Research library

2026-07-29 to 2026-08-28 · 23 sessions

The early warning is as good as the full signal — and that is not a promotion

A quiet heads-up fired when a large position is only part-way unwound. It grades level with the loud signal it precedes, which makes it real and still does not justify making it loud.

VerdictNEGATIVE EVIDENCE for promotion · stays quiet

In simple words

What happened

The partial warning hits 53% of the time against 50% for the full signal that already earns a loud alert — statistically indistinguishable, and with the largest typical peak of the three stages. So the warning is measuring something real. It is not measuring something better, and "as good as the thing we already have" is not a reason to add another loud alert to a feed we recently spent effort quieting.

Who would pay?

A large position rarely closes at once. If a partial exit is detectable, the warning arrives before the full exit does — and the earlier warning is worth more, provided it is not just noise.

How it was tested

All three stages of the same unwind lifecycle were graded the same way over the same window, on real contract prices, collapsed to one observation per name, day and side. The comparison is between stages of one process rather than against an unrelated benchmark.

What would prove it wrong?

The pre-registered failure mode was noise: more than 1.5 fires per session per side would have meant the warning was firing on ordinary churn. It fires 0.87 times per session, comfortably inside the ceiling.

Leader-collapsed evidence

What each setup actually did

19 independent partial warnings, against 24 early and 22 full signals over the same window

StageEpisodesHit rateLower boundMedian peakCurrent treatmentDecision
PARTIALPartial unwind detected (the early warning)1953%+115.7%32%Quiet, 0.87 fires/sessionNEGATIVE EVIDENCE
EARLYEarly unwind stage2458%+101.1%39%QuietNEGATIVE EVIDENCE
FULLFull unwind (already a loud alert)2250%+97.6%31%LoudNEGATIVE EVIDENCE

What we keep

Learnings

  1. 01

    Parity is a real result and it is not a promotion. The honest summary is that the early warning is not noise — which is a different claim from saying it should be louder.

  2. 02

    Set the noise ceiling before you look. Fire rate was the pre-registered way this could have failed, it was checked first, and it passed — which is what makes the outcome comparison worth reading at all.

  3. 03

    Compare stages of one process rather than reaching for an outside benchmark. All three stages share a lane, a window and a grading rule, so the differences between them are about the stage and nothing else.

  4. 04

    A standing instruction to reduce alert volume is itself evidence. When a new candidate grades level with an existing alert, the default should be to keep it quiet.

  5. 05

    Typical peak favours the partial warning (+116% against +98%), but peaks are not captured returns, so that difference does not carry the decision.

The partial warning runs and stays quiet. It keeps recording, so if it separates from the full signal on a larger sample the case can be made then.
Back to all studies