Romeo V6 — High-Risk Ensemble Trading Model for XAUUSD

Model Details

Model Description

Romeo V6 is a high-risk ensemble machine learning model designed for predicting price movements in XAUUSD (Gold vs US Dollar) futures on 15-minute intraday data. It combines tree-based models (XGBoost and LightGBM) with an optional Keras neural network head to generate trading signals. The model outputs a probability score for long (up) trades, and the backtester handles entry/exit logic, position sizing, and risk management with aggressive settings for higher returns and volatility.

• Model Type: Ensemble Classifier (XGBoost + LightGBM + optional Keras NN) - High-Risk Configuration
• Asset: XAUUSD (Gold Futures)
• Strategy: Smart Money Concepts (SMC) with technical indicators
• Prediction Horizon: 15-minute intraday (next bar direction)
• Framework: Scikit-learn, XGBoost, LightGBM, TensorFlow/Keras

Model Architecture

• Ensemble Components:
  • XGBoost Classifier: Gradient boosting on decision trees.
  • LightGBM Classifier: Efficient gradient boosting with leaf-wise growth.
  • Optional Keras Neural Network: Dense layers with custom SumAxis1Layer to replace anonymous Lambda for serialization.
• Features: 31 canonical features including technical indicators (SMA, EMA, RSI, Bollinger Bands) and SMC elements (order blocks, volume profiles).
• Serialization: Tree models saved in joblib .pkl format; Keras model in native .keras format.
• Weights: Ensemble weights stored in artifact for weighted probability averaging.

Intended Use

• Primary Use: Research, backtesting, and evaluation on historical XAUUSD data.
• Secondary Use: Educational purposes for understanding ensemble trading models.
• Out-of-Scope: Not financial advice. Do not use for live trading without proper validation, risk controls, and regulatory compliance.

Factors

• Relevant Factors: Market volatility, economic indicators affecting gold prices (e.g., USD strength, inflation data).
• Evaluation Factors: Tested on unseen data; robustness scanned across slippage, commission, and threshold parameters.

Metrics

• Evaluation Data: Unseen 15m intraday data (out-of-sample).
• Metrics:
  • Initial Capital: 100
  • Final Capital: 162.31
  • Total Return: 62.31%
  • Total Trades: 207
  • Win Rate: 49.28%
  • Avg PnL per Trade: 0.3010

Training Data

• Source: Yahoo Finance (GC=F) historical data.
• Preprocessing: Feature engineering with technical indicators and SMC concepts.
• Split: Trained on historical data; evaluated on unseen fresh dataset.

Quantitative Analyses

• Robustness Scan: Tested under normal conditions (commission 0.02%, slippage 0.5 pips) and difficult market conditions (commission 0.05%, slippage 1.0 pips).
  • Normal: Win Rate 50.51%, Return 62.80%, Trades 198.
  • Difficult: Win Rate 31.82%, Return -16.38%, Trades 198.
• M2M Equity: Per-bar mark-to-market equity calculation for accurate risk metrics.

Ethical Considerations

• Bias: Model trained on historical data; may not account for future market changes or black swan events.
• Risk: High volatility in forex; potential for significant losses.
• Transparency: Full disclosure of assumptions, limitations, and evaluation.

Caveats and Recommendations

• Limitations: High-risk position sizing (5% risk per trade) leads to higher volatility and drawdown. Simplified position sizing; small-account behavior may differ with margin rules. Historical backtests not indicative of future results.
• Recommendations: Use with stop-loss, diversify, and consult financial advisors. Validate on your own data before use. This model is for high-risk tolerance traders.

Usage

Loading the Model

import joblib
artifact = joblib.load('trading_model_romeo_daily.pkl')
features = artifact['features']  # Canonical feature list
models = artifact['models']      # Dict of XGBoost/LightGBM models
weights = artifact['weights']    # Ensemble weights

Making Predictions

import pandas as pd
# Prepare df with features matching artifact['features']
X = df[features].fillna(0)  # Fill missing features with 0
probabilities = sum(weight * model.predict_proba(X)[:, 1] for model, weight in zip(models.values(), weights.values())) / sum(weights.values())
signals = (probabilities > threshold).astype(int)  # threshold e.g. 0.5

Backtesting

Use v6/backtest_v6.py with high-risk settings (risk_per_trade=0.05, threshold=0.4) on custom data. It aligns features automatically.

Requirements

• Python 3.8+
• scikit-learn, xgboost, lightgbm, tensorflow, joblib

Files

• trading_model_romeo_daily.pkl: Main artifact.
• romeo_keras_daily.keras: Optional Keras model.
• README.md: This model card.
• metadata.json: Structured metadata.

Contact

For issues or contributions: https://github.com/JonusNattapong/AITradings-samsam