Applying deep reinforcement learning (DRL) algorithms such as Deep Hedging to financial markets relies on the availability of large amounts of realistic market data. On a daily time scale the amount of financial data for a single underlying is insufficient for training a DRL agent. In this talk, two novel approaches to simulating financial markets by using a rough paths perspective are presented. The first market simulator presented by Dr. Blanka Horvath pairs conditional variational autoencoders (cVAEs) with signatures and allows sampling realistic spot price paths conditional on the market’s state by inverting sampled market signatures. Afterwards, Magnus Wiese presents a market simulator for generating the dynamics of spot price paths and a high-dimensional grid of discrete local volatilities (DLVs) in a robust fashion. Here, the market simulator is split into two learning modules: a compression algorithm for encoding the grid of DLVs into a low-dimensional orthonormal representation by leveraging autoencoders and signature cumulants, and a generative modelling algorithm for learning to generate spot prices and the encoded grid of DLVs conditional on market information.