Can High‐frequency Trading Strategies Constantly Beat the Market?

• Date: 01 April 2016
• Author: Viktor Manahov
• DOI: http://doi.org/10.1002/ijfe.1541
• Published date: 01 April 2016

CAN HIGH-FREQUENCY TRADING STRATEGIES CONSTANTLY BEAT

THE MARKET?

VIKTOR MANAHOV

,

*

,†

The University of York, York, YO10 5GD, UK

ABSTRACT

Policymakers are still debating whether or not high-frequency trading (HFT) is beneficial or harmful to financial markets. We

develop four artificial stock markets populated with HFT scalpers and aggressive high-frequency traders using Strongly Typed

Genetic Programming trading algorithm. We simulate real-life HFT by applying Strongly Typed Genetic Programming to real-

time millisecond data of Apple, Bank of America, Russell 1000 and Russell 2000 and observe that HFT scalpers front-run the

order flow generating persistent profits. We also use combinations of forecasting techniques as benchmarks to demonstrate

that HFT scalping strategies anticipate the trading order flow and constantly beat the market. Copyright © 2015 John Wiley

& Sons, Ltd.

Received 27 February 2015; Revised 16 October 2015; Accepted 21 October 2015

JEL CODE: F3; G10; G12; G14; E47

KEY WORDS: high-frequency trading; algorithmic trading; forecasting and simulation; forecasting methods; market regulation

1. INTRODUCTION

Before the introduction of computers, all trading was performed between humans, often in person on a trading

floor. Over the last few years, we have witnessed rapid transformation of the trading process of securities and other

financial instruments. The traditional floor-based trading is replaced by huge amounts of automated trading. The

application of computers to high-frequency trading (HFT) allows the use of different computer algorithms that ex-

ecute trading orders at superhuman speed (Goldstein et al., 2014). While the blink of a human eye lasts 400 ms, the

trading speed races occur at the level of microseconds (millionths of a second) and even nanoseconds (billionths of

a second). More recently, sophisticated HFT algorithms are able to execute trading orders in 10 μs, which means

40 000 back-to-back trades in the blink of an eye.

However, most studies related to the topic lack the ability to determine which trading messages come from HFT

(Egginton et al., 2012; Hirschey, 2013; Goldstein et al., 2014), making it difficult to estimate the profitability of

high-frequency traders (HFTrs). This is because no publicly available dataset, including NASDAQ 120, which

is the most used research dataset, allows one to directly identify all HFT (Baron et al., 2012). Eggington et al.

(2012) went even further, suggesting that it is hardly possible to identify trading orders generated by computer al-

gorithms in the US equity markets, and therefore, all studies up-to-date used proxies to estimate the level of algo-

rithmic trading in the USA. Also, the number of variables is huge, and the relationships between these variables are

cause–effect imposing another research hurdle (Felker et al., 2014). Despite the fact that HFTrs generate annual

gross trading profits between $2.8bn and $4.1bn (Brogaard, 2011), there appears to be no objective empirical study

that measures the profitability of HFT strategies.

In contrast, this study implements a special adaptive form of the Strongly Typed Genetic Programming (STGP)

and real-time millisecond data of Apple (the most capitalised company), Bank of America (heavily traded by

*Correspondence to: Viktor Manahov, Heslington East, The University of York, York, YO10 5GD, UK.

†

E-mail: viktor.manahov@york.ac.uk

Copyright © 2015 John Wiley & Sons, Ltd.

International Journal of Finance & Economics

Int. J. Fin. Econ. 21: 167–191 (2016)

Published online 25 November 2015 in Wiley Online Library

(wileyonlinelibrary.com). DOI: 10.1002/ijfe.1541

HFTrs) and two indices —Russell 1000 (large-cap stocks) and Russell 2000 (small-cap stocks) —to determine the

level of profit HFTrs generate.

The STGP (described in Appendix A) represents sophisticated trading algorithm that successfully replicates

real-life HFT strategies. According to Dunis et al. (2013), genetic programming models perform remarkably well

even in simple trading exercises. Moreover, Wah and Wellman (2013) argue that questions about HFT implications

are inherently computational in nature because the speed of trading reveals details of internal market activities and

the structure of communication channels.

We reproduce the HFT strategies in an artificial stock market environment where the exact level of profitability

can be measured. In the early 2000s, HFT scalping strategies were initially performed on Chicago futures markets

and then spread over to the equity markets. They originated as relatively simple spread detecting tools that learned

the order book depth and posted on the best bid/ask and then quickly moved to the other side (Patterson, 2012).

These straightforward flipping strategies evolved through time to become modern HFT scalping strategies that

nowadays dominate electronic exchanges, gaining favourable queue position and generating a huge amount of can-

celled orders. The aim of the HFT scalping strategies is to gain a favourable queue position —any particular

scalping strategy must have a high probability of entering the trade and an equally high probability either of exiting

for spread or, if the spread cannot be gained, of immediate exiting in order to avoid losses. The main functional

mechanism of the scalping strategies involves stepping ahead of supply-and-demand imbalances present in market

depth of every round trip trade for the purpose of capturing a micro-spread by closing on the other side for a tick or

alternatively to scratch out by closing on the same side (Bodek, 2013).

Using STGP, we replicate the interactions between HFT scalpers and aggressive HFTrs and compare their fore-

casting performance under the same underlying trading order streams. In other words, we simulate real-life trading

sessions that allow us to avoid the obstacles in the studies discussed earlier. Moreover, we avoid the need of

predicting which trades will be profitable, which is the biggest obstacle in real quantitative trading. Our empirical

findings have important implications for market regulators, academics and the general public.

To summarize, the contribution of this study is threefold. First, this is the first study to use an innovative trading

algorithm and real-time millisecond data to provide concrete empirical evidence of HFT profits. We estimate in

precise quantitative terms the daily profits of HFT after taking into account realistic transaction costs, providing

an advantage over existent studies, such as that of Brogaard (2010), which observed HFT activities in the aggregate

data only, thus preventing them from calculating the exact profitability of HFT.

In order to measure the statistical accuracy and trading, we compare the predictive ability of HFT scalpers and ag-

gressive HFTrs with benchmarks of forecast combination approaches such as support vector regression (SVR),

Kalman filter and least absolute shrinkage and selection operator (LASSO). Second, we take into account all the issues

in previous studies as potentially affecting the reliability oftrading results. The presence of 100000 artificial traders in

our experiment ensures forecasting model stability and lower sensitivity to random factors. All artificial traders l earn

from their experience, evaluating the profitability of trading rules based on their predictive power than in-sample fit.

We avoid data-snooping biases by allowing all trading rules to be evaluated and executed by artificial trad ers.

The remainder of this paper is organized in the following way: Section 2 comprises the background to and lit-

erature review on the topic. Section 3 presents the experimental design of the four artificial stock markets and the

forecasting models we use. Section 3 also describes the data in our experiment. In Section 4, we examine the HFT

scalpers’trading activity and profitability. Finally, Section 5 concludes the paper. Additional clarifying and tech-

nical material can be found in Appendix A to Appendix C.

2. RELATED EMPIRICAL LITERATURE

The vast majority of HFT strategies accumulate only a small profit per trade. Some arbitrage strategies are capable

of generating profits close to 100% of the time, but most of the HFT strategies follow the law of averages principle.

Such strategies might capitalise on only 51% of the trading, but these trades are repeated many times per trading

day resulting in substantial and consistent profits (Jones, 2013). However, studies examining the precise level of

HFT profits are rather rare with most of them using a dataset NASDAQ makes available to academics.

Hendershott and Riordan (2011) analyse the profits of 25 of the largest HFT on the market between 2008 and

2009 and reports average gross trading revenue of $2351 per stock per trading day. However, the authors examine

VIKTOR MANAHOV168

Copyright © 2015 John Wiley & Sons, Ltd. Int. J. Fin. Econ. 21: 167–191 (2016)

DOI: 10.1002/ijfe