Implementation And Evaluation Of An Order Flow Imbalance Trading Algorithm

Carl Reed Jessen
Northwestern University, Predictive Analytics, Students

December 15, 2015

Abstract:

Building upon the success of their 2010 model, in 2014 Cont et al. published a follow-on paper titled The Price Impact of Order Book Events which found that Order Flow Imbalance (OFI) derived from the limit order book model has a statistically significant correlation to contemporaneous price movement at very short time frames . The purpose of this work is to build and test a predictive model based on Cont’s descriptive work. More specifically, if change in OFI in time interval Screenshot_2 is significantly correlated with price change within the same timeframe, does this correlation similarly hold true on a forward looking basis if price change is advanced one period in the future, Screenshot_1?

Implementation And Evaluation Of An Order Flow Imbalance Trading Algorithm – Introduction

Historical Context of High-Frequency, Low Latency Trading

Since the late 1980’s, electronic trading has been taking an ever increasing share of the global securities exchange market and providing market participants with ever lower trade latencies. In 1989, the world’s first high frequency trading firm, Automated Trade Desk was able execute orders in 1 second, faster than any human trader at the time, through a satellite dish bolted to the roof of the garage the firm was founded in (Philips 2013). A little more than a decade later, in 2000, execution speeds had fallen to 25 milliseconds. By 2010, execution speeds had fallen below 1 millisecond(Cont 2011). Today, electronic traders hold a dominant position in the markets.

An interesting result of the world’s evolution from an entirely manual market to electronic based market systems is that data describing supply, demand, and price behavior in securities markets is being increasing recorded. It is now possible to watch price discovery develop in real time, play it back, and analyze it from every angle. The formulaic and mechanical nature of electronic trading makes statistical analysis of price discovery at very short time intervals possible.

A great deal of interest has developed around modeling imbalances in the market. Specifically, analysis of limit order book dynamics at short intervals has become a topic of interest given the availability of ever increasing granularity of data. Previous academic work by Lee and Ready 1991, and Benedikstdottir 2006, and has shown that changes in order book information can be predictive of future prices changes. These methodologies however, are difficult or impossible to apply under live trading constraints.

In 2010, Rama Cont and his coauthors proposed a stochastic model of the limit order book which conceptualized it as a queuing system. This “stylized version limit order book” model contemplates a limit order book as a continuous-time Markov process in which limit orders arrive and wait in a queue until removed from the book by either cancellation or matched with a marketable order. The model enables an observer to determine the volume of limit orders at each given price level at any given point in time1. Cont’s model is motivated by the desire to use information on the current state of the order book to analyze short term price behavior in a given security.

The Cont model has a set of meaningful advantages for those attempting to analyze trading behavior at subsecond and submillisecond timeframes. It can be estimated quickly using high frequency price data, it creates a model which shares the same features as an empirical order book making easy for a human understand, and the analytical mathematics are trivial given standard computational toolsets.

In the original 2010 paper, Cont et al. found that they were able to meaningful predict changes in price midpoint, execution of an order at the best bid before the best ask quote moves, and execution of both a buy and a sell order at the best quotes before the price moves using a two-sided Laplace transform. Others have also found the Cont model effective and have built upon it.

Lee and Kim applied the Cont model to Korean KOSPI 200 futures market with a slight modification and found modest success2. Avellaneda et al. determined that this model can be used as a baseline to estimate the amount of hidden liquidity in a given marketplace allowing trading venues to be ranked in terms of their “information content”3.

Building upon the success of the 2010 model, in 2014 Cont et al. published a follow-on paper titled The Price Impact of Order Book Events which found that that Order Flow Imbalance (OFI) derived from the limit order book model has a statistically significant correlation to contemporaneous price movement at very short time frames4. The purpose of this work is to build and test a predictive model based on Cont’s descriptive work. More specifically, if change in Order Flow Imbalance in time interval Screenshot_2 is significantly correlated with price change within the same timeframe, does this correlation similarly hold true on a forward looking basis if price change is advanced one period in the future, Screenshot_1?

Order Flow Imbalance
Order Flow Imbalance

See full PDF below.