MQL5 MACHINE LEARNING 01: Neural Networks For Algo-Trading

build a neural network in Excel to visualize the forward pass, using moving average and RSI oversold as inputs, with weights, bias, and a sigmoid activation to predict bullish signals.

Compute mean squared error as the squared difference between predicted and expected outputs, then learn how gradient descent and backpropagation adjust weights across epochs to minimize loss.

Learn how to perform a backward pass to adjust weights and bias in a neural network using the learning rate and cost function to reduce mean squared error.

Discover how linear algebra underpins machine learning in MQL5 by modeling data with vectors and matrices, understanding vector spaces, linear transformations, and practical data manipulation.

declare matrices and vectors in mql5 with and without size, using matrix and vector vessels of double or float, specify rows and columns, initialize data, and verify with prints.

Learn to declare and initialize matrices and vectors in MQL5 using EI matrices, full, ones, zeros, and identity functions, including an initialization function and practical printing in MetaTrader 5.

Learn how to copy data into matrices and vectors using the equal sign, copy function, and assign operator, including from matrices and arrays, and handle matrix-to-vector and vector-to-matrix conversions.

Copy time series data, including candlestick and indicator data, into vectors in Mql5, then into matrices, using RSI handles and copy indicator buffer.

Explore matrix and vector operations, including addition, subtraction, multiplication, and division, with emphasis on element-wise versus matrix multiplication, size compatibility, and using vectors for faster programs.

Explore matrix manipulations including transposition, reshaping, and extracting data from rows, columns, and diagonals, with applications to neural networks.

designs a supervised feedforward neural network in MQL5 to classify markets as bullish or bearish using RSI and Bollinger Bands inputs, one hidden layer, sigmoid activation, and backpropagation.

Build a neural network based expert advisor in the meta editor with hourly training and five minute trading. Configure general inputs, symbol, and chart visuals for testing.

Detect new bars by comparing old and current bar counts from history and run the on tick logic only when a new bar appears, boosting performance and reducing lag.

Create indicator inputs for RSI 14 and Bollinger Band 20 with deviation 2, and build a four-column input matrix for neural network training on 1000-bar lookbacks and live data.

Initialize indicator handles and Bollinger band settings, verify enough training paths (8000) via terminal info, and enforce chart path limits for robust neural network training in MQL5.

Initialize indicator and price vectors, implement get live data for trading and collect training data for neural networks, and build input matrix from Bollinger band, RSI, and open/close prices.

Normalize the input matrix with per-column min-max scaling to 0-1, speeding neural network training and improving prediction by standardizing all features.

Initialize networks for algo-trading by using a random seed to initialize weights and biases, create weight and bias matrices, and produce bitrate and sell signals from RSI and Bollinger bands.

Convert a vector to a matrix to support hidden layer calculations, validate column compatibility, and fill the matrix from the vector in row-major order.

Convert vectors to matrices and matrices to vectors using an assign function with autocast to support hidden layer calculations in neural networks for algo-trading in MQL5.

Compute neuron calculations for the hidden layer of a neural network using an activation function enum. Prioritize sigmoid activation, compute dot-product of inputs and weights, add bias, then apply activation.

Execute a forward pass for the neural network using live data, normalizing the vector with min-max, converting to a matrix, and producing a probability vector and tracking backpropagation status.

Prepare for back propagation by implementing a zero-division guard and a pattern-search routine that counts matching elements across two vectors for a confusion matrix.

Develop a void function that removes an index from a vector by reference, creating a vector of size n-1 and copying all except target in confusion matrix and backpropagation.

Develop functions to remove a matrix column or row in MQL5 by constructing a new matrix and copying data, preparing data for algo-trading workflows.

Define a confusion matrix structure to store accuracy, precision, recall, f1 score, and related metrics for algo-trading models, handling true/false positives and negatives across varying class sizes for vector conversion.

populate the confusion matrix by validating actual and predicted vector sizes, then nest loops to fill rows as actual values and columns as predicted values using the search patterns function.

Calculate model accuracy from the confusion matrix by dividing true positives and true negatives by all predictions, then compute precision for each class using check zero and normalization.

Compute recall, or sensitivity, by extracting the class row from the confusion matrix, resizing the recall vector, and calculating true positives over true positives plus false negatives with NaN checks.

Compute specificity, the true negative rate, for each class from the confusion matrix by removing the current class row and column and calculating true negatives and false positives.

Compute the F1 score for each class using the harmonic mean of precision and recall, implemented per element with a matrix object.

Compute class supports from a confusion matrix by summing row values, validating results, and preparing for subsequent average and weighted average calculations in neural networks for algo-trading.

Compute the mean of precision, recall, specificity, and f1 score within a confusion matrix, then derive a weighted average using class support to assess model accuracy for algo-trading.

Create a classes function that extracts unique values from an input vector, preserving order, to support back propagation in neural networks for MQL5 machine learning.

Implement a one hot encoding function to convert categorical data into a matrix for backpropagation. It assigns ones to the correct class and zeros elsewhere, preparing data for backpropagation.

Choose loss functions for backpropagation, including binary cross entropy, categorical cross entropy, mean squared error, and categorical hinge, to minimize prediction error over 100 epochs with a 0.04 learning rate.

Implement a batch input forward pass by looping over the input matrix, applying the live forward pass per row, and building a classes vector via argmax.

Implement a backpropagation training loop in MQL5, updating weights and biases via gradients, performing forward passes, one-hot targets, and epoch-based evaluation.

Train a neural network using backpropagation with a train model function that collects data and updates the model, and display price-direction predictions on the main chart with a configurable arrow.

Train a neural network model for algo-trading, apply backpropagation once, evaluate with epochs, accuracy, precision, recall, F1, and a confusion matrix, then plot arrows predicting price movement.

Display probability signals from a neural network for algo-trading using a 0.55 probability threshold to indicate bullish or bearish moves and update arrow indicators based on the old signal.

Train and visually test a neural network in mql5 to predict price movement, draw bullish or bearish arrows, and evaluate performance on historical data.

Improve the neural network by training on bars of a defined length, add a ranging market class with RSI and Bollinger Bands, and design an MQL5 expert advisor.

Immerse in feedforward neural networks, understand architecture, training, and their real-world applications in algo-trading. Embrace curiosity, stay learning as artificial intelligence evolves, and build reliable expert advisors for forex.