Podcast: Blanka Horvath and Gordon Lee on market generator models

In this Quantcast episode, we talk with Blanka Horvath, finance lecturer at King’s College London and co-winner of Risk.net’s ‘Rising star in quant finance’ award, and Gordon Lee, executive director with the quantitative analytics group at UBS.

The conversation focuses on Horvath and Lee's involvement in the development of generator models in finance.

A generator model is a neural network-based technique used to produce a synthetic dataset that closely resembles the statistical properties and dependence structures of the original dataset.

Our discussion follows a previous Quantcast, in which Alexei Kondratyev and Christian Schwarz explained how a restricted Bolzman machine works and what it can be used for.

Horvath has been working on a class of market generators that use ‘variational auto-encoders’ to encode the information contained in a data sample in only a few parameters, and then use those parameters to create new datasets with the same properties as the original.

In a joint work with Hans Buehler, Oxford University's Terry Lyons, Imanol Perez Arribas and Ben Wood, Horvath developed the variational auto-encoder method, and created a version of it that can work with a very low number of data samples. It is able to produce paths, such as price paths, in a way that is consistent with the structure of the time series. The authors used the theory of signatures – a concept introduced by Lyons.

While the formal definition of signatures is rather technical, Lee offers here a simple explanation of how they operate and explains why this generator model is relevant to the calculation of data-intensive quantities such as that of valuation adjustments.

Horvath adds that they are a versatile and very quick to compute. “Signatures are an extremely powerful way of encoding the features of the path” she says, adding, “They capture the most important components of the path, and if you take more and more it will converge to the original.”

The practicality of this research project was clear from the beginning. “We started looking into producing artificial data that we could feed into the deep hedging engine,” explains Horvath, referring to the data-driven methodology developed by Buehler and his co-authors to obtain a hedging strategy without using the classic Black-Scholes formula.

With regards to this particular application, the introduction of generator models allows the user to overcome issues related to datasets having a short history, which is typical of most financial time series. Twenty years of daily prices produce only about 5,000 data points, while the training of the reinforcement-learning algorithm of deep hedging requires many more.

Index

00:00 Intro

02:15 Blanka and Gordon introduce the market generators

05:12 What market generators are suitable for finance?

09:44 What are variational auto-encoders?

16:00 An explanation of the concept of signatures

22:35 Application of signatures

28:00 Application of variational auto-encoders

30:00 Will data-driven models replace the classic models?

33:30 Next steps in research

To hear the full interview, listen in the player above, or download. Future podcasts in our Quantcast series will be uploaded to Risk.net. You can also visit the main page here to access all tracks, or go to the iTunes store or Google Podcasts to listen and subscribe.