Institutional Intraday Trading System

Architecture

intraday_trader/
├── data/
│   ├── data_loader.py        OHLCV (yfinance + NIFTYBEES volume proxy)
│   └── data_cleaning.py      IST filter, OHLC fix, volume clip
├── features/
│   ├── feature_engineering.py  65 institutional features (no lookahead)
│   └── feature_selection.py    MI + XGBoost blended rank selection
├── labeling/
│   └── barriers.py           Triple-barrier (vol-adjusted) + meta-labeling
├── models/
│   ├── ensemble.py           XGBoost + LightGBM + LR blend (0.50/0.35/0.15)
│   ├── evaluation.py         Walk-forward fold summary
│   ├── train.py              Standalone training script
│   ├── predict.py            Standalone inference script
│   ├── labeling.py           Re-export of labeling primitives
│   └── saved/                Serialised model + features.json
├── strategy/
│   ├── signal_filter.py      Regime-aware threshold + EV + MTF + cooldown
│   ├── signal_generator.py   Meta-signal layer (gates primary signals)
│   ├── position_sizing.py    ADX-regime routing (vol-target vs fixed-frac)
│   └── risk_management.py    Pre-trade risk gate
├── risk/
│   ├── risk_manager.py       Sizing, trailing stop, daily halt, consec losses
│   ├── position_sizing.py    Fixed-fraction / vol-target / half-Kelly
│   ├── slippage.py           Base bps + volume-impact model
│   └── transaction_costs.py  Full Zerodha NSE intraday cost breakdown
├── backtest/
│   ├── backtester.py         Event-driven, slippage+brokerage+latency, trail stop
│   ├── performance.py        Sharpe, drawdown, expectancy, equity curve
│   ├── multi_asset.py        Portfolio Sharpe across universe
│   └── stress_test.py        Monte Carlo (200 permutations)
├── execution/
│   └── paper_trading.py      5-min poll loop, live signal + order sizing
├── reports/
│   └── report_generator.py   PNG equity curve + text summary
├── notebooks/
│   └── research.ipynb        EDA, feature correlation, model diagnostics
├── utils/
│   ├── config.py             Central config (all params here)
│   └── logger.py             Dual stdout + file logger
├── main.py                   Orchestrator (backtest|multi|diagnose|stress|paper|train)
└── requirements.txt

Quick Start

pip install -r requirements.txt

# 1. Diagnose first — check data quality and label distribution
python3 main.py --mode diagnose --symbol RELIANCE.NS

# 2. Full backtest with report
python3 main.py --mode backtest --symbol RELIANCE.NS

# 3. Multi-asset portfolio backtest
python3 main.py --mode multi

# 4. Monte Carlo stress test
python3 main.py --mode stress --symbol RELIANCE.NS

# 5. Paper trading (live 5-min loop)
python3 main.py --mode paper --symbol RELIANCE.NS

# 6. Train only (saves model)
python3 models/train.py --symbol RELIANCE.NS

# 7. Predict latest bars
python3 models/predict.py --symbol RELIANCE.NS --bars 10

Key Design Decisions

Data Pipeline

• NIFTY volume fix: ^NSEI has zero volume from yfinance. Fixed by fetching NIFTYBEES.NS (liquid ETF) as volume proxy — aligned and scaled 10×.
• IST timestamp filter: only 09:15–15:30 bars kept; pre/post market stripped.
• No leakage: all features use .shift(1) or bar-t only values.

Signal Assignment (symmetric dead zone)

• BUY if prob ≥ (1 − dynamic_threshold)
• SELL if prob ≤ dynamic_threshold
• HOLD otherwise
• Threshold is ADX-regime dependent: lower in trend, higher in range.

Features (65 total → 30 selected)

• Price structure: swing H/L, breakout strength, consolidation range, trend persistence
• Volume: spike follow-through, VW momentum, order-flow imbalance, exhaustion
• Volatility: ATR percentile, ratio (expansion), Hurst proxy, hvol
• Microstructure: OFI (close-low / range), tick imbalance, candle delta
• Context: VWAP deviation + slope, MTF EMA alignment (9/21/50/200), session fraction
• Market-wide: NIFTY return, rolling beta (30-bar), corr30, relative strength, idx_trending

Labeling

• Triple-barrier: pt_mult=1.5×ATR, sl_mult=1.0×ATR, max_hold=12 bars
• Timeout bars → NaN (not forced to 0) — cleaner signal
• Fallback to 3-bar forward return if positive rate < 10%
• Meta-labeling available: primary model → direction; meta model → whether to trade

Ensemble

• XGBoost + LightGBM + Logistic Regression (blend 0.50/0.35/0.15)
• Per-fold scale_pos_weight from actual class ratio
• Threshold tuned per fold via F1 maximisation (range 0.40–0.80)
• Isotonic calibration on 15% holdout (if ≥500 samples)

Regime-aware filtering

• ADX ≥ 25 → trend mode → threshold = 0.52 (more trades)
• ADX < 25 → range mode → threshold = 0.62 (more selective)
• MTF alignment score must be ≥ 2/3 for BUY, ≤ 1/3 for SELL
• EV filter: prob × 1.5 − (1−prob) × 1.0 > min_ev
• Volume filter: vol_ratio ≥ 0.5 × 20-bar avg
• Cooldown: 3 bars after each trade

Risk Management

• Position size: (capital × 1%) / (1.0 × ATR); capped at 10% of capital
• Trailing stop: ratchets up by 0.8 × ATR as price moves favourably
• Daily halt: triggers at −2% daily P&L
• Consecutive loss halt: triggers at 4 losses in a row
• EOD close-out: all positions closed at session end

Transaction Costs (Backtester)

• Slippage: 5 bps base + volume-impact k√(order/ADV)
• Brokerage: min(₹20, 0.03%) per side (Zerodha)
• STT: 0.025% on sell side; Stamp: 0.003% on buy side
• Execution latency: 1 bar (signal on bar T → fill on bar T+1 open)

Performance Metrics Output

• Sharpe ratio (annualised, daily P&L basis)
• Max drawdown
• Win rate
• Expectancy (₹ per trade)
• Profit factor
• Equity curve PNG + text report in reports/