Generative Models for Synthetic Market Data

In Deep Learning for Market Data you met the central catastrophe of machine learning in finance: the effective sample size is tiny. Twenty years of daily returns is a few thousand autocorrelated, overlapping, non-stationary rows — a starvation diet for a model class that learned its tricks on billions of words and millions of images. There are only two honest ways out of a data shortage. The first is to need less data (simpler models, harder regularization — the discipline of the last course). The second is the audacious one this course is about: manufacture more market. If you cannot find more history, generate it.

That sounds like cheating, and done carelessly it is. But synthetic data is now a serious, load-bearing tool in quantitative finance — for augmenting thin training sets, for generating stress scenarios the historical tape never delivered, for sharing data without leaking private positions, and for backtesting on regimes that have not happened yet. The catch, and the spine of this whole course, is that a generator is only as trustworthy as your ability to evaluate it — and evaluating fake data is far harder than making it. A generator that quietly memorizes its training set passes every distribution test you throw at it and then leaks the future into every backtest downstream. Most of the field’s pain lives in that trap.

This course assumes you carry three things from the prerequisites: the effective-sample-size and overfitting discipline from Deep Learning for Market Data, the simulation toolkit (GBM, bootstrap, variance reduction) from Monte-Carlo Finance, and the stylized facts — fat tails, volatility clustering, autocorrelation structure — from Time-Series Finance. We build on all three, always asking the same skeptical question the ML-for-alpha creed taught you: is this real, or did I just fool myself? The arc:

• Why generate market data — the data-scarcity problem restated as a generation problem, the four real uses (augmentation, scenario generation, privacy, synthetic backtesting), and the two philosophies for making data: learn a generator from history versus simulate a mechanism. The warning that frames everything: a generator you train on your history can leak that history straight back into your tests.
• The stylized facts a generator must reproduce — the non-negotiable checklist (fat tails, volatility clustering, the slow decay of absolute-return autocorrelation, the leverage effect, aggregational Gaussianity, gain/loss asymmetry) and why a plain Gaussian random walk fails almost all of them. This is the rubric every later model is graded against.
• Classical simulators: the baseline to beat — geometric Brownian motion, the block and stationary bootstrap, and regime-switching / hidden-Markov models. Cheap, interpretable, leak-resistant, and already good enough to embarrass a sloppy neural generator — the honest control group.
• GANs for financial time series — the generator-versus-discriminator game, the QuantGAN (TCN-based) and TimeGAN lineages, conditional and Wasserstein variants, and the two demons that haunt them on small financial data: training instability and mode collapse.
• Variational autoencoders for market data — the encode-to-a-latent, decode-back recipe, the reconstruction-versus-KL trade-off, why a smooth latent space buys you interpolation and scenario knobs, and why VAEs tend to blur the very fat tails you most need.
• Diffusion & score-based models for price paths — the noise-then-denoise paradigm that now dominates image generation, adapted to returns: forward corruption, learned reverse denoising, score matching, and the honest question of whether the state of the art in pictures earns its keep on a few thousand noisy returns.
• Evaluating synthetic data without fooling yourself — the hardest lesson. Distributional and dependence tests, the discriminative (“can a classifier tell real from fake?”) and predictive (“train-on-synthetic, test-on-real”) scores, the memorization and nearest-neighbour tests that catch a copying generator, and the leakage minefield where a generator trained on your full history inflates every backtest it touches.

By the end you’ll be able to pick a generator for a job, defend it against the right baseline, and — most importantly — run the evaluation that tells you whether your manufactured market is a genuine asset or an elaborate way to lie to yourself. A graded final exam runs the whole discipline back at you, one locked question at a time.