MQL5 Neural Network Quantitative Trading System: Application of ONNX Models in MetaTrader

Core Project Information

• Original Author/Maintainer: yordanoweb
• Source Platform: GitHub
• Project Name: mql_neural
• Core Objective: Deploy Python-trained neural network models to MetaTrader 5 (MT5) via ONNX format to enable automated AI trading in foreign exchange and stock markets
• Technical Value: Address the deployment challenges of Python machine learning models to the MT5 platform and demonstrate the practical application of machine learning in quantitative trading

Quantitative Trading Background

Quantitative trading uses mathematical models and algorithms to identify trading opportunities. Traditional strategies are based on technical indicators and statistical arbitrage but struggle to capture complex nonlinear market patterns.

Deployment Challenges

As a mainstream trading platform, MT5 lacks the native ability in its MQL5 language to directly load Python models.

Solution

The project uses ONNX as a bridge to achieve seamless integration between Python-trained models and the MQL5 system.

ONNX Introduction

ONNX is a cross-platform model exchange standard that supports export from frameworks like PyTorch/TensorFlow, with advantages including framework independence, high-performance inference, cross-language deployment, and model compression.

Core Components

1. MQL5 Expert Advisor (EA): Data collection, inference, and trade execution
2. Python training scripts: Preprocessing, feature engineering, model training, and ONNX export
3. ONNX model file: Standardized storage of trained models

Loading Methods

• Compile as resource file: Model bundled with EA for convenient deployment
• Dynamic loading: Update models without recompilation for flexible iteration

Backtesting Configuration

• Instrument: NASDAQ100
• Timeframe: M5 (5-minute candlesticks)
• Test mode: Real tick/OHLC data
• Cost considerations: Slippage, commissions, spreads

Feature Engineering

It is inferred from the naming that three core features are used, common types include price features (open, high, low, close), technical indicators (MA/RSI), statistical features (return rate/volatility), etc.

Data Quality

Financial time series are non-stationary/heteroscedastic and require preprocessing (log transformation/standardization), and there is a lot of noise that needs to be distinguished from signals.

Overfitting Risk

High market noise easily leads models to learn random patterns, requiring out-of-sample testing and walk-forward analysis.

Market Regime Changes

Changes in market states (bull/bear/volatility) can invalidate models, requiring regular retraining or online learning.

Execution Issues

Slippage, latency, and insufficient liquidity affect actual returns, and backtesting cannot fully simulate the real market.

Multi-Instrument and Multi-Timeframe

• Multi-instrument portfolio: Diversify risk across forex, indices, and commodities
• Multi-timeframe: Combine daily/hourly signals
• Cross-market arbitrage: Utilize price deviations of related assets

Model Enhancement

• Ensemble learning: Improve robustness with multiple models
• Reinforcement learning: Directly learn trading decisions
• Attention mechanism: Capture long-range dependencies
• Model quantization: Improve inference speed

Risk Management

• Position management: Dynamically adjust positions
• Stop-loss and take-profit: Control single-trade risk
• Drawdown control: Adjust strategies during consecutive losses
• Correlation monitoring: Avoid highly correlated positions

Project Significance

Fills the gap between the Python ecosystem and MQL5, lowers the threshold for AI quantitative development, reuses Python libraries, enables rapid prototype validation, and protects strategy logic.

Future Outlook

The improvement of the ONNX ecosystem and enhanced AI support in MT5 will promote the application of deep learning quantitative strategies, and mql_neural is an excellent open-source reference in the AI quantitative field.