Front‐Running Scalping Strategies and Market Manipulation: Why Does High‐Frequency Trading Need Stricter Regulation?

• Author: Viktor Manahov*
• Date: 01 August 2016
• Published date: 01 August 2016
• DOI: http://doi.org/10.1111/fire.12103

The Financial Review 51 (2016) 363–402

Front-Running Scalping Strategies and

Market Manipulation: Why Does

High-Frequency Trading Need Stricter

Regulation?

Viktor Manahov*

The University of York

Abstract

Regulators continue to debate whether high-frequency trading (HFT) is beneficial to

market quality. Using Strongly Typed Genetic Programming (STGP) trading algorithm, we

develop several artificial stock markets populated with HFT scalpers and strategic informed

traders. We simulate real-life trading in the millisecond time frame by applying STGP to

real-time and historical data from Apple, Exxon Mobil, and Google. We observe that HFT

scalpers front-run the order flow,resulting in damage to market quality and long-term investors.

To mitigate these negative implications, we propose batch auctions every 30 milliseconds of

trading.

Keywords: high-frequency trading, market regulation, market efficiency,algorithmic trading,

evolutionary algorithms, genetic programming

JEL Classifications: G10, G12, G14, G15, G18, G19, G20, G23, G28, G29

∗Corresponding author: Management School, The Universityof York, Heslington East, YorkYO10 5GD,

United Kingdom; Phone: +44 1 904 325847; Fax: +44 1 904 32502; E-mail: viktor.manahov@york.ac.uk.

I thank Richard Warr(the editor) and the three anonymous referees for their constructive comments, which

significantly improved the paper.

C2016 The Eastern Finance Association 363

364 V.Manahov/ The FinancialReview 51 (2016) 363–402

1. Introduction

Wehave witnessed a great transformation in recent years, from a world in which

humans traded face-to-face to one in which computers trade with other computers

(Angel, 2014). The use of computers in high-frequency trading (HFT) has increased

over time, resulting in computer algorithms that execute electronically targeted trad-

ing strategies at superhuman speed (Goldstein, Kumar and Graves, 2014). While the

blink of a human eye lasts 400 milliseconds the current trading speed races occur at

microseconds (millionths of a second) and even nanoseconds (billionths of a second).

Ding, Hanna and Hendershott (2014) suggest that the average time it takes to execute

a market order is approximately 300 microseconds. However, the apparent lack of

conclusive evidence has enabled HFT to operate with limited regulatory understand-

ing, forcing policy makers worldwide to debate whether HFT is beneficial or harmful

to market efficiency (Manahov, Hudson and Gebka, 2014).

Most empirical work lacks the ability to identify which trades and quotes come

from HFT making it difficult to examinehow HFT affects the market and other market

participants (Egginton, Van Ness and Van Ness, 2012; Hirschey, 2013; Goldstein,

Kumar and Graves, 2014). This is because no publicly available data set, including

NASDAQ 120, allows researchers to identify all HFT directly (Baron, Brogaard and

Kirilenko, 2012). Egginton, Van Ness and Van Ness (2012) and Goldstein, Kumar

and Graves (2014) argue that is hardly possible to identify orders generated by

HFT computer algorithms in the U.S. equities markets, and most previous studies

use proxies to measure the level of algorithmic trading and HFT.1In addition, the

huge number of variables and the very complicated cause-effect relation between

these variables and potential outcomes imposes further research obstacles (Felker,

Mazalov and Watt, 2014).

In contrast, this study uses a special adaptive form of Strongly Typed Genetic

Programming (STGP) and real-time and historical millisecond data from the three

most capitalized companies—Apple, Exxon Mobil, and Google. This demonstrates

how HFT scalpers step ahead of investororders, thus breaching the price-time priority

in trade execution and generating a cascade of canceled orders that lead to a deteri-

oration of market quality. Wah and Wellman (2013) argue that questions regarding

the implications of HFT are inherently computational in nature, because the speed

of trading indicates details of internal market activities and the structure of commu-

nication channels. The STGP, described in the Appendix, is a sophisticated trading

algorithm that is suitable for the successful replication of HFT scalping strategies

(the successful application of STGP and Genetic Programming [GP] has been shown

1Some of these proxies include the rate of electronic message traffic normalized by trading volume

(Hendershott, Jones and Menkveld, 2011; Viljoen, Westerholm and Zheng, 2014), the adoption of pro-

prietary data to identify HFT companies (Brogaard, Hendershott, Hunt and Ysusi, 2014) and the use of

account-level trade-by-trade data for identifying HFT (Baron, Brogaard and Kirilenko, 2012; Brogaard,

Hendershott and Riordan, 2013; Hendershott and Riordan, 2013).

V.Manahov/ The FinancialReview 51 (2016) 363–402 365

in several trading exercises).2Usually, financial markets are studied using various

econometric tests and mathematical models based on rational market participants

(representative rational agents). However, the empirical characteristics of financial

markets cannot be fully investigated by such statistical methods. LeBaron (2004)

suggests that large moves, excess kurtosis, and market crashes can only be explored

in a model where market participants and their trading strategies are able to adapt

and adjust over time. In reality, prices of all financial instruments are established by

a large diversity of boundedly rational market participants with different decision-

making rules, risk preferences and time horizons. The complex nature and dynamics

of these market participants and their price formation behavior require a simulation

method that comprises multiple heterogeneous traders and an artificial stock market.

Perhaps the greatest advantage of STGP comes from the ability to execute real-

time trading orders rapidly. When trading is deemed HFT, it involves the implemen-

tation of sophisticated trading algorithms for submitting and canceling orders rapidly

and frequently (Goldstein, Kumar and Graves, 2014). The availability of real-time

data enables forward testing (real-time simulation), where submitted trading orders

change the market depth, and these changes can be observed in real-time. This cannot

be simulated in backtesting with historical data because no econometric technique is

capable of recreating the market reaction to a market depth change. More importantly,

scalping strategies require forward testing with real-time data.3In addition, there is

always broker latency, such as exchange latency or Internet connection latency, with

constantly changing values (it can be 100 milliseconds in one particular moment of

time, and one millisecond five seconds later). Consequently, this process cannot be

simulated in backtesting with historical data by specifying a fixed latency time (e.g.,

100 milliseconds).4

2Dunis, Laws and Karathanasopolous (2013) note that GP models perform remarkably well even in

simple trading exercises. Potvin, Soriano and Valee (2004) use GP to generate short-term trading rules

on the stock markets. Between 1981 and 1995, Neely, Weller and Dittmar (1997) applied GP to six

exchangerates, reporting strong evidence of significant out-of-sample excess returns. Allen and Karjalainen

(1999) developed genetic algorithm to find positive excess returns in the out-of-sample test period of the

Standard & Poor’s (S&P) Composite Stock Index. Dempster and Jones (2001) applied GP to a set of U.S.

Dollar/British Pound spot foreign exchange tick data from 1994 to 1997 to find profitable trading rules

in the presence of realistic transaction costs. Kaboudan (2000) performed another trading exercise with

GP and six stocks to generate relatively high returns on investment.Manahov, Hudson and Hoque (2015)

applied STGP to several financialinstruments to provide new evidence of return predictability. El-Telbany

(2015) implemented GP to forecast Egyptian stock market returns and demonstrates the capability of GP

in generating accurate predictions.

3For instance, a scalping strategy with averageprofit per trade of one basis point (Strategy 1) will generate

fewer realistic results in backtesting with historical data than a buy and hold strategy with 500 basis

points per trade (Strategy 2). This is because one or two basis point variations between backtesting and

forward-testing will havelittle impact on Strategy 2, while the backtesting performance of Strategy 1 could

be completely different from real-time trading due to these variations.

4While broker latency has little impact on buy and hold trading, HFT requires a short holding time. For

example, the average holding time in HFT is several milliseconds (Strategy 1). Buy and hold trading

(Strategy 2) can have a holding time of hours or days. In Strategy 2, the latency factor is negligible;