Trade signing in fast markets

• Date: 01 August 2020
• Author: Madhuparna Kolay,Allen Carrion
• DOI: http://doi.org/10.1111/fire.12218
• Published date: 01 August 2020

DOI: 10.1111/fire.12218

ORIGINAL ARTICLE

Trade signing in fast markets

Allen Carrion1Madhuparna Kolay2

1Department of Finance, Insurance,and Real

Estate, University of Memphis, Memphis,

Tennessee

2PamplinSchool of Business, University of

Portland,Portland, Oregon

Correspondence

AllenCarrion, Fogelman College Administration

Building,Room 402, University of Memphis,

Memphis,TN 38152.

Email:a.carrion@memphis.edu

Abstract

This study assesses the accuracy of trade signing algorithms in fast

trading environments using NASDAQ and NYSE trade and quote

data. Using data that contain true trade signs, we show that the Lee

and Ready algorithm outperforms the tick rule and classifies trades

at least as well as in earlier studies from slower trading environ-

ments, even in subsamples where the market is particularly fast.

We conclude that trade signing remains viable in fast markets, and

that the use of quote data continues to increase trade classification

accuracy.

KEYWORDS

high-frequency trading, Lee and Ready algorithm, market

microstructure, trade classification

JEL CLASSIFICATIONS

C18, G10

1INTRODUCTION

In many research applications, it is important to distinguish whether stock trades are initiated by the buyer or the

seller.Recent studies use this determination to examine the trading strategies of high-frequency traders (HFTs)(Benos,

Brugler, Hjalmarsson, & Zikes,2017), order flows around the 2010 Flash Crash (McInish, Upson, & Wood, 2014), the

impact of make–takefees on market quality and trading behavior (Battalio, Corwin, & Jennings, 2016), anomaly trading

costs (Novy-Marx & Velikov, 2016), the asset pricing implications of the noninformational components of order flow or

inventory effects (Chung & Huh, 2016; Kang & Lee, 2014), and resiliency,defined as the time required for temporary

price impacts to decay (Bessembinder,Carrion, Tuttle, & Venkataraman, 2016). Microstructure data do not often indi-

cate which side initiates trades, so this must be inferred by the researcher with a trade signing algorithm. When both

trade and quote data are available, the predominant approach in the literature has been the Leeand Ready algorithm

(Lee & Ready,1991), which we describe in more detail below.1

1Asof June 1, 2019, Google Scholar lists 3,156 papers citing Lee and Ready (1991).

Financial Review.2020;55:385–404. wileyonlinelibrary.com/journal/fire c

2019 The Eastern Finance Association 385

386 CARRION ANDKO LAY

It is well-documented that speeds have increased dramatically in modern equity markets.2This has led some to

question the continued reliability of the Lee and Ready algorithm (LR algorithm hereafter). According to Easley,López

de Prado, and O’Hara(2012, p. 1466; ELO [2012] hereafter):

In a high-frequency setting, trade classification is much more difficult …order splitting is the norm, cancellations

of quotes and orders arerampant, and the sheer volume of trades is overwhelming …Because the BBO changes

many times between trades, many contracts exchangedat the same price in fact occurred against the bid and

the offer.In this high-frequency world, applying standard algorithms over individual transactions is problematic.

In short, ELO (2012) question both the continued use of quotes to sign trades and the practice of individual trade

signing altogether.They propose a new trade classification technique (bulk volume classification [BVC]) that operates

on aggregated “bars” instead of individual trades.3While ELO (2012) specifically study futures markets, the concerns

they raiseapply to the equity markets as well, and have been met with varied responses from researchers. Many recent

studies simply ignore these criticisms and proceed to sign tradesusing conventional techniques. Some studies continue

to sign trades using the LR algorithm but exclude data from fast marketsin robustness tests (Chung & Huh, 2016) or

choose their sample periods to avoid fast market issues entirely (Abad & Pascual, 2015). Other papers test, use, or

extendtechniques suggested by ELO (2012) and the companion papers, at least implicitly accepting the possibility that

the LR algorithm and other individual tradesigning algorithms should be replaced (Chakrabarty et al., 2015; Panayides,

Shohfi, & Smith, 2019; Poppe, Moos, & Schiereck, 2016).

In this study,we assess the accuracy of the LR algorithm and the tick test in a recent sample of NASDAQ trades and

quotes that identifies true trade signs. This data set is particularly well-suited for this analysis because it also identi-

fies HFT participation, which is a potential source of trade classification inaccuracy,and also contains millisecond (ms)

timestamps. We investigate the following research questions: Does the LR algorithm still accurately sign individual

trades? Does it outperform the simple tick rule, which can be implemented without quotation data? Is millisecond data

necessary for the algorithm to perform well? In cases where only second level resolution data are available, does the

interpolation method proposed by Holden and Jacobsen (2014) offer an improvement over simply matching trades

with the last quote in the prior second?

Our main results are as follows. In the NASDAQHFT database, the LR algorithm outperforms the tick rule when 1-s

timestamps are used, correctly classifying 86.88% of the trades in our sample compared to 78.62% for the tick rule.

When millisecond data are employed, the accuracy improves to 93.57%, which is considerably better than the accu-

racy rate reported in earlier studies conducted in slower markets. Consistent with the conjectures in ELO(2012), the

1-s version tends to be less accuratein fast market conditions. But, even for these trades its performance remains com-

parable with that reported in prior studies and superior to that of the tick test. We find that the Holden and Jacobsen

(2014)interpolation procedure underperforms the basic LR algorithm in 1-s data and does not consistently outperform

thetickrule.

We also conduct robustness tests using the NYSE trade and quote (TAQ) data set, which is widely used by

researchers. The NYSE TAQ data set may be sequenced less accurately than the NASDAQ HFT data set because it

combines trades and quotes from multiple trading venues. Although the TAQdata do not contain true trade signs, we

employ a matching procedure to utilize the true trade signs provided by NASDAQ to conduct this analysis. We find

that the advantage of millisecond timestamps disappears in the TAQdata. LR classification accuracy is actually slightly

higher using TAQ data with 1-s timestamps than with millisecond timestamps. Both procedures classify trades with

a high level of accuracy (85.65% and 84.84%, respectively), but neither approaches the extremely high accuracyrate

2Hasbrouck and Saar (2013) report that HFTscan cancel or submit orders in reaction to market events within 2–3 ms, and Carrion (2013) reports that HFTs

participatein 68.3% of the dollar trading volume. Also see Angel, Harris, and Spatt (2011), O’Hara (2015), and Hasbrouck (2018), among many others.

3BVCis more fully developed and validated in Easley, López de Prado, and O’Hara (2016). BVC is also tested in Chakrabarty,Pascual, and Shkilko (2015) and

Andersenand Bondarenko (2014, 2015).