ASQ NeuralNet Pure MQL5 Neural Network Library - library for MetaTrader 5

ASQ NeuralNet is a complete neural network library written 100% in native MQL5 — no DLLs, no Python bridge, no external dependencies. Build, train, and run inference with multi-layer neural networks directly inside MetaTrader 5.

WHAT YOU GET

A fully functional deep learning framework for MQL5 developers, including:

— Dense matrix algebra engine with 40+ operations (multiply, transpose, Hadamard, He/Xavier initialization, NaN detection) — 13 activation functions with analytical derivatives: ReLU, LeakyReLU, ELU, SELU, Sigmoid, Tanh, Softmax, Swish, Mish, GELU, Softplus, HardSigmoid, Linear — Dense layers with forward and backward propagation, dropout, and gradient clipping — 3 optimizers: SGD (with momentum), Adam, AdamW (decoupled weight decay) — 7 learning rate schedulers: Constant, Step Decay, Exponential, Cosine Annealing, Linear, ReduceOnPlateau, Warmup, Cyclic LR — 5 loss functions: MSE, MAE, Huber, Cross-Entropy, Binary Cross-Entropy — Full training pipeline with mini-batch SGD, Fisher-Yates shuffle, and epoch logging

QUICK START

Building a network takes 6 lines of code:

CNeuralNetwork net; net.Init(32); // 32 input features net.AddLayer(64, ACT_RELU); // Hidden layer 1 net.AddLayer(32, ACT_RELU, 0.2); // Hidden layer 2 + dropout net.AddLayer(3, ACT_SOFTMAX); // Output: BUY/SELL/HOLD net.Build();

Train with one call:

net.SetOptimizer(OPT_ADAM, 0.001); net.SetLoss(LOSS_CROSS_ENTROPY); net.Fit(trainX, trainY, 100, 32);

Predict with one call:

int action = net.PredictClass(features); // 0=BUY, 1=SELL, 2=HOLD

USE CASES

— Train classification models for trading signals (BUY/SELL/HOLD) — Price direction regression with MSE or Huber loss — Pattern recognition on candlestick formations — Market regime detection (trending / ranging / volatile) — Feature importance analysis — Q-value function approximation for reinforcement learning agents

PERFORMANCE

— Inference latency: < 0.1ms for typical architectures (under 1000 parameters) — Memory: proportional to total parameters (5KB for a 32→64→32→3 network) — No dynamic allocation during inference — Numerically stable: NaN detection, gradient clipping, safe Softmax with max-subtraction

INSTALLATION

Place the 5 library files in MQL5/Include/AlgoSphere/NeuralNet/ and include the main header:

#include <AlgoSphere/NeuralNet/NN_Network.mqh>

The demo script demonstrates matrix operations, activation functions, XOR classification, and synthetic market direction prediction.

LIBRARY FILES

— NN_Matrix.mqh (908 lines) — Dense matrix algebra engine — NN_Activations.mqh (300 lines) — 13 activations + derivatives — NN_Layer.mqh (374 lines) — Dense layer with forward/backward/dropout — NN_Optimizer.mqh (454 lines) — SGD/Adam/AdamW + 7 LR schedulers — NN_Network.mqh (734 lines) — Complete feedforward network with training — ASQ_NeuralNet_Demo.mq5 (283 lines) — 4 runnable demonstrations

Total: 3,053 lines of pure MQL5.

TECHNICAL NOTES

— Weight initialization: He Init for ReLU-family, Xavier for Sigmoid/Tanh — Box-Muller transform for normal distribution (MQL5 native MathRand) — Softmax + Cross-Entropy gradient shortcut (ŷ - y, avoids full Jacobian) — Inverted dropout (scaled during training, identity during inference) — Fisher-Yates shuffle for mini-batch training — Gradient norm clipping per layer (default max norm = 1.0)

Built by AlgoSphere Quant.

Tags / Keywords

neural network, deep learning, machine learning, AI, MQL5, library, matrix, backpropagation, Adam optimizer, classification, regression, trading signals, pure MQL5, no DLL, feedforward, activation functions, softmax, cross-entropy, dropout