Kernel Support Vector Machine Learning of Limit Order Book Dynamics for Short Term Price Prediction

Abstract

Support Vector Machines (SVMs) are a prevalent methods of financial time series forecasting. Significant research has been done training SVMs on historical price data to predict future prices over daily time frames. However, the research on intraday financial time series prediction is limited. The methods used in academic studies up to this point seem disconnected from the reality of modern High Frequency Trading. In this study, we train support vector machines on high frequency order book data from the Chicago Mercantile Exchange. We use a combination of novel feature sets, prediction parameters, kernels, and data samples to train support vector machines. We find that support vector machines trained only on order book data have limited but non-trivial predictive power. Additionally, we conclude that the radial basis function and linear kernels are the most powerful for time series prediction using the order book, and that the order book contains more predictive value for short-term predictions than long term.