spb/anomaly-atlas Public License

1---2project: anomaly-atlas3document: expD_leadlag_scan/hypothesis4author: Simon-Pierre Boucher5contact: contact@spboucher.ai6data_source: hfmarketdata.io7created: 2026-08-128modified: 2026-08-129status: final10---1112# Hypothesis — expD_leadlag_scan1314*Pre-specified 2026-08-12 before the scan ran. Alignment primitives15(interval-fresh masks, both-fresh de-artifacting) passed the §8.1 gate16first — the gate caught an end-freshness-only mask bug (Epps attenuation)17before real data.*1819```text20Hypothesis21  Tests H07/H08 (scan half) and previews H09/H11. Expectations:22  (i) raw LOCF joins will show SPY "leading" everything, in proportion to23  staleness (T3); (ii) on both-fresh subsamples, liquid and sector-ETF24  pairs will show NO lead-lag at ±1..3 min in 2014-2015 (H0825  expected-negative); (iii) 2006-2007 fresh pairs MAY retain a small lead26  (pre-decay regime, Chordia et al. 2005); (iv) ES→SPY: any 1min-visible27  lead is expected to be tiny or absent, and must be invariant to the three28  futures splice variants; (v) SPX→SPY: the raw lead persists even29  both-fresh (index prints are CONTENT-stale though print-fresh) — the H1130  artifact case.3132Falsification criterion33  (ii) falsified if any liquid/sector both-fresh pair survives FDR(5%) at34  |lag| ∈ 1..3 in 2014-2015. (iv) falsified if the ES→SPY fresh lead35  changes sign/significance across splice variants (that would mean the36  measurement is splice-dominated, not economic).3738Artifact null(s)39  T3 non-synchronicity: measured DIRECTLY as artifact share =40  raw_xcorr[+1] − fresh_xcorr[+1] per pair. Day-block bootstrap null41  (resampled days, n=200, seed=42) for fresh xcorr[±1] vs 0.4243Method (grid pre-declared)44  Pairs vs SPY: 11 liquid (AAPL MSFT NVDA AMZN GOOGL META TSLA JPM XOM UNH45  QQQ), 9 sector ETFs (XLF XLE XLK XLV XLI XLY XLP XLU XLB), 10 random46  (first 10, sorted, of the seed-42 random-30). Plus ES→SPY (all three47  splice variants) and SPX→SPY. Windows: W1 2006-01-01→2008-01-01,48  W2 2014-01-01→2016-01-01 (ES/SPX: W2 only — history starts 2008).49  1min RTH grids; raw-LOCF vs both-fresh (interval-fresh) treatments;50  lags ±3; min 200 covered days and 10k fresh pairs per cell.51  Correction: BH-FDR over all fresh xcorr[+1] and xcorr[-1] tests jointly.52  Cell count ≤ (30×2 + 4 + 1) ≈ 65 pairs-windows × 2 lags.5354Result55  Run 20260812T070742Z (198 requests, 8.4M rows, 49 pair-cells, 98 tests).56  2006-2007: broad both-fresh leads survive FDR — SPY leads every sector ETF57  by +0.07..+0.148 at +1min and liquid names by +0.01..+0.03. 2014-2015:58  collapse — liquid/sector fresh leads shrink to [-0.020, +0.046]; a handful59  remain FDR-significant at tiny magnitudes (XLF +0.028, JPM +0.046,60  MSFT -0.015 NEGATIVE); sparse names keep larger fresh leads (BKE +0.075,61  HTD +0.073). ES->SPY: fresh corr(SPY_{t-1}, ES_t) = -0.032, FDR, and62  IDENTICAL across all three splice variants (-0.0323 each). SPX->SPY: SPY63  leads SPX +0.132 both-fresh with artifact share ~0 (SPX prints every64  minute, fresh_frac 0.996) — synchronization does NOT remove it. Median65  raw-vs-fresh artifact share ~0 for liquid/sector; for ultra-stale names66  the raw LOCF join DILUTES the lead (HTD raw +0.048 vs fresh +0.073).6768Interpretation69  (Level 0.) Clause (iii) held: 2006-2007 retained real minute-scale70  diffusion — and its 2014-2015 collapse is the cleanest decay measurement71  of the project so far (XLK +0.148 -> +0.019). Clause (ii) FALSIFIED as72  pre-registered: several liquid/sector pairs survive FDR in 2014-2015 —73  statistically nonzero but economically tiny residual structure; to expF.74  Clause (iv) held: the ES->SPY effect is splice-invariant, a genuine75  basis-reversion-like microstructure candidate. Clause (v) held: the SPX76  "lead" is content-staleness, not print-staleness — synchronizing on print77  times cannot fix a computed index (H11 candidate confirmed in scan form).78  New T3 nuance: LOCF joins INFLATE leads for mid-stale names (expB) but79  DILUTE them for ultra-stale names (zero-return mass) — the artifact is80  non-monotone in staleness.8182Next experiment83  expE calendar scan (pre-counted budget); then expF correction battery.84```85