Emiel Lemahieu | Market Generator models

Market generators are generative machine learning (ML) models with the specificity of modeling financial markets - such as stocks’ spot markets, options’ vol surfaces and limit order books (market microstructure).

The topic has seen a recent surge in interest, and this literature review provides a (non-exhaustive) list of the most important contributions in the field. Moreover, it discusses the following generative ML architectures in detail:

Generative Adversarial Networks (GAN)
Variational Autoencoders (VAE)
Restricted Boltzmann Machines (RBM)
Generative Moment Matching Networks (GMMN)
Normalizing flow-based autoencoders (NF)

Additionally, the literature review describes how each model can be conditioned on some exogenous state variables X to compose a conditional VAE (CVAE), conditional GAN (CGAN), etc. architecture. A good example would be to condition the simulation on the state-of-the-economy by including leading macro-economic indicators. In a portfolio construction context, the differences in optimal portfolios resulting from the conditional simulation can then be attributed to the conditions and sensitivities analyzed. A very simple example is included in the literature review.