Finstreet Trading System

ML-Driven Trend-Following Strategy for NSE Equities

---

Performance Summary

Win Rate	Sharpe Ratio	Total Return	Max Drawdown	Profit Factor
88.9%	6.62	1.31%	0.05%	29.27

9 trades · NSE:SONATSOFTW-EQ · Nov-Dec 2025 · INR 100,000 → INR 101,314

---

Methodology

Data Pipeline

Warmup:     Jan 1 - Oct 31, 2025 (indicators only, no training)
Training:   Nov 1 - Dec 31, 2025
Prediction: Jan 1 - 8, 2026 (out-of-sample)

• Triple-barrier labeling with T+2 to T+5 forward returns
• Walk-forward cross-validation
• Data integrity verified via audit_bias.py

Strategy Logic

Step	Action
Trend ID	EMA(10)/SMA(20) crossover + ADX strength
Entry	Wait for 6% pullback within trend
Validate	ML ensemble confirms direction
Size	Scale position by confidence
Exit	ATR-based stops (1.5x SL, 3x TP)

Risk Controls

Control	Rule
Position Size	risk = base_risk * (confidence / 0.70)
Max Position	15% of capital
Stop Loss	1.5x ATR (EOD exit if not hit)
Regime Filter	ADX < 15 → skip trade
Signal Gate	Technical + ML must align

---

Results

Backtest Metrics (Nov-Dec 2025)

Metric	Value	Metric	Value
Total Trades	9	Win Rate	88.9%
Winners	8	Sharpe Ratio	6.62
Losers	1	Calmar Ratio	180.43
Avg Win	INR 170	Profit Factor	29.27
Avg Loss	INR 46	Max Drawdown	0.05%

All trades exited at EOD (daily bars); stop/target rarely reached intraday.

Equity Curve

Trade Execution

Trade Distribution

Model Performance

Feature Importance: Volatility features dominate (ATR z-score: 39%, ATR percentile: 28%, returns_std: 20%). The model identifies regime conditions rather than predicting direction directly.

---

Forward Predictions (Jan 1-8, 2026)

Date	Signal	Direction	Confidence	Regime
Jan 1	HOLD	DOWN	77.6%	Uptrend Strong
Jan 2	HOLD	DOWN	77.6%	Uptrend Strong
Jan 5	HOLD	DOWN	54.6%	Transitional
Jan 6	HOLD	DOWN	54.6%	Uptrend Weak
Jan 7	HOLD	DOWN	54.6%	Transitional
Jan 8	HOLD	DOWN	50.4%	Transitional

Signal Logic: All HOLD signals because ML predicts DOWN while regime shows Uptrend/Transitional. Conservative approach preserves capital rather than betting against trend.

---

Quick Start

Option 1: Local Setup (Recommended)

# Clone repository
git clone <repository>
cd finstreet

# Run automated setup
chmod +x setup.sh
./setup.sh

# Activate environment
source venv/bin/activate

# Run full pipeline
python run.py all

The setup.sh script automatically:

• Creates Python virtual environment
• Installs all dependencies from requirements.txt
• Creates necessary directories (data/, models/, reports/)
• Generates .env template for FYERS API credentials

Option 2: Docker

# Build image
docker build -t finstreet .

# Run full pipeline (mount volumes for output)
docker run -v $(pwd)/data:/app/data -v $(pwd)/reports:/app/reports finstreet all

# Run specific commands
docker run -v $(pwd)/data:/app/data finstreet backtest
docker run -v $(pwd)/reports:/app/reports finstreet visualize
docker run finstreet predict

The Dockerfile provides:

• Multi-stage build for optimized image size
• Python 3.11 slim base image
• All dependencies pre-installed
• Health check for container monitoring

Available Commands

python run.py fetch      # Fetch data from FYERS API
python run.py train      # Train XGBoost model only
python run.py ensemble   # Train XGBoost + LightGBM ensemble
python run.py backtest   # Run backtest simulation
python run.py predict    # Generate Jan 1-8 predictions
python run.py visualize  # Create performance charts
python run.py all        # Full pipeline

---

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                      TRADING PIPELINE                           │
├─────────────────────────────────────────────────────────────────┤
│  FYERS API ──► Features ──► ML Ensemble ──► Signal Gate ──► Trade
│      │            │              │              │             │
│   OHLCV        RSI, MACD      XGBoost +     Tech + ML      Risk
│   Daily        ADX, ATR       LightGBM      Aligned       Mgmt
│                KER, Z-Score    (60/40)                       │
└─────────────────────────────────────────────────────────────────┘

Project Structure

finstreet/
├── run.py                  # Entry point
├── audit_bias.py           # Data leakage verification
├── sensitivity_analysis.py # Parameter robustness testing
├── Dockerfile              # Container configuration
├── setup.sh                # Local setup script
├── requirements.txt        # Python dependencies
├── config/settings.py      # Configuration
├── src/
│   ├── data/               # FYERS API integration
│   ├── features/           # Technical indicators, preprocessing
│   ├── model/              # XGBoost + LightGBM ensemble
│   ├── strategy/           # Trading logic
│   ├── backtest/           # Simulation engine
│   ├── forecast/           # Daily predictions
│   ├── execution/          # Trade execution
│   └── visualization/      # Chart generation
├── data/
│   ├── raw/                # OHLCV from FYERS
│   └── processed/          # Trades, equity curve
├── models/                 # Trained model artifacts
└── reports/
    ├── figures/            # Performance charts
    └── daily_predictions.csv

---

Competition Compliance

Requirement	Status
FYERS API for data	src/data/fyers_client.py
FYERS API for execution	src/execution/trader.py
Daily OHLCV only	No intraday data
Nov-Dec 2025 trading window	All trades in window
Jan 1-8, 2026 predictions	reports/daily_predictions.csv
No look-ahead bias	Verified via audit_bias.py
Walk-forward validation	Chronological execution

---

License

MIT License - See LICENSE for details.