tick size constraints, high-frequency trading, and liquiditytick size constraints, high-frequency...
TRANSCRIPT
Tick Size Constraints, High-Frequency Trading, and Liquidity
Chen Yao and Mao Ye1
January 20, 2015
Abstract
We argue that the uniform one-cent tick size imposed by SEC rule 612 drives speed
competition in liquidity provision for low-priced securities. A large relative tick size (one
cent divided by price) constrains price competition, creates higher revenue in liquidity
provision, and encourages HFTs to achieve time priority over non-HFTs at constrained prices.
Our difference-in-differences test using splits/reverse splits as exogenous shocks
demonstrates that a large relative tick size harms liquidity and encourages HFT liquidity
provision. These results undermine the rationale for maintaining a uniform tick size and the
recent policy proposal to increase the tick size to five cents.
1Chen Yao is from the University of Warwick and Mao Ye is from the University of Illinois at Urbana-Champaign. Please send all correspondence to Mao Ye: University of Illinois at Urbana-Champaign, 340 Wohlers Hall, 1206 South 6th Street, Champaign, IL, 61820. E-mail: [email protected]. Telephone: 217-244-0474. We thank Jim Angel, Shmuel Baruch, Robert Battalio, Dan Bernhardt, Jonathan Brogaard, Jeffery Brown, Eric Budish, John Campbell, John Cochrane, Amy Edwards, Robert Frank, Thierry Foucault, Harry Feng, Slava Fos, George Gao, Paul Gao, Arie Gozluklu, Joel Hasbrouck, Frank Hathaway, Terry Hendershott, BjΓΆrn HagstrΓΆmer, Yesol Huh, Pankaj Jain, Tim Johnson, Charles Jones, Andrew Karolyi, Nolan Miller, Katya Malinova, Steward Mayhew, Albert Menkveld, Maureen OβHara, Neil Pearson, Richard Payne, Andreas Park, Josh Pollet, Ioanid Rosu, Gideon Saar, Ronnie Sadka, Jeff Smith, Duane Seppi, Chester Spatt, Clara Vega, Haoxiang Zhu, an anonymous reviewer for the FMA Napa Conference, and seminar participants at the University of Illinois, the SEC, CFTC/American University, the University of Memphis, the University of Toronto, HEC Paris, the AFA 2014 meeting (Philadelphia), the NBER market microstructure meeting 2013, the Midway Market Design Workshop (Chicago Booth), the 8th Annual MARC meeting, the 6th annual Hedge Fund Conference, the Financial Intermediation Research Society Conference 2013, the 3rd MSUFCU Conference on Financial Institutions and Investments, the Northern Finance Association Annual Meeting 2013, The Warwick Frontiers of Finance Conference, the China International Conference in Finance, the 9th Central Bank Workshop on the Microstructure of Financial Markets, and Market Microstructure: Confronting many Viewpoints Conference in Paris for their helpful suggestions. We also thank NASDAQ OMX for providing the research data. This research is supported by National Science Foundation grant 1352936. This work also uses the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number OCI-1053575. We thank Robert Sinkovits and Choi Dongju of the San Diego Supercomputer Center and David OβNeal of the Pittsburgh Supercomputer Center for their assistance with supercomputing, which was made possible through the XSEDE Extended Collaborative Support Service (ECSS) program. We also thank Jiading Gai, Chenzhe Tian, Rukai Lou, Tao Feng, Yingjie Yu, Hao Xu and Chao Zi for their excellent research assistance.
1
The literature on high-frequency trading (HFT) identifies two overarching effects of
speed competition: 1) price competition effect: speed allows high-frequency traders (HFTs)
to be the low-cost providers of liquidity; 2) information effect: speed enables HFTs to trade
on advance information and adversely select slow traders (Jones (2013) and Biais and
Foucault (2014)). We contribute to the HFT literature by establishing the tick size constraints
as a new channel for speed competition. From this perspective, speed enables HFTs to
achieve time priority over non-HFTs when price competition is constrained. We argue that
the resulting arms race in speed is a natural consequence of the uniform one-cent tick size
imposed by SEC 612 of regulation NMS, which hinders price competition, generates rents for
liquidity provision, and encourages speed competition to establish time priority at constrained
prices, particularly for securities with a low price or a large relative tick size. This newly
identified channel for HFT leads, in turn, to a new perspective on the policy debate on market
structure: the current policy debate focuses on whether and how to pursue additional
regulation of HFT; our paper, however, demonstrates that HFT can be a response to an
existing regulation.
The main economic mechanism examined in this paper involves a trade-off for
liquidity providers between establishing price and establishing time priority with regard to
the relative tick size. The NASDAQ market studied in this paper is organized as an electronic
limit-order book whereby liquidity is provided by limit orders. The execution precedence of
displayed limit orders follows a price/time priority rule. Price priority means that limit orders
offering better terms of tradeβlimit sells at lower prices and limit buys at higher pricesβ
execute ahead of limit orders at worse prices. Time priority means that limit orders at the
same price are executed in the order which they are submitted. Time priority rewards first-
movers who provide liquidity at a given price. SEC rule 612 (the Minimum Pricing
Increment) of regulation NMS restrict the pricing increment to be no smaller than $0.01 if the
security is priced equal to or greater than $1.00 per share. The uniform one-cent tick size
implies that the cost of establishing price priority over existing limit orders increases with the
relative tick size. For example, holding all else equal, the cost of establishing price priority
over existing limit orders for a $5 stock (with a relative tick size of 20bps) is twenty times
greater than that for $100 stocks (with a relative tick size of 1bps). A large relative tick size
2
can generate three interrelated effects: 1) it hinders price competition on the part of liquidity
providers by imposing a high cost on establishing price priority, and thus it leads liquidity
providers, who would have differentiated themselves by quoting at different prices under a
small relative tick size, to quote the same price. 2) It generates higher rents for liquidity
provision as it stops liquidity providers from bidding the securities to their marginal
valuation. 3) It encourages speed competition on the part of liquidity providers to establish
time priority, as the precedence of execution for limit orders at the same price is determined
by time of submission.
We find that a small relative tick size encourages non-HFTs to establish price priority
over HFTs. We first show that HFTs and non-HFTs are more likely to differentiate on the
best quotes they provide when the relative tick size is small, indicating that, for these cases,
the price priority rule is sufficient to determine the liquidity provision precedence between
HFTs and non-HFTs. More importantly, stocks with a smaller relative tick size are associated
with higher incidences of non-HFTs establishing best quotes relative to that of HFTs. A large
relative tick size, on the other hand, encourages HFTs to establish time priority over non-
HFTs. For instance, for large stocks with a large relative tick size, HFTs and non-HFTs
coincide on the best quotes they provide as much as 95.9% of the time, indicating that for
these incidences, the time priority rule needs to be deployed to determine the execution
sequence. We show that HFTs enjoy a higher percentage of trading volume due to their limit
ordersβ obtaining time priority over price priority than do non-HFTs. The time priority of
HFTs here refers to that non-HFTs display the same quotes at the time of the execution of
HFT orders, but HFTs win execution priority over non-HFTs. As the relative tick size
increases, the difference between HFTsβ and non-HFTsβ percentage of trading volume due to
time priority widens. The comparative advantage that HFTs enjoy by establishing time
priority drives the main results of our paper: the proportion of liquidity provided by HFTs
relative to that of non-HFTs increases with the relative tick size.
Proponents of widening the tick size argue that increasing the tick size would control
the growth of HFT and increase liquidity (Weild, Kim and Newport (2012), but we find that
an increase in the tick size harms liquidity and encourages HFT liquidity provision. The
causal impact of the relative tick size on liquidity and HFT liquidity provision is
3
demonstrated using a difference-in-differences test. The test uses ETF splits/reverse splits as
exogenous shocks to relative tick sizes. The control group involves ETFs that track the same
index, but do not experience splits/reverse splits. We first show that an increase in the relative
tick size after splits harms liquidity. The proportional quotes spread widens; the quoted depth
at the best price increases, but the proportional effective spread, or the actual transaction cost
for liquidity demanders, increases. The impact of splits on liquidity is an important question
that has met with mixed results in the literature (Berk and Demarzo (2013)). We contribute to
this literature by studying the question in a clean setting in which securities with identical
fundamentals but with no splits are used as controls. The results for liquidity also facilitate
the understanding of the economic mechanism underlining the tick size constraints channel.
Stock splits lead to a coarser price grid in a proportional sense, so traders operating at
different price levels before splits may quote the same price after splits. This mechanism
leads to an increase in the proportional spread and the length of the queue to supply liquidity
at the more constrained best bid and offer (BBO). As a result, splits favor HFTs since they
discourage price competition while encouraging speed competition. Reverse splits, on the
other hand, create a finer price grid in the proportional sense, which implies that liquidity
providers who were constrained at the same price can now differentiate themselves by prices,
thus shortening the queue at the best quote. As a result, reverse splits encourage price
competition while discouraging speed competition in liquidity-providing activities. We also
find that reverse splits increase liquidity. We caution that the negative correlation between
HFT liquidity providing activity and liquidity after splits/reverse splits should not be
interpreted as implying that HFTs harm liquidity, because the shock is imposed on the
relative tick size but not HFT activity. Our results do indicate, however, that the effect of the
relative tick size should be taken into account in future studies of HFT on liquidity.
We contribute to the burgeoning literature on HFT by establishing tick size
constraints as a new channel for speed competition. The price competition channel argues
that a speed advantage enables faster traders to provide better quotes, or speed competition
results in price competition. However, the price competition channel cannot explain why
HFT liquidity provision is more active for stocks with large relative tick sizes, as we find that
a large relative tick size is not associated with more incidences in which HFTs establish price
4
priority relative to non-HFTs. The evidence is instead consistent with the tick size constraints
channel, as we find that a large relative tick size leads HFTs and non-HFTs to quote the same
price and helps HFTs establish time priority. More surprisingly, we find that a reduction in
the tick size increases the proportion of best quotes provided by non-HFTs relative to HFTs,
implying that non-HFTs are more likely to provide better quotes when the tick size
constraints are less binding.
As regards the information channel, the literature has documented that quicker access
or quicker response to information generates rents and arms races to exploit such rents (Biais,
Foucault and Moinas (2013) and Budish, Cramton and Shim (2013)). This paper finds that
rents can also be extracted from the tick size: specifically, stocks with a larger tick size have
higher revenue in liquidity provision per dollar volume. It is natural to expect that both HFTs
and non-HFTs have higher incentives to supply liquidity for low-priced stocks, because their
large relative tick sizes increase revenue from limit orders and cost for market orders.
However, the speed advantage of HFTs implies that for limit orders at the same price, orders
from HFTs are more likely to be executed due to their ability to establish time priority.
Therefore, speed allocates the rents created from the tick size. The findings are consistent
with an argument offered by the CEO of the HFT firm Tradeworx, Manoj Narang, who said
that βthe finer the trading increment the more important price priority becomes relative to
time priority. In other words, if the market was designed to trade at continuous (vs. discrete)
non-penny increments you could always win a trade by quoting the best price and the speed
game would be non-existent.β2 We argue that the tick size is one of the drivers of the arms
race in speed, which in turn can directly impact the real economy by reallocating physical
and human capital resources to industries specialized in latency reduction.
More broadly, our paper contributes to the long-standing body of literature that
examines the impact of price controls. Rockoff (2008) summarizes four possible responses
when controls prevent prices from adjusting to natural levels: queuing, black markets,
evading, and rationing. We observe speed competition as a form of queuing through which
traders with the capacity to trade at high speeds compete for a position at the front of the
2 See Credit Suiseeβs HFT 101 with Tradeworx, by Ashley Serrao, Christian Bolu, Dina Shin, and Peter Tiletnick.
5
queue. This economic intuition provides insights into the policy debate on HFT as well as
tick size. Current policy debates on HFT focus on whether additional regulation is required
for HFT. We argue, however, that HFT is a natural response to an existing regulation. We
question the rationale of uniform tick size, as it generates a heterogeneous incentive to
engage in price and speed competition for stocks with different relative tick sizes. We also
question the recent initiative to increase the tick size, which is encouraged by the Jumpstart
Our Business Startups Act (the JOBS Act). The plan for a pilot program that increases the
tick size of 1200 small stocks to five cents has been proposed for comments.3 Proponents of
increasing the tick size first argue that an increase in the tick size would control the growth of
HFT and improve liquidity, which is at odds with our empirical evidence. Another argument
for increasing the tick size is that a larger tick size increases market-making revenue and
supports sell-side equity research and, eventually, increases the number of IPOs (Weild, Kim
and Newport, 2012). Economic theories suggest that constrained prices should facilitate non-
price competition, but we doubt that non-price competition would take the form of competing
on providing the best research, especially when speed exists as a more direct form of non-
price competition. We suggest that the SEC consider running a pilot program that decreases
the tick size for large stocks in addition to the current pilot program that would involve
increasing the tick size to five cents for small stocks.
This paper is organized as follows. Section I discusses the studyβs hypotheses as well
as the empirical strategy for testing them. Section II describes the data used in the study.
Section III presents preliminary results pertaining to the relationship between the relative tick
size and the proportion of liquidity provided by HFTs using double sorting. Section IV
examines the determinants of HFT liquidity provision and revenue using regression analysis.
Section V examines causal relationships between the relative tick size, liquidity, and HFT
using a difference-in-differences test. Section VI provides additional robustness checks of our
results. Section VII concludes the paper and discusses the policy implications.
3 http://www.sec.gov/rules/sro/nms/2014/34-73511.pdf
6
I. Testable hypothesis and empirical design
The main hypothesis of this paper is that a lower nominal price, or a larger relative
tick size, leads to a larger proportion of liquidity provided by HFTs relative to the liquidity
provided by non-HFTs.
Hypothesis 1: An increase in the relative tick size increases the proportion of trading
volume with HFTs as liquidity providers relative to the volume with non-HFTs as liquidity
providers.
The economic mechanism behind hypothesis 1 pertains to the trade-off between price
priority and time priority. A large relative tick size hinders traders from bidding securities to
their marginal valuations. Liquidity providers, who were able to differentiate themselves by
quoting differential prices under a small relative tick size, may have to quote the same prices
under a large relative tick size. The precedence of liquidity provision regarding quotes at the
same prices is determined by time priority, a comparative advantage enjoyed by HFTs. It is
important to note that an increase in the relative tick size may increase the incentive to submit
limit orders for both HFTs and non-HFTs, but non-HFTs are less likely to establish time
priority when they quote the same price as HFTs. Therefore, non-HFTs are βcrowded outβ if
they are less likely to post the order at the top of the queue.
Note as well that hypothesis 1 would not necessarily be true if a speed advantage also
helped HFTs establish price priority. If HFTs could quote better prices than non-HFTs, their
ability to offer strictly better prices would be less binding with a small relative tick size.
Therefore, a reduction in the relative tick size might increase the proportion of liquidity
provided by HFTs. This notion is reflected in a recent editorial report by Chordia, Goyal,
Lehmann, and Saar (2013) that βHFTs use their speed advantage to crowd out liquidity
provision when the tick size is small and stepping in front of standing limit orders is
inexpensive.β The price competition channel has established three mechanisms through
which HFTs are able to provide better prices for liquidity than non-HFTs: HFTs have lower
operations costs (Carrion (2013)) and their speed advantage reduces adverse selection risk
(Hendershott, Jones and Menkveld (2011)) or facilitates better inventory management
(Brogaard, HagstrΓΆmer, NordΓ©n and Riordan (2014)). These three mechanisms suggest that
HFTs can establish price priority when the tick size constraints are less binding. A less
7
benevolent possibility for the ability of HFTs to provide more liquidity when the relative tick
size is small is that they can conduct penny jumping (or quote matching), a practice that
involves stepping ahead of existing limit orders by slightly undercutting their prices. Harris
(2003) mentions that penny jumping is more profitable with a small tick size, and traders who
engage in penny jumping must respond to changes in market conditions faster than passive
traders who they run in front of, so penny jumpers tend to be computerized traders in screen-
based markets.4 Therefore, whether a large relative tick size would increase the proportion of
liquidity provided by HFTs relative to the proportion provided by non-HFTs remains an
empirical question.
Motivated by the argument presented above, our paper further investigates whether a
speed advantage facilitates the establishment of price priority for HFTs, especially when the
relative tick size is small, which leads to hypothesis 2:
Hypothesis 2: The smaller the relative tick size, the higher the probability that the
best quote is provided by HFTs rather than non-HFTs.
The key challenge in testing the tick size constraints hypothesis lies in addressing
possible endogeneity originating from omitted variables or simultaneity bias (Roberts and
Whited, (2012)). Biases induced by omitted variables arise from excluding observable or
unobservable variables that are 1) correlated with the nominal price and 2) have explanatory
power for HFT liquidity provision. Interestingly, in a recent paper Benartzi, Michaely,
Thaler, and Weld (2009) find that the average nominal price for a share of stock on the New
York Stock Exchange has remained roughly constant since the Great Depression. They also
find that several popular hypotheses pertaining to nominal prices, such as the marketability
hypothesis, the pay-to-play hypothesis (or optimal tick size hypothesis), and the signaling
hypothesis, cannot explain nominal prices. Particularly, they do not find that firms actively
manage their nominal price to achieve optimal relative tick size. If firms could choose their
optimal relative tick sizes, they would aggressively split their stocks when the tick size
changes from 1/8 to 1/16 and then to one cent. Such aggressive splits have not occurred in
reality. The main cross-sectional result Benartzi, Michaely, Thaler, and Weld (2009)
4 Lawrence Harris, Trading and Exchanges: Market Microstructure for Practitioners, (Oxford University Press, Oxford), 250.
8
established is that large firms have higher prices. 5 This finding provides the empirical
grounding for our double-sorting test examining the cross-sectional variation in proportional
HFT liquidity-providing activity with respect to the relative tick size in each market cap
category. In addition, to examine how the relative tick size affects the proportion of HFT
liquidity-providing activity related to the total liquidity-providing activity, we perform a
multivariate regression controlling for other variables that may impact the nominal price or
HFT liquidity provision. Notably, we further develop a clean difference-in-differences
specification to test this hypothesis. The test uses splits/reverse splits as exogenous shocks to
the nominal prices of ETFs, where the ETFs that split/reverse split are in the treatment group
and ETFs that track the same index but do not split/reverse split are in the control group.
In our scenario, simultaneity bias arises when an increase in the proportion of
liquidity provided by HFTs leads to a lower nominal price, or when both the proportion of
HFT liquidity provision and the nominal price are determined in equilibrium. Such a case of
reverse causality is unlikely to exist in reality: to the best of our knowledge, no evidence
indicates that HFT liquidity provision has the first-order effect of reducing the nominal price.
In addition, the difference-in-differences test involving ETFs splits/reserve splits alleviates
the simultaneity bias concern.
The difference-in-differences test is particularly interesting not only because the
shocks are exogenous but also because the change in HFT activity after splits/reverse splits
relative to control group activity should not be driven by informational factors. We
acknowledge the importance of the information channel, and the diversity of HFTs
(HagstrΓΆmer and NordΓ©n, (2013)) implies that HFT activities can be driven by any of several
factors. However, ETFs tracking the same index have the same fundamental information, and
this identification strategy allows us to establish the tick size constraints channel as one
additional driver of speed competition in addition to information.6
5 Benartzi, Michaely, Thaler, and Weld (2009) also find that a firm splits when its price deviates from those of other firms in the same industry. However, there are no splits in our 117 NASDAQ HFT sample. Also, it is not clear how the price of other firms in the same industry is determined. We control for the industry by time fixed effects in the regression section. 6 Certainly, one can argue that when the relative tick size is larger, the information about the depletion of the depth at the best quotes becomes more important. See Foucault, Kadan and Kandel (2013).
9
The textbook in corporate finance by Berk and Demarzo (2013) states that the
literature finds mixed results on the impact of splits on liquidity. Our difference-in-
differences test contributes to this literature by studying the question in a clean setting in
which securities with identical fundamentals but with no splits/reverse splits are used as
controls. Our conjecture regarding liquidity is motivated by the theoretical argument of
Foucault, Pagano, and RΓΆell (2013), which shows that a large tick size increases the spread as
well as the depth at a given price level. Therefore, we have our third hypothesis:
Hypothesis 3: Splits increase the quoted spread, depth at the best price, and effective
spread as well as HFT liquidity-providing activity; reverse splits reduce the quoted spread,
depth at the best price, and effective spread as well as HFT liquidity-providing activity.
The hypothesis pertaining to reverse splits is the mirror image of the hypothesis
pertaining to splits under tick size constraints.
The final hypothesis relates to revenues taken by HFTs. One fundamental question in
finance is whether activities in financial markets affect real resource-allocation decisions.
The literature has revealed several indirect mechanisms through which the market
microstructure affects real decisions. 7 Speed competition attracts the attention of a broader
finance and economics audience because of its direct impact on real resource allocation. A
recent article in the Financial Times estimates that a 1-millisecond advantage is worth up to
$100 million in annual gains.8 The common belief is that profits are generated by quick
access to information. The policy debate, as a consequence, focuses on the fairness and
legitimacy of HFTs having ready access to information. Our paper takes a different route: we
argue that regulating the tick size can also lead rents to faster traders and drive the arms race
in speed. Harris (1994) and Foucault, Pagano, and RΓΆell (2013) argue that market-making
profits increase with the tick size. Anshuman and Kalay (1998) argues that all economic rents
from the spread are eliminated in equilibrium by competing liquidity providers with
7 One line of research relates market microstructure to asset pricing, which identifies three channels: liquidity, information risk, and ambiguity (OβHara (2003);; Easley and OβHara (2010)). The other related strand of the literature connects liquidity and price discovery to corporate decisions. For liquidity on corporate investment, see Fang, Noe, and Tice (2009) and Ellul and Pagano (2006), among others. Bolton and von Thadden (1998) and Maug (1998) associate liquidity with corporate governance. Bond, Edmans and Goldstein (2011) offer a useful survey on the informational effects of market prices on firms. 8 βSpeed fails to impress long-term investors,β Financial Times, September 22, 2011.
10
continuous price. The tick size creates positive expected revenue by posing a barrier to such
competition. The above argument leads to our fourth hypothesis:
Hypothesis 4: Liquidity-providing revenue increases with the tick size.
Notice that both HFTs and non-HFTs should have stronger incentives to submit limit
orders for low-priced stocks if the revenue from limit orders increases with the relative tick
size, but HFTsβ speed advantage gives them a comparative advantage for establishing time
priority.
II. Data and institutional details
This paper uses three main datasets: a NASDAQ HFT dataset, the NASDAQ
TotalView-ITCH with a nanosecond time stamp, and Bloomberg. CRSP, Compustat and
I/B/E/S are also used to calculate stock characteristics. The sample period is October 2010
unless indicated otherwise.
A. Sample of stocks and NASDAQ HFT data
The NASDAQ HFT dataset provides information on limit-order book snapshots and
trades for 117 stocks selected by Hendershott and Riordan. The original sample includes 40
large stocks from the 1000 largest Russell 3000 stocks, 40 medium stocks ranked from 1001β
2000, and 40 small stocks ranked from 2001β3000. Among these stocks, 60 are listed on the
NASDAQ and 60 are listed on the NYSE. Since the sample was selected in early 2010, three
of the 120 stocks were delisted before the sample period (BARE, CHTT and KTII). Panel A
of Table I contains the summary statistics for the 117 stocks.
Insert Table I about Here
The limit-order book data offer one-minute snapshots of the book from 9:30 a.m. to
4:00 p.m. for each sample date and stock. The displayed liquidity for each of these 391
snapshots is classified into two types: liquidity provided by HFTs and liquidity provided by
non-HFTs. This paper focuses on analyzing the displayed quotes from HFTs and non-HFTs
at the best bid and ask.9 The trade file provides information on whether the traders involved
9 At each price level, hidden orders have lower priority than displayed orders, even if the hidden orders were entered before displayed orders.
11
in each trade are HFTs or non-HFTs. In particular, trades in the dataset are categorized into
four types, using the following abbreviations: βHHβ: HFTs who take liquidity from other
HFTs;; βHNβ: HFTs who take liquidity from non-HFTs;; βNHβ: non-HFTs who take liquidity
from HFTs;; and βNNβ: non-HFTs who take liquidity from other non-HFTs. Therefore, the
trade data allow us to break out trading volume into two types: volume due to liquidity
provision from HFTs and volume due to liquidity provision from non-HFTs.
NASDAQ data is the best publicly available HFT data on the U.S. equity market, but
it suffers from three limitations. Nevertheless, as we illustrate below, the data limitations are
unlikely to affect the findings in this paper.
The first limitation of the data is that the NASDAQ identifies a firm as an HFT firm if
it engages only in proprietary trading. HFT desks in large and integrated firms (e.g., Goldman
Sachs and Morgan Stanley) may be excluded because these institutions also act as brokers for
customers and engage in proprietary low-frequency strategies, so their orders cannot be
uniquely identified as HFT or non-HFT business. One other omission involves orders from
small HFTs that route their orders through these integrated firms (Brogaard, Hendershott and
Riordan (2014)). Nevertheless, the inclusion of some HFTs in the non-HFT group tends to
bias the estimate of their differences towards zero. NASDAQ HFT data thus underestimate
the true differences between HFTs and non-HFTs and attenuate our findings. The fact that we
still detect economically and statically significant differences between their activities
demonstrates the robustness of our results.
Second, the data contain information on NASDAQ trading only, not on other
exchanges. As Battalio, Corwin, and Jennings (2014), Yao and Ye (2014) and Chao, Yao and
Ye (2014) indicate, one way to establish price priority over standing limit orders on the
NASDAQ is to submit a limit order at the same price to a taker/maker market.10 For two limit
orders at the same price, one submitted to the NASDAQ and the other submitted to a
taker/maker market, if liquidity takers consider the liquidity rebate, the one submitted to the
taker/maker market has price priority over the one submitted to the NASDAQ. However, the
exclusion of other exchanges is unlikely to affect our argument regarding the economic trade- 10 EDGA and the Boston Stock Exchange are two examples of taker/maker markets. On the NASDAQ market, liquidity providers are subsidized, but liquidity takers are charged in the exchange, while in a taker/maker market liquidity providers are charged while liquidity takers are subsidized.
12
off between price priority and time priority for stocks with heterogeneous price levels. To see
this, consider the following two facts: 1) The minimum price variation in a taker/maker
market is one cent as well for stocks above $1. In percentage terms, the minimum variation is
higher for lower-priced stocks, and the cost to establish price priority within a taker/maker
market decreases with the price. 2) The cost of establishing price priority by moving an order
from NASDAQ to a taker/maker market is higher for low-priced stocks, because the make-
take fee of given exchanges are also uniform for stocks priced above $1. For example, in
October 2010, the most active liquidity provider obtained a rebate of 0.295 cents for limit
orders on the NASDAQ, whereas liquidity providers needed to pay 0.02 cents per share for
limit orders on the Boston Stock Exchange. Therefore, the cost of establishing price priority
by posting an order at the same price on the Boston Stock Exchange is 0.315 cents for stocks
priced above $1. In percentage terms, the cost is higher for lower-priced stocks. In this
regard, the fee and rebate function similarly to the tick size. In short, the economic trade-off
captured in this paper holds within markets as well across markets, and the existence of
taker/maker markets should not confound the economic argument of this paper.
Finally, data based on snapshots do not provide a dynamic view of the limit-order
book. This prevents us from separating the presence of the best quotes in the limit-order book
that are due to one traderβs actively improving quotes from those that are due to other tradersβ
passively withdrawing from the best quotes. Fortunately, the NASDAQ also provides us with
five days of data based on all quote updates from HFTs and non-HFTs from February 22,
2010 through February 26, 2010. The robustness checks in section VI using this small data
sample show consistent results.
B. Sample of stocks and NASDAQ ITCH data
We use a difference-in-differences approach to clearly identify the causal impact of
the relative tick size on HFT liquidity provision. The test uses leveraged ETFs that have
undergone splits/reverse splits as the pilot group, and uses leveraged ETFs that track the same
indexes and do not split/reverse split in our sample period as the control group. Leveraged
ETFs are issued prominently by Proshares and Direxion, and often appear in pairs that track
the same index but in opposite directions. For example, the ETFs SPXL and SPXS both track
13
the S&P 500, but SPXL amplifies S&P 500 returns by 300% while SPXS does so by -300%.
These twin leveraged ETFs usually have similar nominal prices when launched for IPOs, but
the return amplification results in frequent divergence of their nominal prices after issuance.
The issuers often use splits/reverse splits to keep their nominal prices aligned with each other.
We search Bloomberg Database to collect information on leveraged ETF pairs that
track the same index with an identical multiplier, and the data are then merged with CRSP to
identify their splitting/reverse splitting events. We identify 11 splits and 33 reverse splits
from January 2010 through December 2013.11 Reverse splits occur more frequently, because
their issuers are often concerned about the higher trading cost of low-priced ETFs.12 Our
empirical analysis provides evidence that supports this concern.
Since the NASDAQ HFT dataset does not provide HFT information for ETFs, we
compute HFT activities based on methodologies introduced by Hasbrouck and Saar (2013)
using NASDAQ-ITCH data. We also use ITCH data to construct a limit-order book at
nanosecond-scale resolution, upon which the calculation of liquidity is based. Details on how
to link these messages can be found in Gai, Yao, and Ye (2013) and Gai, Choi, OβNeal, Ye,
and Sinkovits (2013). Summary statistics for leveraged ETFs are presented in Panel B of
Table I.
III. Double sorting
Our preliminary analysis starts with examining the cross-sectional pattern of liquidity
provision with respect to the relative tick size in each market cap category. We sort the 117
stocks first into small, medium, and large groups based on the average market cap in
September 2010, and each group is further divided into low, medium, and high sub-groups
based on the average closing price in September 2010. Section III.A presents the results for
the provision of best quotes by HFTs and non-HFTs with respect to the relative tick size, and
Section III.B presents the percentage of volume with HFTs as liquidity providers with respect
to the relative tick size.
11 To ensure sufficient trading volume in these ETFs, we use leveraged ETFs that experience at least 50 trades each day. 12 βWhy has ProShares decided to reverse split the shares of these funds?β (http://www.proshares.com/resources/reverse_split_faqs.html)
14
A. Provision of the Best Quotes
The uniform one-cent tick size hinders price competition involving liquidity providers,
particularly for stocks with low prices, or stocks with large relative tick sizes. It is natural to
expect that a small relative tick size facilitates price differentiations between non-HFTs and
HFTs, but it is not clear in the literature whether a smaller tick size leads HFTs to establish
strictly better prices than non-HFTs do or vice versa. The price competition channel has
established three mechanisms through which HFTs are able to provide better prices for
liquidity than non-HFTs: HFTs have lower operational costs (Carrion (2013)), and their
speed advantage either reduces adverse selection risk (Hendershott, Jones and Menkveld
(2011)) or facilitates better inventory management (Brogaard, HagstrΓΆmer, NordΓ©n and
Riordan (2014)). Holding all other things equal, these three mechanisms suggest that the
presence of HFTs as unique providers of best quotes relative to slow traders would increase,
or at least not decrease, with the reduction in the relative tick size, because the constraints on
offering strictly better prices is less binding with a small relative tick size. Also, a small
relative tick size may facilitate penny-jumping from HFTs, which can also increase the
presence of HFTs in the best quotes.
The one-minute snapshots of the limit-order book in the NASDAQ high-frequency
book enable us to examine the providers of the best quotes with respect to the relative tick
size. The data indicate the depth provided by both HFTs and non-HFTs at each bid and ask
price. Our analysis starts by categorizing the best price (the bid and the ask are treated
independently) in these 391 minute-by-minute snapshots for each stock and each day into
three groups according to the following criteria: 1) the best price is displayed by HFTs only,
2) the best price is displayed by non-HFTs only, and 3) the best price is displayed by both
HFTs and non-HFTs. The number is then averaged across all the stocks in each portfolio for
each day, resulting in 21 daily observations for each of the 3-by-3 portfolios. Column 1 of
Table II presents the percentages of time that HFTs are unique providers of the best quotes,
Column 2 presents the percentages of time that non-HFTs are unique providers of the best
quotes, and column 3 presents the percentages of time that HFTs and non-HFTs both display
the best quotes. Columns 1, 2, and 3 sum to 100 percent. Column 4 demonstrates the ratio of
column 2 to column 1. Column 5, defined as column 2 minus column 1, shows the
15
differences in percentage between non-HFTs as unique providers of the best price and HFTs
as unique providers of the best price. Column 6 contains the statistical inferences for column
4 based on 21 daily observations.
Insert Table II about Here
Table II shows that HFTs and non-HFTs are more likely to quote identical prices for
stocks with large relative tick sizes. The best quotes provided by HFTs and non-HFTs are the
same 95.9% of the time for low-priced large stocks, while best quotes provided by HFTs and
non-HFTs are the same only 45.5% of time for high-priced large stocks. In other words, price
priority is sufficient to differentiate two types of traders only 4.1% of the time for low-priced
large stocks, but it is sufficient 54.5% of the time for high-priced large stocks. This result
implies that a large relative tick size is associated with a lower level of price differentiation.
This evidence is consistent with the notion that price competition is more constrained for
stocks with a large relative tick size.
Surprisingly, we find that, as the relative tick size decreases, non-HFTs play an
increasingly more prominent role than HFTs in providing the best price, though the pattern is
not monotonic for small stocks (columns 5 and 6). The evidence does not support hypothesis
2 that the smaller the relative tick size, the higher the probability that the best quote is
provided by HFTs rather than non-HFTs.13
A recent paper by OβHara, Saar, and Zhong (2013) finds that HFT is more likely to
improve the BBO for stocks with large relative tick sizes than for stocks with small relative
tick sizes. Their paper differs from ours from two perspectives. First, OβHara, Saar, and
Zhong (2013) examine the cross-sectional variation of HFTs within their own group, while
we compare the activity of HFTs with that of non-HFTs. Second, and more importantly, our
13 The main purpose of the current paper is to empirically document the comparative advantage of HFTs over non-HFT in terms of price or time priority. Formal modeling of the rationale that makes non-HFTs more willing to establish price priority is beyond the scope of the paper. Nevertheless, we offer the following two possibilities. First, a speed advantage may reduce HFTsβ incentives to improve quotes. The U.S. market prioritizes price over time. HFTs can achieve time priority when they quote the same price as non-HFTs, which reduces their incentive to undercut the price; non-HFTs are likely to lose time priority when they quote the same price as HFTs, which gives them a greater incentive to improve the price to gain price priority. Second, a recent paper by Frazzini, Israel, and Moskowitz (2012) suggests that a large fundamental institutional investor designs trading algorithms to provide rather than demand liquidity. These natural buyers or sellers can offer more aggressive limit orders when price competition is less constrained.
16
paper focuses on the trade-off between price priority and time priority. The fact that HFTs are
more likely to be the first to improve the BBO does not necessarily imply that they are more
likely to establish price priority. To the best of our knowledge, the literature does not
distinguish between the following two scenarios. In the first scenario, both HFTs and non-
HFTs concurrently intend to improve the BBO, but HFTs move faster and thus gain the first
place in the queue for the new BBO, while non-HFTs join the queue for the new BBO after
HFTs do. In the second scenario, only HFTs intend to improve the BBO, and non-HFTs
neither match nor undercut that price. HFTs improve the BBO in both cases, but only the
second case entails price competition, whereas the first case pertains to speed competition
between HFTs and non-HFTs. Therefore, to draw a conclusion about who is the best quote
provider, we need to examine the chances that one type of trader provides a better price than
the other type, particularly when the constraints on price competition are less binding.
Table II presents the presence of HFT and non-HFT on the best quotes and, as a
robustness check, we also examine the best depth, or the quantity provided at the best price in
Table III. We denote the displayed depth provided by HFTs and non-HFTs as HFTdepthitm,
NonHFTdepthitm, where i is the stock, t is the date, and m is the time of day. The share-
weighted average first sums the HFT liquidity provision for all stocks in the portfolio and
then divides the number by the total liquidity provision for all stocks in the portfolio for each
day.14
The average depths from HFTs and non-HFTs for stock i on day t are:
π»πΉπππππ‘β = 1π π»πΉπππππ‘β
ππππ»πΉπππππ‘β = 1π ππππ»πΉπππππ‘β
(1)
The depth provided by HFTs relative to the total depth of portfolio J on day t is:
14 The depth result is share-weighted. We also try equal-weighted depth and the results are similar.
17
π»πΉπππππ‘βπ βπππ = β π»πΉπππππ‘βββ (π»πΉπππππ‘β + ππππ»πΉπππππ‘β )β
(2)
Table III shows the average percentage of depth at the BBO provided by HFTs for
each of the market capβbyβrelative tick size portfolios. The result indicates that the
percentage of depth provided by HFTs increases monotonically with the relative tick size.
The depth from HFTs runs as high as 55.66% for large stocks with large relative tick sizes,
while the figure is only 35.07% for large stocks with small relative tick sizes. The difference
is 20.59% and the t-statistic based on the 21 observations runs as high as 22.10. The depth
percentage provided by HFTs is 39.73% for mid-cap stocks with large relative tick sizes,
while the figure is 24.61% for mid-cap stocks with small relative tick sizes. The difference is
15.13% with a t-statistic of 22.88. As the percentage of depth at the BBO offered by non-
HFTs is 1 minus the percentage of depth at the BBO offered by HFTs, it is easy to see that
stocks with a smaller relative tick size have a larger ratio of depth at the BBO offered by non-
HFTs to that offered by HFTs.
Insert Table III about Here
B. Tick size constraints and volume
NASDAQ high-frequency data indicate, for each trade, the maker and taker of
liquidity. Recall that NHit, HHit, HNit, and NNit are the four types of share volumes for stock i
on each day t. For each portfolio J on day t, the volume with HFTs as liquidity providers
relative to total volume is defined as:15
π»πΉππππππππ = β (ππ»β +π»π» )β (ππ»β +π»π» +π»π + ππ ) (3)
Table IV shows the average percentage of volume with HFTs as liquidity providers
for each of the market capβbyβrelative tick size portfolios. The result demonstrates that the
volume with HFTs as liquidity providers increases with the relative tick size. For example,
about half of the volume is due to HFTsβ being liquidity providers for large stocks with large
relative tick sizes, but only 35.93% of the volume is due to HFTsβ being liquidity providers
15 The equal-weighted average yields similar results. As a falsification test, in section VI.B we also perform double sorting for the percentage of volume with HFTs as liquidity takers.
18
for large stocks with small relative tick sizes. The difference is 14.03% with a t-statistic of
15.54 for the 21 observations.
Insert Table IV about Here
The pattern documented in table IV cannot be explained by the price competition
channel. Table II shows that HFTs have the lowest probability to establish price priority for
stocks with large relative tick sizes. For example, HFTs are the unique best quote providers
only 1.6% of the time (the first row of table II) for large stocks with large relative tick size.
However, table IV shows that proportion of volume with HFTs as liquidity providers is the
highest for this portfolio. The pattern is nevertheless consistent with the tick size constraints
channel: non-HFTs and HFTs quote the same price 95.9% of the time for large stocks with
large relative tick size, where time priority plays a prime role in determining the execution
sequence. As the relative tick size decreases, the volume with non-HFTs as liquidity
providers increases, which also dovetails with the result shown in Table II that a small
relative tick size facilitates non-HFTs to establish price priority. Section VI.B provides
further evidence that HFTs are more likely to establish time priority than non-HFTs,
especially when the relative tick size is large.
IV. Regression analysis
The multivariate regressions in this section further confirm that a large relative tick
size results in a higher percentage of liquidity provided by HFTs. We control for additional
variables to overcome omitted variable bias. A sufficient condition for omitted variable bias
to occur is that the missing variables are correlated with both the nominal price and HFT
liquidity provision. We address the issue of omitted variables based on the necessary
condition, and control for variables that can be potentially correlated with either the nominal
price or HFT liquidity provision.16 We also demonstrate in this section that revenues taken
for liquidity provision are also higher for stocks with large relative tick sizes.
16 To the best of our knowledge, the literature has not documented any variables correlated with both the nominal price and HFT liquidity provision.
19
A. Control variables
The nominal price literature suggests that the industry norm is important in choosing
the nominal price. Benartzi, Michaely, Thaler, and Weld (2009) find that a firm may
split/reverse split if its price deviates from the industry average. We control for this average
using an industry fixed effect, whereby industries are classified using the FamaβFrench
classification of 48 industries. Although Benartzi, Michaely, Thaler, and Weld (2009) argue
that other hypotheses cannot explain nominal prices, to run a robustness check we
nevertheless take five lines of studies in the nominal price literature into consideration, three
of which suggest additional control variables for our analysis.17 The summary of control
variables from these hypotheses is presented in Table I.
The optimal tick size hypothesis argues that firms choose the optimal tick size
through splits/reverse splits (Angel (1997)). Anshuman and Kalay (2002) also build a
theoretical model to demonstrate the existence of optimal tick size under certain parameter
values. This hypothesis implies that firm characteristics can determine both HFT liquidity
provision and the relative tick size. However, the optimal tick size hypothesis has been
rejected in the following experiments. We have not seen that firms aggressively split their
stocks to maintain their optimal relative tick size when the nominal tick size changes from
1/8 to 1/16 and then to one cent. (Benartzi, Michaely, Thaler, and Weld (2009)). Nevertheless,
we include the idiosyncratic risk, and the number of analysts that may affect the choice of the
optimal tick size, from this study.18
The marketability hypothesis argues that a lower price appeals to individual traders.
Tests of this hypothesis yield mixed results.19 Nevertheless, we include the measure of small
17Two other lines of research do not suggest additional variables to control for in our study. The catering hypothesis proposed by Baker, Greenwood, and Wurgler (2009) discusses time-series variations in stock prices: firms split when investors place higher valuations on low-priced firms and vice versa, but our analysis focuses on cross-sectional variation. Campbell, Hilscher, and Szilagyi (2008) find that an extremely low price predicts distress risk, but the 117 firms in our sample are far from default, and the distress risk should not affect the the ETFs sample in Section V. 18 Angel (1997) also argues that the relative tick size also depends on whether a firm is in a regulated industry, and this effect has been taken care of by including an industry-by-time fixed effect. It uses firm book value as a control for size, which is similar to the market cap for which we have controlled. When book value is included as an additional control, the results are similar. 19 Some papers find that individuals prefer low-priced stocks (Dyl and Elliott,( 2006)), whereas Lakonishok and Lev (1987) find no long-term relationship between nominal prices and retail ownership. Byun and Rozeff (2003) suggest that if there are any short-term effects of low prices, they are very small.
20
investor ownership suggested by Dyl and Elliott (2006), which is equal to the logarithm of
the average book value of equity per shareholder.
The signaling hypothesis states that firms use stock splits to signal good news. The
empirical literature, however, does not reach a definitive conclusion as to whether splits serve
as a signal or, if so, what types of news prompt firms to signal.20 In addition, the 117 stocks
in our sample do not split during our sample period. Nevertheless, we use the probability of
informed trading (PIN) proposed by Easley, Kiefer, OβHara and Paperman (1996) to control
for information asymmetry.
We then include additional control variables from the HFT literature in our regression,
such as turnover and volatility following Hendershott, Jones, and Menkveld (2011). The
literature has reached a consensus that HFTsβ speed advantage allows them to reduce the
pick-off risk by canceling their quotes before being adversely selected, but it is intriguing to
further examine whether HFTs take a higher or lower market share for stocks with higher
PINs. Therefore, in addition to its role as a proxy for the signaling hypothesis, the coefficient
estimate for the PIN also sheds light on the activeness of HFT on stocks with higher
asymmetric information. The regression also includes past returns as the independent variable
to examine the impact of returns on HFT liquidity provision. 21 The method used for
calculating the variables from the HFT literature is also summarized in Table I.
B. Regression results for HFT liquidity provision.
The regression specification is
π»πΉππ£πππ’ππ , = π’ , + π½ Γ π‘πππ , + Ξ Γ π , +π , (4) where π»πΉππ£πππ’ππ , is the proportion of volume with HFTs as liquidity providers,
calculated as the volume with HFTs as the liquidity providers over the total trading volume
for stock i on date t.22 π’ , is the industry-by-time fixed effect.23 The key variable of interest,
π‘πππ , , is the daily inverse of the stock price for stock i. π , are the control variables 20 See Brennan and Copeland (1988), Lakonishok and Lev (1987), and Kalay and Kronlund (2013). 21 We thank an anonymous reviewer for the Texas Finance Festival for the suggestion of adding past returns. 22 We also analyze who provides the best price using multinomial logit and probit models and obtain similar results. The results are not reported but are available upon request. 23 We do not include firm fixed effects because the focus of this paper is on understanding the cross-sectional variation in HFT market-making activity, and firm fixed effects defeat this purpose (Roberts and Whited (2012)).
21
presented in Table I. The results are presented in table V. Column 1 presents the results for
the effect of the relative tick size with market cap as a control variable. Column 2β6 include
additional control variables from each hypothesis discussed in Section IV.A. Column 2
controls for variables that affects HFT; column 3 includes the control variable suggested by
the marketability hypothesis; column 4 contains control variables from the relative tick size
hypothesis; column 5 controls PIN for information asymmetry and column 6 controls for past
returns. Column 7 incorporates control variables from all hypotheses.
Insert Table V about here
Table V confirms that HFTvolume increases with the relative tick size, consistent with
our tick size constraints hypothesis. For example, column 7 show that a one-standard-
deviation increase in the relative tick size is associated with a 3.7% increase in the percentage
of volume with HFTs as liquidity providers (0.956*0.039), representing 13% of the increase
relative to the mean of HFTvolume ((0.956*0.039)/0.28). The table also shows that stocks
with larger market caps have higher percentages of liquidity provided by HFTs. The
insignificant coefficients before other variables do not necessarily imply that these variables
have no impact on the absolute magnitude of the liquidity provision of HFTs, but should be
interpreted as implying that these variables have no statistically significant impact on the
ratio of HFT liquidity provision to total volume.
Columns 5 and 7 show that a higher PIN, alone or combined with other control
variables, reduces HFTvolume. This intriguing result suggests that HFTs do not participate in
liquidity provision to a greater extent relative to non-HFTs for stocks with higher information
asymmetry. The main takeaway from Table V is that HFTs concentrate their activity on
stocks with large relative tick sizes and lower information asymmetry relative to non-HFTs.
C. Relative tick size and revenues
The fact that speed competition is more intense for stocks with large relative tick sizes
suggests that revenue from liquidity provision is higher for these stocks (hypothesis 4). We
22
examine revenue instead of profits because our calculation cannot include all costs of
liquidity provision such as investment in infrastructure or human capital.24
Our revenue measure comes from Brogaard, Hendershott, and Riordan (2014),
Menkveld (2013), and Baron, Brogaard, and Kirilenko (2014). The HFT liquidity provision
revenue for an individual stock for a certain time interval v is defined as
π = β(π»πΉπ ) + πΌππ Γ π (5)
The revenue comes from two components. The first term, β β(π»πΉπ ), captures total cash
flows throughout interval v, with n indicating each of the N transactions within each
interval.25 The second term, often referred as βpositioning profit,β cumulates value changes
associated with net position. In our analysis, the inventory cumulated in interval v,
πΌππ_π»πΉπ , is assumed to be cleared at the end of the interval midpoint quote π . The
revenue measure depends on the frequency of inventory clearance. Brogaard, Hendershott
and Riordan (2014) assume that inventory is cleared daily at the closing midpoint. There is
also empirical evidence showing that the inventory of HFTs crosses zero multiple times
within the day. For example, Brogaard, HagstrΓΆmer, NordΓ©n and Riordan (2014) finds that
the inventory of fast HFTs can cross zero 13.316 times per day. Kirilenko, Kyle, Samadi, and
Tuzun (2014) estimate that HFTs reduce half of their net holdings in 137 seconds. Therefore,
we calculate the revenue taking v at varying lengths: one-minute, five-minute, 30-minute and
one-day lengths, respectively.
Since we are interested in the market marking profit per dollar volume per day, we
calculate the unit profit as:
ππππ‘π ππ£πππ’π = β π /π·πππππ (6)
where v is the total number of intervals for each trading day and π·πππππ is the total dollar
volume with HFTs as liquidity providers for each stock on each day. For example, if the
24 Menkveld (2013) and Baron, Brogaard and Kirilenko (2014) estimate capital employed by HFTs in their sample 25 We also add a liquidity rebate when calculating the first part of the profit. NASDAQ has a complex fee structure and we use a fee of 0.295 cents per share, but the results are similar at other fee levels.
23
interval v is taken to be 30 minutes long, cash flows are calculated for each of the 30-minute
intervals and inventories accumulated are emptied at the end of the 30-minute interval; the
total number of intervals V equals 13.
We test hypothesis 4 using the following specification.
ππππ‘π ππ£πππ’π , , = π½ Γ π»πΉπππ’πππ¦ , , + π½ Γ π‘πππ ,
+π’ , + Ξ Γ π , +π , , (7) In the specification, ππππ‘π ππ£πππ’π , , is the unit revenue for each stock i on date t for
trader type n assuming inventory cleared at frequency v. We have two daily observations for
each stock at each frequency v: a unit revenue for HFTs and a unit revenue for non-HFTs.
These unit variables are affected by a number of controls (π , ) and two key variables of
interest: relative tick size and HFT dummy. π»πΉπππ’πππ¦ , , equals 1 for the unit revenue of
HFTs and 0 otherwise. π‘πππ , , is the relative tick size of stock i at day t. π’ , is the
industry-by-time fixed effect. π , are control variables presented in Table I.
Table VI shows that the unit revenue for HFTs is higher than that for non-HFTs if the
inventory is cleared at one minute, five minutes or 30 minutes. For example, if inventory is
clearer at every 30 minutes, the unit revenue of HFT is 0.42 basis points higher than that of
the non-HFT. The economic and statistical significance of the HFTdummy increases with the
frequency of inventory clearance. Indeed, one important feature of HFTs is their βvery short
time-frames for establishing and liquidating positionsβ SEC (2010).
The main result in Table VI is that the unit revenue for liquidity provision increases
with the relative tick size. For example, column 3 shows that a one-standard-deviation
increase in the relative tick size is associated with a 0.72 basis-point increase in the unit
revenue measure. Therefore, both HFTs and non-HFTs should have a stronger incentive to
provide liquidity for low-priced stocks, because their large relative tick sizes increase revenue
from limit orders and cost for market orders. However, the speed advantage enables HFTs to
establish time priority and thus they are able to obtain the higher rents created by large
relative tick sizes. Indeed, the previous sections show that HFT liquidity provision is more
active for stocks with larger relative tick sizes, and this section reveals that stocks with large
relative tick sizes yield greater revenue from liquidity provision. Combining these two
24
results, our paper reveals that speed advantage allows HFTs to take a larger market share for
stocks with higher rents from liquidity provision.
Insert Table VI about Here
Our results pertaining to revenue also shed light on policymaking. The usual
interpretation of the revenue of HFTs is that it is due to their information advantage. Such an
interpretation generates debate on the legitimacy of their information (Hu, Pan and Wang
(2013)). Our paper takes a different route on this debate. Our results suggest that rents can
also be extracted from the tick size regulation, and speed serves as the mechanism for
allocating the rents. The rents still exist even if we impose restrictions on HFT. Assume the
extreme scenario in which HFTs are banned: the rent still rewards whoever can post an order
at the front of the limit-order book, especially for stocks with large relative tick sizes. Also,
the recent policy proposal to increase the tick size to five cents would increase the rents,
which can drive another round of speed competition.
V. Identifications using ETFs
This section examines the causal impact of the relative tick size on HFT liquidity-
providing activity (hypothesis 1) using a difference-in-differences test. The same
identification strategy also facilitates the analysis of the causal relationship between the
relative tick size and liquidity (hypothesis 3). The question of liquidity is not only important
in its own right, but it also offers additional economic insights to help us understand the
relationship between the relative tick size and the proportion of liquidity provided by HFTs.
Section V.A illustrates how the proportion of the liquidity provided by HFTs and market
liquidity are measured. Section V.B presents the difference-in-differences test using the
splits/reverse splits of leveraged ETFs as exogenous shocks to the relative tick size.
A. Measure of HFT activity and liquidity
The NASDAQ HFT dataset does not contain HFT information for ETFs. Therefore,
we use NASDAQ-ITCH data to construct a proxy for the proportion of HFT liquidity
provision. The standard proxy for the absolute level of HFT activity is the βstrategic runβ
25
proposed by Hasbrouck and Saar (2013).26 Hasbrouck and Saar (2013) find that strategic run
has a very high correlation with the level of HFT activity. As our paper focuses on the HFTs
liquidity-providing activity relative to the total liquidity-providing activity, we normalize the
HFT activity by the total trading volume. We construct a variable, RelativeRun, which is
equal to the log value of strategic run divided by the log value of dollar volume in NASDAQ
market.27
Table VII demonstrates that RelativeRun is a good proxy for the proportion of
liquidity provided by HFTs. We calculate RelativeRun for the 117 stocks using ITCH data
and compare it with the proportion of liquidity provided by HFTs calculated using NASDAQ
HFT data. Table VII presents the cross-sectional correlation between RelativeRun and the
percentage of trading volume with HFTs as liquidity providers, HFTvolume (making). The
Pearson correlation is as high as 0.837 and the spearman correlation is as high as 0.835. The
correlations are much higher than two other widely used HFT proxies: the quote-to-trade
ratio (Angel, Harris and Spatt (2010 and 2013)) and negative dollar volume divided by total
number of messages (Hendershott, Jones, and Menkeveld (2011) and Boehmer, Fong, and
Wu (2013)), both of which are based on the intuition that HFTs tend to cancel more orders
than non-HFTs. Surprisingly, these two measures have negative cross-sectional correlations
with the proportion of liquidity provided by HFTs.28 One additional contribution of our paper
26 A strategic run is a series of submissions, cancellations, and executions that are likely to form an algorithmic strategy. The link between submissions, cancellations, and executions are constructed based on three criteria: (1) Limit orders with their subsequent cancellations or executions are linked by reference numbers provided by data distributors. (2) When an inference is needed to decide whether a cancellation is linked to either a subsequent submission of a nonmarketable limit order or a subsequent execution that occurs when the same order is re-sent to the market priced to be marketable, we follow Hasbrouck and Saar (2013), and infer such a link when a cancellation is followed within 100 milliseconds by a limit-order submission of the same size and same direction or by an execution of a limit order of the same size but in the opposite direction. (3) If a limit order is partially executed and the remainder is cancelled, we apply criterion (2) based on the canceled quantity. Strategic run is the sum of the time length of all strategic runs with 10 or more messages divided by the total trading time of that day (Hasbrouck and Saar (2013)). 27 We use log form so that the results are less sensitive to outliers (Wooldridge (2006)). 28 These two measures can be good proxies for distinguishing trader types if the comparison is made within the same security or to measure time-series variation in HFT activity. However, using these two proxies in cross-sectional comparison can be misleading because of the relative tick size. The statement by Manoj Narang, CEO of the HFT firm Tradeworx explains why HFT activity can be negatively correlated with cancellation rates: βThe wider the trading increment, the more size starts aggregating at each price level and the more ties you have in prices as there are fewer prices to be at. This elevates the importance of time priority becomes relative to price priority thereby making the role of speed more important. In addition, wider trading increments also reduce cancellation rates as the fair price of a security can now fluctuate in a wider band.β
26
is that we put forward the notion of carefully interpreting results that use quote-to-trade ratios
and negative dollar volume divided by total number of messages in cross-sectional
comparisons. Despite the common intuition that HFTs tend to cancel more orders than non-
HFTs, HFT activity can be negatively correlated with cancellation.
Table VII show that RelativeRun has a positive correlation with the proportion of
liquidity taken by HFTs (HFTvolume (taking)), calculated as percentage of trading volume
with HFTs are the liquidity takers. The result indicates that RelativeRun may also capture
some liquidity-taking activity on the part of HFTs. However, its high correlation with HFT
liquidity-making activity and its low correlation with HFT liquidity-taking activity imply that
it is a better proxy for liquidity-making activities than for liquidity-taking activities. 29
Therefore, we use RelativeRun as a proxy for liquidity-making, though we are aware that it
may capture liquidity-taking HFT activity to a small degree.
Insert Table VII about Here
Stock market liquidity is defined as the ability to trade a security quickly at a price
that is close to its consensus value (Foucault, Pagano, and RΓΆell (2013)). The spread is the
transaction cost faced by traders, and is often measured by the quoted bid-ask spread or the
trade-based effective spread. Depth reflects the marketβs ability to absorb large orders with
minimal price impact, and is often measured by the quoted depth. These liquidity measures
come from a message-by-message limit-order book we construct from ITCH data.30
The quoted spread (Qspread) is measured as the difference between the best bid and
ask at any given time. The proportional quoted spread (pQspread) is defined as the quoted
spread divided by the midpoint of the best bid and best ask prices. In addition to earning the
quoted spread, a liquidity provider also obtains a rebate from each executed share from the
29 RelativeRun can be a better proxy for liquidity making than for liquidity taking for the following reasons. Pure submissions of market orders are not considered strategic runs, because all βrunsβ start with limit orders. The 10 message cut-off and the time weight also increase the correlation of RelativeRun with patient liquidity-providing algorithms. Impatient liquidity-demanding algorithms may use limit orders, but these algorithms are more likely to switch to market orders once the initial limit orders fail to be executed. Therefore, strategic runs that arise from liquidity-demanding algorithms should contain fewer messages. Even if they contain more than 10 messages, it is natural to expect that they span a shorter period of time and carry lower time weight in RelativeRun. 30 We are aware that measures from a limit-order book are not sufficient for statics liquidity. See Jones and Lipson (2001) for a discussion.
27
NASDAQ. Therefore, we compute two other measures of the quoted spread: Qspreadadj and
pQspreadadj, which are spreads adjusted by the liquidity supplierβs rebate.31 Specifically,
ππ πππππ , = ππ πππππ , + 2 β ππππ’ππππ‘π¦ πππππ πππππ‘π (8) πππ πππππ , = (ππ πππππ , + 2 β ππππ’ππππ‘π¦ πππππ πππππ‘π)/ππππππππ‘ (9)
All four of these quoted spreads are weighted by the duration of a quote to obtain the daily
time-weighted average for each stock.
Our main measures of depth are the time-weighted average of displayed dollar depth
at the BBO. The effective spread measures the cost of trading against the actual supply of
liquidity (SEC (2012)). The effective spread (Espread) for a buy is defined as twice the
difference between the trade price and the midpoint of the best bid and ask price. The
effective spread for a sell is defined as twice the difference between the midpoints of the best
bid and ask and the trade price. The proportional effective spread (pEspread) is defined as the
effective spread divided by the midpoint. Also, a liquidity demander on the NASDAQ pays
the taker fee.32 Therefore, we compute the fee-adjusted effective spread and the fee-adjusted
proportional effective spread:
πΈπ πππππ = πΈπ πππππ + 2 β ππππ’ππππ‘π¦ π‘ππππ πππ (10) ππΈπ πππππ = (πΈπ πππππ + 2 β ππππ’ππππ‘π¦ π‘ππππ πππ)/ππππππππ‘ (11)
B. Difference-in-Differences test using leveraged ETF splits
This section uses a difference-in-differences test to establish the causal relationship
between the relative tick size, market liquidity, and the percentage of HFT liquidity provision.
ETF splits provide us with a cleaner environment than stock splits, allowing us to isolate the
effects of the tick size constraints. Stock splits may be motivated by information (Grinblatt,
Masulis, and Titman (1984), Brennan and Hughes (1991), Ikenberry, Rankine and Stice
(1996)), but the ETF splits are much less likely to be motivated by informational reasons and,
even if they are, their pairs that track the same index but do not split, provide an ideal control.
Among ETFs, splits/reverse splits are more frequent for leveraged ETFs. The bear-and-bull
31 For each stock on each day, the liquidity makerβs rebate is 0.295 cents per execution, but the results are qualitatively similar at other rebate levels. 32 We set the taker fee at 0.3 cents per share.
28
leveraged ETFs for the same index are usually issued by the same company at similar IPO
prices, but large cumulative movements of an index result in the divergence of nominal
prices. The issuers of leveraged ETFs usually conduct splits/reverse splits to align the
nominal prices of bull-and-bear ETFs.
The regression specification for the difference-in-differences test is:
π¦ , , = π’ , + πΎ , + π Γ π· , , + π Γ πππ‘π’ππ , , + π , , (12) where π¦ , , is HFT liquidity-providing activity or the liquidity measure for ETF j in index i at
time t. π’ , , the index-by-time fixed effects, controls for the time trend that may affect each
index. πΎ , are ETF fixed effects that absorb the time-invariant differences between two
leveraged ETFs that track the same index i. One such time-invariant difference is the
divergence of nominal prices before the splits/reverse splits. After controlling for index-by-
time and ETF fixed effects, the only major differences between the bull-and-bear ETFs that
track the same indexes are returns in each period, πππ‘π’ππ , , . The key variable in this
regression is π· , , , the treatment dummy, which equals 0 for the control group. For the
treatment group, the treatment dummy equals 0 before splits/reverse splits and 1 after
splits/reverse splits. Therefore, coefficient π captures the treatment effect. The leveraged bull
ETF is in the treatment group, and the leveraged bear ETF is in the control group if the
leveraged bull ETF splits, and vice versa.
To derive an unbiased estimate of coefficient π, the actual split/reverse split must be
uncorrelated with the error term in equation (12). This does not mean that the actual
split/reverse split must be exogenous. Equation (12) includes an index-by-time fixed effect,
an ETF fixed effect, and a contemptuous ETF return. The estimate of coefficient π remains
unbiased even if the actual split/reverse split is related to these particular explanatory
variables. Two stylized facts are then important for establishing the unbiasedness of π.
First, the motivation for ETF splits/reverse splits is transparent. The issuer of ETFs
tends to align the nominal price of the twin leveraged ETFs, so they decide to split an ETF
when it has a drastically higher nominal price than its pair, or reverse-split an ETF when it
has a drastically lower nominal price than its pair. Such a difference in price can be captured
by the ETF fixed effect, and the estimate of coefficient π remains unbiased. In fact, as we
29
control for both index-by-time fixed effects and ETF fixed effects, the estimation will be
biased only if the actual split/reverse split is somehow related to the contemporaneous
idiosyncratic shocks to the percentage of HFT liquidity-provision activity (Hendershott,
Jones and Menkveld (2011)).
It is then helpful to notice that the schedule for executing a split/reverse split is
predetermined and announced well before the actual split/reverse split. Our event window is
five days before and five days after the actual splits, and the event window sits after the
announcement of the splits. 33 Thus it seems highly unlikely that the split/reverse split
schedule could be correlated with idiosyncratic shocks to HFT liquidity provision in the
future (Hendershott, Jones and Menkveld (2011)). Also, a closer examination of these
announcements reveals that fund companies often conduct multiple splits/reverse splits on the
same day for ETFs tracking diversified underlying assets.34 Such a diversified sample further
mitigates the concern that the split decision is correlated with ETF specific idiosyncratic
shocks.
Panel A of Table VIII reports the splits results. Quoted spreads without rebate
adjustment (Columns 1) and with rebate adjustment (Columns 2) both decrease by 10.565
cents following splits. Column 3 shows that the proportional quoted spread increases by
2.129 basis points. Without the relative tick size frictions, we would expect the nominal
quoted spread to decrease by the same percentage as the nominal price, keeping the
proportional spread unchanged. However, with the increased relative tick size, the percentage
decrease in the nominal quoted spread is less than the percentage decrease in the nominal
price,35 which leads to an increase in the proportional quoted spread. Column 4 shows that
the proportional quoted spread with rebate adjustment experiences an even larger increase
(3.804 basis points). Indeed, the liquidity rebate is held constant during these events, which
implies a higher rebate per stock price after splits. In this sense, the economic impact of a
33 Using a longer time window creates overlap between the pre-announcement and pre-split periods, but we find similar results. 34 For example, the announcement made on April 9, 2010 involves splits ranging from oil, gas, gold, real estate, financial stocks, and basic materials to Chinese indices. 35 Technically, the nominal stock price may also be subject to tick size constraints, but the magnitude of the relative tick size has a greater impact on the nominal quoted spread than on nominal price, because the nominal price is much higher than the nominal quoted spread.
30
liquidity rebate resembles that of the tick size, the inclusion of which leads to a further
increase in the proportional quoted spread.
Column 5 shows that the average dollar depth at the BBO increases by 45,000 dollars.
A large relative tick size after splits implies that traders who were able to quote a range of
prices under a finer grid may have to quote the same price under a coarser grid, which
lengthens the queue at the best price. The literature generally agrees that securities with lower
quoted spreads and greater depth are more liquid. Because splits lead to higher quoted
spreads and greater depth, the key variable of interest turns out to be the effective spread, the
measure of the actual transaction costs incurred by liquidity demanders. The nominal
effective spread decreases by 6.826 cents following a split, but the proportional effective
spread, or the transaction cost for a fixed transaction dollar amount, increases by 2.073 basis
points and, after fee adjustment, 3.777 basis points. Therefore, we find that liquidity
decreases as the cost of actual transactions increases. This result is consistent with the
findings of Conroy, Harris and Benet (1990), Schultz (2000) and Kadapakkam,
Krishnamurthy, and Tse (2005) that splits harm liquidity, but the ETFs tracking the same
index provides an ideal control group for cleaner identification.
Our main results in panel A are presented in Column 10, which shows an increase in
the proportion of HFT liquidity-providing activity in the treatment group after splits relative
to the control group. Since the analysis has controlled for the index-by-time fixed effect, an
increase in the proportion of HFT liquidity-providing activity cannot be attributed to the
change in the underlying fundamentals of the leveraged ETF. Specifically, the increased HFT
activity cannot be ascribed to information events, because the treatment and control groups
have the same underlying information. The ETF fixed effects and the return variable further
control for other possible differences between ETFs in the same pair.36 We argue that the
increase in HFT liquidity-provision activity is driven by changes in the relative tick size.
Following the same intuition as discussed previously, splits lead to coarser price grids, which
can force traders who have quoted different prices before splits to quote the same price after
splits, thus causing an increase in the length of the queue at the new but more constrained 36 One such example is the clientele effect. For example, short sellers may prefer a bear ETF because of the short sale constraints for the bull ETF. The ETF fixed effects also control for other differences between the control and treatment groups before the splits/reverse splits, such as their divergence in nominal price.
31
best price. Once liquidity providers are in the same price queue, the time priority rule
determines the execution sequence, which intensifies speed competition. Taken as a whole,
splits encourage HFT liquidity provision, as a large relative tick size impedes price
competition while encouraging speed competition.
Insert Table VIII about Here
Panel B shows that reverse splits generate a pattern that is the opposite of that of splits.
Columns 1β4 show that the quoted spread increases but the proportional quoted spread
decreases after reverse splits. The patterns hold with or without rebate adjustment. We
ascribe this result to the reduction in the relative tick size following the reverse split. Given a
fixed nominal spread, after reverse splits, the proportional spread has new incremental units;
the newly added units of increment encourage price differentiation. The depth at the best
price decreases by 374,000 dollars, implying a shorter queue to provide liquidity. This result,
along with the reduction in the proportional quoted spread, constitutes evidence that traders
who were forced to quote the same price can now choose to differentiate themselves by price,
leading to a reduction in depth offered at the new BBO. Columns 6β9 show that the effective
spread increases but the proportional effective spread decreases, suggesting that the
transaction cost for liquidity demanders falls after reverse splits. Column 10 reports a
reduction in HFT activity following reverse splits. A reduction in the relative tick size
enhances price competition and reduces the queue for supplying liquidity, which weakens the
incentives to be at the top of the queue at the constrained price. Therefore, we see less HFT
liquidity-providing activity following reverse splits.
In summary, we find that splits decrease liquidity whereas reverse splits increase
liquidity. We also find that splits attract HFT liquidity-providing activity whereas reverse
splits reduces HFT liquidity-providing activity. These results should not be interpreted to
mean that HFTs have a negative impact on liquidity provision, because the tests are based on
exogenous shocks to the tick size but not HFT activity. The correct economic interpretation is
that a large relative tick size reduces liquidity but attracts HFT liquidity provision. Our results,
however, do reveal the importance of controlling for the relative tick size when considering
the causal relationship between HFT liquidity provision activity and liquidity.
32
VI. Robustness Tests
This section provides several robustness checks for the main results presented in
previous sections. Section VI.A checks the validity of alternative hypotheses to explain the
prevalence of HFTβs as liquidity providers on stocks with large relative tick sizes. Section
VI.B provides additional evidence pertaining to the comparison of price and time priority
based on a sample of data that provides all the quote updates from HFTs and non-HFTs from
February 22, 2010 through February 26, 2010. This small sample of data provides a dynamic
view of the market. We are able to provide additional evidence that HFTs enjoy higher
percentages of volume due to limit orders establishing time priority relative to non-HFTs,
especially when the relative tick size is large. Using the same dataset, Section VI.B also
addresses the concern that the presence of non-HFTs regarding the best quotes is due to HFTsβ
canceling or executing orders instead of non-HFTs actively improving the quotes.
A. Cross-sectional variation of HFT activity on liquidity taking
Concerns may arise if HFTs prefer low-priced stocks for reasons other than their large
relative tick sizes. For example, HFTs may prefer low-priced stocks because with them they
can trade the same number of shares with less capital, or HFTs may be the counterparts of
less sophisticated traders and less sophisticated traders may favor low-priced stocks. These
incentives, however, affect not only HFTsβ liquidity-providing behavior but also their
liquidity-taking behavior. Thus, if these motivations drive the cross-sectional variation of
HFT liquidity provision, we should observe that HFT liquidity taking and HFTβs activities
overall display the same cross-sectional pattern as that of HFT liquidity-providing activity.
This section checks the validity of these alternative hypotheses as well as of other hypotheses
that have similar effects on liquidity provision and liquidity taking.37
37 We are aware that the make-take fee structure is a possible mechanism for generating alternative liquidity-making and taking patterns. Because the maker rebate and take fee are also uniform for stocks priced above $1, the rebate and fees are higher in percentage for stocks with lower prices. However, as discussed in section II, the economic mechanism underlying the fee is the same as the relative tick size, which is the trade-off between price and time priority.
33
We follow the methodology in Table IV, and use the Nasdaq HFT trade file for 117
stocks in Oct 2010 to compute the volume with HFTs as liquidity takers relative to total
volume for each portfolio J on day t:
π»πΉπππππ‘πππ = β (π»πβ +π»π» )β (ππ»β +π»π» + π»π + ππ ) (13)
In addition, we compute the volume with HFTs engaged either as liquidity providers or
liquidity takers relative to total volume for each portfolio J on day t:
π»πΉπππππππ‘β = β (π»πβ + ππ» +π»π» )β (ππ»β +π»π» + π»π + ππ ) (14)
Table IX Panel A shows the average trading volume percentage due to HFT liquidity
takers for each of the market capβbyβrelative tick size portfolios. The result demonstrates
that the percentage of volume with HFTs as liquidity takers decreases with the relative tick
size. If HFTsβ preference for low-priced stocks is driven by low capital requirements or a less
sophisticated clientele, it is unlikely that we would observe that the HFT liquidity-making
and liquidity-taking activities display opposite cross-sectional variations with respect to the
relative tick size.
Panel B of table IX reveals an interesting pattern for overall HFT activity. Although
HFT liquidity-providing and liquidity-taking activities have opposite cross-sectional
variations with respect to the relative tick size, the overall HFT activity follows the same
cross-sectional variation as the HFT liquidity-making activity for large- and medium-sized
stocks, while following the same cross-sectional variation as the HFT liquidity-taking activity
for small-sized stocks. This result can be driven by varying dispersions of HFT liquidity-
making and liquidity-taking activities for stocks of varying market caps. For example, for
large-sized stocks, the HFT liquidity-making activity for stocks with large relative tick sizes
is 14.03% higher than that for stocks with small relative tick sizes (Table IV), but HFT
liquidity-taking activity for stocks with large relative tick sizes is only 4.17% lower than that
for stocks with small relative tick sizes (Panel A of Table IX). The difference in percentage
can contribute to the increasing share of overall HFT activity with respect to relative tick
sizes for large stocks. The same reasoning can be applied to medium- and small-sized stocks.
Insert Table IX about Here
34
In summary, the HFT activity overall does not exhibit a clear relationship with the
relative tick size since HFT liquidity-providing and liquidity-taking activities exhibit the
opposite cross-sectional variations with respect to the relative tick size. Proponents of a wider
tick size argue that a large tick size would control the growth of HFT (Weild, Kim, and
Newport (2012)). This argument is not consistent with the results displayed in Table IX.
Trading volumes of large- and medium-sized stocks tend to be larger than those of small-
sized stocks; thus if we take the market as a whole, a larger tick size is associated with higher
overall HFT activity.
B. Additional robustness check on price v.s. time priority
The tick size constraints channel argues that speed enables HFTs to achieve time
priority over non-HFTs when price priority is unable to differentiate two types of traders due
to the tick size regulation. The previous sections have established two results. First, a large
relative tick size is associated with higher frequency of incidences in which the best quotes
provided by HFTs and non-HFTs quotes are the same, while a small relative tick size is
associated with higher frequency of incidents in which the best quotes are uniquely provided
by non-HFTs relative to HFTs. Second, the portfolio in which HFTs serve most actively as
the liquidity providers for executed orders is also the portfolio in which HFTs and non-HFTs
provide the same best quotes the majority of the time when HFTs enjoy their comparative
advantage in establishing time priority over non-HFTs. These two results are established
using snapshots of limit-order books and executed trades, and thus are susceptible to two
criticisms. The first criticism corresponds to the speed competition between HFTs and non-
HFTs. Although the result that HFTs enjoy time priority over non-HFTs when the two types
of traders quote the same price is very much to be expected, and the result from the
percentage of volume with HFTs as the liquidity providers is consistent this claim, the nature
of the data we deploy in previous sections prevents us from providing direct evidence for this
claim. The second criticism corresponds to the price competition between HFTs and non-
HFTs. The data cannot discriminate whether the presence of non-HFTs in the best quotes is
due to non-HFTsβ actively undercutting HFTs or to HFTsβ withdrawing from best quotes.
35
Fortunately, NASDAQ provides us with five days of a sample that allows us to
dynamically examine the limit-order book. We use this sample to address the two concerns
mentioned above. Section VI.B.1 describes the data. Section VI.B.2 examines whether HFTs
are more likely to establish time priority than non-HFTs, and also tests how such a difference
in establishing time priority varies with the relative tick size. Section VI.B.3 examines
whether non-HFTs are more likely to actively establish price priority by undercutting HFTs.
B.1. Data
The NASDAQ top-of-book data contain the collective best HFTβdisplayed quotes,
the collective best non-HFTβdisplayed quotes, and their corresponding aggregate sizes. There
is a data record any time there was a change in either the HFT or non-HFT quote. The data is
available only for the week of Feb 22β26, 2010, so we use the data mainly for a robustness
check instead of showing the results in the main section. The sample for our analysis contains
the same 117 stocks as in the previous sections.
For each record, we are able to differentiate whether the change to the top-of-book is
made by an order added to the book or by an order that has disappeared from the book. An
order can disappear from the limit-order book by being either executed or canceled. Since the
NASDAQ also provides trading records with trade identifiers (HH, HN, NH, NN) for the 117
stocks during this sample period, we match the top-of-book data and the trade data by
millisecond timestamp, sign of the order (trade), price of the order (trade), size of the order
(trade), and type of liquidity providers (HFTs or non-HFTs). We classify the ones that we
identify as having disappeared from the limit-order book and matches that can be found in the
trade data as executions of displayed orders.
B.2. Time priority of HFT traders
Although the top-of-book limit-order data do not provide information on the time
precedence queue at each price, we are still able to infer whether an order is executed due to
time priority or price priority. If, at the time of execution, only a unique type of trader (HFT
or non-HFT) provides quotes at the execution price, we classify this trade as executed
because its liquidity provider has price priority. If, at the time of execution, both HFTs and
36
non-HFTs provide quotes at the execution price, we classify this trade as executed because its
liquidity provider has time priority. This can be supported by the reasoning that although the
other type of trader provides liquidity at the same price as the executed order, displayed
orders from the other type of trader fail to be executed because they have lower (time)
priority. By this classification, each trade from displayed orders is classified into one of the
four types: 1) the liquidity-providing HFT has price priority, 2) the liquidity-providing HFT
has time priority, 3) the liquidity-providing non-HFT has price priority, and 4) the liquidity-
providing non-HFT has time priority.
To test which type of trader is more likely to have time priority over the other type,
and how this relation changes with the relative tick size, we perform the following regression:
ππππ πππππππ‘π¦ , , = π½ Γ π»πΉπππ’πππ¦ , , + π½ Γ π‘πππ , + π½ Γ π‘πππ , Γ π»πΉπππ’πππ¦ , , +π’ , + Ξ Γ π , +π , , (15)
In the specification, ππππ πππππππ‘π¦ , , is the proportion of volume due to the liquidity
provider having time priority, calculated as the volume due to the liquidity provider having
time priority over the liquidity providerβs total volume, for stock i on date t of liquidity
provider type n.38 We have two daily observations for each stock: one for HFTs and one for
non-HFTs. The dependent variable is impacted by control variables (π , ) and three key
variables of interest: relative tick size, an HFT dummy, and their
interaction. π»πΉπππ’πππ¦ , , equals 1 for HFTs and 0 otherwise. π‘πππ , is the relative
tick size of stock i at day t minus the average of the relative tick size of the sample. We
demean the relative tick size to facilitate the interpretation of π½ , which captures the
difference between HFTs and non-HFTs for stocks with the average relative tick size. 39 π½
measures the relationship between the dependent variable and relative tick size for non-HFTs.
The interaction term π‘πππ , Γ π»πΉπππ’πππ¦ , , . captures the differences between HFTs
and non-HFTsβ percentage of volume due to having time priority for stocks with different
relative tick sizes. π’ , is the industry-by-time fixed effect. π , are control variables presented
in Table I.
38 The number is defined as missing if the trading volume from liquidity provider type n is 0 for stock i on date t. 39 Without demeaning the data, π½ is interpreted as the difference in profit between HFTs and non-HFTs for stocks with zero tick size, or infinite price.
37
Table X shows that the larger the relative tick size, the higher the percentage of
volume due to liquidity providers with time priority. The table also shows that HFTs enjoy a
higher percentage of volume due to achieving time priority than that enjoyed by non-HFTs.
For a stock with average tick size, HFT has about 8% more trades due to time priority than
that for non-HFTs. Note that the larger the relative tick size, the larger the difference between
HFTβs percentage of volume due to obtaining time priority and non-HFTsβ percentage of
volume due to obtaining time priority, as the positive significant coefficient of the interaction
term between the relative tick size and the HFT dummy variable indicates.
Insert Table X about Here
In summary, an increase in the relative tick increases the proportion of the dollar
volume due to time priority, and HFTs enjoy a higher proportion of the dollar volume due to
time priority than non-HFTs, especially when the relative tick size is large. All findings are
consistent with the tick size constraints hypothesis.
B.3. Active update of quotes
Using a minute-by-minute snapshot of the limit-order book, Table II shows that as the
relative tick size goes down, the best quotes are more likely to be provided by non-HFTs
relative to HFTs. Blume and Goldstein (1997) distinguish the active from the passive ways in
which an order can become the best bid (or ask). The first, the active way, is by actively
posting a bid (or ask) within the spread and thereby narrowing the spread. The second, the
passive way, is through lowering the collective best quotes or through the collective
withdrawal of other orders at the best quotes, thus leaving the remaining order at the very
best bid (or ask). Since orders actively improving the BBO or orders passively canceled or
executed can both affect the state of the limit-order book, one concern regarding the finding
in Table II is that the finding is due to orders which passively change the top-of-book limit
order. For example, the best quotes established by non-HFTs may simply be stale quotes, or
when non-HFTs fail to update their quotes when there is an information event. To address
this concern, we focus exclusively on orders which actively improve the BBO, and examine
how their activities vary with the relative tick size.
38
We use NASDAQ top-of-book displayed limit orders for the week of Feb. 22β26,
2010 to calculate the dollar size of each order which actively improves BBO for the 117
stocks on each day. For each update in the top-of-book, we check whether the best bid price
after the top-of-book update is higher than the previous best bid price, or whether the best ask
price after the top-of-book update is lower than the previous ask price. If so, we regard the
BBO as having been actively improved. Since the data allow us to identify the types of
traders who actively improve the BBO, we aggregate the dollar sizes for all orders that
actively improved BBO for HFTs and non-HFTs, respectively.40 The specification of the
regression is as follows:
π·ππππππππ§π , = π½ Γ π‘πππ , + Ξ Γ π , + π’ , +π , (16)
where π·ππππππππ§π , is the proportional difference between non-HFTsβ and HFTsβ
aggregate dollar size of orders that actively improve the BBO. This is calculated as
for stock i on date t, where π·πππππ§π denotes the aggregate dollar size
of orders that improve the BBO for liquidity provider type j. The variable of interest, π½,
measures how the percentage difference between HFTsβ and non-HFTsβ active dollar size
improvement varies with the relative tick size. π’ , is the industry-by-time fixed effect. π ,
are the control variables presented in Table I.
Table XI displays the regression output. The result shows that a decrease in the
relative tick size leads to an increase in the percentage of active dollar size improvement from
non-HFTs relative to HFTs. Therefore, our result that the proportion of liquidity provided by
non-HFTs relative to that of HFTs decreases with the relative tick size is robust even when
only active improvement of quotes is taken into account.
Insert Table XI about Here
40 When weighted by size, the case in which traders improve the best quotes by only 1 share carries less weight. The aggregated dollar size is not weighted by time. This is due to the fact that for those orders, the duration of their remaining in the limit-order book cannot be determined. The top-of-book data provide the information on the cumulative depth, but without the information on order ID.
39
VII. Conclusion
We contribute to the HFT literature by providing a tick sizeβbased explanation of
speed competition in liquidity provision. We find that a large relative tick size constrains
price competition and generates higher revenue for liquidity provision. The constrained price
competition also helps HFTs achieve time priority when they quote the same price as non-
HFTs. As a consequence, HFTs provide a larger fraction of liquidity for low-priced stocks in
which a one-cent uniform tick size implies a larger relative tick size. We argue that such a
cross-sectional pattern cannot be explained by either of two existing HFT channels: the price
competition channel or the information channel.
The tick size constraints channel provides new insights into the policy debate over
HFT. The current policy debate focuses on whether additional regulation is required and, if
so, how to pursue it. Our paper points in a new direction: HFT may simply be a consequence
of the existing tick size regulation. One possible policy solution would be tick size
deregulation instead of additional regulation. Specifically, our paper questions the uniform
tick size rationale, which generates a heterogeneous relative tick size.
This paper also questions the rationale to increase the tick size. The JOBS Act
encourages the SEC to examine the possibility of increasing the tick size, and a pilot program
is under way for less liquid stocks. Proponents of a larger tick size have offered three
rationales for this position (Weild, Kim and Newport (2012)). First, a larger tick size should
control the growth of HFT, and second, a larger tick size should increase liquidity. Our paper
shows that a larger tick size encourages HFT without improving liquidity. The third
argument for increasing tick size is that a larger tick size would lead to more IPOs. However,
the causal impact of the tick size on IPOs has not been proved by academic research using
clean identification. We believe that an increase in the tick size would create more rents for
traders who establish time priority, and would fuel another round of the arms race in speed.
Our paper has several limitations. First, we show that non-HFTs are more likely to
establish price priority than HFTs when the relative tick size is small. However, our data do
not provide the identity of the non-HFTs who undercut the price of HFTs. One such
possibility is that some non-HFTs in the sample are fundamental investors who use limit
orders to minimize their transaction costs, and they can afford more aggregative limit orders
40
because they are natural buyers or sellers of a security. This conjecture can be tested with
account-level data. Second, we consider only displayed orders in our analysis because hidden
orders have lower execution priority than displayed orders at the same price. It would be
interesting to examine the use of hidden orders by HFTs and non-HFTs.
Our paper could be extended in various ways. First, the extant literature on HFT finds
that technological shocks that enhance trading speed improve liquidity. It is commonly
believed that liquidity improvement arises from traders with higher speed. However, this
paper finds that non-HFTs provide larger proportion of best quotes relative to that of HFTs
when the relative tick size is small. An intriguing conjecture is that liquidity improvement
can arise from traders who choose not to upgrade in the face of technology shocks. Slow
traders lose time priority when competing with faster traders and thus they have stronger
incentives to undercut the price; on the other hand, faster traders enjoy the speed advantage
for establishing time priority and thus they have weaker incentives to improve quotes. It
would be interesting to test this conjecture using account-level data. Second, current
theoretical work on speed competition focuses on the role of information. Models using
discrete prices can be constructed to indicate the value of speed and the impact of tick size
constraints on market quality. Third, speed competition is not the only market design
response to tick size constraints (Buti, Consonni, Rindi, Wen and Werner (2014) and Chao,
Yao and Ye (2014)). The literature on market microstructure focuses on liquidity and price
discovery under certain market designs, but market design is also endogenous, and it should
prove fruitful for examining why certain market designs exist in the first place. Finally, the
SEC recently announced a pilot program for increasing the tick size for a number of small
stocks, believing that the tick size may need to be wider for less liquid stocks. We encourage
the SEC to consider decreasing the tick size for liquid stocks in the pilot program,
particularly for large stocks with lower prices.
41
References
Angel, James J., 1997, Tick Size, Share Prices, and Stock Splits, Journal of Finance 52, 655β
681.
Angel, James, Lawrence Harris, and Chester Spatt, 2010, Equity trading in the 21st century,
Working paper, Marshall School of Business.
Angel, James, Lawrence Harris, and Chester Spatt, 2013, Equity trading in the 21st century:
an update, Working paper, Marshall School of Business.
Anshuman, V. Ravi, and Avner Kalay, 1998, Market making with discrete prices. Review of
Financial Studies, 11 81-109.
Anshuman, V. Ravi, and Avner Kalay, 2002, Can splits create market liquidity? Theory and
evidence. Journal of Financial Markets 5 83-125.
Baker, Malcolm, Robin Greenwood, and Jeffrey Wurgler, 2009, Catering through Nominal
Share Prices, Journal of Finance 64, 2559-2590.
Baron, Matthew, Jonathan Brogaard, and Andrei Kirilenko, 2014, Risk and return in high
frequency trading, Working paper, Princeton University, University of Washington
and Massachusetts Institute of Technology.
Battalio, Robert H., Shane A. Corwin, and Robert H. Jennings, 2014, Can brokers have it all?
on the relation between make take fees & limit order execution quality, Working
paper, University of Notre Dame and Indiana University Bloomington.
Benartzi, Shlomo, Roni Michaely, Richard H.Thaler, and William C. Weld, 2009, The
nominal share price puzzle, Journal of Economic Perspectives 23, 121β142.
Berk, Jonathan, and Peter DeMarzo, 2013, Corporate Finance, 3rd editon (Prentice Hall Press,
NJ).
Biais, Bruno, and Thierry Foucault, 2014, HFT and market quality, Bankers, Markets &
Investors 128, 5-19.
Biais, Bruno, Thierry Foucault, and Sophie Moinas, 2014, Equilibrium High-Frequency
Trading, Journal of Financial Economics, Forthcoming
Blume, Marshall, and Michael Goldstein, 1997, Quotes, order flow, and price discovery,
Journal of Finance 52, 221-244.
42
Boehmer, Ekkehart, Kingsley Y. L., Fong, and Juan Wu, 2012, International Evidence on
Algorithmic Trading, Working paper, University of Georgia and University of New
South Wales.
Bolton, Patrick, and Ernst-Ludwig von Thadden, 1998, Blocks, liquidity, and corporate
control, Journal of Finance 53, 1-25.
Bond, Philip, Alex Edmans, and Itay Goldstein, 2011, the real effects of financial markets,
Working paper, NBER.
Brennan, Michael, and Thomas Copeland, 1988, Stock splits, stock prices, and transaction
costs, Journal of Financial Economics 22, 83-101.
Brennan, Michael, and Patricia Hughes, 1991, Stock prices and the supply of information,
Journal of Finance 46, 1665-1691.
Brogaard, Jonathan, BjΓΆrn HagstrΓΆmer, Lars L. NordΓ©n, and Ryan Riordan, 2014, Trading
fast and slow: colocation and market quality, Working paper.
Brogaard, Jonathan, Terrence Hendershott, and Ryan Riordan, 2014, High frequency trading
and price discovery, Review of Financial Studies, Forthcoming.
Budish, Eric, Peter Cramton, and John Shim, 2013, The High-Frequency Trading Arms Race:
Frequent Batch Auctions as a Market Design Response, Working paper, University of
Chicago and University of Maryland.
Buti, Sabrina, Francesco Consonni, Barbara Rindi, Yuanji Wen, and Ingrid M. Werner, 2014,
Tick size: theory and evidence, Working paper, University of Toronto.
Byun, Jinho, and Michael S. Rozeff, 2003, Long-run Performance after Stock Splits: 1927 to
1996, Journal of Finance 58, 1540β6261.
Campbell, John Y., Jens Hilscher, and Jan Szilagyi, 2008, In search of distress risk, Journal
of Finance 63, 2899-2939.
Carrion, Allen, 2013, Very fast money: High-frequency trading on the NASDAQ, Journal of
Financial Markets 16, 680-711.
Chao, Yong, Chen Yao, and Mao Ye, 2014, Two-Sided Platforms, Make-Take Fees and
Market Fragmentation, Working paper, University of Louisville, University of
Warwick and University of Illinois at Urbana-Champaign.
43
Chordia, Tarun, Amit Goyal, Bruce N. Lehmann, and Gideon Saar, 2013, High-Frequency
Trading, 2013, Journal of Financial Markets, 16, 637-645.
Conroy, Robert, Robert Harris, and Bruce Benet, 1990, The Effects of Stock Splits on Bid-
Ask Spreads, Journal of Finance 45, 1285-1295.
Cordella, Tito, and Thierry Foucault, 1999, Minimum Price Variations, Time Priority, and
Quote Dynamics, Journal of Financial Intermediation 8,141-173.
Dyl, Edward A., and William B., Elliott, 2006, The Share Price Puzzle, Journal of Business
79, 2045-2066.
Easley, David, and Maureen, O'Hara, 2010, Microstructure and Ambiguity, Journal of
Finance 65 , 817β1846.
Easley, David, Nicholas M., Kiefer, Maureen OβHara, and Joseph B., Paperman, 1996οΌ
Liquidity, Information, and Infrequently Traded Stocks, Journal of Finance 51, 1405-
1436.
Ellul, Andrew, and Marco Pagano, 2006, IPO Underpricing and After-Market Liquidity,
Review of Financial Studies 19, 381-421.
Fang, Vivian W., Thomas H. Noe, and Sheri Tice, 2009, Stock market liquidity and firm
value, Journal of Financial Economics 94, 150-169.
Foucault, Thierry, Ohad Kadan, and Eugene Kandel. 2013, Liquidity cycles and make/take
fees in electronic markets." The Journal of Finance 68: 299-341.
Foucault, Thierry, Ohad Kadan, and Eugene Kandel, 2005, Limit Order Book as a Market for
Liquidity, Review of Financial Studies 18, 1171-1217.
Foucault, Thierry, Marco Pagano, and Alisa RΓΆell, 2013, Market Liquidity: Theory, Evidence
and Policy (Oxford University Press, Oxford).
Frazzini, Andrea, Ronen Israel, and Tobias Moskowitz, 2012, Trading Costs of Asset Pricing
Anomalies, Working paper.
Gai, Jiading., Choi, Dong Ju, Choi, David O'Neal, Mao Ye, and Robert, Sinkovits, 2013,
Fast Construction of Nanosecond Level Snapshots of Financial Markets, Concurrency
and Computation: Practice and Experience 26, 2149-2156.
Gai, Jiading, Chen Yao, and Mao Ye, 2013, The Externalities of High Frequency Trading,
Working paper.
44
Grinblatt, Mark, Ronald W., Masulis, and Sheridan, Titman, 1984, The Valuation Effects of
Stock Splits and Stock Dividends. Journal of Financial Economics 13, 461-490.
HagstrΓΆmer, BjΓΆrn, and Lars NordΓ©n, 2013, The diversity of high-frequency traders, Journal
of Financial Markets 16, 741-770.
Harris, Lawrence, 1994, Minimum Price Variations, Discrete Bid/Ask Spreads and
Quotation Sizes, Review of Financial Studies 7, 149-178.
Harris, Lawrence, 2003, Trading and Exchanges: Market Microstructure for Practitioners
(Oxford University Press, Oxford).
Hasbrouck, Joel, and Gideon Saar, 2013, Low-latency Trading, Journal of Financial Markets.
16, 646-679
Hendershott, Terrence, Charles M. Jones, and Albert J. Menkveld, 2011, Does Algorithmic
Trading Improve Liquidity? Journal of Finance 66, 1β33.
Hirschey, Nicholas H., 2012, Do High-Frequency Traders Anticipate Buying and Selling
Pressure? Working paper, London Business School.
Hu, Grace Xing. , Pan, Jun, and Wang, Jiang, Early Peek Advantage?, 2013, Working Paper,
University of Hong Kong and MIT.
Ikenberry, David, Graeme Rankine, and Earl Stice, 1996, What Do Stock Splits Really Signal?
Journal of Financial and Quantitative Analysis 31, 357-375.
Jones, Charles., 2013, What do we know about high-frequency trading? Working paper,
Columbia University, New York, NY.
Jones, Charles, and Marc Lipson, 2001, Sixteenths: direct evidence on institutional execution
costs, Journal of Financial Economics 13, 253-378.
Kadapakkam Palani-Rajan, Srinivasan Krishnamurthy, and Yiuman Tse , 2005, Stock Splits,
Broker Promotion and Decimalization. Journal of Financial and Quantitative
Analysis.
Kalay, Alon., and Mathias Kronlund., 2014, The market reaction to stock split
announcements: earnings information after all, Working paper, Columbia University.
Kirilenko, Andrei, Albert S. Kyle, Mehrdad Samadi, and Tugkan Tuzun. 2011, The flash
crash: The impact of high frequency trading on an electronic market. Working Paper,
University of Maryland.
45
Lakonishok, Josef, and Baruch. Lev, 1987, Stock Splits and Stock Dividends: Why, Who and
When, Journal of Finance 42, 913β932.
Maug, Ernst, 1998, Large Shareholders as Monitors: Is There a Trade-Off between Liquidity
and Control? Journal of Finance 53, 65β98.
Menkveld, Albert, 2013, High frequency trading and the new market makers, Journal of
Financial Markets 16, 712-740.
OβHara, Maureen, 2003, Presidential address: Liquidity and price discovery, Journal of
Finance 58, 1335β1354.
OβHara, Maureen, Gideon Saar, and Zhuo Zhong, 2013, Relative Tick Size and the Trading
Environment, Working paper, Cornell University.
Roberts, Michael, and Toni Whited, 2012, Endogeneity in empirical corporate finance,
Working paper, University of Pennsylvania and University of Rochester.
SandΓ₯s, Patrik, 2001 Adverse selection and competitive market making: Empirical evidence
from a limit order market. Review of Financial Studies 14, 705-734.
Schultz, Paul, 2000, Stock Splits, Tick Size, and Sponsorship, Journal of Finance 55, 429-
450.
U.S. Securities and Exchange Commission (SEC), 2010, Concept release on equity market
structure.
U.S. Securities and Exchange Commission (SEC), 2012, Report to Congress on Decimalization.
Weild, David, Edward Kim, and Lisa Newport, 2012, The trouble with small tick sizes,
Working paper.
Wooldridge, Jeffrey M., 2006, Introductory econometrics 3rd edition, Thomson South-
Western.
Yao, Chen and Mao Ye, 2014, Tick Size Constraints, Market Structure, and Liquidity,
Working paper, University of Illinois.
46
Table I. Summary Statistics This table presents the summary statistics for the various samples used in this paper. Panel A presents the summary statistics for stocks contained in the NASDAQ HFT sample in October 2010. Panel B provides the summary statistics for the leveraged ETF sample used in the difference-in-differences test, in which the event window is 5 days before splits/reverse splits from 2010 through 2013. Panel C lists the summary statistics for stocks in the NASDAQ HFT sample from February 22, 2010 through February 26, 2010 used in the robustness test section. All the variables are measured for each stock per day unless otherwise indicated. HFTvolume is the trading volume with HFTs as liquidity providers over the total trading volume. tickrelative is the reciprocal of price. logmcap is the log value of market capitalization. turnover is the annualized share turnover. volatility is the standard deviation of open-to-close returns based on the daily price range, that is, high minus low, proposed by Parkinson (1980). logbvaverage is the logarithm of the average book value of equity per shareholder at the end of the previous year (December 2009). idiorisk is the variance on the residual from a 60-month beta regression using the CRSP Value Weighted Index. age (in 1k days) is the length of time for which price information is available for a firm on the CRSP monthly file. numAnalyst is the number of analysts providing one-year earnings forecasts calculated from I/B/E/S. pastreturn is the monthly return for each stock in September 2010. PIN is the probability of informed trading. UnitRevenue is the revenue for liquidity provision per dollar volume. return is the contemporaneous daily return. Qtspd represents the time-weighted quoted spread. pQtspd represents the time-weighted proportional quoted spread. Qtspdadj and pQtspdadj represent the time-weighted quoted spread and the time-weighted proportional quoted spread after adjusting for fees. SEffspd represents the size-weighted effective spread and pSEffspd represents the size-weighted proportional effective spread. SEffspdadj and pSEffspdadj represent the size-weighted effective spread and the size-weighted proportional effective spread after adjusting for fees. Depth represents the sum of the time-weighted dollar depth at the best bid and ask. RelativeRun is the proxy for percentage of liquidity provided by HFTs.
Mean Median Std. Min. Max. Panel A. NASDAQ HFT Sample
HFTvolume 0.28 0.267 0.135 0 0.728 tickrelative 0.05 0.036 0.039 0.002 0.192 logmcap 21.891 21.402 1.885 19.371 26.399 turnover 2.293 1.744 2.136 0.074 33.301 volatility 0.015 0.013 0.009 0.002 0.099 logbvaverage 13.148 13.196 2.112 8.822 17.881
idiorisk 0.013 0.008 0.016 0.001 0.139 age (in 1k days) 9.617 7.565 7.737 0.945 30.955 numAnalyst 13.545 11 9.92 1 48 pastreturn 0.105 0.096 0.071 -0.077 0.336 PIN 0.12 0.115 0.051 0.021 0.275 UnitRevenue (in bps)
1-min Interval 0.57 0.2 3.33 -32.87 88.89
47
5-min Interval 0.38 0.2 4.17 -29.81 145.62 30-min Interval 0.47 0.36 6.43 -69.21 145.62 Daily Closing 1.4 1.28 18.54 -323.38 215.43
Panel B. Leveraged ETF Sample
return 0 0 0.04 -0.25 0.25 Qtspd (in cent) 4.93 2.55 5.48 1.00 33.88 pQtspd (in bps) 10.56 9.76 5.84 1.77 33.17 Qtspdadj (in cent) 5.52 3.14 5.48 1.59 34.47 pQtspdadj (in bps) 13.11 11.87 7.19 2.44 44.67
SEffspd (in cent) 3.47 1.79 3.98 0.67 32.90 pSEffspd (in bps) 7.95 6.89 4.89 1.50 34.81 SEffspdadj (in cent) 4.07 2.39 3.98 1.27 33.50 pSEffspdadj (in bps) 10.54 9.05 6.67 2.16 43.28 Depth (in mn dollars) 0.35 0.11 0.87 0.01 6.99 RelativeRun 0.22 0.22 0.05 0.03 0.38
48
Table II. Who Provides the Best Quotes? This table displays the percentages of time that the best bid and ask quotes to the NASDAQ limit-order book are provided by HFTs and non-HFTs, respectively. The sample includes 117 stocks in the NASDAQ HFT data from October 2010. Stocks are sorted first into 3-by-3 portfolios by average market cap and then by average price from September 2010. For each portfolio and each trading day, we calculate the percentage of time that the best quotes are provided uniquely by HFTs, the percentage of time that the best quotes are provided by uniquely non-HFTs, and the percentage of time that the best quotes are provided by both HFTs and non-HFTs. Column (1) presents the average percentage of time that the best quotes are provided uniquely by HFTs, column (2) presents the average percentage of time that the best quotes are provided uniquely by non-HFTs, and column (3) presents the percentage of time that the best quotes are provided by both HFTs and non-HFTs. Column (4) shows the ratio of column (2) figures to column (1) figures. Column (5) shows the difference between column (1) figures and (2) figures. t-statistics for column (5) based on 21 daily observations are presented in column (6). *, **, and *** represent statistical significance at the 10%, 5%, and 1% levels, respectively.
(1) (2) (3) (4) (5) (6)
Relative Tick Size HFT Only
Non-HFT Only
HFT & Non-HFT Ratio Non-HFT
minus HFT t-stat
Large Cap
Large (Low Price) 1.6% 2.5% 95.9% 1.55 0.9%*** 7.27
Medium (Medium Price) 11.9% 18.6% 69.6% 1.57 6.7%*** 14.01
Small (High Price) 16.8% 37.7% 45.5% 2.25 20.9%*** 37.81
Middle Cap
Large (Low Price) 18.0% 15.2% 66.8% 0.84 -2.9%*** -3.52
Medium (Medium Price) 20.0% 56.6% 23.4% 2.83 36.6%*** 36.03
Small (High Price) 20.7% 63.7% 15.7% 3.08 43.0%*** 67.15
Small Cap
Large (Low Price) 11.3% 54.7% 34.1% 4.86 43.4%*** 27.55
Medium (Medium Price) 20.2% 55.8% 24.0% 2.77 35.7%*** 30.11
Small (High Price) 18.6% 70.7% 10.7% 3.80 52.1%*** 66.79
Total 15.4% 41.7% 42.9% 2.62 26.3%*** 18.31
49
Table III. Market Share of BBO Depth Provided by HFTs This table presents the percentages of depth at the NASDAQ best bid and offer (BBO) provided by HFTs. The sample includes 117 stocks in NASDAQ HFT data from October 2010. The stocks are sorted first by average market cap and then by average price in September 2010. To calculate the share-weighted average for each portfolio on each day, we aggregate the number of shares provided by HFTs at the BBO and then divide it by the total number of shares at the BBO for that portfolio. t-statistics are calculated based on 21 daily observations. *, ** and *** represent statistical significance of large-minus-small differences at the 10%, 5%, and 1% levels, respectively.
Large Medium Small Large-Small
Relative Tick Size Relative Tick Size Relative Tick Size Relative Tick Size t-stat
(Low Price) (Medium Price) (High Price) (Low-High Price)
Large Cap 55.66% 45.44% 35.07% 20.59%*** 22.10
Middle Cap 39.73% 29.24% 24.61% 15.13%*** 22.88
Small Cap 25.78% 23.02% 20.78% 5.00%*** 3.18
L-S Cap 29.88%*** 22.43%*** 14.29%***
t-statistics 18.84 17.92 16.80
50
Table IV. Percentage of Volume with HFTs as the Liquidity Providers This table presents the trading volume percentage due to HFTs as liquidity providers. The sample includes 117 stocks in the NASDAQ HFT data from October 2010. The stocks are sorted first by average market cap and then by average price in September 2010 into 3-by-3 portfolios. To calculate the volume-weighted average for each portfolio on each day, we aggregate the volumes due to HFT liquidity providers and then divide that figure by the total volume for that portfolio. t-statistics are calculated based on 21 daily observations. *, **, and *** represent statistical significance at the 10%, 5%, and 1% levels of large-minus-small differences, respectively.
Large Medium Small Large-Small
Relative Tick Size Relative Tick Size Relative Tick Size Relative Tick Size t-stat
(Low Price) (Medium Price) (High Price) (Low-High Price)
Large Cap 49.96% 38.23% 35.93% 14.03%*** 15.54
Middle Cap 39.30% 24.03% 24.33% 14.97%*** 18.74
Small Cap 24.11% 18.88% 18.49% 5.62%*** 5.38
L-S Cap 25.84%*** 19.35%*** 17.43%***
t-statistics 21.33 21.76 18.22
51
Table V. HFT Liquidity Provision This table presents the results of the regression of percentage of trading volume with HFTs as liquidity providers on the relative tick size. The regression uses the NASDAQ HFT data sample as of October 2010. The regression specification is:
π»πΉππ£πππ’ππ , = π’ , + π½ Γ π‘πππ , + Ξ Γ π , +π , where π»πΉππ£πππ’ππ , is the trading volume with HFTs as the liquidity provider over total trading volume for stock i on date t. π‘πππ , is the inverse of the stock price for stock i on day t. π’ , represents industry-by-time fixed effects. The definitions of the control variables π , are presented in Table 1. We report estimated coefficients and heteroskedasticity robust standard errors, clustered by firm. Standard errors are in parentheses. ***, **, and * indicate the statistical significance at the 1%, 5%, and 10% levels. Dep. Variable HFTvolume
(1) (2) (3) (4) (5) (6) (7)
tickrelative 1.034*** 1.035*** 0.943*** 1.074** 0.982*** 1.033*** 0.956***
(0.38) (0.37) (0.34) (0.42) (0.37) (0.38) (0.36) logmcap 0.054*** 0.051*** 0.051*** 0.047*** 0.046*** 0.054*** 0.031***
(0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) turnover 0.008** 0.007**
(0.00) (0.00) volatility -0.719 -0.536
(0.64) (0.59) logbvaverage 0.003 0.005
(0.00) (0.00) idiorisk -0.359 -0.547
(0.63) (0.46) age 0.002 0.003*
(0.00) (0.00) numAnalyst 0.001 0.001
(0.00) (0.00) PIN -0.469** -0.407*
(0.22) (0.21)
pastreturn -0.013 -0.034 (0.11) (0.09)
R2 0.613 0.622 0.634 0.618 0.626 0.613 0.664 N 2352 2352 2268 2352 2352 2352 2268 Industry*time FE Y Y Y Y Y Y Y
52
Table VI. Tick size and Liquidity Provision Revenue This table presents the regression of the unit revenue of liquidity provision on the relative tick size. The regression uses the NASDAQ HFT sample as of October 2010.The regression specification is:
ππππ‘π ππ£πππ’π , , = π½ Γ π‘πππ , + π½ Γ π»πΉπππ’πππ¦ , , + π’ , + Ξ Γ π , + π , , Columns (1)β(4) present regression results for daily unit profit measured on the assumption that inventories are emptied at one-minute, five-minute, 30-minute intervals and at daily closing. π‘πππ , is the inverse of the stock price for stock i on day t. π»πΉπππ’πππ¦ , , is equal to 1 if the revenue measure is from HFTs and 0 otherwise. π’ , represents the industry-by-time fixed effects. The definitions of the control variables π , are presented in Table 1. We report estimated coefficients and heteroskedasticity robust standard errors, clustered by firm. Standard errors are in parentheses. ***, **, and * indicate the statistical significance at the 1%, 5%, and 10% levels. Dep. Variable Unit Revenue (in bps) (1) (2) (3) (4) 1-min Interval 5-min Interval 30-min Interval Daily Closing
tickrelative 20.125*** 19.024*** 18.677*** 21.561
(4.68) (4.21) (4.25) (17.55)
HFTdummy 0.828*** 0.761*** 0.420** -0.093
(0.13) (0.14) (0.17) (0.52)
logmcap -0.268 -0.160 -0.129 -0.233
(0.17) (0.16) (0.16) (0.51)
turnover -0.113** -0.085 0.038 -0.178
(0.05) (0.07) (0.09) (0.20)
volatility -0.807 -19.509 -91.450*** -179.865**
(14.92) (21.25) (22.08) (73.51)
logbvaverage 0.029 0.036 0.017 0.314
(0.07) (0.08) (0.07) (0.19)
idiorisk -3.551 2.883 11.624** 26.079
(7.79) (6.03) (5.02) (20.86)
age 0.005 -0.008 -0.028* -0.033
(0.02) (0.02) (0.02) (0.05)
numAnalyst 0.042* 0.040** 0.033* 0.063
(0.02) (0.02) (0.02) (0.06)
PIN 0.604 0.797 2.895 9.835
(3.05) (3.15) (3.47) (8.87)
pastreturn -2.409 -1.658 1.373 9.321*
(2.06) (1.72) (1.95) (5.59)
R2 0.260 0.219 0.214 0.194
N 4484 4484 4484 4484
Industry*time FE Y Y Y Y
53
Table VII. Correlation Test for the Proxy of the Percent of HFT Activity This table presents the cross-sectional correlations between the proxies for the percentage of liquidity provided by HFTs and the HFT activity calculated from the NASDAQ HFT dataset. The HFT activity measures include the percentage of volume with HFTs as liquidity providers, HFTvolume (making), and the percentage of volume with HFTs as liquidity takers, HFTvolume (taking). The proxy for the percentage of liquidity provided by HFTs include RelativeRun, Quote-to-Trade Ratio and the Dollar Volume (in $100)βtoβMessage Ratio multiplied by -1. P-values are shown under correlation coefficients; *, **, and *** denote statistical significance at the 10%, 5% and 1% levels, respectively.
HFT Measures HFTvolume (making) HFTvolume (taking)
Panel A. Pearson Correlation
RelativeRun 0.837 0.377
<.0001 <.0001
Quote / Trade Ratio -0.390 -0.082
<.0001 0.378
-Trading Vol. (in $100) / Message Ratio -0.252 -0.274
0.006 0.003
Panel B. Spearman Correlation
RelativeRun 0.835 0.361
<.0001 <.0001
Quote / Trade Ratio -0.589 -0.154
<.0001 0.0969
-Trading Vol. (in $100) / Message Ratio -0.554 -0.506
<.0001 <.0001
54
Table VIII. Difference-in-differences Test Using Leveraged ETF Splits (Reverse Splits) This table presents the results of difference-in-differences tests using leveraged ETF split (and reverse split) events from 2010 to 2013. The event windows are 5 days before and 5 days after the events. The regression specification is:
π¦ , , = π’ , + πΎ , + π Γ π· , , + π Γ πππ‘π’ππ , , + π , , π’ , is the index by time fixed effects and πΎ , is the ETF fixed effects for ETF j of index i. The treatment dummy π· , equals 0 for the control group. For the treatment group, π· equals 0 before splits/reverse splits and 1 after the splits/reverse splits. Returni,t,j is the return for ETF j of index i on day t. Qtspd represents the time-weighted quoted spread. pQtspd represents the time-weighted proportional quoted spread. Qtspdadj and pQtspdadj represent the time-weighted quoted spread and the time-weighted proportional quoted spread after adjusting for fees. SEffspd represents the size-weighted effective spread and pSEffspd represents the size-weighted proportional effective spread. SEffspdadj and pSEffspdadj represent the size-weighted effective spread and the size-weighted proportional effective spread after adjusting for fees. Depth represents the time-weighted dollar depth at the BBO in millions of dollars. RelativeRun is the proxy of the percentage of liquidity provided by HFTs. We report estimated coefficients and heteroskedasticity robust standard errors, clustered by ETFs. Standard errors are in parentheses. ***, **, and * indicate the statistical significance at the 1%, 5%, and 10% levels. Panel A. Split Sample
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Qtspd Qtspdadj pQtspd pQtspdadj Depth SEffspd SEffspdadj pSEffspd pSEffspdadj RelativeRun (in cent) (in cent) (in bps) (in bps) (in mn) (in cent) (in cent) (in bps) (in bps)
π· -10.565*** -10.565*** 2.129* 3.804*** 0.045 -6.826*** -6.826*** 2.073** 3.777** 0.008*
(2.17) (2.17) (1.07) (1.33) (0.05) (1.59) (1.59) (0.91) (1.42) (0.00)
return -8.451*** -8.451*** -9.417*** -10.060*** 0.037 -4.671*** -4.671*** -5.103** -5.757* -0.016
(1.07) (1.07) (1.44) (2.29) (0.03) (0.67) (0.67) (2.07) (3.12) (0.01) R2 0.950 0.950 0.957 0.952 0.956 0.932 0.932 0.929 0.926 0.964 N 220 220 220 220 220 220 220 220 220 220 Index*time FE Y Y Y Y Y Y Y Y Y Y ETF FE Y Y Y Y Y Y Y Y Y Y
55
Panel B. Reverse Split Sample
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Qtspd Qtspdadj pQtspd pQtspdadj Depth SEffspd SEffspdadj pSEffspd pSEffspdadj RelativeRun
(in cent) (in cent) (in bps) (in bps) (in mn) (in cent) (in cent) (in bps) (in bps)
π· 1.415** 1.415** -3.822*** -6.917*** -0.374** 0.837* 0.837* -4.088*** -7.235*** -0.016**
(0.66) (0.66) (0.79) (1.38) (0.17) (0.46) (0.46) (0.87) (1.48) (0.01)
return 0.076 0.076 -7.758*** -10.672*** 0.877 2.498 2.498 -0.768 -3.731 -0.038
(3.05) (3.05) (2.25) (3.43) (0.73) (1.71) (1.71) (2.82) (4.37) (0.03)
R2 0.908 0.908 0.932 0.896 0.780 0.867 0.867 0.845 0.842 0.897 N 660 660 660 660 660 660 660 660 660 660 Index*time FE Y Y Y Y Y Y Y Y Y Y ETF FE Y Y Y Y Y Y Y Y Y Y
56
Table IX. Percentage of Volume with HFTs as the Liquidity Takers and Percentage of Volume with HFTs as the Liquidity Takers or Liquidity Makers This table presents the trading volume percentage with HFTs as liquidity takers and trading volume percentage with HFTs as liquidity providers or liquidity takers. The sample includes 117 stocks in NASDAQ HFT data from October 2010. The stocks are sorted first by average market cap and then by average price from September 2010 into 3-by-3 portfolios. Panel A displays the volume-weighted percentage of trading volume with HFTs as liquidity takers. To calculate the volume-weighted average for each portfolio on each day, we aggregate the volumes with HFTs as liquidity takers and then divide that figure by the total volume for that portfolio. Panel B displays the volume-weighted percentage of trading volume with HFTs engaged as either liquidity providers or takers. To calculate the volume-weighted average for each portfolio on each day, we aggregate the volumes due to HFT liquidity providers or HFT liquidity takers and then divide that figure by the total volume for that portfolio.t-statistics are calculated based on 21 daily observations. *, **, and *** represent statistical significance at the 10%, 5%, and 1% levels of large-minus-small differences, respectively.
Large Medium Small Large-Small
Relative Tick Size Relative Tick Size Relative Tick Size Relative Tick Size t-stat
(Low Price) (Medium Price) (High Price) (Low-High Price)
Panel A. Percentage of Trading Volume with High-frequency Traders as Liquidity Takers
Large Cap 44.11% 47.57% 48.28% -4.17%*** -5.31
Middle Cap 39.19% 42.40% 46.97% -7.78%*** -6.25
Small Cap 27.65% 33.36% 39.64% -11.99%*** -10.69
L-S Cap 16.46%*** 14.21%*** 8.64%***
t-statistics 13.76 13.04 8.47
Panel B. Percentage of Trading Volume with High-frequency Traders as Liquidity Makers or Takers
Large Cap 73.72% 69.03% 68.14% 5.58%*** 9.30
Middle Cap 64.46% 57.39% 61.45% 3.01%** 2.55
Small Cap 45.61% 46.51% 51.41% -5.80%*** -3.92
L-S Cap 28.11%*** 22.53%*** 16.73%***
t-statistics 20.03 18.75 15.13
57
Table X. Dollar Volume due to Time Priority with Respect to Tick Size This table presents the regressions of the proportion of dollar volume due to limit orders having time priority on the relative tick size. The regression uses the NASDAQ HFT data sample of February 22β26, 2010.The regression specification is:
ππππ πππππππ‘π¦ , , = π½ Γ π»πΉπππ’πππ¦ , , + π½ Γ π‘πππ , +π½ Γ π»πΉπππ’πππ¦ , , Γ π‘πππ , + π’ , + Ξ Γ π , +π , ,
ππππ πππππππ‘π¦ , , is the volume due to the liquidity provider having time priority over the liquidity providerβs total volume for stock i on date t of liquidity provider type n. π‘πππ , is the daily inverse of the stock price. π»πΉπππ’πππ¦ , , equals 1 if the revenue measure is from HFTs and 0 otherwise. π’ , represents the industry-by-time fixed effects. The definitions of the control variables π , are presented in Table 1. We report estimated coefficients and heteroskedasticity robust standard errors, clustered by firm. Standard errors are in parentheses. ***, **, and * indicate the statistical significance at the 1%, 5%, and 10% levels.
Dep. Variable Volume Due to Time Priority (1) (2) (3) (4) (5) (6) (7)
tickrelative 3.069*** 3.231*** 3.075*** 3.407*** 3.089*** 3.073*** 3.529*** (0.47) (0.49) (0.49) (0.56) (0.49) (0.47) (0.61) HFTdummy 0.083*** 0.083*** 0.082*** 0.082*** 0.083*** 0.083*** 0.081*** (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) tickrelative*HFTdummy 0.551** 0.551** 0.590** 0.571** 0.556** 0.552** 0.613** (0.28) (0.28) (0.29) (0.28) (0.28) (0.28) (0.30) logmcap 0.098*** 0.092*** 0.095*** 0.059*** 0.094*** 0.099*** 0.051*** (0.01) (0.01) (0.01) (0.02) (0.01) (0.01) (0.02) turnover 0.009* 0.008* (0.01) (0.00) volatility -3.963*** -3.506*** (1.32) (1.32) logbvaverage -0.007 0.000 (0.01) (0.01) idiorisk -1.134 -1.064 (0.86) (1.24) age 0.006*** 0.006*** (0.00) (0.00) numAnalyst 0.007* 0.006* (0.00) (0.00) pin -0.282 -0.271 (0.34) (0.32) pastreturn -0.073 -0.006 (0.16) (0.15) R2 0.644 0.655 0.655 0.685 0.647 0.645 0.699 N 1154 1154 1094 1144 1154 1154 1084 Industry*time FE Y Y Y Y Y Y Y
58
Table XI. Active Quote Improvement This table presents the regression of the proportional difference between non-HFTsβ and HFTsβ dollar sizes of orders which actively improve the NASDAQ BBO on the relative tick size. The regression uses the NASDAQ HFT data sample as of February 22β26, 2010. The regression specification is:
π·ππππππππ§π , = π½ Γ π‘πππ , + Ξ Γ π , + π’ , +π , where π·ππππππππ§π , is non-HFTs' minus HFTsβ dollar sizes of orders which actively improve the NASDAQ BBO over the sum of the two terms for stock i on day t. π‘πππ , is the daily inverse of the stock price. π’ , represents the industry-by-time fixed effects. The definitions of the control variables π , are presented in Table 1. We report estimated coefficients and heteroskedasticity robust standard errors, clustered by firm. Standard errors are in parentheses. ***, **, and * indicate the statistical significance at the 1%, 5%, and 10% level. Dep. Variable non-HFTs Minus HFTsβ Active Quote Improvement
(1) (2) (3) (4) (5) (6) (7) tickrelative -2.434** -2.499** -2.170** -2.267** -2.437** -2.374** -2.105* (0.97) (0.97) (1.00) (1.11) (0.99) (0.95) (1.18) logmcap -0.117*** -0.111*** -0.109*** -0.118*** -0.101*** -0.115*** -0.084* (0.02) (0.02) (0.02) (0.04) (0.03) (0.02) (0.04) turnover -0.015 -0.012 (0.01) (0.01) volatility 2.462 2.563 (3.80) (3.71) logbvaverage -0.019 -0.028 (0.02) (0.02) idiorisk 0.004 -0.210 (1.17) (1.56) age -0.002 -0.005 (0.01) (0.01) numAnalyst 0.003 0.003 (0.01) (0.01) pin 0.866 1.071 (0.70) (0.71) pastreturn -0.305 -0.202 (0.38) (0.41) R2 0.459 0.464 0.477 0.450 0.468 0.462 0.487 N 585 585 555 580 585 585 550 Industry*time FE Y Y Y Y Y Y Y