Reading Liquidations as a Contrarian Signal: When the Forced Seller Is Finished

What a liquidation feed actually reports

This article is an application of one signal family to one signal. If you have not read how to read crypto market signals without fooling yourself, start there: liquidation data belongs to the structural family, and the characteristic failure of that family is sign ambiguity, where the same reading is consistent with opposite outcomes. Everything below is an attempt to remove that ambiguity for one specific reading.

The mechanics of a liquidation itself are covered in how crypto liquidations happen: a venue's risk engine takes control of a position whose margin has fallen below the maintenance threshold, prices the trigger off the mark rather than the last trade, and closes the position on the book. That article owns the formulas, the venue penalties, insurance funds, and auto-deleveraging. This one starts at the moment the engine submits an order, because that order is the thing the public data records.

The reporting unit is an order, not a position and not a trader. Binance's futures liquidation stream publishes an order object: symbol, side, order type, time in force, original quantity, price, average price, order status, last filled quantity, and accumulated filled quantity. The documented example is a LIMIT order with immediate-or-cancel time in force. That tells you three things at once. The engine is submitting a real order to the book rather than settling internally. The order carries a deadline rather than a target price. And a single trader's position can produce several of these objects, while a single object can be far larger than any individual account, because what you are counting is engine output, not people. Nor is it volume in the ordinary sense: the closing trade also appears in the venue's normal trade tape, so liquidation notional is a tagged subset of volume rather than an addition to it.

Venues do not agree on what the fields mean. Bybit's current liquidation stream publishes a different object entirely: position side, executed size, and bankruptcy price, pushed on a 500ms cadence. Bybit documents its side field as the position side and spells out that a Buy value means a long position was liquidated. Binance documents its side field simply as Side, on an object that also carries order type and time in force, so it is describing the order the engine submitted. Same letter, different referent.

A long flush in this data is a burst of forced closes of long positions, which means the engine is selling. A short squeeze is the mirror: forced closes of short positions, engine buying. That definition is about which side was liquidated, not about which direction price moved afterwards, and conflating the two is where most of the folk wisdom in this area starts.

Why aggregate liquidation totals are understated

Start with the conclusion, because it reframes every threshold rule in this article. Absolute liquidation dollar totals are not comparable across venues or across time. Relative spikes within one venue's own series, under one unchanged publication rule, still carry information.

The reason is that the public feeds are deliberately throttled. Binance's derivatives change log records that on 27 April 2021 the liquidation order streams stopped pushing realtime order data and began pushing snapshot data at a maximum frequency of one order push per second, for both USD-margined and coin-margined futures. The stream documentation states the same rule from the other side: for each symbol, only one liquidation order inside each 1000ms window is pushed as the snapshot, and if no liquidation happens in that interval nothing is pushed at all.

Bybit's history is the same problem with a different date. Until February 2025 its public liquidation topic pushed one liquidation per second. On 20 February 2025 Bybit's changelog introduced a new topic covering the full set of liquidations occurring on the exchange and noted that the older, one-per-second topic could be discarded.

A worked understatement

The numbers here are constructed to show the shape of the distortion, not to describe any dated event.

Suppose a cascade on one symbol runs for 40 seconds. The risk engine submits roughly 600 separate immediate-or-cancel orders during that span, totaling $95 million of closed notional. Under a one-per-second snapshot rule, the public stream can emit at most 40 messages, each carrying one of those orders. If the sampled orders average $220,000, the feed reports:

40 messages x $220,000 = $8,800,000 reported
$8,800,000 / $95,000,000 = 9.3% of the actual forced flow

What this example actually tells you

The number of reported events is capped by duration, not by size. Forty seconds of cascade can never produce more than forty published events on that venue, no matter how many orders the engine actually sent. Size still enters through whichever order gets sampled, but the count does not, so two cascades of very different magnitude that happen to last a similar number of seconds can report similar totals. That also makes the understatement variable rather than constant. If the shortfall were a fixed fraction you could scale it away; instead it depends on how many orders were compressed into each second, which is exactly the variable you were trying to measure. A brief, violent cascade is understated far more than a slow one.

Comparison breaks in both directions, and for different reasons. Across time, an aggregate series drawing on Bybit's public stream saw its coverage of that venue change in February 2025 for reasons that had nothing to do with price or positioning, so unless the aggregator restated its history a chart spanning that date contains a discontinuity that looks like a change in market behavior. Across venues, composition is a choice: Coinglass's aggregated liquidation history endpoint takes an exchange list as a parameter, so two people quoting the total for the same hour can both be reading their screen correctly and disagree by a wide margin.

The actionable version. Never write a threshold in dollars. Write it as a percentile of one venue's own trailing distribution for one symbol, re-baseline whenever that venue changes its publication rule, and treat any cross-venue aggregate as a headline rather than as an input.

Liquidation heatmaps are models, not observations

This is the most common category error in this space, and it is worth stating flatly: a liquidation heatmap is a model output. It is not a record of liquidations that happened.

Coinglass describes its heatmap as a tool for predicting potential large-scale liquidation points, calculated from market data including trading volume and leverage usage. Its liquidation heatmap API returns price levels alongside a leverage value, which is the tell: the map is generated by taking open interest, assuming a distribution of leverage across it, applying the standard maintenance-margin arithmetic, and plotting where positions built under those assumptions would be force-closed. Coinglass states that the displayed levels are a relative intensity indicator and that actual liquidation amounts may be lower than the levels shown.

Three assumptions sit inside every such map.

Leverage is assumed, not observed. No venue publishes per-position leverage. The distribution is inferred, and the inference is the single largest source of error in where the bands land.

Positions are treated as static. A trader who adds margin, reduces size, or is stopped out manually removes a level from the map without any liquidation ever occurring. The map has no way to see that happen, so it goes on showing an overhang that no longer exists.

Cross-margin positions do not have a single liquidation price. For an account whose collateral is shared across positions, the level depends on the rest of the account, which the model cannot see. Every such position is placed on the map at a price that is at best a guess.

What a heatmap is legitimately for, and what it is not

It describes geography, not destination. The map is a statement about the shape of latent supply and demand: if price reached this level, forced flow of roughly this relative intensity would appear there. That makes it a stop-placement input above all. If your stop sits inside a dense estimated band, expect a worse fill than your slippage model assumes, and consider whether the band is a reason to place the stop wider and size smaller.

It is not evidence that price will travel to the band. The claim that liquidity acts as a magnet is an assertion about intent that the model contains no evidence for. The heatmap has no time axis and produces no probability. It answers "what is there" and is silent on "will price go".

It is public, which makes it reflexive. Every participant can see the same bands. That cuts both ways: the levels are visible to anyone who would like to push price into them and equally visible to anyone positioned to defend them. A widely published map is not a private edge.

Why a flush can stop and reverse a move

Here is the spine of the whole idea, and it is not the one usually given.

The contrarian read is not that the crowd was wrong. It is that the forced seller is finished.

That distinction matters because it tells you exactly what to check. Four things are happening inside a flush.

The flow is price-insensitive. The risk engine is not seeking a good fill. Binance's documented liquidation orders carry immediate-or-cancel time in force, which is the mechanical expression of "close this now, at whatever the book gives." A seller operating under a deadline is not expressing a view about value. That matters after the event as much as during it, because it means the selling had no price target to return to.

The overhang is removed. Before the flush, every leveraged long sitting inside that price band was a future forced seller at a known level. Their presence was a claim on future supply. After the flush, they no longer exist as positions. The supply that was hanging over the level has been converted into trades that already happened.

Open interest is destroyed rather than transferred. This is the observable version of the previous point, and it is the single most important confirmation variable in this article. Open interest in crypto covers the price and open-interest combinations and, critically, why you must read open interest in coin or contract units rather than USD, since a USD-denominated series falls simply because price fell.

The buyer on the other side chose to be there. Someone absorbed those immediate-or-cancel orders at that price without being compelled to. That is a discretionary bid, and its presence is visible in the book as absorption rather than as a wall; how to read a crypto order book explains why resting size is a claim and filled size is a commitment.

What this actually tells you

The signal is about the seller's remaining capacity, not about the crowd's opinion. Framings like "everyone was long, so buy the flush" get the mechanism backwards and will lead you straight into the failure case, because a market where the leverage is immediately rebuilt still has everyone long and no exhausted seller. The failure mode is written into the mechanism itself: if forced sellers were only a small part of total supply, finishing them changes very little, and a stop-and-reverse only follows when the forced seller was the marginal seller.

The actionable version. Before you treat a flush as a setup, write down who you believe the marginal seller is and what observation would prove they are still active. If your answer is "leveraged longs clustered at this level," then the open-interest trajectory is your test and you already know what disproof looks like. If you cannot name the seller, you do not have a setup, you have a headline.

The conditions that have to hold before a flush is a setup

These are conditions to check, not a prediction. Each one is a way the mechanism above can be absent.

An extended prior trend in the direction being flushed. The overhang has to have been built before it can be cleared. A flush inside a two-week range clears very little, because very little accumulated.

Funding already reset or flipped. Funding rates are the running cost of the crowded side. If funding is still strongly positive after a long flush, the crowd was not removed, it was trimmed. The reset is the price signal that the paying side has been reduced.

Open interest sharply lower in coin terms, and staying lower. The word doing the work is staying. A single reading taken at the low tells you positions closed. It does not tell you whether new ones opened at the same level ten minutes later. This condition can only be evaluated with a second observation, hours later.

Concentration in time. One violent burst is one clearing event. The same notional distributed across a session is a sequence of sellers who are still arriving, which is a completely different situation wearing the same daily total.

A flush into structure that existed before the flush. The level has to be identifiable on the chart before the event, using the ordinary tools in how to read a candlestick chart. If you can only see the level after the low is printed, you have drawn the structure around the outcome.

When the contrarian read fails, and how to tell in real time

Three quite different events produce a liquidation spike, and on the feed alone they are close to indistinguishable.

Three observations separate them while the event is still live.

Open-interest trajectory. Destroyed and staying down is the confirmation. Rebuilding between bounces is the disconfirmation, and it is the specific reason the cluster-retest case eats contrarian buyers: the level looks cleared each time, and each time new leverage has taken the vacated seats, so the same price band is loaded again.

Funding trajectory. Reset or flipped supports the read. Still extreme in the original direction contradicts it directly, because the crowd whose exit you were counting on has not exited.

Which market led. If the perpetual traded below the spot index through the low and funding printed negative, the selling was derivative-led and the forced seller is plausibly the marginal seller. If spot printed the low first while the perpetual held at or above the index, real supply is leading and the forced flow is downstream of it. In the second case, finishing the forced seller changes nothing, because the seller who matters is in the spot market and is under no deadline. That gap between the perpetual and the index is the basis, and its sign through the low is the cheapest available test of who was driving.

One more discipline. Extreme flushes are rare events, which means a screen that fires often is not measuring the same thing you designed it to measure. The base-rate arithmetic in how to read crypto market signals applies here without modification: a rule that catches most reversals can still be wrong most of the times it fires, and the false-positive rate is the number nobody publishes.

Two flushes with the same headline number

Both cases below are constructed. The figures are chosen to isolate the variables that separate the two outcomes and are not measurements of any dated event. The real charts circulated as examples of this setup were almost all selected after their outcomes were known, which is precisely why constructed cases are more informative here: a case built to isolate a variable can hold everything else equal, and a screenshot cannot.

The design of the comparison is the point: the two cases report an identical headline liquidation figure.

Walking through the two cases

In Case A, the sequence is internally consistent. A long build ran for weeks, funding paid for it the whole time, and then 78% of an extreme hour arrived inside nine minutes. That compression is what a risk engine working through a cluster looks like: it is not making decisions, it is emptying a queue. Open interest falls 22% and is still down 20% six hours later, which means the closed positions were not replaced. Funding flips negative, which means the paying side is now the shorts. The perpetual trading below the index through the low says the selling originated in the derivative, not in the spot market. Every one of those observations is a different way of saying the same thing: the seller who was forced is gone and nobody took their place.

In Case B, the identical headline number describes something else entirely. The largest nine-minute window holds only 21% of the hour, so the flow arrived as a sequence rather than as a queue being emptied. Open interest falls 8% and, by the six-hour reading in the table above, is back within 2% of its starting level — a recovery already well underway inside four hours, which is the observable signature of new leveraged longs stepping into the seats the liquidated ones vacated. Funding never resets. And the perpetual holds at the index while spot prints the low first, which places the marginal seller in the spot market, where nothing about a liquidation cascade constrains their behavior.

Figure 1 and Figure 2 draw the two cases on one shared price scale. The flush candle is deliberately identical in both — same open, same high, same low, same close — because that is the honest version of the problem.

Read the two figures in three steps.

1. Compare the third candle in each. Open 101, high 101.5, low 88, close 95, in both figures. Same range, same long lower wick, same close well off the low. Whatever you believe a flush candle tells you, it tells you the same thing in both cases.
2. Compare the two candles after it. Case A recovers to a close of 103, above where the flush began. Case B closes at 83, a further 13% below the flush close. These are the outcomes the contrarian trade is betting between, and nothing in the candle that preceded them distinguished one from the other.
3. Now recall what did. Open interest was down 22% and staying down in Case A, down 8% and fully rebuilt in Case B, and that separation was readable roughly four hours in — while price in both figures was still inside the flush candle's range. The evidence was in the positioning data, not in the chart.

This is also why the constructed-case framing matters. A screenshot of a real flush shows you one path and cannot show you the path that did not happen, so it can never demonstrate that two identical-looking events diverged. Holding the candle fixed and varying only what follows is the one way to make that point visible.

What these two cases actually tell you

The one row that is identical is the headline liquidation number. Every row that distinguishes the two cases is a row about something else: timing, open-interest persistence, funding, and basis. The liquidation print is the reason to look. It is not the evidence.

Open-interest persistence is the highest-value single observation. Case B is distinguishable from Case A within four hours by open interest alone, without reference to price. That is unusual and worth exploiting: it gives you a test that does not require you to have been right about direction.

The actionable version. Compute two numbers on every flush: the share of the window's notional inside its densest few minutes, and the ratio of current open interest to its pre-event level at a fixed check-in later. Both are available before you would normally be forced to decide, and if your process requires an entry inside the first ten minutes neither is available to you and you are trading the headline. Record them for every event whether or not you trade it. Those two series are the closest thing to a personal base rate you can build from data you actually have.

What the evidence actually supports

Be straight about this, because the rest of the internet will not be.

There is no robust public body of evidence establishing that this setup is tradeable. What exists is a well-understood mechanism, which this article has described, plus a very large amount of anecdote and hindsight-selected charts. Those are different kinds of thing. A mechanism tells you an effect is possible and gives you the conditions under which it should be present. It does not tell you how often those conditions produce the outcome, what the payoff distribution looks like, or whether the edge survives costs.

The closest thing to a systematic public study we could verify is a July 2026 preprint that examines seven major Bitcoin liquidation cascades from 2022 to 2025 using minute-level price data and five-minute open-interest and order-flow data. Its findings are worth reading carefully, and so are its limits.

What it finds. No state variable carries an early-warning signature across all seven events. The author proposes a two-type structure: events with gradual endogenous buildup, where price shows a critical-slowing-down signature, and sudden exogenous news shocks, where that signature is absent. A leverage-based pattern that fit the October 2025 event inverted completely when tested out of sample on August 2024. There is one positive result, and it is worth quoting accurately because it cuts against the folk reading rather than for it: a compression in taker order-flow variance precedes all six usable cascades and survives placebo testing at the population level, but the author describes it as too weak to warn on any individual event. The paper explicitly disclaims that any particular trading rule would have helped.

What it does not establish. It is a preprint rather than peer-reviewed work, the sample is seven events, and it studies the run-up to a cascade rather than the aftermath. It therefore does not test the contrarian claim at all. What it does support is the premise of this article, that cascades are not one kind of event, and it supplies a concrete instance of the risk that governs everything here: a pattern that fit one cascade inverted on another.

Why the examples you see are systematically misleading

Selection is applied after the outcome. A flush that reversed is a screenshot with arrows on it. A flush that continued is not posted at all. Every example in your feed passed a filter that could only be applied once the outcome was known, so the population you are learning from has had its failures removed by construction.

The accounts posting them have survived a similar filter. Traders whose version of this setup did not work stopped posting, so the visible distribution of practitioners is the surviving tail rather than the population. And the setup is unusually easy to narrate after the fact: every large down move contains liquidations, so every reversal can be told as a flush that cleared the overhang and every continuation as a flush that was not big enough. A story that fits both outcomes is not a signal.

How to build a liquidation screen you can repeat

This is a method for deciding what to look at, not a tool for predicting price. The first three steps produce candidates. The remaining four are what stop a candidate from becoming a loss you cannot explain.

The time stop in step six deserves emphasis, because it is the step traders skip. The premise of this trade is that a specific seller finished at a specific moment. That premise has a natural expiry. If it has not produced a move within the window in which the absence of that seller should have mattered, the premise has stopped being the thing driving price, whatever the chart looks like.

Conclusion

Liquidation data is the most-watched positioning feed in crypto and the most systematically misread, and the misreading starts with what the data physically is. It is a stream of forced-close orders that venues chose to publish under rules they wrote for their own reasons and have changed more than once. Binance caps its public feed at one order per symbol per second and has done since April 2021; Bybit only began publishing its full set in February 2025. Any absolute dollar figure built on top of that is a lower bound of unknown and event-specific tightness, which makes it useful as a rank against one venue's own history and useless as a quantity.

The second misreading is the confusion between the recorded feed and the heatmap. The feed reports things that happened. The heatmap is a model that takes open interest, assumes a leverage distribution nobody outside the venue can observe, and plots where positions built under those assumptions would be closed. It answers where forced flow would appear if price went there. It contains no information about whether price goes there, and treating a dense band as a destination is the single most expensive error available in this dataset.

The mechanism behind the contrarian read is real and it is narrow. A risk engine sells with a deadline rather than a price target, so its flow is price-insensitive; the leveraged longs sitting above a cluster were a claim on future supply, and once they are closed that claim is extinguished; the buyer who absorbed the flow chose to be there. The correct summary is not that the crowd was wrong. It is that the forced seller is finished. State it that way and the test writes itself: open interest that falls and stays down means the seats were not refilled, and open interest that recovers means they were and the next flush is already being loaded.

What the evidence supports is thinner than the confidence with which this setup is usually presented. There is a well-understood mechanism, there is a preprint finding that cascades are heterogeneous and that patterns fitting one event invert on another, and there is an enormous volume of charts selected after their outcomes were known. That is not a measured edge, and you should not size it as one. Build the screen so that it is repeatable, express every threshold as a percentile of one venue's own history rather than a dollar amount, scale events by open interest, require the concentration and the confirmations, write the invalidation and the time stop before you enter, and log the firings you skipped. The discipline is not decoration around the signal. Given how little the public evidence actually establishes, the discipline is the only part of this you can rely on.

Frequently asked questions

Sources and further reading

Primary sources opened and verified for this article:

• Binance — Derivatives API change log, including the 2021-04-27 liquidation stream snapshot change
• Binance — All Market Liquidation Order Streams, coin-margined futures
• Bybit — V5 API changelog, including the 2025-02-20 all-liquidation topic
• Bybit — All Liquidation websocket stream reference
• Coinglass — How to use liquidation heatmaps to assist trading
• Coinglass — Liquidation heatmap API reference
• Coinglass — Aggregated liquidation history API reference
• Garcia Seuma — Early-warning signals across seven crypto-perpetual liquidation cascades, arXiv preprint 2607.27070

Related CoinBeaver articles:

• How crypto liquidations happen
• How to read crypto market signals without fooling yourself
• Open interest in crypto
• How funding rates work
• How to read a crypto order book

This article is educational and is not financial advice. The understatement calculation and both flush cases are constructed illustrations, chosen to isolate specific variables; they are not measurements of any dated market event, and no figure in them should be read as a historical record. Exchange stream rules, aggregator venue coverage, and heatmap methodologies change without notice, so verify the current publication rules of any data source before building a threshold on top of it. The cited preprint has not been peer-reviewed and studies the run-up to seven cascades rather than their aftermath.