Background Subtraction for Pattern Recognition in High Frequency Financial Data

Abstract

Financial markets produce massive amounts of complex data from multiple agents, and analyzing these data is important for building an understanding of markets, their formation, and the influence of different trading strategies. We introduce a signal processing approach to deal with these complexities by applying background subtraction methods to high frequency financial data to extract significant market making behavior. In foreign exchange, for prices in a single currency pair from many sources, we model the market as a low-rank structure with an additive sparse component representing transient market making behavior. We consider case studies with real market data, showing both in-sample and online results, for how the model reveals pricing reactions that deviate from prevailing patterns. We place this study in context with alternative low-rank models used in econometrics as well as in high frequency financial models and discuss the broader implications of the melding of background subtraction, pattern recognition, and financial markets as it relates to algorithmic trading and risk. To our knowledge this is the first use of high-dimensional signal processing methods for pattern recognition in complex automated electronic markets.