📉 #3 effective-n deflation + BH-FDR GROUNDED

The honest-significance instrument: autocorrelated data has fewer independent points than it looks, so count them properly (HAC effective-n), then correct for many tries (Benjamini-Yekutieli FDR). It grounds on a family of 7 real features — and reproduces SEAL's independent measurement to the decimal.

35%

real N_eff / N (HAC deflation)

2.34×

naive SE understatement (= SEAL)

≤ 0.07

FDR on all nulls (BY, ≤ q)

4 / 20

instruments grounded

Preflight (resource-only): load1 0.22 · 43 GiB available · si/so ~0 · ClickHouse active, readonly=2. Compute wall ~4 s, RSS ~400 MB. Return is genuinely archetype-d: excess kurtosis 37.7, vol-clustering ACF(r²) 0.66.

In plain English

870,000 bars sounds like a lot of evidence — but consecutive bars move together, so they're worth far fewer independent observations. This instrument counts the real number (HAC), and it finds the real data is worth only about 35% of its face value. Ignoring that (the naive count) would inflate a feature's t-statistic by up to 2.34× — manufacturing confidence out of thin air. It then corrects for testing many features at once using the arbitrary-dependence FDR rule (Benjamini-Yekutieli), which is needed because the features move together.

Gate status — all pass

Gate	Result
Effective-n recovery	N_eff_HAC/N = 0.96 (iid shuffle) → 0.35 (real, deflates)	PASS
FDR control (HAC + BY)	complete-null rejection 0.043 / 0.057 / 0.067 (white/block/circular) ≤ q	PASS
Necessity (deflation)	real-aligned naive-SE understatement median 1.55; vwap 2.34 (= SEAL)	PASS
Power (known-positive)	real look-ahead detected; N_min 3,000	PASS
Known-duplicate	exact clone → identical IC & p-value	PASS
Envelope (U-shape)	even dependence marginal IC ~0 → FNR 1 (blind, route to #11/#18)	PASS

Two honest findings

(1) Benjamini-Yekutieli, not Hochberg. The 7 features are cross-correlated (shared microstructure); plain BH over-rejects erratically under that dependence, so §7's BH(-Y) (arbitrary-dependence FDR) is the correct, frozen procedure — and it controls the FDR (0.043–0.067 ≤ q).
(2) The GARCH-specific scalar failure is muted for a rank IC. The return is genuinely heavy-tailed + vol-clustered (archetype d), but ranks tame exactly those extremes, so the scalar AR(1) deflator does not fail dramatically. The clean, unmuted necessity is therefore naive-vs-HAC (the SE understatement) — recorded, not hidden.

Verify-before-report — principled corrections

HAC lag was recomputed on the 40k subsample (=17) → frozen Bucket-2 lag 30 (autocorrelation range is a data property, not a subsample-size one).
BH → BY (arbitrary-dependence FDR, §7) for the cross-correlated family.
Necessity framing → the SEAL-anchored real-aligned SE-understatement (deflation matters on the persistent signal, not the signal-free null).
Understatement on the full 870k series — the 40k subsample under-captured long-range autocorrelation (vwap 1.58 → 2.34 at full N, matching SEAL exactly).

Verdict & what's next

GROUNDED — the second realness-axis instrument.

Next iteration → #6 SFI single-feature OOS (usefulness / the per-feature return object): train on one feature over CPCV, does it beat a coin toss out-of-sample?