NeuroNetworksBook

A comprehensive library for building, training, and deploying neural networks directly in MQL5 for MetaTrader 5.

Features

• Wide Variety of Layers: Supports fully connected (Base), Convolutional, LSTM, Attention, Multi-Head Attention, GPT-like blocks, Dropout, and Batch Normalization.
• OpenCL Acceleration: Many layers and training algorithms are optimized with OpenCL for high-performance computing.
• Flexible Architecture: Build complex network architectures using a collection of layer descriptions.
• Loss Functions: Multiple loss functions supported (MSE, Cross-Entropy, etc.).
• Optimization Algorithms: Support for various optimizers like SGD, Momentum, Adam, and more.
• Python Integration: Includes Python scripts for validating gradients and testing logic against standard implementations.

Project Structure

• Include/realization/: Core MQL5 library files.
  • neuronnet.mqh: Main dispatcher class CNet.
  • neuronbase.mqh, neuronconv.mqh, neuronlstm.mqh, etc.: Individual layer implementations.
  • bufferdouble.mqh: Data buffer management with OpenCL support.
  • opencl_program.cl: OpenCL kernels for GPU acceleration.
• Experts/: Example Expert Advisors using the library.
• Scripts/NeuroNetworksBook/: Test scripts and Python validation tools.
• Include/about_ai/: Educational resources and additional implementations.

Training and Deployment Workflow

The library provides a streamlined process for training and deploying models.

1. Data Preparation

Use scripts in Scripts/NeuroNetworksBook/initial_data/ to generate training data (CSV). For example, create_initial_data.mq5 can export indicator values and target outcomes.

2. Training

Use the Scripts/NeuroNetworksBook/TrainAndDeploy.mq5 script:

• Set StudyFileName to your exported CSV.
• Configure architecture and hyperparameters.
• The script will train the model and save it as a .net file in MQL5/Files.

3. Backtesting and Deployment

Use Experts/ea_template.mq5:

• Set Model to your trained .net file.
• The EA will load the model, prepare real-time data, and execute trades based on network predictions.

---

Code Cleaning

The library has been updated for better stability:

• Improved Error Handling: CNet::GetResults and CNet::Create now include robust pointer checks.
• Unified Macros: CBufferType and TYPE macros in defines.mqh ensure consistency across the library and Expert Advisors.
• Refactored EA Template: Fixed logic for data input and result retrieval to match the core library's expected behavior.

---

Quick Start

To use the library in your Expert Advisor:

1. Include the main header:

   #include <NeuroNetworksBook\Include\realization\neuronnet.mqh>
   

2. Define your network architecture:

   CArrayObj *descriptions = new CArrayObj();
   
   // Input layer
   CLayerDescription *desc = new CLayerDescription();
   desc.type = defNeuronBase;
   desc.count = 10; // Number of inputs
   desc.activation = None;
   descriptions.Add(desc);
   
   // Hidden layer
   desc = new CLayerDescription();
   desc.type = defNeuronBase;
   desc.count = 20;
   desc.activation = ACT_TANH;
   descriptions.Add(desc);
   
   // Output layer
   desc = new CLayerDescription();
   desc.type = defNeuronBase;
   desc.count = 1;
   desc.activation = ACT_SIGMOID;
   descriptions.Add(desc);
   

3. Initialize and use the network:

   CNet net;
   if(net.Create(descriptions)) {
       // Feed forward
       CBufferDouble *inputs = new CBufferDouble();
       // ... fill inputs ...
       net.FeedForward(inputs);
   
       // Get results
       CBufferDouble *results;
       net.GetResults(results);
   }
   

Development

This project is part of a book on Neural Networks in Algorithmic Trading. Feel free to explore the code, run tests in Scripts/, and adapt the EA templates in Experts/.