Is the liquidation heatmap accurate?

It is a fair question and it is asked constantly. It is also almost never answered, because answering it requires the vendor to measure their own product and publish a number that might not flatter them. We did that, so this article can answer it directly — including for our own tool.

First, 'accurate' is two different questions

The first is whether the map shows where stop orders genuinely sit. The second is whether price does anything notable when it arrives at those levels. They have very different answers, and conflating them is where most of the confusion comes from.

Question one: are the levels real?

For spot forex, gold, indices and equities, nobody can verify this, including us. Stop orders rest inside individual brokers, there is no central book to observe, and no venue publishes them. The one public dataset that ever showed real retail stop placement — a broker's own order book endpoint — was withdrawn in 2024. Every heatmap on these markets is therefore a model.

Crypto derivatives are different: exchanges do publish forced-liquidation prints, so real data exists there. But it is worth knowing that the best-known crypto heatmap is still a model. CoinGlass's own documentation describes its heatmap as calculating liquidation levels from market data and assumed leverage amounts, and notes that actual liquidation amounts may be lower than the levels shown. That is the same class of artefact as ours, built by a bigger team on better inputs.

So the honest answer to question one is: no heatmap in this category is showing you verified stop orders. They are all estimates. Treat any product that implies otherwise with suspicion.

Question two: does price do anything there?

This one is testable, so we tested it. The method matters, because the obvious test gives a flattering and meaningless answer.

If you simply ask 'did price reach the level', you will get a high number — over half the time at ordinary distances — and it tells you nothing. Price wanders, so any level at that distance gets touched at a similar rate. The comparison only becomes meaningful against a control: an arbitrary price the same distance away, mirrored to the other side. That holds distance constant, so what is left is whatever the level construction contributes.

Across roughly 33,000 of our own historical levels, our modelled bands were reached 52.9% of the time against 53.3% for the control. Order blocks: 14.7% against 15.6%. Fair value gaps: 15.6% against 17.5%. We then asked whether price reverses at a level once it arrives — among 4,244 touched levels, 59.8% saw a reversal, against 59.7% at the control.

We also tested the classic alternatives: previous-day highs and lows, swing extremes, round numbers. All flat. The full method and figures are on our methodology page.

What that means if you use these tools

• Do not use a heatmap as a directional signal. Ours does not predict direction, and we have found no construction that does.
• Do use it to see where your own stop would be sitting relative to the crowd — that is a real and different question, and it does not require the levels to be predictive.
• Be sceptical of any accuracy claim that is not measured against a control. A raw hit rate at a short distance is close to meaningless.
• If a vendor publishes no accuracy figure at all, that is the industry norm rather than a red flag by itself — but it does mean you have no basis for the price you are paying.

Why publish this about our own product?

Because the alternative is selling something a customer can disprove. Anyone with the data and a couple of hours can run the test we ran, and a paid tool whose central claim collapses on inspection does not survive contact with a serious buyer.

What we are left with is smaller and true: one consistent modelling method across gold, forex, indices, equities and crypto, in a single terminal, showing where the crowd's stops plausibly sit. That is worth something to a trader deciding where to place a stop. It is not worth anything to a trader looking for a signal, and we would rather say so.