Algorithmic trading and market quality

John Paul Broussard, PhD, CFA, FRM, PRM

Estonian Business School
A. Lauteri 3, 10114 Tallinn, Estonia
and Rutgers University
Camden, New Jersey 08102, USA
[email protected]

Andrei Nikiforov
Rutgers School of Business – Camden
227 Penn Street, Camden, NJ 08102, USA
[email protected]

Sergey Osmekhin1
Department of Finance and Statistics, Hanken School of Economics
P.O. Box 479, FI-00101, Helsinki, Finland
[email protected]

August 14, 2020

Abstract:
A unique data set from NASDAQ OMX Nordic allows a deep analysis of trader types’ activity and
provides evidence on the roles played in the trading ecosystem. We specifically investigate the
impact of algorithmic traders on market quality relative to the activities of other market
participants under various conditions. We find that relative to other traders, algorithmic traders
contribute to lower spreads, especially during highly volatile markets, and provide more shares
traded at the NBBO. We also identify the main determinants of algorithmic traders’ liquidity
provisions and order cancellation patterns.

Acknowledgements:
The authors thank Frank Hatheway, Henrik Husman and staff of the NASDAQ OMX Nordic research group,
particularly Jakob Hallmer and Petter Dahlström, for their help and data provision. We appreciate Niklas Ahlgen,
Timo Korkeamäki, Juha Joenväärä and Terrence Hendershott for fruitful discussions and comments. Osmekhin
gratefully acknowledges the financial support from the NASDAQ OMX Nordic Foundation and the Ella and Georg
Ehrnrooth Foundation. Broussard was Distinguished Fulbright Professor and Visiting Research Professor , Hanken
School of Economics when effort began, and acknowledges the financial support of Fulbright, Fulbright Finland,
and Hanken.

1
These authors contributed equally to this work.

Page 1 of 35
Electronic copy available at: https://ssrn.com/abstract=3673881
Algorithmic trading and market quality
Abstract:
A unique data set from NASDAQ OMX Nordic allows a deep analysis of trader types’ activity and
provides evidence on the roles played in the trading ecosystem. We specifically investigate the
impact of algorithmic traders on market quality relative to the activities of other market
participants under various conditions. We find that relative to other traders, algorithmic traders
contribute to lower spreads, especially during highly volatile markets, and provide more shares
traded at the NBBO. We also identify the main determinants of algorithmic traders’ liquidity
provisions and order cancellation patterns.

JEL classification: G20, G14, L10

Page 2 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

Algorithmic trading and market quality
Highlights:
· Our study has two unique features: It identifies the entirety of algorithmic
traders in the market and it studies their impact relative to other trader
types.
· We contribute to three strands of literature. First, we find that algorithmic
traders, as a group, are contributing to tighter spreads and more depth.
· Second, we find that algorithmic traders do NOT withdraw from the market
during highly volatile time periods.
· Finally, we investigate the determinants of market making behavior and
order cancellation activities.

Page 3 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

Algorithmic trading and market quality

1. Introduction

Recent technological developments have transformed financial markets and the way
assets are traded. Trading floors have almost completely been replaced by electronic trading
platforms (Jain, 2005). The highly automated trading processes have brought a new set of
benefits and challenges. The benefits include faster execution, lower costs, and often higher
liquidity. It also led to the creation of a new type of market participant: the algorithmic trader,
and as a group, otherwise known as “ALGOs.” The challenges are increased scrutiny from
researchers and market regulators with regards to overall impact on markets. As noted by
Abergel et al (2012), the research on the impact of algorithmic traders on market quality has
been split with some papers pointing to algorithmic traders predatory behaviours, while others
find mostly beneficial effects.

This paper contributes to the literature on how algorithmic traders affect market quality
by utilizing unique data provided by NASDAQ OMX Nordic. The data allow for clean separation
of traders who utilize algorithms to conduct their trading, which includes HFT (high-frequency
trading firms) as well as “slower” frequency algorithmic traders, non-algo professional traders,
and retail trading classifications. Our findings are summarized as follows.

First, algorithmic traders decrease the spread on average by almost 10% relative to
retail traders. Algorithmic traders’ contribution to spread reduction is at least 50% greater than
that of the next “beneficial” group – institutional traders. We also find that as turnover sharply
increases, it is the institutional and professional traders who exhibit aggressive “picking up” of
limit orders, which widen the spread, not the algorithmic traders.

Second, we find that consistent with the more recent literature (Brogaard et al 2018),
that during highly volatile days, neither algorithmic trader or institutional traders impact
spread, or volume of shares traded at the NBBO. Professional non-algorithmic traders,
however, appear to negatively affect market quality during top quintile volatility days. We also

Page 4 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

find that in general, higher spreads encourage market making activity, and higher volatility
reduces it.

Third, we present evidence that retail traders supply liquidity to the other three types,
with algorithmic traders being most active in removing liquidity. This finding is unique since our
study includes all types of algorithmic traders. Finally, we show that share turnover, the size of
the firm, the relative bid-ask spread, and the total volume of shares in the order book predict
higher rate of order cancellation.

The paper proceeds as follows. Section 2 reviews the literature and connects our effort.
Section 3 presents data specification and research methodology. Section 4 presents and
discusses the results. Section 5 concludes. Additional clarifying material is found in the
Appendix.

2. Literature Review

Hendershott and Riordan (2013) address the question of whether trades and quotes
posted by algorithmic traders contain value-relevant information. They find algorithmic traders
contribute to price discovery. Brogaard, Hendershott and Riordan (2014) examine the role of
HFT in price discovery and show that overall, HFTs increase market efficiency. They document
that HFT tend to trade in the direction of the permanent price changes and against the
temporary price deviations. Karajczyk and Murphy (2019) look at market making to large
institutional traders and find that HFTs decrease bid-ask and price impact for small uninformed
orders, but increase the costs for large informed institutional traders. Van Kervel and Menkveld
(2019) confirm that HFTs supply liquidity to the market, but quickly reverse positions when they
detect persistent one-sided buying or selling. Informed trading costs are increased, but price
discovery is accelerated. Finally, Hu and Murphy (2020) studied closing auctions usually
conducted on the floor of the NYSE. When the floor closed due to the COVID-19 pandemic,
causing auction execution to be completely automated, the quality of auctions substantially
improved as measured by indicative closing auction prices and smaller order imbalances. We
also find that under most conditions, algorithmic traders (broadly defined as HFTs, but also all

Page 5 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

other types of algorithmic traders) contribute to improved market liquidity as measured by
average bid ask spread and depth.

Our paper also contributes to the literature on how heightened volatility impacts the
behavior of algorithmic traders. Anecdotal evidence (e.g. Kirilenko et al 2017) suggests that HFT
market makers withdraw from markets when they experience higher than usual volatility. Ait-
Sahalia and Saglam (2016) link elevated price volatility with smaller depths, reflecting market
makers lower capacity to bear risk in increased volatility environments. However, further
research finds that connecting negative behavior of algorithmic traders to heightened volatility
is not as direct. Hasbrouck and Saar (2013) study whether volatility influences algorithmic
traders or vice versa. They use 10-minute volatility intervals for NASDAQ stocks and find that
algorithmic trading decreases short-run volatility. Brogaard et al (2018) show that during very
extreme price events, HFTs supply liquidity, but only if it is a single-stock event. When multiple
stocks are breaking down, HFTs begin demanding liquidity. They find no evidence that HFTs
cause the events themselves.

Similarly, Ait-Sahalia and Brunetti (2020) look at every transaction in the SP500 futures
market, and find that HFTs do not induce volatility jumps. In fact, volatility and jumps are more
prevalent when HFTs trade less intensely. They find that spikes in volatility cause HFTs to
increase trading activity as higher levels of noise generate more opportunities. Zhou, Kalev and
Frino (2020) also find that stocks with higher HFT trading experience fewer price drops, or
surges, when the market is especially volatile. This is consistent with HFT minimizing price
pressures and mitigating transitory price errors. We also find that algorithmic traders, as a
broad category relative to other trader types, do not cause spreads to widen during high
volatility events

Another effect of the rapid and substantial growth of algorithmic trading is increased
competition among trading platforms. Competition has triggered a sharp decline in trading fees
(Colliard and Foucault, 2012). Exchanges and other trading venues started charging different
fees for primary counterparty and secondary counterparty trades. The traders who provide
liquidity by providing competitive limit orders (the so-called “makers”), obtain a rebate from

Page 6 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

the exchange, while the party that accepts the existing price (the so-called “takers”) will pay an
extra fee. Hence, the usually termed “maker-taker” arrangement.

Hagströmer and Norden (2013) study the behavior of market making HFT and
opportunistic HFT at NASDAQ OMX Stockholm. They find that market making constitutes a
larger share of HFT trading volume. Malinova and Park (2016) study the impact of maker-taker
fees on market quality on the Toronto Stock Exchange. They find that a breakdown of exchange
fees into maker-taker fees does not impact liquidity. It is the total fee that matters. This was
predicted earlier by Colliard and Foucault (2012), who show that only changes in the total
trading fee (maker and taker fee together) retained by the exchange affect liquidity and trading
volume. Foucault, Kadan and Kandel (2013) develop a model that provides an explanation for
maker-taker pricing. They argue that lower trading fees increase the trading rate when the level
of competition among liquidity providers is low compared to among liquidity demanders.

This paper employs two unique features that distinguish it from prior research. First, we
can cleanly separate market participants into four main groups: 1) algorithmic traders, who
utilize their own proprietary algorithms for all trading activities (this group includes HFTs)2; 2)
institutional traders, who employ only NASDAQ’s approved algorithms (such as VWAP, TWAP,
etc) to execute their trades; 3) professional investors who do not use algorithms; and 4) retail
traders. Second, clear separation of the four trader types allows us to study the impact of all
traders on market quality, viewing them as an integral part of the entire trading ecosystem.
Thus, we can judge how algorithmic trading activities impact market quality relative to other
market participants. Our paper also investigates the determinants of algorithmic traders’
decisions to add or remove liquidity, as well as order cancellation activity.

Our primary finding is that algo traders decrease the spread on average by almost 10%
relative to retail traders. It is the institutional and professional traders who aggressively trade
picking up limit orders and widening the spread the most, not algorithmic traders. In line with
recent literature (Brogaard et al., 2018), neither algorithmic trader nor institutional traders

2
One downside of this feature is that we do not know the composition of this group. How many are market
makers, opportunistic traders, or both? Some studies that utilize a similar dataset to ours but with more
granularity allows more in-depth analysis of this specific issue (e.g. Hagstromer and Norden, 2013)

Page 7 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

impact the spread, or the volume of shares traded at NBBO in higly volatility environments. The
only group that affects spreads negatively during days in the top volatility quintile is a group of
professional traders.

We also present evidence that retail traders supply liquidity to the other three trader
types. This finding is unique in the literature, and makes a contribution to the literature since
our study includes all types of algorithmic traders. Finally, we investigate the impact of order
cancellation in our data. We show that share turnover, the size of the firm, the relative bid-ask
spread, and the total volume of shares in the order book predict higher rate of order
cancellation.

3. Data Specification and Methodology

The analysis is based on a unique and proprietary dataset provided by NASDAQ OMX
Nordic research group. The paper uses an 18-month window from 1 June 2010 to 30 November
2011 where market quality as well as pricing data for each day/firm are obtained. The full
sample contains 382 trading days.Trading days that have no trades on one of the exchanges at
NASDAQ OMX Nordic (an example is country specific holidays) are excluded from the sample.

The trading volume on NASDAQ OMX can be attributed to four main types of traders
classified by the exchange: 1) algorithmic traders, who utilize their own proprietary algorithms
for all trading activities (this group includes HFTs); 2) institutional traders, who employ only
NASDAQ’s approved algorithms (such as VWAP, TWAP, etc) to execute their trades; 3)
professional investors who do not use algorithms; and 4) retail traders. Figure 1 shows the
composition of traders by Type in the market during the sample period as a percent of total
volume over time.

FIGURE 1 ABOUT HERE

Figure 1 clearly shows that the fraction of total proprietary algorithmic volume has been
steadily increasing at the expense of the other three types of traders over the sample period
from 06/2010 to 12/2011.

Page 8 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

 Figure 2 separates the fraction of total trading volume generated by proprietary
algorithmic trading (ALGO). The figure shows that algorithmic trading activity (the segment at
the top of the Figure 1) almost doubled over the sample period.
FIGURE 2 ABOUT HERE

The data provided by NASDAQ OMX Nordic contain the following variables:

“Date” is a daily date variable indicator for the period 1 Jun 2010 – 30 Nov 2011.
“Order book” is the NASDAQ trading symbol for an individual stock.
“Acks” is the amount of partial executed and then cancelled orders.
“Executions” is the amount of orders that executed completely.
“Cancels” is the amount of orders that were cancelled without any partial execution.
“Account type” contains information about member account type and price list
specification, including algorithmic trading discount properties. “Account type” allows
distinguishing automated traders from others.
“Liquidity” captures whether the trade is adding or removing liquidity.
“Volume” is the traded amount of the security.
“RTWAS” corresponds to relative time weighted average spread.
“TVOB Volume” and “TVOA Volume” are total volume value of best and all levels in
orderbook, respectively.3

Information about Last/Open/High/Low prices and Market Capitalization for each stock
are also included.

The total number of traded symbols on NASDAQ OMX Nordic is over 7000 during the
sample period, however, most securities traded rather sparsely. We filter the data according to
the procedure of Hendershott, Jones and Menkveld (2011), and retain only the symbols that
have at least 21 trades during each trading day. Stocks with an average closing share price of
less than €2 are removed from the sample, as are stocks with an average closing share price of
more than €500. The resulting sample contains observations for 156 unique stocks
correpsonding to 458,813 observations.

TABLE 1 ABOUT HERE

3
“TVOB Euro” and “TVOA Euro” are total value in Euro of best and all levels in orderbook, respectively. If trades
are in different currencies, the daily FX rates from the Swedish Central Bank are used to convert all amounts into
Euros.

Page 9 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

 Table 1 presents summary statistics for all major variables used in the study. Panel A
lists the statistical characteristics of the “trading” variables. TOTVOL is average daily number of
shares traded and TURN is the turnover, the ratio of shares traded to the total shares
outstanding. Both variables have a minimum of zero. The maximum TURN in the sample is over
26% of the outstanding shares corresponding to a volume of over 42 million shares. The
average daily TURN is around 0.04% corresponding to roughly 150,000 shares. The smallest
company in the sample has a Market Capitalization of 20 million Euros and the largest exceeds
66 billion Euros. The average company has a mean Market Capitalization of 4.5 Billion Euros.
The lowest average price in the sample is just above 2 Euros and the highest is just under 120
Euros with the average price of about 17.5 Euros. The next three variables represent
executions. The mean number of partial executions (ACKS) is just under 10,000 per day (median
is under 2,000). The average number of fully executed orders is 1,074 (median of only 440) and
the number of cancelled orders is 11,017 (median of 1,622).

Panel B of Table 1 presents the market “quality” variables. The relative Bid-Ask spread
weighted by time (RTWAS) ranges from 0.03 to 9.58% with the average of 0.27 (median of
0.17). The average total number of shares offered at best quotes (TVOBV) is around 4,000,
which is roughly 1/30 of the average total number of shares offered at all quotes (TVOAV) at
36,700 shares. The other two variables (TVOBR and TVOAR) are scaled by the price in Euros. VLT
is the daily volatility variable defined as (HIGH – LOW)/CLOSE. The variable shows the daily
range of the stock’s price scaled by the closing price. This variable helps identify the daily
activity in the stocks even when the open and the close prices ended up the same. Martens and
van Dijk (2007) show that this measure of daily volatility, called “realized range,” is a more
efficient estimator of daily volatility than realized variance (square root of daily return). For an
average stock in our sample, prices change almost 3% a day (the median is 2.5%). The minimum
daily move is 0.28% while maximum daily move is over 58%. Since most variables are highly
skewed to the right, we took log transformations of all major variables for the purpose of
multivariate analyses.

Page 10 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

 Finally, Panel C of Table 1 shows the statistical characteristics of several variables
filtered by trader type. We can see that the average daily TURN is the highest for non-algo
professional traders at 0.07%, followed by retail investors at 0.04%, with algo traders and
institutional traders not far behind. Even though the number of fully executed orders (EXEC) is
comparable across the different trader types (from 745 for institutional automated traders to
1,388 for professional manual traders), the number of partially executed and cancelled orders is
more than ten times higher for ALGO (at 30 thousand partially executed and 37 thousand
canceled daily orders) vs the other trader types partially executed and canceled approximately
the same number of orders (at around 3,000 per day.) This finding is consistent with algorithmic
traders use of rapid submission and cancellation of orders to “feel” the market (Griffith and Van
Ness, 2020). The next variable, MAKING, shows the proportion of orders that supply liquidity
(making the market as opposed to taking the market) relative to the total traded volume.
Professional and retail traders on average are liquidity suppliers, submitting limit orders, while
institutional and algo traders are on average liquidity takers4. The final variable in Panel C, the
CORR, is the correlation between liquidity “making” and “taking” orders for each type. If a
trader consistently uses both limit and market orders every day, the CORR will be close to one,
while leaning on one side (such as only “making” liquidity one day and then “taking” the next)
will result in lower correlation value. The table shows that algorithmic (ALGO) and institutional
types (PROF) are more balanced in their use of “making” vs “taking” orders relative to the other
two types (INST and RETL) who tend to prefer one mode over the other on different days.

Table 2, documents the correlations between the main market variables used in the
paper. Not surprisingly, the average total number of shares available at the best bid/ask,
TVOBV, is positively correlated with the time weighted spread, RTWAS, since wider spread
makes it more attractive to place limit orders inside or at NBBO. However, other factors, such
as volatility could be the reason for the increase in spread, which would make placing more

4
This finding might seem contrary to the existing literature where algorithmic traders are seen primarily as
liquidity suppliers. However, our data does not distinguish between HFTs and non-HFTs algorithmic traders and
thus combine both the market makers and the opportunistic algorithmic traders as well as “lower frequency”
aggressive proprietary algorithms from banks and investment firms. When taken as a whole, pure market making
HFTs are barely half of all algorithmic traders (Hagstromer and Norden, 2013). Thus, our paper provides a wider
window on the activity of ALL algorithmic traders and not just the market making HFTs.

Page 11 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

limit orders at NBBO riskier. The spread, RTWAS, and the volatility, VLT, are highly correlated
with a correlation coefficient of 0.32. The correlation between RTWAS and TVOBV is
significantly positive but very close to zero. Perhaps surprisingly, when looking at the
correlation between the relative spread and the total number of shares available as limit
orders, the relation becomes negative. This could be a volatility driven effect. When we look at
the correlation between total turnover, TURN, with the relative spread, RTWAS, we find that it
is slightly negative, which makes because higher spreads tend to reduce trading activity
(Chordia, Roll, Subrahmanyam, 2011). Finally, the ratio of making to taking activity (Making) has
a somewhat positive correlation with the relative spread indicating that when the spreads are
higher, more traders are willing to provide liquidity by posting limit orders. However, the
correlation is pretty small (significant at 1% level) due to potentially a mitigating effect of the
higher volatility and thus increased risk.

TABLE 2 ABOUT HERE

In the next section we consider a series of multivariate tests to shed light on the activity
of algorithmic traders relative to the other trader types controlling for various factors. We aslo
perform robustness tests.

4. Results and discussion

4.1 The effect of different types of traders on relative spreads

To investigate the effect of algorithmic traders on market quality, we fit a simple model
(1), where market quality measures are regressed on dummy variables representing the four
trader types while controlling for turnover, volatility, log price and log market capitalization.

(1)’s Dependent Variableit is one of the market quality measures RTWAS, TVOB, and TVOA.
TURN is the turnover on day t, VLT is the market volatility defined as daily (High -Low)/Close for

Page 12 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

each stock/day. Finally, the Dummy stands for the indicator variable identifying each trader
type. We use the retail trades dummy RETL as the base to which we compare the effect of the
other three types. All variables, except for Dummies, are log-transformed.

Table 3 presents the results of the regression for all three dependent market quality
variables per model (1). The dummy variable for the algorithmic traders is negative and
significant and indicates that the presence of algorithmic traders reduces spreads by around
9.5% relative to the retail traders. The effect of institutions is also positive for the spread, with
an estimated reduction of almost 6% relative to retail traders. Professional traders, on the other
hand, increase the spread by over 5% relative to the retail traders. Thus, this simple model
suggests positive effect of algorithmic traders on the time weighted relative spread, which is
consistent with other literature on the topic (e.g. Brogaard et al., 2014). The control variables
are also statistically significant. It appears that one percent increase in turnover (TURN)
decreases the spread by 0.13%. This is not surprising as higher volume promotes liquidity.
Volatility, on the other hand, has profoundly negative effects on the spread. Each percentage
point increase in volatility widens the spread by almost a quarter percentage point (0.243%)
holding everything else constant. The two other dependent variables (TVOBV and TVOAV) show
similar dynamics.

Algorithmic traders’ activity is associated with the increase in both the total volume of
shares traded at the NBBO, as well as at all other levels by about 6% and 19%, respectively,
relative to retail traders. Institutional investors also contribute positively, increasing the
number of shares at NBBO by around 6%, and at other levels by about 18%, similar to those for
algorithmic traders. The manual professional traders seem to have negative impact on the
spreads by reducing both numbers relative to the retail traders. Volatility affects the shares at
NBBO much stronger (0.34% reduction for every 1% increase in volatility) than shares at all
levels (0.28% reduction).

TABLE 3 ABOUT HERE

Page 13 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

 There is one variable, PRICE, that has unexpected signs across the three dependent
variables. In the relative spread regression, one percent increase in price of a stock, decreases
the relative spread by around 0.20%. This would make sense, as lower priced shares tend to be
more speculative and thus riskier. However, since we control for volatility, there must be
another explanation5. This is especially puzzling since the one percent increase in price
decreases the shares available at NBBO by over 1% and the shares available at all price levels by
essentially the same amount.

4.2. The sensitivity of ALGO traders to increased turnover

Since at least some of the algorithmic traders thrive on quick “in and out” style of
trading (e.g. HFTs), we next investigate the effect of rising (and falling) total volume on the
effect of algorithmic traders on market quality variables relative to the base group: the retail
traders. We fit the following model, where the dummy variable indicator for each type of
traders interacts with the log of total daily turnover for each stock.

where the Dependent Variableit in (2) is one of the market quality measures RTWAS, TVOB and
TVOA. TURN is the log of total daily turnover, Xit is a vector of control variables including log
Volatility, Log Price and Log Market Cap.

Table 4 shows the results of regressions generated by (2). To interpret the positive
coefficient on the interaction between ALGO and TURN we need to consider a hypothetical
level of turnover. For example, at the 75% level of log(TURN) =log(0.053%) =-3.275, the effect of
algorithmic traders on spread relative to retail traders is (-0.071) + (-3.275)*(0.003) =-0.0805 or
approximately 8% reduction in spreads. Similarly, at the average level of turnover, the effect of
institutional traders leads to the reduction of about 1.7% while the activity of professional

5
One possible explanation is the regulation whereby some institutions will avoid low priced stocks to stick with the
“blue chips”. This behavior increases the demand and trading volume of higher priced stocks and could lead to the
reduction in spreads unrelated to their lower volatility (e.g. Green and Hwang, 2009).

Page 14 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

traders is associated with an increased spread, of around 7.7%. This might be an indicator of
how aggressively professional traders pursue the profit opportunities that might be associated
with higher turnover.

Looking at the other two variables, we can see that algorithmic and institutional traders
activity reduce the number of shares at the NBBO and all levels, while the professional traders
activity is associated with increasing number of shares available at NBBO and at all levels as
turnover increases. For example, at the 75th percentile of turnover level (log(0.053%) = -3.27),
the activities of algorithmic traders are associated with a significant reduction of shares at
NBBO of about 40% relative to the activities of retail traders. Institutional traders reduce that
number by about 9%, and professional discretionary trading leads to increase in shares offered
at both NBBO and all levels by about 9% and 5% respectively.

TABLE 4 ABOUT HERE

A possible interpretation of this finding is that during times of higher than usual volume,
which is likely associated with fast moving markets6, algorithmic traders utilize their speed to
transact at the NBBO with smaller sizes, thus leading to both tighter spreads and fewer number
of shares offered at any one point.

4.3. The sensitivity of algorithmic traders to volatility

Several recent papers suggested that volatility might “scare off” algorithmic traders
precisely when their involvement is especially needed. Huang and Wang (2008, 2010) develop a
theoretical framework in which market crashes emerge endogenously when a sudden order
imbalance overwhelms the insufficient risk-bearing capacity of market makers. Further, Ait-
Sahalia and Saglam (2016) link elevated price volatility with low willingness to supply liquidity
by “high frequency” intermediaries, reflecting their risk aversion to high volatility. As a result,
market tends to “dry up” when volatility spikes (e.g. Kirilenko at all., 2017). We investigate this
issue by fitting the following model of interaction between trader types and stock’s volatility:
6
Our volatility measure (VLT) captures only the daily range and does not account for multiple directional changes
in price during the day, which could be accompanied by higher uncertainty and hence turnover.

Page 15 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

where the Dependent Variableit in (3) is one of the market qualities measures RTWAS, TVOB and
TVOA. VLT is the log of stock’s relative range during the day. Xit is a vector of control variables
including log Turnover, Log Price and Log Market Cap.

Table 5, Panel A shows the results for market quality variable RTWAS (relative time-
weighted average spread). Perhaps surprisingly, the coefficients for both ALGO and Institutional
traders are not significantly different from zero when both types are interacted with volatility.
In other words, relative to retail traders, neither algorithmic nor Institutional traders influence
the spreads as volatility moves up or down. Having said that, the activities of discretionary
professional traders is strongly associated with increased spreads, suggesting aggressive limit
orders by this trader type during more volatile times. The other two measures of market quality
(TVOBV and TVOAV) show similar results7 .

To check the possibility that the influence of algorithmic traders might only be visible at
higher levels of volatility, we sorted all stock/days into quintiles based on volatility and refit our
regression analysis by interacting the type of traders indicator variable with the indicator
variable for the five volatility quintiles from the “calmest” (Q1) to the “stormiest” (Q5) of
environments. The results are shows in Table 5, Panel B.

TABLE 5 ABOUT HERE

First, the dummy indicator variables are all significant in both economic and statistical
senses. The volatility quintiles also demonstrate monotonic behavior by increasing the spreads
relative to Q1, the calmest periods. The spread increases by about 10% for Q2 over Q1. It goes
up by another 6% from Q2 to Q3 with the largest increase in spreads associated with the
“stormiest” days when spreads are almost 36% higher compared to the calmest days (Q1).
Turning our attention to the interaction terms we see that the activity of algorithmic and
institutional traders during the higher volatility periods (Q3, Q4, and Q5) is associated with only

7
The data for these two variables are available upon request.

Page 16 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

small and statistically insignificant increase in spreads. The only group of traders with significant
impact on spreads is the discretionary professional traders. It seems their activity during higher
volatility periods (Q4 and Q5) significantly impacts spreads by widening them, perhaps due to
their aggressive picking of limit orders during this “stormy” phases. However, the data provide
no evidence that either algorithmic or institutional investors impact the quality of the market
negatively as volatility increases.

4.4. The effect of market making activities

In this section, we investigate how adding and removing liquidity by the four types of
traders impact market quality variables. Prior research showed that HFTs tend to be
overwhelmingly in the “market making” category. However, our study includes ALL algorithmic
traders comprised of market making as well as opportunistic HFT and non-HFT funds. While our
data allow for clear separation of the algorithmic from non-algorithmic trades, we cannot easily
distinguish among the diverse taxonomy of algorithmic trading modalities. Other studies that
had access to much more granular (although much shorter horizon) datasets from NASDAQ
OMX (e.g. Hagstromer and Norden, 2013) show that among HFTs, multiple trading strategies
are present and range from pure market makers to pure arbitrageurs (aggressive opportunistic
traders) with multiple strategies falling in between. Moreover, HFTs tend to constantly switch
from one strategy to the next making it difficult to study HFTs behaviour at the individual level.
Instead, we look at the aggregate market making (or taking) actions among ALL “species” of
algorithmic traders. To that end, we fit the following model:

where the Dependent Variableit in (4) is market property measures RTWAS, TVOB and TVOA.
TYPE and MAKE are dummy indicator variables that identify various types of traders and

Page 17 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

whether the liquidity is added (MAKE =1) or removed (MAKE =0), respectively. Xit is a vector of
control variables, including log Volatility, logTurnover, Log Price and Log Market Cap.

Takers, who remove liquidity, enter their market orders while accepting the existing
price. The effect of liquidity removal activity should be an increased bid-ask spread. Makers,
who add liquidity by posting their limit orders should lead to a decreased bid-ask spread.

The results of four specifications of regression (4) are shown in Table 6. Specification (a)
where we run regression (4) without interacting MAKE and TYPE indicate that market making
activity has significantly reduced the relative bid-ask spread. Trades that add liquidity seem to
improve the spread by 0.004%. Next, we explore if there is a trader type specific effect on
spreads by re-fitting the above model (4) by interacting the TYPE and MAKE indicator variables.
Specification (b) in Table 6 presents the impact of the interaction terms on RTWAS, the relative
spread, where the base case are retail traders who remove liquidity.

TABLE 6 ABOUT HERE

The coefficient on MAKE turns positive (compared to the basic specification), indicating
the retail investors who remove liquidity worsen the bid-ask spread, an expected result. All
interactions between the type of traders and MAKE variable are statistically significant and
negative. Among the three types, algorithmic traders have the biggest positive effect, followed
by institutional traders, and then professional manual traders. Specifically, algorithmic traders,
when they provide liquidity, improve the spread by approximately 0.10% relative to the retail
traders (adding the coefficients on TYPEALGO, MAKE, and the interaction together). The
Institutional liquidity providers improve the spread by half that at 0.05%. Perhaps surprisingly,
the activity of professional traders who provide liquidity is associated with an increase in spread
by almost 0.07%. This finding provides more evidence (along with the results on extreme
volatility) that the trading activity of professional manual traders seem to be associated with
worse overall liquidity.8 Fitting the regressions for the other two market quality variables, the

8
Since we do not investigate causality, one interpretation of this result (and the one related to the volatility in the
previous section) is that professional manual traders only step in when the liquidity is already “bad”. In other
words, they start posting limit orders when the bid ask spread is already large.

Page 18 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

volume of shares at NBBO (specification (c)) and at all levels (specification (d)), shows a similar
picture. Algorithmic traders make the biggest contribution to market quality by increasing the
number of shares at NBBO by about 6.5% (adding the coefficients on TYPEALGO, MAKE and the
interaction term together), followed by the institutional investors at 6.1% improvement with
professional traders contributing to a negative 7.3% decline relative to the retail investors. The
evidence here suggests that algorithmic traders’ activity is associated with tighter spreads and
more shares available at NBBO and all levels, relative to the retail investors, when algorithmic
traders are engaged in market making pursuits.

4.5 What determines the making/taking activity of traders?

In this and the next sections, we investigate other variables that measure market
quality: the fraction of liquidity supplying orders and the rate of cancelled orders.

(5) describes the maker-taker turnover rate and indicates relative maker-taker activity. The
variable increases when the relative activity of adding liquidity is going up and decreases when
the relative activity of making declines. To examine the impact, the following regression is run:

where the Dependent Variableti in (6) is MAKE_FRAC as in (5). Xti is a vector of control variables
including log Turnover, Volatility, Log Price and Log Market Cap as well as the log-transformed
values of RTWAS, TVOAV, and TVOBV. The results of the regression (6) are presented in Table 7.

First, we tried to fit the OLS regression but because both dependent variables are ratios,
the residuals exhibit strong heteroscedasticity (studentized Breusch-Pagan test rejects the null
of homoscedasticity at 1% level). Instead of OLS, we fit a more appropriate model for a
proportion, the beta regression model. Ferrari and Cribari-Neto (2004) have proposed using
beta regression model for continuous response with values confined to the unit interval (0, 1).

Page 19 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

The ML estimates of the coefficients of the beta regression model for market making are
presented in Table 7, Panel A. As in the previous regression models, the retail traders serve as
the base case.

We can see that the coefficient on trader type dummy variables are negative indicating
that, on balance, all three types of traders remove liquidity relative to the Retail traders, who
serve as the net suppliers of liquidity to these three types. Algorithmic traders’s effect is almost
twice as big as that of the institutional traders which, in turn, is 40% larger than that of the
professional traders. This result echoes the univariate findings from Figure 3, which presents
the distribution of MAKE_FRAC ratio for each trader type. As in the figure, the multivariate
results of this section suggest that taken as a whole, algorithmic traders are more likely to
remove liquidity than to add to it. Paradoxically, this removal is very “gentle” as it does not
seem to contribute to worsening of any of the market quality variables (RTWAS, TVOBV,
TVOAV).

FIGURE 3 ABOUT HERE

Four other independent variables positively influence market making activities. The high
bid-ask spread (RTWAS) has the highest contribution, followed by the total number of shares
available at all levels of limit book (TVOAV), followed by price. Perhaps higher priced stocks
attract more market making activities when the turnover, volatility, and relative spread are held
constant. Finally, the higher volume has a positive relation to increased market making.

The variables that adversely affect market making activities are (from strongest to
weakest): market volatility, the total volume of shares at NBBO (TVOBV), and the market cap of
the firm. These findings are not surprising and agree with the established literature: higher
volatility makes it risky to place limit orders since providers of liquidity can be adversely
selected against in a likelier unfavorable move. Also, when there are more shares available at
NBBO it makes little sense to make markets at that level. The effect of the market cap is
puzzling although the value of the coefficient is small.

TABLE 7 ABOUT HERE

Page 20 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

4.6 What determines the rate of canceled orders

We next tackle the issue of cancelled orders. When orders are submitted, but then
cancelled, market participants get an inflated estimate of available liquidity. When liquidity is
this “ephemeral” participants act more cautiously. Thus, higher rates of cancelled orders
increase uncertainty and decrease overall market quality. We compute the ratio of Cancelled-
to-Total orders for each stock-day-type and estimate the following regression using the beta
regression model approach (6) utilized in the prior section, but replacing the dependent
variable with CANCELLED_FRAC.

Table 7, Panel B results indicate that algorithmic traders have the biggest influence on
cancelled orders, which is not surprising, since they are responsible for over 70% of them. The
two other trader types: Institutional and professional traders, have a negative effect on
cancelled orders compared to the base group: the retail traders. In general, the coefficients for
other variables in this model are much smaller than those in the previous model as algorithmic
traders explain most of the variation in the cancelled orders. However, all other variables have
statistically significant impact on the rate of order cancellation. Share turnover, the size of the
firm, the relative bid-ask spread and the total volume of shares in the order book predict higher
rate of cancellation. It is well known (see Hasbrouck and Saar, 2013 and Ait-Sahalia, Yacine, and
Mehmet Saglam, 2016) that higher bid-ask spread induce algorithmic traders to place orders

inside the spread near the bid or the offer and then quickly cancel them to “fish out” a most
current “true” price (see e.g. Hasbrouck and Saar, 2013). Higher number of shares available in
the order book might also encourage higher number of cancelled orders to prevent them from
becoming “stale.” The effect of market cap, although small, is positive indicating larger order
cancellation share for larger firm.

Volatility, price, and the total number of shares at NBBO have a negative effect on the
rate of cancellations. Higher volatility might result in less “fishing” behavior as the “true” value
is constantly changing and the shares are hit and lifted as soon as they are placed, thus making

Page 21 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

cancellations unnecessary. The negative influence of stock price could be explained if
algorithmic traders concentrate their activities in the lower priced shares for liquidity reasons9.

5. Conclusion

We study the impact of algorithmic trading activity on market quality. Our approach has
two unique features. First, using an exclusive data set from NASDAQ OMX Nordic, we
distinguish between four main types of traders: 1) algorithmic traders, who utilize their own
proprietary algorithms for all trading activities; 2) institutional traders, who employ NASDAQ’s
approve algorithms (such as VWAP, TWAP, etc.) to execute their trades; 3) professional
investors who do not use any algorithms; and 4) retail traders. Second, this separation allows us
to judge how the entirety of algorithmic trading activities impacts the market relative to other
market participants in the trading ecosystem

We contribute to three strands of literature. First we start with a basic question of how
the presence of algorithmic trading activity affects market quality, finding that algorithmic
traders decrease the spread by an average of almost 10% relative to retail traders. Algorithmic
traders’ contribution to spread reduction is at least 50% greater than that of the next
“beneficial” group – institutional traders. We also find that as turnover sharply increases, it is
the institutional and professional traders who exhibit aggressive “picking up” of limit orders and
thus widen the spread, not the algorithmic traders.

Second, we contribute to the literature of volatility and algorithmic traders. When we

investigate the effect of increased volatility, we find that consistent with the more recent
literature (Brogaard et al 2018), neither algorithmic trader nor institutional traders impact the
spread, or the volume of shares traded at NBBO. The only group that affects the market quality
negatively during the days in the top volatility quintile are professional non-algorithmic traders.

Our third contribution is to the literature on the determinants of market making and
order cancellation behavior by traders. We present evidence that retail traders supply liquidity

9
For the same reason that firms split their shares when the share price rises “too much”. For example, Lin Singh,
and Yu (2009) show that stock splits reduce costs and improve liquidity.

Page 22 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

to the other three types with algorithmic traders being the most active in removing liquidity.
This finding is unique since our study includes all types of algorithmic traders. We find that in
general, higher spreads encourage market making activity, and higher volatility reduces it.
Finally, we investigate the impact of order cancellation in our data. We show that share
turnover, the size of the firm, the relative bid-ask spread, and the total volume of shares in the
order book predict higher rate of order cancellation.

These results have implications for exchanges, regulators and traders. Non-algorithmic
trading participants, together with algo-traders, represent a symbiotic system where all sides
benefit. Non-algorithmic traders gain from a lower bid-ask spread, and algo-traders profit from
their unique ability to utilize speed. The behavior of the non-algorithmic professional traders in
the presence of algorithmic trading needs to be studied further to determine overall interaction
and impact.

Page 23 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

References:

Abergel F., Bouchaud J.-P., Foucault T., Lehalle C.-A. and Rosenbaum M. (2012) “Market
microstructure: Confronting Many Viewpoints” Wiley.

Aït-Sahalia, Yacine and Celso Brunetti. (2020). High frequency traders and the price process,
Journal of Econometrics, Volume 217, Issue 1, Pages 20-45.

Brogaard, J., Carrion, A., Moyaert, T., Riordan, R., Shkilko, A. and Sokolov, K. (2018). High
frequency trading and extreme price movements. Journal of Financial Economics, 128(2),
pp.253-265.

Brogaard, J., Hendershott, T. and Riordan, R. (2014). High-frequency trading and price
discovery. The Review of Financial Studies, 27(8), pp.2267-2306.

Tarun Chordia, Richard Roll, Avanidhar Subrahmanyam. (2011) Recent trends in trading activity
and market quality, Journal of Financial Economics, Volume 101, Issue 2, pp. 243-263.

Colliard, J.-E., and T. Foucault (2012) “Trading fees and efficiency in limit order markets”,
Review of Financial Studies 25, 3389-3421.

Ferrari, S., & Cribari-Neto, F. (2004). Beta regression for modelling rates and proportions.
Journal of applied statistics, 31(7), 799-815.

Foucault, T., O. Kadan, and E. Kandel (2013) Liquidity Cycles and Make/Take Fees in Electronic
Makets, Journal of Finance, 1, 299-341.

Green, T. Clifton and Byoung-Hyoun Hwang. (2009), Price-based return comovement, Journal of
Financial Economics, Volume 93, Issue 1, 2009, Pages 37-50.

Hagstromer B. and Norden L. (2013) “The diversity of high-frequency traders” Journal of

Financial Markets, 16, 741-770

Hasbrouck J. and Saar G. (2013) “Low-Latency Trading”, Journal of Financial Markets, 16, 646-
679

Hendershott, T., C. Jones, and A.J. Menkveld (2011) “Does Algorithmic Trading Increase
Liquidity?” Journal of Finance 66:1-33.

Hendershott T., and Riordan R. (2013) “Algorithmic trading and the Market for Liquidity”,
Journal of Financial and Quantitative Analysis 48, 1001-1024.

Hu, Edwin and Murphy, Dermot,2020. Vestigial Tails: Floor Brokers at the Close in Modern
Electronic Markets. Available at SSRN: https://ssrn.com/abstract=3600230.

Page 24 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

Korajczyk, R.A. and Murphy, D. (2019) High-frequency market making to large institutional
trades. The Review of Financial Studies, 32(3), pp.1034-1067.

Kauppalehti (2012) “Pörssi epäille väärinkäytöksiä robottikaupassa” 16 February.

Kirilenko, A., Kyle, A.S., Samadi, M. And Tuzun, T. (2017), The Flash Crash: High‐Frequency
Trading in an Electronic Market. The Journal of Finance, 72: 967-998.

Lin, Ji-Chai, Ajai K. Singh, Wen Yu. (2009) Stock splits, trading continuity, and the cost of equity
capital,Journal of Financial Economics, Volume 93, Issue 3, Pages 474-489.

Malinova K. and Park A (2015) “Subsidizing Liquidity: The Impact of Make/Take Fees on Market
Quality”, Journal of Finance, 2, 509-536.

Martens, Martin, and Dick van Dijk. (2007) Measuring volatility with the realized range, Journal
of Econometrics, Volume 138, Issue 1, 2007, Pages 181-207.

Osmekhin S. (2016) “Exchange fees as a cost of trading at NASDAQ OMX”, Working Paper

Van Kervel, V. and Menkveld, A.J. (2019) High‐frequency trading around large institutional
orders. The Journal of Finance, 74(3), pp.1091-1137.

Zhou, Hao, Petko S. Kalev, and Alex Frino, (2020) Algorithmic trading in turbulent markets,
Pacific-Basin Finance Journal, Volume 62, 2020, 101358.

Page 25 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

Table 1. Descriptive statistics.

The descriptive statistics for all major variables during our sample period from June 2010 to December 2011.
TOTVOL is the average daily trading volume in shares. TURN is the ratio of shares to the total shares outstanding,
MCAP is the market capitalization in billions of Euros. ACKS are partially executed orders, EXEC – the fully executed
orders, CANCELS are orders that were cancelled, and PRICE is the average daily price in euros. Panel B describes
the market quality variables. RTWAS is the relative Bid-Ask spread weighted by time, TVOBV and TVOAV are the
average number of shares at the Best and ALL levels of order book respectively. TVOBR and TVOAR are the same
variables multiplied by the price in euros. VLT is a volatility variable measures as the difference between High and
Low price during a day scaled by the Closing price. Panel C presents some variable characteristics separated by
trader type where ALGO represents all trades completed through proprietary algorithms, INST – all trades
submitted through standard and exchange-approved execution algorithms (VWAP, TVAP, etc.), PROF – all orders
that were submitted “manually”, and RETL – are all orders that come from the retail investors.

Panel A. Trading Variables

TOTVOL TURN MCAP, €B ACKS EXEC CANCELS PRICE
Min. 0 0.000% 0.02 1 1 1 2.19
1st Qu. 7,175.00 0.008% 0.71 431.00 133.00 338.00 6.92
Median 28,870.00 0.023% 1.64 1,873.00 440.00 1,622.00 11.54
Mean 149,361.00 0.044% 4.45 9,530.00 1,074.00 11,017.00 17.67
3rd Qu. 105,740.00 0.051% 4.42 6,310.00 1,293.00 5,745.00 20.05
Max. 42,481,766 26.380% 66.20 2,559,620 67,772 2,590,058 119.61
N 458,813 458,813 458,813 458,813 458,813 458,813 458,813

Panel B. Market Quality Variables

RTWAS TVOBV TVOBR TVOAV TVOAR VLT, %
Min. 0.03 31.70 890.00 337.00 6,030.00 0.28%
1st Qu. 0.11 812.70 11,256.00 21,997.00 264,582.00 1.72%
Median 0.17 1,868.90 19,156.00 51,819.00 614,936.00 2.47%
Mean 0.27 4,004.80 36,700.00 125,078.00 1,332,398.00 2.97%
3rd Qu. 0.37 3,997.20 40,990.00 128,732.00 1,608,029.00 3.61%
Max. 9.58 2,033,229 4,211,052 11,071,906 43,494,537 58.02%
N 458,813 458,813 458,813 458,813 458,813 458,813

Panel C. Variables by Trader type

TRADER TYPE TURN ACKS EXEC CANCELS MAKING CORR
1 ALGO 0.037% 30,030 1,184 37,045 42.0% 0.946
2 INST 0.029% 2,477 745 2,249 47.1% 0.829
3 PROF 0.068% 3,978 1,388 3,485 52.2% 0.979
4 RETL 0.040% 3,010 978 2,729 55.2% 0.858

Page 26 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

Table 2. Correlations.

The table presents Spearman correlations between the main market quality variables. TWAS is the
relative Bid-Ask spread weighted by time, TVOBV and TVOAV are the average number of shares
at the Best and ALL levels of order book respectively. VLT is a volatility variable measures as the
difference between High and Low price during a day scaled by the closing price. TURN is the
number of shares traded scaled by the number of shares outstanding. MAKING is the average
ratio of the trading volume that supplies liquidity to the total trading volume for the day.

RTWAS TVOBV TVOAV TURN MAKING VLT

RTWAS
1
TVOBV
0.0189 1
TVOAV
(0.1798) 0.5702 1
TURN
(0.0641) 0.0264 0.1207 1
MKTK
0.0184 0.0105 0.0522 0.0402 1
VLT
0.3202 (0.0137) (0.0010) 0.2221 0.0048 1
Note: All correlations are significant at 5% level.

Page 27 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

Table 3. The effect of four trader types on market quality measures.

The table presents the results of regression model (1). RTWAS is the relative Bid-Ask spread
weighted by time, TVOBV and TVOAV are the average number of shares at the Best and ALL
levels of order book respectively. VLT is a volatility variable measure as the difference between
High and Low price during a day scaled by the closing price.

RTWAS
term RTWAS TVOBV TVOAV
(Intercept) 6.87*** 0.0337*** 3.84***

TYPEALGO -0.0945*** 0.0572*** 0.192***

TYPEINST -0.0599*** 0.0599*** 0.177***

TYPEPROF 0.0527*** -0.055*** -0.149***

log(TURN) -0.127*** 0.13*** 0.355***

log(VLT) 0.2427*** -0.342*** -0.285***

log(PRICE) -0.0697*** -1.2062*** -1.0480***

log(MCAP) -0.3996*** 0.4862*** 0.5460***

Note: ***, **, and * denote the level of significance at 1%, 5%, and 10% respectively

Page 28 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

Table 4. The effect of volume interaction

The table presents the results of regression model (2) where we interact the trader type dummy
variable and the daily turnover, TURN defined as the ratio of number of shares traded to the total
number of shares outstanding. RTWAS is the relative Bid-Ask spread weighted by time, TVOBV and
TVOAV are the average number of shares at the Best and ALL levels of order book respectively.
VLT is a volatility variable measure as the difference between High and Low price during a day
scaled by the closing price.

term RTWAS TVOBV TVOAV

(Intercept) 6.837*** -0.00642 3.895***

TYPEALGO -0.071*** -0.652*** -1.250***

TYPEINST 0.009 -0.182*** -0.549***

TYPEPROF 0.095*** 0.204*** 0.203***

log(TURN) -0.131*** 0.153*** 0.413***

log(VLT) 0.243*** -0.342*** -0.285***

log(PRICE) -0.070*** -1.208*** -1.051***

log(MCAP) -0.399*** 0.497*** 0.567***

TYPEALGO*log(TURN) 0.003** -0.079*** -0.161***

TYPEINST*log(TURN) 0.008*** -0.028*** -0.083***

TYPEPROF*log(TURN) 0.005*** 0.0329*** 0.046***

Note: ***, **, and * denote the level of significance at 1%, 5%, and 10% respectively

Page 29 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

Table 5. The sensitivity of algorithmic traders to volatility

The table shows the effects of interacting the trader types with volatility from (3). TURN defined
as the ratio of number of shares traded to the total number of shares outstanding. RTWAS is
the relative Bid-Ask spread weighted by time, which is the results presented in this Table.
(TVOBV and TVOAV are the average number of shares at the Best and ALL levels of order book
respectively, whose results are available upone request). VLT is a volatility variable measure as
the difference between High and Low price during a day scaled by the closing price. Volatility
quintiles are calculated by sorting all stock/days based on VLT from Q1 (the “calmest”) to Q5
(the “stormiest”)

Panel A. Using continuous volatility measure. Panel B. Using Volatility Quintiles

estimate Estimate
(Intercept) 6.868*** (Intercept) 5.881***
TYPEALGO -0.111*** TYPEALGO -0.092***
TYPEINST -0.067*** TYPEINST -0.061***
TYPEPROF 0.101*** TYPEPROF 0.042***
log(VLT) 0.241*** Q2 0.096***
log(TURN) -0.127*** Q3 0.157***
log(PRICE) -0.070*** Q4 0.230***
log(MCAP) -0.400*** Q5 0.362***
TYPEALGO * log(VLT) -0.005 log(TURN) -0.124***
TYPEINST * log(VLT) -0.002 log(PRICE) -0.073***
TYPEPROF * log(VLT) 0.013*** log(MCAP) -0.401***
TYPEALGO * Q2 0.0029
TYPEINST * Q2 0.0059
TYPEPROF * Q2 0.0048
TYPEALGO * Q3 0.0017
TYPEINST * Q3 0.0039
TYPEPROF * Q3 0.0083
TYPEALGO * Q4 0.0013
TYPEINST * Q4 0.0028
TYPEPROF * Q4 0.0123*
TYPEALGO * Q5 -0.0055
TYPEINST * Q5 -0.0006
TYPEPROF * Q5 0.0204***

Note: ***, **, and * denote the level of significance at 1%, 5%, and 10% respectively

Page 30 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

Table 6. The effect of market making activities on market quality variables.

The table presents the results of (4) where in addition to the trader type dummy variable we also add
MAKE dummy variable. MAKE is an indicator variable (supplied by the NASDAQ OMX) that
signifies whether the trade was adding (MAKE =1) or removing (MAKE =0) liquidity. Daily
turnover, TURN defined as the ratio of number of shares traded to the total number of shares
outstanding. RTWAS is the relative Bid-Ask spread weighted by time, TVOBV and TVOAV are the
average number of shares at the Best and ALL levels of order book, respectively. VLT is a
volatility variable measure as the difference between High and Low price during a day scaled by
the closing price.

RTWAS RTWAS TVOBV TVOAV

term (a) (b) (c) (d)
(Intercept) 6.877*** 6.855*** 0.052** 3.884***

TYPEALGO -0.094*** -0.058*** 0.022*** 0.099***

TYPEINST -0.059*** -0.039*** 0.033*** 0.093***

TYPEPROF 0.052*** 0.063*** -0.065*** -0.179***

MAKE -0.004*** 0.030*** -0.028*** -0.072***

log(TURN) -0.127*** -0.128*** -0.342*** -0.286***

log(VLT) 0.242*** 0.243*** 0.131*** 0.356***

log(PRICE) -0.069*** -0.070*** -1.206*** -1.048***

log(MCAP) -0.399*** -0.400*** 0.486*** 0.546***

TYPEALGO:MAKE -0.074*** 0.071*** 0.188***

TYPEINST:MAKE -0.043*** 0.056*** 0.172***

TYPEPROF:MAKE -0.020*** 0.020*** 0.059***

Note: ***, **, and * denote the level of significance at 1%, 5%, and 10% respectively

Page 31 of 35

Electronic copy available at: https://ssrn.com/abstract=3673881

Table 7. Determinants of market making activity and order cancellation.

The table presents the results of beta regression (6) and where the dependent variables are changed to
MAKE_FRAC and CANCELLED_FRAC. MAKE_FRAC is a ratio of liquidity supplying volume to the
total volume for each stock/day (as presented in equation (5)) and CANCELLED_FRAC is a ratio
of cancelled orders to total orders for each day/stock. Both equations are estimated with beta
regression. Daily turnover, TURN defined as the ratio of number of shares traded to the total number
of shares outstanding. VLT is a volatility variable measure as the difference between High and
Low price during a day scaled by the closing price. MAKE is an indicator variable (supplied by
the NASDAQ OMX) that signifies whether the trade was adding (MAKE =1) or removing (MAKE
=0) liquidity. RTWAS is the relative Bid-Ask spread weighted by time, TVOBV and TVOAV are the
average number of shares at the Best and ALL levels of order book, respectively.

Panel A. Market Making Determinants Panel B. Cancelled Orders Determinants

term estimate Term estimate

(Intercept) -0.316*** (Intercept) -1.499***
TYPEALGO -0.460*** TYPEALGO 0.459***
TYPEINST -0.259*** TYPEINST -0.077***
TYPEPROF -0.167*** TYPEPROF -0.139***
log(TURN) 0.087*** log(TURN) 0.020***
log(VLT) -0.119*** log(VLT) -0.050***
log(PRICE) 0.106*** log(PRICE) -0.020***
log(MCAP) -0.026*** log(MCAP) 0.053***
log(RTWAS) 0.225*** log(RTWAS) 0.042***
log(TVOBV) -0.108*** log(TVOBV) -0.078***
log(TVOAV) 0.210*** log(TVOAV) 0.051***

Note: ***, **, and * denote the level of significance at 1%, 5%, and 10% respectively

Page 32 of 35
Electronic copy available at: https://ssrn.com/abstract=3673881
FIGURE 1

The distribution of total trading volume by trader types during the sample period from 06/2010
to 12/2011. ALGO stands for the fraction of volume generated by the proprietary trading
algorithms. INST – the fraction of trading volume generated by traditional execution algorithms
approved by the exchange (VWAP, TWAP, etc.). PROF – stands for the fraction of total volume
generated by discretionary “manual” trading originating from trading firms and institutions.
And RETL is the fraction of the trading volume originated from the retail brokerage accounts.

Page 33 of 35
Electronic copy available at: https://ssrn.com/abstract=3673881
FIGURE 2

The fraction of algorithmic trading volume from 06/2010 to 12/2011. The smoothed line is
smoothed conditional mean estimated with Loess function.

Page 34 of 35
Electronic copy available at: https://ssrn.com/abstract=3673881
FIGURE 3

The distribution of liquidity “making” vs “taking” TURN for the four trader types. We can see
that while the INST, PROF, and RETL types exhibit normal-like distribution, the ALGO “making”
turnover is highly skewed. Thus, ALGO appear to “spend” a lot less time in the liquidity
“making” mode with occasional bursts of limit orders illustrating the occasional dominance of
opportunistic behavior in this group of traders.

Page 35 of 35
Electronic copy available at: https://ssrn.com/abstract=3673881