PPO Model for XAUUSD Gold Trading

This repository contains a Reinforcement Learning model trained using Proximal Policy Optimization (PPO) for trading XAUUSD (Gold vs US Dollar) on 15-minute timeframes.

Model Details

• Model Type: PPO (Proximal Policy Optimization)
• Framework: Stable-Baselines3
• Environment: Custom Gym environment for XAUUSD trading
• Training Data: Historical XAUUSD data from 2004 to 2025 (resampled to 15-min bars)
• Total Timesteps: 1,000,000
• Position Sizing: Base 5.0 oz, Max 7.5 oz
• Initial Capital: 200 USD
• Transaction Cost: 0.65 USD per oz

Performance Metrics (Test Set)

• Average Daily Profit: 51.46 USD
• Win Rate: 69.0%
• Max Drawdown: 12.0%
• Sharpe Ratio: 7.56
• Average Trades per Day: 2.66

Features Used

• Log Return
• RSI (14-period)
• Moving Averages (short/long)
• Bollinger Bands
• MACD
• Volume indicators

Source Code

• GitHub: https://github.com/JonusNattapong/Reinforcement-Learning-for-Gold-Trading

Usage

Loading the Model

Below are two safe ways to load the trained policy depending on what you have available.

Option A — Load the full Stable-Baselines3 model (.zip)

from stable_baselines3 import PPO
from stable_baselines3.common.vec_env import VecNormalize
import os

# Create or reconstruct an environment similar to the one used for training
# e.g. `env = make_your_env(...)` — replace with your env factory
env = ...

# If you saved VecNormalize separately, load and wrap your env first
if os.path.exists("models/vecnormalize.pkl"):
	vec = VecNormalize.load("models/vecnormalize.pkl", env)
	vec.training = False
	vec.norm_reward = False
	env = vec

# Load the full model (policy + optimizer state)
model = PPO.load("models/ppo_xauusd.zip", env=env)

Option B — Load weights saved as SafeTensors into a fresh PPO policy

from safetensors.torch import load_file
import torch
from stable_baselines3 import PPO
from stable_baselines3.common.vec_env import VecNormalize
import os

# Create or reconstruct the same environment used for training
env = ...

# If you have VecNormalize statistics, load them and wrap the env
if os.path.exists("models/vecnormalize.pkl"):
	vec = VecNormalize.load("models/vecnormalize.pkl", env)
	vec.training = False
	vec.norm_reward = False
	env = vec

# Instantiate a PPO model with the same policy architecture
model = PPO("MlpPolicy", env)

# Load SafeTensors state dict and convert values to torch.Tensor if needed
raw_state = load_file("models/ppo_xauusd.safetensors")
state_dict = {k: (torch.tensor(v) if not isinstance(v, torch.Tensor) else v) for k, v in raw_state.items()}

# Load weights into the policy
model.policy.load_state_dict(state_dict)

# Ensure the model has the same env wrapper
model.set_env(env)

Notes:

• Option A is preferred when ppo_xauusd.zip is available (it contains the entire SB3 model).
• Option B is useful when only the policy weights were exported as SafeTensors. Ensure the policy architecture and observation/action spaces match the original training setup.
• Always set vec.training = False and vec.norm_reward = False when running inference.

For Full Inference

To use the model for trading, you'll need to:

1. Set up the trading environment (XAUUSDTradingEnv)
2. Load VecNormalize stats
3. Run predictions

Note: This is a simulation model. Use with caution in real trading.

Training Configuration

• Learning Rate: 0.0003
• Batch Size: 256
• Gamma: 0.99
• GAE Lambda: 0.95
• Clip Range: 0.2
• Entropy Coefficient: 0.01

Files

• ppo_xauusd.safetensors: Model weights in SafeTensors format
• vecnormalize.pkl: VecNormalize statistics for observation normalization

License

MIT License

Disclaimer

This model is for educational and research purposes only. Trading involves risk, and past performance does not guarantee future results. Always backtest and validate before using in live trading.