Implementing and Evaluating Backpropagation Networks with MQL5 Matrices
Summary
The article introduces supervised backpropagation and shows how matrix and vector operations can express neural-network forward passes and weight updates in MQL5. It reviews neurons, layer weights, activation functions, bias inputs, and gradient descent, then describes reusable classes for constructing and training fully connected networks. The matrix approach supports batch processing and is presented as a way to keep the implementation close to the underlying formulas. The article also discusses using the classes in scripts, indicators, and Expert Advisors, including visualization and saving or restoring networks.
A trading demonstration tests predictions on held-out periods, including data outside the training sample. The reported system remains unprofitable on new data, and the author treats this as evidence that the particular predictive idea did not succeed, while arguing that the toolkit can still help evaluate ideas technically. The demonstration does not establish that neural networks generally fail; results may reflect noisy or uninformative inputs and the limited factor search. The article emphasizes that model architecture cannot compensate for weak data.
Key ideas
- Backpropagation trains network weights by propagating output error through differentiable activation functions.
- Matrix operations express layer calculations and batch processing in a compact MQL5 implementation.
- Reusable classes are presented for network construction, training, visualization, and persistence.
- The trading demonstration loses on unseen data, so its predictive idea is not validated.
- Network performance depends on the information and quality of the input data, not only the algorithm.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.