feat(research): broaden+overlay deploy-book validation — walk-forward + capacity head-to-head vs liquid-61
Walk-forward stress-test of Step 3's single-split lead: is a BROAD universe +
spread-aware overlay a deploy-book UPGRADE over the live liquid-61 ($10M) Bybit
book? READ-ONLY SHADOW eval — never mutates the live book.
- broaden_overlay_validate.py: expanding-window folds (knobs+breadth selected on
TRAIN, evaluated on disjoint TEST → no leakage); per-fold head-to-head live vs
broad+overlay NET; bottleneck deployable-capital estimator (capacity ratio
invariant to participation_frac); corr + concentration robustness; verdict
UPGRADE / CAPACITY-TRADEOFF / OVERFIT / NO.
- CLI `fxhnt broaden-overlay-validate` (--min-dollar-vols, --live-bound,
--n-folds, --participation-frac): per-fold table + aggregate + capacity verdict.
- Reuses spread_overlay (engine + live per-coin weights), bybit_liquidity_sweep
(metrics), liquid_universe. 17 TDD tests (no-leakage folds, baseline
reconciliation, capacity monotonicity, verdict falsification, WriteTripwire,
determinism, CLI). Full suite 1593 passed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
03c356d5d5