ML Liquidity

Screening every unused column in features_full_v2 against the base 78 found exactly ONE family that lifted beta-HEDGED Sharpe while beta FELL: the 12 liquidity columns (hedged +0.36 → +0.99, beta 0.44 → 0.39). This book forward-tests that, at the slowest rebalance that still clears a pre-registered bar.

• What it survivedrandom-block control — 14 random 12-column blocks from the unused pool, liquidity at the 100th percentile, z=+3.72 (null mean −0.03, max +0.24). Same control that killed the earlier "options carry the ml edge" claim.
• Cost gatelift essentially FLAT from 10bp to 100bp one-way (+0.63/+0.63/+0.62/+0.60); median ADV of picks moved only $12.1M → $11.0M (0.91x). At 100bp the base book goes NEGATIVE (−0.20) while this stays +0.41.
• What carries it`illiquidity` (Amihud) is the single necessary column: +0.27 added alone, −0.21 when dropped. Cross-sectional IC −0.029 (t=−6.6) — ordinary predictor magnitude, not a leak.
• ⚠ What it FAILEDcross-model transfer: adding these columns to precision25's 79 features lifted hedged by only +0.09. And Deflated Sharpe charged for the day's ~145 selection events is 0.135 against a 0.95 gate. This is NOT a validated edge.
• Why run it anywayDSR says "unprovable on this dataset after that much searching", not "false". The only instrument that settles it is forward data — hence a zero-risk paper book from 0.00%, not funding.
• Config rule (fixed before the sweep)forward-test only if some rebalance ≥ 10d keeps lift ≥ +0.40 at ≤ 160 legs. 5d/301 legs/+0.65, 10d/152/+0.63, 21d/74/+0.49 — 21d chosen as the fewest legs that clears ($136/leg on $10k).