XGBoost Bank Nifty Feature Engineering

Bank Nifty XGBoost models fail or succeed on features. Retail copy-paste indicators underperform because they ignore index-specific structure: weekly expiry, Thursday settlement, OI concentration at ATM +-2 strikes, and a retail-heavy participant base that creates repeatable sentiment in the Option Chain. This page documents a tested feature set: atm_iv (ATM implied volatility), pcr_strikes (put-call ratio across top 5 strikes), oi_buildup (change in OI at ATM +-2, direction of money), vix_india (India VIX z-score), theta_decay_rate (avg theta per day to expiry), max_pain_distance (spot distance from max-pain strike). Each feature is computed from the Dhan/retail feed daily, lagged by 1 bar to avoid leakage. The predictor outputs P(direction | features) per strike; a risk filter then decides action. Code: import lightgbm as lgb; model = lgb.LGBMClassifier(n_estimators=400, learning_rate=0.05, max_depth=5); model.fit(X_train, y_train). The point is not the model but the features: get these right and even a shallow tree ranks Nifty direction better than a blind option buy.