Using Wasserstein GANs to Generate More Realistic Financial Returns

Summary

The document summarizes research applying Wasserstein generative adversarial networks (WGANs) to financial return series and comparing them with standard GANs. WGANs replace the original GAN’s distribution-distance measure with Wasserstein distance, which the summary says helps reduce gradient problems and makes training progress easier to assess. The stated motivation is to address unstable training, nonconvergence, and mode collapse in conventional GANs.

The reported experiments generate daily Shanghai Composite returns and monthly S&P 500 returns. Generated series are assessed using distributional and temporal properties, including autocorrelation, heavy tails, volatility clustering, leverage effects, variance-ratio tests, and the Hurst exponent; dynamic time warping is used to compare diversity. The summary reports that WGAN samples better match several observed characteristics and are more diverse than standard GAN samples, with differences across the two markets and sampling frequencies. These are reported comparisons, not evidence that generated data improves trading outcomes. The underlying paper’s full text is not included, limiting review of its setup, metrics, and caveats.

Key ideas

• WGANs use Wasserstein distance to compare real and generated data distributions.
• The approach is presented as a response to training instability and mode collapse in standard GANs.
• The summarized tests cover daily Chinese index returns and monthly US index returns.
• The comparison evaluates market patterns such as volatility clustering and long-range dependence.
• The reported WGAN samples improve realism and diversity, but trading utility is not demonstrated.

Tags

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