Using GANs to Generate Financial Time Series and Test Strategy Overfitting

Summary

The report introduces generative adversarial networks (GANs) for quantitative finance. A generator creates synthetic market data while a discriminator tries to distinguish it from real data; alternating training aims to make the generated series resemble the training distribution. The report also outlines GAN benefits and practical difficulties, including training instability, imbalance between the two networks, black-box behavior, and mode collapse.

The authors generate index return and price series across several markets and frequencies, then compare them with Bootstrap and GARCH simulations using six properties, such as autocorrelation, heavy tails, volatility clustering, leverage effects, and gain-loss asymmetry. They report that GAN series reproduce these features more fully than the comparison methods. A moving-average timing example shows how synthetic scenarios can help assess strategy overfitting. The report also discusses synthetic training data and conditional GANs for forecasting as potential uses. Results depend on the tested indices and evaluation measures; the report acknowledges GAN training challenges and says forecasting applications remain less developed.

Key ideas

• A GAN trains a generator and discriminator in alternating competition to approximate the distribution of observed data.
• The report evaluates synthetic financial series using six statistical properties, including volatility clustering and gain-loss asymmetry.
• Its tests compare GAN output with Bootstrap and GARCH simulations across equity indices and sampling frequencies.
• The authors report stronger reproduction of tested features by GANs, while noting instability and mode collapse as limitations.
• Synthetic market scenarios can be used to probe strategy overfitting, illustrated with a dual moving-average timing strategy.

Tags

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