Adapting a Restricted Boltzmann Machine for MQL5 Trading Signals

Summary

The article introduces the conventional restricted Boltzmann machine (RBM) as a network that learns hidden representations through positive and negative passes, with weights adjusted by contrastive divergence. It sketches an MQL5 implementation using Gibbs sampling, neuron states, weights, and biases, but explicitly says this orthodox version is for exploration and has not been tested.

The main experiment uses an unconventional alternative: a smaller hidden layer between matching input and output layers, trained to reconstruct the input in an unsupervised manner. The resulting hidden-layer weights or activations are then used as features for an MQL5 expert signal. The article reports optimization and walk-forward testing on GBPUSD at a four-hour timeframe over stated historical periods, but provides no basis for treating the results as conclusive. It recommends further customization and longer testing with broker real-tick data, noting limits in input selection and computational efficiency.

Key ideas

• A conventional RBM learns hidden representations through upward and downward passes and weight updates by contrastive divergence.
• The article presents an MQL5 implementation sketch of Gibbs sampling and parameter updates, while stating that this version is untested.
• An alternative trains a smaller hidden layer to reconstruct its input, providing a compact unsupervised feature-learning approach.
• The learned representation is applied to an expert signal and evaluated through optimization and walk-forward testing on GBPUSD.
• The reported results are preliminary and require broader testing, input refinement, and attention to computational efficiency.

Tags

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