Podcast: Man Group’s Zohren on forecasting prices with DeepLOB

Deep learning model can project prices around 100 ticks into the future

Limit order book (LOB) data has long been a source of fascination and frustration for market makers and high-frequency traders. The log of bids and offers in the market undoubtedly contains pertinent information about the next tick in the price of an asset. But the data is so voluminous that it is cumbersome to analyse using traditional techniques.

Quants at the Oxford-Man Institute are using deep learning to tackle this problem. Our guest for this episode of Quantcast is Stefan Zohren, an associate professor at the Oxford-Man Institute and principle quant in charge of futures and foreign exchange execution research at Man Group’s central trading division. He has been working on the project with Zihao Zhang and Stephen Roberts at the Oxford-Man Institute.

Quantcast – a Risk.net Cutting Edge podcast · Stefan Zohren – 26/11/21

 

The technique they have developed is called DeepLOB. In essence, it uses ideas borrowed from image recognition and applies them to financial data. The underlying principle is that an algorithm can recognise shapes in the LOB in the same way it identifies images.

“Imbalances [in the LOB] are the kind of features that these models are able to extract,” says Zohren.

DeepLOB can predict forecast a price at a single future point in time and has already been adopted by a number of financial institutions, including Man Group. In their most recent paper, Zohren and Zhang describe how the technique can be adapted to generate multi-period forecasts.

The innovation in their latest work is the implementation of techniques, such as sequence-to-sequence and attention models, that can learn longer-term relationships and extract features with longer memories. This allows forecasts to be projected further into the future.

The forecasting horizon of the adapted model can extend to about 100 ticks into the future, which in a liquid market might translate to around 30 seconds.

Zohren explains this methodology could be combined with reinforcement learning to compute the paths of optimal execution. That might be the subject of a future study.

Index

00:00 LOBs and the problem of unmanageable data

03:00 Using deep learning to manage LOB data

07:07 Sequence-to-sequence, attention models and long memory

15:10 Real-world applications

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.

Now also available on Spotify.