SSRN Id3281324 PDF

This Draft: September 16, 2019
The Cost of Exposing Large Institutional Orders

to Electronic Liquidity Providers
Robert Battalio Brian Hatch Mehmet Saǧlam

Mendoza College of Business Lindner College of Business Lindner College of Business
University of Notre Dame University of Cincinnati University of Cincinnati
rbattali@nd.edu brian.hatch@uc.edu Mehmet.saglam@uc.edu
(574) 631-9428 (513) 556-7076 (513) 556-9108
Abstract: We use a novel dataset to examine the impact of exposing institutional orders to electronic
liquidity providers (ELPs). We present empirical evidence that marketable pieces of large parent orders are
routed to ELPs, seemingly to avoid paying liquidity fees on exchanges. This routing decision results in
lower net effective spreads for these child orders, but leads to higher execution shortfall for the parent order.
We obtain causal evidence by utilizing the parent orders of investors who disallow the broker to route their
child-orders to ELPs. Our analysis suggests this cost increase is due to information leakage about the parent
order.
Keywords: Electronic Liquidity Providers, Transactions Costs, Order Anticipation

JEL Classification: G12, G14.
We thank Jeff Bacidore, Yashar Barardehi (discussant), Bill Harts, Andy Puckett (discussant), seminar
participants at the Ohio University and University of Notre Dame, and participants at the SEC’s 6th Annual
Conference on Market Regulation and the 2019 WFA meetings. An earlier version of this paper was
circulated under the title, “The Cost of Routing Orders to High Frequency Traders.” Robert Battalio is the
corresponding author.
Electronic copy available at: https://ssrn.com/abstract=3281324

Jeffries LLC estimates that approximately 40% of institutional trading is in the form of orders
directed to broker-dealer algorithms that use different strategies to achieve their trading goals. 1 Algorithms
designed to execute orders at the volume-weighted average price (VWAP) over some period of time
typically slice a large ‘parent’ order into child orders that are released into the market over time. Hendershott
et al. (2011) write that “algorithms often use a mix of active and passive strategies, employing both market
and limit orders” and they rely on smart order routers (SORs) to determine where each child order should
be routed. 2 SORs can route child orders to three different types of trading centers: registered exchanges,
alternative trading systems (ATSs) mainly operating as dark pools and electronic liquidity providers (ELPs)
functioning as proprietary trading firms. 3,4 Sofianos (2007) writes that order routing logic should optimize
execution across all of these trading centers to minimize information leakage.
Liquidity fees and rebates introduce another level of complexity that must be considered when
evaluating the performance of smart routers. As the fees/rebates vary across exchanges, dark pools and
ELPs, the SEC’s Equity Market Structure Advisory Committee notes that broker-dealers who charge fixed
commissions (e.g., do not pass back fees and rebates directly to their customers) have “an incentive to route
to the venue with the highest rebate, rather than diligently search out the venue likely to deliver the best
execution of its customer’s order.” Sofianos et al. (2011) note that brokers who minimize trading costs for
client orders are penalized by a 0.04 basis point reduction in net revenue while clients of brokers that
minimize all-in venue fees are penalized by a 0.26 basis point increase in trading costs. 5
1
See https://www.sec.gov/comments/10-222/10222-321.pdf.
2
Johnson (2010) writes that “an algorithm is a set of instructions for accomplishing a given task.” He notes that the
rules determining the type, the price and quantity for each of these child orders are often based on a mixture of
historical and live market data. The algorithm is responsible for monitoring each child order, adjusting or cancelling
as and when it becomes necessary. See Bertsimas and Lo (1998) for theoretical derivations of optimal trading
strategies that minimize the expected cost of trading a large equity position over time.
3
Registered exchanges are referred to as “lit markets” whereas all other trading centers constitute “dark markets.”
Throughout the paper, we will use the term “dark pool” and “ATS” interchangeably.
4
For example, Credit Suisse operates an SOR that utilizes all of the thirteen exchanges, nineteen ATSs and eight
ELPs. The full list can be accessed at https://www.credit-suisse.com/media/assets/sites/aes/doc/aes-us-order-
handling-guidelines.pdf.
5
In a comment letter dated May 24, 2018, BabelFish Analytics writes that “clients that specifically instruct brokers to
remove rebate-driven trading behaviors from their algorithms achieve significantly lower trading costs that result in
higher returns to their investors.” In a May 16, 2018 website post, Clearpool notes that in a recent analysis of VWAP

A broker operating a SOR can avoid liquidity fees by entering an agreement to send marketable
child orders to one or more ELPs. 6 However, sending one or more child orders directly to an ELP can alert
the ELP that a parent order is being worked. 7 Because institutional orders have price impact (see Keim and
Madhavan (1998)), ELPs can profit from the early knowledge that an institution is seeking to buy or sell a
large block of stock regardless of the institution’s motivation for trading. SEC (2015) writes that by offering
cheap liquidity, ELPs will have a “high position on taker routing tables” that will allow the ELPs “to interact
with the first tranche of a large market order, thus allowing the traders to detect the earliest signs of a
potential price move and quickly adjust their quoting or trading strategies on other markets.” Once aware
that a large parent order is being worked, the trading of ELPs may exacerbate the price impact of the parent
order and lead to inflated trading costs. As described in SEC administrative proceedings against large U.S.
broker-dealers in recent years, many institutions seek to avoid interacting with ELPs when executing large
parent orders. 8 In this paper, we empirically examine the impact that routing child orders to ELPs has on
the execution quality of the parent order.
We use a novel dataset of institutional parent orders which can be routed to the ELPs after obtaining
the permission of the investor in the pre-trade phase. The dataset includes more than 20,000 parent orders
and 2.5 million child order trades occurring between January 1, 2011 and March 31, 2012. The average
parent order is sizable at around $1 million and corresponds to roughly 2% of the volume realized during
algorithms, they found that the fee-sensitive algorithm missed the VWAP and underperformed relative to the fee
agnostic algorithm.
6
See SEC Administrative Proceeding File No. 3-18549, which notes that “when orders were executed by ELPs,
Merrill Lynch avoided the access fees typically charged by exchanges while receiving commissions from customers.”
7
van Kervel and Menkveld (2019) present evidence that HFTs operating on the Nasdaq OMX stock exchange in
Sweeden are unable to detect the presence of a large parent order early in its life. The fact that the HFTs do not have
any information on the counterparty to their trades makes it more difficult to detect institutional order flow. Similarly,
Baxter (2017) writes that “as dark pools are not required to immediately disclose execution or trade information,
predatory firms are – without pinging orders – less able to front-run dark market trades.” The author also notes that
institutions can eliminate or reduce the effect of pinging orders by utilizing a minimum order size in dark pools.
8
See Administrative Proceeding File No. 3-18766, which notes that “many market participants, particularly
institutional firms, sought to avoid trading against HFT” in Citi Match, a dark pool run by Citigroup Global Markets.
Also see Administrative Proceeding File No. 3-18549, which states that certain customers of Merrill Lynch “were
concerned that orders routed to ELPs could be subject to information leakage” and that “certain customers specifically
requested that their orders not be sent to ELPs.” The proceeding also noted that “when orders were executed by ELPs,
Merrill Lynch avoided the access fees typically charged by exchanges while receiving commissions from customers.”

the parent order execution. The dataset identifies where each child order is executed and allows us to exactly
identify the set of ELPs who provide liquidity to each parent order. These ELPs include the largest high-
frequency traders during our sample period: Citadel, D.E. Shaw, Getco, Knight, Sun Trading and Two
Sigma. The average ratio of trade volume executed by ELPs is 5.6% and more than 60% of the parent orders
have at least one child order filled by this group of ELPs.
Our analysis of the trading costs incurred by individual marketable child orders suggests that, net
of exchange fees, ELPs provide the cheapest liquidity when measured by spread costs. While this finding
seemingly rationalizes the routing of marketable child orders to ELPs, it ignores how the underlying stock
price moves during the life of a parent order. Adverse price movements after the exposure of a parent order
to ELPs may increase the total trading costs by offsetting the savings in spread costs. We next examine
when child orders are routed to ELPs and find that a high-share of ELP trades occur in the early stages of
the execution, which provides the ELPs with knowledge that can be used to employ order anticipation
trading strategies. 9 Consistent with assertion that the ELPs in our sample utilized their knowledge that a
parent order was being worked at the expense of that parent order, we find the underlying price of the stock
being purchased (sold) by a sample parent order is more likely to rise (fall) while it is being worked when
the order is exposed to ELPs. More specifically, despite the fact that ELPs offer inexpensive liquidity to
individual child orders, parent orders that source more ELP liquidity have higher average trading costs, all
else being equal. This finding is robust to controlling for venue fees and rebates, proxies for trade
complexity, various control variables that can co-vary with liquidity and stock, day and investor fixed-
effects. In the most conservative case, if 1% of the parent order is routed to ELPs rather than to lit stock
exchanges, trading costs increase by roughly 10%.
Since the routing decision is determined in equilibrium, quantifying a causal effect is empirically
challenging. For example, the broker may be routing the most toxic orders to the ELPs and if these variables
9
SEC (2010) explains that “one example of an order anticipation strategy is when a proprietary firm seeks to ascertain
the existence of one or more large buyers (sellers) in the market and to buy (sell) ahead of the large orders with the
goal of capturing a price movement in the direction of the large trading interest (a price rise for buyers and a price
decline for sellers).” See SEC Release No. 34-61358.

are not properly controlled in the regression, this may bias the estimated impact of the ELP exposure. Our
dataset addresses this concern as it includes a set of investors whose parent orders are never routed to an
ELP. We exploit this variation in the dataset to test the causal relationship and find that parent orders
exposed to ELPs have higher implementation costs compared to matched parent orders whose child orders
are not exposed to ELPs. Further, there is an active investor in the dataset who has child orders executed
by ELPs only in the first part of our sample period, suggesting the investor strategically disallowed the
broker to route its child orders to ELPs after a particular date. We design a difference-in-differences
regression to exploit this variation and document a statistically significant decrease in this investor’s
transaction costs after eliminating the ELP exposure.
We next investigate the specific channels of the increase in trading costs when child orders are
exposed to ELPs. First, we examine whether there is additional cost when child orders are exposed to ELPs
in the early stages and find that all else equal, early exposure to ELPs is costlier than a comparable late
exposure to ELPs. Second, we find that exposing the child orders to multiple ELPs increases the
implementation costs when compared to a same level of exposure with a single ELP. Both of these empirical
findings highlight the information leakage channel associated with ELP exposure.
We provide empirical evidence that suggests routing marketable child orders in a manner that
reduces venue fees is associated with elevated trading costs for parent orders. For brokers that do not pass
through venue fees directly to their clients, this result presents a dilemma as routing orders to minimize
parent order trading costs reduces their net revenue. While Battalio et al. (2016) provide results that suggest
that the use of differential fee schedules by different equity trading venues creates agency problems for
retail brokers, our results indicate this agency problem extends to institutional brokers as well. Thus, our
results provide empirical justification for the SEC’s recent decision to extend broker routing information to
institutional orders. Our results also provide economic justification for the SEC’s decision to allow broker-
dealers to exclude certain types of liquidity providers from their dark pools.
The remainder of this paper is organized as follows. Section II provides a brief literature review.
In Section III, we describe the dataset and provide its summary statistics. We specifically examine the

patterns of the liquidity provided by the ELPs during the lifetime of the parent-order. Section IV provides
a brief rationale on why ELPs are utilized. In Section V, we present multivariate regression results on the
relationship between overall trading costs and ELP exposure. Section VI utilizes the client aversion to ELPs
to establish a causal link between ELP exposure and execution costs. Section VII investigates the
underlying mechanisms that can explain the increase in trading costs and presents empirical evidence
consistent with the information leakage channel. Finally, we conclude in Section VIII.
II. Related literature.
This paper is related to three strands of the microstructure literature. First, there are a number of
studies that examine how the potential conflicts of interest arising from payment for order flow and make-
and-take fees impact trading costs and liquidity provision. Battalio et al. (2016) find that several retail
brokers monetize their order flow by selling market orders to wholesalers and routing their limit orders to
exchanges that offer large liquidity rebates. They present evidence that routing orders in this manner
adversely impacts retail limit order execution quality. Bacidore et al. (2011) compare the performance of
the broker smart routers and find that their performances do not vary significantly for marketable orders.
However, they find that for nonmarketable orders, routers that aim to maximize rebates perform worse than
routers that maximize fill rates.
In a contemporaneous paper, Anand et al. (2019) use FINRA’s proprietary Order Audit Trail
System (OATS) database to empirically examine the relation between broker routing, venue ownership and
execution outcomes. They examine over 350 million institutional orders in 300 stocks that were placed by
43 active institutional brokers in October 2016. The average-sized order in their analysis is 1,348 shares.
Their data do not indicate, however, the identity of the trader or the strategy (e.g., VWAP, implementation
shortfall, percentage of volume) utilized to execute the order. Among other things, the authors present
evidence that the larger institutional brokers route a disproportionate number of child orders to their own
dark pool and that this type of order routing leads to inflated trading costs. In contrast, our study examines
20,335 parent orders seeking to trade an average of 26,000 shares that were placed by 146 unique investors
and were executed via a single broker’s VWAP algorithm. While we are unable to compare and contrast

the effectiveness of the routing strategies used by different institutional brokers, we are able to hold broker
and algorithm constant to examine whether exposing parent orders to ELPs inflates trading costs.
Second, there is a growing literature studying the relationship between institutional trading costs
and activities of high-frequency traders (HFTs). In this strand, the most closely related study is van Kervel
and Menkveld (2019), who use child executions obtained from four institutional investors and trade level
data from the Nasdaq OMX in Sweden that identifies the counterparties to every execution. They present
evidence that HFTs first provide liquidity to and later trade alongside large, informed institutional orders
and show that this type of behavior increases institutional trading costs. In contrast to van Kervel and
Menkveld (2019), our dataset involves the trading activity of a larger investor universe with 146 distinct
clients that utilize a specific broker’s single trading algorithm. The benefit of focusing on one algorithm is
that it removes the variation in execution costs due to heterogeneity across brokers and the aggressiveness
in which their algorithms work orders. van Kervel and Menkveld (2019) construct the net flows of HFT
trading around institutional orders by using the trade reports provided by Nasdaq OMX, a public lit
exchange which has 65% of the market share. The authors exclude 11.5% of their institutional orders
because they generated both buy and sell orders and they are unable to observe HFT activity in dark pools.
We have perfect knowledge of the parent orders with exact start- and end-time which allows us to exactly
compute the statistics around the execution interval.
There are significant differences in contribution between this paper and our study. First, van Kervel
and Menkveld (2019) investigate the cost dynamics of large orders by studying the interaction between the
aggregate HFT trades while we study the impact of the direct information leakage when a child order is
directly routed to an ELP for execution. In van Kervel and Menkveld (2019), it is not clear how the HFT
detects the presence of the large-order whereas the detection is almost immediate in our study due to the
routing of the order. Second, van Kervel and Menkveld (2019) do not claim causality whereas we identify
causal relationship by studying the variation through the investor’s ability to disallow the broker to route
his order to ELPs. Finally, our dataset provides more granular data at the institutional side and has

information about ELP child order trades in the non-lit venues. Conversely, van Kervel and Menkveld
(2019) has high-quality data for HFT trading activity occurring on a stock exchange.
Korajczyk and Murphy (2019) show that HFT liquidity provision is significantly lower for stressful
large institutional orders using order-level data from Investment Industry Regulatory Organization of
Canada, a regulatory organization for Canada’s equity markets. As in the case of van Kervel and Menkveld
(2019), this dataset does not exactly identify how a parent order is split into child orders. HFTs initially
provide liquidity to the large order but then compete with it due to inventory management and back-running.
Our paper complements these studies by directly studying an established routing relationship between a
broker and a set of known ELPs. Since some investors opt out of the routing relationship, we obtain causal
evidence between exposure to ELPs and execution costs. To our knowledge, this is the first paper that
studies the impact of such institutional routing agreements on investors’ trading costs.
Finally, our paper is related to the understanding the market quality implications of aggregate dark
venue executions (e.g., Buti et al., 2011; Comerton-Forde and Putniņš, 2015; Foley and Putniņš, 2016). All
of the ELP trades in our dataset show up as dark pool trades in the publicly available TAQ dataset. Exchange
code ‘D’ is used in the TAQ dataset to identify all trading within ATSs as well as internalized trades at the
broker-dealers. There are a few studies that use this classification to examine the impact of dark pool trades
on market quality (e.g., O’Hara and Ye, 2011; Hatheway et al., 2017; Farley et al., 2018). Given the special
nature of the routing relationships, it is not clear how to aggregate them with the rest of the dark pool trading
activity. The granularity of our dataset allows us to study the liquidity implications of a unique group of
dark-labeled trades that was not possible to study in the past literature.
III. Data and summary statistics
A. Description of the institutional order data.
We compile the data from several sources. Stock returns, volume, outstanding shares and prices
come from the Center for Research in Security Prices (CRSP). Intraday trade and quote data come from the
Trade and Quote (TAQ) database. Institutional parent orders and the corresponding child order executions

are provided by the execution desk of a large investment bank. We next describe this dataset in detail and
provide institutional details about the execution strategy.
The investment bank, hereafter referred to as the broker dealer, is one of the top five providers of
execution services by market share. This data set was originally obtained to study the implications of
investor heterogeneity on the estimation of the price impact (see Sağlam et al., 2019). For this purpose, the
parent order executions in the sample have been selected from an active subset of the broker dealer’s clients.
An investor is considered active if he has at least 100 and at most 500 VWAP parent order executions in
S&P 500 stocks between January 2011 and March 2012 that are fully executed and take at least 10 minutes
to execute. 10
With these criteria, we obtain 22,074 parent orders and 2.6 million child order trades. At the parent
order level, most of the variables are based on the execution horizon. These statistics include order size,
direction of the order (buy or sell), order start and end times, participation rate (the ratio of order size to the
total volume during the trading interval), arrival price (NBBO mid-quote at the start time of the parent order
execution), average execution price, proportional quoted spread and mid-quote volatility. Further, we have
detailed information on the child order executions generated by each parent order. For each child order
execution, the dataset includes the trade time (timestamped to the millisecond), trade size, executing venue,
and the trade price. We do not know whether the executed child order is a market or a limit order and we
do not have unexecuted child orders.
We exclude parent order executions which have less than 5 child order executions or have value
less than $50,000 at the arrival time of the order which correspond to approximately 1,500 parent order
executions. We also exclude an additional 200 executions with missing entries of participation rate, spread,
10
The rationale behind these filters are explained in detail in Sağlam et al. (2019) but we briefly summarize them here
for completeness. First, VWAP orders are passively managed so the performance of the trade would be mainly based
on investor’s timing ability rather than the broker’s skill. Second, S&P 500 stocks are very liquid and relatively
homogeneous subset of the stock universe and it would be unlikely to have an investor trading with insider
information. Third, investors with small number of executions (< 100) are eliminated as their short-term trading skill
may not be estimated reliably with fewer observations. Investors with high number of executions (> 500) are also
filtered to have a balanced data set across clients and prevent any specific investor from fully driving the estimation
procedure. Finally, short executions are often very small in size that may get executed with a single market order.

volatility, or duration. The final sample consists of 20,335 parent orders, 9,856 buy and 10,479 sell orders
placed by 146 distinct investors.
The orders originate from a diverse pool of investors, such as institutional portfolio managers,
quantitative investment funds, internal trading desks and brokers who aggregate their retail order flow. The
dataset only reports the masked identity for each investor, thus it is not possible to know the underlying
trading strategy they are following. For ease of reference, we will refer our investor universe as
“institutional investors.”
We do not know the compensation agreement between the individual clients and our data provider.
However, the broker-dealer informed us that there are two common practices: fixed commission and pass-
through. In the fixed commission scheme, the client pays a fixed fee per share traded and any accumulated
fees or rebates are the broker’s responsibility. In the pass-through scheme, the client pays the fees and
receives the rebates. In our data set, there is no indication that the VWAP algorithm is different across
investors using different compensation schemes. Further, the broker explicitly noted to us that some of their
clients chose this broker to maximize their rebate revenue.
Each of the sample parent orders is executed by a volume-weighted average price (VWAP)
algorithm. The broker informed us that this trading algorithm is the most commonly employed strategy,
constituting roughly 50% of all of the broker’s execution volume. The algorithm slices the parent order into
child orders based on the historical volume curve realized over the past month during the planned execution
horizon.
Institutional clients can shape the parameters of the VWAP algorithm with a limited set of pre-
trade instructions that may affect the ultimate execution cost. Clients select the size and side of the parent
order as well as the start time and targeted order completion time. Clients can also instruct the algorithm to
avoid routing child orders to an ELP or to dark pools. The broker-dealer refers to ELPs as “Liquidity
Partners” in its user interface. While we do not have access to any of these pre-trade instructions, we can
examine where a client’s child order executions occur to infer whether the client consistently avoids ELPs

or specific dark pools. Once the order is initiated, the clients do not have any control over the creation,
duration, and routing of each child order.
Our data provider furnished us with information that allows us to use the Financial Information
Exchange (FIX) protocol tags in the dataset to identify the venue on which each child execution occurs.
These data reveal the presence of six ELPs in our dataset: Citadel, D.E. Shaw, Getco, Knight/Trimark, Sun
Trading, and Two Sigma. As we do not observe any child executions from D.E. Shaw after January 7, 2011,
we effectively have five active ELPs during our sample period.
Finally, we do not know the exact payments that our data provider received from the ELPs for our
broker’s child orders. However, our broker informed us that it did not incur any explicit cost when an ELP
executed a child order. We understand that child orders routed by our data provider to the ELPs were to
either be executed at or within the prevailing National Best Bid and Offer (NBBO) or rejected. Although
we do not have any data on rejection rates, our data provider explicitly told us that its routing agreements
with the ELPs specified that rejection rates cannot be high.
B. Summary Statistics.
Table I contains summary statistics for the sample of 20,335 parent orders placed into our broker’s
VWAP algorithm between January 2011 and March 2012. The average parent order in our sample has a
market value of just over $1 million. Parent order sizes range from $50,000 to $62.864 million. The average
(median) parent order generates 1,278 (60) child order executions. We find that on average, roughly 69%
of the parent order is filled aggressively and 31% is filled passively. 11
In general, trading costs are increasing in a parent order’s participation rate, measured as the ratio
of the parent order’s trading volume to the trading volume of the underlying stock during the parent order’s
life. The average (median) sample parent order consumed 1.8% (0.6%) of the volume traded during the
parent order’s life. The average and the median parent order in our sample had a life of three hours and
thirty-eight minutes (e.g., 0.52 x 6.5 hours).
11
We define a buy (sell) child execution as a passive fill if the transaction price occurs at a price that is lower (higher)
than the midpoint of the execution-time NBBO.
10

On average, exchanges execute 84.5% of the shares in an average parent order, while broker
operated dark pools and ELPs respectively execute 9.6% and 5.6% of the shares. The median parent order
uses ELPs to execute 2.63% of its shares and does not trade any shares in broker operated dark pools.
Overall, 99% of the sample parent orders execute at least one child order on an exchange, while 61% of the
parent orders have at least one child trade with an ELP. Less than half of the parent orders receive a child
order execution in the broker operated dark pool.
In terms of order size and participation rate, our parent order execution data are very similar to
other datasets employed in the broader microstructure literature. For example, Anand et al. (2011) find the
average daily participation rate of the institutions in the Ancerno institutional order database is 2.1% of the
total market volume between 1999 and 2008 (1.0% in 2008). Compared to the Ancerno data, our dataset
has information on the exact start- and end-time of the parent order execution (i.e., order duration), the
interval volume during the execution period of the parent order (that helps us compute participation rate),
the algorithm type (i.e., VWAP), and interval return. 12 More importantly, we have information on the
executed child orders, e.g., the price, quantity and execution venue of the child order, which enables us to
study the impact of interacting with ELPs in the dark.
Korajczyk and Murphy (2019) and van Kervel and Menkveld (2019) also employ institutional
trading data (from Canada and Sweden, respectively) to examine the impact of high-frequency trading on
parent order execution costs. Korajczyk and Murphy (2019) report an average trade size of $2.2 million
with a participation rate of 2.5%. van Kervel and Menkveld (2019) examine roughly 5,000 parent orders
(inferred from individual child order trades) from Sweden during a similar time period as our data and
report an average trade size of $2.2 million corresponding to a participation rate of 3.6%. Overall, these
similar statistics support the representativeness of our institutional trading data.
12
Hu et al. (2017) report that Ancerno dataset has client identifiers only through 2011.
11

C. ELP Liquidity Provision.
Due to the sensitivity of the topic, the data providers were not able to share the exact logic behind
the VWAP algorithm’s decision to route child orders to ELPs. In this section, we examine routing outcomes
to better understand the variation in the algorithm’s use of ELPs within the life of a parent order and across
parent orders placed by different institutional institutions.
All else equal, the potential for information leakage is greatest the earlier that one or more ELPs
become aware that a parent order is being worked. To examine when the VWAP algorithm sources ELP
liquidity, we compute the fraction of shares executed by ELPs in various deciles of the parent order. The
first 10% of the shares executed are in the first decile and the last 10% of the shares executed are in the last
decile. Figure 1 (top left chart) illustrates that the algorithm’s use of ELPs varies throughout the life of a
parent order, as the fraction of volume executed by ELPs has a U-shaped pattern. This suggests that the
VWAP algorithm sources ELP liquidity early in a parent order’s life, which potentially allows the ELPs to
engage in order anticipatory trading strategies throughout most of the parent order’s life.
Figure 1 (top right chart) reveals that the algorithm’s use of exchanges varies across time, with the
percentage of volume executed on exchanges being highest in the middle of a parent order’s life. Figure 1
(bottom center chart) suggests the algorithm’s use of the broker’s dark pool is relatively constant throughout
the parent order’s life. Together, these results suggest there is a negative correlation between the sourcing
of exchange and ELP liquidity. Indeed, we find that the correlation between the percentage of volume
executed by ELPs and exchanges in the first and tenth deciles are -50% and -54%, respectively.
We expect that all of the child orders executed by ELPs were marketable (e.g., the ELPs sold shares
at the offer and bought shares at the bid). Conversely, in an effort to meet or beat the VWAP benchmark
and generate liquidity rebates, we expect the algorithm to route at least some nonmarketable (passive) child
orders to the exchanges. To understand the extent to which the algorithm routes child orders to ELPs after
passive child orders go unexecuted on exchanges, we compute the percentage of volume executed by
exchanges within each decile of the parent order’s life that resulted from a marketable child order. The
results presented in Figure 2 suggest that an increase in the ELPs’ share of child order fills leads to a
12

decrease in the share of passive and aggressive executions in the exchanges. During the later stages of the
parent order execution, it is expected that the passive share would decline. Given that there is no large spike
in aggressive fills in the higher deciles, it appears the algorithm is not simply canceling unexecuted child
orders on exchanges and replacing them with marketable child orders that are routed to the ELPs.
We do not have the pre-trade instructions of the investors so we cannot exactly identify the set of
clients who prohibit the broker from exposing their orders to ELPs. However, we can infer this group of
investors with a simple procedure. Analyzing the ELP trades at the investor level, we find that 4 out of 146
investors have no child orders executed by an ELP. These clients have individually 280, 228, 62 and 40
parent orders consisting of 40,098, 19,997, 1,153 and 2,397 individual child order executions, respectively.
These 610 parent orders are executed on 224 distinct stocks on 217 different dates, so it is almost certain
that these investors have chosen not to source ELP liquidity. Further, three of these clients have also opted
out of dark pool executions, implying a strategic desire to trade in lit venues. In addition, we have two
additional clients who have only one child order executed by an ELP during our sample period. These
clients have individually 180 and 166 parent orders consisting of 64,251 and 13,136 child executions. For
the first of these clients, the lone ELP interaction occurs on the first parent order placed by the client during
our sample period. Collectively for these two clients, the ratio between ELP-exposed parent orders to the
number of total parent orders is only 0.6%. For the remaining 140 clients, this ratio is greater than 12%
with a corresponding mean (median) of 61% (65%). Thus, ELP exposure of these two clients is abnormally
low. Consequently, we will also label their 344 parent orders (excluding the two parent orders with ELP
trades) as Not Exposed to ELPs during their life. Based on this classification, our sample contains 954
parent orders in the Not Exposed group and 19,379 parent orders in the Exposed group.
We now examine the differences in the venues on which child orders are executed for these two
groups of parent orders. First, we study the shares of exchange and broker dealer owned dark pool
executions of the Not Exposed group as a function of volume deciles. Figure 3 confirms our earlier
conjecture about the substitution between exchange and ELP executions when we compare these plots to
those in Figure 1. When a parent order is not exposed to ELPs, we observe the percentage of shares executed
13

on exchanges in the first 9 volume deciles are roughly the same, suggesting that child orders that would
have been routed to ELPs are routed instead to the exchanges when a client prohibits ELP liquidity sourcing.
Next, we look at the parent order statistics for the Not Exposed and Exposed samples. Table II
provides the order characteristics of each client in Not Exposed group and compares the average
characteristics across two samples. We specifically examine the participation rate, the duration of the parent
order, the stock price volatility while the parent order is being worked, the percentage of the parent orders’
dollar volume executed in the broker owned dark pool, and the percentage of the parent orders’ dollar
volume executed by an ELP. On average, the participation rate (duration) of the Exposed parent orders is
76 basis points lower (70 minutes lower) than the orders that were not exposed to ELPs. There is no
difference in volatility across the two sets of parent orders. Given that trading costs are generally increasing
in participation rate, all else equal, these statistics suggest that the Exposed parent orders will have lower
all in trading costs. In Section VI, we will utilize this variation of ELP exposure at the investor level to
examine its impact on implementation costs.
IV. Rationalizing the decision to route individual child orders to ELPs.

We first examine the impact of sourcing ELP liquidity on the cost of individual child order
executions. The exchanges with the largest market share of trading use the make/take pricing model to
arrange trades, which imposes a liquidity fee on liquidity demanding orders. Conversely, based on evidence
provided in the SEC settlement documents, the ELPs do not charge liquidity fees and often offer rebates to
attract marketable orders. Broker operated dark pools tend to charge low liquidity fees and/or offer rebates
to attract marketable orders, but the probability of a child order execution in a broker operated dark pool is
lower than that offered by ELPs. Thus, brokers seeking to minimize liquidity fees will route marketable
child orders to dark pools and ELPs before routing them to exchanges.
Exchanges, dark pools, and ELPs each offer price improvement to marketable orders (i.e.,
execution prices within the prevailing National Best Bid and Offer (NBBO)). Brokers typically negotiate
with ELPs as to the amount of price improvement that their marketable child orders will receive and the
amount the ELPs will pay for the order flow. On exchanges and in dark pools, price improvement typically
14

occurs when a marketable child order interacts with a hidden limit order whose limit price is within the
NBBO. Thus, in contrast to ELP liquidity provision, price improvement is probabilistic on exchanges and
in dark pools.
We use the relative effective spread, measured as the signed difference between the execution price
and the midpoint of the execution-time NBBO normalized by the midpoint of the execution-time NBBO,
to measure execution quality for marketable child orders. We also compute the relative effective spread net
of the typical exchange liquidity fee, since some of the data provider’s clients pay/receive the fees/rebates
generated by their child order executions. To compute the fee-adjusted spreads, we use the fee schedules
reported by the lit exchanges using the baseline values. These schedules report the default fees or rebates
associated with liquidity addition or removal. We assume that each passive (aggressive) child order is
subject to the fee or rebate corresponding to adding (removing) liquidity. Finally, we assume that child
orders executed by ELPs and child orders executed in dark pools do not pay a liquidity fee. Table A.I reports
our fee assumptions for each venue that appears in our data set.
Table III illustrates the average raw and net effective spreads of the marketable child orders
executed in the various venues. We also report aggregate statistics by grouping venues under “inverted
exchanges,” “ELPs,” “make-take exchanges,” “other dark pools” and the “broker’s own dark pool.” As a
group, ELPs do not seem to offer substantially better prices for marketable orders if we rank the venues by
the average nominal effective spread. However, when we account for rebates and fees, make-take
exchanges become substantially costlier than ELPs. Make-take exchanges are roughly 40% more expensive
than ELPs when compared in terms of net effective spread. While the inverted exchanges (e.g., exchanges
that offer rebates for marketable orders) and other dark pools have lower average net effective spreads than
the average ELP, for many stocks and/or in many market conditions these venues are not offering to trade
at the NBBO. Overall, this univariate analysis implies that ELPs can be an attractive venue choice by
allowing the broker to save from take fees for marketable orders. We next examine how the cost of the
parent order execution differs as a function of the exposure to ELPs in a multivariate setting.
15

V. Exposure of parent orders to ELPs and overall trading costs.
A. Measure of parent order execution quality.
Perold (1988) introduced the Implementation Shortfall (IS) measure to quantify the difference
between the performance of a theoretical and the implemented portfolio. Over the years, IS has been
extensively used as a proxy for institutional trading cost (see, for example, Anand et al., 2011, 2013). It is
computed as the normalized difference between the volume-weighted average price of all child orders and
the price of the asset prior to the start of the execution as proxied by the NBBO midpoint. More formally,
for the ith parent order in our sample, we compute IS as follows:
1 𝑁𝑁𝑖𝑖
� ∑ 𝑃𝑃 𝑄𝑄 �− 𝑀𝑀𝑖𝑖,0
𝑄𝑄𝑖𝑖 𝑗𝑗=1 𝑖𝑖,𝑗𝑗 𝑖𝑖,𝑗𝑗
𝐼𝐼𝐼𝐼𝑖𝑖 = 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 (𝑃𝑃𝑃𝑃𝑖𝑖 ) , (1)
𝑀𝑀𝑖𝑖,0
where sign (POi ) is 1 if the ith parent order is a buy order and is -1 if it is a sell order, Qi,j is the size of the
jth child order of the ith parent order, Pi,j is the price at which the jth child order is executed, Ni is the number
of child order executions for the ith parent order, and Mi,0 is the midpoint of the NBBO that was prevailing
when the parent order starts being executed. Similarly, we will use ISAdj to denote the IS that is adjusted
for estimates of venue fees and rebates.
Note that we can decompose the IS into two terms where the first term is the share-weighted
effective spread of the parent order’s child executions normalized by the midpoint of the NBBO prevailing
when the first child order trades and the second term measures the trade-weighted drift in the underlying
stock price over the duration of the parent order’s life: 13
𝑁𝑁 𝑃𝑃𝑖𝑖,𝑗𝑗 − 𝑀𝑀𝑖𝑖,𝑗𝑗 𝑄𝑄𝑖𝑖,𝑗𝑗 𝑁𝑁 𝑀𝑀𝑖𝑖,𝑗𝑗 − 𝑀𝑀𝑖𝑖,0 𝑄𝑄𝑖𝑖,𝑗𝑗

𝑖𝑖
𝐼𝐼𝐼𝐼𝑖𝑖 = 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 (𝑃𝑃𝑃𝑃𝑖𝑖 ) ∑𝑗𝑗=1 � 𝑀𝑀𝑖𝑖,0
𝑖𝑖
� 𝑃𝑃𝑃𝑃 + 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 (𝑃𝑃𝑃𝑃𝑖𝑖 ) ∑𝑗𝑗=1 � 𝑀𝑀𝑖𝑖,0
� 𝑃𝑃𝑃𝑃 , (2)
𝑖𝑖 𝑖𝑖
All else equal, the second term of this equation (which is not affected by exchange fees/rebates) can be
larger when order anticipators become aware of the presence of a parent order.
13
See Holden, Jacobsen, and Subrahmanyam (2014).
16

We use ISAdj to denote the implementation shortfall that has been adjusted for the assumed
liquidity fee or rebate incurred by child orders. We again use the values reported in Table A.I in the Online
Appendix for our fee assumptions for each venue that appears in our data set.
B. ELP exposure and trading costs.
Execution costs can be a function of multiple trade-level and stock-level characteristics, thus, to
test formally whether each ELP proxy is associated with higher costs, we run the following multivariate
regression at the parent order execution level with a rich set of control variables:
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑠𝑠𝑠𝑠𝑠𝑠 𝐶𝐶𝐶𝐶𝑠𝑠𝐶𝐶𝑖𝑖 = 𝛼𝛼 + 𝛽𝛽 𝐸𝐸𝐸𝐸𝑃𝑃 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇𝐸𝐸𝑖𝑖 + 𝜃𝜃1 𝐵𝐵𝑃𝑃𝐵𝐵𝑃𝑃 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇𝐸𝐸𝑖𝑖
+ 𝜃𝜃2 𝑃𝑃𝐶𝐶ℎ𝐸𝐸𝑇𝑇 𝐵𝐵𝑃𝑃 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇𝐸𝐸𝑖𝑖 + 𝜃𝜃3 𝑃𝑃𝑇𝑇𝑠𝑠𝑠𝑠𝑠𝑠𝑃𝑃𝐸𝐸 𝐸𝐸𝐸𝐸𝐸𝐸ℎ. 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇𝐸𝐸𝑖𝑖 + ∑𝑗𝑗 𝛿𝛿𝑗𝑗 𝐶𝐶𝐶𝐶𝑠𝑠𝐶𝐶𝑇𝑇𝐶𝐶𝐶𝐶𝑗𝑗,𝑖𝑖 + 𝜖𝜖𝑖𝑖 , (3)
where Trading Costi is either implementation shortfall or effective spread along with their fee adjusted
versions. ELP Exposurei, BODP Exposurei, Other DP Exposurei and Passive Exch. Exposurei are
respectively the percentage of the ith parent order’s child orders executed by an ELP, broker’s own dark
pool, other dark pools and passive executions in exchanges. Note that we have to exclude Aggressive Exch.
Exposurei from the regression, as the sum of all venue exposure variables is equal to 1. With this regression
model, 𝛽𝛽 × 0.01 will capture the additional cost incurred when 1% of a parent order is executed by ELPs
instead of being filled aggressively in exchanges.
The control variables include the parent order’s participation rate, the logarithm of the arrival price,
the volatility of the NBBO midpoint throughout the duration of the parent order, the duration of the parent
order execution, and turnover. Finally, we include stock, client, and calendar day fixed-effects. Participation
rate and order duration can control for the urgency of the trade and client-fixed effects can control over the
different trading strategies or the skill level of the investor that may be correlated with the price movements
during the execution. Finally, price level and volatility may also affect the execution rates of the ELPs and
can affect the total cost of the order, thus we also include them as controls. This set of control variables is
17

also consistent with the prior literature in examining institutional trading costs. 14 Throughout the paper, we
compute standard errors by clustering at the stock and calendar day level.
Table IV reports the regression results. Consistent with our univariate analysis, ELP Exposure is
not a significant variable in explaining raw effective spreads. Further, the estimated coefficient on ELP
Exposure is negative and statistically significant when the spread costs are adjusted for fees and rebates.
This finding again rationalizes the routing decision to ELPs in the multivariate setting. However, when we
account for the price drift in addition to spread costs in the form of IS and ISAdj, we observe that the
estimated β coefficient is positive and statistically significant at the 0.01 level. Given the ELPs offer
aggressive child orders lower net effective spreads, these results suggest the trade-weighted drift in the
underlying stock price over the duration of the parent order’s life is increasing in ELP Exposure. This result
holds independent of whether or not the fees and rebates generated by the child orders are passed through
to the institutional client.
To understand the exact economic impact of 𝛽𝛽, consider the following example. Suppose that in
the benchmark case, 100% of a parent order execution is traded in exchanges with aggressive orders and
now you consider having an ELP exposure of 5.6% (the mean value of ELP exposure in the data). In this
hypothetical example, compared to the benchmark case, IS would increase by (0.056)(44.45)=2.49 bps.
The economic importance can be also quantified by comparing β to the coefficient on Participation Rate,
one of the most important drivers of execution costs documented in the literature (e.g., Almgren et al.,
2005). We observe that β is of the same order of magnitude as the Participation Rate coefficient. This
suggests that an increase in the percentage of a parent order executed by ELPs has a similar impact on
institutional trading costs as an increase in trading aggressiveness.
Our regressions also analyze the relationship between other dark pool trading activity and
institutional trading costs. In contrast to the strong pattern we identified with ELP fills, we find no evidence
14
For example, Almgren et al. (2005) use participation rate and volatility and van Kervel and Menkveld (2019) use
order duration, volatility, turnover, client and stock fixed effects as controls in their respective analyses of institutional
trading costs.
18

that exposing parent orders to the broker’s dark pool inflates trading costs. Similarly, the coefficient on
other dark pool exposure is negative but insignificant. Overall, the different signs of the coefficients
highlight the potential heterogeneity of the effects of dark pool executions and call for further research. For
example, there are a few studies that use TAQ classification to examine the impact of dark pool trades on
market quality (e.g., O’Hara and Ye, 2011; Hatheway et al., 2017; and Farley et al., 2018) but the TAQ
data cannot differentiate between these two different types of fills. One takeaway from this analysis is that
the potential positive effect of dark pools is biased downwards if one classifies the dark pool trades from
the TAQ database.
C. Robustness Checks.
GETCO accounts for approximately 70% of the trades (60% of the dollar volume) that are executed
by ELPs. To examine whether the results presented in Table IV are driven by exposure of parent orders to
GETCO, we decompose our measure of ELP trading activity into two components: the percentage of child
orders executed by GETCO and the percentage of child orders executed by ELPs other than GETCO and
re-estimate Equation 3. We present the results in Table A.II in the Online Appendix. As is the case when
all ELP child executions are aggregated, we find that exposure to GETCO and exposure to the other ELPs
is associated with inflated trading costs for the parent order. This suggests child order executions by
GETCO are not driving our results.
IS uses the NBBO midpoint prior to the execution of the first child order as a benchmark price to
compute the execution cost measure. One popular ex-post benchmark for the average execution price is the
volume-weighted average price (VWAP) during the parent-order’s trading interval. Madhavan (2002)
argues that VWAP slippage may not be a reliable proxy of execution quality, as early aggressive trading
would significantly affect the realized VWAP. Nevertheless, we examine the robustness of our findings
using VWAP slippage as a measure of overall trading costs. Table A.III in the Online Appendix reports the
main regression results using VWAP slippage and its fee-adjusted version. In both cases, the estimated
coefficient on ELP Exposure is highly significant implying the robustness of our results with respect to a
different proxy.
19

VI. Evidence from client aversion to ELPs
Our findings in the previous section point to a strong positive correlation between proxies of ELP
trading activity and parent order execution costs which may not directly imply causal relationship. There
may be an omitted variable that is correlated with execution costs and ELP trading activity which may bias
the regression coefficients. In order to mitigate this concern, we pursue two identification strategies. First,
we utilize the set of 954 parent order executions from six institutional investors who have no more than one
child order executed by an ELP during our sample period. We assume that these investors chose not to
directly source ELP liquidity for their parent orders. Second, we examine the trading costs incurred by a
particular institutional client before and after she seemingly stops allowing the broker to route her child
orders to ELPs.
A. Clients without ELP exposure.
Recall that Table II provides the order characteristics of the distinct six clients who have zero
(clients 1 through 4) or abnormally small exposure to ELPs (clients 5 and 6). First, we observe that each
client has quite different order characteristics. For example, client 6 has a large participation rate and
relatively short duration, which may lead to higher execution costs at the parent order level. Conversely,
client 2 has a relatively small participation rate but longer duration, which can lead to lower trading costs.
In contrast to clients 4, 5 and 6, who execute between 7% and 15% of their child orders in the our data
provider’s dark pool, clients 1, 2, and 3 have no exposure to the broker’s dark pool. One takeaway from
this heterogeneity is that there does not seem to be a common pattern across these six investors aside from
the fact that they avoid ELPs. 15
The clients choosing not to expose their orders to ELPs, may be following different trading
strategies with regards to which stock to buy or sell and when to trade. Ultimately, we would like to compare
15
Although the investor identities are masked, the description field regarding the parent order identifier may signal
the true identity of the investor. For three of these six clients, we observe that texts like “BNP,” “SUSQ” and “FID”
appear suggesting that these orders may be coming from BNP Paribas, Susquehanna International Group and Fidelity
Investments, respectively. For the remaining clients, we observe “AMC,” “AMMN” and “TXP” appearing in their
parent order identifiers. Overall, this group of investors does not seem to be part of the broader ELP group. We thank
Andy Puckett for suggesting this check.
20

the execution costs of two investors who follow identical trading strategies but only differ in their exposure
to ELPs. We employ an exact matching procedure to obtain such a sample of treated and control groups.
For each parent order i in the Not Exposed group, we search for a matched parent order in the Exposed
group (without replacement) using the following algorithm. First, the executed stock and the trade direction
(buy or sell) must be the same. Second, the dates of the executions of the ith parent order and the matched
parent order must be within five trading days. We screen the Exposed group using these criteria. If there is
no match, ith parent order will not be matched. If there are multiple matches, we use the parent order with
the closest number of shares to be executed. We were able to match 364 (out of 954) parent orders using
this exact procedure. We will label these groups of parent orders, NoELP and ELP, respectively. This exact
matching procedure addresses the potential timing and stock-selection ability of the clients in the Not
Exposed group.
Next, we compare the characteristics of the NoELP and ELP matched samples. Table V provides
the detailed comparison. We specifically check BODP and Passive Exchange exposure, participation rate,
logarithm of the arrival price, volatility, and quoted spread. We observe that the differences in logged price,
volatility, and quoted spread are not statistically significant implying the success of the exact matching for
these characteristics. We observe that parent order executions in the NoELP group have lower BODP
exposure, and larger participation rate, which would typically lead to higher execution costs.
We now test whether parent orders in the NoELP group have lower implementation shortfall
compared to the ones in the ELP group by running the following multivariate regression:
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑠𝑠𝑠𝑠𝑠𝑠 𝐶𝐶𝐶𝐶𝑠𝑠𝐶𝐶𝑖𝑖 = 𝛼𝛼 + 𝛽𝛽 𝑁𝑁𝐶𝐶𝐶𝐶 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇 𝐶𝐶𝐶𝐶 𝐸𝐸𝐸𝐸𝑃𝑃𝑠𝑠𝑖𝑖 + ∑𝑗𝑗 𝛿𝛿𝑗𝑗 𝐶𝐶𝐶𝐶𝑠𝑠𝐶𝐶𝑇𝑇𝐶𝐶𝐶𝐶𝑗𝑗,𝑖𝑖 + 𝜖𝜖𝑖𝑖 , (4)
where the cost measure is the raw and net IS and the control variables include BODP and Passive Exchange
exposure, participation rate, and order duration.
Table VI reports the regression results. In both specifications, the coefficient on
𝑁𝑁𝐶𝐶𝐶𝐶 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇 𝐶𝐶𝐶𝐶 𝐸𝐸𝐸𝐸𝑃𝑃𝑠𝑠 is negative and highly significant. The estimates point to an additional cost of
approximately 12 bps in the parent order executions with ELP trading relationship. Overall, this evidence
21

provides a causal interpretation to the observed positive correlation between trading costs and ELP
exposure. Our results are robust to different matching procedures (e.g., propensity score matching).
B. The client that stops exposing its orders to ELPs.
Large institutional clients like Fidelity and Vanguard can experiment with pre-trade instructions to
maximize their execution performance. These clients have the ability to statistically examine the relative
performance of parent orders that are exposed to ELPs. For each client in our sample, we evaluate the time
series of child executions to determine whether the client either begins or stops exposing it orders to ELPs.
None of the clients in our analysis avoid ELPs in the early part of the sample period and source ELP liquidity
in the latter part of the sample period. We do, however, find one institutional investor who seemingly allows
its orders to source ELP liquidity in the early part of our sample period, but refrains in the latter part - none
of the client’s 40,000+ child orders generated by more than 100 parent orders executed after March 18,
2011 are routed to ELPs. This sharp switch suggests the client has prohibited the broker from routing the
child orders to the ELPs. Furthermore, the client’s child orders execute in the broker’s and in other dark
pools throughout the sample period. The fact that the client appears to stop sourcing ELP liquidity is
consistent with what one would expect if the client detected that parent orders with ELP exposure had
higher trading costs than those that were not exposed to ELPs. In this section, we test this hypothesis.
Between January 2011 and March 2011 this client has 40 parent orders (5,239 child order trades)
that have sporadic exposure to ELPs. On March 18, 2011, in his last parent order execution during this
period, there is a child order executed by an ELP suggesting that the client consents for ELP exposure.
During this period, 67.5% of the client’s parent order executions have ELP exposure and 4.2% of the trades
are executed in the broker’s own dark pool. The client’s next parent order execution after this period occurs
on April 6, 2011 and starting from this execution, the client’s child orders are not routed to ELPs in 111
distinct parent orders consisting of 40,479 child order trades. The client still continues to trade in the dark
pools as 4.1% of the client’s orders are executed in the broker’s own dark pool in this no-ELP period. These
statistics strongly imply that the client switched its instructions on ELP routing after March 18, 2011.
22

Given the switching decision, one may wonder whether the client changes his overall trading
strategy. We compare various parent order and stock-level statistics between the pre- and post-switch period
and report the statistics in Table A.IV in the Online Appendix. We specifically check BODP and Passive
Exchange exposure, participation rate, logarithm of the arrival price, volatility, and quoted spread. Since
volatility and quoted spread can change significantly, we check whether these statistics for the client remain
stable over time by computing the z-scores of his executed stock volatility and quoted spread by normalizing
with the daily means and standard deviations of the statistics. We find that apart from the ELP exposure,
all of the parent order and stock-level characteristics are very similar suggesting that the client does not
seem to follow a different trading strategy in the post-switch period.
Given that the same client switches from active ELP exposure to zero ELP exposure, we can design
a difference-in-differences framework to formally test the impact of ELP exposure on execution costs. Let
Switching Client Order take a value of 1 for parent order executions generated by the switching client’s
orders and zero otherwise. Post 20110318 is an indicator variable that takes a value of 1 for executions after
March 18, 2011. Switching Client Order after 20110318 is an indicator variable that takes a value of 1 if
the execution is from the switching client and occurs after March 18, 2011. In the control group, we can
use the parent order executions in the Exposed group that have ELP exposure throughout the data period.
We then run the following diff-in-diff regression using the same set of control variables and stock, day and
client fixed effects as in Equation (5):
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑠𝑠𝑠𝑠𝑠𝑠 𝐶𝐶𝐶𝐶𝑠𝑠𝐶𝐶𝑖𝑖 = 𝛼𝛼 + 𝛽𝛽 𝐼𝐼𝑆𝑆𝑠𝑠𝐶𝐶𝐸𝐸ℎ𝑠𝑠𝑠𝑠𝑠𝑠 𝐶𝐶𝐶𝐶𝑠𝑠𝐸𝐸𝑠𝑠𝐶𝐶 𝑃𝑃𝑇𝑇𝑇𝑇𝐸𝐸𝑇𝑇 𝑇𝑇𝑎𝑎𝐶𝐶𝐸𝐸𝑇𝑇 20110318𝑖𝑖 + 𝜅𝜅 𝑃𝑃𝐶𝐶𝑠𝑠𝐶𝐶20110318𝑖𝑖 +
𝜃𝜃 𝐼𝐼𝑆𝑆𝑠𝑠𝐶𝐶𝐸𝐸ℎ𝑠𝑠𝑠𝑠𝑠𝑠 𝐶𝐶𝐶𝐶𝑠𝑠𝐸𝐸𝑠𝑠𝐶𝐶 𝑃𝑃𝑇𝑇𝑇𝑇𝐸𝐸𝑇𝑇𝑖𝑖 + ∑𝑗𝑗 𝛿𝛿𝑗𝑗 𝐶𝐶𝐶𝐶𝑠𝑠𝐶𝐶𝑇𝑇𝐶𝐶𝐶𝐶𝑗𝑗,𝑖𝑖 + 𝜖𝜖𝑖𝑖 (5)
Table VII reports the regression results using raw and net implementation shortfall as the cost
metrics. We find that 𝛽𝛽 is negative and statistically significant for both measures implying that after the
switch, the client realized lower trading costs. Given that the client seems to be following the same trading
strategy except disallowing ELP exposure, these findings provide further causal support for the cost
increase in parent order executions with exposure to ELPs.
23

VII. How do ELPs increase transaction costs?
In this section, we conduct two additional analyses to further explore whether the evidence is
consistent with the claim that ELPs use their knowledge that an institutional order is being worked to engage
in order anticipatory trading strategies to the detriment of the institutional order. Since early interaction
with a parent order gives ELPs more time to exploit this information, we first examine whether intense ELP
interaction with a parent order early in its life is associated with inflated trading costs. As we expect
information leakage to be more likely when multiple ELPs are aware that an order is being worked, we next
examine whether interacting with multiple ELPs is associated with higher trading costs.
A. Early versus late ELP exposure.
In Figure 1, we observe the U-shaped pattern for ELP fills. Given the positive relationship with
execution costs and ELP exposure, we expect that early exposure to ELPs would lead to higher execution
costs. This finding would be consistent with the information leakage channel. To investigate this
explanation formally, we decompose our ELP trading activity proxy into three components: the percentage
of child executions by ELPs in the first decile (ELP Exposure D1), the tenth decile (ELP Exposure D10),
and in deciles 2 through 9 (ELP Exposure D2_D9). By construction, the sum of these variables equals ELP
Exposure.
To examine this hypothesis, we re-run our main specification (see equation 3) with the three
decomposed ELP trading variables replacing ELP Exposure:
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑠𝑠𝑠𝑠𝑠𝑠 𝐶𝐶𝐶𝐶𝑠𝑠𝐶𝐶𝑖𝑖 = 𝛼𝛼 + 𝛽𝛽𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸 𝐸𝐸𝐸𝐸𝑃𝑃 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇𝐸𝐸 𝐵𝐵1𝑖𝑖 + 𝛽𝛽𝐿𝐿𝐸𝐸𝐿𝐿𝐿𝐿 𝐸𝐸𝐸𝐸𝑃𝑃 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇𝐸𝐸 𝐵𝐵10𝑖𝑖 +
𝛽𝛽𝑀𝑀𝑖𝑖𝑀𝑀𝑀𝑀𝐸𝐸𝐿𝐿 𝐸𝐸𝐸𝐸𝑃𝑃 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇𝐸𝐸 𝐵𝐵2_𝐵𝐵9𝑖𝑖 + ∑𝑗𝑗 𝛿𝛿𝑗𝑗 𝐶𝐶𝐶𝐶𝑠𝑠𝐶𝐶𝑇𝑇𝐶𝐶𝐶𝐶𝑗𝑗,𝑖𝑖 + 𝜖𝜖𝑖𝑖 , (6)
where the cost measure is the raw and net IS and the control variables include BODP and Passive Exchange
Exposure, Participation Rate, and Order Duration. We expect that βEarly > βLate if ELPs are utilizing order
anticipatory trading strategies.
Table VII reports the regression results and verifies our conjecture. Using a Wald test, we confirm
that βEarly is statistically greater than βLate for each specification. Consistent with our hypothesis, we also
24

observe that βMiddle is statistically larger than βLate. Overall, these findings point to higher information
leakage with earlier ELP fills.
B. Competition between ELPs.
Hendershott and Madhavan (2015) write that “revealing trading intentions to many potential
counterparties can lead to costly information leakage.” Recall that 61.1% of all parent orders have at least
one child order executed by an ELP. We compute the breakdown of this statistic with regards to distinct
ELPs that provide liquidity to a parent order and find that 25.3% of the parent orders interact with one ELP,
21.4% trade with two ELPs, 11.7% source liquidity from three ELPs, 2.4% transact with four ELPs, and
0.25% trade with five of the six sample ELPs. None of our parent orders interact with each of the six sample
ELPs. If ELPs engage in order anticipatory trading strategies, one might expect transactions costs to be
higher when a parent order is exposed to multiple ELPs.
Conversely, Bessembinder et al. (2016) extends the model of Brunnermeier and Pedersen (2005)
for resilient markets in which the immediate price impact of trades may be transitory. In this model, in
addition to the same-side trading before the liquidation, the strategic traders trade in the opposite direction
compared to the direction of the parent order. This theory illustrates that this opposite-side trading can
decrease the liquidator’s transitory price impact. This benefit to the liquidator from strategic trading persists
at any level of market resiliency if there are multiple strategic traders. Given that there are six distinct ELPs
in the data, if the broker routes orders to multiple ELPs, this theory would suggest a potential decrease in
execution costs compared to the case where there is a single ELP taking advantage of the order flow
information.
To examine the impact that exposing parent orders to multiple ELPs has on trading costs, we run
the following regression:
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑠𝑠𝑠𝑠𝑠𝑠 𝐶𝐶𝐶𝐶𝑠𝑠𝐶𝐶𝑖𝑖 = 𝛼𝛼 + 𝛾𝛾1 𝑀𝑀𝐶𝐶𝑇𝑇𝐸𝐸 𝐶𝐶ℎ𝑇𝑇𝑠𝑠 2 𝐸𝐸𝐸𝐸𝑃𝑃𝑠𝑠 + 𝛾𝛾2 𝑀𝑀𝐶𝐶𝑇𝑇𝐸𝐸 𝐶𝐶ℎ𝑇𝑇𝑠𝑠 3 𝐸𝐸𝐸𝐸𝑃𝑃𝑠𝑠 + 𝛽𝛽 𝐸𝐸𝐸𝐸𝑃𝑃 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇𝐸𝐸𝑖𝑖 +
𝜃𝜃1 𝐵𝐵𝑃𝑃𝐵𝐵𝑃𝑃 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇𝐸𝐸𝑖𝑖 + 𝜃𝜃2 𝑃𝑃𝐶𝐶ℎ𝐸𝐸𝑇𝑇 𝐵𝐵𝑃𝑃 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇𝐸𝐸𝑖𝑖 + 𝜃𝜃3 𝑃𝑃𝑇𝑇𝑠𝑠𝑠𝑠𝑠𝑠𝑃𝑃𝐸𝐸 𝐸𝐸𝐸𝐸𝐸𝐸ℎ. 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝑠𝑠𝐸𝐸𝑇𝑇𝐸𝐸𝑖𝑖 +
∑𝑗𝑗 𝛿𝛿𝑗𝑗 𝐶𝐶𝐶𝐶𝑠𝑠𝐶𝐶𝑇𝑇𝐶𝐶𝐶𝐶𝑗𝑗,𝑖𝑖 + 𝜖𝜖𝑖𝑖 , (7)
25

where the cost measure is IS and ISAdj and the control variables include ELP Exposure, BODP Exposure
and Passive Exchange Exposure, Participation Rate, and Order Duration. We also include the binary
variables, More than 2 (3) ELPs which take a value of one if the parent order is exposed to more than 2 (3)
distinct ELPs, to examine the incremental impact of routing to multiple ELPs.
Table IX reports the regression results. We find that the coefficients on More than 2 ELPs and
More than 3 ELPs are positive and significant. These findings suggest that after controlling for the ratio of
ELP child order executions, having additional competing ELPs does not decrease execution costs. These
findings are instead more consistent with higher information leakage when the orders are routed to multiple
ELPs.
VII. Conclusion.
This paper examines the costs of routing relationships between a broker and ELPs from the
perspective of an institutional investor. Typically, brokers do not pass the fees and rebates associated with
executing parent orders directly through to their institutional clients. This creates an incentive for brokers
to reduce the fees and/or increase the rebates generated when working institutional orders. One common
way for brokers to reduce fees is to route marketable child orders to ELPs, who provide child order
executions at or within the prevailing NBBO and, unlike many exchanges, do not charge liquidity fees. A
potential cost of this routing strategy is that ELPs become aware that a large parent order is being worked
in the market place and can use this information to engage in profitable order anticipatory trading strategies
to the detriment of the parent order. We use parent orders executed by a single VWAP algorithm to examine
whether the sourcing of ELP liquidity ultimately leads to inflated trading costs.
We first present evidence consistent with the idea that routing marketable child orders to ELPs
reduces net effective spreads relative to the major exchanges that charge positive liquidity fees. These
results, however, ignore the potential impact that ELPs can have on the underlying stock price over the life
of the order. We next examine how exposing parent orders to ELPs effects the implementation shortfall for
parent orders. We obtain robust evidence suggesting parent orders that source liquidity from ELPs incur
higher transactions costs than those that avoid ELPs. This finding is robust to controlling for venue fees
26

and rebates along with other control variables including stock, day and investor fixed-effects. In the most
conservative case, our results suggest that if 1% of a parent order is routed to ELPs instead of the stock
exchanges, the implementation shortfall for the order will increase by roughly 12%.
We establish a causal relationship by examining the parent order executions from a set of investors
who strategically opt out of the routing agreement with the ELPs. We exploit this variation in the dataset
to test the causal relationship between ELP exposure and execution costs. Further, we also exploit the fact
that one of our sample institutions seemingly ceases to allow our data provider from sourcing ELP liquidity
in the latter part of our sample period. Using both of these identification strategies, we find statistically
significant differences in execution costs between ELP-exposed and ELP-excluded parent orders.
Investigating the underlying mechanism for the cost increase in detail, we examine the cost of exposing the
child orders to ELPs in different stages of a parent order’s life and find that early exposure is substantially
costlier. We also present evidence that exposing a parent order to multiple ELPs leads increases the parent
order’s implementation shortfall. Each of these findings is consistent with the information leakage channel.
More generally, our results suggest that there is a cost to sourcing liquidity directly from ELPs.
When an ELP interacts with a single child order on an exchange, it does not know where the order
originated. The results of van Kervel and Menkveld (2019) suggest that ELPs can detect the presence of an
institutional order (with error) by interacting with multiple child orders on lit exchanges. When an
institutional broker sources liquidity from an ELP with whom it has an ongoing relationship, the ELP
immediately becomes aware that a parent order is being worked. This is because the ELP can see the broker
from which the order originated. Thus, while the relationship allows the broker to obtain low cost liquidity
from ELPs for individual child orders, it also allows ELPs to more quickly ascertain that a large institutional
order is being worked and to use that knowledge to earn trading profits at the expense of the institutional
order.
27

References
Almgren, Robert, Chee Thum, Emmanuel Hauptmann, and Hong Li, 2005. Direct estimation of equity
market impact, Risk 18, 57-62.
Anand, Amber, Paul Irvine, Andy Puckett, and Kumar Venkataraman, 2011. Performance of institutional
trading desks: An analysis of persistence in trading costs, The Review of Financial Studies 25, 557-598.
Anand, Amber, Paul Irvine, Andy Puckett, and Kumar Venkataraman, 2013. Institutional trading and stock
resiliency: Evidence from the 2007-2009 financial crisis, Journal of Financial Economics 108, 773-797.
Bacidore, Jeff, Hernan Otero, and Alak Vasa, 2011. Does smart routing matter?, Journal of Trading 6, 32-
37.
Battalio, Robert, Shane Corwin, and Robert Jennings, 2016. Can brokers have it all? On the relation
between make-take fees and limit order execution quality, Journal of Finance, 2193-2237.
Baxter, Brian, 2017. The securities black market: Dark pool trading and the need for a more expansive
Regulation ATS-N, Vanderbilt Law Review 70, 311-337.
Bertsimas, Dimitris and Andrew Lo, 1998. Optimal control of execution costs, Journal of Financial
Markets 1, 1-50.
Bessembinder, Hendrick, Allen Carrion, Laura Tuttle, and Kumar Venkataraman, 2016. Liquidity,
resiliency and market quality around predictable trades: Theory and evidence, Journal of Financial
Economics 121, 142-166.
Brunnermeier, Markus and Lasse Pedersen, 2005. Predatory trading, Journal of Finance 60, 1825-1863.
Buti, Sabrina, Barbara Rindi and Ingrid M. Werner, 2011. Diving into dark pools, University of Toronto
working paper.
Comorton-Forde, Carole and Talis Putniņš, 2015. Dark trading and price discovery, Journal of Financial
Economics 118, 70-92.
Farley, Ryan, Eric Kelley, and Andy Puckett, 2017. Dark trading volume and market quality: A natural
experiment, University of Tennessee working paper.
Foley, Sean and Talis Putniņš, 2016. Should we be afraid of the dark? Dark trading and market quality,
Journal of Financial Economics 122, 456-481.
Hatheway, Frank, Amy Kwan, and Hui Zheng, 2017. An empirical analysis of market segmentation on U.S.
equity markets, Journal of Financial and Quantitative Analysis 52, 2399-2427.
Holden, Craig, Stacey Jacobsen, and Avanidhar Subrahmanyam, 2014. The empirical analysis of liquidity,
Foundations and Trends in Finance 8, 263-365.
Hu, Gang, Koren Jo, Yi Wang, and Jing Xie, 2018. Institutional trading and Abel Noser Data, Journal of
Corporate Finance 52, 143-167.
28

Keim, Donald and Ananth Madhavan, 2019. The cost of institutional equity trades, Financial Analysts
Journal 54, 50-69.
van Kervel, Vincent and Albert Menkveld, 2019. High-frequency trading around large institutional orders,
Journal of Finance 74, 1091-1137.
Korajczyk, Robert and Dermot Murphy, 2019. High-frequency market making to large institutional trades,
Review of Financial Studies 32, 1034-1067.
Hendershott, Terrence, Charles Jones, and Albert Menkveld, 2011. Does algorithmic trading improve
liquidity?, Journal of Finance 66, 1-33.
Hendershott, Terrence and Ananath Madhavan, 2015. Click or call? Auction versus search in the over-the-
counter market, Journal of Finance 70, 419-447.
Johnson, Barry, 2010. Algorithmic Trading & DMA: An Introduction to Direct Access Trading Strategies,
Myeloma Press, London, England.
Madhavan, Ananth, 2002. VWAP Strategies, Transaction Performance, 32-38.
O’Hara, Maureen and Mao Ye, 2011. Is market fragmentation harming market liquidity?, Journal of
Financial Economics 100, 459-474.
Perold, Andre, 1988. The implementation shortfall: Paper versus reality, Journal of Portfolio Management
14, 4-9.
Saǧlam, Mehmet, Ciamac Moallemi, and Michael Sotiropoulos, 2019. Short-term trading skill: An analysis
of investor heterogeneity and execution quality, Journal of Financial Markets 42, 1-28.
SEC, 2010. Concept Release on Equity Market Structure, SEC Release 34-61358.
SEC, 2015. Memorandum from the SEC’s Division of Trading and Markets to the SEC’s Market Structure
Advisory Committee Regarding Maker-Taker Fees on Equity Exchanges, October 20.
Sofianos, George, 2007. Dark pools and algorithmic trading, in Algorithmic Trading, A Buy-Side Handbook
(2nd Edition), The Trade, July 7.
Sofianos, George, Juan Xiang, and Ali Yousefi, 2011. Smart routing, all-in shortfall, and optimal order
placement, Goldman Sachs Street Smart 42, 1-9.
29

Figure 1. Share of ELPs, exchanges and the broker’s dark pool in executed volume deciles.
Notes: We plot the percentages of ELP (top left), lit-venue (top right) and the broker’s dark pool (bottom) executions in various deciles of
executed volume, e.g., ELPs’ share of each 1000 shares for a parent order consisting of 10,000 shares.
10 10
9 9
8 8
ELP Share (%)
ELP Share (%)

7 7
6 6
5 5
4 4
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Volume Deciles Volume Deciles
12
10
Broker DP Share (%)
1 2 3 4 5 6 7 8 9 10
Volume Deciles
30

Figure 2. Share of passive and aggressive fills in executed volume deciles.
Notes: We plot the ratios of passive (left) versus aggressive (right) lit-venue executions in various deciles of executed volume.
30
55
LIT (Aggressive) Share (%)

LIT (Passive) Share (%)
28
26
50
24
22 45
20
18 40
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
31

Figure 3. Share of exchanges and the broker’s dark pool in executed volume deciles when ELPs are not sourced.
Notes: We plot the ratio of lit exchanges (left) and the broker’s dark pool (bottom) executions in various deciles of executed volume in the Not
Exposed group.
97
96
6
Broker DP Share (%)

95
LIT Share (%)
94 4
93
92 2
91
90 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
32

Table I. Summary statistics
Notes: This table presents descriptive statistics for the 20,335 parent order executions placed with a large institutional broker between January 2011
and March 2012. There are 9,865 buy orders and 10,479 sell orders in 498 of the S&P 500 stocks. Participation rate is measured as the ratio of the
parent order’s trading volume to the overall trading volume of the underlying stock over the period of time that the parent order is being worked.
Average Min. 25th Pctl. Median 75th Pctl. Max.
Value of parent order ($ millions) 1.015 0.050 0.131 0.343 1.001 62.864
# of child order executions 127.8 5 26 60 148 4,533
% of parent order executed w/ aggressive child executions (trades) 69.31 0.00 60.00 70.00 79.17 100.00
% of parent order executed w/ passive child executions (trades) 30.69 0.00 20.83 30.00 40.00 100.00
Participation rate 0.018 0.000 0.002 0.006 0.0019 0.521

Parent order duration (% of trading day) 0.52 0.03 0.16 0.52 0.90 1.00
% of parent order executed on an exchange (trades) 84.51 0.00 80.00 91.38 97.22 100.00
% of parent order executed in a broker operated dark pool (trades) 9.57 0.00 0.00 0.00 10.00 100.00
% of parent order executed by an ELP (trades) 5.60 0.00 0.00 2.63 7.69 100.00
% of parent orders with an execution by an exchange 99.27 0.00 100.00 100.00 100.00 100.00
% of parent orders with an execution in a broker operated dark pool 48.27 0.00 0.00 0.00 100.00 100.00
% of parent orders with an execution by an ELP 61.10 0.00 0.00 100.00 100.00 100.00
33

Table II. Descriptive statistics for parent orders that avoid ELPs
Notes: Participation Rate is measured as the ratio of the parent order’s trading volume to the overall trading volume of the underlying stock over the
period of time that the parent order is being worked. Volatility is measured as the volatility of the midpoint of the NBBO over the parent order’s
life. BODP Exposure is the percentage of a parent order’s child executions that occur in the broker’s own dark pool. ELP Exposure is the percentage
of a parent order’s number of trades with ELPs. The table below presents averages for parent orders placed by each of the six clients that seemingly
avoid sourcing ELP liquidity, the entire sample of parent orders placed by clients that appear to avoid sourcing ELP liquidity, and the sample of
parent orders placed by clients that allow their parent orders to interact with ELPs.
Not
Client 1 Client 2 Client 3 Client 4 Client 5 Client 6 Exposed Diff
Exposed
# of parent orders 280 228 62 40 179 165 954 19,379 18,427
Participation rate (%) 1.82 0.84 0.95 1.02 2.03 7.50 2.52 1.76 0.76**
Parent order duration 0.31 0.70 0.04 0.48 0.18 0.25 0.36 0.53 -0.18***
Volatility (%) 1.81 1.38 1.84 1.89 2.53 1.40 1.78 1.49 0.28
BODP Exposure 0.00 0.00 0.00 12.46 9.22 5.90 3.27 9.87 -6.60***
ELP Exposure 0.00 0.00 0.00 0.00 0.00 0.00 0.00 5.87 -5.87***
34

Table III. Relative and fee-adjusted net effective spreads for marketable child orders by venue
Notes: Appendix Table I reports our fee assumptions for each venue that appears in our data set.
# of Child Effective Spread Net Effective Spread
Venue Type
Executions Average (bps.) Std. Error Average (bps.) Std. Error
BYX Stock Exchange: BYX Inverted Exch. 2,776 2.15 0.04 1.94 0.04
Nasdaq BX Stock Exchange: BX Inverted Exch. 2,668 2.55 0.05 2.33 0.05
Citadel: CDRG ELP 15,199 1.88 0.01 1.88 0.01
Knight Securities: TRIM ELP 5,386 1.90 0.02 1.90 0.02
Knight Securities: NITE ELP 1,214 2.03 0.05 2.03 0.05
Getco: GFLO ELP 97,415 2.29 0.01 2.29 0.01
D.E. Shaw: SHAW ELP 555 2.61 0.12 2.61 0.12
Two Sigma: SOHO ELP 1,969 2.81 0.05 2.81 0.05
Sun Trading: FSOM ELP 724 3.03 0.10 3.03 0.10
Broker-owned dark pool: BODP Broker’s D.P. 93,745 4.33 0.02 4.33 0.02
BIDS ATS: BIDS Other D.P. 400 0.85 0.05 0.85 0.05
Level ATS: EBXL Other D.P. 6,798 2.29 0.03 2.29 0.03
EDGA Stock Exchange: EDGA Low Fee Exch. 9,049 2.16 0.02 3.18 0.03
New York Stock Exchange: NYSE Make/Take Exch. 230,185 2.11 0.01 2.91 0.01
Nasdaq Stock Exchange: Nasdaq Make/Take Exch. 614,059 2.13 0.00 3.05 0.00
Archipelago Stock Exchange: ARCX Make/Take Exch. 289,624 2.14 0.00 3.07 0.00
EDGX Stock Exchange: EDGX Make/Take Exch. 65,885 2.23 0.01 3.20 0.10
BZX Stock Exchange: BZX Make/Take Exch. 94,401 2.28 0.01 3.50 0.01
Archipelago Stock Exchange: ARCA Make/Take Exch. 6,701 3.04 0.03 4.60 0.05
ELPs 122,462 2.24 0.01 2.24 0.01

Make/Take Exchanges 1,300,855 2.15 0.00 3.08 0.00
Inverted Exchanges 5,444 2.34 0.03 2.13 0.03
Other Dark Pools 7,419 2.21 0.03 2.21 0.03
Broker’s Dark Pool 93,745 4.33 0.02 4.33 0.02
35

Table IV. Multivariate analysis of the relationship between trading costs and ELP exposure.
Notes: For a given child order execution, effective spread is measured as the difference between the trade
price and the execution-time NBBO midpoint divided by the execution-time NBBO midpoint. Net effective
spread is the effective spread net of the assumed fee or rebate charged by the executing venue.
Implementation Shortfall is computed as the normalized difference between the volume-weighted average
child order execution price and the NBBO midpoint prior to the start of the execution. Net Implementation
Shortfall is adjusted for assumed fees and rebates. ELP Exposure is the percentage of a parent order’s trades
with ELPs. BODP Exposure (Other DP Exposure) is the percentage of a parent order’s child executions
that occur in the broker’s own dark pool (in dark pools that our data provider does not own). Passive
Exchange Exposure is the percentage of a parent order’s child executions that provide liquidity on a lit
stock exchange. Participation Rate is measured as the ratio of the parent order’s trading volume to the
overall trading volume of the underlying stock over the period of time that the parent order is being worked.
Volatility is measured as the volatility of the midpoint of the NBBO over the parent order’s life. Duration
is the fraction of the trading day that the parent order is worked. Turnover is the ratio of the number of
shares traded during the life of the parent order to the outstanding number of shares in thousands. Each
regression includes stock, client and calendar day fixed-effects. Standard errors are given in parentheses
and are adjusted by double-clustering on stock and day.
Effective Net Effective Implementation Net Implementation

Spread Spread Shortfall Shortfall
0.14 -0.45** 44.45*** 43.85***
ELP Exposure
(0.26) (0.23) (8.56) (8.56)
-1.48*** -2.21*** -7.81* -8.53**
BODP Exposure
(0.21) (0.22) (4.13) (4.13)
-0.40 -1.36 -39.20 -39.78
Other DP Exposure
(0.64) (0.85) (29.73) (29.06)
-3.29*** -4.23*** 6.37 5.35
Passive Exchange Exposure
(0.17) (0.18) (5.24) (5.24)
1.99*** 2.28*** 52.09*** 52.33***
Participation Rate
(0.58) (0.62) (19.85) (19.85)
-0.94*** -1.35*** 4.14 3.78
Log (Price)
(0.31) (0.33) (4.79) (4.79)
9.20 6.00 56.37 56.99
Volatility
(12.07) (15.58) (314.67) (314.74)
0.05 0.09 -3.99 -3.98
Duration
(0.08) (0.09) (8.68) (8.68)
-0.002 -0.003 0.37 0.37
Turnover
(0.003) (0.004) (0.38) (0.38)
N 20,335 20,335 20,335 20,335

Adjusted R2 0.13 0.19 0.10 0.10
36

Table V. Summary statistics for parent orders submitted by clients who avoid ELPs and their matches.
Notes: ELP Exposure is the percentage of a parent order’s trades with ELPs. BODP Exposure is the
percentage of a parent order’s child executions that occur in the broker’s own dark pool. Passive Exchange
Exposure is the percentage of a parent order’s child executions that provide liquidity on a lit stock exchange.
Participation Rate is measured as the ratio of the parent order’s trading volume to the overall trading volume
of the underlying stock over the period of time that the parent order is being worked. Log (Price) is the log
of the midpoint of the NBBO prevailing when the parent order is placed with the broker. Volatility is
measured as the volatility of the midpoint of the NBBO over the parent order’s life. Quoted Spread is the
time-weighted percentage bid-ask spread over the parent order’s life. Standard errors are adjusted by
double-clustering on stock and day.
Statistic No ELP ELP Difference p-value

# of parent orders 364 364
ELP Exposure (%) 0.00 6.31 -6.31 < 0.01
BODP Exposure (%) 2.99 10.11 -7.12 < 0.01
Passive Exchange Exposure (%) 0.370 0.312 0.058 < 0.01
Participation Rate (%) 2.41 1.22 1.19 < 0.01
Log (Price) 3.934 3.935 -0.001 0.98
Volatility (%) 1.51 1.60 0.09 0.28
Quoted Spread (bps) 4.10 4.20 -0.10 0.50
37

Table VI. Multivariate analysis of trading costs of parent orders that are not exposed to ELPs and their
matches.
Notes: Implementation Shortfall is computed as the normalized difference between the average child order
execution price and the NBBO midpoint prior to the start of the execution. Net Implementation Shortfall is
adjusted for assumed fees and rebates. Not Exposed to ELPs is an indicator variable that is set equal to one
if the client submitting the parent order does not source ELP liquidity. BODP Exposure is the percentage
of a parent order’s child executions that occur in the broker’s own dark pool. Passive Exchange Exposure
is the percentage of a parent order’s child executions that provide liquidity on a lit stock exchange.
of the underlying stock over the period of time that the parent order is being worked. Duration is the fraction
of the trading day that the parent order is worked. Standard errors are given in parentheses and are adjusted
by double-clustering on stock and day.
Implementation Shortfall Net Implementation Shortfall
-12.22** -12.07**
Not Exposed to ELPs
(5.69) (5.69)
-19.44 -20.23
BODP Exposure
(19.85) (19.86)
24.33 22.65
(18.97) (18.97)
157.50* 156.57*
Participation Rate
(83.13) (83.16)
2.85 2.81
Duration
(7.84) (7.84)
N 728 728
Adjusted R2 0.01 0.01
38

Table VII. Difference-in-differences regression of trading costs incurred by the client who stops sourcing
ELP liquidity.
Notes: All executions generated by parent orders in the Exposed group of orders are included in this analysis.
Implementation Shortfall is computed as the normalized difference between the average child order execution price
and the price of the asset prior to the start of the execution. Net Implementation Shortfall is adjusted for assumed fees
and rebates. Switching Client Order takes a value of 1 for executions generated by the switching client’s orders. Post
20110318 takes a value of 1 for executions after March 18, 2011. Switching Client Order after 20110318 takes a value
of 1 if the execution is from the switching client and occurs after March 18, 2011. BODP Exposure (Other DP
Exposure) is the percentage of a parent order’s child executions that occur in the broker’s own dark pool (in a dark
pool that is not owned by the data provider). Passive Exchange Exposure is the percentage of a parent order’s child
executions that provide liquidity on a lit stock exchange. Participation Rate is measured as the ratio of the parent
order’s trading volume to the overall trading volume of the underlying stock over the period of time that the parent
order is being worked. Log (Price) is the log of the midpoint of the NBBO prevailing when the parent order is placed
with the broker. Volatility is measured as the volatility of the midpoint of the NBBO over the parent order’s life.
Duration is the fraction of the trading day that the parent order is worked. Turnover is the ratio of the number of shares
traded during the life of the parent order to the outstanding number of shares in thousands. Each regression includes
stock, client and calendar day fixed-effects. Standard errors are given in parentheses and are adjusted by double-
clustering on stock and day.
-32.27** -32.23**
Switching Client Order after 20110318
(15.77) (15.77)
49.23*** 49.27***
Post 20110318
(11.86) (11.86)
-1.44 -1.38
Switching Client Order
(10.97) (10.97)
-13.56*** -14.19***
BODP Exposure
(3.85) (3.85)
-58.50** -58.94**
Other DP Exposure
(29.16) (29.16)
-0.09 -1.01
Passive Exch. Exposure
(4.69) (4.69)
46.13*** 46.47***
Participation Rate
(15.96) (15.96)
1.45 1.15
Log (Price)
(4.34) (4.34)
67.60 68.25
Volatility
(252.00) (252.00)
-7.60* -7.56*
Duration
(3.95) (3.95)
0.49* 0.49*
Turnover
(0.29) (0.29)
N 19,386 19,386
2
Adjusted R 0.10 0.10
39

Table VIII. The cost of early versus late exposure to ELPs.
execution price and the price of the asset prior to the start of the execution. Net Implementation Shortfall is
adjusted for assumed fees and rebates. ELP Exposure in Decile 1 is the percentage of child executions by
ELPs in the first ten percent of the parent order’s executed volume. ELP Exposure in Decile 10 and ELP
Exposure in other deciles are defined analogously. BODP Exposure (Other DP Exposure) is the percentage
of a parent order’s child executions that occur in the broker’s own dark pool (in a dark pool that is not
owned by the data provider). Passive Exchange Exposure is the percentage of a parent order’s child
executions that provide liquidity on a lit stock exchange. Participation Rate is measured as the ratio of the
parent order’s trading volume to the overall trading volume of the underlying stock over the period of time
that the parent order is being worked. Log (Price) is the log of the midpoint of the NBBO prevailing when
the parent order is placed with the broker. Volatility is measured as the volatility of the midpoint of the
NBBO over the parent order’s life. Duration is the fraction of the trading day that the parent order is worked.
Turnover is the ratio of the number of shares traded during the life of the parent order to the outstanding
number of shares in thousands. Each regression includes stock, client and calendar day fixed-effects.
Standard errors are given in parentheses and are adjusted by double-clustering on stock and day.
55.84*** 55.21***
ELP Exposure in Decile 1
(9.45) (9.45)
26.58*** 25.93***
ELP Exposure in Decile 10
(9.14) (9.15)
33.04*** 32.40***
ELP Exposure in other deciles
(9.70) (9.71)
-7.20 -7.97
BODP Exposure
(4.44) (4.44)
-43.33 -43.97
Other DP Exposure
(31.77) (31.78)
8.26 7.18
(5.49) (5.49)
59.98*** 60.23***
Participation Rate
(19.86) (19.87)
5.22 4.85
Log (Price)
(4.98) (4.97)
155.34 155.93
Volatility
(326.57) (326.63)
-4.07 -4.06
Duration
(7.66) (7.66)
0.50 0.50
Turnover
(0.32) (0.32)
N 20,335 20,335
40

Table IX. The cost of exposing to multiple ELPs.
adjusted for assumed fees and rebates. More than 2 ELPs (More than 3 ELPs) is an indicator variable that
equals one if the parent order sourced liquidity from at least two different ELPs (three different ELPs).
BODP Exposure (Other DP Exposure) is the percentage of a parent order’s child executions that occur in
the broker’s own dark pool (in a dark pool that is not owned by the data provider). Passive Exchange
measured as the volatility of the midpoint of the NBBO over the parent order’s life. Duration is the fraction
of the trading day that the parent order is worked. Turnover is the ratio of the number of shares traded
during the life of the parent order to the outstanding number of shares in thousands. Each regression includes
stock, client and calendar day fixed-effects. Standard errors are given in parentheses and are adjusted by
double-clustering on stock and day.
5.51*** 5.51***
More than 2 ELPs
(1.41) (1.41)
5.93** 5.95**
More than 3 ELPs
(2.77) (2.77)
33.30*** 32.69***
ELP Exposure
(8.26) (8.26)
-5.24 -5.96
BODP Exposure
(4.44) (4.06)
-32.15 -32.73
Other DP Exposure
(28.54) (28.54)
7.26 6.24
(5.21) (5.21)
48.34** 48.56**
Participation Rate
(19.95) (19.96)
4.84 4.48
Log (Price)
(4.81) (4.82)
43.25 43.85
Volatility
(312.86) (312.92)
-6.01 -6.00
Duration
(8.58) (8.58)
0.37 0.37
Turnover
(0.37) (0.37)
N 20,335 20,335
41

Online Appendix to “The Cost of Exposing Large Institutional Orders to
Electronic Liquidity Providers”
This online appendix contains additional list of tables referenced in the main document.
Table A.I Fee assumptions (per share) for child order executions in each venue.
Table A.II Multivariate analysis of the relationship between VWAP slippage and ELP
exposure.
Table A.III Multivariate relationship between trading costs and exposure to GETCO and the
other ELPs.
Table A.IV Summary statistics for parent orders submitted by the switching client before and
after the switch date.
42

Table A.I. Fee assumptions (per share) for child order executions in each venue.
Notes: Negative numbers refer to rebates. BODP refers to the broker’s own dark pool.
Liquidity Providing Liquidity Demanding

Venue Type
Trades Trades
Archipelago Stock Exchange: ARCA Make/Take Exch. − $0.0025 $0.0030
Archipelago Stock Exchange: ARCX Make/Take Exch. − $0.0025 $0.0030
BZX Stock Exchange: BZX Make/Take Exch. − $0.0020 $0.0030
EDGX Stock Exchange: EDGX Make/Take Exch. − $0.0020 $0.0030
Nasdaq Stock Exchange: Nasdaq Make/Take Exch. − $0.0020 $0.0030
New York Stock Exchange: NYSE Make/Take Exch. − $0.0014 $0.0027
EDGA Stock Exchange: EDGA Low Fee Exch. $0.0003 $0.0003
Virtu Americas: GFLO ELP n.a. $0.0000
Citadel: CDRG ELP n.a. $0.0000
Trimark: TRIM ELP n.a. $0.0000
Knight Securities: NITE ELP n.a. $0.0000
D.E. Shaw: SHAW ELP n.a. $0.0000
Two Sigma Securities: SOHO ELP n.a. $0.0000
Sun Trading: FSOM ELP n.a. $0.0000
BIDS ATS: BIDS Other D.P. $0.0000 $0.0000
Level ATS: EBXL Other D.P. $0.0000 $0.0000
Broker-owned dark pool: BODP Broker’s D.P. $0.0000 $0.0000
Nasdaq BX Stock Exchange: BX Inverted Exch. $0.0020 − $0.0006
BYX Stock Exchange: BYX Inverted Exch. $0.0018 − $0.0008
43

Table A.II. Multivariate relationship between implantation shortfall and exposure to GETCO and the other
ELPs.
adjusted for assumed fees and rebates. GETCO Exposure is the percentage of child executions by GETCO.
Other ELP Exposure is the percentage of child executions by ELPs other than GETCO. BODP Exposure
(Other DP Exposure) is the percentage of a parent order’s child executions that occur in the broker’s own
dark pool (in a dark pool that is not owned by the data provider). Passive Exchange Exposure is the
percentage of a parent order’s child executions that provide liquidity on a lit stock exchange. Participation
Rate is measured as the ratio of the parent order’s trading volume to the overall trading volume of the
underlying stock over the period of time that the parent order is being worked. Log (Price) is the log of the
midpoint of the NBBO prevailing when the parent order is placed with the broker. Volatility is measured
as the volatility of the midpoint of the NBBO over the parent order’s life. Duration is the fraction of the
trading day that the parent order is worked. Turnover is the ratio of the number of shares traded during the
life of the parent order to the outstanding number of shares in thousands. Each regression includes stock,
client and calendar day fixed-effects. Standard errors are given in parentheses and are adjusted by double-
clustering on stock and day.
Implementation Net Implementation

Shortfall Shortfall
34.92*** 34.31***
GETCO Exposure
(8.92) (8.92)
89.67*** 89.10***
Other ELP Exposure
(15.61) (15.60)
-7.26* -7.98*
BODP Exposure
(4.14) (4.13)
-37.35 -37.93
Other DP Exposure
(29.05) (29.05)
6.82 5.81
(5.26) (5.25)
54.51*** 54.75***
Participation Rate
(19.88) (19.88)
4.15 3.78
Log (Price)
(4.76) (4.76)
50.49 51.11
Volatility
(314.23) (314.30)
-4.08 -4.07
Duration
(8.69) (8.69)
0.37 0.37
Turnover
(0.38) (0.38)
N 20,335 20,335
44

Table A.III. Multivariate analysis of the relationship between VWAP slippage and ELP exposure.
Notes: VWAP Slippage is computed as the normalized difference between the average child order
execution price and the volume-weighted average price observed in the market during the parent order’s
lifetime. Net VWAP Slippage is adjusted for assumed fees and rebates. ELP Exposure is the percentage of
a parent order’s trades with ELPs. BODP Exposure (Other DP Exposure) is the percentage of a parent
order’s child executions that occur in the broker’s own dark pool (in a dark pool that is not owned by the
data provider). Passive Exchange Exposure is the percentage of a parent order’s child executions that
provide liquidity on a lit stock exchange. Participation Rate is measured as the ratio of the parent order’s
trading volume to the overall trading volume of the underlying stock over the period of time that the parent
order is being worked. Log (Price) is the log of the midpoint of the NBBO prevailing when the parent order
is placed with the broker. Volatility is measured as the volatility of the midpoint of the NBBO over the
parent order’s life. Duration is the fraction of the trading day that the parent order is worked. Turnover is
the ratio of the number of shares traded during the life of the parent order to the outstanding number of
shares in thousands. Each regression includes stock, client and calendar day fixed-effects. Standard errors
are given in parentheses and are adjusted by double-clustering on stock and day.
VWAP Net VWAP

Slippage Slippage
4.70*** 4.10***
ELP Exposure
(1.15) (1.28)
-0.96* -1.68***
BODP Exposure
(0.55) (0.53)
9.90 9.32
Other DP Exposure
(4.73) (7.71)
-1.85*** -2.87***
(0.67) (0.79)
3.15 3.40**
Participation Rate
(2.86) (1.69)
0.46 0.09
Log (Price)
(0.55) (1.08)
142.92* 143.56*
Volatility
(11.15) (82.79)
-0.28 -0.27
Duration
(0.39) (0.78)
0.01 0.01
Turnover
(0.01) (0.06)
N 20,335 20,335
2
Adjusted R 0.08 0.08
45

Table A.IV. Summary statistics for parent orders submitted by the switching client before and after the
switch date
Notes: ELP Exposure is the percentage of a parent order’s trades with ELPs. BODP Exposure is the
percentage of a parent order’s child executions that occur in the broker’s own dark pool. Passive Exchange
measured as the volatility of the midpoint of the NBBO over the parent order’s life. Quoted Spread is the
time-weighted percentage bid-ask spread over the parent order’s life. We compute the z-scores by
normalizing the raw variables with their daily means and standard deviations. Standard errors are adjusted
by double-clustering on stock and day.
Statistic Pre Post Difference p-value

# of parent orders 40 111
ELP Exposure (%) 6.19 0.00 6.19 < 0.01
BODP Exposure (%) 4.22 4.13 0.09 0.95
Passive Exchange Exposure 0.39 0.36 0.03 0.21
Participation Rate (%) 2.05 2.05 0.00 0.99
Log (Price) 3.853 3.813 0.039 0.70
Volatility (z-score) -0.13 -0.13 0.00 0.99
Quoted Spread (z-score) -0.03 0.06 -0.09 0.55
46

SSRN Id3281324 PDF

Uploaded by

Copyright:

Available Formats

SSRN Id3281324 PDF

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SSRN Id3281324 PDF

Uploaded by

Copyright:

Available Formats

This Draft: September 16, 2019

The Cost of Exposing Large Institutional Orders

Robert Battalio Brian Hatch Mehmet Saǧlam

Keywords: Electronic Liquidity Providers, Transactions Costs, Order Anticipation

Electronic copy available at: https://ssrn.com/abstract=3281324

execution across all of these trading centers to minimize information leakage.

Electronic copy available at: https://ssrn.com/abstract=3281324

the execution quality of the parent order.

Electronic copy available at: https://ssrn.com/abstract=3281324

have at least one child order filled by this group of ELPs.

exchanges, trading costs increase by roughly 10%.

Electronic copy available at: https://ssrn.com/abstract=3281324

transaction costs after eliminating the ELP exposure.

Electronic copy available at: https://ssrn.com/abstract=3281324

II. Related literature.

routers that maximize fill rates.

Electronic copy available at: https://ssrn.com/abstract=3281324

compute the statistics around the execution interval.

Electronic copy available at: https://ssrn.com/abstract=3281324

III. Data and summary statistics

A. Description of the institutional order data.

Electronic copy available at: https://ssrn.com/abstract=3281324

provide institutional details about the execution strategy.

do not have unexecuted child orders.

Electronic copy available at: https://ssrn.com/abstract=3281324

placed by 146 distinct investors.

clients chose this broker to maximize their rebate revenue.

Electronic copy available at: https://ssrn.com/abstract=3281324

duration, and routing of each child order.

we effectively have five active ELPs during our sample period.

with the ELPs specified that rejection rates cannot be high.

of the parent order is filled aggressively and 31% is filled passively. 11

thirty-eight minutes (e.g., 0.52 x 6.5 hours).

Electronic copy available at: https://ssrn.com/abstract=3281324

order execution in the broker operated dark pool.

study the impact of interacting with ELPs in the dark.

similar statistics support the representativeness of our institutional trading data.

Electronic copy available at: https://ssrn.com/abstract=3281324

parent orders placed by different institutional institutions.

Electronic copy available at: https://ssrn.com/abstract=3281324

Electronic copy available at: https://ssrn.com/abstract=3281324

examine its impact on implementation costs.

IV. Rationalizing the decision to route individual child orders to ELPs.

Electronic copy available at: https://ssrn.com/abstract=3281324

Electronic copy available at: https://ssrn.com/abstract=3281324

A. Measure of parent order execution quality.

for the ith parent order in our sample, we compute IS as follows:

for estimates of venue fees and rebates.

stock price over the duration of the parent order’s life: 13

𝑁𝑁 𝑃𝑃𝑖𝑖,𝑗𝑗 − 𝑀𝑀𝑖𝑖,𝑗𝑗 𝑄𝑄𝑖𝑖,𝑗𝑗 𝑁𝑁 𝑀𝑀𝑖𝑖,𝑗𝑗 − 𝑀𝑀𝑖𝑖,0 𝑄𝑄𝑖𝑖,𝑗𝑗

Electronic copy available at: https://ssrn.com/abstract=3281324

B. ELP exposure and trading costs.

instead of being filled aggressively in exchanges.

Electronic copy available at: https://ssrn.com/abstract=3281324

to the institutional client.

institutional trading costs as an increase in trading aggressiveness.

Electronic copy available at: https://ssrn.com/abstract=3281324

the TAQ database.