Mean-reversionKilled

Betting the after-hours price gap “snaps back” in 30 minutes

26.6% vs 73.4%

reversion accuracy vs just assuming it persists

In plain English

We assumed a price gap would close within half an hour; across 5,800 cases it usually didn’t, the gap tends to stick around, not snap back.

What was claimed

When the perp's price drifts away from the real-world price it tracks, the gap snaps shut within half an hour, so trade the snap.

The bar we locked

the gap closes more than half the time

What the data said

just 27%, the gap usually sticks, it doesn’t close

Why it's dead

The mark-vs-oracle basis on HIP-3 equity perps (the gap between the price the perp trades at and the external reference price it is meant to track) doesn't revert, it persists. The signal was directionally backwards, and fully powered: more than enough data for the answer to be real, not a sample-size fluke.

The detail

n=5,814 events across 12 instruments / 4 asset classes. Binary 30-min reversion accuracy was 0.266 against a naive-persistence baseline of 0.734 (the score you'd get by just assuming the gap stays put), a −46.7 point gap. Robust to leave-one-out (drop any one symbol and the result moved <0.01) and to the locked sensitivity fork (the one alternative test spec we committed to in advance). A clean falsification, not a sample problem.

Kill date: 2026-06-09
Sample: n=5,814
Method: Pre-registered live test
Verdict: fully-powered falsification