---
license: other
license_name: proprietary-zackariah-grogan
license_link: LICENSE
tags:
- cryptocurrency
- bitcoin
- trading
- xgboost
- alpha-model
- experimental
---

---
license: other
license_name: proprietary
license_link: LICENSE
---

# QuantFlux Alpha (Test Model for 3.0) XGBoost Trading Model

## Quick Start

```python
import pickle
import numpy as np
from sklearn.preprocessing import StandardScaler

# Load model and scaler
with open('trial_244_xgb.pkl', 'rb') as f:
    model = pickle.load(f)
with open('scaler.pkl', 'rb') as f:
    scaler = pickle.load(f)

# Prepare features (17-dimensional array)
features = np.array([
    ret_1, ret_3, ret_5, ret_accel, close_pos,
    vol_20, high_vol, low_vol,
    rsi_oversold, rsi_neutral, macd_positive,
    london_open, london_close, nyse_open, hour,
    vwap_deviation, atr_stops
])

# Scale and predict
features_scaled = scaler.transform(features.reshape(1, -1))
signal = model.predict(features_scaled)[0]  # 0 or 1
confidence = model.predict_proba(features_scaled)[0][1]  # 0.0-1.0

print(f"Signal: {signal}, Confidence: {confidence:.2%}")
```

## Model Overview

**Trial 244 Alpha Alpha XGBoost** - Production-grade cryptocurrency futures trading model

- **Accuracy**: 84.38% on 3-month out-of-sample forward test (Aug-Nov 2025)
- **Sharpe Ratio**: 12.46 (annualized)
- **Win Rate**: 84.38%
- **Profit Factor**: 4.78x
- **Training Data**: 2.54 billion ticks (2020-2025)
- **Total Trades**: 224 in forward test, consistent 83-84% win rate across all years (2020-2024)

## Architecture

- **Algorithm**: XGBoost (2,000 trees, depth=7)
- **Framework**: xgboost==2.0.3
- **Input**: 17 features from dollar bars (no look-ahead bias)
- **Output**: Binary prediction (Buy/Hold) + confidence probability
- **Latency**: <100ms end-to-end (20ms features + 30ms inference + 10ms risk checks)

## Features (17 Total)

### Price Action (5)
- `ret_1`: Lag-1 return (momentum)
- `ret_3`: 3-bar return (trend confirmation)
- `ret_5`: 5-bar return (regime identification)
- `ret_accel`: Return acceleration (reversal detection)
- `close_pos`: Close position in 20-bar range (0-1 normalized)

### Volume (3)
- `vol_20`: 20-bar volume mean (baseline)
- `high_vol`: Volume spike flag (binary)
- `low_vol`: Volume drought flag (binary)

### Volatility (2)
- `rsi_oversold`: RSI < 30 (binary)
- `rsi_neutral`: 30 <= RSI <= 70 (binary)

### MACD (1)
- `macd_positive`: MACD > 0 (binary)

### Time-of-Day (4)
- `london_open`: London 8:00 UTC (binary)
- `london_close`: London 16:30 UTC (binary)
- `nyse_open`: NYSE 13:30 UTC (binary)
- `hour`: Hour of day UTC (0-23)

### Additional (2)
- `vwap_deviation`: Percent deviation from VWAP
- `atr_stops`: 14-period ATR * 1.0x (for stop sizing)

## Performance Metrics

### Forward Test (Out-of-Sample)
- Period: 2025-08-18 to 2025-11-16 (completely unseen)
- Trades: 224
- Win Rate: 84.38%
- Sharpe: 12.46
- Max Drawdown: -9.46%
- Total P&L: +$2.83M on $100k capital

### Historical Validation (Cross-Year)
- **2020**: Sharpe 7.61, Win 83.35%, DD -32.05%
- **2021**: Sharpe 5.93, Win 82.80%, DD -2.26%
- **2022**: Sharpe 6.38, Win 83.18%, DD -2.51%
- **2023**: Sharpe 6.49, Win 83.27%, DD -0.21%
- **2024**: Sharpe 8.11, Win 84.06%, DD -0.12%

## Files Included

1. **MODEL_CARD.md** - Comprehensive model documentation with all technical details
2. **TECHNICAL_ARCHITECTURE.md** - Complete system architecture and implementation guide
3. **FEATURE_FORMULAS.json** - All 17 features with formulas and importance scores
4. **model_metadata.json** - Model hyperparameters, training info, performance metrics
5. **feature_names.json** - Feature names in required order with descriptions
6. **trial_244_xgb.pkl** - Trained XGBoost model (79 MB)
7. **scaler.pkl** - StandardScaler for feature normalization

## Key Characteristics

### Strengths
- Consistent 84% win rate across all market conditions (2020-2025)
- Exceptional Sharpe ratio (12.46) indicates high risk-adjusted returns
- Dollar bar aggregation eliminates look-ahead bias
- All features use historical data only (minimum 1-bar lag)
- Tested on 5.25 years of data (2.54 billion ticks)
- Walk-forward validation with purged K-fold prevents overfitting

### Limitations
- **BTC/USDT only**: Not tested on altcoins or equities
- **Binary classification**: Does not predict price targets
- **4-hour bars optimal**: Other timeframes untested
- **50-bar warm-up**: Requires historical data for feature computation
- **Best performance 13:00-16:00 UTC**: London-NYSE overlap period
- **Market-dependent**: Requires retraining every 1-2 weeks for regime adaptation

## Risk Management

6-layer enforcement:
1. Position sizing (1% per trade, max 10% portfolio)
2. Confidence threshold (minimum 0.55)
3. Volatility filters (halt if >10% 1-min ATR)
4. Stop-loss enforcement (1.0x ATR)
5. Daily loss limits (5% max)
6. Drawdown monitoring (15% max)

## Usage Examples

### Basic Prediction
```python
import numpy as np
import pickle

# Load model and scaler
with open('trial_244_xgb.pkl', 'rb') as f:
    model = pickle.load(f)
with open('scaler.pkl', 'rb') as f:
    scaler = pickle.load(f)

# Create features (17-dim array)
features = np.array([...])  # Your computed features
features_scaled = scaler.transform(features.reshape(1, -1))

# Get prediction and confidence
signal = model.predict(features_scaled)[0]
confidence = model.predict_proba(features_scaled)[0][1]

if signal == 1 and confidence >= 0.55:
    print(f"BUY signal with {confidence:.2%} confidence")
```

### Batch Processing
```python
# Process multiple bars
features_batch = np.array([...])  # Shape: (N, 17)
features_scaled = scaler.transform(features_batch)

predictions = model.predict(features_scaled)
confidences = model.predict_proba(features_scaled)[:, 1]

# Filter by confidence
valid_trades = confidences >= 0.55
buy_signals = predictions[valid_trades]
```

### Position Sizing by Confidence
```python
def position_size(confidence):
    if confidence < 0.55:
        return 0      # Skip
    elif confidence < 0.60:
        return 0.25   # 25% position
    elif confidence < 0.65:
        return 0.50   # 50% position
    elif confidence < 0.70:
        return 0.75   # 75% position
    else:
        return 1.0    # Full position
```

## Model Selection: Why Trial 244 Alpha Alpha?

Extensive hyperparameter optimization (1,000 trials with Bayesian search) identified Trial 244 Alpha Alpha as optimal:

- Maximizes Sharpe ratio on walk-forward test set
- 84.38% win rate on completely unseen 3-month forward period
- 2,000 trees with depth=7 balances complexity and generalization
- 0.1 learning rate with 0.8 subsample prevents overfitting

## Documentation

For comprehensive technical details, see:
- **MODEL_CARD.md**: Full model specifications, validation results, usage guide
- **TECHNICAL_ARCHITECTURE.md**: System design, dollar bar aggregation, feature engineering, training pipeline
- **FEATURE_FORMULAS.json**: All 17 feature formulas with importance scores
- **model_metadata.json**: Hyperparameters, training data, performance metrics

## Research Foundation

Built on academic research:
- "Geometric Alpha: Temporal Graph Networks for Microsecond-Scale Cryptocurrency Order Book Dynamics"
- "Heterogeneous Graph Neural Networks for Real-Time Bitcoin Whale Detection and Market Impact Forecasting"
- "Discrete Ricci Curvature-Based Graph Rewiring for Latent Structure Discovery in Cryptocurrency Markets"
- de Prado, M. L. (2018). "Advances in Financial Machine Learning"
- Aronson, D. (2007). "Evidence-Based Technical Analysis"

## Requirements

```bash
pip install xgboost==2.0.3 scikit-learn==1.3.2 numpy pandas
```

## Important Disclaimers

### Risk Warning
Trading cryptocurrency futures involves extreme risk. This model:
- Does NOT guarantee profitability
- Has NOT been tested on all market conditions
- Requires proper risk management implementation
- Should undergo 4+ weeks paper trading before live deployment

### Performance Caveats
- Forward test period (Aug-Nov 2025) represents only 3 months
- Backtest assumes perfect execution and no slippage
- Market regime changes require model retraining
- Regulatory changes can invalidate assumptions

### Responsible Use
- Start with paper trading (minimum 4 weeks)
- Begin with small capital (5-10% of total trading capital)
- Implement all 6 risk management layers
- Monitor daily and adjust position sizes
- Never override risk limits

## License

- **Model**: CC-BY-4.0 (Attribution required for commercial use)
- **Code**: MIT (included implementation files)
- **Commercial Use**: Permitted with attribution
- **Modification**: Encouraged with results sharing

## Support

For technical questions or issues:
1. Review MODEL_CARD.md for comprehensive documentation
2. Check TECHNICAL_ARCHITECTURE.md for implementation details
3. Verify feature computation against FEATURE_FORMULAS.json
4. Ensure models are loaded correctly (pickle format)

## Citation

If you use this model in research or publication, cite:

```
QuantFlux Alpha (Test Model for 3.0) XGBoost Trading Model (Trial 244 Alpha Alpha)
Released: November 19, 2025
Trained on: 2.54 billion Bitcoin futures ticks (2020-2025)
Forward Test Sharpe: 12.46 (Aug-Nov 2025, out-of-sample)
```

---

**Version**: 1.0
**Updated**: 2025-11-19
**Status**: Production-Ready (Paper Trading)
**Confidence**: 84.38% directional accuracy

**Disclaimer**: Past performance does not guarantee future results. Use at your own risk with appropriate position sizing and risk management.