iteration 18 · 2026-07-23 · usefulness axis · laptop-drives-bigblack
📊 #19 Quantile monotonicity + PT/RW MR CHECKPOINT · in-progress
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. The mechanics are right (5/6 base-seed gates), but three genuinely-hard sub-problems on real 3s-return data need careful work before it grounds.
.042
FPR (per-test permutation)
Preflight (resource-only): load1 3.43 (elevated, other users; no heavy nasimubd job) · 42 GiB available · si/so ~0 · ClickHouse active, readonly=2. 5c/5G/no-swap capped.
State across 5 seeds
| Gate (§7 row 19) | Base seed | Seed range | |
|---|
| Admit demo | PT p 0.01, MR% 0.778 | MR% 0.556–0.889 | fragile |
| Power @ IC=.05 | 1.00 | 1.00 | robust |
| Both nulls FPR ∈ [.03,.07] | 0.042 / 0.042 | bp 0.008–0.375 | heavy-tie |
| RW trap (rare-veto ≤ .05) | plain PT 1.0, RW 0.10 | RW 0.000–0.167 | leaks |
| Collinear decoy collapses | 1.00 | 0.97–1.00 | robust |
| Envelope FNR (non-monotone) | veto 0.90, inter 1.0 | 0.83–1.00 | robust |
What is solid (real progress)
- The PT/RW split is correct. The formal PT test = a studentized linear-trend contrast; significance by a per-test block-permutation p-value (permute the target within the feature's bins) → an exact, well-calibrated FPR in principle (0.042/0.042 base), autocorrelation-aware. Power(PT) 1.0 at IC=.05.
- The RW-FWE substance is the both-tails requirement: a genuine monotone relation separates BOTH tails from the center; a one-sided rare-veto separates only the top.
- Decoy collapse (0.97–1.0) and the envelope FNR (rare-veto + interaction-only missed → route to #11) are robust.
Three open items (genuinely hard on real 3s returns)
- Heavy-tie binning. Heavy-tie features (aggregation_density, lookback_hurst) give degenerate quantile bins → FPR corrupts to 0.375 on one seed (the #4 heavy-tie class). Fix: full-spread-only null pool or sub-ULP jitter.
- The monotone exemplar is extremes-dominated. Even IC=0.30 gives a bin profile huge at the tails and ≈0 in the middle (
[-0.82, ~0×8, +0.80]) — not monotone-throughout — so MR% swings 0.556–0.889 by seed. Fix: a robustly monotone exemplar (IC~0.5 or a real driver).
- The rare-veto trap leaks (RW 0.0–0.167; need ≤.05). Fix: require ≥k significant increments spread across the range (a true RW-FWE step-down).
Resolution (next firing, pre-registered): tie-robust binning → robustly-monotone admit exemplar → sharper rare-veto trap → seed sweep, then a terminal verdict. The usefulness axis is already complete (#4, #5, #7); this is the 4th usefulness instrument.