Using Conditional GAN Synthetic Time Series in Machine Learning Backtests

Summary

This tutorial describes generating synthetic multivariate financial time series with SDV’s PAR synthesizer, a conditional autoregressive GAN approach. It fits sequences using asset identifiers and stable context fields, with returns and volume as modeled variables. The example reconstructs price paths from generated returns, builds technical indicators, and trains random forest classifiers in a walk-forward process. The stated motivation is to expand the paths available for testing when historical observations are scarce. The method requires multiple asset series with context information; it is not intended for a single asset series alone.

The tutorial reports that its machine learning approach had lower volatility and maximum drawdown than buy-and-hold in its example, while buy-and-hold had higher annualized return and stronger Sharpe, Calmar, and Sortino ratios. These claims are specific to the described experiment, whose settings and sample construction limit generalization. GAN training can be slow or unstable, and mode collapse can reduce sample diversity. The author recommends multiple generated paths, attention to transaction costs, and careful model selection; synthetic observations do not replace validation on real out-of-sample data.

Key ideas

• The PAR synthesizer generates multivariate sequences conditioned on asset context, so the approach needs data for multiple assets.
• The example combines generated returns and volume with technical indicators and random forest classification in walk-forward testing.
• Synthetic paths can expand scenario coverage when historical data is limited, but their usefulness depends on the generated data quality.
• The reported comparison favors buy-and-hold on several return and risk-adjusted measures, while the machine learning strategy has lower volatility and drawdown.
• Training time, instability, and reduced sample diversity are stated challenges, and results are specific to the example.

Tags

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