---
project: anomaly-atlas
document: expH_oos_stability/analysis
author: Simon-Pierre Boucher
contact: contact@spboucher.ai
data_source: hfmarketdata.io
created: 2026-08-12
modified: 2026-08-12
status: reviewed
---

# Analysis — expH_oos_stability

Run: `results/expH_oos_stability/20260812T072518Z/results.json`. The
validation split (2016-2021), frozen before the first scan, was opened here
for the first time (418 requests, 18.8M rows). The holdout (2022→) remains
sealed. All four pre-registered questions resolved:

| Question | Train | Validation | Verdict |
|---|---|---|---|
| Q-b nothing survives κ=1 | 0 survivors | **0 survivors** | negative finding **replicates** |
| Q-a CKX 1day net@0.25 | κ*=0.28 | +16.3 bp/d CI [1.4, 26.5], but 2019-21 CI [−19.5, +22.0] | **fails robustness clause** — not confirmed |
| Q-c ES→SPY fresh corr | −0.032 (FDR, splice-invariant) | **+0.003** (p=0.68) | **does not replicate** |
| Q-d daily reversal family | VR30 excess −0.143 | **+0.023** | crisis-regime story **confirmed** |

## Reading

1. **The ladder closes on the honest outcome.** Artifact nulls → search
   correction → costs → OOS: nothing survives all four. This is the
   charter's anticipated result ("a well-documented 'it's all artifacts'
   is a success"), now demonstrated rather than assumed — and it becomes
   the atlas's first Level-2 entry.
2. **CKX was exactly what it looked like.** The lone cost survivor decays
   within validation (κ* 0.445 → 0.263), fails its pre-registered
   two-sub-period clause, and lives in a name whose capacity rounds to
   zero. Not escalated; the sealed holdout is not spent on it.
3. **ES→SPY is a lesson in regime-fragility**: a splice-invariant,
   FDR-significant train effect that flips sign OOS. Splice-invariance
   guards against one artifact — it says nothing about stationarity.
4. **The artifact replicates even though the anomalies do not.** SPY still
   leads SPX on validation (+0.080 vs +0.132 in train) — content staleness
   is structural. That contrast (artifacts stable, "alphas" not) is the
   project's cleanest single sentence.

## Hand-off

Two atlas entries created via `tools/new_finding.py` (F001 negative
finding, Level 2; F002 SPX artifact demonstration, Level 2). Candidate 02
(turn-of-month) and candidate 01 (intraday momentum) keep their protocols
for future work; the holdout stays sealed until a claim earns it.