Feature Leakage in Options ML

Leakage = using future info by mistake. In options ML it hides as: same-bar close as label with same-bar features, using settlement price in features, not lagging OI, including the traded bar in the training window. Audit: plot feature importance; if 'close' or 'next_oi' dominates, you leaked. Fix: lag every feature by 1 bar vs the label (next-day direction). Walk-forward with strict lagging exposes leakage immediately (out-of-sample collapses). Publishing the leakage audit is the moat -- it proves the result is real, not a bug wearing a PnL suit.