LightGBM for Nifty 50: Complete Indian Market Guide

Nifty 50 with LightGBM

Nifty 50 is India's benchmark index. LightGBM's speed makes it ideal for rapid development and tuning on Nifty data. This guide covers Indian market-specific features and implementation.

Indian Market Features

def create_nifty_features(nifty_data, option_chain_data, vix_data):
    """Create Nifty-specific features."""
    features = pd.DataFrame(index=nifty_data.index)
    
    # Price features
    features['nifty_returns_1d'] = nifty_data['Close'].pct_change()
    features['nifty_returns_5d'] = nifty_data['Close'].pct_change(5)
    features['nifty_returns_20d'] = nifty_data['Close'].pct_change(20)
    
    # Technical indicators
    features['rsi_14'] = calculate_rsi(nifty_data['Close'], 14)
    features['macd'] = calculate_macd(nifty_data['Close'])
    features['bb_width'] = calculate_bollinger_band_width(nifty_data['Close'])
    
    # India VIX
    features['vix_level'] = vix_data['Close']
    features['vix_change'] = vix_data['Close'].pct_change()
    features['vix_rank'] = vix_data['Close'].rolling(252).rank(pct=True)
    
    # Option chain features
    features['pcr'] = option_chain_data['put_oi'] / option_chain_data['call_oi']
    features['max_pain'] = option_chain_data['max_pain']
    features['distance_to_max_pain'] = (nifty_data['Close'] - option_chain_data['max_pain']) / nifty_data['Close']
    
    # Global features (SGX Nifty)
    features['sgx_nifty'] = get_sgx_nifty()  # Singapore Nifty futures
    features['us_futures'] = get_us_futures()  # S&P 500 futures
    
    # FII/DII data
    features['fii_flow'] = get_fii_flow()  # Foreign Institutional Investors
    features['dii_flow'] = get_dii_flow()  # Domestic Institutional Investors
    
    return features

LightGBM Configuration for Nifty

params = {
    'objective': 'binary',
    'metric': 'auc',
    'boosting_type': 'gbdt',
    'num_leaves': 15,
    'max_depth': 4,
    'learning_rate': 0.02,
    'feature_fraction': 0.8,
    'bagging_fraction': 0.8,
    'bagging_freq': 5,
    'lambda_l1': 0.1,
    'lambda_l2': 1.0,
    'min_child_samples': 30,
    'verbose': -1
}

# Walk-forward training
scores = []
for i in range(252, len(X) - 1):
    X_train = X[:i]
    y_train = y[:i]
    X_val = X[i-20:i]
    y_val = y[i-20:i]
    X_test = X[i:i+1]
    y_test = y[i:i+1]
    
    train_data = lgb.Dataset(X_train, label=y_train)
    val_data = lgb.Dataset(X_val, label=y_val)
    
    model = lgb.train(
        params,
        train_data,
        num_boost_round=1000,
        valid_sets=[val_data],
        callbacks=[lgb.early_stopping(50), lgb.log_evaluation(0)]
    )
    
    pred = model.predict(X_test)
    scores.append(pred[0])

Walk-Forward Results

LightGBM on Nifty 50 (2020-2025):

• Walk-forward AUC: 0.57 ± 0.03
• Training time: 18 seconds per fold
• Total training time: ~10 minutes

Trading Signals

def generate_nifty_signals(model, features, threshold=0.58):
    """Generate Nifty trading signals."""
    proba = model.predict(features)
    
    if proba > threshold:
        return 'BUY_CE'  # Buy Call Option
    elif proba < (1 - threshold):
        return 'BUY_PE'  # Buy Put Option
    else:
        return 'NO_TRADE'

SEBI Compliance

This article is for educational purposes only. Trading in Nifty options involves substantial risk of loss. The author, Shakti Tiwari, is NISM-Series-XII certified. Past performance does not guarantee future results. All trading decisions are your own responsibility.

Sector Weighting Features

Nifty 50 is not one market, it is a weighted bundle of sector moods. Feature engineering that respects the structure:

• Compute the index's sector-consolidated weights (financials, IT, energy, auto, FMCG) as rolling features, since the same Nifty level can be built by very different sector pictures.
• Add the top-5 constituent's returns and the breadth ratio (advancers/decliners) to surface divergence between the index line and the bodies inside it.
• Include the Bank Nifty/Nifty ratio, a classic risk-on risk-off dial for the Indian session.

India Macro Features That Move the Tape

Nifty models that ignore macro logic memorise the calendar wrong. Useful, honest columns:

• Policy date proximity: days until the MPC, Budget, or election events, because IV and drift cluster around them.
• Rupee moves versus the dollar in rolling bins; a fast-import is a Nifty headwind that appears in the model as a rupee-feature interaction.
• G-Sec yield direction and the 10-year slope; rates repricing is the quiet hand under all valuation spread.

Expiry-Session Features

The Indian weekly-expiry structure leaves fingerprints an ML model learns quickly:

• Thursday session features: pre-15:00 drift, max-pain distance, and gamma positioning that pulls price toward OI-heavy strikes.
• Friday-open behaviour after expiry day: roll-over effects and the new-week IV reset.
• Time-of-day and day-of-week dummies: modest individually, persistently useful when the model races them against macro flags.

FII/DII Flows as a Crowd Slice

Indian markets are the rare venue where every day's institutional buying and selling is published. Use the data deliberately:

• FII net buys in cash and index futures as features, along with their rolling z-score; extreme FII selling extends crashes, extreme buying anticipates rebounds.
• DII net flows as the often-mean-reverting counterpart; the two are a natural pair of contrarian-plus-momentum columns.
• Combine without double counting: FII index-futures flow and cash flow are correlated and the model can burn importance on the duplication.

A Trend-Following Layer, Built to Survive

On Nifty 50, an honest LightGBM baseline is a trend-and-quality layer, not a magic reversion machine:

1. Target: next-session sign of the 5-day forward return, modelled on features from the families above.
2. Gate entries by the model's conviction (probability band) and the session window (avoid overnight on event days).
3. Trade only liquid instruments: Nifty futures or ATM-weekly options, with defined-risk spreads preferred for options legs.

Walk-forward results on Nifty 50 data reward model humility: the layers that last combine macro proximity, sector breadth and crowd-flow features with strict cost models and crisis-era stops. The index itself is a long-horizon business machine; the model's edge is calendar and crowd timing, and only when both are justified by the tape.