Research · test #1
Do liquidation heatmaps predict price? Undecided
The popular claim: price gets "pulled" toward the price levels where the most leveraged positions would be liquidated. We wrote the rules down before running anything, then tested them. First 30 days: no edge. The forward test is collecting data now.
The rules (fixed before the run)
- Coins: BTC and ETH perpetuals on Binance. A check every 4 hours.
- From our liquidation map, take the biggest cluster 0.5-3% above price and the biggest 0.5-3% below. If one side is at least 1.5x heavier, that is the target.
- The control is a level at the same distance on the opposite side. If price simply drifts, it hits target and control equally often.
- Outcome: which one price touches first within 24 hours.
- An edge needs: target first in at least 55% of cases, statistically significant (p < 0.01), above what random side assignment produces, on both the past data and the forward period, and in both directions.
First result (21 Sep - 21 Oct 2026)
| Events | Target first | p-value | |
|---|---|---|---|
| BTC | 43 | 62.8% | 0.063 |
| ETH | 62 | 50.0% | 0.551 |
| Total | 105 | 55.2% | 0.165 |
With random target sides, 99% of shuffles stay below 61.0% - our 55.2% sits well inside that noise. The split by direction gives it away: "target below" won on both coins, "target above" did not (ETH: 34.6%). That is the market drifting down during the month, not a magnet - a real pull would work both ways.
What would change our mind
The forward test uses only data collected after 21 October 2026. We will publish the verdict once at least 200 events have resolved, whatever it says. Until then the label stays undecided.
Limits of this test
- The map is a model: leverage mix and long/short split are assumptions. A different model could behave differently.
- One month is one market regime.
- Binance publishes at most one liquidation per symbol per second, so real liquidation volumes are understated (this affects only our model-vs-reality check, not the main test).