SinGAN for Financial Time Series: Single-Sample Generation and Evaluation

Summary

This study adapts SinGAN, a multiscale generative adversarial network trained on one sample, to generate financial return sequences. Its pyramid trains progressively across scales: coarse layers represent broad, low-frequency structure, while finer layers capture local, high-frequency detail. A patch-based discriminator and reconstruction loss help preserve local characteristics and stabilize training. The motivation is to address the sample scarcity and sequence-length constraints that arise when conventional GANs train on short time-series slices.

The authors compare SinGAN with WGAN-generated short sequences joined into full-length series across daily Chinese equity indices and monthly US equity and foreign exchange data. They report similar results for both methods on the CSI 300, but mode collapse for WGAN on the shorter STAR 50 history. For monthly S&P 500 and EUR/USD series, SinGAN’s spectra more closely resemble the originals, while WGAN misses some medium- and long-period signals. The authors caution that the model may overfit and can fail when market patterns change; synthetic sequence similarity does not establish forecastability or strategy profitability.

Key ideas

• SinGAN uses a pyramid of GANs to learn sequence structure from coarse scales to fine scales.
• The method can generate sequences at lengths chosen for a task, including lengths matching the original series.
• In the reported comparisons, WGAN suffered mode collapse on a short equity history, while SinGAN better matched selected monthly frequency patterns.
• The study evaluates synthetic data resemblance, not trading returns or predictive performance.
• Overfitting and changing market behavior limit how confidently historical patterns can be carried forward.

Tags

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