GAN-Generated Market Scenarios

A backtest only ever sees the one history that actually happened, so a strategy can look robust simply because it never faced a scenario the real market avoided. A generative adversarial network (GAN) offers a way to manufacture additional, statistically realistic market paths: a "generator" network learns to produce synthetic return sequences, while a "discriminator" network is trained simultaneously to tell real historical sequences apart from generated ones. As the two compete, the generator gets better at producing paths a discriminator can't distinguish from genuine market data, including realistic volatility clustering and fat tails, not just simple noise.

Running a strategy against hundreds of these synthetic paths, rather than the single historical path, gives a distribution of outcomes instead of one number, exposing fragility that a single backtest can hide.

A worked example

A strategy backtested on 10 years of historical data shows a Sharpe ratio of 1.2 with no drawdown worse than 15%. Run against 500 GAN-generated synthetic paths that share the historical data's statistical properties but aren't literal repeats of it, the same strategy might show a median Sharpe ratio of 0.9 with a worst-case drawdown across paths of 35%, revealing that the historical path was, by chance, kinder to this strategy than a typical draw from the same underlying process would be.