Generative Models for Time Series in Finance

A generative model is a statistical model of the joint probability distribution. We built a generative model for univariate time series in finance using a Variational Autoencoder (VAE) neural network architecture. We test the model in SP500 and the Heston Model widely used for option pricing and hedging in the financial industry. In both contexts the generative model is able to correctly learn the density according to our qualitative and quantitative tests . Users have to consider the potential non-stationarity in this context and potentially enrich the learning process. Applications of this generative models include risk management, data augmentation to be used in supervised learning applications and can alleviate the problem of small data in finance. In future researchwe intend to investigate the multivariate case extremely important for portfolio and risk management as well as testing other generative architectures