Lu Chung

Abstract

We present a neural network based generative model for producing financial time series. The model is trained with Maximum Mean Discrepancy (MMD) using a signature kernel as the loss function. The noise input into the model is generated by a moving average model fitted using maximum likelihood estimation on log returns that are “Gaussianised” with the inverse Lambert W transformation. We show that the time series generated by such a model is able to capture most stylised facts of financial time series. Finally, we show an application of the model by using the synthetic data to train a reinforcement learning based agent to trade.

Our speaker

I am a PhD student at the Asian Institute of Digital Finance within the National University of Singapore. My research interests are in using machine learning for portfolio management and trading. This includes the study of techniques in reinforcement learning, deep learning and generative modelling. Prior to starting the PhD program, I spent more than a decade with Credit Suisse (acquired by UBS since) in various roles related to financial derivatives.

To become a member of the Rough Path Interest Group, register here for free.