Building predictive models is the cornerstone of quantitative finance. However, two major obstacles exit: non-stationary data-generating distribution, and low signal-to-noise ratio. In this talk we will try to address these issues head-on by developing novel deep generative models. Deep generative models have been proven to have the ability to learn extremely complicated data-generating distributions, either explicitly or implicitly. More importantly, they can offer the ability to sample reliably from learned distributions. They open up the door to building robust and reliable predictive models in finance, as well as constructing further strategies based on these models.