MQL5 Neural Network Library

This repository demonstrates a complete workflow for training and deploying neural networks directly inside MetaTrader 5. The goal is to show that the MQL5 language can handle custom machine learning models without relying on external tools.

This proof-of-concept aims to design and implement a reliable workflow for neural network training within the MQL5 ecosystem—an environment that typically lacks native machine learning support. By doing so it enables real-time trading strategies to leverage custom models without external dependencies.

Project Highlights

• Native training and inference in MQL5
• Serialization of model weights for easy deployment
• Modular design supporting multiple network architectures
• Example Expert Advisor showing how to automate the process

Supported Architectures

The library currently implements three common network types, all inheriting from the INeuralNetworkModel interface:

• Multilayer Perceptron (MLP) – a basic feedforward network suitable for simple patterns.
• Recurrent Neural Network (RNN) – processes sequences and retains short-term memory.
• Long Short-Term Memory (LSTM) – an RNN variant that captures long-range dependencies.

Adding new architectures follows the same interface so the workflow can be extended easily.

Workflow Overview

The training pipeline consists of four stages:

1. Training Phase – Configure the network on the Strategy Tester Inputs tab with TrainMode set to true. Here you choose the architecture, set the training start date and other parameters. Training runs until the MAE target is reached using the Adam optimizer.

   • Inputs tab for selecting the model, enabling training and defining initial settings.
   • Example of configuring the historical range for learning.
   • Starting the Strategy Tester to launch the optimization process.

2. Persistence Phase – After training completes, the weights are saved to a binary .bin file under the platform's common Files directory.

   • Click the Files button in MetaEditor to browse generated files.
   • Use the Common Folder option to access shared data.
   • Open the My_AI_Models directory created by the EA.
   • Each .bin file contains the serialized weights for a trained network.

3. Model File and Training Log – The binary file itself appears unreadable if opened directly, which is normal because weights are stored in a compact format. This binary representation loads quickly inside MT5, making it efficient for the workflow.

   • Example .bin file contents shown for demonstration purposes.
   • Strategy Tester log illustrates how MAE decreases over epochs.

4. Inference Phase – With TrainMode set to false, the Expert Advisor loads the saved weights and begins producing live predictions.

   • The EA uses the stored model to forecast price movement on new data.

This separation allows you to train once and deploy the same model on any chart or in live trading without modification.

Repository Layout

C_MLP.mqh               // Multi-layer perceptron implementation
C_RNN.mqh               // Simple recurrent network implementation
C_LSTM.mqh              // Long short-term memory implementation
INeuralNetworkModel.mqh // Interface for all models
WorkFlow_test.mq5       // Example Expert Advisor demonstrating the workflow

Getting Started

1. Copy the .mqh and .mq5 files into your platform's MQL5 directory.
2. Open WorkFlow_test.mq5 in MetaEditor and adjust parameters as needed (such as training epochs or MAE target).
3. Run the Expert Advisor in Strategy Tester with TrainMode = true to generate a weight file.
4. Set TrainMode = false and run again on the desired symbol and timeframe to execute live predictions.

Trained weights are stored under My_AI_Models inside the platform's common Files directory so they can be reused across accounts and charts.

Feature Engineering

The proof-of-concept uses only two input features—closing price and RSI—but the architecture is designed for easy expansion. Future updates may incorporate additional indicators (MACD, Bollinger Bands, volume) or raw OHLCV data.

Future Plans

• Support for advanced architectures such as GRU, CNN, or Transformer
• Reinforcement learning agents for dynamic risk management
• A richer library of technical indicators

License

This code is provided for educational and research purposes. Use it at your own risk when trading live markets.