W-DCGAN for Generating Multi-Asset Financial Time Series

Summary

This study examines whether generative adversarial networks can produce realistic multi-asset financial time series. It combines the convolutional architecture of DCGAN with the Wasserstein-distance loss used by WGAN: convolutional layers replace some conventional network components, while the Wasserstein objective is intended to improve training stability and reduce gradient and mode-collapse problems.

The reported comparison uses returns from the S&P 500, Shanghai Composite, and EURO STOXX 50, assessed with nine single-asset and five multi-asset indicators. Plain DCGAN performs poorly on several features, including autocorrelation, leverage effects, return asymmetry, and cross-asset dependence. WGAN and W-DCGAN perform better overall, with W-DCGAN reported to have a modest edge on some measures, including asymmetry, the Hurst exponent, and rolling correlations. These results concern statistical resemblance under the study’s chosen metrics; they do not establish trading profitability, and the available text does not provide the full paper’s implementation details or broader validation.

Key ideas

• DCGAN changes GAN architecture through convolutional layers and related network design choices.
• W-DCGAN pairs that architecture with a Wasserstein-distance loss.
• The study evaluates generated data using single-asset and cross-asset statistical indicators.
• The reported tests favor W-DCGAN and WGAN over plain DCGAN on several time-series features.
• Statistical similarity in the tested markets does not demonstrate investment performance.

Tags

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