A Framework for Implementing Predictive Analytics in the High Frequency Trading Realm

Abstract

In this paper, we explore a framework for applying machine learning techniques at the high frequency trading level. In particular, we explore how machine learning algorithms such as recurrent neural networks and elastic-nets can be useful in the high frequency domain for predicting bid and ask movements. Specifically, we focus on analyzing the performance of such models during a high volatility scenario, namely the flash crash of 2010 - we find that elastic nets can be powerful for predicting the volatile price movements that begin to occur soon after the start of a flash crash. We also discover specific features which can warn us about upcoming volatile price movements. Finally, we propose and test a prototype policy for setting off alarms before potential flash crashes, and also discuss the limitations and potential extensions of our methodology.