Podcast: Kondratyev and Schwarz on generating data

Market generator models may aid areas of finance where data is limited or sensitive, by generating new data with the same statistical properties.

L to R: Mauro Cesa, Christian Schwarz and Alexei Kondratyev

In this episode of Quantcast, I talk to Alexei Kondratyev, head of the data analytics group at Standard Chartered Bank, and his colleague Christian Schwarz, executive director in the same team.

Kondratyev and Schwarz have co-authored The market generator, in which they introduce a technique to generate synthetic market data that preserves the statistical properties of the original data.

It is a big leap from traditional simulation models, such as those that use Monte Carlo. “The difference between our approach and what was traditionally done in Monte Carlo simulation frameworks is that rather than relying on parametric modelling, we advocate a non-parametric modelling,” Kondratyev explains, implying that the learning process of market scenarios, risk factors and probability are also non-parametric.

The model uses a specific type of generative neural network, called the restricted Boltzmann machine (RBM), to sample from an existing dataset of risk factors and generate a synthetic dataset that replicates the dependence structure of the original data. It is a key achievement that was not possible with parametric models. The ability to capture varying correlations between risk factors is crucial for managing risk and portfolios, and even more so around tail events, when linear correlations break down.

Asked why it is only now possible to implement computationally demanding techniques such as this one, Schwarz says: “There’s a lot more computing power available now that allows us to calibrate this type of machine learning model.

The potential applications of the model are numerous and go beyond the simulation of market data for risk management or other financial purposes.

“You can think of the market generator as a data anonymiser,” says Kondratyev. “For example, data anonymisation in the medical data context is a very important application.”

Schwarz adds that another application is connected to reinforcement learning techniques such as agent-based ones, where the environment where the agent is supposed to learn and act upon is simulated. “In situations where you do not have a lot of data, you can potentially learn the probability distribution of the underlying data and have the agent learn from both real data and synthetic data.”

Index

00:00 Intro

01:45 What is a market generator?

03:15 What are the principles behind this new simulation technique?

08:05 How does your market generator work?

12:22 Application to FX markets and capturing of dependence structures

16:50 Non-financial applications

19:47 Model validation

21:45 Other market generator models

22:43 Model development

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.