Vittus Mikiassen
7a419eb6fc
Added weekly PPO training performance statistics and observations for the training period from June 2024 to June 2026.
4.7 KiB
MT5 XAUUSD LSTM PPO Trading Bot
A MetaTrader 5 reinforcement learning trading bot for XAUUSD (Gold).
The bot combines an LSTM neural network for sequence learning with Proximal Policy Optimization (PPO) to learn trading decisions directly from historical market data.
Unlike traditional bots that rely on fixed rules, the PPO agent learns when to Buy, Sell or Hold from thousands of market examples.
The program should be in a folder or the desktop where the 5m csv file with OHLC data from https://www.kaggle.com/datasets/novandraanugrah/xauusd-gold-price-historical-data-2004-2024 is, in training it will then make a LSTM-PPO-saves folder where the training is saved (used for making decisions, also in testing).
Features
Reinforcement Learning
- LSTM policy network
- PPO training
- Continuous online training
- Live inference on MT5
- Automatic checkpoint saving/loading
Technical Indicators
- EMA 7
- EMA 21
- EMA Difference (Momentum)
- ADX
- +DI
- -DI
- Stochastic
- VWAP
- VWAP Bands
- VWAP Slope
- Volume Moving Average
Smart Money Concepts
- Bullish Order Blocks
- Bearish Order Blocks
- Bullish Fair Value Gaps
- Bearish Fair Value Gaps
- Bullish Rejection Blocks
- Bearish Rejection Blocks
- Equal Highs
- Equal Lows
- Market Breaks
- Indecision Candles
PPO State Features
Current state contains:
- OHLC
- EMA trend
- Momentum
- ADX trend strength
- DI Direction
- Stochastic
- VWAP
- VWAP Bands
- VWAP Position
- VWAP Slope
- Volume MA
- Order Blocks
- Fair Value Gaps
- Rejection Blocks
- Equal Highs/Lows
- Buy Score
- Sell Score
Current Strategy
Current reward structure:
- Take Profit: 20 pips
- Stop Loss: 40 pips
- Risk/Reward: 1 : 0.5
The current focus is high-probability momentum trades rather than large swing trades.
Training
The PPO agent trains continuously over historical MT5 data.
Example metrics during training:
- Win rate: 70–85%
- Profit Factor: 1.5–3+
- Weekly performance: typically 10–30R during training (varies by market conditions)
These figures are training statistics only and are not guarantees of future performance.
Quarterly stats
Weekly PPO Training Performance by Quarter
Training period: ~June 2024 – June 1, 2026 (102 rolling weekly training windows)
| Period | Approx. Dates | Avg Weekly R | Avg PF | Avg Max DD | Avg Recovery Factor |
|---|---|---|---|---|---|
| Q1 | Jun 2024 – Sep 2024 | 6.20R | 1.10 | 8.80R | 2.04 |
| Q2 | Sep 2024 – Dec 2024 | 7.39R | 1.12 | 9.80R | 2.25 |
| Q3 | Dec 2024 – Mar 2025 | 16.89R | 1.25 | 8.14R | 5.42 |
| Q4 | Mar 2025 – Jun 2025 | 43.73R | 1.29 | 10.81R | 9.20 |
| Q5 | Jun 2025 – Sep 2025 | 36.54R | 1.38 | 8.94R | 8.72 |
| Q6 | Sep 2025 – Dec 2025 | 75.07R | 1.38 | 8.62R | 19.42 |
| Q7 | Dec 2025 – Mar 2026 | 164.37R | 1.82 | 7.23R | 54.69 |
| Q8* | Mar 2026 – Jun 1, 2026 | 126.55R | 1.49 | 8.76R | 34.54 |
*Q8 contains the final 11 weeks of training.
Observations
- Profit Factor increased from approximately 1.10 during the earliest training period to 1.4–1.8 in the later periods.
- Average weekly drawdown remained relatively stable between 7R and 10R despite substantially higher returns.
- Recovery Factor improved significantly over time, indicating that profitability increased faster than drawdown.
- The strongest performance occurred during the final two quarters while maintaining comparable risk characteristics.
Requirements
- Python 3.11+
- MetaTrader 5
- MetaTrader5
- pandas
- numpy
- torch
Install dependencies:
pip install MetaTrader5 pandas numpy torch
Running
Train:
python mt5-xau-lstm-ppo-bot.py --train
Live trading:
python mt5-xau-lstm-ppo-bot.py --test
Train and trade simultaneously:
python mt5-xau-lstm-ppo-bot.py --train --test
To train download a m5 xauusd csv file form kaggle or dukascopy and place the folder in "download" relative to the directory the mt5-xau-lstm-ppo-bot.py is in.
Project Goals
Current work focuses on:
- Improving PPO policy learning
- Better feature engineering
- Dynamic trade management
- VWAP and volume analysis
- Smart Money Concept detection
- MFE/MAE prediction research
- Higher timeframe context
Disclaimer
This project is for educational and research purposes only.
Trading leveraged products involves substantial risk. Always test thoroughly on historical data and demo accounts before risking real capital.