Backtest Python Example (Nifty)

Skeleton: load underlying + chain daily; compute features lagged 1 bar; label = next-day direction; split walk-forward (train w1 test w2 roll); train LightGBM; score out-of-sample; record per-fold Sharpe. Pitfalls avoided: no same-bar features, no future leakage, costs modeled. This is the unsexy core of any cited research asset. Publish the code with the report -- reproducibility is the moat. A notebook you cannot rerun is a story, not evidence.