DetectorKilled

Trade the forced sell-off at the squeeze moment

+1.12 bps

average return on fresh data, not statistically real, and 10× weaker than in testing

bps = basis points = a hundredth of a percent (100 bps = 1%).

In plain English

Betting that forced sell-offs keep falling worked in old data but fell apart on fresh data, the effect shrank tenfold and flipped direction.

What was claimed

When the exchange's vault force-closes positions the price keeps falling, so sell into the squeeze.

The bar we locked

> ~18 bps after cost, stable out-of-sample

What the data said

+1.12 bps, not statistically significant, 10× weaker than in testing

Why it's dead

A real mechanism, the exchange's HLP vault (its in-house liquidity pool) forcibly closes positions during cascades, that completely failed to survive a clean OOS split (out-of-sample: judging the signal on data it was never fitted to). The in-sample magnitude, how big the effect looked on the data used to find it, collapsed 10× and the signal flipped sign on fresh data.

The detail

OOS holdout Mar 18–25. Exploratory (in-sample) signal: −11.95 bps mean at 5m. OOS mean: +1.12 bps, p=0.449 (chance alone produces a result this size almost half the time), not significant, and ~10× too small for the 18 bps cost hurdle even at face value. The collapse from in-sample to OOS is the classic overfitting signature (the model memorised its history instead of learning something real).

Kill date: 2026-03-26
Sample: fresh data, Mar 18–25
Method: Pre-registered live test
Verdict: 10× in-sample collapse