Convolutional neural networks applied to high-frequency market microstructure forecasting

Abstract

Highly sophisticated artificial neural networks have achieved unprecedented performance across a variety of complex real-world problems over the past years, driven by the ability to detect significant patterns autonomously. Modern electronic stock markets produce large volumes of data, which are very suitable for use with these algorithms. This research explores new scientific ground by designing and evaluating a convolutional neural network in predicting future financial outcomes. A visually inspired transformation process translates high-frequency market microstructure data from the London Stock Exchange into four market-event based input channels, which are used to train six deep networks. Primary results indicate that con-volutional networks behave reasonably well on this task and extract interesting microstructure patterns, which are in line with previous theoretical findings. Furthermore, it demonstrates a new approach using modern deep-learning techniques for exploiting and analysing market microstructure behaviour.