Batch Normalization Choices and Activation Matching for Trading Neural Networks

Summary

The article surveys three normalization approaches for neural networks used in trading systems: standard scaling, feature scaling, and robust scaling. It discusses commonly proposed benefits of normalization, including steadier training and improved convergence, while noting debate over the explanation for those effects. Feature scaling maps values into a bounded zero-to-one range, and the text emphasizes matching normalization output ranges with compatible activation functions such as sigmoid or softmax. Standard-score scaling is unbounded, while robust and feature scaling are bounded alternatives.

The author describes incorporating normalization choices into an MQL5 multilayer perceptron and comparing normalized models with an unnormalized control. The tests use EURJPY daily data from 2023 and also describe testing GBPCHF on a four-hour timeframe during 2023, but the supplied excerpt omits the numerical result tables. The conclusion presents normalization as potentially helpful but computationally costly and sensitive to choices of scaling, activation, network size, and layers. The reported discussion does not provide enough figures to establish which configuration performs best or whether gains generalize.

Key ideas

• The article compares standard, feature, and robust scaling for neural network inputs or layer data.
• Feature scaling bounds values between zero and one, so activation ranges should be considered when pairing methods.
• Standard-score scaling is unbounded and may produce large values that complicate training.
• The described MQL5 experiments compare normalized networks with an unnormalized control, but numerical outcomes are absent from the excerpt.
• Normalization may add computational cost, and its effects depend on the scaling method, activation, and network design.

Tags

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