•
Lezmi et al. (2020) use Boltzmann machines and GANs to simulate extra market paths so a strategy’s backtest is less of a one-history lottery. Plain-language intro. Not investment advice.
Takeaway. A single historical path is a tiny sample of what your rule might have lived through. Lezmi, Roncalli, and co-authors train restricted Boltzmann machines and GANs on returns so you can draw synthetic scenarios, re-run the strategy, and see whether the edge survives distributional stress—not just the luck of 2010–2019.
Improving Robustness of Trading Strategy Backtesting with Boltzmann Machines and Generative Adversarial Networks (Edmond Lezmi, Joseph Simonian, Jiali Xu, 2020; arXiv:2007.04838) is a risk-lab paper, not a signal recipe.
Fit for learners: Open-PDF classic/working paper on risk — useful for after-work Python quant learners building intuition before coding.
This Hub note is a plain-language intro. Not investment advice.
Walk-forward still uses one realised world. Overfit rules look brilliant on that world and collapse on the next. Classical bootstrap often breaks dependence (volatility clusters, crashes). You need extra paths that keep fat tails and serial structure without claiming you know the true data-generating process.
This is a concrete reason to learn generative models after you can already code a naive backtest: the object of interest is the metric’s sampling distribution, not a prettier equity curve.
Tiny evening exercise: from 252 daily returns, resample blocks of 5 days 200 times; plot a histogram of a toy long-only Sharpe. Then write one sentence on why a GAN might (or might not) beat block bootstrap. Learning only.
Landing: arXiv abs
Not investment advice. This Hub note is educational only. GAN/Boltzmann backtests are academic robustness tools—not a recommendation to trade any generated scenario or strategy.
Single course CTA: SoloQuant Academy’s Introduction to Python Programming—loops and lists first; then resampling a return series is a natural weekend project.