Activation Functions for Custom Trading Optimization Criteria
Summary
The article explores using neural network activation functions to build custom criteria for ranking strategy parameter sets in MetaTrader optimization. It argues that returning zero for an undesirable pass can disrupt genetic search, while unrestricted combinations of performance metrics can produce scores that grow excessively. ReLU and Softplus illustrate limitations such as zero-valued regions or unbounded output; Sigmoid and Tanh are presented as bounded alternatives that can give graded scores around chosen thresholds. Offsets shift the curves toward desired metric values, while rescaling can support criteria for both minimum and bounded ranges.
Examples use trading statistics such as Recovery Factor and trade count to illustrate how a metric could contribute to a custom score. The article frames this as an initial, empirical exploration and defers detailed weighting, normalization, and practical combinations to later installments. It does not provide comparative optimization results establishing that these criteria improve robustness or trading performance, so the functions are scoring design ideas requiring validation.
Key ideas
- Returning zero for rejected optimization passes may remove useful candidates from a genetic search.
- Unbounded combinations of outcome metrics can create excessively large scores.
- Sigmoid and Tanh functions can map metric deviations into bounded, gradual scores.
- Offsets can align a curve's scoring transition with a chosen target or threshold.
- The proposed criteria are exploratory and need testing before being treated as evidence of robust strategies.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.