This repository contains a deep neural network (DNN) model for predicting forex commodity price movements and a trading bot that uses the model with the OANDA API for live trading. The bot is designed to trade forex pairs such as USD/JPY, leveraging historical candlestick data and machine learning to make informed trading decisions.
Note: This project is for educational purposes only, and caution should be exercised when/if live trading.
- Overview
- Features
- Setup Instructions
- Usage
- Model Details
- Backtesting and Performance
- Live Trading
- Example
This project combines financial data analysis, machine learning, and automated trading. The DNN model predicts price direction using engineered features and a custom architecture. The trading bot integrates the model with the OANDA API for real-time trading.
- Data Collection: Uses OANDA API to fetch historical candlestick data.
- Feature Engineering: Includes MACD, EMA, and custom lagged features.
- DNN Model: Binary classification for price movement direction.
- Backtesting: Evaluates strategy performance on historical data.
- Live Trading: Executes trades in real-time based on model predictions.
- Python 3.8+
- OANDA Practice Account (API access)
- Required Python packages:
pip install tensorflow keras matplotlib pandas numpy oandapyV20
- Use the model.py script to train the DNN model.
- Model parameters and training results are saved in the models/ directory.
- Use the model_inference.py script for live trading.
- Adjust the instrument and parameters in the script as per your model's training setup.
- Inputs: Historical price data, MACD, EMA, SMA, Volume, Min, Max, lagged features.
- Architecture: Fully connected layers with dropout and optional L1 regularization.
- Optimizer: Adam with a learning rate of 1e-5.
- Loss Function: Binary cross-entropy.
- Outputs: Probability of upward price movement.
- Backtesting is implemented in
model.pyusing a trading strategy based on model probabilities. - Results include cumulative returns and strategy visualization.
- The trading bot fetches real-time data using the OANDA API.
- Executes long/short positions based on model predictions.
- Handles trading parameters like granularity, thresholds, and lot sizes.
A Model was trained on the USD/JPY instrument with 15-minute candles. Historical data was collected from 2024-01-01, with features EMA_200, MACD_line, SIGNAL_line, min and max. Results were as follows:
Across the randomised two-month period, the model achieved a net return (before fees) of 3.398%, marginally exceeding buy-hold returns.