Liquidation cascade study

We tested one popular crypto trading idea once, and first proved the test could say yes.

Sector: Quantitative research · Year: 2026

Summary

An in-house study on Binance perpetual futures. The idea: when forced liquidations push a market too far, the price snaps back far enough to trade. We registered one hypothesis, sealed half the data, set the cost bar at twice what we think is realistic, and wrote down in advance that a miss means we close the study. Before looking at a single result, the testing pipeline had to prove it could find a real effect and reject a fake one. It caught a flaw in our own work on the way. The idea then failed the test decisively.

Challenge

Most crypto backtests flatter themselves. They peek at the future by accident, they quietly try many variants until one works, and they assume costs that no one pays. A negative result is only worth something if the test that produced it could have produced a positive one. We wanted a result we could trust in either direction, and a record that would stop us asking the same question again without new information.

Approach

  • Chose a single hypothesis on ten of the largest USDT-margined perpetual contracts on Binance, built from the exchange’s public archive. Forced liquidations are not published there, so the events are inferred, and the registration says so.
  • Fixed everything in advance and fingerprinted it: twelve months for development, six months sealed and never downloaded, a cost bar of twice the realistic figure, and the sentence “a result below the gate is a null, and the response is to close the study, not to re-parameterise.”
  • Stated the power up front. With about forty market-wide event days as the real sample, a null was the expected outcome, and a null would show only that this sample could not detect an edge.
  • Built the pipeline, then tested the pipeline before testing the idea. A replay test hides the future and checks that nothing changes. An oracle given deliberate foresight must pass. Eight coin-flip sets must all fail.
  • Failed the first replay test. A one-minute overlap between the window used to detect an event and the bar used to enter it changed 215 of 321 events. No return had been computed. We corrected the bar alignment in a dated, hashed amendment and rebuilt the event list.
  • Passed every validation check on the corrected pipeline: no differences across 50 random replays, all 312 events unchanged by an entry-bar cut, the oracle passed, all eight coin-flip sets failed. Only then did the real data run, once.

Outcome

The idea failed decisively, and the study is closed. Annualised Sharpe was −1.64 against 1.0 needed, and the clustered t-statistic was −1.95 against 2.20. Five of six statistical criteria failed, and the lower cost assumption did not rescue it. There is one detail worth keeping. Before costs the average trade gained about as much as the cost bar, but a single crash day produced the three largest winners, and without the ten best trades the rest lost money. That is a description, not a reason to re-tune, and we did not. The six-month holdout was never opened. What remains is a test harness proven to detect a real effect and reject noise, and a pre-registered null that will stop the question being asked again without new information. No client capital is involved and no event definition or trading parameter is published.

Stack

Python, pandas, NumPy, Parquet

Published papers

  • A Test That Could Say Yes — A pre-registered test of liquidation-cascade reversal in crypto futures found no tradable edge, and proved beforehand that it could have found one. 8 pages, PDF.

More work