ML K5/H60

The deployed ml book uses 78 features with K=10 and a 20-day hold — a configuration inherited from research/110 and never tested against alternatives. A 75-cell K x hold map showed that cell sits in a poor region of its OWN surface. This book changes nothing but the construction: same 78 features, top-5/bottom-5, 60-day hold.

• ChangeK 10→5, hold 20→60 days. Features identical to deployed ml — no feature selection, so no selection risk in the comparison
• ⚠ LABEL LEAKAGE 2026-07-30The target is close.shift(-HOLD) but training selected date < d, so a row at d-30 carried a label built from prices at d+29 — inside the window being traded. No purge in any script. Re-run with training cut off HOLD sessions before each fold: candidate 1.90 → 0.76, deployed 1.41 → 0.46.
• Why it hit this book hardestthe bias scales with HOLD (a 60-day label leaks 60 days, a 20-day label 20), so it predicted the exact finding this book was built on. It also explains why the K/hold result failed to transfer to precision25 — real structure transfers, leak artifacts do not.
• Edge is now inside the noiseadvantage over the deployed cell fell +0.49 → +0.30, while the purged seed range is 0.34. With 3 seeds the two configs are indistinguishable. The original "3 seeds spread 0.01" stability claim was itself an artifact — leaked labels make predictions deterministic across seeds.
• Forward record IS validat live prediction time shift(-HOLD) is NaN for the last HOLD sessions, so dropna() removes exactly the leaking rows. The live model is naturally purged. This paper book is a sound forward test; it was simply advertised against a number it could never hit.
• Real-world corroborationthe deployed ml book: unpurged backtest +21%/yr, purged +6.9%, LIVE -8.92%. Reality is tracking the purged figure or worse.
• Honest caveatINCEPTION is 2026-07-30. An earlier 07-15 start was removed: the ETF denylist and 50%-move guard were written because 2026-07-15 showed SOXS +974%, so scoring that window would credit the book for a cleaning rule fitted to it. Earlier periods are backtest and were part of the data the K/hold cell was chosen on — they are deliberately NOT shown as live. Paper only, no orders.