Exploration of Machine Learning in Quantitative Trading: Practical Research on PPO, SAC, and XGBoost

This project is a systematic quantitative trading research repository focusing on exploring the application of reinforcement learning (PPO, SAC) and traditional machine learning (XGBoost) in trading strategies. It adopts a three-stage research process: exploration, validation, and implementation, including feature engineering experiments, forward backtesting, and signal generation tests, providing a complete workflow paradigm for quantitative trading research.

In the field of quantitative trading, the application of machine learning technology in market prediction and strategy generation is a popular direction. This project is developed by racoope70, with a tech stack covering mainstream algorithms. It adopts a three-stage research process: exploration (rapid iterative experiments), validation (backtesting evaluation), and implementation (production-level code refactoring), enabling a smooth transition from research to application.

Technical Models:

• Reinforcement Learning: PPO (Policy Gradient Method, ensuring training stability), SAC (Maximum Entropy Framework, balancing exploration and exploitation);
• Traditional Machine Learning: XGBoost (Gradient Boosting Tree, suitable for modeling financial time-series data). Research Methodology:

1. Exploration Phase: Feature engineering and model prototype development;
2. Validation Phase: Forward backtesting, historical data backtesting, simulated trading evaluation;
3. Implementation Phase: Refactoring into production-level code pipelines.

One of the core tasks of the project is feature engineering experiments, constructing rich features for financial time-series data:

• Technical indicators (moving averages, RSI, MACD, etc.);
• Statistical features (volatility, skewness, kurtosis, etc.);
• Time-series features (lag features, rolling window statistics, etc.);
• Signal generation tests (buy/sell signal trigger conditions). These features provide high-quality inputs for model training.

The reliability of quantitative strategies depends on the quality of backtesting. The project uses the forward backtesting method:

1. Rolling training window: Training the model with fixed-length historical data;
2. Out-of-sample testing: Evaluating performance using data after the training window;
3. Time progression: Moving the window to simulate real strategy updates. This method effectively detects overfitting and provides reliable performance estimates.

Research Value: The project's methodology clearly separates the three stages, maintaining exploration flexibility, establishing strict validation standards, and forming reusable code assets, providing a structured learning path for quantitative trading developers. Summary: This is a research-oriented open-source project that fully demonstrates the quantitative trading research process, and has important reference value for understanding the application boundaries and best practices of machine learning in the financial field.