Neural Network Basics and a Single-Neuron Trading EA Example
Summary
The article introduces artificial neurons and multilayer networks, describing inputs, weights, weighted sums, activation functions, and how learning changes weights. It explains scaling inputs to a bounded range and compares step, sigmoid, and hyperbolic tangent activations. For sigmoid and tangent, it illustrates adjusting the weighted sum by a coefficient to spread outputs across a more useful portion of each function’s range.
A small MQL5 example combines weighted inputs and an activation function in a one-neuron Expert Advisor, then compares a MACD-based neural EA with a standard MACD sample on EURUSD hourly data. The reported test gives the neural version higher net profit and recovery factor, but also higher drawdown, lower profit factor and Sharpe ratio, and many more trades. The example is explicitly primitive, with only one neuron; the reported historical test does not establish robustness or general performance. The article’s normalization example also uses the observed data minimum and maximum, so practical use requires care about how scaling is determined for future inputs.
Key ideas
- A neuron combines inputs and their weights before applying an activation function to produce an output.
- Input normalization puts variables with different scales into a bounded range before they enter a network.
- Sigmoid and hyperbolic tangent outputs can be adjusted by scaling their weighted-sum argument.
- The article demonstrates a single-neuron EA and reports mixed results against a standard MACD EA in one historical test.
- A small example and one test period do not establish that neural networks will outperform conventional trading rules.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.