Researchers at the Oxford-Man Institute of Quantitative Finance (OMI) have used Graphcore’s Intelligence Processing Unit (IPU) to dramatically accelerate the training of advanced price prediction models, using techniques which are typically plagued by computational bottlenecks when run on other types of processor.
The IPU’s designed-for-AI architecture allowed the OMI team to reduce the training times for their multi-horizon forecasting models to the point where they could deliver significant commercial advantage by more accurately estimating market price movements. Such models can be used in the development of alpha for fast trading and in market making strategies.
Abstract:
In this work, researchers Zihao Zhang and Stefan Zohren from the Oxford-Man Institute of Quantitative Finance at the University of Oxford design multi-horizon forecasting models for limit order book (LOB) data by using deep learning techniques.
They adapt encoder-decoder models to generate forecast paths over multiple time steps. An encoder reads through the raw LOB data to extract representative features and a decoder steps through the output time step to generate multi-step forecasts. The experiments suggest that the method delivers superior results compared to state-of-art algorithms. This is due to the iterative nature the decoder delivers, which yields better predictive performance over long horizons as short-term estimates are fed into next prediction through an autoregressive structure.
Encoder-decoder models rely on complex recurrent neural layers that often suffer from slow training processes. Researchers address this problem by using a novel hardware IPU developed by Graphcore which is specifically designed for machine intelligence workload. They conduct a comparison between GPUs and IPUs to benchmark their training speed on modern deep neural networks for LOB data, and observe that IPUs leads to an acceleration that is significantly faster than common GPUs.
Such speed-ups in training time could open up a wide variety of applications, for example, application of online learning or reinforcement learning in the context of market-making, as such a high-frequency trading strategy has strict requirements on communication latency. It would be interesting to deploy IPUs to such setups and test their computational efficiency. Also, the encoder-decoder structure can be applied to a reinforcement learning framework.
“Reinforcement learning algorithms provide an excellent framework to apply such multi-horizon forecasts in an optimal execution or market making setting. Given the computational complexity of such algorithms, the speed-ups using IPUs might even be larger in this setup,” said Zohren in a blog post.