📈 #4 Rank IC + ICIR + HAC-t CHECKPOINT · in-progress

The rank information coefficient: does the feature carry monotone forward-predictive signal — and is it significant once you correct for autocorrelation? This opens the usefulness axis. Injection bug fixed; Admit + power pass; the remaining gates move to the right level of aggregation.

2 / 5

gates pass (injection fixed)

0.030 / 0.081

achieved IC vs targets .03/.08

H 0.809

vwap period-IC ≈ target .79 ✓

2.19×

period naïve/HAC @lag30

Preflight (resource-only): load1 0.13 · 42 GiB available · si/so ~0 · ClickHouse active, readonly=2. 5c/5G/no-swap capped, single-thread BLAS.

State

Gate (§7 row 4)	Result	Target
Admit demo	achieved IC 0.081, ICIR 0.75, HAC-t 21, sign-stable	\\|t\\|≥3, IC≥.01, ICIR≥.5	PASS
Power @ IC=.03 / T=750	achieved IC 0.0295, power 1.0	≥ .8	PASS
Null FPR ≤.01 & naïve ≥2×	HAC FPR 0.003 ✓; bar-level naïve/HAC 1.0	≥2×	wrong level
N_min(H≈.79) ≥ 2×N_min(.5)	bar contribution near-white (H 0.52)	ratio ≥2	wrong level
Envelope FNR (non-mono / interaction)	non-monotone missed ✓; interaction marginal	FNR=1 both	OPEN

Three construction errors caught (verify-before-report)

Injection bug — fixed. The signal-injection used a raw real feature as noise; two features have heavy ties (aggregation_density, lookback_hurst) → they add almost no rank variation, so a tiny injected signal dominated the ranking (achieved IC 0.97 / 0.68 instead of 0.03). Fix: use the globally-permuted return as noise → achieved 0.030 / 0.081.
The bar-level contribution is near-white — a real finding. zr(f)·zr(rr) has lag-1 autocorrelation ≈ 0 for every feature, and a circular-shifted persistent feature gives a naïve/HAC FPR ratio of 1.0. BTC 3s-returns are near-white at lag 1, so the HAC correction is not load-bearing at the bar level here.
The necessity + H≈.79 envelope live at the PERIOD-IC series. The per-period (100-bar) IC series of a real predictive feature is long-memory: vwap H_raw 0.809 ≈ the §7 target 0.79 (shuffled 0.555), and the naïve-over-HAC understatement reaches ≥2× at period-lag 30–50 (vwap 2.19@30, all three ≥2.16@50).

feature	mean period-IC	H_raw	H_shuffled	naïve/HAC @lag30	@lag50
vwap_close_deviation	−0.625	0.809	0.555	2.19	2.57
ofi	+0.108	0.736	0.553	1.90	2.16
aggression_ratio	+0.140	0.726	0.560	1.96	2.24

Resolution (next firing, pre-registered)

Rewrite the significance test to the period-IC series (HAC-t on the 750-length IC series with a frozen period-lag); null-FPR size ≤.01 via block-perm null, the ≥2× naïve-over-HAC necessity on the real persistent period-IC.
Freeze the period-lag (lag 30 → vwap 2.19; lag 50 → all ≥2.16) with a long-memory (H=0.81) justification — no lag-shopping.
N_min(H) at the period level: periods to detect vwap's signal, raw (H 0.81) vs shuffled (H 0.55) → ratio ≥2.
Balanced interaction-only exemplar; seed-robustness sweep before GROUNDED.