ML Trading Bot - Panic Cop v2

A machine learning-based position sizing system that dynamically reduces exposure during drawdown conditions using a RandomForest classifier.

⚠️ Disclaimer

This software is for educational and research purposes only. Trading involves substantial risk of loss. Past performance does not guarantee future results. Never risk more than you can afford to lose.

Overview

This system implements a "panic cop" mechanism that:

1. Identifies Risk Conditions: Detects when the trading system is in a vulnerable state (drawdown, loss streaks)
2. ML-Based Decision Making: Uses RandomForest to predict when to reduce position size
3. Conservative Approach: Only acts during clearly defined risk scenarios
4. Statistical Validation: Validates edge through Monte Carlo simulation

Key Features

• Gated Activation: ML only activates during specific market regimes and risk thresholds
• Conservative Teacher Labels: Training labels based on deep drawdowns and extended loss streaks
• Probability Threshold: High confidence requirement (default 0.75) before size reduction
• Monte Carlo Validation: Statistical proof that the ML is providing edge vs random
• Configurable Parameters: All thresholds and settings externalized

Installation

Requirements

pip install -r requirements.txt

Python Version: 3.8+

Required Libraries

• pandas >= 1.3.0
• numpy >= 1.21.0
• scikit-learn >= 1.0.0

Usage

Basic Usage

python ml_panic_cop.py your_trading_data.csv

With Custom Output File

python ml_panic_cop.py your_trading_data.csv --output results.csv

With Custom Configuration

python ml_panic_cop.py your_trading_data.csv --config config.json

As a Module

from ml_panic_cop import PanicCopML

# Initialize with custom config
config = {
    "panic_mult": 0.50,
    "prob_threshold": 0.75,
    "dd_gate": -0.15,
}

bot = PanicCopML(config)
results = bot.run_full_pipeline("data.csv", "output.csv")

Data Requirements

Your input CSV must contain these columns:

Column	Description
date	Trade date (parseable datetime)
ret_pct	Return percentage per trade
size_final	Position size (e.g., 4x leverage)
q_confidence	Signal confidence metric
volume_ratio_x	Volume-based feature
regime_x	Market regime classification

Important: The system only activates during regime_x == "bull_trending"

Configuration

Default Parameters

{
  "panic_mult": 0.50,         // Reduce size to 50% when panic triggers
  "prob_threshold": 0.75,     // Require 75% ML confidence
  "dd_gate": -0.15,           // -15% drawdown enables panic eligibility
  "loss_gate": 6,             // 6 consecutive losses enables panic eligibility
  "teacher_dd": -0.20,        // -20% DD for training labels
  "teacher_loss": 3,          // 3 losses for training labels
  "teacher_bigloss": -0.05,   // Single -5% loss for training labels
  "train_split": 0.70,        // 70% train, 30% test
  "n_runs_mc": 2000,          // Monte Carlo simulations
  "rf_n_estimators": 300,     // Random Forest trees
  "rf_max_depth": 4,          // Tree depth (prevents overfitting)
  "random_state": 42          // Reproducibility seed
}

Creating a Config File

Create config.json:

{
  "panic_mult": 0.60,
  "prob_threshold": 0.80,
  "dd_gate": -0.12
}

How It Works

1. Feature Engineering

The system creates these features from your trading data:

• dd_4x: Current drawdown from equity peak
• loss_streak: Consecutive losing trades count
• roll_loss3: 3-trade moving average of sized returns
• roll_std5: 5-trade rolling standard deviation
• q_confidence: Your signal confidence metric
• volume_ratio_x: Your volume feature

2. Panic Eligibility Gate

ML can only activate when:

(drawdown <= -15% OR loss_streak >= 6) AND regime == "bull_trending"

3. Teacher Labels (Conservative)

Training examples labeled as "panic=1" when:

eligible AND drawdown <= -20% AND (loss_streak >= 3 OR single_loss <= -5%)

4. ML Prediction

• RandomForest trained on eligible trades only
• High probability threshold (75%) required
• Position size reduced to 50% when panic triggers

5. Monte Carlo Validation

Shuffles panic decisions randomly among eligible trades 2000 times to prove the ML selections beat random chance.

Statistical Significance:

• p-value < 0.05: Strong evidence of edge
• |z-score| > 1.96: Statistically significant at 95% confidence

Output

The system generates a CSV with these additional columns:

• eq_4x: Equity curve without panic
• dd_4x: Drawdown without panic
• loss_streak: Consecutive losses
• roll_loss3, roll_std5: Rolling features
• panic_eligible: Boolean for eligibility
• teacher_panic: Training labels
• panic_prob: ML probability of panic
• panic_flag_ml: Final panic decision
• size_ml_panic: Adjusted position size

Performance Metrics

The system outputs:

=== Performance Comparison ===
4x gas equity: 2.45, max DD: -32.5%
ML panic equity: 2.78, max DD: -28.3%

=== Monte Carlo Validation ===
Real ML equity: 2.78
Random mean: 2.52
p-value: 0.0123
z-score: 2.34

Interpretation Guide

Good Results

• ML equity > 4x equity
• ML max DD < 4x max DD
• p-value < 0.05
• |z-score| > 1.96

Warning Signs

• p-value > 0.20 (likely no edge)
• Very few panic triggers (< 1% of trades)
• No improvement in drawdown

Tuning Advice

If panic triggers too often:

• Increase prob_threshold (0.75 → 0.85)
• Tighten dd_gate (-0.15 → -0.18)
• Increase loss_gate (6 → 8)

If panic never triggers:

• Decrease prob_threshold (0.75 → 0.65)
• Loosen teacher thresholds
• Check that you have "bull_trending" regime periods

If overfitting:

• Reduce rf_max_depth (4 → 3)
• Increase rf_n_estimators (300 → 500)

Safety Considerations

What This Code Does NOT Do

❌ Connect to any broker or exchange
❌ Execute live trades
❌ Store credentials or API keys
❌ Access the internet
❌ Modify your original data files

What You Should Never Commit

• Actual trading data (CSV files)
• Broker credentials or API keys
• Trading results or performance data
• Trained models with real data

The .gitignore is configured to protect you, but always double-check before pushing.

File Structure

ml-trading-bot/
├── .gitignore              # Protects sensitive data
├── README.md               # This file
├── requirements.txt        # Python dependencies
├── ml_panic_cop.py         # Main ML system
├── config.json             # Configuration (optional)
└── example_config.json     # Example configuration

Development

Running Tests

# Test with example data
python ml_panic_cop.py example_data.csv --config example_config.json

Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Never commit real trading data
4. Submit a pull request

License

[Your chosen license - MIT, GPL, etc.]

Support

For issues or questions:

• Open an issue on GitHub
• Check existing issues for solutions

Acknowledgments

Built with:

• scikit-learn for ML
• pandas for data manipulation
• numpy for numerical operations

---

Remember: This is a research tool. Always paper trade extensively before considering live use. Never risk capital you can't afford to lose.