▲ Work in progress. This site is built and researched in the open: pages, numbers and findings may change or be wrong, and the record has gaps. Not advice. More
The Lab
Ideas traders use, put to the record
← the Lab · signals tested · write-up
#67 Real liquidation clusters: magnets, turns, or neither? Not shown
Past prices and recorded liquidations; no real orders were placed. A first look, reported and not judged: the judged looks are fixed for 9 January and 9 April 2027.
The question
Liquidation "heat maps" are a common sight on trading charts. The idea behind them is that where many leveraged positions were forced out, the price is drawn back (a magnet) or turns (a wall). An earlier study here tested modelled liquidation levels, worked out from swing points and leverage, and found no edge after fees. This one tests levels where liquidations actually happened:
- Is a level where liquidations clustered touched again more often than chance?
- Does fading the price at that level pay after fees?
How we tested
- Bitcoin, Ether and Solana perpetuals. Recorded liquidations from Binance, Bybit and OKX, added together in US dollars.
- Two sources of liquidations:
- the free sample days from Tardis.dev, the 1st of each month from January 2020 (Solana from October 2020): 82 days for Bitcoin and Ether, 73 for Solana. These are used to choose one setting, then shown for orientation;
- our own record of every liquidation, continuous from 11 October 2026. Only 17 days so far.
- Prices: Binance's public one-minute candles (data.binance.vision).
- Everything was fixed before any result was read: the question, the settings, the controls, the pass marks and the dates of the judged looks.
- A cluster: each hour, the last 6 hours of liquidations are sorted into price steps 0.25% wide. A step counts as a cluster if its total is in the top 20% of the steps that had any liquidations, and it sits at least 0.3% from the price. Each cluster is counted once, in the hour it first appears. (The top 20% was chosen from 20%, 10% and 5% on the Tardis days. It is the setting whose clusters were touched most often compared with the mirror level.)
- Two controls for every cluster:
- the mirror: the same distance on the other side of the price;
- the busiest price: the 0.25% step on the cluster's side where the most volume traded in the same 6 hours. A liquidation can only happen where price traded, so this asks whether liquidations add anything beyond "price was busy there".
- Magnet: does the price touch the level within 4 hours?
- Turn: at the first touch, take the opposite side with a limit order at the level (0.02% fee), a stop 1% beyond it, and close at market 4 hours later (0.055% fee each way). The same trade is run on the controls.
- The uncertainty: whole days are resampled 2,000 times (a block bootstrap: Efron, 1979, and Künsch, 1989, for blocks), and the middle 95% of the results is shown as an interval.
What we found
1. Clusters were touched about as often as the mirror level. Over 4 hours, on the Tardis days:
| clusters | touched: cluster | mirror | busiest price | cluster minus mirror (95% interval) | |
|---|---|---|---|---|---|
| Bitcoin | 1,007 | 52% | 47% | 56% | +4.7 points (−0.3 to +9.6) |
| Ether | 1,170 | 52% | 55% | 57% | −2.9 points (−8.0 to +2.4) |
| Solana | 1,196 | 55% | 55% | 64% | −0.3 points (−6.1 to +5.5) |
Every interval spans zero. In our own 17 days the result was the same: Bitcoin +7.2 points (−3.0 to +17.1), Ether +0.7, Solana +0.6, each from 166 to 312 clusters.
Bands are 95% intervals from resampling whole days; a band that spans zero cannot be told from zero.
2. The busiest price was touched more often than the cluster, on all three coins (4 to 8 points, every interval below zero on the Tardis days). Part of this is distance: the busiest price sat closer to the price than the clusters did (median 0.51–0.58% away, against 0.62–1.04%). So this comparison is tilted against the clusters, and the mirror is the fairer one. The mirror shows nothing either.
3. Fading a cluster lost a little after fees. Net per trade on the Tardis days: Bitcoin −0.04%, Ether −0.16%, Solana +0.00%. In our 17 days: −0.09%, −0.14%, −0.07%. Fading the busiest price instead did about as well or as badly; no difference could be told from zero.
Bands are 95% intervals from resampling whole days; a band that spans zero cannot be told from zero.
4. No split rescued it. Clusters made of longs' liquidations and of shorts' liquidations, each exchange on its own, 1 and 24 hours, and hours when open interest rose or fell: all were printed (81 rows). None showed a steady effect across coins and sources. The few intervals that clear zero do not repeat on another coin or in the other source, which is what chance gives over this many rows.
What this means, and its limits
- In this data, a price where liquidations clustered behaved like any other price at the same distance. Fading it lost a little after fees.
- Limits:
- Binance's public feed reports at most one liquidation per coin per second, so its totals are a floor.
- The Tardis days are one day a month: many clusters, but few days.
- 17 continuous days is too short for a verdict. That is why the judged looks are fixed for later.
- The "busiest price" control is not matched on distance (finding 2). This was decided in advance and is kept as it was, and the mirror is the fair comparison.
What comes next
- 9 January 2027: the record from 28 October 2026 to 8 January 2027 is judged against the pass marks fixed in advance.
- 9 April 2027: the next three months, alone, judged again.
- A test passes only if it passes both times, on at least two of the three coins.
Credits
- Liquidations: Tardis.dev's free monthly sample days; our own record of Binance's, Bybit's and OKX's public liquidation feeds.
- Prices: Binance's public data archive (data.binance.vision).
- Bradley Efron, "Bootstrap Methods: Another Look at the Jackknife", Annals of Statistics, 1979: the bootstrap.
- Hans R. Künsch, "The Jackknife and the Bootstrap for General Stationary Observations", Annals of Statistics, 1989: resampling in blocks.
- The clusters, the two controls, the fade and the look dates are our own design. The earlier study of modelled levels: "Liquidation levels after a turn: targets, or just distance?" on this site.
- PerpGremlin and Claude, Real liquidation clusters: magnets, turns, or neither?, TerminalGremlin, 2026: the tests, the code and the words.
Educational research from public market data; any trading discussed is paper, never real. Nothing here is advice. The full notice.