Generative Models for Synthetic Market Data — Final Exam

This is the graded finale for Generative Models for Synthetic Market Data, and it threads through everything the course built. You opened with the uncomfortable truth that markets hand you exactly one realized history — autocorrelated, regime-haunted, and far smaller in effective sample size than its row count suggests — and that this scarcity is what makes a synthetic-data generator worth chasing, for augmentation, scenario and stress generation, privacy-preserving sharing, and synthetic backtesting (while never forgetting that a generator trained on your history can quietly leak it back into your tests). You then learned the rubric any honest generator must satisfy — fat tails, volatility clustering, slow-decaying absolute-return autocorrelation, the leverage effect, aggregational Gaussianity, gain/loss asymmetry — and watched a Gaussian random walk fail nearly all of them. From there you climbed the toolkit: the classical simulators (GBM, block and stationary bootstrap, regime-switching HMMs) that form the leak-resistant baseline; the GANs (QuantGAN, TimeGAN, WGAN-GP) with their minimax game and their twin demons, mode collapse and memorization; the VAEs with their ELBO, reparameterization trick, and tail-blurring habit; and the diffusion and score-based models that denoise from pure noise yet sample slowly and stay data-hungry. Finally you assembled the evaluation gauntlet — distributional and dependence tests, the discriminative score, train-on-synthetic TSTR, the nearest-neighbour memorization test, and the leakage minefield — and learned the iron law that no single test suffices. No hints are shown, each answer locks the moment you submit, and your score stays hidden until the very end.

Course Recap

Big picture

Generative Models for Synthetic Market Data — the whole arc

Synthetic Market Data
- 1 · Why generate data
  - Only one realized history, tiny effective sample
  - Four uses augmentation stress privacy backtesting
  - Learn a generator vs simulate a mechanism
  - Leakage trap generator memorizes your history
- 2 · Stylized facts rubric
  - Fat tails and excess kurtosis
  - Volatility clustering in absolute returns
  - Slow decay long memory and leverage effect
  - Gaussian random walk fails almost all
- 3 · Classical simulators
  - GBM thin tails no clustering
  - Block bootstrap cannot exceed history
  - Regime switching HMM calm and turbulent
  - The leak resistant baseline to beat
- 4 · GANs
  - Generator versus discriminator minimax game
  - QuantGAN TimeGAN conditional and WGAN
  - Mode collapse misses crashes
  - Memorization copies training data
- 5 · VAEs
  - Encoder latent distribution decoder reconstructs
  - ELBO reconstruction plus beta KL
  - Reparameterization trick mu plus sigma times noise
  - Blurs fat tails posterior collapse if beta too large
- 6 · Diffusion and score models
  - Forward process adds noise reverse denoises
  - Score view learns gradient of log density
  - Stable and high fidelity covers modes
  - Slow sampling data hungry transfer uncertain
- 7 · Evaluating synthetic data
  - Distributional and dependence stylized fact tests
  - Discriminative score classifier AUC near half
  - TSTR train on synthetic test on real
  - Nearest neighbour memorization test and leakage CV

Seven lessons, one ladder: from why we fake market data to how we prove the fake is honest.

One run, one shot

This is a graded, irreversible exam. There are 24 questions, shown one at a time. The instant you submit a question it locks — there is no Back button, no retry, and no Restart. A wrong answer simply fails that question and the exam moves on; you cannot revisit it. Your running score is hidden until the final screen. The pass mark is 70%. Some questions accept more than one correct option — read every option before you commit, because once you submit you own the answer.

Question 1 of 32

Markets have run for over a century, yet practitioners still complain of 'data scarcity' when building generative models. What is the core reason a long price history is still effectively small?

Old price data is corrupted by survivorship bias and therefore unusable.Daily sampling throws away too many digits of precision to be informative.There is plenty of data; scarcity is a myth invented to justify deep learning.There is only one realized path of history, and its observations are autocorrelated and regime-dependent, so the effective sample of independent stress events is tiny.

Select an answer to continue.

Key Takeaways

Where this leaves you on the quant ladder

Pass or fail, you now hold a skill most practitioners skip straight past: you can build a market-data generator and prove whether it deserves trust. You know why one realized history is too little data, you can recite the stylized-facts rubric a generator must satisfy and watch a Gaussian random walk flunk it, and you can place every tool — GBM, bootstrap, regime-switching, GANs, VAEs, diffusion — on a spectrum from transparent-but-rigid to flexible-but-opaque, with mode collapse, tail-blurring, and memorization as the failure modes to fear. Most of all you carry the iron law of evaluation: distributional, discriminative, TSTR, and nearest-neighbour memorization tests each catch a different lie, no single one is enough, and the whole thing must run inside a purged, embargoed, trial-deflated protocol or the synthetic data simply leaks your history back into your backtest. From here the ladder climbs into full simulation-driven risk and portfolio systems — and you arrive with the habit that matters most: trusting the evaluation gauntlet over a pretty histogram.