Accuracy: how we measure whether the map is right

Every liquidation map is an estimate. Nobody publishes every trader's position and leverage, so the real question is not whether a map looks convincing but how closely it matches what happens next. liqlin measures that against liquidations actually observed on exchanges and against Hyperliquid's real positions, always on data the model never saw while it was being fitted, and publishes the result for each pair.

What it means for a liquidation heatmap to be right

The map says, for every price level, how much leveraged notional would be liquidated if the mark price reached it. Each time price trades through a zone, the model records an event with two numbers: the pressure it had predicted across the levels crossed, and the notional actually liquidated on that side during that minute. If the map is any good, more predicted pressure should line up with more liquidations.

The prediction is stored before the model removes the positions at the crossed levels. Every event therefore compares what the map said beforehand with what happened afterwards, never the other way round.

Which real liquidations the liquidation map is measured against

Exchanges publish different things, and that decides what can be measured on each.

ExchangeWhat it publishesHow we use it
BybitEvery liquidation (allLiquidation)Real notional per minute, so sizes can be compared
BinanceA sample: at most one liquidation per symbol per second, the latestLower bound: ranking and whether anything was liquidated
OKXA sampleLower bound, as with Binance
HyperliquidNo liquidation stream, but every position with its liquidation priceA real map to compare shape against

During a cascade, a sampled stream can show three liquidations out of two hundred. Comparing predicted dollars with observed dollars on Binance or OKX would mostly measure the sampling, so there we compare ranking and presence instead. Only minutes when the capture was actually listening count: a minute with no capture is a gap, not a minute without liquidations, and counting it would blame the map for something nobody saw.

Hit rate and rank correlation: the core accuracy metrics

The hit rate asks a plain question: when price crossed a zone the map flagged as heavy, were there liquidations?

hit_q = P(observed > 0 | predicted ≥ percentile q)

It is computed at the 50th, 75th, 90th and 95th percentiles of predicted pressure. hit_95 looks only at the top 5% of events by pressure; hit_50 at the top half. A map that carries information has hit_95 well above hit_50, meaning it is right more often exactly where it is most confident. If the two are close, the colours on the map are not telling you much.

The second metric is Spearman's rank correlation between predicted pressure and observed notional, over events that had liquidations. It compares order rather than size, which is why it survives sampling: even if Binance only shows an unknown fraction, the big events should still rank near the top of both lists. On Bybit, where the stream is complete, a size-aware loss is added with a free scale factor, and that factor is never reported as accuracy.

Walk-forward testing: never fit and score on the same days

A model tuned on a period always looks good on that period. For the score to mean anything, every evaluation uses days the fit never touched:

  • Calibration window: twenty-one days, where parameters are searched.
  • Embargo: one day left unused, so nothing from the fit leaks into the evaluation.
  • Evaluation window: the following seven days, with parameters frozen.
  • Step: the cut moves forward seven days and the whole thing repeats.

The figure that matters is the average across out-of-sample cuts, never the in-sample fit. Before scoring, the model is replayed from scratch with the frozen parameters, so the map that arrives at each cut was built only from data that came before it.

Liquidation heatmap vs. Hyperliquid's real positions

Hyperliquid lets anyone read each open position along with its liquidation price. Adding them up by level gives the real map, which is then compared with the map the same model infers from Hyperliquid's open interest, trades and mark price.

shape = W₁(inferred map, real map), per side, in bp, within ±5% of mark

Wasserstein-1 distance reads as how many basis points, on average, you would have to shift the inferred map to make it match the real one. Each side is normalised before comparing, and cosine similarity between the two maps is reported alongside it. It is recomputed every time the real map is rebuilt, so it is the one accuracy figure available from day one, before any calibration. It always comes with coverage, the share of open interest the scanned positions add up to; below one half, the figure is indicative only and the screen flags it.

When a new version is promoted, and what "uncalibrated" means

Model parameters, such as the weight of each leverage tier, decay, and how entries and closes are spread, are recalibrated weekly. A new version replaces the active one only if, out of sample, it passes all four checks:

  • It improves the main loss by at least 3%; for a pair without complete liquidation data, it instead raises hit_95 by at least two points or improves rank correlation by 3%.
  • It does not make rank correlation worse.
  • It does not worsen shape distance on Hyperliquid by more than 10%.
  • It does not lower hit_95 by more than two points.

If any check fails, the version is kept as a candidate for manual review. When one is promoted, parameters are not swapped on a running map: the map is rebuilt by replaying thirty days with the new values. A state built with one set of parameters and continued with another would be a map no model ever produced.

Until a pair has a promoted calibration, it runs on starting values measured from real Hyperliquid positions and carries a visible "uncalibrated" badge. Calling the model ready also requires, for BTC: at least eight weeks of live data, three complete walk-forward cuts, out-of-sample Spearman above 0.55, hit_95 of at least 0.70 with a lead of 0.30 over hit_50, and shape distance under 25 bp on Hyperliquid. Those are design targets, not results. If they are still out of reach after twelve weeks, the model's structure gets reviewed rather than its parameters tuned further.

Where liquidation map accuracy is published

Results will be published by exchange, pair and mode once pairs have their first calibration: the hit rate of the latest promoted calibration, always out of sample, how many liquidation events there were over the last seven days, and whether the exchange publishes every liquidation or only a sample. They will update weekly, with each calibration run.

We do not quote an accuracy number here. It depends on the pair, the exchange and the week, and the only one worth reading is the published one, with its context. What we can explain is how it is computed, and that it never comes from the same data the model was fitted on.

Frequently asked questions

Why not measure against Binance liquidations alone?

Binance publishes a sample: at most one liquidation per symbol per second. That is enough to rank events and tell whether liquidations happened, not how much was liquidated. Sizes are checked against Bybit, which publishes all of them, and shape against Hyperliquid's real positions.

What is the difference between hit_95 and hit_50?

hit_95 is the hit rate across the 5% of zones the map rated heaviest; hit_50 across the top half. A useful map is clearly right more often in its strongest zones than in average ones.

What does the "uncalibrated" badge mean?

The pair has no promoted calibration yet. Its map runs on starting values measured from real Hyperliquid positions, but it has not yet passed the promotion rules on its own data.

Does high accuracy mean price will go to that zone?

No. Accuracy measures whether liquidations were there when price crossed a flagged zone. It says nothing about where price goes next and is not a recommendation to trade.

Related maps

Guides to read the map