Equity book · gate study · 16 Mar – 31 Aug 2026

Does G3 Earn What It Throws Away?

G3 requires 30-day momentum of at least 25%. It is the only one of the eight gates never measured on anything but the trades it already let through, and it refuses 96% of everything that clears G1, G2 and G4. The threshold is a round number. This asks what the round number buys. The answer is nothing measurable, and the honest response is to leave it alone anyway.

46,404 setups, no momentum cut32,007 after G4 4 thresholds x 2 breadth armswithin-period split + permutation
population admitted by G3
4%
1,399 of 32,007
what that buys
+0.34pp
kept minus refused, mean return
is it chance
p=0.43
20,000 permutations
rank correlation
−0.079
a ranking gate needs this positive
The shape

Momentum does not rank returns. The curve is a U.

Mean return by 30-day momentum at entry, across every setup that cleared G1, G2 and the 5% stop cap. No relative strength filter and no earnings filter, so there is no look-ahead in this chart at all.

0%5%G3 cuts here — everything left of this line is refused4.32under 0n=13,4523.560 to 10n=10,9443.1510 to 15n=3,2342.7215 to 20n=1,9651.9820 to 25n=1,0132.6425 to 30n=5655.0730+n=834
The band G3 admits is the flat part of the curve. The worst returns in the whole population sit between 15% and 30% momentum — precisely where the gate draws its line. Only the 30%+ band stands out, and the strongest band of all is the one with negative momentum, which the gate refuses outright.
momentum bandn meanmedianwin rate
under 013,452+4.32%-1.73%40%
0 to 1010,944+3.56%-1.92%38%
10 to 153,234+3.15%-2.79%34%
15 to 201,965+2.72%-3.19%32%
20 to 251,013+1.98%-3.17%30%
25 to 30565+2.64%-3.23%30%
30+834+5.07%-2.84%34%
Every median is negative and no win rate reaches 41%. This is a population of mostly-losers with a fat right tail, which is what a breakout book looks like before the eight gates and twelve slots are applied to it. Means are driven by the tail.
The test

It does not survive the within-period split

A cross-section is not evidence until it survives one. Split the same population by period and the difference the gate makes changes sign:

periodsetups kept minus refusedpreads as
Mar–May11,095+2.21 pp0.006the gate helps
Jun–Jul14,045−2.76 pp0.0002the gate hurts, and by more
Aug6,867+0.40 pp0.59nothing
all three32,007+0.34 pp0.43nothing
Two significant results pointing opposite ways is not two findings. It is one non-finding, and the pooled number is the average of a coin landing on each side. This is the fifth proposal in two weeks to die here.
The tempting mistake

Loosening the threshold is measurably worse

The obvious move from "the gate costs 96% of admissions and buys nothing" is to lower it. The data refuses that too, and more clearly than it refuses the gate:

G3 thresholdadmits kept minus refusedp
≥ 25 (shipped)4%+0.34 pp0.43
≥ 208%−0.60 pp0.07
≥ 1514%−0.89 pp0.0005
Lowering the bar walks the gate into the worst part of the curve. A single threshold cannot express a U-shaped relationship, and every threshold available makes the gate worse than leaving it where it is.
Verdict

No change. Record it and leave the gate alone.

G3 at 25 is not earning its exclusions — it admits four percent of the population for a difference indistinguishable from zero. That is a real finding and it is not a licence to act. The effect flips sign between periods, which means the pooled estimate is an artefact, and every alternative threshold tested is worse than the one shipped.

The gate stays. What changes is that it is now measured rather than inherited, and it joins the list of rules whose cost is known and whose benefit is not. If admissions remain the binding constraint on the book after the v156 stop change, G3 is the first place to look — but it needs a rule shaped like the curve, not a lower line through the middle of it.

Caveats, stated rather than buried: the window is five and a half months of one market. The secondary arms in the script apply today's RS and EPS to March setups and are therefore optimistic by roughly five times at the level — they are reported because the ordering across momentum bands is identical in both, which is the only thing this study rests on.

← Trading book · 2026-09-23_g3_momentum.py · 23 September 2026