Synthetic market data
Manufactured price history used to train or evaluate a strategy when the real thing is too short. Financial data is finite in a way image or text data isn’t: there is exactly one history, it arrives at one bar per bar, and the regime that produced it keeps changing — so the training set can’t be grown by collecting more, only by inventing more.
Two sources in this spoke reach for it independently:
- Augmentation — Koshtenko’s robot expands ~1,000 hours into ~5,000 training examples with noise, time-shifting, scaling and inversion (flipping the series so a long setup becomes a short one). Cheap, and each derived example is a transform of a real one.
- Generation — trademaster ships Market-GAN, and Koshtenko proposes GANs to synthesize 25 years of history he doesn’t have. Ambitious, and the failure mode is the interesting part.
The circularity problem
A generator learns the statistical structure of the history you already have, then produces more of it. Train a strategy on that output and you’ve tested against your own model’s idea of a market, not against a market. It widens the sample, not the distribution — which is precisely the thing that kills strategies live, where the distribution shifts. Augmentation is the more honest version, and even there inversion assumes markets are symmetric between up and down, which is a real assumption about market behaviour smuggled in as a preprocessing step.
No source in the corpus tests whether a strategy trained on synthetic data survives real fills. Until one does, treat synthetic data as a regularizer — a way to stop the model memorizing the one history it has — rather than as evidence about behaviour it has never seen.
Related
multithreaded-ml-trading-robot · trademaster · supervised-learning-trading · backtesting · market-regime-analysis