GPy-Trade Generalist PPO Trading Agent
A hierarchical multi-asset generalist reinforcement learning agent for cryptocurrency trading built on top of Stable-Baselines3 (PPO), optimized with pre-trained LSTM trend classifiers, MLP volatility regressors, and an Isolation Forest anomaly detector.
Model Details
- Framework: PyTorch & Gymnasium
- Algorithm: PPO (Stable-Baselines3)
- Assets: BTCUSDT, ETHUSDT, SOLUSDT, BNBUSDT, XRPUSDT
- Base Models:
- Classifier: LSTM (Price direction prediction)
- Regressor: MLP (Future return volatility prediction)
- Anomaly Detector: Isolation Forest (Volatility filter)
- Observations: Standardized z-score technical indicators (RSI, MACD, Bollinger Bands, ATR, SMAs) and Cumulative Volume Delta (CVD) to capture order book flow, plus asset-specific
symbol_id.
Files included
ppo_trading_agent.zip: Trained PPO model weightsscaler.pkl: StandardScaler fitted on base featuresanomaly_detector.pkl: Isolation Forest weightsclassifier.pt: PyTorch LSTM direction classifierregressor.pt: PyTorch MLP volatility regressor
