Introducing Improving Robustness of Trading Strategy Backtesting with Boltzmann/GANs for after-work quants

Takeaway. A notebook Sharpe on one history is a point. Lezmi–Roche–Roncalli–Xu train market generators—restricted Boltzmann machines, GANs, convolutional Wasserstein models—to simulate multi-asset paths that keep moments, dependence (copulas), and autocorrelation, then look at the distribution of backtest stats. Robustness, not a prettier equity curve.

Improving the Robustness of Trading Strategy Backtesting with Boltzmann Machines and Generative Adversarial Networks (Edmond Lezmi, Jules Roche, Thierry Roncalli, Jiali Xu, 2020; arXiv:2007.04838).

Fit for learners: One historical path is not a distribution: RBM/GAN market generators turn a backtest number into a risk histogram.

This Hub note is a plain-language intro. Not investment advice.

The problem

You overfit a factor or a rule on 2010–2019, print Sharpe 1.4, and ship. That number is one path. Bootstrap-of-returns often breaks lag structure; plain multivariate normals kill fat tails and clustering. Risk of a strategy is the spread of outcomes, not the in-sample mean.

The idea / core intuition

• Learn P(returns) with an RBM (likelihood) or a GAN (minimax generator vs discriminator); they also review Wasserstein training that is less brittle.
• Synthetic series should match mean, vol, skew, kurtosis, cross-sectional dependence, and ACF.
• Re-run the same strategy on many draws → histogram of Sharpe, drawdown, turnover.
• Goal: risk management of quant strategies (smart beta / ARP), not generating fake prices to trade.

A stress tool. If your edge vanishes on generated twins, it was probably the original sample.

Why after-work Python learners care

You already know np.random.normal. This paper is why that is not a market. Even a toy GAN is heavy; start by checking four moments and a corr matrix on shuffled vs block-bootstrap paths. Then read why they needed generators.

Tiny evening exercise: resample 252 daily returns with np.random.choice (i.i.d.) vs overlapping 20-day blocks; compare Sharpe histograms of a dummy MA rule. Learning only.

One-evening reading checklist

1. What a market generator must preserve besides volatility.
2. RBM likelihood vs GAN two-player training in one sentence each.
3. Why a distribution of Sharpes beats one backtest number.
4. Name one failure mode of i.i.d. bootstrap for strategies with memory.

Download PDF

Download PDF

Landing: arXiv abs

Important disclaimer

Not investment advice. This Hub note is educational only. The paper is academic research on synthetic scenarios—not a recommendation to trade GAN-generated markets or any strategy.

What this is not

• Not CEF / subsidy framing
• Not Quant or Macro course CTA
• Not a “print infinite backtests” pitch

Next step

Single course CTA: SoloQuant Academy’s Introduction to Python Programming—histograms and loops over simulated lists come before GANs.