Liquidation Heatmaps: How They Work, How to Read Them, What They Get Wrong

What does a liquidation heatmap actually show?

A liquidation heatmap plots price on one axis and, usually, time on the other. Color marks how much leveraged exposure is estimated to be force-closed at each price. Bright bands mean more. A liquidation map or levels chart is the same idea with the time axis removed: bars per price, at one moment. If leverage or margin are new to you, the Learn course covers the groundwork, and our derivatives overview covers the venues.

The chart invites a reading it can't support on its own. It looks like an order book, and it is not one. No order sits at a bright band. A band says that if price reached that level, some positions would probably be closed by force. Every question worth asking follows from what produced the color, so that comes first.

How do vendors build liquidation heatmaps? What their own docs say

We read each vendor's documentation rather than the many third-party guides that repeat each other. The shared point: centralized exchanges don't publish account-level liquidation prices, so a cross-exchange map has to estimate them. No vendor we checked publishes a full formula or a weighting scheme.

That last row needs a plain statement of where other tools are better. For a map of every Hyperliquid position, HyperDash and Kiyotaka describe wider coverage than ours. For exposure across Binance, Bybit and OKX, a modeled heatmap is the only kind that exists, because no one outside those exchanges can see their accounts. A model isn't a flaw when the real data is locked away. It only becomes a problem when a reader takes it for a measurement.

Why can Hyperliquid publish a real liquidation price?

Hyperliquid's public info endpoint returns any address's positions, including a liquidationPx field. We confirmed this against the live API on November 5, 2026. Its liquidation docs also give the formula, which we restate here:

liq_price = price − side × margin_available / position_size / (1 − l × side)

Here side is 1 for a long and −1 for a short, and l is 1 divided by the maintenance leverage. For an isolated position, margin_available is that position's margin minus its maintenance requirement. For a cross position, it is the account value minus the maintenance requirement. That one difference explains most of what follows.

The leverage setting is 20x in every row, and the distance runs from under 4% to no liquidation price at all. A setting chosen on the screen doesn't fix where a cross position gets closed. Hyperliquid's margining docs say cross is the default and that leverage is checked when a position opens. After that, the account's collateral does the work.

The last two rows explain something we saw in the live data. Hyperliquid's liquidation page doesn't address this case, so this is our reading of the formula: a long can have no liquidation price, because even a price of zero wouldn't exhaust the collateral, and a short always has one, because price has no ceiling. Among our tracked whales, every BTC and ETH position with no published price was a long: 62 in BTC ($69.4M) and 34 in ETH ($30.6M). Those positions can't appear on any price-based map, however real they are.

Modeled vs real: how far does the leverage formula land from the published price?

This is the test we can run and vendors can't, since it needs the true answer. We took every tracked whale position in BTC and ETH on Hyperliquid and computed the simple formula from that position's real entry price and real stated leverage. Then we compared that with Hyperliquid's published liquidation price.

This is not a test of any vendor's product. No vendor's model is this formula applied to these accounts. Vendors don't know a position's entry or leverage, and they spread estimates across many assumed settings, which may average out some of the error. What this measures is how much information the formula needs, by giving it more inputs than an outside model could have.

The result splits cleanly by margin type. Where the formula has its required input, it comes close: isolated positions land within a few points. Where the true input is the whole account, it misses by tens of points. Cross positions are also most of the money here, about 88% of BTC and 95% of ETH among positions with a published price.

Three more ways to see the same gap:

• Positions the formula says are already gone. For 30 BTC positions ($130.9M) and 49 ETH positions ($303.4M), the formula's liquidation price was already past the current price, yet the position was open. Every one was cross margin.
• How crowded the area near price looks. The formula put $761M of BTC exposure within 10% of the current price. The published prices put $89M there. For ETH, the figures were $733M and $30M. Those totals exclude the positions in the bullet above.
• The 25x, 50x and 100x grid. Hyblock names these as high-leverage settings, and the simple formula puts them 4%, 2% and 1% from entry. Only 3.7% of BTC whale size and 3.4% of ETH whale size had a real liquidation price within 5% of entry. Nothing in BTC was within 2.5%. Hyperliquid's public market data also caps BTC at 40x and ETH at 25x, so 50x and 100x don't exist on this venue. Those settings reflect venues with much higher limits: Kingfisher's leverage views run up to 125x.

Where do tracked whales' published liquidation prices sit around the current price?

Here is the real map for our tracked whales, as a snapshot. Positions and marks were read at 12:33 UTC on November 5, 2026, with BTC at $80,080 and ETH at $2,409.90. Red bars are long positions, which get closed by forced selling as price falls. Green bars are short positions, closed by forced buying as price rises. The same data refreshes on our BTC and ETH pages.

Most of this size sits far from the current price. The nearest BTC whale liquidation price was 3.1% away, and the nearest for ETH was 4.4%. By size, the median position was 44.2% from price in BTC and 42.7% in ETH. The high-leverage settings above imply levels within a few percent of entry, and our whale data doesn't look like that. Two explanations are possible and we can't separate them: smaller traders may crowd in near price, or models may place levels closer than they really are.

What this map doesn't say: it says nothing about which way price goes, and it isn't a list of places price is heading. Larger bands below than above, or the reverse, describe the current mix of positions and nothing more.

Are liquidation clusters a crowd, or one wallet?

A bright band reads as a crowd. On real data we can count who is behind it. We split the area within 30% of the current price into bands 1% of price wide, and looked at the five largest in each coin.

One wallet held more than half of every top-five BTC band and four of the top five in ETH. Across all occupied bands, 25 of 36 in BTC and 21 of 32 in ETH came from a single wallet. These figures are for whales only, so the bands would hold more wallets on a full-market map. Even so, the point stands, and HyperDash makes the same one in its own guide: a large bar can belong to one account and vanish in one transaction. That trader can add margin, reduce size or close at any time. A modeled heatmap has no way to show this, because it has no wallets to count.

How quickly does a real liquidation map change?

Modeled heatmaps often look fixed, with bands that stay until price touches them. Real levels behave differently, especially on cross margin. To measure this, we read the same positions a second time, 65 minutes later, at 13:38 UTC on November 5. BTC's mark fell 0.47% over that span, to $79,702, and ETH's fell 0.85%, to $2,389.30. It was a quiet hour.

Every level moved, even though most traders did nothing: only 27 of 275 BTC positions present at both reads changed size. That follows from the formula. A cross position's liquidation price depends on account value, so it moves as the account's open profit moves. Isolated positions barely shifted, as the formula says they should.

One band shows how misleading a chart can be. The $53.7M BTC band at $62,462–$63,263 in the table above read $0 at the second read. A quick look would suggest someone had closed or defended the position. In fact the position size hadn't changed at all. As price fell, the account's value slipped, and the liquidation price rose about $477, from roughly $63,164 to $63,641, past the edge of the band. The trader did nothing, and the bar simply landed in the next row.

How much selling or buying would a band really cause?

Band size isn't the same as an order that hits the market in one go. Hyperliquid's liquidation docs describe several details that a bar on a chart hides:

• Partial closes. For a position above $100K, only 20% goes to the book as a market order at first, followed by a 30-second cooldown during which further liquidation orders can cover the whole position.
• A backstop. If equity falls below two-thirds of the maintenance margin without a successful book close, the position can pass to a liquidator vault instead of the book.
• Mark price, not last trade. Liquidations use a mark price combining outside exchange prices with Hyperliquid's own book, so a brief spike on a chart doesn't necessarily trigger anything.
• Margin tiers. Maintenance needs rise for very large positions. In the public margin table we read on November 5, BTC's maximum leverage drops from 40x to 20x above $150M.

For scale, we read Hyperliquid's public BTC order book at 12:44 UTC on November 5, grouped into $100 price levels. It held about $139.6M of bids and $125.4M of asks within 1% of the mid price, and $209.8M and $180.6M within 2%. A 20% slice of the $64.8M short band above is about $13M, roughly a tenth of that ask depth. Take this only as a sense of scale. It's one reading of a book that changes constantly, liquidity can thin out quickly in a fast move, and the book at $89,000 wouldn't look like the book at $80,000 if price ever got there. Forced orders can also trigger others, which is how the October 10, 2025 cascade we covered elsewhere played out, far beyond what one band's size would suggest.

What do liquidation heatmaps get wrong?

Modeled maps

• Unknown leverage. Vendors assume it. CoinGlass's API reference and Hyblock's academy both describe calculating levels at various leverage amounts.
• Unknown collateral. This is the bigger gap in our test above, and we didn't see it discussed in the vendor docs we read. Knowing a trader's leverage setting wouldn't fix it for cross-margin accounts.
• Closed positions. Hyblock says its heatmap tracks where levels open but not where they close, so real liquidations come in lower. CoinGlass's guide also warns that actual amounts may be lower than shown.
• Relative scales. CoinGlass and Kingfisher both describe intensity, not dollars. A bright band can't be compared across coins or across vendors, and Coinperps notes that sizes differ between data providers.
• Missing venues. Coinperps says CME isn't modeled, and spot, options and ETF flows don't appear.

Real-position maps

• One venue. This is the limit on our figures, so here it is in numbers. Our tracked whales held $1.69B in BTC positions and $1.81B in ETH, counting longs and shorts together. CoinGecko's public derivatives feed, read at 12:44 UTC, lists about $66B of BTC and $43B of ETH perpetual open interest across venues. That's self-reported, may count sides differently by venue, and covers perpetuals only, so CME's dated futures aren't in it. As a rough scale, then, our whales are about 2.6% of the BTC figure and 4.2% of ETH. Hyperliquid as a whole is about 4.7% and 5.8% ($3.09B and $2.49B). Price is set across all of these markets, and a real map sees one of them.
• Prices that move. A cross position's price shifts with any change in the account. In March 2025, press reports described a Hyperliquid whale whose 40x BTC short was publicly targeted. Reports also said the whale added margin during the episode, which moved the level (Decrypt).
• No stop orders. Liquidations aren't the only forced-looking orders. Resting stops are separate, and HyperDash shows them in a separate widget.
• Survivors only. A snapshot shows positions that exist now, and it doesn't tell you about the ones that closed earlier.

Ground truth

The usual check on a heatmap is reported liquidation totals, but those have their own holes. Binance's documentation describes its public stream as a snapshot of the latest liquidation order per symbol, updated every 1,000 milliseconds. Hyperliquid's co-founder argued in October 2025 that this could undercount by up to 100 times in bursts (The Block). That 100x is his estimate, from a competitor, and not a measured figure, and The Block's report carried no Binance reply. The stream design itself is documented.

Is price drawn to liquidation clusters? What the evidence shows

The claim that bright bands act as magnets is everywhere in trading content. We looked for tests of it, and two different claims need separating:

• Attraction: price is more likely to reach a level because liquidations are there.
• Acceleration: once price reaches a level, forced orders make the move faster or larger.

Acceleration has reasonable support by analogy and from mechanism. Attraction, the claim traders usually mean, has very little.

Vendors don't agree on this, and none we read reports a test of it. CoinGlass's learning article calls these areas magnetic. HyperDash's guide says a band doesn't predict where price goes, Kingfisher presents its clusters as context and not as predictions, and OpenMarket's post says its map isn't a forecast.

Can traders deliberately aim for a level? Sometimes they try. In March 2025, a pseudonymous trader publicly called for a group to push BTC high enough to liquidate a 40x Hyperliquid short with a published liquidation price near $85,592. BTC reached about $84,584 and the position wasn't liquidated when Cointelegraph reported it, citing 10x Research's description of coordinated stop-hunting. Also that month, in the JELLY episode, accounts took opposing positions and used a self-liquidation method to pass a large short to Hyperliquid's vault, per Arkham. Both show that public levels can draw attention, and that the incentive in Brunnermeier and Pedersen's model isn't imaginary. A single attempt that failed, and a thin-market case in a small token, don't establish how often this works on BTC.

Can our data test it? Not yet, and we'd rather say that than imply otherwise. Our Hyperliquid tracking began on November 4, 2026, and we store each position's latest liquidation price rather than an archive of past ones. Across our 18 hourly market readings to 13:19 UTC on November 5, BTC's mark stayed between $79,699 and $80,680. The nearest BTC whale level was 2.68% away at our second read, so no large band has been tested. That is a sample of nothing.

What would a good test of a liquidation heatmap look like?

Most evidence offered for heatmaps is a chart with an arrow on it. Here's why that proves nothing. If daily volatility is σ and the path has no drift, the chance of touching a level at distance d within T days is about 2 × (1 − Φ(d / (σ√T))), where Φ is the normal distribution function. Using BTC's daily volatility of 2.06% over the past 90 days (from CoinGecko daily prices, November 5, 2026):

So a band 5% away gets touched within a week about a third of the time without any magnet. Screenshot the hits, and a heatmap looks prophetic. A fair test has to beat that baseline. We'd want five separate questions, because a map can pass one and fail another:

1. Location. Are the modeled levels where real ones are? Only Hyperliquid allows this directly: compare a vendor's Hyperliquid map with published prices at the same moment.
2. Size. Does modeled size track liquidations that followed? Coinfuty's test is this type, subject to the holes in reported data.
3. Attraction. Is a band reached more often than a matched placebo?
4. Acceleration. After a touch, is the move faster or larger than after touching a placebo level?
5. Economic size. Is any measured effect large compared with fees, funding and slippage? A past effect also doesn't guarantee a future one.

And the traps that make a test look better than it is:

• Hindsight data. Use maps archived at the time, not a chart rebuilt later. Some views count only levels price hasn't touched, as Hyblock describes for its profile, so a later look at the chart isn't the map a trader saw.
• A matched placebo. Compare each band with random levels at the same distance, on the same side, in the same volatility. A random level elsewhere isn't a fair control.
• Levels built from price. Modeled levels sit at fixed offsets from where trading took place, so they're mechanically tied to recent price. Any tendency of price to revisit its past range would look like a magnet. A control should shift the map by a set amount, or use the same offsets from a different period.
• Overlap. Hourly checks on a 7-day window aren't independent. Count episodes, and adjust the uncertainty for overlapping windows.
• Many tries. Band definitions, windows and coins make many combinations. Fix one before looking, and report all of them. Our lesson on measuring a signal honestly covers this in more detail.
• Splits. Report calm periods and cascade periods apart. A handful of huge days can carry the whole average.

Hyperliquid offers a rare setup for this: real levels instead of estimates. Archive every position's liquidation price every few minutes for a year, then run questions three and four against placebos. We found no one who has published that. It wouldn't settle things for Binance, but it would be the first direct measurement on real data.

How should a liquidation heatmap be read?

None of this makes heatmaps useless. They're a rough description of where forced exits may exist, so they work better as context than as a map of destinations. A checklist:

1. Identify the source. Modeled or real? Which venues? If the page doesn't say, assume modeled.
2. Check the scale. Relative intensity or dollars? Don't compare brightness across coins or vendors.
3. Check the time window. A longer lookback can pile up older levels, and on a model that doesn't track closes, some of those positions may no longer exist.
4. On real data, count wallets. One account behind a bar means that bar can vanish at any moment.
5. Separate location from direction. A band tells you what may happen if price gets there, not whether it does.
6. Compare with a base rate. When someone shows a band being hit, ask how often a random level at that distance gets hit too.
7. Check your own number. If you hold a position, the only liquidation price that matters to you is the one your venue reports for your account.

How do you build a Hyperliquid liquidation map yourself?

It's straightforward for any set of addresses, and you don't need us for it. Send a POST request to Hyperliquid's public info endpoint with {"type": "clearinghouseState", "user": "0x..."}. Each entry in assetPositions has a position with coin, szi (signed size), entryPx, positionValue, leverage and liquidationPx. Then:

1. Choose your addresses, and write down the rule. The endpoint is per address, so your choice of addresses decides what your map means. Ours are Hyperliquid accounts with $1M or more, without market makers and accounts that never traded.
2. Treat an empty liquidationPx as “no price”, never as zero, and count those positions separately.
3. Keep the side from the sign of szi, because long and short bands behave differently.
4. Group by a band width set as a percentage of price, and read the mark price at the same moment.
5. Keep the timestamp on every reading, and store your snapshots. Without an archive, you can't test anything later.

If you'd rather not run this, the tools in the table above do it at different scales. GMX is different: our GMX pages have no liquidation chart, because its public data doesn't include a liquidation price per position, and we show real figures rather than estimates.

What can't our data show?

• Whales outside our list. Our Hyperliquid list comes from the public leaderboard, so it misses other accounts. It also misses builder-deployed markets and sub-accounts.
• The market. Tracked whales on one venue, roughly 2.6% of reported BTC and 4.2% of reported ETH perpetual open interest. Most open interest sits elsewhere.
• Smaller traders. Our figures start at $10K positions in accounts mostly of $1M or more. The stated-versus-real gap may look very different lower down.
• History. Tracking began November 4, 2026. These are snapshots, not trends, and nothing here is a track record.
• Why a position exists. A long here might hedge a short elsewhere. We can see the position, not the reason, and we don't name any wallet's owner.
• Precision on Hyperliquid's own figures. Its docs note that estimates can differ from outcomes as book liquidity changes. We report the published field as it stood.

Bottom line

Modeled heatmaps are a reasonable answer to a hard problem: exchanges hide positions, so vendors estimate. Vendors say this in their docs more clearly than most people quoting them do. The trouble starts when a relative, assumption-driven estimate gets read as a list of real orders, or as a destination.

Real data on Hyperliquid shows how much can hide inside those assumptions. For our tracked whales on cross margin, stated leverage said 5% and published prices said over 40%, and the largest bands were usually one wallet. That doesn't prove that a Binance heatmap is wrong. It does show that leverage alone can't locate a liquidation for a well-funded account.

On attraction, the honest position is that no one has shown it with published numbers, as far as we could find. Acceleration is better supported, but mostly by analogy. Until someone archives real levels and runs a placebo test, a band is a description of positions at a moment. That's useful context, and it stops there.

Related reading

Crypto Whale Options and Derivatives

Deribit, put/call ratios, famous trades and the October 10, 2025 cascade in full.

Hyperliquid and Ethereum Whale Bridge Flows

How capital moves between Ethereum and Hyperliquid, and how the venue is built.

How to Spot Fake Whale Activity

The same skepticism applied to spot flows: what looks like a whale and isn't.

Whale Concentration Risk Methodology

How to read concentration in a token's holders, a close cousin of the one-wallet problem.

Learn: Measuring a Signal Honestly

Baselines, sample sizes and the traps that make a backtest look better than it is.

Learn: Spotting Noise in Whale Data

Telling a real pattern from a chance one in on-chain data.

Whale derivatives → Hyperliquid whale positions → ETH whale positions → GMX whale positions →