GAN Learning-Rate Schedules in an MQL5 Trading Signal

Summary

The article examines how learning-rate choices affect a simple generative adversarial network used as an MQL5 trading signal. It describes fixed and scheduled rates, including step decay, exponential decay, and cyclical adjustment. In broad terms, a fixed rate is simple and predictable but may converge poorly; decaying rates reduce update size over training, while cyclical schedules vary it over a cycle. The generator’s output and the discriminator’s decision are combined to produce long or short conditions.

The author compares schedules in strategy tests using EURJPY on the daily timeframe over 2023, holding the symbol, period, and a compact network architecture consistent. Performance is defined narrowly as total profit with recovery factor considered. The article says results vary with learning rate, but the provided text omits much of the individual test evidence and does not establish durable predictive value. It also cautions that longer evaluation periods and other network designs could change conclusions; further schedules are deferred to a later discussion.

Key ideas

• GAN training pairs a generator with a discriminator, and both contribute to the trading signal logic.
• A fixed learning rate is easy to implement and compare, but can limit convergence and adaptation.
• Step decay lowers the rate at set epoch intervals, while exponential decay reduces it progressively.
• Cyclical scheduling varies the learning rate through a repeating range.
• The reported comparison uses one currency pair, daily data, a single test year, and a simple network, limiting generalization.

Tags

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