iteration 19 · 2026-07-23 · usefulness axis · laptop-drives-bigblack
📊 #19 Quantile monotonicity + PT/RW MR GROUNDED
Does the mean forward return increase monotonically across feature quantile bins (healthy), or spike in one tail (rare-veto)? The Patton–Timmermann / Romano–Wolf monotonic-relation test. All six gates pass, robust across 5 seeds — the iter-18 checkpoint's fragility (2/5) is fully resolved. Supersedes iter 18.
RW 0.00
rare-veto controlled (plain 1.0)
.05
FPR (per-test permutation, ∀H)
Preflight (resource-only): load1 2.06 · 42 GiB available · si/so ~0 · ClickHouse active, readonly=2. 5c/5G/no-swap capped.
The six gates
| Gate (§7 row 19) | Result (seed range) | Target | |
|---|
| Admit demo | trend-t ~120, MR% 0.778–0.889, sign-stable | p<.05, MR%≥7/9 | PASS |
| Power @ IC=.05 | 1.00 | ≥ .8 | PASS |
| Both nulls FPR ∈ [.03,.07] | bp 0.037–0.060, bt 0.047–0.055 | ∈[.03,.07] ∀H | PASS |
| RW trap (rare-veto ≤ .05) | plain PT 1.0, RW 0.000–0.017 | ≤ .05 | PASS |
| Collinear decoy collapses | 1.00 | ≥ .9 | PASS |
| Envelope FNR (non-monotone) | rare-veto + interaction 1.00 | measured | PASS |
The three checkpoint fixes (all principled, not threshold-shopping)
- Heavy-tie features can't be quantile-binned → dropped from the family. Mostly-single-value features give chronological bins (ties broken by position) that pick up return drift → FPR corrupted to 0.375. Grounded on full-spread features (the #4 heavy-tie class).
- A genuinely-monotone exemplar → IC=0.50. At IC=0.30 the profile is extremes-dominated; IC=0.50 is monotone throughout (
[-1.06, -0.50, -0.14, …, +0.50, +1.05]) so MR%≥7/9 every seed.
- The RW-FWE trap sharpened → both halves of the profile must trend. A genuine monotone separates across the whole range; a one-sided rare-veto's low half is flat → rejected (RW 0.000–0.017, while plain PT over-rejects 1.0).
FPR calibration (the ∀H point). The analytic N(0,1) threshold over-rejects the autocorrelation-preserving null (a real feature's bins + a block-permuted target both carry autocorrelation → FPR 0.13). The per-test block-permutation p-value (p≤.05, B=199) is exact and autocorrelation-aware → both nulls control at ~.05 every seed; R=600 keeps the estimate inside the tight [.03,.07] band.
Published blind spot (envelope). The MR test is monotone-only by design → it MEASURES (not requires) the FNR on useful-but-non-monotone signals (rare-veto, interaction-only, FNR 1.0) → route to #11 (CMI) / detector track. ★ Usefulness axis roster complete (#4, #5, #7, #19). Next: conditional axis — #11 CMI.