How accurate are liquidation heatmaps? What the map knows and what it guesses

Open a liquidation heatmap and the colours look like evidence. Bright bands above and below price, a glow that thickens where "the liquidations are". Most readers take the picture as a record of something that happened, or as a list of orders waiting to be hit. It is neither. A liquidation heatmap is two different things drawn on one canvas: a model of where open positions would be force-closed if price reached them, and, on the better maps, a tape of the liquidations that actually printed. The first is arithmetic on assumptions. The second is a measurement with holes in it.

This piece is about the assumptions and the holes. We pulled one week of BTC perpetual liquidations from MarketTrace's own tape, 8 to 15 October 2026, 19,162 prints and $464M across Binance, Bybit and OKX, and used it to check six things people take for granted when they read a liquidation map. The basic reading guide is a separate post, how to read a liquidation heatmap. This is the part that comes after it.

Two layers, two kinds of knowledge

A weather map and a rain gauge both tell you about rain. One is computed from a model. The other is a bucket that got wet. A liquidation heatmap is both at once, and the trouble starts when readers forget which layer they are looking at.

The expected layer is the model. On MarketTrace it is liqmap_model_v1, five rules written out on the methodology page. In plain words: when open interest rises across Binance, Bybit and OKX, the model assumes new positions were opened at that price, splits them into longs and shorts by who was hitting the market at the time (the taker buy ratio), and splits each side into equal thirds at 25x, 50x and 100x. Each third gets a liquidation level from its leverage and the venues' maintenance-margin rates. When open interest falls, every open lot shrinks by the same fraction. When price crosses a level, that part of the lot leaves the map. Everything is counted in price buckets about 0.06% wide, which at $81,600 is a $50 step.

The realized layer is the gauge: liquidation prints as the three venues published them, placed at the trade price at that moment where one is known.

The hit-rate is where the two meet, and it is the only number that can say how accurate the model was. More on that at the end.

1. "The bright bands are liquidations that happened"

They are not. Nothing in the expected layer has happened. Exchanges do not publish the liquidation prices of open positions, so no map, from any provider, can draw where the actual liquidations are waiting. What it can draw is where they would be if the model's assumptions held: positions opened where open interest rose, at the leverage the model assumed, held without anyone adding margin.

This is why maps from different providers share a shape and disagree on intensity. The same open-interest series, run through a different leverage mix or a different margin table, puts the glow at a different distance from price. Every map is a different recipe on the same ingredient.

Read a bright band as: if price got here, this much open interest would be at its liquidation point under these assumptions. Not as "this much will be liquidated here", and never as "this is where it was".

2. "The map knows how much leverage people use"

It assumes it. The leverage distribution of open positions is unobservable from outside. The exchanges know it and nobody else does, so a map has to pick.

The arithmetic of that pick is short. For a long with leverage L and a maintenance-margin rate m, the liquidation level sits about 1/L minus m below entry. With m around 0.4% to 0.5%, the first tier on BTC at the large venues, and taking 0.4% for the example, a 100x long opened at $81,600 is at its liquidation level 0.6% lower, near $81,110. A 50x long: 1.6% lower, near $80,290. A 25x long: 3.6% lower, near $78,650. A 10x long is 9.6% away, near $73,740, which at the usual zoom of a 24-hour map is near or beyond the edge of the chart. (MarketTrace's model uses 0.5% for all three venues, which moves each level 0.1% closer to entry.)

Change the assumption and the bands jump. MarketTrace's map makes one and states it on the page: equal thirds of 25x, 50x and 100x, a bracket around the common retail range rather than a point estimate. Treat the bands as a range of where forced orders could appear.

Two consequences follow. The map leans toward high leverage: 10x and 5x positions sit so far from price that they rarely render, so a map that looks quiet below price may be looking at the wrong leverage. And the map cannot see margin being added. A trader who tops up margin moves their liquidation level without touching open interest, and the band stays where it was. If you want your own number rather than a modelled one, the liquidation price calculator does the per-venue arithmetic.

3. "Every liquidation shows up as a dot"

The realized layer is a measurement, and the measuring instruments are the exchanges' public feeds. Each has its own gap.

Binance documents its liquidation stream as a snapshot: "For each symbol, only the latest one liquidation order within 1000ms will be pushed as the snapshot." In a cascade, where hundreds of positions die inside a second, the public feed carries one print for that second. Bybit's allLiquidation topic pushes liquidations in 500 ms batches. MarketTrace's coverage notes treat OKX's public channel as a sample. Hyperliquid publishes no liquidation feed; at the time of writing it holds 19% of BTC open interest, and the map draws nothing for it.

The Binance gap is visible in our week. Rank the week's minutes by dollars liquidated and take the top five. Binance printed 35, 22, 35, 31 and 19 liquidations in those minutes. Bybit printed 469, 145, 297, 375 and 295. Binance is the largest venue by open interest, and in those same five minutes it carried $78M of the $108M that printed. Its count just stops climbing at the ceiling of the feed.

So the count is a floor. The dollar figure is a floor too, since prints that were never pushed are never summed. Any site that totals liquidations across venues is adding a batched feed (Bybit) to a capped one (Binance) to a sampled one (OKX) and calling the sum "crypto liquidations". The number is useful. It is also smaller than reality by an amount nobody outside the exchanges can measure.

One week, one asset, three venues:

Source: MarketTrace liquidation archive, BTC, 2026-10-08 08:00 to 2026-10-15 08:00 UTC. Dollars are quantity times the venue's reference price.

Binance: a third of the prints, two thirds of the dollars. Part of that is venue size. Part is that the feeds do not even agree on what one liquidation is. Binance publishes one order per push; Bybit publishes executed size in batches, so a large position there can arrive as a run of small prints. Of the 43 prints of $1 million or more in the week, 41 were on Binance, two on OKX and none on Bybit. A count of prints is a count of feed messages. A map that sizes its dots by print count is drawing the feeds.

4. "The dollar figure is what got sold"

A liquidation print carries a size and a price, and the dollar figure on every map and every aggregator is one multiplied by the other. The price is the soft part.

Bybit documents the price in its liquidation stream as the bankruptcy price: the level at which the position's margin would be fully gone. Binance publishes the price of the liquidation order (p in its stream). Neither is where the forced order traded, and no venue's public feed says where it did. The liquidation engine takes a position over at the liquidation price; the bankruptcy price sits about one maintenance margin further along, in the direction of more loss.

The gap shows up in the data. Over seven days of BTC liquidations on MarketTrace's tape, Binance's reported price sat a median 0.39% past the same venue's last trade at that second, below it for a long and above it for a short. Bybit's sat 0.31% past it. That is about one maintenance margin. OKX's matched the last trade to within 0.01%. A map that draws each print at its reported price puts the Binance and Bybit dots a step past where the positions died, longs below the market and shorts above it, each about six $50 buckets out on BTC. MarketTrace draws each print at the same venue's trade within a second of the liquidation where it has one, which is nearly always, and inside that step's price range where it does not.

Sizes have the same softness. Large positions are liquidated in steps. When the market cannot absorb a forced order at better than its bankruptcy price, the insurance fund covers the gap or auto-deleveraging closes positions on the other side. The print you see is one slice of that process, priced at a reference level. MarketTrace labels its totals accordingly in the API: estimated_base_qty_x_source_reference_price. An estimate of notional. What the forced order received stays inside the exchange.

This is also the whole answer to "why does site A say $200M and site B say $350M for the same day". Different venue lists. Different handling of Binance's one-per-second snapshot. Different reference prices. Different deduplication. None of them is wrong in its own terms, and none of them is the number.

5. "Liquidation happens when the candle touches the band"

The candle you watch is last price on one venue. Liquidation engines do not use it. Binance states it directly: "Binance uses the Mark Price to prevent unnecessary liquidations and discourage market manipulations by bad actors." Mark price is built from an index across spot venues and a moving average of the basis; it moves more slowly than last price and ignores a single venue's wick. Bybit and OKX trigger on their mark prices for the same reason.

So a wick through a bright band with no prints under it does not mean the model failed. Last price went there. Mark price did not. The opposite happens too: in a fast move the mark can cross a level while the candle on your venue has not quite reached it, and prints appear under a band the chart never touched. On top of that, the band itself is a 0.06% bucket, and the three venues' margin tables differ, so "the level" is a short range rather than a line.

MarketTrace's model tests crossings against each step's high and low of a consolidated price, not a single venue's close: a part of a lot leaves the map when the step's low reaches its long level or the step's high reaches its short level. That is a fair test for a map drawn in steps of a minute or more. It is still a test against a price series the liquidation engines do not use, and the hit-rate inherits that.

6. "A liquidation heatmap is a liquidity heatmap"

They share a colour scheme and half a word, and they are opposites.

A liquidity heatmap, the Bookmap-style order-book heatmap, shows resting limit orders: real, visible in the book right now, and cancellable in a millisecond, which is the whole subject of spoofing and fake walls. A liquidation heatmap shows modelled forced-close points. They stay invisible until triggered and cannot be cancelled; the only way to move one is to add margin or close the position. One is what people are offering to do. The other is what the exchange would do to them.

The two interact, which is why the confusion persists. A cluster of long liquidation levels below price is a cluster of future market sell orders. If the book is thin there, those sells travel further, which is the mechanism behind a cascade and behind many of the moves that get called a stop hunt. Thick resting bids at the same price would absorb them. Read the two maps together and never read one as the other. The book side lives on the depth page; the forced-flow side on the liquidation heatmap.

What accuracy would have to mean

Any map can be called accurate. For the word to be checkable, it has to mean something specific.

Take every expected level that price crossed in a window. For each crossing, look at the realized tape: did liquidations of the same side print within one time step and 0.3% of that level, in an amount of at least $100,000 or 1% of what the model expected there? The share of crossings where they did is the hit-rate. MarketTrace computes it hourly for every asset, window and leverage preset, from the day the map went live, and publishes the figure only after three gates: 90 days of tape, less than 1% of minutes lost to feed gaps, and 30 days without a change to the model. Until then the page shows a day counter instead of a number.

Three things will push that number around, on any map, not only this one. The Binance snapshot and the OKX sample bias it downward: liquidations that happened but were not pushed count as misses. Positions closed or re-margined before price arrived are not counted as misses, but they were never going to be liquidations either; the model drew a band for something that quietly left. And some share of the hits may be self-fulfilling, since traders who read the same maps place orders at the same bands.

The May guide, how to read a liquidation heatmap, listed four other misreadings: the map as a price target, the magnet, the map as a substitute for the live tape, and one provider's output as universal. All four stand. The magnet claim in particular deserves a number rather than an argument, and the hit-rate is that number. A map that does not publish one is asking you to take the glow on faith.

What the map is for

Used as what it is, a liquidation heatmap answers a narrow question well: if price reaches this level, how much forced flow could appear, on which side, under which leverage assumption. It is a map of potential fuel, to be read against the real book and the real tape.

The habits follow from the six points above. Read the leverage assumption as a bracket. Check the coverage line before trusting a total. The venue list is the denominator. During a move, watch the realized layer to see whether the fuel is burning, and switch to the live liquidation tape when the move is fast. When a band is crossed cleanly with no prints, ask about mark price before you blame the model.

The free BTC liquidation heatmap shows the last 24 hours with both layers, the coverage line and the hit-rate counter. For a side-by-side with a predicted-only map, see CoinGlass liquidation heatmap vs MarketTrace. The same tape is queryable by agents over MCP: get_liquidations_recent returns up to 24 hours of liquidation totals by side, with the notional semantics spelled out, and the MCP server is free.

FAQ

Are liquidation heatmaps accurate?

Nobody can say until the publisher prints a hit-rate, so ask for one before trusting any map. The expected layer is a model on an assumed leverage mix. The realized layer is a partial tape. MarketTrace computes a hit-rate every hour and will publish it once there are 90 days of tape, under 1% feed gaps and 30 days of an unchanged model.

Does a liquidation heatmap show real orders?

No. It shows modelled levels at which open positions would be force-closed. Real resting orders are shown by an order-book heatmap, which is a different tool and often a different page.

Why do liquidation totals differ between sites?

Because each site adds different feeds. Binance pushes at most one liquidation per symbol per second, OKX publishes a sample, Hyperliquid publishes no feed; reference prices differ (bankruptcy, order, or trade price); deduplication differs. Every total is a floor and an estimate at the same time.

What leverage does a liquidation heatmap assume?

It depends on the provider. MarketTrace splits each new lot into equal thirds at 25x, 50x and 100x and says so on the page. At $81,600 that puts long liquidation levels about 0.6%, 1.6% and 3.6% below entry.

Can I be liquidated without price touching the level on my chart?

Yes. Liquidation triggers on mark price, an index-based price. The last trade on your venue is a different number. Your chart can wick through a level with no liquidation, and a liquidation can fire while your chart stops short of it.

Is the MarketTrace liquidation heatmap free?

Yes. The 24-hour map, both layers, the coverage line and the methodology are free with no login. Longer windows are planned for a Pro tier that is not on sale yet.

Data note and sources

Tape: BTC perpetual liquidations from Binance, Bybit and OKX, 2026-10-08 08:00 to 2026-10-15 08:00 UTC, 19,162 prints, $464.3M estimated notional, from MarketTrace's liquidation archive (/api/liquidations/range). Dollar values are quantity times the venue's reference price. The price gaps in section 4 come from the same archive: BTC liquidations over the seven days to 2026-10-15, median signed gap between each venue's reported price and its own last trade within one second (Binance −0.39% for longs, +0.39% for shorts; Bybit −0.31% and +0.32%; OKX within ±0.003%). Open-interest shares from MarketTrace get_market_state, 2026-10-23 09:45 UTC: Binance $7.75B, Bybit $5.11B, Hyperliquid $3.58B, OKX $2.57B.

• Binance, Liquidation Order Streams, USDⓈ-M Futures WebSocket documentation.
• Bybit, All Liquidation, v5 WebSocket public topic.
• Binance, What Are Mark Price and Price Index in USDⓈ-Margined Futures?
• MarketTrace, liquidations methodology: data sources, liqmap_model_v1, realized layer, hit-rate.

Descriptive microstructure, not financial advice.