Building and Training Feedforward Neural Networks in MQL5

Summary

This document describes a native MQL5 library for creating, training, and running feedforward neural networks inside MetaTrader 5. It outlines dense matrix operations, activation functions, layers with dropout and gradient clipping, optimizers, learning rate schedules, loss functions, and a mini-batch training pipeline. Example applications include signal classification, price direction prediction, candlestick pattern recognition, market regime detection, and reinforcement learning value estimates.

The technical notes mention He and Xavier weight initialization, Fisher-Yates shuffling, inverted dropout, and a simplified gradient for the Softmax and cross-entropy combination. The document also reports sub-millisecond inference for typical networks under 1,000 parameters and gives a memory estimate for one example architecture. These are library claims rather than independently documented benchmarks. It provides no trading dataset, out-of-sample evaluation, or evidence that models trained with the library generate profitable signals; users must supply suitable data and validate their models.

Key ideas

• The library supports defining and training dense neural networks entirely in MQL5.
• It includes several optimizers, learning rate schedules, activation functions, and loss functions.
• Training features include mini-batches, shuffling, dropout, and gradient clipping.
• The document lists trading use cases but provides no empirical evidence of trading performance.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.