Conditional GAN Time-Series Forecasting

Article MQL5 articles

Summary

The article adapts a conditional generative adversarial network to forecast financial time series within an MQL5 expert advisor. It describes a generator and discriminator, both based on a multilayer perceptron: the generator uses prior close-price changes to predict the next change, while the discriminator evaluates whether a pairing of those inputs and a label is real or generated. Generator training weights its prediction error using discriminator output so it learns both to forecast and to produce samples the discriminator accepts.

The proposed signal class normalizes inputs, trains as new bars arrive, and exposes parameters for learning rate, epoch count, and training-set size. The article presents the design as a prototype, but the supplied excerpt gives no clear quantitative evidence of forecast or trading performance. It explicitly leaves network architecture selection, weight persistence, and alternative training schedules unresolved. In particular, training on every bar may adapt to changing conditions or may overfit noise, so the author treats the system as a starting point rather than a trade-ready method.

Key ideas

A conditional GAN pairs prior price changes with a forecast label for discriminator evaluation.
The generator's training objective incorporates discriminator output alongside forecast error.
The example uses an MLP generator and discriminator and trains on each new bar.
Architecture selection, weight saving, and training frequency remain open design choices, and the excerpt provides no performance validation.

Conditional GANs for Financial Time-Series Forecasting

Summary

Key ideas

Tags