Liquidity Provision and Market Making by HFTs∗

Report prepared for the Investment Industry Regulatory Organization of Canada

Katya Malinova† Andreas Park‡

September 8, 2015
Executive Summary

1. We classify high-frequency traders based on reaction speeds, buy-side institutions

based on inventory and use of (off-exchange arranged) dealer-crosses, and retail
traders based on the usage of seek-dark-liquidity orders. Based on similarities in
trading characteristics, we group HFT IDs with a cluster analysis. We identify
(voluntary) market-makers based on the extent to which the buy and sell volume of
their passive order submissions is balanced.
2. In our sample, HFTs account for 26% of active trades and 53% of passive trades,
buy-side institutions account for 48% of active trades and 39% of passive trades.
3. HFTs account for around 80% of all orders, around 65% of all orders that improve
the best price and 70-80% of all orders that match the prevailing best prices.
4. The order cancellation rate of existing orders by (HFT) market makers is higher in
the first 10-50 milliseconds after a trade compared to half or a full second after the
trade, which documents the quote-fade phenomenon. Similarly, in the first 10ms
after a trade, there is an increase in aggressive order submissions that are in the
same direction as the trade. This relation is positively related to the volume of the
order and the directional volume that the trader has already traded on the day.
5. We find (weak) evidence that there is transmission of activity between fundamentally
linked securities (e.g., stocks to preferred shares, or between stocks with different
voting rights). We observe that trades in the frequently traded security have price
impact and affect the order cancellation behavior by HFT market makers in the
less-frequently traded security.
6. For TSXV securities, HFT participation is lower, and HFTs trade in only about half
of all TSXV securities. Retail trading participation, on the other hand, is higher in
TSXV than in TSX securities, and it is higher than HFT participation.

∗
We thank the Investment Industry Regulatory Organization of Canada (IIROC) and in particular
Victoria Pinnington and Helen Hogarth.
†
University of Toronto, [email protected].
‡
University of Toronto, [email protected] (corresponding author).
I. Introduction and Summary of Findings

In this report we describe and analyze the market-making behaviour of high frequency
traders. We describe how high-frequency market makers submit quotes relative to posted
prices, and we analyze how market makers react to trades by canceling their existing
quotes and by eliminating others’ stale quotes with aggressive, marketable orders. We are
particularly interested in understanding whether and how the cancellation and aggressive
order submission behaviour relates to the directional volume of the trader who submitted
the trade. Finally, we analyze whether there is transmission of activities among related
securities, such as between stocks with different voting rights.
We classify traders into several groups, using trading in a comprehensive sample
of 307 highly liquid securities from January to May 2013. We are interested in retail,
buy-side, and high-frequency traders. Our classification approach builds on Comerton-
Forde, Malinova, and Park (2015): we classify a trader as retail based on the usage of
an order type that only retail traders are allowed to used, buy-side based on the signed
dollar-volume that the trader accumulates over a number of days, and high frequency
based on the traders’ ability to react to market events quickly. We additionally classify
any trader as a (voluntary) market maker if this trader regularly posts similar passive
volume on both sides of the market on many days and across many securities.
We first study the distributional features of submitted orders relative to the best
prices on the marketplace where the orders are posted. The Canadian marketplaces are
separately but anonymously identified in the data, and we label them alphabetically. We
observe that about 52% of all orders are submitted at the local best prices, 7% strictly
improve the best prices, 37% are worse than the local best prices, and 4% are marketable.
There are differences across markets: on market B, which is the largest by value traded
and which also has the lowest quoted spread, only 3-4% of all orders improve the local
best bid and offer prices; on this venue, the quoted spread is regularly constrained by
the 1 cent minimum tick size. For market A, almost 10%, and for market C, slightly
over 10% of orders improve the price. A notable fraction of orders are submitted as
immediate-or-cancel (IOC) orders. Many of these orders are not marketable, but since
their intent is to seek immediate execution, we classify such orders as aggressive in our
subsequent analysis.
We then study the order submission behaviour of high-frequency traders and market

1
makers. HFTs submit the bulk of all passive orders (just under 80%), and they account
for about 65% of all price-improving orders. There are, however, differences between
the marketplaces. HFTs are most active on markets B and C, at the time the largest
and the third largest market by value traded. On market A, they account for only 50%
of price improving orders in early 2013, but after a major technological change to this
market in April 2013, HFTs account for around 70% of all price improving orders. We
further observed an offsetting drop in the activities of market-makers that we did not
classify as HFT. The details of the trader-level data suggest that some of the non-HFT
market makers on that venue were likely part of an HFT strategy that our classification
approach could not detect. In the subsequent by-trade analysis, we thus study all market
makers as a group.
We next focus on HFT behaviour subsequent to trades. Theoretical models of market
making predict1 that market makers adjust their quotes subsequent to trades, as (a) these
trades may reveal information about fundamentals and (b) taking an inventory exposes
the market maker to risk. One would thus expect that market makers cancel existing
quotes, and we study the cancellation behavior after trades. We also study whether
market making traders submit aggressive orders, e.g., to “take out” other traders’ stale
quotes. In our analysis, we further ask, whether the cancellation and aggressive order
submission rates depend on the size of the trade, on the type of trade (smart-order-
routed trades vs. “normal” trades), on the type of liquidity provider (market maker vs.
non market-maker), and on how much the entity that submitted the marketable order
has already traded during the day.
We perform this part of the analysis for a smaller subset of securities and focus on
highly liquid, non-interlisted2 securities that are in the TSX60 index, and we focus on
two measures: cancellations in the opposite direction of trades (e.g., cancellations of
sell-orders following a buy)3 and aggressive order submissions in the same direction of
1
See, for instance, Kyle (1985), Glosten and Milgrom (1985), Glosten (1994).
2
In Canada, cross-listed securities are commonly referred to as “’interlisted” securities.
3
Opposite-side cancellations can be interpreted as being indicative of “stepping aside” of possible
future trades from the same trader. Theoretically, however, when moving a quote, a market maker
would cancel orders on both sides of the market resubmit orders on both sides at different prices.
Indeed, we ran components of our analysis also for same-side cancellations and found that these orders
are cancelled, too. This part of the analysis is both computationally intensive and difficult to present
concisely when discussing both same-side and opposite-side cancellations. To simplify the exposition,
we focus only on opposite-side cancellations.

2
the trade (e.g., marketable or IOC buy orders following a buy). The cancellation rate
beyond the first 10ms decays rapidly. To the naked eye, which can at best observe market
movements after 250ms, the quotes would appear to have been cancelled concurrently
with the trade.
A formal regression analysis confirms the observations from the aggregate data: In
this part of the analysis, we additionally address two questions. First, we ask whether
the cancellation and aggressive order submission rates of market makers differ between
“normal” trades and trades that were submitted via a smart-order router (SOR), where
we classify a trade as stemming from a SOR if the order is from the same trader and hits
multiple markets at around the same time. Second, we ask whether the cancellation and
aggressive order submission rates are affected if the marketable order executes against a
market maker’s passive order. We find that trades generally have a positive, significant
effect on cancellation and aggressive order submission rates. SOR trades generally have
a larger impact except in the very first milliseconds after the trade; similarly for trades
that have a market maker on the passive side.
As a next step, we analyze whether a trader who has already traded a lot in the
same direction can be detected and thus triggers a stronger reaction by market makers.
To address this question, we estimate the effect of the market order submitter’s net
inventory on the cancellation and order submission rates.4 We perform this part of
the analysis by aggregating observations by the both trade-volume and net-inventory
deciles for the security and type of trader. Here, too, we observe that the effect on
the cancellation rate is strongest in the first 10-50ms after the trade and that there
is an increasing relation between the traders’ net inventories and the market makers’
cancellation rates and submission rates of aggressive orders.5
Overall, our analysis indicates that after, say, trading with a buyer, market-makers
cancel their sell orders quickly and submit aggressive buy orders. This latter behavior
can be interpreted as market makers either trading in anticipation of future orders or
taking advantage of and eliminating mis-priced, stale quotes. The more buys the buying
trader had already submitted, the stronger this effect.
4
We control for market-wide data for the security such as daily return, volume traded, and aggregate
net inventory to ensure that our results are not driven by general market developments.
5
The only exception is the case when the incoming trade was submitted by a non-market-making
HFT. For these traders, there is a non-negative aggressive order submission rate, but this rate does not
depend on this type of traders’ net inventory.

3
 We further explore (HFT) market-making across different securities. Specifically, we
analyze how trading in a frequently traded security (e.g., heavily traded BBD.B shares)
affects trading and quoting behaviour in related, but less-frequently traded securities
(e.g., more lightly traded BBD.A shares). For this part of the analysis, we look at
three classes of non-interlisted securities: dual-class shares, preferred shares, and deben-
tures. For each of these we observe that trades in the more frequently-traded securities
commonly affect trading behaviour in the related security in the sense of positive price
impacts and positive effects on market makers’ order cancellation rates.
Finally, we apply some components of our analysis to TSX Venture-listed securities.
We use the trader classification from our main analysis for TSX Venture-listed securi-
ties to take the advantage of the information from behaviour in much more frequently
traded securities where speed may be more important and where we may thus have more
opportunities to observe behaviour that reveals the ability to react quickly.6 Overall,
we observe trading in about 2,500 TSXV listed securities, though only about 1,300 of
these exhibit HFT trading. Of all transactions, around 20% have HFTs on the passive
side; aggressive trades by HFTs are negligible. Yet, HFTs still account for a very large
fraction of the price improving and BBO-matching orders (45% and 50%, respectively).
TSXV-securities have a much higher fraction of retail trader activities than TSX-listed
securities (20+% of all active trades and 10% of passive trades), and the shares of retail
exceeds the HFT fractions.
The remainder of this report is organized as follows: Section II. describes the in-
stitutional background and the organization of the different marketplaces. Section III.
describes the data and the sample. Section IV. outlines our classification approach,
Section V. presents our findings on the behaviour of the classified traders. Section VI.
analyzes the reaction of market-making traders to trades, Section VII. discusses how
market-making HFT help transmit information from trades across related securities.
Statistics concerning HFT trading in TSXV securities are in Section VIII.
6
The downside is that we may be missing behaviour by trader IDs that only trade in TSXV securities.

4
II. The Institutional Setting

A. Core rules governing trading in Canada

The Toronto Stock Exchange (TSX) is the primary listing venue for large compa-
nies in Canada, small and mid-cap companies are typically listed on the TSX Venture
exchange. As in other major markets around the world, trading in TSX-listed stocks
is fragmented across multiple exchanges and Alternative Trading Systems (ATS), and
many TSX-listed companies are also listed on U.S. exchanges. Securities trading and
the activities of market participants in Canada are regulated by the Investment Industry
Regulatory Organization of Canada (IIROC), the members of the Canadian Securities
Administrators, and are governed by the Universal Market Integrity Rules (UMIR).
Most of the core elements of the UMIR are similar to those governing trading in
the U.S. equities markets. Brokers and marketplaces are required to respect the order
protection rule, which mandates that orders must be routed to the marketplace with
the best-priced orders available on lit markets. Brokers are also subject to obligations
regarding best execution for client orders.
In the context of our study, there are three critical differences between trading rules in
the U.S and Canada. First, the order protection rule in Canada applies to the whole-of-
book rather than the top-of-book as is the case in the US. Second, Canada also imposes
a strict version of an order exposure rule, with few exceptions. This rule requires that
client orders below a certain size be immediately sent to a marketplace that publicly
displays prices. This rule severely limits the practice of broker internalization, which oc-
curs when a broker trades against their customer’s order instead of sending the order to
a public marketplace, and the practice of selling retail orders to market makers. Third,
unlike the US, several Canadian marketplaces are allowed to offer broker-preferencing
on the market’s order book. This practice allows incoming orders to a marketplace
to match with other orders from the same broker-dealer ahead of similarly priced or-
ders from other broker-dealers, without regard to time priority. To take advantage of
broker-preferencing, brokers must elect to publicly display broker IDs when submitting
their orders.
A significant volume of trades is pre-arranged off-exchange, before entering orders
on a public marketplace. These trades must still be executed on a public marketplace,

5
respecting all the applicable rules. There are usually very few such deals each day, but
they are large and, on average, account for roughly 10-12% of trading value. We omit
such trades from most of our analysis.

B. Marketplaces and their trading rules

The data in our sample contains observations for ten marketplaces. These mar-
ketplaces are separately, but anonymously identified in our data, and we label them
alphabetically. During our sample period, marketplaces A, B, and C together account
for about 90% of the dollar volume traded, respectively; marketplace D, which is a dark
pool, accounts for about 5% of dollar volume. The remaining marketplaces account for
no more than 2% market share each.
Marketplace A operates a public limit order book and a dark pool facility. The lit
market of A allows lit and partially hidden (iceberg) limit orders. Broker-preferencing is
allowed provided the broker chooses to publicly display its broker ID when submitting
the order. In most of our analysis, we only consider the lit portion of this market.
Marketplace B is a lit market that operates as a public limit order book. Broker-
preferencing is allowed, provided the broker chooses to publicly display its broker ID
when submitting the order. Traders may post lit, partially hidden (iceberg), and fully
hidden orders, including “mid-point” orders, which are pegged to execute at the floating
midpoint of the NBBO. Marketplace B operates a so-called electronic liquidity provider
program (ELP) that offers favorable trading fees to traders who routinely provide liq-
uidity in a sufficiently large number of stocks.
Marketplace C is a lit market that operates as a public limit order book. Like
marketplace B, it allows lit, iceberg, and fully hidden orders, which may be pegged to
the midpoint. Marketplace C does not offer broker-preferencing.
Marketplace D is a dark pool that allows traders to interact using two types of
orders. First, traders may submit passive dark orders that remain in the dark pool until
they are executed or cancelled. Second, traders may submit aggressive, liquidity taking
orders that are either executed immediately against a passive dark order or cancelled.
All trades are priced at the midpoint of the NBBO.
Marketplaces E, F, and G operate as public limit order books. These markets are
very small and account for less than 2% of daily value each. Markets F and G operate

6
inverted, taker-maker pricing.
Marketplaces H and I are dark pools. During our sample period, marketplace H is
an institutional-only venue, marketplace I offers periodic matching with 1-second random
NBBO prices. These venues account for negligible shares of volume and we omit them
from our analysis.
Finally, marketplace J operates as a public limit order book and trades only secu-
rities that are listed on this venue, and trading on this marketplace thus does not feature
in our main analysis.
Towards the end of our sample period (end-May 2013), a new marketplace started
operating. The special feature of this new venue was the so-called “taker-maker” or
“inverted” pricing schedule under which the liquidity taking side of a transaction would
be paid a fee rebate and the liquidity making side would pay a fee. Trading in May in
this market was sparse and to avoid any confounding effects, we omit this venue from
the analysis.

III. Data and Sample

Data. The data for this study is provided by the Investment Industry Regulatory
Organization of Canada (IIROC). The dataset contains detailed records on all trades,
orders, order cancellations, order amendments, and updates to marketplaces best bid and
offer quotes from IIROC’s real-time surveillance system, for all trading on all regulated
Canadian marketplaces. Each order-related record includes, in particular:

• The marketplace where the order was sent (masked).

• Size, price, and the direction (buy or sell) of an order.

• Broker ID (masked), user ID (masked), and account type (e.g., specialist, client,
options-trader, or inventory).

• Other characteristics, including the duration of an order (for instance, good-

till-cancel or immediate-or-cancel), whether an order was transparent or non-
transparent, whether the order was a seek-dark-liquidity order, and a unique iden-
tifier for each order.

7
For trades, the data additionally specifies the aggressive and passive (liquidity-providing)
side of a trade. The data also identifies the aforementioned intentional broker-crosses,
which we omit from the analysis. The information for marketplaces, brokers and users is
masked in the sense that IIROC provides a scrambled identifier. The masking is applied
consistently so that the same marketplace, broker and user are always assigned the
same identifier. Marketplaces time-stamps are reported with millisecond precision for
our sample period. Brogaard, Hendershott, and Riordan (2014), Korajczyk and Murphy
(2014), Comerton-Forde, Malinova, and Park (2015) and Devani, Tayal, Anderson, Zhou,
Gomez, and Taylor (2014) contain further information of the data.
Large Sample. We base our analysis on the period from January 1 to May 31,
2013. We end the sample at the end of May for two reasons. First, subsequent to the
integration of Alpha into the TMX Quantum platform, a large number of high-activity
trader IDs disappeared. At the same time, several new high activity IDs appeared
(for the same brokers), and the time horizon of the data is too short to characterize
many of the new IDs. Second, IIROC’s public market share statistics illustrate that a
new marketplace rapidly gained market share. The entry of this marketplace has been
associated with changes in behavior that we might not be able to fully capture as our
data ends too early (end June 2013).
We base our analysis on the 307 securities that are classified as “highly-liquid” se-
curities by IIROC during the entire sample period. Loosely, a security qualifies as
highly-liquid for a given day if over a 60-day period it traded more than 100 times per
trading day and had an average trading value of at least $1M. IIROC compiles a list of
highly-liquid securities daily; we include a security in our sample if that security is on
the list of highly liquid securities at the end of each month in our sample period. We ap-
plied no further filters, in particular, there are no corrections for stock splits, corporate
actions, halts, etc.
We use the entire set of securities for the classification of traders and for the general
description of trading behavior. For our analysis of market-making behavior we restrict
attention to two subsamples (partly due to data processing constraints).
Sub-Sample 1: non-interlisted TSX60 Securities. We determine the inter-
listing status based on the June 2013 TSX e-review publication. Of the 307 securities
in our sample, 87 are interlisted, 222 are not interlisted. Of the latter 222 securities,

8
17 are S&P/TSX60 constituents, and these comprise our SubSample 1. We restrict
attention to non-interlisted securities because for these securities we know all the trades.
For interlisted securities, on the other hand, observed cancellations could be driven by
trades on U.S. markets. For this sub-sample, we further restrict the sampling horizon
to March and May 2013. It is part of the public record, that in April 2013, the TMX
Group integrated Alpha Exchange in its Quantum platform. It is not clear to us when
exactly TMX performed the move and whether all securities where affected at the same
time. We thus omit April from the sample.
Sub-Sample 2: “Fundamentally Linked” Securities. For our analysis of mar-
ket making in fundamentally-linked securities, we consider non-interlisted, highly liquid
securities that also have TSX listings of dual-class shares, preferred shares, or deben-
tures. These pairs of securities are linked by common fundamentals. Specifically, we
have 8 non-interlisted securities from our list that have dual class listings, 102 stock-
preferred share pairs, and 79 stock-debenture pairs.7 As for SumSample 1, we restrict
the sampling horizon to March and May 2013.
Outliers. We eliminated two days from our sample: January 21 (Martin Luther
King Day) and May 27 (Memorial Day); these days are public holidays in the U.S., and
trading activity on Canadian markets on such days is very low.

IV. Trader Classification

All traders access the marketplaces via brokers. We base our classification on the
analysis of order submission and trading behavior by trader IDs, where we define a
trader ID as the combination of broker ID plus user ID, plus the account type (client,
specialist, inventory, option market maker, and non-client).
The user ID is the most granular identification that is available to regulators in
Canada; IIROC researchers describe the usage of user IDs in detail in their research
reports (IIROC (2012), Devani, Tayal, Anderson, Zhou, Gomez, and Taylor (2014), and
Devani, Anderson, and Zhang (2015)). According to these research reports, market-
places assign user IDs, and an ID may identify a single trader, a business stream (for
example, all orders that originate through a broker’s online discount brokerage system),
7
The most obvious pair would, of course, be options and their underlying. However, we do not have
access to derivative market data.

9
or a client that accesses trading venues directly (through a direct market access (DMA)
relationship). It is our understanding that the brokers separate different types of order
flows (e.g., retail vs. institutional) by user ID. For DMA clients, IIROC requires dedi-
cated IDs. However, according to Devani, Tayal, Anderson, Zhou, Gomez, and Taylor
(2014), a DMA client may be assigned more than one user IDs, for instance, to trade
through multiple brokers or to trade on different marketplaces, and they may choose to
use multiple user IDs for business or administrative purposes.
We classify traders based on trading characteristics during our entire sample period
in all securities from the large sample.
We group traders into four general categories: HFT, retail, buy-side, and other. The
“other” category includes trader IDs that we were not able to classify as HFT, retail, or
buy-side. The group of HFT and “other” are then each segmented into market-making
and non-market-making traders. In this report, we do not present information about
the behavior of non-market-making “other” traders.
Our general approach is to provide a classification for each trader based on the
trader’s behaviour in a large number of securities. Thus, for instance, to classify as
a market maker, a trader needs to exhibit market-making behaviour for a range of
securities. In our opinion, algorithmic strategies are scalable in these sense that they
can be applied to many securities and it is our intention to capture core characteristics
of an algorithmic strategy. We acknowledge that this approach will likely exclude firms
that employ security-specific, heterogenous strategies.
Table I provides summary statistics on these trader ID groups. We have a total of
4,892 trader IDs in our classification sample. Figure 1 shows the distribution of trader
categories and we discuss the details of our categorization in what follows.
Retail. On marketplace A, seek dark liquidity (SDL) orders are exclusively available
to retail investors. The use of SDL orders is the choice of the broker, not the customer.
We extract all trader IDs that use SDL orders from the complete database (which spans
January 1, 2012 to June 30, 2013 and all securities that trade on Canadian equity
markets), and we classify these traders IDs as retail. In our sample, we observe 134 of
these IDs. We know with certainty that these trader IDs are used to trade order flow
from retail investors, but there may be other trader IDs that are assigned for order flow
from retail investors that are not captured by our classification.

10
 Buy-Side Institutional: We conjecture that buy-side institutions will be involved
in large pre-arranged trades and accumulate large inventory positions. We therefore use
two criteria to identify buy-side institutions.
First, we extract all trader IDs that involve a client account and that are involved
in a so-called ”intentional cross”. An intentional cross is a trade, usually a large one,
that is pre-arranged off-exchange by a brokerage, for instance to match two client orders
or take an inventory from a client via a liability desk. These trades typically involve
human intervention.
Second, we search for trader IDs that accumulate very large inventory positions in
non-interlisted securities. We determine each trader’s maximum cumulative position for
the sample period, assigning a zero inventory at the beginning of the period. We focus
on non-interlisted securities to reduce the possibility that a seemingly large inventory
position of an entity is offset by an equally large position elsewhere. Since a trader
may buy in one jurisdiction and sell in another, for instance, to exploit an arbitrage
opportunity, it is imaginable that a Canada-only position is off-set by a U.S.-based
position. We acknowledge that this classification is imperfect, for instance, because
a trader ID may trade on behalf of multiple retail clients who jointly accumulate a
large position or because a DMA client may use different trader IDs for buying and
selling securities. To mitigate these imperfections, we set a high bar for the required
cumulative position. We classify the trader ID as a buy-side institution if its maximum
cumulative position during the classification period exceeds $100,000,000 in absolute
value for one security during our entire sample period. We classify 778 trader IDs as
buy-side institutions.
High Frequency: The critical component of high frequency trading is that trading
is automated and that traders have the ability to react quickly to market conditions.
Definitions used by various regulators or policy institutions (e.g., BAFin in Germany,
the European Commission, or the S.E.C.) often include a further requirement that HFTs
use many orders, in particular in relation to their trades. In our opinion, using orders or
order-to-trade ratios biases the classification against traders or strategies that use only
marketable orders.
We focus on reaction speeds as the main metric to identify HFTs, and we use reaction
times that are faster than human reaction times (the average duration of a single blink

11
of a human eye is 100-400 milliseconds, according to the Harvard Database of Useful
Biological Numbers). We further require that trader IDs exhibit fast reaction times
across many orders and trades, and in many securities. We use three criteria to quantify
a trader ID’s reaction speed.
Our first criterion is the trader ID’s median order-to-cancel time. The order-to-cancel
time is the time from the submission to cancellation of the same order; for the purpose
of this classification, we exclude immediate-or-cancel (IOC) orders, because their order-
to-cancel time is determined by the processing speed of the marketplace.
Our second criterion is the number of trade and order messages that a trader ID
submits during a short interval after a daily scheduled public information release. We
focus on the first 500 milliseconds after 3:40 p.m., which is when the TSX first publishes
the imbalance between the buy and sell orders in its market-on-close facility.
The closing price for TSX-listed securities is determined in a multi-stage process.
Before 3:40 p.m., traders may submit market buy and sell orders tagged as market-
on-close orders. These orders will trade at the 4:00 p.m. closing price. At 3:40 p.m.,
the TSX publishes the imbalances of buy and sell orders, and traders then have the
opportunity to submit priced limit orders to trade at the market-on-close to off-set the
market order imbalance. The market-on-close imbalance is indicative of the closing price
and may help predict behavior over the last 20 minutes of trading.8
In aggregate, there is a significant spike in trades immediately after the publication
of the market-on-close imbalance, though this spike may not be visible or pronounced on
a stock-by-stock basis. Comerton-Forde, Malinova, and Park (2015) includes a plot of
the by-minute number of trades (Figure 1 in their paper), aggregated over all securities
in their sample (which is similar to ours, albeit for a different time horizon). The dataset
that is provided to us by IIROC does not contain information on the market-on-close
announcement. Thus, we are not able to determine the time between the publication
of the market-on-close imbalance and a trader’s action at the millisecond level. For this
reason, we classify trader IDs as HFTs based on their actions during a relatively long
interval of 500 milliseconds after the announcement.
Our third criterion is the fraction of orders and (non-IOC) cancellations that a trader
submits very quickly after a change in the order book that was not triggered by the trader
8
The publication at 3:40 p.m. is merely the first publication. Between 3:40 and 4:00 p.m., TMX
regularly publishes updates of the prevailing imbalance.

12
him-/herself but by another trader.
For each trader ID, stock and day we compute the median order-to-cancel speed, and
we compute the total number of orders and aggressive trades during the 500 milliseconds
after 3:40 p.m. Furthermore, for each trader ID and security, we computed the number
of orders and cancellations that this trader submits and that this trader submits within
1 millisecond following another visible order submission by a trader other than him-
/herself in that security. A trader ID is classified as HFT

1. if the median of the trader ID’s median stock-day order-to-cancel speeds is below
250 milliseconds, or

2. if the trader ID submits more than 1,000 orders or is involved in more than 500
aggressive transactions in the first 500 milliseconds after the market-on-close pub-
lication across all securities in our classification sample during our classification
period, or

3. if the trader ID submits more than 85% of its orders and cancellations within 1
millisecond of some other trader’s order submission.

We classify a total of 101 trader IDs as HFT, though in each month, there are only
around 80 active HFT IDs.
In Devani, Tayal, Anderson, Zhou, Gomez, and Taylor (2014), IIROC researchers
discuss their classification of HFTs. In their data, they had direct information about
a subset of the existing HFT IDs in the Canadian market (49 IDs) and they used the
knowledge of these IDs to apply machine learning techniques to identify further IDs as
HFT. Their report does not list the details of the criteria, but Figure 1 in their report
shows that “speed” in various forms is a decisive criterion. For their sample period from
March to June 2013, which overlaps with ours, they identify 98 IDs as HFT, which is
close in number to our 101 HFT IDs.
Cluster Analysis of HFT Groups. The by-trader data displays some pronounced
similarities among subsets of traders in the sense that traders have very similar charac-
teristics for instance, in terms of numbers of trades and orders, or number of securities
traded. As IIROC researchers Devani, Tayal, Anderson, Zhou, Gomez, and Taylor (2014)
highlight, HFT firms may use several trader IDs for their strategies. For instance, it is

13
possible that an HFT firm uses one trader ID to submit buy orders and another to sub-
mit sell orders. Taking together, these two IDs may have a perfectly balanced end-of-day
inventory, whereas individually their inventory is imbalanced.
The usage of multiple IDs is particularly important and presents a challenge for an
analysis of market making behavior, in particular with respect to inventory positions and
order submission behavior. We thus group together trader IDs using a cluster analysis
approach to detect similarities in behaviour. Specifically, we use the following criteria:
the average per day number of securities traded, the average per day per stock number
of trades and orders, the average daily order imbalance and trade imbalances per stock,
the median order-to-cancel time, the total number of orders and trades submitted in the
500ms after 3:40 p.m., and the average percentage of orders and cancels submitted within
1ms of another trader ID’s order submission. For these nine criteria, critk , k = 1, . . . , 9,
we then compute the pair-wise absolute-value distance for traders i and j as follows9

9
X |critki − critkj |
Distij = . (1)
k=1
critki + critkj

We use visual analysis of the pair-wise distances in an Excel table to identify the clusters.
Figure 2 displays the pairwise pair-wise distances based on these criteria, using color-
coding to highlight the pairwise distances, where darker colors indicate smaller pair-wise
distances. The color-coding in the figure shows that there are groups of securities that
have small pair-wise distances. Using a maximum pairwise distance of threshold of .25,
we identify four clusters of IDs. Notably, members within a group all have the same
underlying broker (but cluster groups have different brokers).
Market-Making Index. A defining feature of a market maker is that the trader
posts on both sides of the market and stands ready to trade. We thus expect that a
market-making trader would submit passive buy and sell orders on both sides of the
market so as to earn the bid-ask spread (and maker-taker fees) on as many trades as
possible. To account for strategies that are linked across multiple IDs, we treat each
HFT trader ID-cluster as a single entity. We then compute for each trader-cluster, day,
9
As a convention, when both criteria are 0 for some k for i and j, then we set the distance to 0.

14
and security, the market maker index defined as
 
 passive buy order volume − passive sell order volume 
market maker index = 
 . (2)
passive buy order volume + passive sell order volume 

By construction, this index is between 0 and 1, where an index close to 0 indicates that
the trader’s order submission behaviour is well-balanced. A trader’s (or trader group’s)
market making index is the median index over all days and securities.10
Upon visually examining the data, it is apparent that for HFT IDs (including HFT
groups), there is a structural break for an imbalance score of 0.2 for our sample. We
thus classify a trader ID (cluster) as market-making if the trader ID (cluster) has an
imbalance score below 0.2. We identify 50 HFTs as market makers. In addition, we
identify 88 non-HFT IDs as (non-HFT) market makers.11
Inventories of Individual HFT Trader IDs. A common perception is that high-
frequency trading firms aim to hold no or only very small overnight inventories. We
observe that most trader IDs that we classify as HFT hold substantial median end-of-
day inventories, even in non-interlisted securities. Furthermore, several of the fastest
trader IDs that we classify as HFT trade more than 85% passive, have order-to-trade
ratios in the 99th percentile, and yet hold median inventories of 70% or more of their
daily trading volume.
This observation highlights the importance of understanding the usage of trader and
user IDs in different jurisdictions and in different datasets. In Canadian markets, a
single DMA client may use multiple trader IDs (IIROC (2012) and IIROC (2014)), and
it is thus possible that an HFT firm is assigned multiple user IDs. Furthermore, a single
user ID may be used for trading activity of multiple entities, for instance, for all the
brokerage’s retail order flow (which is balanced, on average). As a consequence, low end-
10
For traders with low median scores, the averages are similar to the medians. Many traders frequently
post perfectly unbalanced scores. By using the median, we can ensure that this high frequency of
unbalanced submissions is properly reflected in the score.
11
As discussed in the introduction, it is possible that some of these latter market makers are, in HFTs,
and that we were just not able to identify them based on our speed criteria. Indeed, at the beginning
of May 2015, a number of IDs with similar trading characteristics retired from posting at market A
while at the same time, some IDs in market A showed a significant increase in activity to the point
that the number of orders from the disappeared IDs almost coincides with the increase in the number
of orders for the HFT IDs. It is thus imaginable that these non-HFT IDs were part of a market making
strategy that used different IDs on different markets and that were consolidated when markets A’s and
B’s system effectively merged.

15
of-day inventories are neither a necessary nor a sufficient attribute of an HFT trader ID
in our dataset.
HFT vs. designated market making. One question that often arises in debates
in industry and that may be relevant when interpreting our results is whether designated
market makers, such as stock specialists (sometimes referred to as odd-lot traders, iden-
tifiable through the usage of a specialist (ST) account), options market makers (OT
account), or ETF market makers are classified as HFT. These traders interact with the
market in a high-speed environment and may thus exhibit behavior that looks similar
to that of an HFT (in terms of reaction speed, or, using the S.E.C. or E.C. definitions,
uses many orders and has a high order-to-trade ratio). In contrast to an HFT, however,
they have an obligation to make a market under all circumstances.
In our classification approach, we included no special provision to exclude such des-
ignated market makers. None of the 101 traders that we classify as HFT are ST or
OT accounts. Of the classified HFT, four are non-client (NC) accounts, twenty are
broker-inventory (IN) accounts, and the remainder are client (CL) accounts. NC traders
account for .6% of HFT trades and .8% of HFT orders and their activities are thus
unlikely to affect our results. It is not fully clear to us under which circumstances a
trader uses the IN vs. the ST account. One reasonable possibility is that some of the
IN accounts are linked to, for instance, ETF or options market makers in the sense that
such market makers use the ST/OT account in the assigned security and hedge using
the IN account. Three of the twenty HFT IN accounts are “linked” to ST accounts in
the sense that for the same broker and userid, there are activities under an ST and an IN
account. These three linked ST accounts trade in ETFs. Two of these three combined,
however, account for less than 0.04% of HFT trades and less than 0.003% of orders. Even
if these two are ETF market makers, their activities are unlikely to affect our results
significantly. The remaining account is additionally linked to an OT account and trades
in a large number of securities using the ST account identifier, including infrequently
traded ones.12 This behavior suggests that this account may bundle a large number of
electronic-trading activities (options, ETF and regular market making). We further note
that the ST account of this userid is not classified as HFT or market-making, and thus
the activities of this trader in its direct market making role would not affect our results.
12
We believe that Korajczyk and Murphy (2014) study the impact of this trader becoming a DMM
in November 2012.

16
V. Characteristics of Traders

In this section, we describe characteristics of the trader groups for our classification
sample of 307 securities.

A. Trades and Orders.

Our HFT group displays the characteristics that are typically associated with high
frequency traders. Figure 3 plots the time series of the types’ shares of orders. As
a group, HFTs account for over 80% of all passive orders. Figure 3 plots the overall
share of active and passive value traded (Panel A) and transactions (Panel B); this
figure is based upon the sum of all trading in the 307 sample securities over the sample
horizon from January to May 2013. HFTs account for 25% of all aggressive and 55%
of all passive transactions. For value traded, HFTs account for a similar percentage of
aggressive trades, but for a lower percentage of passive volume (41%). Buy-side traders
account for 45-50% of all aggressive trades, and 35% of passive trades. Retail traders
represent around 12% of aggressive volume.
Non-HFT market makers account for around 4% and 8% of active and passive volume
respectively. On an average day, the average HFT trades in 111 securities.
Panel A in Figure 4 indicates that there is an uptick in HFT order submissions
starting at the end of April 2013, coinciding with a drop in non-HFT market maker
submissions. We are able to identify that this change is driven by changes in order
submission behaviour on marketplace A. At the time, this marketplace’s technology was
integrated with that of marketplace B. Examining the data in detail we observed that
several non-HFT market makers disappeared from market A. At the same time, some
of the existing market-making HFTs increased their activity to the point where the
April order volume of the disappearing IDs coincided with the increase in volume for
the existing HFTs. It thus appears the the group of the market-making non-HFT IDs
may, in fact, be part of a cluster of market-making HFTs. We thus have to admit the
possibility that we were unable to classify some HFT IDs as HFT. As such, in our formal
regression analysis we will combine HFT and non-HFT market-making activities.

17
B. Order Competitiveness.

We conjecture that a market maker would post quotes competitively in the sense that
they match or improve upon the best available prices. To assess the competitiveness of
quoting behaviour we analyze, for each trader ID cluster, where in the limit order book
traders place passive orders relative to the prevailing prices. We determine the best
prevailing local bid and ask prices for each order and we differentiate between marketable
orders (the buy-price matches or exceeds the ask or the sell-price matches or is below
the bid-price), immediate-or-cancel, and passive orders. Immediate-or-cancel orders can
be non-marketable, but they are not passive because these orders do not enter the book
and they are submitted with the aim of an immediate execution. These orders thus do
not provide liquidity to subsequent traders, and we consider them to be aggressive in
nature. For the best prices we use the local price for the venue where the order is posted.
We then compute, for each trader ID cluster, how many orders and how much order
volume the trader submits at the best price, one tick higher/lower, 2 ticks higher/lower,. . . ,
and 5 or more ticks higher/lower. In the data, a fraction of orders is unpriced; one ex-
ample are dark midpoint orders which are pegged to the prevailing midpoint. We omit
these orders from the data.
Panel B in Figure 4 shows the distribution of orders, where we group them into four
categories to simply the exposition: better-than-best, at-the-quote, worse-than-best,
and marketable. As the figure highlights, fewer than 10% of all orders are submitted
at prices that improve the best prices available at that venue. One explanation for this
small fraction is that bid-ask spreads in frequently traded securities are narrow, often
at one cent.
Panel C in Figure 4 plots the time series of fractions of price-improving orders by HFT
for the main markets A, B, and C. We observe that for markets B and C, HFTs comprise
of around 65% of all price-improving orders, whereas for market A this fraction is much
lower until May 2013, when the HFT fraction increased to around 70%, for reasons that
we discuss above. Panel D of Figure 4 shows the time series for the fraction of “at-best”
orders by HFT. The pattern is similar to Panel A, except that on market C, the fraction
of “at-best” submissions by HFTs is between 75-80%.
Finally, considering the composition of a trader type’s order submissions, we observe
that HFTs that are also market makers submit around 11.5% of their orders improving

18
the best price and 58% at the best price. Notably, buy-side institutions, submit a larger
fraction of their orders improving the price, around 21%, and they submit a further 48%
of their orders at the best price; this behavior suggests that buy-side traders want to get
trades done and thus use relatively more aggressively priced orders.

VI. Reactions to Trades

A. Aggregate Observations

Practitioners often describe the phenomenon of the so-called quote-fade, a situation

when following a trade, available liquidity at the best prices disappears market-wide.
Models of asymmetric information, such as Glosten and Milgrom (1985) and Glosten
(1994), will predict that a market maker would adjust quotes after observing trades,
usually by adjusting posted prices upward after buys and downward after sells. The
reason for the adjustment is that trades, on average, convey some information about
the fundamental value. For instance, a sell (at least on average) reveals that the seller
believes the stock to be overvalued and thus, upon observing a sell, the market maker
adjusts the price downward. And alternative view is that as a market maker takes an
inventory, e.g., when buying from a seller, then he accepts a liability. If the stock were
to fall in price, he would lose. Assuming that the market maker is risk averse, he would
accept further inventories only at lower prices. Both information and risk-aversion would
thus imply that market makers adjust quotes on the opposite side of the market relative
to the trade, i.e., they should adjust the bid following a sell and the ask following a
buy.13 If market makers quote on multiple markets, then they will adjust these quotes
on multiple venues. As a first step, we thus ask if this immediate adjustment is visible
in the data.
Using SubSample 1, we compute, for every trade, the volume cancelled by market
makers during the 2, 5, 10, 50, 100, 500, and 1,000 milliseconds subsequent to a trade and
in the opposite direction of the trade (e.g., cancellations of passive buy-orders following
a marketable sell order), where we distinguish between trades that were initiated by
13
One can further make a case that the market maker should also adjust the quote on the same side
of the market, where the latter argument usually relies on a competition-based argument. We focus on
the opposite side here to simplify the exposition.

19
buy-side, retail and non-market making HF traders. We then compute the cancellation
rate x milliseconds after the trade as follows: for buyer-initiated trades, we compute
the total volume of cancelled sell-orders within x milliseconds of the trade by market
makers. We scale this number by (a) the number of milliseconds since the trade and
(b) the sum of all order-volume by market makers for the day for the security. Scaling
by market-maker order volume allows us to better compare results across the different
securities. It is computationally impractical to compute cancellations on a linear scale
(i.e., for every millisecond), and by dividing by passage-time, we can provide a better
assessment of the cancellation rate per millisecond.
Panel A in Figure 5 plots the aggregate distribution of opposite-side cancellations for
Subsample 1. It illustrates that the cancellation rate is largest in the first 10 milliseconds
after a trade, and the rate is 6 to 7 times smaller after 100 milliseconds. To the naked
eye, which cannot process information in intervals smaller than 250 milliseconds, the
quotes would thus indeed appear to disappear instantly.
By canceling quotes, a market maker aims to ensure that he is not “picked off” in
the future. However, it is possible that there are other traders’ orders in the book at
now “wrong” prices, e.g., buy orders that are at bid prices that became “too high” after
a sell. These stale quotes can present a profit opportunity if the fundamental value has
moved sufficiently. Thus assuming that non-HFTs are slow to react to the information
contained in prices and that such traders’ orders are present in the order book, we should
observe that fast traders, including HFT market makers, take advantage of stale quotes
and trade against them. An alternative view is that market makers have predictive power
over future order flow and thus trade in anticipation of such flow. Namely, large buy-side
orders (“parent” orders) are usually not traded in one large chunk, but they commonly
get split into smaller (“child”) orders that are traded over time. If, by observing the
order flow, market makers can detect the presence of a large parent order (for instance,
because the buy-side child orders are traded with a poorly designed algorithm), then
market makers may try to trade ahead of the large order and, as a consequence, would
trade against existing orders in the book. Both views presented here would result in a
situation where, subsequent to the trade, HFT market makers submit aggressive trades
in the direction of the trade.
We compute the submission rate of aggressive orders that are in the direction of the

20
trade analogously to the cancellation rate. We specifically consider not only marketable
orders (which trade) but also immediate-or-cancel orders (which were presumably sub-
mitted with the intention to trade but which may not find an execution). As with
cancelled volume, we compute the aggressive volume submitted 2, 5, 10, 50, 100, 500,
and 1,000 milliseconds subsequent to a trade, and we scale it by the market maker
aggressive volume for the day for the security and the passage time since the trade.
Panel B in Figure 5 plots the aggregate distribution of same-side aggressive order
submissions. As can be seen, the rate is also largest in the first 2,5, and 10 milliseconds
after a trade, and the rate is much smaller at longer horizons.

B. Trade-by-Trade Regression Analysis

Having analyzed the aggregate cancellations and order submission rates, we seek to
confirm these casual observations with a formal regression analysis. We approach the
problem from different angles, using different types of independent variable specifications
for our dependent variables, the cancellation and order submission rates.
We additionally analyze whether it matters if the trade was part of a smart-order-
routed trade and whether it matters if the liquidity provider of the trade was a market-
maker. We perform the regressions analysis in this subsection based on trade-by-trade
observations. As the trading day progresses, it may be possible for traders, in particular
market makers who are active in many trades, to detect whether a particular trader
accumulated a position. It is thus imaginable that over the day (and thus over larger
cumulative volume) cancellation rates increase. In the next subsection, we aggregate by
cumulative volume deciles to further study this issue.
Smart-Order Routed Trades. We classify a trade as stemming from a smart order
router (SOR) if all of the following three conditions are met: (1) trade originated from
the same trader, (2) the trade involves transactions on at least two different marketplaces
within 5 milliseconds and (3) if the time from the first to the last transaction does not
exceed 9 milliseconds. We aggregate the total value for such trades, and we consider the
cancellation and new order submission volume based on the time of the first transaction

21
that is part of the trade. We then estimate the following regression

3
X
DVt+m = α × SORt |volt | + κ × (1 − SORt )|volt | + βj controljt + ǫt , (3)
i=1

where DVt+m is the dependent variable that measures the volume of cancelled orders in
the opposite direction of the time t trade in the m milliseconds following the trade at
time t; volt is the time t dollar-volume (aggregated by market order); the three control
variables controljt and are the log of the daily cumulative volume in the security up until
the trade, the signed trade imbalance up to but not including that trade (in percent),
and the security’s return since the open (in percent). Variable SORt is a dummy that is
1 if the trade at time t is a SOR trade and 0 otherwise. For this regression, we will test
whether αj = κj .
Which type of liquidity provider? Similarly to smart order routed trades, we
assess whether the cancellation rate depends on the type of trader who supplied liquidity.
For this test, we interact the value with dummies for market maker and non-market
maker passive traders

3
X
DVt+m = α × mmt |volt | + κ × (1 − mmt )|volt | + βj controljt + δi + ǫt , (4)
i=1

where mmt is a dummy that is 1 if the trade at time t had a market maker on the passive
side (for any of the transactions) and 0 otherwise.
Results. Table II presents the results for the case where we do not distinguish between
the types of trader who initiated the order (except that we only consider marketable
orders from buy-side, retail and non-market-making HFT). To simplify the exposition,
we multiplied the dependent variable with 1,000,000.
We observe a positive impact of a trade’s dollar-volume on the cancellation and
aggressive order submission rates, and we observe that the effects on the rates are largest
in the first 5-10 milliseconds after the trade. We also observe that having a market maker
on the passive side or the trade being classified as a SOR trade increases the cancellation
rate, except in the first 5ms. For aggressive order submission rates, there is a similar
pattern for trades with HFT on the passive side in the sense that in the first 5-10
milliseconds after the trade, trades without an market maker on the passive side have

22
higher cancellation rates and that after the first 10 milliseconds, the order reverses. For
SOR trades, the estimated coefficient is always smaller, though the order of magnitude
is similar. Notably, SOR trades are three-times larger than non-SOR trades, and thus
when a SOR trade occurs, the observed impact will be larger. For most specifications,
the estimated coefficients are statistically significantly different. Since we aggregate the
data based on the first transaction of a trade, it is possible that the difference in the first
five milliseconds stems from trading system latency in the sense that we start counting
cancellations too early.

C. Which type of trader demands liquidity?

We determine the type of trader for the aggressive order and the trader’s daily ab-
solute net position up until and including this trade. We consider incoming trades only
from buy-side traders, retail traders, and non-market-making HFT; we omit market-
making HFTs because their trading and cancellation activities may be jointly deter-
mined. We perform the analysis by aggregating the data across dollar-volume and
inventory-deciles. Namely, for each security and for each type of volume (dollar value,
absolute value of directional volume across all traders and by group of traders (buy-
side, retail and non-market-making HFT) we determine the deciles. We then compute
for each day, stock and decile the averages for the cancellation rates, same-side aggres-
sive order submission rates, and the control variables (market-wide trade imbalance,
return, and volume traded). We aim to establish whether or not the cancellation rate
of market-makers depends on the volume that the trader had accumulated (expressed
by the trader’s cumulative volume) before he initiated the trade. We then estimate the
following regression

3
i X
DV d,m = α + β × decile + γj controljt + δ × VIXt + ǫt , (5)
i=1

where DV d,m is the dependent variable that measures the average daily volume of can-
celled orders in the opposite direction of the trade for security i on day d in the m
milliseconds following the relevant trade; decile is the decile (1 to 10) for the cumula-
tive volume for the type of trader who submitted the market order; VIXt is the daily
realization of the U.S. volatility index VIX; the three control variables controljt are the

23
log of the daily cumulative volume in the security up until the trade, the signed trade
imbalance up to but not including that trade (in percent), and the security’s return since
the open (in percent). We estimate this regression separately for all traders and for the
three types of incoming trader (buy-side, retail, non-market making HFT). Transactions
by the same trader in the same direction within a time span of 9 milliseconds are con-
sidered to be a single trade. Standard errors in this regression are double-clustered by
date and firm to control for cross-sectional and serial correlations.
Results. Table III displays the summary statistics for the deciles of directional volume
for all traders, buy-side and non-market-making HFT. Figures 6 and 7 plot the average
cancelled and aggressively submitted volume for all traders, buy-side and non-market-
making HFT trades relative to the deciles for the cancellation and order submission
volume 2, 5 and 10 milliseconds after the trade. Buy-side traders generally have (much)
larger cumulative inventory than other types of traders, and the decile values for buy-
side mimic those of all traders combined. If the trader’s cumulative position matters,
then there should be a positive relationship between deciles and cancellation rates. As
can be seen, for all traders and for buy-side traders separately the rate is indeed increas-
ing,14 which suggests that as a buy-side trader accumulates an inventory, market makers
withdraw more volume from the market subsequent to trades. For non-market-making
HFT, however, the pattern is less clear: Panel C in Figure 7 indicates that there may
be only a weak relationship of accumulated inventory and cancellations, and Panel D
indicates that there may be no relationship of accumulated inventory and aggressive
order submissions.
Table IV presents the results for the case where we aggregate across all traders (either
by trading value or by directional volume). For all subsequent regression outcome tables,
we multiplied the dependent variable by 100,000 to ease the exposition. We observe a
positive impact of (directional) volume, and we also observe a positive slope, implying
that higher (directional) volume has a (proportionally) larger impact on the cancellation
and new order submission rates. Furthermore, we observe that the estimated effects for
cancellations are largest for the first 50 milliseconds (i.e., for the 2ms, 5 ms, 10ms and
50ms regressions) and for aggressive, same-side submission for the first 10ms. Indeed,
14
In a robustness check, we also estimated a specification with an estimate for each decile. Then we
checked whether the coefficient for the 10th decile differed from the one for the first decile. The results
are consistent.

24
the estimated coefficients are larger by an order of magnitude compared to those for
“later” (100ms+) cancellations.
Tables V displays the estimation results for the regressions by trader type. To sim-
plify the exposition, the table only shows the estimates for α, the intercept, and β, the
slope coefficient for the deciles. Thus each combination of α and β in the table represents
the result from a single regression.
For trades that stem from non-market making HFT, the slope coefficient is positive
and significant, except for long delays (500ms+), and the intercept for cancellations after
the trade is insignificant. For aggressive order submissions after the trade the slope and
most intercepts are insignificant, but positive. These results suggest that even those the
cancellation rate after HFT trades is positive, the net inventory of the HFT does not
affect the reaction of market makers. It is important to note that the decile values for
HFT and buy-side are different in the sense that buy-side traders generally accumulate
much more directional volume.
For trades from buy-side traders, the estimated slope for cancellations with regards
to the decile is positive and significant. Since the slopes are largest for the first four
estimations (2, 5, 10 and 50 ms), the cancellation effect of the volume is strongest
shortly after the trade. Since the intercept is non-negative (and the lowest decile value
is 1), the regression additionally confirms the trade-by-trade regression w.r.t. the overall
positive effect of a buy-side trade on the market-maker cancellation rate. For same-side
aggressive trades, the effect is strongly positive for buy-side traders, in particular for the
first three estimations (2, 5 and 10 ms). Since the intercept is positive and significant,
we conclude that there is a positive effect of a trade on same-side order submissions and
that this effect is positively related to the size of the inventory of the buy-side trader.
For trades from retail traders, the results are similar to those of buy-side traders.
Taken together, our decile analysis suggests that (a) market makers quickly move
the quote in response to trades, in particular from the buy-side, and (b) market makers
submit aggressive orders in the direction of the buy-side trade; furthermore, both effects
are amplified by the size of the inventory that the trader has already accumulated and
by the dollar-size of the incoming order.

25
VII. Market Making and Arbitrage

In this section we examine trading behaviour in securities that are linked by the same
underlying fundamentals. We consider three types of possible security pairs: stocks
with unequal voting rights, stocks with preferred shares, and stocks with debentures
(unsecured, exchange-traded debt).
We will explain our approach using the example of dual-class shares. When a firm has
dual-class shares —typically with different voting rights— one share class usually trades
more frequently than the other. An example in the Canadian context is Bombadier
Inc.: its BBD.B shares are among the most frequently traded securities, its BBD.A
shares trade less often. Even though they may represent the same claim on future cash
flows, stocks with unequal voting rights may trade at different prices, but over intra-
day horizons, it is unlikely that this value difference fluctuates strongly; see Shive and
Schultz (2010) and references therein. Intra-day price movements in both securities
should thus be highly correlated, except for extraordinary circumstances (e.g., when
corporate actions trigger a change in the difference between the two classes of shares).
At the high frequency level, when prices for one security move, the other should follow
suit, putting pressure on those traders that make the market in that security. Consider
a hypothetical firm for which its Class A securities trade frequently and its Class B
securities trade less frequently. We will study how trades in the Class A securities affect
trading and order submissions in the Class B securities. Consequently, in our regressions,
the dependent variables relate to the less-frequently traded securities (Class B in our
example here) whereas the explanatory variables relate to trading in the more-frequently
traded security (Class A).
We first observe that, as expected, there are fewer trades in the higher-voting rights
share class, preferred shares, and debentures compared to the other voting rights share
class, common shares, and equities respectively. For preferred shares, debentures, and
dual-class shares, there are, on average 125, 78 and 208 times as many trades in the
frequently-traded securities; see Table VI.
Next, we establish the price impact of trading in the more frequently traded securities
(Class A) on the less-frequently traded (Class B) securities. We compute price impacts
based on constructing the national best bid and offer price for the class B securities. We
define the price impact of a trade in the frequently traded security f on the infrequently

26
traded security i at time t as

price impactit = qtf (mit+x − mit )/mit , (6)

where qtf is 1 if the trade in f was a buy and −1 if it was a sell; mit is the midpoint
of the bid-ask spread in security i at time t; and mit+x is the midpoint of the bid-ask
spread in the security x units of time after time t, where we compute the price impact
for 1, 5 10, and 100 milliseconds, 1, 5, 10, and 30 seconds and 1,5, and 10 minutes. In
our regression analysis, we compute the volume-weighted averages per stock per day.
Finally, analogously to Section VI. we compute, for each trade in Class A securities,
how many trades and order cancellations follow in Class B securities by market makers
in the 2, 5, 10, 50, 100, 500 and 1,000 milliseconds after the Class A trade. For all
measures, we further differentiate between trades by the different types of traders.
We perform the analysis at the daily level, where, as in the last section, we scale
the daily observations by the daily market maker order volume and the time. We then
perform t-tests for the coefficients of interest. A statistically significant observation
indicates that the measure for buy-side, non-market-making HFT and retail respectively
is positive; the standard errors that we employ in these t-tests are clustered by firm and
date. In the last section, where we analyzed market making behavior following a trade
for the same security, we considered cancellations in the opposite direction of the trade.
Here, we study different securities and we include orders, cancellations and trades in both
directions because the direction of a strategy likely depends on the specific arbitrage
relation; for this reason, we also did not perform an analysis of same-side aggressive
order submissions.
Results. Our results for the daily price impact estimation are in table Table VII;
standard errors for this regression are clustered by security-pair and date. We first
observe that, in dual class shares, trades in the more frequently traded shares have a
positive and significant price impact for the less-often traded shares. For preferred shares
and debentures, we observe no clear pattern in terms of sign, size, or significance of the
estimated coefficients.
Our results for the daily estimations for canceled orders for debentures, preferred
shares and dual-class shares are in Table VIII. For debentures, we observe positive, but
insignificant effects of the trades on cancellation rates in the frequently-traded securities.

27
For preferred shares and dual class shares, we observe positive and significant effects for
cancellation rates. For all three classes of securities, consistent with the observations for
“isolated” securities, we observe that the cancellation effects are strongest in the first
10ms after the trade.

VIII. HFT in TSX Venture-Listed Securities

We discuss the extent of high-frequency trading in securities that are listed on the
Toronto Stock Exchange Venture, TSXV. The TSXV is the junior market to the TSX,
and its purpose is to bring young, smaller companies to the public market. Securities in
the TSXV tend to have small market capitalization and they often trade at sub-penny
prices. Nevertheless, even for low-priced stocks, the tick size is at least $0.005, making
it potentially beneficial to earn marking making profits with a low capital commitment.
Like TSX-listed securities, TSXV-listed securities can trade on any of Canada’s other
marketplaces. In this appendix, we present numbers for trading on all marketplaces
combined.
For this part of the analysis, we determined all TSXV-listed securities that traded
at any point between January and May 2013. We did not separately classify trader IDs
based on their trading in TSXV securities. Instead, we retained the trader groups from
the main analysis in the sense that trader IDs that trade in TSXV securities are classified
based on their trading in Large Sample securities. There are a number of reasons for why
we did not apply our classification method using trading in TSXV securities. Most of
the TSXV-listed securities trade infrequently, and it would thus be difficult to generate
enough data to assess reaction times. Furthermore, most TSXV securities commonly do
not qualify for the market-on-close facility, and we thus cannot classify traders based
on their reaction speed to the information about market-on-close imbalances. Since our
classification is based on broad, absolute measures, traders that act as HFTs in TSX
securities are also reasonably HFTs in TSXV securities. It is, however, possible that
some HFT IDs may specialize in TSXV-listed securities, and using the traders classified
for TSX-listed securities could thus lead to an underestimation of the extent of high
frequency involvement.
On average only around 8% of all passive trades in TSXV-listed securities were
by HFTs, and only 2% of aggressive trades. However, there is some significant cross-

28
sectional variation. Of the 2,478 TSXV securities, 1,266 have no HFT trading, and 1,676
have no aggressive HFT trades. Of those that do have some form of HFT trading, the
fraction of HFT trading is positively related to the trading volume in these securities
(this observation does not imply a causal relationship). Figure 8 plots the percentage
of total value traded by HFT (aggressive/passive in Panels A/B) in TSXV securities
against the total value (in logs).
Figure 9 shows the HFT participation as a fraction of all transactions and value.
Compared to the main market, HFT activity in TSXV stocks is lower, with only 20% of
transactions and less than 10% of value involving an HFT on the passive side (aggressive
volume is negligible). As Panel C indicates, however, they submit about 55% of all
orders, and they are thus noticeable in these securities. Panel D highlights the share
of HFT and other traders’ orders relative to best prices. HFTs generally submit more
than half of the orders that are submitted at the best price or that are improving the
best price.
Overall, there are many securities that see no HFT activity at all, but for the venture
market as a whole, HFT activity will be noticeable, in particular with regard to the
posting of aggressive, competitive quotes.

29
REFERENCES

Brogaard, Jonathan, Terrence Hendershott, and Ryan Riordan, 2014, Market integration
and high frequency intermediation, Discussion paper, IIROC Working Paper.

Comerton-Forde, Carole, Katya Malinova, and Andreas Park, 2015, Regulating dark
trading: Order flow segmentation and market quality, Discussion paper, IIROC Work-
ing Paper.

Devani, Baiju, Lisa Anderson, and Yifan Zhang, 2015, Impact of the dark rule amend-
ments, Discussion paper, IIROC Working Paper.

Devani, Baiju, Ad Tayal, Lisa Anderson, Dawei Zhou, Juan Gomez, and Graham W.
Taylor, 2014, Identifying trading groups – methodology and results, Discussion paper,
IIROC Working Paper.

Glosten, Lawrence R., 1994, Is the electronic open limit order book inevitable?, The
Journal of Finance 49, 1127–1161.

, and Paul R. Milgrom, 1985, Bid, ask and transaction prices in a specialist
market with heterogenously informed traders, Journal of Financial Economics 14,
71–100.

IIROC, 2012, The HOT study, Discussion paper, Investment Industry Regulatory Or-
ganization of Canada.

Korajczyk, Robert, and Dermot Murphy, 2014, High frequency market making to large
institutional trades, Discussion paper, IIROC Working Paper.

Kyle, Albert S., 1985, Continuous auctions and insider trading, Econometrica 53, 1315–
1336.

Shive, Sophie, and Paul Schultz, 2010, Mispricing of dual-class shares: Profit opportu-
nities, arbitrage, and trading, Journal of Financial Economics 98, 524–549.
 Distribution of Trader Types

99
131 88
778

3796

HFT MM buy−side other retail

Figure 1
Trader Types
 Figure 2
Visualization of Two-way clusters for HFT Groups
The figure represents the table of pairwise distances computed in equation (1) for the 101 trader IDs that are classified as HFT. Smaller
numbers are represented by darker cell-shadings.
 Distribution of Transactions by Types Distribution of Value by Types

100
100
80

80
60

60
40

40
20

20
0

0
active passive active passive

HFT MM buy−side retail others HFT MM buy−side retail others

Panel A: Transactions Panel B: Value

Figure 3
Distribution of Trades by Trader-Types
The figure presents the distribution of aggregate trades (Panel A) and value traded (Panel B) for our large sample for the time from
January to May 2013, split by classified trader-types.
 Distribution of Aggressive Orders by Types

80

50
60

40
%orders

30
40

20
20

10
0

20 40 60 80 100

0
February−May 2013 atbest better mktable worse

HFT Buyside non−HFT MM retail HFT MM buy−side retail

Panel A: Times Series all Order Volume Panel B: Overall Distribution

80

80
% orders submitted at best by HFT
% improving orders that are HFT
70

70
60

60
50

50
40

40
0 20 40 60 80 100 0 20 40 60 80 100
day day

Market A Market B Market C Market A Market B Market C

Panel C: HFT as a percentage of improving Panel D: HFT as a percentage of at-best submissions

Figure 4
Distribution of Order Volume by Trader Type and Aggressiveness
Panel A plots the time series of the distribution of aggregated order volume. As of early May 2013, there was an upward shift in HFT
order volume and a simultaneous downward shift in non-HFT market maker volume. This shift was largely caused by changes to the
order submission behavior on market A: there, a number of market-making, non-HFT IDs stopped posting while at the same time
number of previously classified HFT-market-making IDs increased their quoting behavior. The same shift is visible in Panels C and D,
which plot the distribution of price-improving and at-best orders by trading venue. Taken together we thus hypothesize that several
non-HFT and HFT market makers are, in fact, linked which is why we combine all market makers when we assess the reactions to
trades. Panel B plots the distribution of orders by aggressiveness and type. Stacked up, the bars would add to 100%. Aggressiveness
is defined relative to the prevailing best price posted on the respective trading venue.
 Distribution of Opposite−Size Cancelations per unit of time Distribution of Same−Side Aggressive Orders per unit of time

60
2
1.5

40
1

20
.5
0

0
ms0002 ms0005 ms0010 ms0050 ms0100 ms0500 ms1000 ms0002 ms0005 ms0010 ms0050 ms0100 ms0500 ms1000

buyside retail non−MM HFT buyside retail non−MM HFT

Panel A Panel B
Figure 5
Cancellation and order-submission rates in response to trades.
The figures plot the opposite-side cancellation (Panel A) and same-side aggressive order submission (Panel B) rates by market-making
traders. The rates are split by whether the trade that preceded the cancellations was by a retail, a buy-side or non-market-making
HFT.
 Market−makers’ cancellation vs. trader inventory at order submission Market−makers’ cancellation vs. volume at order submission

2000

1500
average opposite−side cancelled volume

average opposite−side cancelled volume

1500

1000
1000

500
0 500

0
0 2 4 6 8 10 0 2 4 6 8 10
quantiles of |cumulative volume| quantiles of vol

2ms 5ms 10ms 2ms 5ms 10ms

Panel A Panel B
Market−makers’ aggressive orders vs. trader inventory at order submission Market−makers’ aggressive orders vs. volume at order submission
800

600
average same−side aggressive volume

average same−side aggressive volume

600

400
400

200
200 0

0
0 2 4 6 8 10 0 2 4 6 8 10
quantiles of |cumulative volume| quantiles of volume

2ms 5ms 10ms 2ms 5ms 10ms

Panel C Panel D
Figure 6
Cancellations and Order Submissions by Deciles of the Order Volume
and Incoming Trader’s Cumulative Inventory – all traders.
The figures plot the average opposite-side cancellation (Panels A and B) and same-side aggressive order submission (Panels C and
D) volume by market-making traders against the dollar-volume of the incoming order and the deciles of the absolute inventory of the
submitting trader. Deciles are computed across all trades per security. We do not scale volume here (thus for more passage-time, the
volume will be higher). We note that the dollar-volume deciles are very similar for the first 4 deciles.
 Market−makers’ cancellation vs. buy−side inventory at order submission Market−makers’ aggressive orders vs. buy−side inventory at order submission

2000

800
average opposite−side cancelled volume

average same−side aggressive volume

1500

600
1000

400
500

200
0

0
0 2 4 6 8 10 0 2 4 6 8 10
quantiles of |cumulative volume| quantiles of |cumulative volume|

2ms 5ms 10ms 2ms 5ms 10ms

Panel A Panel B
Market−makers’ cancellation vs. HFT inventory at order submission Market−makers’ aggressive orders vs. HFT inventory at order submission

1000
2000
average opposite−side cancelled volume

average same−side aggressive volume

800
1500

600
1000

400
500

200
0

0 2 4 6 8 10 0 2 4 6 8 10
quantiles of |cumulative volume| quantiles of |cumulative volume|

2ms 5ms 10ms 2ms 5ms 10ms

Panel C Panel D
Figure 7
Cancellations and Order Submissions by Deciles of the Incoming Trader’s
Cumulative Inventory — by trader type.
The figures plot the average opposite-side cancellation (Panel A and C) and same-side aggressive order submission (Panel B and D)
volume by market-making traders against the deciles of the absolute inventory of the submitting trader. Deciles are computed with
regard to the trader-types. The rates are split by whether the trade that preceded the cancellations was by a buy-side or non-market-
making HFT. In contrast to Figure 5, we do not scale volume here (thus for more passage-time, the volume will be higher).
 40
80

30
%passive value by HFT
60
% active value by HFT

20
40

10
20
0

0
0 5 10 15 20 0 5 10 15 20
log(value) log(value)

Panel A: Aggressive HFT vs. Log-Value Panel B: Passive HFT vs. Log-Value
Figure 8
Participation of HFT in TSXV-listed securities
The figure plots the log of value traded against the share of value traded actively (Panel A) and passively (Panel B) by HFT for our
sample of 2,478 TSXV securities for the time from February 2013 to end-May 2013. The trader classification is based upon the main
sample of highly-liquid TSX-listed securities.
 Distribution of Transactions by Types Distribution of Value by Types

100

100
80

80
60

60
40

40
20

20
0

0
active passive active passive

HFT MM buy−side retail others HFT MM buy−side retail others

Panel A: Transactions Panel B: Value

Distribution of Aggressive Orders by Types

60

40
30
40
%orders

20
20

10
0

20 40 60 80 100
0

February−May 2013 atbest better mktable worse

HFT Buyside non−HFT MM retail others HFT MM buy−side retail

Panel C: Order Volume Panel D: Order Aggressiveness

Figure 9
TSX Venture Stocks: Distribution of Trades and Orders by Trader-Types
The figure plots the log of value traded against the share of value traded actively (Panel A) and passively (Panel B) by HFT for those
TSXV securities that display any form of HFT-trading during the time from February 2013 to end-May 2013. The trader classification
is based upon the main sample of highly-liquid TSX-listed securities.
 Table I
Summary Statistics for Trader Type Activities
The table presents summary statistics for trading activities by the different trader types; all numbers are averages per trader in the
respective group. Total orders and trades are summed over the entire horizon from January to May 2013 for the Large Sample. Average
daily orders are summed across all securities per day.

non-HFT HFT market

All traders HFT Buy-side Retail Others
market makers makers

securities traded 14 67 19 42 59 157 9

average daily trades 569 2,246 943 1,549 2,328 15,222 194
total trades 47,434 173,677 84,674 159,970 150,012 1,500,000 11,343
average daily orders 5,130 25,916 2,282 1,396 38,307 318,902 613
total orders 466,299 2,032,000 215,933 144,629 2,509,000 31,160,000 46,415
%orders within 1ms of market event 35 61 37 35 33 82 33
median order to cancel time 274,124 38,494 193,351 550,132 825,670 11,116 274,494
trades at 15:40 71 2,311 38 9 64 3,897 2
orders at 15:40 317 3,353 125 3 690 23,130 14
average daily trade imbalance 91 79 99 92 26 31 92
average daily order imbalance 93 74 99 92 10 11 95
 Table II
Cancellations after Trades: Smart-Order Routers and Passive MM-HFT
The table presents results from our OLS estimation of equations (3) and (4) which assesses whether the impact of a trade differs
depending on whether or not the aggressive order stemmed from a smart order router (Panel A and C) and whether or not the
passive-side trader was a market maker or not (Panel B and D). The regression is based on SubSample1 and it is estimated for each
marketable order (1,534,305 observations). For most estimates in Panels A and C, we cannot reject a test of equality of the coefficients
for the different trader groups, for most estimates in Panels B and D we can reject the equality hypotheses. All regressions include
three control variables: the log of the daily cumulative volume in the security up until the trade, the signed trade imbalance up to
but not including that trade (in percent), and the security’s return since the open (in percent). Standard errors are in parentheses. *
indicates significance at the 10% level, ** at the 5% level, and *** at the 1% level.

2ms 5ms 10ms 50ms 100ms 500ms 1000ms

Panel A: SOR trades and opposite side cancellations

SOR trade 0.000799*** 0.00382*** 0.00473*** 0.00308*** 0.00168*** 0.000399*** 0.000208***

(0.00009) (0.00022) (0.00027) (0.00016) (0.00009) (0.00002) (0.00001)

non-SOR trade 0.00374*** 0.00450*** 0.00378*** 0.00224*** 0.00125*** 0.000344*** 0.000188***

(0.00014) (0.00020) (0.00017) (0.00009) (0.00005) (0.00001) (0.00001)

Panel B: HFT MM passive and opposite side cancellations

passive MM 0.00315*** 0.00591*** 0.00628*** 0.00365*** 0.00201*** 0.000492*** 0.000261***

(0.00014) (0.00020) (0.00021) (0.00011) (0.00006) (0.00001) (0.00001)

passive other 0.00015 0.000734*** 0.000558*** 0.000924*** 0.000500*** 0.000152*** 0.0000831***

(0.00008) (0.00009) (0.00007) (0.00007) (0.00004) (0.00001) (0.00001)

Panel C: SOR trades and same-side aggressive orders

SOR trade 0.0763*** 0.152*** 0.139*** 0.0416*** 0.0215*** 0.00454*** 0.00233***

(0.01150) (0.01060) (0.00829) (0.00238) (0.00123) (0.00026) (0.00013)

non-SOR trade 0.441*** 0.304*** 0.188*** 0.0490*** 0.0258*** 0.00574*** 0.00304***

(0.03030) (0.01530) (0.00980) (0.00247) (0.00123) (0.00028) (0.00014)

Panel D: HFT MM trades and same-side aggressive orders

passive MM 0.125*** 0.185*** 0.158*** 0.0476*** 0.0250*** 0.00543*** 0.00282***

(0.01280) (0.00370) (0.00610) (0.00169) (0.00089) (0.00019) (0.00010)

passive other 0.453*** 0.285*** 0.166*** 0.0399*** 0.0206*** 0.00441*** 0.00231***

(0.04420) (0.02470) (0.01470) (0.00356) (0.00183) (0.00039) (0.00020)
 Table III
Summary Statistics: Cancellations and Aggressive Orders after Trades by Deciles

P Opposite-side cancelled order volume Same-side aggressive volume

Decile | volume| 2ms 5ms 10ms 50ms 100ms 500ms 1000ms 2ms 5ms 10ms 50ms 100ms 500ms 1000ms

Panel A: All traders

1 292 145 381 628 1,893 2,215 3,186 3,696 84 176 244 352 390 466 517
2 959 170 454 744 2,142 2,467 3,470 3,977 95 203 281 398 434 507 559
3 1,933 177 475 788 2,242 2,577 3,578 4,097 100 219 306 430 462 535 594
4 3,358 189 521 869 2,437 2,768 3,737 4,266 103 233 335 475 506 571 627
5 5,429 211 573 976 2,651 2,995 3,985 4,514 107 248 350 503 533 593 644
6 8,427 215 580 996 2,675 3,027 3,992 4,498 110 255 367 536 566 624 672
7 12,987 219 602 1,049 2,876 3,243 4,195 4,691 120 277 413 616 651 710 755
8 20,600 230 642 1,140 3,093 3,435 4,363 4,888 130 298 478 737 773 840 889
9 36,902 259 787 1,474 3,942 4,337 5,391 5,921 144 359 578 905 972 1,064 1,116
10 84,009 244 814 1,722 4,244 4,653 5,720 6,247 138 423 692 1,057 1,102 1,200 1,274

Panel B: non-market-making HFT

1 139 103 220 334 703 816 1,258 1,498 168 334 487 754 830 1,063 1,229
2 366 114 232 359 776 917 1,430 1,722 211 408 591 924 1,054 1,278 1,431
3 727 161 319 486 1,004 1,186 1,778 2,092 258 478 682 1,143 1,275 1,574 1,757
4 1,052 124 255 404 881 1,041 1,612 1,920 211 438 658 1,119 1,242 1,531 1,718
5 1,802 121 252 413 890 1,198 1,709 2,108 253 496 756 1,248 1,419 1,686 1,929
6 2,711 98 231 382 887 1,024 1,474 1,838 205 405 619 1,057 1,179 1,446 1,669
7 4,455 104 225 397 900 1,050 1,484 1,745 212 417 667 1,050 1,197 1,443 1,774
8 6,829 105 278 467 928 1,086 1,528 1,732 201 419 717 1,139 1,273 1,511 1,704
9 12,697 152 331 498 929 1,040 1,432 1,623 301 558 880 1,381 1,488 1,665 1,859
10 26,505 102 254 442 887 998 1,390 1,600 137 417 784 1,830 1,894 2,042 2,281

Panel C: buy-side
1 330 58 135 194 438 506 757 886 71 150 206 292 322 380 421
2 1,086 69 160 228 489 562 823 950 85 184 252 349 379 434 475
3 2,149 73 167 242 512 584 848 975 92 203 285 395 422 476 524
4 3,670 79 187 273 561 631 893 1,028 98 223 318 446 472 528 579
5 5,844 88 205 306 617 693 956 1,089 103 241 337 487 514 573 620
6 8,941 88 209 315 618 698 956 1,079 110 256 368 530 559 613 657
7 13,599 90 219 329 646 726 989 1,110 119 275 406 601 637 699 745
8 21,388 98 234 360 698 770 1,031 1,166 132 303 483 732 766 832 882
9 38,074 108 277 454 896 985 1,302 1,438 164 376 574 883 951 1,047 1,099
10 86,522 103 299 552 1,028 1,114 1,428 1,562 132 412 684 1,053 1,096 1,185 1,251
 Table IV
Regressions: Cancellations and Aggressive Orders after Trades by Deciles.
The table presents results from our OLS estimation of equation (5). It assesses whether the impact of a trade differs depending on the
decile of the order-submitter’s cumulative volume (Panel A) and directional volume (Panel B) at the time of the trade. The regression
is based on SubSample1 and it is estimated for the stock-day-deciles for all traders. All regressions include four control variables (all
averaged by decile): the log of the average daily cumulative volume in the security up until the trade, the signed trade imbalance up
to but not including that trade (in percent), the security’s return since the open (in percent), and the daily realization of the U.S.
volatility index VIX. Standard errors are in parentheses and they are double-clustered by firm and date. * indicates significance at the
10% level, ** at the 5% level, and *** at the 1% level.

2ms 5ms 10ms 50ms 100ms 500ms 1000ms

Panel A: Directional volume and opposite side cancellations

Slope 1.70*** 1.98*** 2.00** 1.58*** 0.83*** 0.18*** 0.09***

(0.35) (0.60) (0.90) (0.41) (0.21) (0.05) (0.02)
Intercept -82.84 -137.96 -150.78 50.21 30.89 12.06** 7.41**
(109.08) (127.26) (124.30) (57.71) (29.95) (6.09) (3.31)

Panel B: Directional volume and same side aggressive volume

Slope 227.39*** 179.86*** 122.52*** 29.93*** 15.08*** 3.18*** 1.60***

(72.81) (28.38) (21.08) (5.89) (2.99) (0.63) (0.31)
Intercept 17,691.69*** 9,654.24*** 4,763.13*** 1,000.57* 524.81* 121.07** 64.39**
(3,096.71) (1,802.00) (1,521.84) (513.16) (273.12) (56.90) (30.49)

Panel C: Dollar volume and opposite side cancellations

Slope 2.62** 4.04*** 3.52*** 2.90*** 1.58*** 0.42*** 0.23***

(1.19) (1.19) (0.95) (0.25) (0.13) (0.03) (0.02)
Intercept -25.74 -30.53 -21.10 62.36*** 35.99*** 11.68*** 7.06***
(62.21) (58.63) (46.07) (19.59) (10.31) (2.21) (1.32)

Panel D: Dollar volume and same side aggressive volume

Slope 398.18*** 311.35*** 200.44*** 50.62*** 26.57*** 5.82*** 3.06***

(47.52) (29.18) (17.51) (4.27) (2.30) (0.51) (0.27)
Intercept 8,103.25*** 4,315.75*** 2,296.57*** 576.27** 304.07** 70.69*** 39.38***
(1,744.60) (1,172.79) (799.58) (228.08) (119.86) (26.27) (13.82)
 Table V
Regressions: Cancellations and Aggressive Orders after Trades by Deciles.
The table presents results from our OLS estimation of equation (5). It assesses whether the impact of a trade differs depending on
the quintile of the order-submitter’s cumulative inventory at the time of the trade. The regression is based on SubSample1 and it
is estimated for the stock-day-quintile (6,737 observations for buy-side, 5,418 for non-market-making HFT, and 6,291 for retail). All
regressions include four control variables (all averaged by quintile): the log of the average daily cumulative volume in the security up
until the trade, the signed trade imbalance up to but not including that trade (in percent), the security’s return since the open (in
percent), and the daily realization of the U.S. volatility index VIX. Standard errors are in parentheses and they are double-clustered
by firm and date. * indicates significance at the 10% level, ** at the 5% level, and *** at the 1% level.

2ms 5ms 10ms 50ms 100ms 500ms 1000ms

Panel A: Opposite-side cancelled order volume

buy-side slope 1.878311*** 2.089765*** 2.129935** 1.679728*** 0.889664*** 0.210862*** 0.113055***

(0.378322) (0.619629) (0.905711) (0.411327) (0.214700) (0.045083) (0.022218)
intercept -84.740302 -140.480081 -152.870120 46.217414 28.767977 11.466415* 7.026424**
(108.310335) (126.154238) (123.677602) (57.278019) (29.502529) (5.933288) (3.189934)

non-MM HFT slope 1.409190* 1.873467*** 1.549070*** 0.911765*** 0.465798*** 0.084447 0.044462
(0.816054) (0.709918) (0.573410) (0.288212) (0.170468) (0.057924) (0.033926)
intercept -501.438150 -408.846948 -315.444831 -42.357761 -23.404032 -0.714002 0.883114
(349.255846) (281.482068) (222.976829) (89.330541) (49.824929) (10.480873) (6.078769)

retail slope 3.033805*** 3.339257*** 3.137842*** 2.240723*** 1.326900*** 0.344408*** 0.187405***

(0.640917) (0.846633) (0.855842) (0.360650) (0.201783) (0.047876) (0.023550)
intercept -49.722733 -133.171311 -113.890587 111.328841* 65.179915** 19.089377*** 11.060655***
(145.394688) (173.708333) (144.618198) (57.946775) (31.164041) (6.073277) (3.186634)

Panel B: Same-side aggressive volume

buy-side slope 353.867509** 235.803974*** 147.658130*** 36.767947*** 18.738474*** 4.005234*** 2.052980***

(156.638508) (58.467007) (29.782993) (6.630350) (3.358143) (0.700256) (0.356427)
intercept 19262.840585*** 10166.689667*** 4,636.174092*** 950.908143* 495.681849* 116.196233** 61.414496**
(4,517.726679) (2,068.787520) (1,491.000983) (501.075870) (264.236320) (54.748115) (29.667493)

non-MM HFT slope -25.455048 1.136734 35.005358 8.596510 5.456145 1.065000 0.751001
(42.255192) (26.256268) (31.684198) (8.726318) (5.345547) (1.196771) (0.798077)
intercept 7,989.033811** 6,795.741798** 5,814.865456** 1,243.876197 693.614307 168.727052 97.825422
(3,587.990560) (2,959.147648) (2,966.679349) (853.137856) (512.512151) (115.169678) (73.553756)

retail slope 160.835796*** 139.860817*** 100.599065*** 33.481836*** 19.298198*** 4.237356*** 2.167784***

(24.368633) (18.602006) (20.152911) (6.114839) (3.528363) (0.701904) (0.336995)
intercept 13360.216706*** 7,662.132855*** 3,889.792290*** 1,233.983624*** 685.600488*** 166.089020*** 87.265567***
(1,810.690969) (551.222590) (1,036.142418) (459.442243) (257.526106) (55.444537) (29.823494)
 Table VI
Summary Statistics Related Securities
The table presents summary statistics for trading activities in debenture, preferred shares, the higher voting rights share class for dual
class shares and the more frequently traded security (last column). The table is based on SubSample2.

frequently
Mean Stdev max traded

Preferred trades 21 22 545 2,624

volume 7,407 22,771 1,139,600 741,347
value 187,547 570,439 28,600,000 13,800,000

Debentures trades 31 65 413 2,422

volume 19,488 44,728 310,100 1,827,376
value 80,962 158,995 1,381,529 12,700,000

Dual trades 9 15 203 1,869

volume 74,113 120,740 1,891,000 775,860
value 77,818 124,410 1,919,425 8,203,107
 Table VII
Price Impact for Fundamentally Linked Securities

1ms 5ms 10ms 100ms 1s 5s 10s 30s 1min 5min 10min

Panel A: Dual Class Shares

Buy-Side 0.0000 0.0000 0.0000 0.0020*** 0.0033*** 0.0045*** 0.0054*** 0.0073*** 0.0073*** 0.0112*** 0.0118***
[0.8064] [-0.8160] [-0.4745] [3.0274] [2.6539] [3.1843] [3.1869] [3.5267] [3.6548] [3.1281] [2.9697]

non-MM HFT 0.0000 0.0000 0.0000 0.0019* 0.0060** 0.0054*** 0.0073*** 0.0094*** 0.0135*** 0.0293** 0.0319**
[0.5118] [0.3021] [-0.3739] [1.7273] [2.3903] [3.0148] [2.7094] [3.1750] [3.8955] [2.5343] [2.5062]

retail 0.0000 0.0000 0.0002** 0.0073 0.0002 0.0076*** 0.0059** 0.0084*** 0.0113*** 0.0142*** 0.0157**
[1.0714] [1.2723] [2.4663] [1.5694] [0.0715] [3.1030] [2.3573] [3.5398] [3.7271] [2.8729] [2.2111]

Panel B: Preferred Shares

Buy-Side 0.0000 0.0000 0.0000 0.0000** 0.0000 -0.0000** -0.0001** -0.0001* -0.0002* -0.0002 0.0000
[1.3964] [1.0008] [0.9689] [2.1890] [-0.5436] [-1.9891] [-2.4798] [-1.8232] [-1.9392] [-0.9644] [0.0773]

non-MM HFT 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 -0.0001 -0.0002* -0.0004 -0.0001
[0.3410] [-0.5396] [0.0905] [0.1340] [0.8849] [1.1555] [-0.0819] [-1.2779] [-1.7901] [-1.4841] [-0.2758]

retail 0.0000 0.0000 0.0000 0.0000 0.0000 -0.0001 -0.0001 0.0001 -0.0001 0.0002 0.0002
[0.2214] [0.0919] [0.9296] [1.5156] [0.5258] [-1.2956] [-1.5953] [0.4624] [-0.9515] [0.4251] [0.4883]

Panel C: Debentures
Buy-Side 0.0000 0.0001 0.0001 0.0000 -0.0013 -0.0035* -0.0023 -0.0004 0.0008 0.0036** 0.0042*
[-0.8097] [1.5864] [0.8445] [0.0099] [-1.0102] [-1.8927] [-1.5459] [-0.4042] [0.7784] [1.9764] [1.6621]

non-MM HFT 0.0000 0.0003* 0.0001*** -0.0005 -0.0014 -0.0035 -0.0022 -0.0006 0.0003 0.0038 0.007
[0.7663] [1.9551] [3.0732] [-0.2822] [-0.4198] [-1.0016] [-0.7521] [-0.2195] [0.1171] [1.0988] [1.1256]

retail 0.0000 0.0001* 0.0002 0.0002 -0.0019 -0.0047** -0.0031* -0.0012 -0.0001 0.0019 0.002
[0.0888] [1.7147] [1.3363] [0.1093] [-1.0613] [-2.0037] [-1.6451] [-0.9667] [-0.1189] [0.7863] [0.7332]
 Table VIII
Cancellation Rates Across Fundamentally Linked Securities

Cancels 2ms 5ms 10ms 50ms 100ms 500ms 1000ms

Panel A: Debentures

Buy-side 25,337.70 10,135.08 5,067.54 1,013.51 506.7562 101.3566 50.6805

[1.1525] [1.1525] [1.1525] [1.1525] [1.1525] [1.1526] [1.1526]
HFT 4,590.28 1,836.11 918.0574 183.6115 91.806 18.362 9.1816
[1.1453] [1.1453] [1.1453] [1.1453] [1.1453] [1.1453] [1.1454]
retail 3,489.25 1,395.70 697.8506 139.5706 69.7855 13.9611 6.9816
[1.5685] [1.5635] [1.5635] [1.5685] [1.5685] [1.5690] [1.5692]

Panel B: Preferred Shares

Buy-side 3,926.7934*** 1,570.7184*** 785.3613*** 157.0682*** 78.5345*** 15.7062*** 7.8538***

[6.4050] [6.4050] [6.4050] [6.4052] [6.4052] [6.4056] [6.4064]
HFT 1,446.2056*** 578.4805*** 289.2405*** 57.8487*** 28.9231*** 5.7850*** 2.8927***
[5.4669] [5.4668] [5.4663] [5.4669] [5.4666] [5.4671] [5.4678]
retail 1,339.9715*** 535.9893*** 267.9951*** 53.5995*** 26.7983*** 5.3585*** 2.6796***
[6.1955] [6.1955] [6.1955] [6.1956] [6.1955] [6.1952] [6.1962]

Panel C: Dual-Class Listing

Buy-side 959.2655*** 383.7091*** 191.3604*** 38.3763*** 19.1834*** 3.8322*** 1.9213***

[2.8018] [2.8019] [2.8020] [2.8023] [2.3020] [2.7979] [2.8108]
HFT 392.6936*** 157.0847*** 78.5517*** 15.7277*** 7.8680*** 1.5742*** 0.7940***
[2.9130] [2.9132] [2.9139] [2.9203] [2.9233] [2.9317] [2.9700]
retail 57.1960*** 22.3788*** 11.4412*** 2.2906*** 1.1462*** 0.2279*** 0.1160***
[4.1663] [4.1665] [4.1636] [4.1805] [4.1842] [4.3050] [4.3232]
 Table IX
List of Securities used in SubSamples 1 &2

Non-CrossListed Preferred Debentures Dual Class

TSX60

ARX AIM AIM.PR.A PWF PWF.PR.A AFN AFN.DB AC.B AC.A

BBD.B ALA ALA.PR.A PWF PWF.PR.E ARE ARE.DB ACM.A ACM.B
COS ALA ALA.PR.U PWF PWF.PR.F ARE ARE.DB.A ATD.B ATD.A
CPG AQN AQN.PR.A PWF PWF.PR.G AX.UN AX.DB.F BBD.B BBD.A
CTC.A AX.UN AX.PR.A PWF PWF.PR.H AX.UN AX.DB.U CCL.B CCL.A
ERF AX.UN AX.PR.E PWF PWF.PR.I AYA AYA.DB DII.B DII.A
FM AX.UN AX.PR.U PWF PWF.PR.K CHE.UN CHE.DB QBR.B QBR.A
FTS BBD.B BBD.PR.B PWF PWF.PR.L CHE.UN CHE.DB.A TCL.A TCL.B
HSE BBD.B BBD.PR.C PWF PWF.PR.M CRR.UN CRR.DB.B
L BBD.B BBD.PR.D PWF PWF.PR.O CRR.UN CRR.DB.C
MRU BIR BIR.PR.A PWF PWF.PR.P CRR.UN CRR.DB.D
NA CF CF.PR.A PWF PWF.PR.R CSH.UN CSH.DB.B
POW CF CF.PR.C PWF PWF.PR.S CUF.UN CUF.DB.C
SAP CPX CPX.PR.A REI.UN REI.PR.A CUF.UN CUF.DB.D
SC CPX CPX.PR.C REI.UN REI.PR.C CUF.UN CUF.DB.E
SNC CPX CPX.PR.E RON RON.PR.A CUS CUS.DB
WN CU CU.PR.C TCL.A TCL.PR.D CUS CUS.DB.A
CU CU.PR.D VSN VSN.PR.A CUS CUS.DB.B
CU CU.PR.E WN WN.PR.A CWT.UN CWT.DB.B
CU CU.PR.F WN WN.PR.C D.UN D.DB.H
CU CU.PR.G WN WN.PR.D DC.A DC.DB
CWB CWB.PR.A WN WN.PR.E DIR.UN DIR.DB
DC.A DC.PR.A EIF EIF.DB.A
DC.A DC.PR.B EIF EIF.DB.B
DC.A DC.PR.C EIF EIF.DB.C
EMA EMA.PR.A EIF EIF.DB.D
EMA EMA.PR.C EIF EIF.DB.E
FFH FFH.PR.C EIF EIF.DB.F
FFH FFH.PR.E EXE EXE.DB
FFH FFH.PR.G EXE EXE.DB.B
FFH FFH.PR.I FCR FCR.DB.D
FFH FFH.PR.K FCR FCR.DB.E
FTS FTS.PR.C FCR FCR.DB.F
FTS FTS.PR.E FCR FCR.DB.G
FTS FTS.PR.F FCR FCR.DB.H
FTS FTS.PR.G FCR FCR.DB.I
FTS FTS.PR.H FCR FCR.DB.J
FTS FTS.PR.J HR.UN HR.DB.C
GWO GWO.PR.F HR.UN HR.DB.D
GWO GWO.PR.G HR.UN HR.DB.E
GWO GWO.PR.H HR.UN HR.DB.F
GWO GWO.PR.I HR.UN HR.DB.G
GWO GWO.PR.J HR.UN HR.DB.H
GWO GWO.PR.L INE INE.DB
GWO GWO.PR.M INN.UN INN.DB.B
GWO GWO.PR.N INN.UN INN.DB.C
GWO GWO.PR.P INN.UN INN.DB.D
GWO GWO.PR.Q INN.UN INN.DB.E
GWO GWO.PR.R INN.UN INN.DB.F
HSE HSE.PR.A INN.UN INN.DB.G
IAG IAG.PR.A KEY KEY.DB.A
IAG IAG.PR.C KMP KMP.DB.A
IAG IAG.PR.E KMP KMP.DB.B
IAG IAG.PR.F LW LW.DB
IAG IAG.PR.G MRT.UN MRT.DB.A
IFC IFC.PR.A MSI MSI.DB
IFC IFC.PR.C NKO NKO.NT
IGM IGM.PR.B NPI NPI.DB.A
INE INE.PR.A ORT ORT.DB
INE INE.PR.C PGF PGF.DB.A
L L.PR.A PGF PGF.DB.B
LB LB.PR.D PKI PKI.DB
LB LB.PR.E PKI PKI.DB.A
LB LB.PR.F PRE PRE.DB
NA NA.PR.K PXT PXT.DB
NA NA.PR.L RMM.UN RMM.DB.A
NA NA.PR.M RMM.UN RMM.DB.B
NA NA.PR.N RMM.UN RMM.DB.C
NA NA.PR.O RUS RUS.DB
NA NA.PR.P SPB SPB.DB.C
NA NA.PR.Q SPB SPB.DB.D
NPI NPI.PR.A SPB SPB.DB.E
NPI NPI.PR.C SPB SPB.DB.F
POW POW.PR.A SPB SPB.DB.G
POW POW.PR.B STP STP.DB
POW POW.PR.C TCN TCN.DB
POW POW.PR.D TCN TCN.DB.A
POW POW.PR.E TFI TFI.DB
POW POW.PR.F TFI TFI.DB.A
POW POW.PR.G VSN VSN.DB.C