The aim of this paper is to extend Variational AutoEncoders (VAE) to allow for heavy tailed distribution of the latent space and apply them to the problem of market risk of large portfolios. We extend VAE to incorporate a Student-t distribution as the latent distribution and test their performance on synthetic data. We show that allowing for a latent distribution other than normal improves the VAE performance.
Furthermore, we apply this version of VAE for measuring VaR of a large investment portfolio, by compressing the returns to just a few latent factors and simulating their dynamics with a GARCH model. The performance of such a method is better than that involving PCA. Finally, we apply our VAE to the problem of reconstructing implied volatility curves and show its excellent performance also in that application.