Generative Models

Read these first

1. 01Denoising Diffusion Probabilistic Models

   Diffusion models generate realistic data by learning to undo noise: train a network to remove a little static from a slightly corrupted image or price path, then chain thousands of tiny cleanup steps starting from pure static to synthesize something new.

2. 02Evaluating Synthetic Market Data

   A generative model can produce price paths that look fine on a chart and are still statistically useless for backtesting, evaluating synthetic market data means checking the specific quantitative fingerprints real markets leave behind, not just eyeballing the picture.

3. 03Latent Variable Models and the ELBO

   Some data is easiest to explain by inventing a hidden cause behind it, a latent variable, but fitting such a model runs into a probability integral nobody can compute exactly, so training instead maximizes a computable lower bound called the ELBO.

4. 04The Reparameterization Trick

   You can't backpropagate through a coin flip, because randomness has no slope, the reparameterization trick rewrites a random draw as a fixed formula applied to noise, so gradient descent can flow straight through it.

Then the rest

1. 05Generative vs Discriminative Modeling
2. 06Mode Collapse in GANs
3. 07Autoregressive Generation: PixelCNN and WaveNet
4. 08Classifier-Free Guidance
5. 09Conditional GANs
6. 10Conditional Variational Autoencoders
7. 11Latent Diffusion Models
8. 12Posterior Collapse and KL Annealing
9. 13Privacy and Memorisation in Synthetic Data
10. 14QuantGAN for Financial Time Series
11. 15Score-Based Models and Langevin Dynamics
12. 16Stylized Facts as Generative Benchmarks
13. 17The GAN Minimax Objective
14. 18TimeGAN for Sequential Synthetic Data
15. 19Vector-Quantized VAE (VQ-VAE)
16. 20Wasserstein GAN and Gradient Penalty