who both opened at 10x can sit at very different points in the queue an hour later.

A worked ranking on one contract

Take a hypothetical perpetual on a mid-cap token. All figures below are invented to show the ordering; they are not from any live market.

The token has fallen from the $100 area to a mark price of $40. Longs are being liquidated, so the profitable side is the shorts, and the shorts are the queue. Five accounts hold short positions.

Walk one row so the arithmetic is not a black box. Account E is short 800 units from $50. At a mark of $40 the unrealized profit is 800 x $10 = $8,000, and the position is now worth 800 x $40 = $32,000 of notional. E posted $1,000 of margin, so equity is $1,000 + $8,000 = $9,000. Return on notional is $8,000 / $32,000 = 25%. Effective leverage is $32,000 / $9,000 = 3.56x. Multiply: 0.889.

Sorted from the front of the queue:

Figure 1 rebuilds this table as something you can drag instead of only read. It opens on account E's numbers — $1,000 margin, $8,000 profit, $32,000 notional — so it starts at the front of the queue, matching the 0.889 above. Drag margin posted and unrealized profit: those are the only two inputs in the rank formula, and moving either one resorts the "You" row against the fixed five accounts. Now drag position notional on its own and watch the other two readouts instead: return on notional and effective leverage move in opposite directions, one falling as the other rises, while the ADL rank number does not move at all. That is the cancellation from the formula made visible — notional appears in both terms of rank and divides out, so position size by itself cannot change where you sit in the queue.

What this ranking actually tells you

The biggest position is last. Account C holds $200,000 of notional, five times account A's and six times account E's, and sits at the back of the queue. C is the second-largest winner in dollars and is the least exposed to ADL of anyone in the table. The folk belief that ADL hunts whales is wrong; it hunts concentration.

The biggest winner is not first either. Account A has $60,000 of unrealized profit, more than the other four combined, and ranks second. A's profit is so large that it has driven effective leverage below 1x — the position is now smaller than the equity backing it. Winning defuses you, slowly.

The front of the queue is the account whose equity is mostly paper. E posted $1,000 and is sitting on $8,000 of unrealized gain, so 89% of E's equity is profit that has not settled. That is the whole story, and on Binance's published definitions it is literally the whole formula: substitute their two definitions into each other and the notional cancels, leaving unrealized profit divided by equity. For a single-position isolated account that reduces to what fraction of your account is a promise rather than a balance. Bybit's version, which expresses the leverage term as a margin rate rather than against the account balance, does not collapse this neatly, but it ranks in the same direction.

What you should do differently: stop reading the queue as a size ranking and start reading it as a settlement ranking. The exposure that puts you at the front is unsettled profit sitting on a thin margin base, so the lever that moves you is the ratio between those two numbers, not how many contracts you hold and not how clever the trade was.

How the settlement gap is calculated, and why it is the smallest part of the event

Continue the same hypothetical. A bankrupt long has a bankruptcy price of $42, the market has gapped to $40, and the engine needs to close 1,000 units on a venue that settles ADL at the bankruptcy price.

It takes all 800 of account E's units, then 200 of account A's.

• Account E. 800 units close at $42 rather than the $40 mark. Realized profit is 800 x ($50 − $42) = $6,400, against the $8,000 those units were marked at. E is short $1,600, and E's position is gone.
• Account A. 200 units close at $42. Realized profit is 200 x ($100 − $42) = $11,600, against $12,000 at the mark. A is short $400, and A still holds 800 units with $48,000 of unrealized profit.

What this example actually tells you

The haircut is contracts matched multiplied by the price gap, and nothing else. Both accounts lost exactly $2 per unit, because the gap between the settlement price and the mark was $2. E lost $1,600 because 800 units were matched; A lost $400 because 200 were. Profit size does not enter the calculation at any point.

Which means the damage is wildly uneven in relative terms. E's $1,600 is 20% of E's total profit and 1.6 times the entire $1,000 of margin E posted. A's $400 is 0.67% of A's profit. The smaller of the two winners took four times the hit, in dollars, of the trader who had made seven and a half times as much.

And the queue reshuffles as it executes. After the partial close, A's realized $11,600 settles into the balance, so A's rank falls from 0.857 to 0.690 — the same arithmetic that made winning defuse A slowly now does it in one step. Bybit's documentation notes the same effect: a partially deleveraged trader keeps the same margin against a smaller position and may no longer be top-ranked.

And the numbers above are a thin-contract calibration, not a typical one. A $2 gap on a $40 mark is 5%, which is what a violent move through an illiquid book looks like. The gap is set by exactly one thing — how far price travelled past the bankruptcy price before the engine could act — so on a deep major perpetual it is a fraction of a percent, and on a thin altcoin contract it is the number above. The example is calibrated to make the arithmetic visible, not to represent the common case.

Putting the gap next to the move that produced it

The gap is the part of ADL you can compute in advance, which is the only reason to compute it. It is not a measure of what the episode cost you, and on a liquid contract it is usually the smallest number involved.

The gap is bounded. The move is not. ADL fires at the extreme of a dislocation, which is the same instant a discretionary trader would be trying to take profit into. If price reverts once liquidity returns — the common pattern after a forced cascade — the trader closed at $42 kept a fill near the best of the whole episode, and holding through the reversion would have given back a multiple of the gap. Binance's compensation for the October 2025 depeg was calculated as the difference between the market price at 00:00 UTC on 11 October and the liquidation price, which is a venue treating the cascade prints as transient rather than as fair value. A few tenths of a percent of settlement gap does not survive comparison with a move of that size in either direction.

And you were not the one short of liquidity. It is tempting to assume ADL catches you at a moment when you could not have exited anyway. Usually the opposite is true. If you are being deleveraged out of a short, the cascade around you is forced selling, and closing your short means buying — you are the natural counterparty to the flow, not competing with it. The claim that you would have suffered worse slippage closing it yourself does not hold for the side ADL actually targets.

And the continuation you supposedly missed is rarer than it sounds. The obvious objection is that ADL cuts you out of a move that keeps running. Follow the conditions and that case mostly closes itself. ADL on a deep contract requires the insurance fund to fail, which requires a cascade violent enough to gap through bankruptcy prices; a move that violent is a dislocation; dislocations revert when liquidity comes back. So the price you were filled at sits at an extreme that is typically not revisited for some time. For the continuation case to cost you, price has to resume through that extreme almost immediately — close to the negation of the condition that produced the ADL. And on the occasions the move genuinely does continue, it resumes from the reverted level, which means there is a window to put the position back on before it goes.

Which leaves a smaller residual than most writing on ADL implies. For a directional trader on a liquid contract, ADL is close to a non-event and frequently a favor: a fill near the best price of the episode, on a position you were going to take profit on anyway. The residual is real but modest — your resting orders are cancelled, you have to notice and decide, and re-entry is a deliberate act rather than an automatic one. That is an operational nuisance, not a risk to organize a strategy around.

The case where ADL is genuinely dangerous: a hedge that loses a leg

Everything above assumes a directional trader riding a winner. For that trader ADL is usually a non-event with a good fill attached. There is one structure where it is not, and it is the reason ADL deserves attention at all.

Run a hedged book — long spot against short perpetual, the cash-and-carry basis trade, or a long perp on one venue against a short perp on another — and the two legs are not equally exposed. In a crash the short perpetual is the profitable leg. That is precisely the leg ADL reaches for. It closes your hedge and leaves the other side untouched.

You are now long spot, unhedged, in the middle of a cascade, and you did not choose to be.

Why the damage is worse than it looks

You sized for a neutral book. The entire point of running delta-neutral is that it carries notional a directional trader would never take. When one leg is removed, the delta you inherit is not a normal position — it is basis-trade size, suddenly directional, at the worst moment available.

The ranking formula pushes you to the front of the queue. This is the part that makes it structural rather than unlucky. A hedged trader has every reason to post thin margin on the perp leg — why over-collateralize a position that is offset? But rank is profit multiplied by effective leverage, and in a crash that leg has large unrealized profit sitting on a small margin base. That is the exact profile the queue selects for. The hedge structure does not merely fail to protect you; it manufactures a top-of-queue ranking.

You had no stop on it. Nobody stops out a hedged leg, because the hedge was the risk control. So the position that replaces it arrives with no protection attached, and you find out in a fast market.

And on a cross-venue or spot-versus-perp hedge, nothing links the two. The venue running ADL sees one profitable short. It has no knowledge of, and no obligation to, the spot inventory or the offsetting position sitting somewhere else. Same-venue hedge mode is worth checking separately: confirm whether your venue ranks the two legs independently before you rely on holding both directions.

What to actually do about it

For how these strategies are constructed and what else can break them, see the Ethena basis trade and funding rate trading.

Reading the ADL indicator on your own position row

This is the part of ADL you can actually act on, and most traders have never looked at it.

Every major centralized venue exposes your live queue position on the position row itself. It is a small bar of segments, usually five, sometimes labelled "Delev." or "ADL". OKX describes a five-light signal in which five lights means the front of the queue and one light the back. Bybit publishes the mapping in a table: 20th percentile is five lights, 40th is four, and so on down to 100th percentile at one light. Binance shows an equivalent indicator on the position and also exposes it programmatically through an ADL quantile endpoint, so you can alert on it rather than watch it.

The lights are quintiles. Filling one segment means you have crossed into the next fifth of the ranked queue for that contract and that side.

What to do when the indicator fills up

The second step is the one worth slowing down on, because it is a genuine divergence rather than a documentation inconsistency. The two venues define the leverage term differently. Binance uses position notional over account equity, so halving the position halves the notional while the closed half's profit stays in the account: the ratio falls. Bybit uses the position's maintenance margin over the position's own equity, and both numerator and denominator shrink roughly in step: the ratio barely moves. Same mechanism, opposite advice, and the only way to know which applies is to read your venue's formula.

ADL, liquidation, socialized loss, and clawback are four different things

These get used interchangeably and they are not interchangeable. The distinction that matters is whose position or balance is touched, and whether the trader did anything to deserve it.

The ordering is the point. A liquidation ends a position that failed. ADL reduces a position that succeeded. Socialized loss does not touch positions at all — it shrinks what a balance is worth, which is why it can happen without anyone receiving a notification. Clawback is the only one of the four that reaches backwards in time, and it is the reason the queue-based mechanisms exist: venues moved to insurance funds and ADL precisely so that settled profit could be treated as final.

One terminology warning, because it causes real confusion: venue commentary often calls the ADL haircut itself a clawback. Throughout this article, clawback means the narrower historical mechanism in the last row of the table, the reversal of profit that had already settled. If you arrived here looking for "how the clawback is calculated," the worked haircut above is the calculation you want.

For which venue architecture leans on which of these, and why a lending market has no clawback available to it at all, see who pays when a liquidation goes bad.

When does ADL actually fire?

Rarely, and in clusters. The conditions that precede it are recognizable in advance.

The trigger is a fund drawdown, not a fund balance. Bybit publishes the clearest version. ADL for a contract triggers when the drawdown of that pair's insurance fund over the past 8 hours reaches the published trigger line as a share of the fund's highest balance in that period. Their worked example: a fund that peaked at 20,000 USDC over the window, a trigger line of 30% and a stop line of 25%, position margin of 1,000 USDC against 8,000 USDC of realized and unrealized losses. The drawdown ratio is (8,000 − 1,000) / 20,000 = 35%, which exceeds 30%, so ADL triggers and runs until the ratio falls back to 25% or below. A separate condition triggers ADL when the combined balance of several independently pooled contracts falls to or below zero. The threshold point — that ADL can start while a fund is still solvent — is established in the companion article; what Bybit adds is that the relevant quantity is the speed of the drawdown over a rolling window.

That reframes the market conditions to watch. You are not waiting for a fund to empty. You are watching for anything that drains one quickly:

• Open interest concentrated in few accounts on one side. The queue on the other side is short, so a single bankrupt position reaches further down it.
• A thin book relative to that open interest. The gap between the bankruptcy price and where the engine can actually fill is the per-contract haircut, and it is set by depth.
• Funding pinned at an extreme. One-sided positioning means the losing side is crowded and correlated, so liquidations arrive together rather than spread out.
• Spread dislocation or an oracle that disagrees with the book. If the mark price and the tradable price diverge, positions are liquidated at a price the engine cannot obtain.
• A collateral asset that can itself gap. A wrapped or yield-bearing collateral token that depegs turns solvent accounts insolvent without any move in the contract being traded.

For reading the aggregate liquidation data these conditions produce, see reading liquidations as a contrarian signal.

A dated event, cited only to what is documented

On 10 October 2025 the crypto derivatives market went through a broad forced-liquidation cascade. What can be sourced precisely is narrow, and it is worth separating from the commentary around it.

Binance's own announcement documents that USDE, BNSOL, and WBETH depegged on its platform between 21:36 and 22:16 UTC on 10 October 2025, and that it compensated users whose positions were liquidated as a result, calculated as the difference between the market price at 00:00 UTC on 11 October and their liquidation price. The venue is documenting a collateral-pricing failure and compensating the liquidations that followed from it, which is a different failure path from a contract simply moving faster than the engine could fill.

A November 2025 preprint by Tarun Chitra, Autodeleveraging: Impossibilities and Optimization, analyzes Hyperliquid's public dataset from the same day and reports that ADL "was used repeatedly to close $2.1 billion of positions in 12 minutes." The same preprint estimates, by comparing production ADL against benchmark allocations, that excess profit lost by profitable traders was between $45.0M and $51.7M, corresponding to roughly $653.6M of positions closed. Those are the paper's estimates, not venue disclosures, and they depend on its choice of benchmark.

What is safe to conclude from this and nothing more: ADL at scale is measured in minutes, not hours, and the pre-event condition that mattered was collateral pricing rather than an obviously overheated contract. Precise open interest and funding readings for the hours before are not publicly documented at the granularity that would let anyone reconstruct the setup, and articles that present them are reconstructing, not reporting.

Where the deficit goes when there is no ADL queue

Some venues route the shortfall somewhere else first. Hyperliquid's documentation describes backstop liquidations running through a liquidator vault that is "a component strategy of HLP," its community liquidity vault, so a deficit is a mark-down on depositors before it is ever a haircut on a winner. Several other decentralized perpetual venues use the same pooled-vault pattern under different names. ADL still exists on these designs — Hyperliquid documents it as a separate mechanism — but it sits one absorber further back, which is why it fires less often. The trade-off is that somebody chose to underwrite it, and the economics of being that somebody are covered in DEX perps and pooled liquidity. The venue-by-venue comparison of who absorbs the deficit lives in who pays when a liquidation goes bad.

How to reduce your ADL exposure before it matters

Everything above collapses into four decisions, and only one of them is made after the indicator lights up. Calibrate the effort to the structure: if you are directional on a deep contract, the first and third of these are worth a few minutes and the rest is optional. If you run a hedged book or trade thin contracts, all four are load-bearing.

Size against the contract's open interest, not against your account. Your queue depth is set by how many other traders hold your direction in that specific contract. A position that is 5% of the open interest in a major perpetual is buried; the same dollar amount in a small altcoin perpetual may be 40% of one side of the book, which means a single bankrupt counterparty reaches you on the first pass.

Treat your leverage tier as a queue setting, not just a liquidation setting. The tier you select determines both your maintenance margin and, through it, your effective leverage as profit accrues. How crypto leverage works covers the tier tables; the point specific to ADL is that the same choice that widens your liquidation distance also moves you back in the queue.

Convert unrealized profit into balance on a schedule. Concentration of unsettled profit is the actual ranking input. Taking partial profit at fixed intervals does two things: it moves money permanently out of ADL's reach, and on venues whose formula uses account equity it lowers your rank as a side effect. On venues whose formula does not, it still reduces the number of contracts that can be matched.

Choose the venue on its backstop, not its fee schedule. Insurance fund depth relative to the contract's open interest, whether the fund is pooled or isolated per contract, whether the venue publishes a drawdown trigger, and whether the ADL price is the bankruptcy price or the mark price are all knowable before you open an account. Bybit publishes its insurance pool balances through its API and on its insurance history page. That is a due-diligence item with the same weight as maker fees.

Conclusion

Auto-deleveraging is the mechanism that keeps a perpetual venue solvent when a liquidation fails and the insurance fund will not or cannot pay. It works by closing positions on the winning side of the same contract, ranked by return multiplied by effective leverage, and settling them at a price the venue publishes in advance.

Three things follow from the mechanics, and they are the three things worth carrying away.

The queue is ranked on concentration, not size. The account at the front is the one whose equity consists mostly of unrealized profit sitting on a thin margin base, which is why the largest position in the worked example above was last in line while an account holding $8,000 of profit outranked one holding $60,000. The lever that moves you is the ratio between profit and margin, not the number of contracts.

The settlement gap is the small part, and treating it as the story gets ADL backwards. It is the quantity matched multiplied by the distance between the settlement price and the mark, it is bounded by how far price ran past the bankruptcy price, and on a deep contract it is a fraction of a percent. Set against the move that produced it — the reason you had a large winner in the first place — it rarely decides anything. It is still worth knowing, because it is the one number available to you in advance and because a venue settling at the mark price is running a cheaper mechanism than one settling at the bankruptcy price. It is not worth fearing.

For a directional trader on a liquid contract, it is close to a non-event. The conditions that produce ADL are the conditions that produce a reversion, so the fill lands near the extreme of the episode and that price is usually not revisited soon. The move continuing straight through it is the rare case, and when it does continue it resumes from the reverted level, which leaves room to re-enter. What is left is an operational nuisance — cancelled orders, a decision to make, a position to put back on — not a reason to trade smaller or avoid a venue.

The exception is structural, and it is the whole reason to read about this mechanism. If you run a hedged book, ADL removes the profitable leg and leaves you holding the other one, unhedged, in exactly the conditions the hedge was built for. The thin margin that makes the structure capital-efficient is the same thing that pushes that leg to the front of the queue, so the strategy manufactures its own selection. That risk does not show up in a backtest of the spread, it cannot be stopped out of, and it is the version worth designing around. Everything else about ADL is worth understanding once and then largely forgetting.

And the indicator on your position row is real information that almost nobody uses. It is a quintile rank inside one contract and one side, it updates live, and on most venues it can be read through the API. Adding margin lowers it everywhere. Whether closing part of your position lowers it depends on your venue's formula, and that is a question you can only settle by reading the venue's own documentation. For a mechanism that reaches into a profitable trade without asking, it is worth twenty minutes before you need it rather than twenty seconds after.

Frequently asked questions

Sources and further reading

Venue documentation opened and verified on 6 August 2026:

• Binance — What Is Auto-Deleveraging (ADL) and How Does It Work?
• Bybit — Auto-Deleveraging (ADL) Mechanism
• OKX — Introduction to Auto-deleveraging (ADL)
• Hyperliquid — Auto-deleveraging
• Hyperliquid — Liquidations and the liquidator vault
• Delta Exchange — Auto Deleveraging
• dYdX — Liquidations and the insurance fund
• Binance — Position ADL Quantile Estimation endpoint
• Binance — Resolution of USDE, BNSOL, and WBETH Price Depeg and Risk Control Enhancements

Research:

• Tarun Chitra — Autodeleveraging: Impossibilities and Optimization (arXiv preprint, 30 November 2025)

Related CoinBeaver articles:

• Crypto liquidations explained
• Who pays when a liquidation goes bad
• How crypto leverage works
• Perpetual futures explained
• Cross versus isolated margin
• DEX perps and pooled liquidity
• Reading liquidations as a contrarian signal

This article is educational and is not financial or trading advice. Every price, position size, margin figure, and ranking score in the worked examples is hypothetical and constructed to demonstrate the mechanism; none of it is live market data or a quote from any venue. ADL rules, insurance fund parameters, settlement prices, and ranking formulas differ by venue and change without notice. Verify the current documentation on the exchange you trade before relying on any rule described here.