This draft: October 29, 2013

Can Brokers Have it all? On the Relation between Make Take Fees & Limit Order Execution Quality*
by Robert Battalio Mendoza College of Business University of Notre Dame rbattali@nd.edu (574) 631-9428 Shane Corwin Mendoza College of Business University of Notre Dame scorwin@nd.edu (574) 631-6026 Robert Jennings Kelley School of Business Indiana University jennings@indiana.edu (812) 855-2696 Extended Abstract: Today, every U.S. equity exchange utilizes a rebate-fee schedule (see Cardella, et al. (2013)), typically charging traders taking liquidity (e.g., marketable orders) and paying traders making liquidity (e.g., nonmarketable limit orders). Using publicly available data, we document that several large retail brokerages route their order flow in a manner that appears to maximize order flow payments: they sell marketable orders and they route limit orders to venues that offer the highest liquidity rebates. As rebates are funded by take fees, the venues with the highest liquidity rebates also have the highest take fees. If limit orders displayed on venues with low take fees trade ahead of a similarly priced limit orders displayed on venues with high take fees, routing orders to maximize rebates might be inconsistent with a brokers fiduciary responsibility to obtain best execution. For a set of proprietary limit orders we obtain from a major broker, we document a negative relation between limit order execution quality and the relative level of liquidity fees. Among other things, we show that when identical limit orders are routed to two venues at the same time, orders routed to the venue with the lower fee execute more frequently and earn higher realized spreads. We conclude our analysis by using the NYSEs Trade and Quote data to show the negative relation between take fees and limit order execution quality extends beyond the limitations of our proprietary dataset. Thus, brokers cannot have it all. Routing limit orders in a manner that maximizes make rebates reduces fill rates and produces less profitable limit order executions.

*The authors gratefully acknowledge research support from the Q-Group. We thank Jeff Bacidore, Peter Bottini, Larry Harris, Paul Jiganti, Bill OBrien, Paul Schultz, Jamie Selway, John Standerfer, and brownbag participants at Indiana University and at the University of Notre Dame for their comments.

Today, every U.S. stock exchange levies fees or pays rebates based on order characteristics (see Cardella, et al. (2013)). In the traditional model, exchanges charge liquidity demanding orders (e.g., marketable orders) a fee that exceeds the rebate they offer liquidity supplying orders (e.g., nonmarketable limit orders). More recently, a few inverted exchanges began charging liquidity suppliers a fee that exceeds the rebate they pay to liquidity demanders.1 Although these differential fee schedules give traders increased flexibility, they are controversial. As noted by the Investment Company Institute in their April 2010 letter to the Securities and Exchange Commission (SEC), brokers may refrain from posting limit orders on a particular exchange because it offers lower liquidity rebates than other markets, even though that exchange offers the best possibility of an execution for those limit orders.2 Although the SECs Order Protection Rule establishes price priority in U.S. equity markets, the rule does not specify who trades first when multiple venues have the best posted price. Angel et al. (2010) note that traditional and inverted make-take fee schedules allow market participants to adjust the priority of their orders at a given or stated limit price. All else equal, if two venues offer the national best bid, one expects liquidity-demanding sellers arriving in the marketplace to route their orders to the venue with the lower take fee. If sufficient selling demand arrives, sellers walk down the limit order books and the limit order traders who purchased the stock at the bid price suffer a loss. However, if the stock price rises before liquidity is exhausted at the national best bid, limit orders on the venue with the higher take fee (and thus, the higher make rebate) become isolated and miss out on profitable trading opportunities. Thus, limit orders routed to venues with high take fees suffer greater adverse selection costs they are more likely to trade when the price moves against them and are less likely to trade when prices move in their favor. This suggests that brokers routing limit orders to venues with the highest take fees (and make rebates) might not be fulfilling their fiduciary responsibility to obtain best execution for their clients. Why might brokers and clients interests diverge? Angel et al. (2010) note that if clients paid/received fees/rebates, brokers generally would send limit orders to the venue that maximizes the
1

The difference between the fee and the rebate is an important source of revenue for exchanges. Given the vigorous competition between U.S. exchanges, there is a high correlation between the level of an exchanges fee and rebate. 2 See http://www.sec.gov/comments/s7-02-10/s70210-138.pdf. 1

likelihood of execution as brokers receive commissions only when orders execute. The typical situation, however, is for a brokerage firm to offer a fixed commission that is set assuming it earns (pays) exchange rebates (fees). All else equal, in a competitive market, brokers that pay the least to exchanges offer the lowest commission. If investors choose brokers based primarily on commissions (perhaps because they lack the sophistication and/or the necessary information to evaluate limit order execution quality), it may be profit maximizing for brokers to maximize liquidity rebates rather than the profitability and the frequency of limit order executions.3 Take fees are not the only factor that influences the frequency with which limit orders execute. In stocks with sufficient intraday volatility, the limit order routing decision might have little or no impact on fill rates as limit orders are likely to become marketable at some point in the trading day. In this situation, there is little cost associated with routing limit orders to venues with high take fees and high liquidity rebates. Moreover, as suggested by UBSs best execution statement,4 algorithms can find it optimal to route marketable orders to venues with high take fees if those venues offer sufficient price and/or depth improvement.5 It might also be the case that venues with high take fees attract marketable orders by offering competitively priced co-location services, which make it easier for algorithms to trade at the quotes they see thereby reducing execution price risk. Thus, the relation between exchange fee schedules and limit order execution quality is an empirical issue. Are marketable orders allocated to venues quoting the best price on the basis of fees (e.g., net price priority)? If so, then limit orders routed to exchanges with high take fees might suffer significant adverse selection costs. Alternatively, do liquidity supply queues adjust in a way such that limit order execution quality is unrelated to the venues rebate-fee structures? In this paper, we investigate these issues using a proprietary dataset to examine limit order execution quality across venues with different fee
3

Indeed, the head of order routing for the retail brokerage Ameritrade recently stated that by routing their market orders to wholesalers and their nonmarketable limit orders to exchanges, Ameritrade could improve its revenue. See Wall Streets Most Sophisticated Order Sender in the September 2010 issue of Traders Magazine. 4 See UBSs Best Execution Statement at http://www.ubs.com/content/dam/static/wmamericas/bestexecution.pdf. 5 Price improvement occurs when a marketable order trades at a price that is better than the posted quote. Depth improvement occurs when a marketable order trades more shares than are advertised in the posted quote. Price and depth improvement occur when a marketable order trades against non-displayed liquidity, which is typically provided by non-displayed limit orders on exchanges. 2

structures and with NYSE TAQ data to evaluate the location and the profitability of limit order executions when multiple venues post the best bid or ask price. If all brokers employ smart routers that accurately assess the likelihood of limit order execution, then the policy implications of our study might be minimal. If, however, some brokers follow order routing strategies designed to maximize maker fees, our work may raise best execution questions. Thus, we begin by investigating data that describe where eleven national brokerages route orders in the fourth quarter of 2012. We interpret the routing decisions of four brokers, Ameritrade, E*Trade, Fidelity and Scott Trade, as being consistent with the objective of harvesting liquidity rebates. Three of these brokers sells market orders to market makers and routes limit orders either to market makers or to EDGX, an exchange offering the largest per share liquidity rebate. In an effort to understand whether routing nonmarketable limit orders to venues with high fees results in diminished execution quality, as hypothesized by Angel et al. (2010), we examine the relation between take fees and limit order execution quality. We begin by analyzing proprietary limit order data obtained from a major investment bank that uses a sophisticated algorithm to route orders. Univariate limit order execution quality statistics are suggestive of a negative relationship between take fees and limit order execution quality. However, it is difficult to use univariate statistics to make assessments of relative limit order execution quality across exchanges, as orders may be routed to different venues under different types of market conditions. We address the endogeneity of the order routing decision by analyzing the association between take fees and the probability that a limit order executes and the relation between take fees and execution speed for executed limit orders in a multivariate setting. All else equal, we find that fill rates for displayed limit orders are lower on exchanges with higher fees. In addition, limit orders executed on venues with high fees take longer to execute than those executed on venues with low fees. These results confirm the negative relation between take fees and limit order execution quality suggested by the univariate results. As the decision to route limit orders to different venues is likely not random, our findings may be affected by a selection bias (see Peterson and Sirri (2002)). We utilize a unique feature of our data to
3

address this issue. Frequently, identically priced limit orders to trade shares of the same stock are routed simultaneously to multiple venues. For these identical orders, differences in fill rates, execution speeds, and adverse selection costs can be linked directly to exchange characteristics such as the make-take fee schedule, hidden liquidity, latency and queue length. Perhaps most important, an analysis of these orders allows us to hold market conditions constant. Where there are sufficient order-pair observations, we conduct horseraces between different exchanges to assess whether the limit order routed to the venue with the lower take fee provides the fastest execution and lower adverse selection costs. In nearly every comparison we make, the venue with the lower take fee wins a higher percentage of horseraces and has less adverse selection. Interestingly, all of the winning venues and even most of the losing venues have positive paper trading profits after five minutes, which suggests our data provider does a nice job managing the adverse selection risk associated with placing non-marketable limit orders. For brokers that do not pass fees/rebates directly through to their customers, the results of our analysis suggest that the decision to route limit orders to a single exchange that offers the highest liquidity rebates is inconsistent with maximizing limit order execution quality. One may, however, question the generalizability of our results as our data are from a single broker, constitute only 1.5% of average daily trading volume, and cover only about one-half of the U.S. equity trading venues. For this reason, we next use the NYSEs TAQ database to make inferences regarding the across-venue execution quality of at-thequote limit orders. If we assume that the quotes displayed on each of the traditional and inverted venues are set by limit orders, we can examine whether take fees influence where trades execute when multiple venues are at the inside quote and the adverse selection associated with these executions. In general, we find that measures of adverse selection are positively correlated with ven ues make rebates/take fees: for at-the-quote limit order executions the inverted venues tend to have the least adverse selection and the high rebate/take fee venues tend to have the most adverse selection. The correlation is not perfect, however, as the NYSE and EDGX do better than expected given their fee structures, suggesting fees are not the only determinant of across-venue differences in limit order execution quality.

For limit order traders that pay/receive liquidity fees/rebates, realized spreads adjusted for fees are the appropriate executed limit order quality metric. For NYSE-listed securities, the inverted venues, the NYSE, and EDGX offer positive average realized spreads net of liquidity rebates/fees for at-the-quote limit order executions. For at-the-quote limit order executions in Nasdaq stocks, only the inverted venues have average net realized spreads that are positive. These findings suggest that brokers seeking to make limit order routing decisions to optimize execution quality may make different routing decisions depending upon whether or not they pass fees/rebates on to their customers. We plan to complete this section by using cross sectional regressions to better understand the types of situations in which the routing of at-the-quote limit orders is most/least important. The remainder of this paper is as follows. In Section II, we provide a brief literature review. In Section III, we use publicly available data to investigate broker routing decisions. In Section IV, we use proprietary order data to examine the relationship between make-take fees and limit order execution quality. In Section V, we use the NYSEs TAQ database to investigate the relationship between take fees and limit order execution quality when multiple venues are at the inside quote. Section VI concludes. II. Related Literature When an assets trading volume is concentrated on one venue, the (limit) order routing decision is trivial. In the United States, the order routing decision became more substantive after the Consolidated Tape was introduced in 1982. By allowing trading venues to benchmark their trades against each others quotes, competition for order flow became more intense. According to the SEC, when making its order routing decisions a broker-dealer must consider several factors affecting the quality of execution, including, for example, the opportunity for price improvement, the likelihood of execution (which is particularly important for customer limit orders), the speed of execution, and the trading characteristics of the security, together with other non-price factors such as reliability and service.6 The SECs 1997 Report on the Practice of Preferencing, contains one of the first cross-venue analyses of retail limit order execution quality. In their study, the SEC documents fill rates, time-to6

http://www.sec.gov/rules/final/34-43590.htm. 5

execution statistics, and adverse selection costs for limit orders that execute on various U.S. exchanges. Their analysis does not, however, control for the endogeneity associated with the order routing decision. In order to ensure that routing retail limit orders away from the NYSE did not degrade execution quality, regional exchanges implemented rules that benchmarked their limit order executions to NYSE limit order executions. Using a methodology that allows them to control for market conditions and order submission strategies, Battalio et al. (2002) find that these rules were effective at producing limit order execution quality that was competitive with the NYSEs. Given the fear that at least some brokers were maximizing order flow payments rather than execution quality, the SEC passed Rule 11ac1-5 (now Rule 605) and Rule 11ac1-6 (now Rule 606) in 2001. Together, these rules are intended to bring sufficient transparency to the market so that investors can determine whether their brokers are making optimal order routing decisions. Rule 605 requires exchanges to produce execution quality statistics on a monthly basis. Rule 606 requires brokers to reveal on a quarterly basis the destinations to which they route orders and whether they receive compensation for their routing choices. Boehmer, Jennings and Wei (2007) find that the routing of marketable order flow became more sensitive to across-venue changes in execution quality after Rule 605 execution quality statistics were available. However, the Rule 605 limit order execution quality statistics are similar to those used in the SECs 1997 study in that they do not control for potential endogeneity issues. We apply alternative metrics to publicly available trade and quote data to gain insights as to the differences in limit order execution quality across exchanges. Foucault and Menkveld (2008) examine how market fragmentation and fee differentials affect limit order execution quality when price priority is not enforced across markets. When the two markets they examine are both at the inside quote, they find that smart routers predominately send orders to the market with the lowest fee. They also find evidence that violations of price priority across the two markets adversely affect liquidity provision. Consistent with the idea that marketable order flow is sensitive to take fees, Cardella, Hao and Kalcheva (2013) find that reductions in relative take fees in U.S. equity markets are associated with increased market share.
6

Colliard and Foucault (2013) and Foucault, Kadan and Kandel (2013) construct theoretical models to investigate how make-take fees affect liquidity supply. Colliard and Foucault consider how an exchange competing with a dealer market should optimally set its make-take fee schedule. They conclude that the net fee (fee minus rebate) has an ambiguous effect on the execution probability for limit orders and the speed with which the limit order queue moves. Foucault, Kadan, and Kandel (FKK) examine whether it is the net fee, or the relative levels of the make and take fees that matter. They argue that exchanges can maximize their trading volume by differentiating their make and take fees. If there is not enough liquidity demand (supply), the venue can decrease (increase) its take fee and its make rebate. We contribute to this literature by providing the first empirical evidence as to the relationship between take fees and multiple dimensions of limit order execution quality. Perhaps due to the availability of superior data, the relation between limit order execution quality and make-take fees has received more attention in the practitioner arena than in academic research. Sofianos et al. (2010) use nonmarketable limit orders placed on six exchanges by the Goldman Sachs smart router, SIGMA, to examine the relation between take fees and limit order execution quality. Using only those limit orders split between two exchanges, Sofianos et al. make pairwise comparisons of execution quality across exchanges. As all non-venue specific factors are the same, they note that differences in execution quality are either due to adverse selection risk or differences in fill rates. They find the venue utilizing an inverted make-take schedule has lower adverse selection costs, faster fills, and higher fill rates than the other five exchanges. Ignoring fees, these results suggest brokers routing nonmarketable limit orders to venues with high take fees disadvantage their clients. In addition to replicating this analysis, we use both proprietary and publicly available data to examine these issues in more general settings. We also examine whether the conclusions of Sofianos et al. change when fees/rebates are passed through to limit order traders. Using a proprietary dataset of orders generated from one of Goldmans execution algorithms, Bacidore, Otero and Vasa (2011) attempt to quantify the benefits of smart routing by evaluating five order routing strategies. When routing limit orders, one strategy follows the nave approach of maximizing
7

rebates and two strategies attempt to maximize fill rates. When routing marketable order flow, one algorithm minimizes take fees while two consider both take fees and hidden liquidity. They present evidence that the nave strategy delivers inferior limit order execution quality, even after accounting for maker rebates. For marketable orders, they find smaller differences in execution quality across algorithms, which they argue is intuitive given the Reg. NMS protections on market orders. This said, in large capitalization stocks they find some benefit to routing on the basis of both fees and the prospect of hidden liquidity. We complement this analysis by identifying retail brokerages whose order routing decisions appear consistent with the objective of maximizing rebates and then examining whether this type of order routing disadvantages limit order traders. III. Broker routing decisions. Market makers profit by selectively purchasing and executing market orders and marketable limit orders. One constraint on a market makers ability to interact with purchased order flow is FINRA Rule 5320, which states that a market maker holding a nonmarketable limit order is prohibited from trading that security on the same side of the market for its own account at a price that would satisfy the customer order.7 Thus, brokers can increase the revenue generated from their order flow by routing marketable orders to purchasers and non-marketable limit orders to venues offering high make rebates. Indeed, if all brokers routed orders in this fashion, purchasers could interact with 100% of their marketable order flow. Whether or not a decision to route non-marketable limit orders to exchanges with high make rebates maximizes limit order execution quality, however, is an empirical question. If limit order execution quality is inversely related to the level of maker rebates, a broker routing most of its nonmarketable limit orders to an exchange with the highest maker rebate might not fulfill its obligation to obtain best execution for customer orders. In this section, we use Rule 606 data to present evidence that some U.S. retail brokerages make order routing decisions that appear to maximize per share maker rebates. How these routing decisions affect limit order execution quality is addressed in subsequent sections of the paper.
7

See FINRAs May 2011 Regulatory Notice 11-24. 8

We begin by identifying brokers appearing in either Barrons or Smart Moneys 2012 Broker surveys. We next use broker websites to collect Rule 606 reports for the fourth quarter of 2012.8 Rule 606 requires that brokers reveal the fraction of their orders that were non-directed (e.g., the customer did not choose the routing destination) and to report the percentage of non-directed orders that were market orders, limit orders, and other orders. A broker must also report the percentage of market, limit, and other orders routed to any venue receiving at least 5% of the brokers non-directed orders. Overall, nine of our brokers route at least a portion of their orders to market makers that provide rebates for the marketable orders they interact with. For brokers that do not want to delegate the handling of customer limit orders in NYSE-listed securities to market makers (and potentially reduce payments for marketable orders) there are several options. There were twelve traditional and four inverted trading venues in 4Q2012.9 The three venues charging the maximum permissible take fee of $0.30 per hundred shares were DirectEdge X (EDGX), Nasdaq Stock Market (NDAQ), and the NYSE-Arca Exchange (ARCA). Of these venues, EDGX offered the highest liquidity rebates. Conversely, the Nasdaq OMX BX (BSX) funded payments of $0.14 per hundred shares to liquidity demanders by charging liquidity supplying limit orders a make fee of $0.18 per hundred shares [Insert Table I about here.] In Table I we document the venues to which ten top retail brokerages route non-directed orders in NYSE-listed securities during the 4Q2012.10 We also report the range of liquidity rebates/fees on each of these venues. As our broker sample size is small, we conduct no statistical tests. Our goal in this section is simply to examine the order routing decisions of popular retail brokers. Of our sample brokers, Charles Schwab, Morgan Stanley, Edward Jones, Just2Trade, and LowTrade route 100% of their non-directed orders to market makers that purchase order flow. By routing all of their orders to market makers, a broker ensures that its limit orders can trade against the purchasers

8 9

See Weber (2003) for an analysis of order routing behavior. Fees are obtained from Traders Magazine and from www.sec.gov. 10 For brevity, we do not include order routing decisions for Nasdaq-listed securities. These results, which are quite similar to those presented in Table I, can be obtained from the authors upon request. 9

marketable orders. Whether or not this is preferable to routing limit orders in a different fashion is an empirical question that we cannot address with our data. Four sample brokerages, Ameritrade, E*Trade, Fidelity, and Scott Trade, route orders in a way that suggests they may be focused on liquidity rebates. Three of these brokers sell the vast majority of market orders and route limit orders either to market makers or to the venue offering the most lucrative liquidity rebates, EDGX. In addition to these venues, Scott Trade also routed some of its limit orders to the Lava ATS. The Lava ATS offers relatively high make rebates ($0.0028/share) and charges a take fee that is more attractive than EDGX. Our Rule 606 data do not distinguish between marketable and nonmarketable limit orders. Thus, we cannot determine whether the bulk of the limit orders routed to market makers (EDGX) are marketable (non-marketable). However, given these brokers choose to sell their market orders rather than route them to venues with take fees, it is likely that the majority of limit orders routed to EDGX are non-marketable. The evidence presented in Table I suggests that there is heterogeneity in order routing decisions. The routing of five of brokerages suggests that they delegate the handing of their limit orders to market makers. Interactive Brokers routing suggests that its smart router found the NYSE, the venue with the lowest non-negative make rebate, to be an attractive venue for its limit orders. However, for four of the brokerages we examine, it appears that fees are a significant determinant in where orders are routed. Whether or not the pursuit of maker fees results in diminished limit order execution quality is the empirical question we address in the remainder of the paper. IV. A. Make-take fees and limit order execution quality: Proprietary order data. Data. To examine limit order execution quality, we obtain order data from a major broker-dealers smart order routing system for October and November 2012. The data include 28,627,467 orders from the broker-dealers algorithmic trading system and orders entered directly by customers.11

11

We treat each order routed to a venue as an independent order. Although it is likely that many of these orders were generated by the same parent order, our data do not allow us to link parent and children orders. 10

Each record contains information about the order, the destination execution venue, a timestamped order history, and information about the order outcome. The order is defined by the ticker symbol, the order side (buy, sell, and short sell), the order size, the amount of that size to be displayed, the time in force (all orders are day orders), the order type (market, market on open, market on close, limit, limit or better, limit on open, and limit on close), and limit price if applicable. The broker-dealer uses seven destination venues; the New York Stock Exchange (NYSE), ARCA, BATS Z Exchange (BZX), Nasdaq OMX BX (BSX), DirectEdge A (EDGA), EDGX, and NADQ. Each order is dated and events in the orders life are time-stamped in microseconds (s - millionths of a second). Recorded events are order time and, if applicable, reject time, first fill time, last fill time, replace time, and cancel time. Should an order receive a full or partial fill, the order history contains the quantity of shares done and the average fill price. We focus on orders arriving during regular market hours (9:30am through 4:00pm) and, for filled orders, filling by 4:02pm. These restrictions reduce the sample size to 28,456,733 orders (99.4% of the original sample). The data appear to be very high quality. Checking for obvious data errors (e.g., outcome time before order time, negative size or quantity done, or quantity done exceeding order size), we exclude only 29 orders. Because we plan to focus our analysis on limit orders, we remove the 125,565 (0.4% of the remaining sample) orders that are not simple limit orders. We also discard orders with an average trade price worse (higher for buy orders or lower for sell orders) than the limit price. We use Daily TAQ data to match quote data to order arrival time (adjusting the datas s time stamps to TAQs millisecond time stamps) using the Holden and Jacobsen (2013) corrections. In order to eliminate possible erroneous orders, we subjectively discard orders with limit prices more than 10 percent from the relevant quoted price (ask price for sell orders and bid price for buy orders). We begin our analysis with 28,221,514 orders involving 6,594 unique ticker symbols. [Insert Table II about here.] Table II provides some descriptive statistics about order characteristics. From Panel A, we see that the average order arrives just prior to 1:00pm, but the open and close are quite busy. Five percent of
11

our orders are placed in the first ten minutes of the trading day and another five percent arrive in the final two minutes. Typical order size is small (mean of 635 shares), but we have some large orders (the 95 th percentile is 2,100 shares). Very little of the typical order is displayed. The mean display size is 85 shares, only about 13% of mean order size, and the majority of orders are non-displayed (more details on display are provided in Panel D). Although the typical order displays little size, there is an order with nearly three-quarters of a million shares displayed. We also have considerable variation on share price that is skewed toward low-priced shares. As shown in Panel B of Table II, about 52% of the our brokers limit orders are buy orders, 27% are sell orders and 21% are short sell orders. In Panel C, we provide additional detail on display choice. Overall, about 63% of orders are non-displayed, about 28% are fully displayed, and 9% are partially displayed. Much of the subsequent analysis separates displayed and non-displayed orders. Finally, Panel B of Table II describes where our (at least partially) displayed and hidden orders were routed.12 EDGX received the most displayed orders (29.7%), while both NDAQ and the NYSE each received more than 23%. NDAQ received the most hidden orders (39.8%), while EDGX received 32%. Overall, five (four) venues received at least 5% of our brokers displayed (hidden) limit orders. Panel B also exposes a limitation of our data in that our broker routed very few orders to inverted venues. [Insert Table III about here.] In Table III, we describe order outcomes. There are four possibilities noted in the data; cancel, fill, replace and reject (the venue rejects the order). Orders might also expire as all are day orders. A given order might have more than one outcome. For example, it is common to have an order fill partially and then have the remainder cancelled or expire. Just over 38% of the orders receive at least a partial fill and the average limit order in our sample executes 103 shares (around 16% of average order size). When an order fills at least one share, the typical order is completely filled and the average order trading any shares is 97% completed. The distribution of average execution price differs considerably from the limit price distribution (see Table II) and suggests
12

Unless stated otherwise, we refer to an order as displayed if at least part of the order is displayed. 12

that limit orders in higher priced stocks are more likely to execute. A majority (58%) of orders are cancelled, and the average (median) time to cancellation is 194 (28) seconds. Ignoring the 116,352 orders that are not actively cancelled lowers the average time to cancel to 158 seconds.13 The average (median) time it takes an order to execute at least one share is 98 (15) seconds. For orders with multiple executions, the average (median) time until the last fill is 130 (20) seconds. A very small number of orders are replaced or are rejected by the destination venue. Panel B of Table III provides a bit more detail on order outcomes conditional on whether limit orders are unfilled, partially filled, or completely filled. Of the 17,455,297 orders receiving no fills, over 90% are cancelled and 8.5% expire. Over 36% of the orders fill completely and another 2% fill partially. Of the 10,766,217 orders receiving at least a partial fill, about 77% have a single execution. There are a non-trivial number of completely filled orders with a cancellation or replacement time stamp. These apparently are instances where cancellation/replacement notice is too late to be effective. The majority of partially filled orders are cancelled. We use Daily TAQ to determine the positioning of each buy (sell) orders limit price relative to the National Best Bid (Offer). Following Rule 605, we assign limit orders to four categories: marketable, inside-the-quote, at-the-quote, and behind-the-quote. A limit order to buy (sell) shares at a price that is equal to or greater (less than or equal to) the National Best Offer (Bid) is marketable. Inside-the-quote limit orders have limit prices that improve prevailing quotes. A limit order to buy (sell) shares at the National Best Bid (Offer) is at-the-quote, and a limit order seeking to buy (sell) shares at prices that are outside the prevailing quotes are behind-the-quote. Given the difference in time stamp precision, (ms for TAQ and s for the proprietary data) we anticipate that our categorization of limit orders is imperfect but not biased in any particular direction. [Insert Table IV about here.] Columns two and three in Panel A of Table IV suggest that there is little difference in the aggressiveness of buy versus sell limit orders. Approximately 11% of the limit orders establish a new
13

The cancellation of unfilled day orders appears to come when our data feed ends at 6:00pm. 13

National Best Bid or Offer while another 60% join the NBBO. In columns four and five of Panel A, we categorize the aggressiveness of sample limit orders as a function of display choice. Consistent with the idea that investors can use quote improving orders to gain price priority (guaranteed by Reg. NMS), a higher fraction of fully and partially displayed orders are inside-the-quote compared to hidden orders. Conversely, non-displayed orders are more frequently behind-the-quote. Panel B of Table IV, describes order aggressiveness and display choice conditional on destination venue. We report the percentage of hidden (displayed) orders sent to a venue at a given pricing aggressiveness as a fraction of all hidden (displayed) orders sent to that venue. There is a clear difference in the types of limit orders sent to traditional versus inverted venues. Consistent with the idea that the inverted venues can be used to gain priority at a price, over 99% of the displayed orders routed to the inverted venues are at-the-quote. Conversely, quote improving orders are much more frequently routed to the traditional venues that offer make rebates. With the exception of the NYSE, most of the displayed limit orders received by these venues are at-the-quote (anywhere from 60.4% to 81.8%). The NYSE receives more behind-the-quote displayed orders (39.8%) than it does at-the-quote displayed orders (39.2%). Overall, at-the-quote orders are the most popular type of hidden order. As is the case with displayed orders, the NYSE receives a mix of hidden orders that differs from the other three venues offering make rebates. Over 76% of the hidden orders routed to the NYSE are behind-the-quote while 17.4% are at-the-quote. Arca has the next highest concentration of behind-the-quote hidden orders at 37.8%. We posit that venues take fees influence limit order execution quality. Specifically, the higher the take fee (and, correspondingly the higher the make rebate), the less attractive it is for liquidity demanders to use the venue. In the sample period, ARCA, EDGX, and NDAQ charged a take fee of $0.30 per hundred shares, BATS was just behind charging $0.29, and NYSE charged $0.23. The inverted venues, BSX and EDGA, paid a rebate to takers. Thus, we anticipate that the inverted venues outperform the traditional venues, that the NYSE outperforms the other traditional venues, that BATS marginally

14

outperforms ARCA, EDGX, and NDAQ, and that ARCA, EDGX, and NDAQ perform similarly. We examine these questions in the remainder of this section. B. Univariate Analysis. We first provide univariate execution-quality statistics separately for displayed (Panel A) and for hidden (Panel B) at-the-quote limit orders by venue in Table V. We focus on at-the-quote orders as it is difficult to control for the actual pricing aggressiveness of behind- or inside-the-quote limit orders in our univariate analysis. We present fill rates, execution speeds, realized spreads, and good fill ratios by venue. Fill rates are order weighted. An order is considered filled if any of the order fills, but recall that 97% of orders that fill at least one share fill the entire order. Execution speeds are in seconds from the order submission time until first fill time for filled orders. Realized spreads are presented from the perspective of liquidity suppliers. Thus, for executed limit buy orders, the realized spread is equal to twice the difference between the midpoint of the bid ask spread prevailing five minutes after the limit order executes and the executed orders limit price. For sell orders, we multiply by -1. Comparisons of realized spreads across venues illustrate whether the gross revenue earned by liquidity providers varies as a function of take fees. If limit orders on venues with low take fees routinely execute prior to those on venues with high take fees, then we expect to find that both fill rates and realized spreads are higher on venues with lower take fees. Introduced by Sofianos and Yousefi (2010), the good fill ratio classifies limit order executions relative to the sign of the realized spread. A limit order execution is classified as a good fill if the orders realized spread is positive (e.g., prices moved in the limit order traders favor after the limit order executes). [Insert Table V about here.] Focusing on fill rates for at-the-quote orders, we find that orders on the BSX are most frequently filled and orders on the high take venues (EDGX, NDAQ and ARCA) are least frequently filled. These results suggest that fill rates are negatively correlated with take fees. Likewise, the mean realized spread and the good fill ratio for displayed orders are highly negatively correlated with take fees. Only the inverted venues have positive mean realized spreads.
15

As hidden orders are by definition not displayed, it is difficult to predict the impact of take fees on hidden limit order execution quality. Consistent with this observation, we find that the correlation between take fees and fill rates for at-the-quote non-displayed orders is much lower than the correlation between take fees and fill rates for at-the-quote displayed orders. However, even for the hidden orders, the two inverted venues have the highest fraction of good fills. Overall, for at-the-quote limit orders, we find that execution quality is decreasing in take fee, especially for displayed orders. This suggests that order routing decisions may have an impact on limit order execution quality. C. Multivariate Analysis. It is unlikely that our sample limit orders are distributed randomly across trading venues. Presumably, limit order routing decisions are made conditional on several factors including stock-specific characteristics, market conditions, and liquidity fees. As a result, it is potentially misleading to draw conclusions from our unconditional statistics. To address this concern, we conduct multivariate analyses to investigate whether a given order fills and, for filled orders, the time required to fill. For each analysis, we conjecture that the relevant independent variables are those dealing with order characteristics, stock attributes, the time of day, and venue traits. We distinguish between orders based on the side the order takes (distinguishing between sell and short sell orders), whether the order is displayed, the size of the order relative to the relevant quoted size (bid size for sell orders and ask size for buy orders), and the limit price in relation to the relevant quoted price (for buy orders the limit price is compared to the bid price and for sell orders the limit price is compared to ask price). Stock characteristics include a proxy for the intensity of trading, the price level, and the volatility of the stock price. Given the frequency with which trading variables are characterized by extremes at open and close, we compute a seconds from mid-day variable to represent time of day. The two venue variables of interest are the average response time for cancellations (a proxy for the technological connectivity of the broker to the exchange) and, of direct interest to our work, the take fee. Thus, we use probit to estimate equation (1) and OLS to estimate equation (2). Both equations are presented below:
16

Fill

= 0 + 1 Sell + 2 Short Sell + 3 Display + 4 Moneyness + 5 Trading Intensity + 6 Price + 7 Volatility + 8 Seconds from Mid-day + 9 Mean Response Time + 10 Order Size + 11 Take Fee + 12 (Display)(Take Fee) + (1)

and Speed = 0 + 1 Sell + 2 Short Sell + 3 Display + 4 Moneyness + 5 Trading Intensity + 6 Price + 7 Volatility + 8 Seconds from Mid-day + 9 Mean Response Time + 10 Order Size + 11 Take Fee + 12 (Display)(Take Fee) + where Fill equals 1 if the order at least partially fills and 0 otherwise; Speed is the number of seconds between Order Time and First Fill Time conditional on at least a partial fill; Sell equals 1 if the order side is Sell and 0 otherwise; Short Sell equals 1 if the order side is Short Sell and 0 otherwise; Display equals 1 if the order is at least partially displayed and 0 otherwise; Moneyness equals (Limit Price/Ask Price) 1 for sell orders and (Bid Price/Limit Price) 1 for buy orders; Trading Intensity equals the number of orders in the proprietary dataset for this stock symbol; Stock Price equals the mean trade price in the proprietary data set for this stock symbol; Volatility equals the standard deviation of the trade price in the proprietary data set for this stock symbol; Seconds from Mid-Day equals the absolute value of Order Time (in seconds since 12:00am) 45900; Mean Response Time equals Out Time from the venue Cancel Time from the broker; Order Size is the number of shares specified in the order; and Take Fee equals the venues take fee during the sample period. We present the results from performing these regressions in Table VI. Given the sample size, all coefficients are statistically significant at beyond the .0001 level. [Insert Table VI about here.] As can be seen in the second column of Table VI, sell orders are more likely to fill than are buy orders while short sell orders are less likely to fill than buy orders. Consistent with the univariate results, displayed orders, small orders, and orders more aggressively priced are more likely to fill. Orders submitted early or late in the trading day are more likely to fill, and orders in stocks with greater trading intensity and higher volatilities are less likely to fill. Venues with faster links to the broker are more likely
17

(2)

to fill. Of particular interest is the association between take fees and the likelihood of filling. For displayed orders, the higher the take fee, the less likely the order is to fill (the sum of the estimate of 11 and 12 is reliably negative). However, for non-displayed orders, the higher the take fee the more likely the order is to fill. The relationship between fill rates and take fees are broadly consistent with the univariate results presented in Panels A and B of Table V. Fill times are examined in the third column of Table VI. Sell orders fill slower and short sell orders fill faster than buy orders. Displayed orders, smaller orders, and orders more aggressively priced fill faster than non-displayed, large and passive orders. Stocks with larger share prices, higher trading intensity and more volatility have faster fills. Fills occur more quickly away from mid-day. Perhaps not surprisingly, venues with quick connections fill orders in less time than those with slower connections. Again, focusing on take fees, we find that orders sent to venues with higher take fees fill more slowly (regardless of whether they are or are not displayed). Overall, we document that the execution quality of limit orders is related to the destination venues take fee. We find that displayed orders on venues with high take fees fill less often than similar orders on venues with low take fees. Non-displayed orders fill more frequently on venues with high take fees. Conditional on filling, orders fill more quickly on venues with low take fees. Together, these results suggest that automatically routing all limit orders to an exchange with the highest take fee is not optimizing limit order execution quality. D. Horseraces. Our proprietary data allow another approach to examining differences in across-venue limit order execution quality while controlling for other factors. We frequently find that pairs of identical orders are simultaneously sent either by the brokers order routing algorithm or directly by customers to different venues. For each pair of identical orders, we define the venue filling the order first while the competing venue still holds the paired order to be winner of the horserace. To shed light on the economic implications of winning a set of horseraces, we examine the percentage of limit order executions that produce positive realized spreads on each of the paired venues. Assuming that at a given price, limit
18

orders that execute first incur lower adverse selection costs, we expect a larger proportion of the identical limit orders executed on the low fee venue to have a positive realized spreads. In other words, the low fee venue should have the higher good fill ratio. To begin this analysis we identify order-pairs that have the same stock symbol, same order date, same order side (buy or sell), same limit price, same order time (to within one ms), but different destination venues. These paired orders, by construction, control for stock characteristics and market conditions as we attempted to do above statistically. For a horserace to produce a winner, we require that at least one of the two orders in the order-pair fill (at least partially) and that both orders in the pair be outstanding at the time of the first fill. Should both orders fill, a venue wins if it fills the order first at least 500 s before the second order executes.14 Should one order in the pair fill and the other be cancelled or replaced, the venue filling the order wins. If one order in the order-pair is rejected, replaced, or cancelled prior to its paired order filling, then we eliminate that pair from our sample. This allows us to focus on order pairs where the losing order retained an apparent trading interest when the competing venue filled the order. We report detailed results of our horseraces in the Appendix conditional on which venue receives the identical order in an order pair first. Figure 1 summarizes the outcomes of our pair-wise horseraces between venues with different take fees using order pairs constructed from all orders. There are sufficient data to make eight pairwise comparisons. [Insert Figure 1 about here.] For each venue-pair, Panel A of Figure 1 presents the number of order pairs routed to and the percentage of order pairs for which both orders execute. Panel A also describes the difference in take fees on the paired venues. The number of identical order pairs routed to different venue pairs varies significantly. For example, the NYSE and NDAQ receive 79,159 identical order pairs while the NYSE and EDGX receive only 719. Take fee differentials range from $0.0001 per share for BZX vs. ARCA,

14

Our data provider considers orders filled within 500 s of one another to be ties. Our results are not sensitive to the exact definition of a tie. 19

BZX vs. NDAQ, and BZX vs. EDGX to $0.0044 per share for BSX vs. NDAQ. Over 60% (less than 10%) of the identical order pairs routed to the NYSE and ARCA (BSX and NDAQ) execute on both venues. For each venue pair, Panel B of Figure 1 presents the percentage of horseraces won by the venue with the lower take fee. A venue wins a horserace if it is first to execute the identical order.15 Panel B reveals that for seven of the eight sets of horseraces, the venue with the lower take fee won more than 50% of the horseraces. The largest margin of victory involves the venue pair with the largest fee differential: BSX vs. NDAQ. BSX wins nearly 100% of the horseraces that do not end in a tie. In the venue pair with the second highest fee differential, the NYSE wins 93.79% of its horseraces with EDGX. The only set of horseraces in which the venue with the higher take fee executes a larger percentage of identical orders first is BZX vs. ARCA. Despite having a take fee that is $0.0001 per share higher than the BZXs, ARCA wins over 76% of the horseraces. Overall, the data presented in Panel B suggest limit order execution quality is lower on venue with higher take fees. We present good fill ratios for each of the eight venue pairs in Panel C of Figure 1. The results are consistent with those in Panel B in that the venue winning a higher percentage of its horseraces has the larger good fill ratio. In particular, the set of horseraces between venues with the largest fee differential, NDAQ and BSX, also produces the largest difference in good fill ratios. Nearly 53% of the identical limit orders executed on the BSX had positive realized spreads compared to just under 42% on NDAQ. In addition, although ARCA won 76% of its horseraces with the BZX (see Panel B), ARCAs good fill ratio is only 0.39% higher than the BZXs. Overall, the results presented in Panels B and C of Figure 1 are consistent with the claim that, for fee differences larger than one penny per 100 shares, limit order execution quality is inversely related to take fees. In only one set of horseraces does the venue with the lower take fee fail to win a majority of its contests. These results confirm our earlier finding that limit order execution quality is inversely related to
15

The number of horseraces resulting in a tie for each venue pair can be inferred by comparing the number of orderpairs in Panels A and B. For example, there were 1,698 1,433 = 265 identical orders routed to the BZX and ARCA that were executed within 500 s of one another. 20

the relative level of make/take fees. Our results also suggest that the commonly used strategy of using EDGX as the sole venue to display limit orders (see Table I) results in diminished execution quality. [Insert Figure 2 about here.] To determine whether the decision to display an order affects the results, we repeat the analysis for only those order pairs in which portions of both orders are displayed. The results are Figure 2. Given the limited number of horseraces, the results in Panels B should be interpreted with caution and are presented to serve as a benchmark against which to evaluate the results we obtain using all orders. There are six different sets of horseraces presented in Panel B. For five of the venue pairs, the venue with lower fee clearly wins a higher percentage of the horseraces. In the sixth, between the NYSE and the BZC, the lower priced NYSE only wins 51.27% of its horseraces. Further, when we focus on displayed orders, the BZX vs. ARCA horserace goes in the predicted direction with BZX winning 69% of the horseraces. Restricting our analysis to displayed order pairs produces results that are largely consistent with our analysis of order pairs constructed from all orders. In most cases, we find that when identical displayed orders are routed to venues with different take fees, the venue with the lower take fee tends to execute more quickly. Thus, are results appear to be robust to the order display decision. E. Caveats. Overall, the results of this section suggest the decision to route the bulk of ones limit orders to a single venue charging the maximum permissible take fee, and correspondingly offering high liquidity rebates, is inconsistent with a brokers fiduciary responsibility to obtain best execution. While suggestive, there are a few reasons why our results may not generalize. First, the executed limit orders in our dataset are from a single broker and they comprise only 1.5% of average daily volume. Second, our data do not span the thirteen U.S. stock venues that utilize make-take or inverted make-take fee schedules. Specifically, our data provider uses inverted venues somewhat sparingly. Finally, our proprietary data do not allow us to examine the types of stocks/situations in which the routing of limit orders to venues with

21

relatively high take fees is most (least) harmful to limit order execution quality. In the next section, we use the NYSEs TAQ database to address these concerns. V. Make-take fees and limit order execution quality: NYSE TAQ data. The NYSEs TAQ database contains a time-stamped recording of each instance that a trading venues best bid or offer price changes or the depth at these prices changes. TAQ also contains a timestamped recording of every trade. Thus, TAQ data can be used to determine whether fees influence where marketable orders (as evidenced by trades) execute when multiple venues have the best posted quote.16 Based on the analysis in the prior section, we expect that venues with lower or negative take fees receive a larger share of marketable orders than venues with high take fees when all venues display the best quote. Unless a venue charging the maximum take fee always has the shortest queue when it is at the best quote or the market always exhausts the aggregate depth at the best quote, such a finding would suggest that brokers seeking to obtain best execution for customer limit orders should not route 100% of their limit orders to an exchange charging the maximum take fee. Since exchanges today predominately operate as electronic limit order books, marketable orders typically execute against standing limit orders.17 As a result, TAQ data are well suited for making inferences as to the execution quality of executed limit orders. By examining the realized spreads of executions that take place at the quoted price when each of the relevant trading venues is posting the best quote, we can examine whether executed limit orders on venues with high take fees face higher adverse selection risk than those executed on venues with low or negative take fees. By examining the percentage of limit order executions for which the five-minute realized spread is positive (e.g., the good fill ratio), we can examine whether the average limit order on each venue encounters short-term ex post regret. To investigate how the routing decisions might be influenced by the decision to pass on make-take fees to the
16

Bessembinder (2003) examines quote-based competition for orders in NYSE-listed securities in June 2000. During this time period, virtually no trading volume in NYSE-listed securities was executed on venues offering make-take pricing. Among other things, he finds that venues realize increased market share when their quote has time priority and when the depth at their quote is large. 17 On venues such as NDAQ, NSX, and the NYSE, some of the reported trades be the result of a liquidity demander interacting with a dealers non-displayed trading interest. For example, internalized trades may be reported to the NSX and to NDAQ. 22

investor, we also assess realized spreads net of make-take fees. If the marginal investor is responsible for the fees and rebates her trades generate, one might expect the adjusted realized spreads to equilibrate across venues. A. Summary Statistics. We currently examine the first five trading days of October 2012 for both NYSE- and Nasdaqlisted stocks.18 For each stock, we determine the best bid and offer for each trading venue in TAQ throughout each trading day. Using these quotes, for each stock we compute the National Best Bid and Offer (NBBO) at each point in time during the trading day.19 Using the NBBO, we employ the Lee-Ready algorithm to determine whether trades are initiated by liquidity demanders or liquidity suppliers. As we are interested in differences in limit order execution quality across exchanges, we eliminate sweep trades since they are designed to trade against all orders posted at the best quote.20 Finally, we eliminate the Nasdaq TRF from our analysis since most of its trades are internalized. Because of the expanded data, we can conduct separate analyses for stocks listed on the NYSE and for stocks listed on Nasdaq. Since the NYSE does not facilitate trades in stocks listed on Nasdaq, the relationship between fees and limit order execution quality may depend on where a stock is listed. Panel A (Panel B) of Table VII presents summary statistics for trading activity, trading costs, and adverse selection costs in NYSE-listed (Nasdaqlisted) stocks. [Insert Table VII about here.] The first four rows of each panel describe the percentage of the trading day that each venues quotes are equal to the NBB, the NBO, either the NBB or the NBO, and the NBB and the NBO, respectively. Blume and Goldstein (1997) document that, for NYSE-listed stocks, the NYSE is at either the NBB or the NBO an average of 99.9% of the trading day and executes 83% of the trades in NYSE-

18 19

We are currently expanding the TAQ analysis to include the entirety of October and November 2012. These screens are similar to those used by Corwin and Schultz (2012). 20 Regulation NMS allows an investor to walk up (down) a venues limit order book if the investor first sends an intermarket sweep order (ISO) to each venue whose bid (offer) is equal to or better than the venues best bid (offer). The execution quality of at-the-quote limit orders that provide liquidity to ISOs is identical since ISOs consume all of the liquidity at a price point. For an in-depth analysis of ISOs, see Chakravarty et al. (2012). 23

listed securities during the twelve months ending in June 1995. Despite the fact that it executes less than 20% of all trades during our sample period, the NYSE is at either the NBB or the NBO an average of 91.37% of the trading day in October 2012. Moving along the first row of Panel A, we see that four traditional exchanges are at one side of the NBBO at least half of the trading day. The percentage of time spent by inverted venues at either the NBB or the NBO ranges from a low of 25.18% on EDGA to a high of 46.01% on BYX. The next four rows of Panel A describe the percentage of all trades, all shares, non-sweep trades, and non-sweep shares in NYSE-listed securities executed by each trading venue. In the analysis to follow, we exclude venues with market shares of less than 1%. Excluding sweep trades, three venues execute more than 10% of trades in NYSE-listed securities: the NYSE, ARCA, and NDAQ. Of the remaining traditional venues, BZX executes around 8% of non-sweep trades and EDGX has a market share of 6%. Market shares on the inverted venues range from 1.6% on EDGA to 3.3% on BYX. Overall, the venues with the larger market share of NYSE-listed trades are also the venues whose quotes are most frequently at the NBBO. The final six rows of Panel A present unconditional execution quality statistics. On average, relative effective spreads are lower on the traditional exchanges than they are on the inverted venues. For example, average effective spreads on the traditional venues range from a low of 6.76 bps on the BZX to a high of 7.68 bps on the NYSE. EDGAs average effective spread is 8.5 bps is the lowest of the i nverted venues. The fact that, on average, dollar effective spreads are lower on inverted venues suggests that this result may be due to differences in the average price of the stocks traded on inverted and traditional venues. Unconditional realized spreads provide an estimate of the gross revenue earned by liquidity providers on the various exchanges. Consistent with Chakravarty et al. (2012), who find that ISO orders tend to be informed, average realized spreads on each venue are higher when we exclude ISO trades. More importantly, we find that both dollar and percentage average realized spreads are monotonically decreasing in take fees. The BSX, with a take rebate of $0.14 per hundred, has an average realized spread
24

of $0.0068, while the three venues charging the maximum permissible take fee have average realized spreads ranging between -$0.0032 and -$0.0055. With one exception, average realized spreads are negative on each of the traditional venues and are positive on each of the inverted venues. For executed limit orders in NYSE-listed securities, these results suggest that limit order execution quality is decreasing in the size of the take fee. More important, of the seven relevant make-take venues we examine, liquidity providers on EDGX earn the lowest realized spreads.21 This again suggests that routing limit orders to a single venue charging the maximum permissible take fee may be inconsistent with the pursuit of best execution. We present these results separately for Nasdaq-listed stocks in Panel B since the NYSE does not quote or trade these securities. Consistent with the results for NYSE-listed securities, the three venues charging the maximum take fee are most frequently at either the NBB or the NBB while the three inverted venues spend the least time at the best quote. The traditional (inverted) venues execute nearly half (just over 7%) of non-sweep trades. As is the case in NYSE-listed stocks, average realized spreads for trades in Nasdaq stocks are positive on the three inverted venues and are negative on the four traditional venues. On average, the BSX again has the highest realized spreads and EDGX has the lowest. Overall, the results presented in Panels A and B suggest that, on average, limit orders executed on venues with high take fees generate negative short-term gross returns. While the exclusive routing of non-marketable limit orders to these venues by brokers who do not pass on fees/rebates to their customers may maximize fee revenue, our univariate statistics suggests this type of routing does not maximize limit order execution quality. B. Across-venue quality of limit order executions when all venues are at the inside quote. We next examine whether fees influence which venue receives a marketable order (evidenced by trades) when multiple venues are at the best quote. Since marketable orders have the choice of obtaining liquidity on one of multiple venues, we expect the limit order routing decision to be more important in these situations. How often does this situation arise? In untabulated results, we find that 12.25% of all
21

The relevant trading venues include Nasdaq BSX, EDGE-A, BATS-Y, NYSE, BATS-Z, Nasdaq, NYSE ARCA, and EDGE-X. Amex, NSX, Chicago, PSX, and CBOE have less than one percent average market share and are excluded from the analysis. 25

non-sweep trades in NYSE-listed stocks occur at the quote when each of the relevant trading venues is at the inside quote. The average (median) NYSE stock has 3.97% (0.24%) of its non-sweep trades executing at the quote when all venues are at that quote. The 90th percentile (maximum) NYSE-listed stock has 15.90% (42.30%) of its non-sweep trades executing at the quote when each of the relevant venues is at the quote.22 For trades executed at the NBB (NBO) when all venues are at the NBB (NBO), we compute the following statistics: each venues market share of trades, each venues average depth at the relevant quote when a trade executes at that quote, each venues average depth at the relevant quote when it executes a trade at that quote, each venues average realized spread, each venues average realized spread net of the make rebate or fee, and each venues good fill ratio. Assuming the quotes on each venue represent limit order trading interest, these statistics allow us to evaluate limit order execution quality for executed atthe-quote limit orders. We can then examine the extent to which limit order execution quality is associated with take fees. [Insert Table VIII about here.] In Table VIII we present execution quality statistics for at-the-quote limit orders executed when all venues are at the relevant quote separately for NYSE-listed and Nasdaq-listed securities. As noted in the bottom two rows, the requirement that trades in NYSE-listed (Nasdaq-listed) stocks occur when eight (seven) venues are at the inside quote reduces the number of stocks in our sample from 1,210 to 850 (from 1,960 to 831). On average, each NYSE-listed (Nasdaq-listed) stock remaining in our sample has 4,254 (2,586) trades executed at the inside quote when all venues are at that quote. While the BSXs market share of all non-sweep trades is 2.80% (see Panel A of Table VII), it climbs to just over 9% when we focus on trades executed when all venues are at the inside quote. Indeed, the aggregate market share of trades executed by inverted venues increases from 9.86% to 22.82% when we restrict our analysis to those non-sweep trades executed when all are at the inside quote. Conversely,

22

We find qualitatively similar results for non-sweep trades in Nasdaq-listed securities. 26

the market share of each of the traditional venues falls when this restriction is imposed. Consistently with Cardella et al. (2012), these results suggest that fees influence the routing of marketable orders. A comparison of columns three and four of Table VIII reveals that the inverted (traditional) venues tend to have more (less) depth when they execute trades than they do when trades execute on other venues at the best quote. Overall, average depths at the inside quote on traditional venues are two to three times the size of the average depths on the inverted exchanges. Ranking the average depth when a venue executes a trade while all are at the inside quote, we find that only the queue length on the NYSE is larger than that on EDGX. Thus, even if marketable orders are randomly allocated across venues at the inside quote, the results suggests brokers seeking to maximize limit order execution quality should not route 100% of their orders to an exchange charging the maximum permissible take fee. A comparison of the realized spreads presented in column six of Panel A of Table VIII with those presented in Table VII reveals that restricting our analysis to non-sweep trades executed when all are at the inside quote increases the average realized spread on each trading venue. The across-venue range in average realized spreads increases from 6.35 bps when we examine all non-sweep trades to 9.5 bps when we examine non-sweep trades executed when all are at the inside quote. This suggests that when prices become locally less volatile, order routing has a greater impact on the quality of limit order executions. When we restrict attention to trades executed when all are at the inside quote, we find at-thequote limit orders executed on the NYSE generate higher average realized spreads than every other venue except the inverted BSX. We also find EDGX has a higher realized spread than BZX, despite the fact that the make rebate (take fee) on EDGX is higher. In contrast, the good fill ratios on the inverted exchanges are all at least as high as the good fill ratio on the NYSE. This suggests that the NYSEs average realized spread may be unduly influenced by outliers. Focusing on the good fill ratios presented in column eight, we see that on average, at-the-quote limit order executions do not generate ex post regret on the three inverted venues or on the NYSE. Conversely, the average at-the-quote limit order executions on BZX, NDAQ, ARCA, and EDGX when all venues are at the inside quote generate negative five-minute returns (e.g., the good fill ratios are less than 50%). The average realized spread on each of these venues is more
27

than 5 bps less than that on the inverted venues and the NYSE. For brokers that do not pass fees and/or rebates directly through to their customers, these results suggest routing non-marketable limit orders in NYSE-listed securities only a single venue charging the highest take fee results in diminished limit order execution quality. For brokerages such as Interactive Brokers that pass fees/rebates onto their customers, the realized spread net of the make fee/rebate is a more appropriate measure of execution quality for executed limit orders. We present the average net realized spread for each venue in column seven of Table VIII. Since average realized spreads are highest (lowest) on the inverted (traditional) venues, it is not surprising that the across-venue dispersion in average net realized spreads is 2.26 bps smaller than the across-venue dispersion in average realized spreads. Once the differential costs associated with providing liquidity on the various exchanges are taken into account, we see that the NYSE offers at-the-quote limit orders the best executions. Indeed, according to this metric, at-the-quote limit orders submitted to the three inverted venues, to the NYSE, and to EDGX generate positive average net realized spreads. The last six columns of Table VIII provide similar statistics for at-the-quote trades in Nasdaqlisted securities when all venues are at the inside quote. Each of the inverted venues offers higher average realized spreads and higher good fill ratios than the traditional venues for executed at-the-quote limit orders when all venues are at the inside quote. Indeed, this relationship even holds for realized spreads net of take fees. Again, these statistics make it hard to rationalize a decision to route the bulk of ones nonmarketable limit orders to a single venue when that venue charges the maximum permissible take fee. If fees were the only determinant of where marketable orders are routed, we expect the BSX (with the largest take rebate) to execute 100% of the trades when all are at the inside. Other factors, however, likely influence the routing of marketable orders. For example, a liquidity demander may route her order to a venue with a relatively high take fee if there is a greater possibility of price improvement or if there is greater execution certainty at the limit. In untabulated results, we find that over 99% of the trades executed at the quote when all venues are at the inside quote occur when the width of the NBBO is equal to $0.01. This suggests that the expectation of price improvement is not a primary factor in
28

determining which venues executed the trades in our sample. To address the latter concern, we examine limit order execution quality statics for trades executed when each venue is at the inside quote and has a depth of at least 2,000 shares (results not reported). With the exception of the NYSE, the inverted venues execution quality for executed at-the-quote limit orders in NYSE-listed securities improves relative to the traditional venues as we impose more stringent depth requirements. Surprisingly, the NYSE has the highest average realized spread and good fill ratio when the 2,000 share depth requirement is imposed. For at-the-quote trades in Nasdaq-listed securities when all are at the inside quote with increased depth, the difference between good fill ratios and average realized spreads on the inverted and the traditional venues becomes more pronounced. Overall, with the exception of trades executed by the NYSE in NYSE-listed securities, the results in Table VIII suggest there is a monotonic relationship between take fees and the execution quality of atthe-quote limit order executions when all venues are at the inside quote. For at-the-quote limit order executions in NYSE-listed securities, the NYSE (which charges a take fee of $0.23 per hundred) routinely outperforms the inverted venues and the BZX underperforms venues with higher take fees. We might expect this result if some brokers/market participants do not find it economical to connect to the inverted venues. For these participants, the NYSE may be the venue with the lowest take fee when deciding where to route their marketable order flow. For Nasdaq-listed securities, we find the venue with the highest take rebate (fee) offers the best (worst) execution quality for at-the-quote limit orders executed when all venues are at the inside quote. Ignoring make rebates/fees, our results suggest that routing non-marketable limit orders to the inverted rather than traditional venues improves execution quality. For brokers that pass fees/rebates on to their customers, executed at-the-quote limit orders in NYSE-listed securities receive the best treatment on the NYSE. This is not the case for Nasdaq-listed securities. Even after make rebates/fees are incorporated into realized spreads, we find at-the-quote limit order executions in Nasdaqlisted securities are the best on the inverted venues.

29

C.

Multivariate analysis of take fees and realized spreads for at-the-quote trades. In the prior section we attempted to hold all else equal by examining at-the-quote limit order

execution quality when all venues were at the inside quote. Admittedly, this sample selection process tilts our analysis toward instances when fees matter most price volatility is relatively low and depth at inside quote is relatively high when all venues are required to be at the inside quote. It is exactly in these instances that we expect (and find) fee differentials to have the biggest impact on limit order execution quality. How does the quality of limit order executions as a function of take fees in more general settings? Are across-venue differences in execution quality greater for low priced stocks, where the minimum variation is a bigger percentage of the stock price? Does the routing decision affect the execution quality of limit orders in volatile stocks? In this section, we use pooled regressions of realized spreads on stock and trade characteristics to examine these questions. We ultimately present the results of four realized spread regressions for all at-the-quote trades and for at-the-quote trades executed when all venues are at the relevant inside quote separately for NYSElisted and for Nasdaq-listed securities. In each of our analyses, a trade is the unit of observation. In each specification, we include intraday time period dummies and controls for average daily volume, trade price, trade size, and daily volatility as measured by the log ratio of the high-to-low stock price. We expect adverse selection risk to be lower for smaller trades in more active stocks. We expect the at-thequote limit order routing decision to be more important in when there depth at the relevant quote is large. Thus, we expect the negative relation between take fees and realized spreads to be greater in lower priced stocks (where the minimum tick size of $0.01 is more likely to be binding) and in less volatile stocks. We define several take fee interaction variables. In one model we include an interaction variable that equals the take fee of the venue executing the trade if the executing venue has the lowest take fee of all venues at the relevant quote when the trade is executed and equals zero otherwise. We expect realized spreads to be decreasing in take fees. However, holding fee constant, we expect realized spreads to be higher when the executing venue offers the cheapest liquidity net of take fees. In a second model, we include an interaction variable that equals the take fee of the venue executing the trade if the aggregate
30

depth at the relevant quote is at least 20,000 shares and equals zero otherwise. We expect the benefit of having a limit order executed on a venue with low take fees to be greater when there is more depth at the relevant quote. In a third model, we include both the take fee of the venue executing the trade and an interactive variable that is equal to the take fee of the venue executing the trade if the trade price is between $1.00 and $6.00 inclusive. We include this variable as the relation between fees and stock price may not be linear. In our final specification, we include five dummy variables that equal one if the trade executes on a venue with a take fee of $x and zero otherwise. We include dummies for take fees of -$0.04, -$0.02, +$0.23, +$0.29, and +$0.30. In this specification, the intercept captures the impact of trades executing on the venue with a take fee of -$0.14. We expect average realized spreads to be decreasing in these take fee dummies. [Insert Table IX about here.] Panel A (Panel B) of Table IX presents our regression results for trades in NYSE-listed (Nasdaqlisted) securities. Regardless of the number of venues at the relevant quote and the exact specification of the regression, realized spreads for both NYSE and Nasdaq-listed stocks tend to be higher for trades that are smaller, in lower priced stocks, in more actively traded securities, and in stocks with higher daily volatility. As expected, for both NYSE- and Nasdaq-listed securities realized spreads are decreasing in take fee but, conditional on the size of the take fee, are higher when the executing venue has the lowest take fee of the venues at the relevant quote (see column 2 of Panels A and B). This suggests venues with lower take fees, either in an absolute or a relative sense, generate more profitable at-the-quote limit order executions. When we restrict our analysis to at-the-quote trades executed when all venues are at the quote, we see that the inverse relation between realized spreads and take fees becomes much stronger when there are at least 20,000 shares offered at that quote (see column 7). Results presented in columns three and seven of Panels A and B suggest that all else equal, the at-the-quote limit order routing decision is more important in lower priced stocks as average realized spreads are significantly lower for at-the-quote trades in low-priced stocks executed on venues with high take fees.
31

In the fifth and the ninth columns of Table IX, we present regression results for the model that includes individual take fee dummy variables. Focusing on Panel A, the results presented in column five suggest that when we examine all at-the-quote trades in NYSE-listed securities, trades executed on the BSX generate realized spreads that are, on average, significantly higher than the realized spreads generated on venues with higher take fees. Indeed, with the exception of the -$0.04 and the -$0.02 fee regimes, statistical tests (not reported) reveal the average realized spread generated by at-the-quote trades is decreasing in take fee. When we restrict our focus to trades in NYSE-listed securities executed when all venues are at the relevant quote (see column nine), our inferences change as realized spreads executed under the -$0.14 and the +$0.23 fee regimes are statistically indistinguishable. Likewise, average realized spreads executed under the -$0.04 and the -$0.02 take fee regimes are also statistically indistinguishable. For at-the-quote trades when all venues are at the relevant quote, the statistical relation between take fees and average realized spreads is as follows: average realized spreads are lowest on the venue with the take fee of $0.29, are higher on the venues with a take fee of $0.30, are even higher on venues with take fees of -$0.02 and $-0.04, and are highest on the venues with take fees of +$0.23 and -$0.14. The fifth and the ninth columns of Panel B present results for trades in Nasdaq-listed securities. Regardless of whether the focus is on all at-the-quote trades or is only on at-the-quote trades when all venues are at the relevant quote, we find a statistically significant negative relation between take fees and average realized spreads (results not reported). In each case, the average realized spread for trades executed under a specific take fee regime is statistically different than the average realized spread for trades executed under the adjacent take fee regimes. In other words, the average realized spread generated by at-the-quote limit order executions in Nasdaq-listed stocks is monotonically decreasing in take fee. What is the economic significance of these results? The results presented in column 7 of Table IX suggest that relative to the venues with take fees of $0.30, the average at-the-quote limit order execution on the venue with the lowest take fee, -$0.14, generated a realized spread that was 7.71 bps higher for NYSE-listed securities and was 9.92 bps higher for Nasdaq-listed securities. Ignoring the fact that the probability of limit order execution is decreasing in take fee, how large of a make rebate would the
32

average at-the-quote limit order investor need to compensate her for executing on the high take fee venue rather than the low take fee venue? The trade-weighted average at-the-quote trade price in NYSE(Nasdaq-) listed securities when all venues are at the relevant quote is $14.98 ($11.80) per share. Thus, the average at-the-quote limit order execution would have to earn a make fee of $1.16 per hundred shares ($14.98*0.000771*100) for NYSE-listed securities and a make rebate of $1.17 per hundred shares ($11.80*0.000992*100) for Nasdaq-listed securities to compensate for the lower realized spreads generated by at-the-quote trades on the venues charging a take fee of $0.30 per hundred shares. Moreover, the make fee would have to be directly passed through to investors. VI. Conclusion. We present evidence that Ameritrade, E*Trade, Scott Trade, and Fidelity make order routing decisions in the 4th quarter of 2012 that appear to maximize the liquidity rebates generated from limit order executions. To the best of our knowledge, none of these brokers makes it a practice to pass exchange fees/rebates through to their customers. As a result, limit order execution quality, not liquidity rebates, should be the primary (only) factor in determining where limit orders are routed. Angel et al. (2010) hypothesize that, when multiple venues are displaying the best quote, limit orders resting on venues that pay low/negative liquidity rebates should execute before those on venues that offer high liquidity rebates. We provide the first empirical analysis of the relation between order flow rebates/fees and limit order execution quality. We show that several national brokerages made routing decisions in 4Q2012 that appear to be consistent with the objective of maximizing order flow rebates. Using both proprietary limit order data and publicly available trade and quote data, we present evidence that suggests limit orders routed to venues with lower liquidity rebates are executed faster and more frequently. We also show that, on average, limit orders executed on venues with low/negative take fees generate higher average and median realized spreads. Finally, even if fees/rebates are passed directly through to the investor, the decision to use a single venue that offers the highest liquidity rebates does not appear to be consistent with the objective of obtaining best execution. So, it appears that brokers cannot have it all.
33

References Angel, James, Lawrence Harris, and Chester Spatt, 2010, Equity trading in the 21 st century, USC Marshall School of Business working paper 09-10. Bacidore, Jeff, Hernan Otero, and Alak Vasa, 2011, Does smart routing matter?, Journal of Trading 6, 3237. Battalio, Robert, Jason Greene, and Robert Jennings, 2002, Does the limit order routing decision matter?, Review of Financial Studies 15, 159-194. Bessembinder, Hendrick, 2003, Quote-based competition and trade execution costs in NYSE-listed stocks, Journal of Financial Economics 70, 385-422 Blume, Marshall and Michael Goldstein, 1997, Quotes, order flow, and price discovery, Journal of Fiannce, 221-244. Boehmer, Ekkehart, Robert Jennings, and Li Wei, 2007, Public disclosure and private decisions: Equity market execution quality and order routing, Review of Financial Studies 20, 315-358. Cardella, Laura, Jia Hao, and Ivalina Kalcheva, 2013, Make and take fees in the U.S. equity market, University of Arizona working paper. Chakravarty, Sugato, Pankaj Jain, James Upson, and Robert Wood, 2012, Clean sweep: Informed trading through intermarket sweep orders. Colliard, Jean-Edouard, and Thierry Foucault, 2012, Trading fees and efficiency in limit order markets, Review of Financial Studies 25, 3389-3421. Corwin, Shane and Paul Schultz, 2012, A simple way to estimate bid-ask spreads from daily high and low prices, Journal of Finance 67, 719-760. Foucault, Thierry, Ohad Kadan, and Eugene Kandel, 2013, Liquidity cycles and make/take fees in electronic markets, Journal of Finance 68, 299-341. Foucault, Thierry, and Albert Menkveld, 2008, Competition for order flow and smart order routing systems, Journal of Finance 63, 119-158. Harris, Larry, 2003, Trading & exchanges: Market microstructure for Practitioners, Oxford University Press, New York, New York. Harris, Larry, 2013, Maker-taker pricing effects on market quotations, working paper, USC Marshall School of Business. Holden, Craig and Stacey Jacobsen, 2013, Liquidity measurement problems in fast, competitive markets: Expensive and cheap solutions, forthcoming in the Journal of Finance. Peterson, Mark and Erik Sirri, 2002, Order submission strategy and the curious case of marketable limit orders, Journal of Financial and Quantitative Analysis 37, 221-241. SEC, 1997, Report on the practice of Preferencing, SEC, Washington, DC.
34

Sofianos, George and Ali Yousefi, 2010, Smart routing: Good fills, bad fills and venue toxicity, Goldman Sachs Equity Execution Strats Street Smart 40, 1-9. Sofianos, George, Juan Juan Xiang, and Ali Yousefi, 2010, Smart routing going to the races: Venue toxicity comparisons, Goldman Sachs Equity Execution Strats Street Smart 41, 1-8. Weber, Bruce, 2013, What holds online markets back? An examination of U.S. stockbrokers Rule 11Ac1-6 reports, University of Delaware working paper.

35

Table I The Order Routing Decisions of Ten Retail Brokers in NYSE-Listed Securities. Charles Schwab, Morgan Stanley, Just2Trade, Edward Jones, & LowTrade

Venue Make/Take EDGX

+$0.23/-$0.30 +$0.32/-$0.29

Order Mix % Mkt % Lmt % Mkt % Lmt % Mkt % Lmt % Mkt % Lmt % Mkt % Lmt % Mkt % Lmt % Mkt % Lmt % Mkt % Lmt % Mkt % Lmt % Mkt % Lmt

Ameritrade

E*Trade

Fidelity

Scott Trade 0% 28%

Interactive Brokers

0% 49%

0% 46%

0% 28%

NDAQ
+$0.20/-$0.30 +$0.29/-$0.30

3% 7% 2% 23% 5% 14% 0% 51% 23% 47% 100% 100% 96% 45% 98% 51% 97% 57% 66% 21%

Arca
+$0.21/-$0.30 +$0.30/-$0.30

BZX
+$0.25/-$0.29 +$0.29/-$0.29

Lava
+$0.24/-$0.28 +$0.27/-$0.28

NYSE
+$0.15/-$0.23 +$0.21/-$0.23

Purchasers
$0.0/<$0.0 $0.0/<$0.0

BYX
-$0.03/+$0.02 -$0.02/+$0.02

EDGA
-$0.06/+$0.04 -$0.05/+$0.04

BSX
-$0.18/+$0.14 -$0.15/+$0.14

Notes: %Mkt (%Lmt) refers to the percentage of a brokers non-directed market (limit) orders routed to the destination. %Lmt includes both marketable and non-marketable limit orders. Brokers are responsible for deciding where non-directed orders are routed. Brokers do not have to disclose destination venues that receive less than 5% of their orders. As a result, percentages may not sum to 100%. Wholesalers/market makers/purchasers includes Citadel, Knight, Citigroup, G1 Execution Services, UBS Securities, National Financial Services, Goldman Global Markets, Two Sigma Securities, and Getco. In the first (second) row under each venue, we provide the venues most (least) attractive liquidity rebate and the venues most (least) attractive take fee. These fees are expressed in cents per hundred shares.

36

Table II Descriptive Statistics Proprietary Limit Order Data Panel A. Order time, size, and price. Variable Order Time Order Size (shares) Display Size (shares) Limit Price (dollars) Average 12:52pm 635 85 15.20 Minimum 9:30am 1 0 0.0435 25th 10:52am 100 0 14.80 Median 12:51pm 100 0 26.89 75th 3:01pm 300 100 46.09 Maximum 4:00pm 999,000 745,547 2,398.97

Panel B. Distribution of buys, sells, and short sells. Order Side Buy Sell Short Sell % of Orders 51.71% 26.88% 21.41%

Panel C. Order display characteristics. % of Orders Full Display Partial Display No Display Panel D. Order destination. Venue EDGX NDAQ ARCA BZX NYSE EDGA BSX Take Fee -$0.29/-$0.30 -$0.30/-$0.30 -$0.30/-$0.30 -$0.29/-$0.29 -$0.23/-$0.23 +$0.04/+$0.04 +$0.14/+$0.14 Displayed Order Make Rebate +$0.23/+$0.32 +$0.20/+$0.29 +$0.21/+$0.30 -$0.25/+$0.29 +$0.15/+$0.21 -$0.06/-$0.05 -$0.18/-$0.15 Hidden Order Make Rebate +$0.23 +$0.10 +$0.21 -$0.10 +$0.15 -$0.06 -$0.18 % of Displayed Orders
(N=10,392,317)

28.07% 8.75% 63.18%

% of Hidden Orders
(N=17,829,197)

29.72% 23.84% 12.61% 5.26% 26.72% 0.20% 1.65%

31.97% 39.76% 11.36% 2.15% 14.46% 0.00% 0.30%

Notes: Take fees and make rebates are expressed in cents per 100 shares. In Panel B, an order is considered to be displayed if it exposes any trading interest to the public.

37

Table III Order Outcomes Panel A. Summary statistics. # of Orders # Shares Executed: All Orders # Shares Executed: (Partially) Filled Orders Average Execution Price Seconds to First Execution Seconds to Last Execution ( > 1 Fills) Seconds to Cancellation: All Cancels Seconds to Cancellation: Actively Cancels Seconds to Order Rejection Seconds until Order is Replaced Ratio of Shares Executed to Order Size Ratio of Shares Executed to Displayed Size 28,221,514 10,766,217 10,766,217 10,766,217 2,452,149 16,431,681 16,315,329 1,682 118,010 10,766,217 4,892,093 Average 103 270 $39.85 98 130 194 158 27 26 0.97 1.51 Min. 0 1 $0.05 <1 <1 <1 <1 <1 <1 0.00 0.00 25th 0 100 $16.53 3 4 5 5 <1 1 1.00 1.00 Median 0 100 $31.08 15 20 28 28 <1 4 1.00 1.00 75th 100 200 $50.39 56 66 76 74 <1 14 1.00 1.00 Max. 745,547 745,547 $2,378.1 23,399 23,399 30,499 2,339 12,877 22,279 1.00 5,000.00

Panel B. Order outcome conditional on fill status. Fill Status No Fill Complete Fill Partial Fill # of Orders 17,455,297 10,231,469 534,748 Rejected 1,681 0 1 Replaced 117,895 8 107 Cancelled 15,853,841 53,083 524,757 One Fill n.a. 7,964,586 349,482 Expired 1,481,880 n.a. 9,883

Notes: By sample construction, outcome time is later than order placement time, no executions take place after 4:02pm, and the number of shares executed cannot be greater than an orders size. When computing seconds to cancellation for all orders, we assume open day orders are cancelled at 6:00pm. Multiple outcomes can happen to a given order in Panel B.

38

Table IV Order Aggressiveness Panel A. Order aggressiveness conditional on order side and display choice. Order Side Limit Order Type Behind-the-quote At-the-Quote Inside-the-Quote Marketable Buy Orders 14.90% 30.74% 5.48% 0.65% Sell Orders 12.32% 29.76% 5.24% 1.04% Display Choice Partially or Fully Hidden Displayed 32.21% 18.56% 58.84% 7.19% 1.88% 63.34% 16.77% 1.38%

Panel B. Order aggressiveness of orders conditional on destination venue and display status. Destination Venue EDGX
$0.30

Traditional NDAQ ARCA

$0.30 $0.30

NYSE
$0.23

BZX
$0.29

Inverted EDGA BSX

-$0.04 -$0.14

Behind-thequote At-the-Quote

Displayed Hidden Displayed Hidden Displayed Hidden Displayed Hidden

5.90%

14.91% 18.33% 39.76%

7.71%

0.10% 15.19% 99.87% 84.17% 0.00% 0.00% 0.24% 0.64%

0.22% 18.46% 99.44% 41.28% 0.22% 21.56% 0.27% 18.79%

14.13% 29.72% 37.80% 76.73% 25.50% 81.77% 60.56% 60.37% 39.19% 88.97% 79.37% 59.65% 51.30% 17.35% 59.86% 11.48% 22.42% 19.78% 20.45% 5.82% 0.92% 1.11% 8.55% 2.40% 2.22% 7.17% 1.73% 3.40% 5.33% 0.68% 0.64% 0.96% 12.88% 2.50% 1.97%

Inside-the-Quote

Marketable

Notes: A limit order to buy (sell) shares at a price that is equal to or greater (less than or equal to) the National Best Offer (Bid). Inside-the-quote limit orders have limit prices that improve prevailing quotes. A limit order to buy (sell) shares at the National Best Bid (Offer) is at-the-quote, and a limit order seeking to buy (sell) shares at prices that are more advantageous than the prevailing quotes are behind-the-quote. Traditional (inverted) venues charge (pay) liquidity demanders and charge (pay) liquidity suppliers. An order is displayed if at least part of the orders size is shown on the venues limit order book. Under each venue, we provide the take fee/rebate for the venues least preferred customer. Take fees are expressed in cents per 100 shares.

39

Table V Limit Order Execution Quality Statistics for At-the-Quote Limit Orders Panel A: Displayed orders. Venue EDGX NDAQ ARCA BZX NYSE EDGA BSX Take Fee/Rebate $0.30 $0.30 $0.30 $0.29 $0.23 -$0.04 -$0.14 # Orders 2,525,513 1,500,240 791,087 486,375 1,088,393 20,529 170,736 Fill Rate 53.96% 47.96% 53.11% 54.66% 56.75% 56.85% 74.51% Fill Speed (seconds) 111 59 61 76 40 105 33 Realized Spread -$0.0021 -$0.0020 -$0.0019 -$0.0028 -$0.0011 $0.0011 $0.0011 Good Fill Ratio 49.32% 49.19% 49.44% 50.42% 50.39% 57.36% 54.81%

Panel A: Hidden orders. Venue EDGX NDAQ ARCA BZX NYSE EDGA BSX Take Fee/Rebate $0.30 $0.30 $0.30 $0.29 $0.23 -$0.04 -$0.14 # Orders 4,523,538 4,228,990 1,038,667 229,131 447,095 654 22,311 Fill Rate 50.92% 35.64% 35.92% 44.10% 35.48% 63.15% 36.07% Fill Speed (seconds) 102 127 66 71 82 73 68 Realized Spread $0.0192 -$0.0045 -$0.0057 -$0.0046 $0.0007 -$0.0031 $0.0110 Good Fill Ratio 46.92% 47.88% 47.12% 48.65% 50.46% 52.30% 57.26%

Notes: A limit order to buy (sell) shares at the National Best Bid (Offer) is at-the-quote. An order is displayed if at least part of the orders size is shown on the venues limit order book. Under each venue, we provide the take fee/rebate for the venues least preferred customer. Take fees and make rebates are expressed in cents per 100 shares. Fill Rate is the ratio of orders that have at least one share executed to the total number of orders. Speed represents the average time that it takes executed limit orders to receive their first (and perhaps only) execution. For executed nonmarketable limit buy orders, the Realized Spread is equal to twice the difference between the midpoint of the bid ask spread prevailing five minutes after order receipt and the limit price. For sell orders, the Realized Spread is equal to twice the difference between the limit price and the bid/ask midpoint five minutes after order receipt. The Good Fill Ratio is the percentage of executed limit orders with positive realized spreads.

40

Table VI Multivariate Analysis of Limit Order Execution Quality We use Probit to examine whether fees are associated with the probability that a limit order executes. We use Ordinary Least Squares to examine the relationship between execution speed and fees. Probability of Limit Order Execution Speed: Execution Filled Orders (N = 28,221,392) (N = 10,766,180) Long Sales Short Sales Order Displayed? Average Trading Volume x 100 Stock Price Stock Price Volatility Order Placement Time x 100 Venue Response Time Moneyness Order Size Take Fee Order Display * Take Fee c/R2 0.0912 - 0.1143 2.1318 - 0.0004 0.0486 - 0.0405 0.0008 - 4.4538 - 797.3 - 0.0109 498.2 - 597.8 0.737 1.0520 - 16.5468 - 73.2609 - 0.1880 - 0.5916 - 0.4178 - 0.4730 - 390.29 59193.0 1.7543 4828.90 13353.0 0.0419

Notes: Sell equals 1 if the order side is Sell and 0 otherwise; Short Sell equals 1 if the order side is Short Sell and 0 otherwise; Display equals 1 if the order is at least partially displayed and 0 otherwise; Moneyness equals (Limit Price/Ask Price) 1 for sell orders and (Bid Price/Limit Price) 1 for buy orders; Trading Intensity equals the number of orders in the proprietary dataset for this stock symbol; Stock Price equals the mean trade price in the proprietary data set for this stock symbol; Volatility equals the standard deviation of the trade price in the proprietary data set for this stock symbol; Seconds from Mid-Day equals the absolute value of Order Time (in seconds since 12:00am) 45900; Mean Response Time equals Out Time from the venue Cancel Time from the broker; Order Size is the number of shares specified in the order; and Take Fee equals the venues take fee during the sample period.

41

Table VII Time at the Inside Quotes, Market Share, and Spread Decomposition by Trading Venue % Time at Inside/ Market Share Time at the Inside: At Bid At Ask At One Side At Both Sides Market Share: % All Trades % All Share Volume % No Sweep Trades % No Sweep Volume NYSE AMEX BSX NSX CHX ARCA NDAQ PSX BZX BYX EDGA EDGX CBOE

Panel A - NYSE Stocks (N=1210) 74.39 73.50 91.37 56.52 19.66 22.28 17.12 18.81 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 20.74 20.88 35.99 5.63 3.52 2.65 2.80 2.07 1,667 0.50 3.13 11.26 11.61 18.58 4.28 0.31 0.34 0.15 0.17 130 -0.62 -2.57 0.03 0.02 0.05 0.00 0.00 0.05 0.00 0.06 56.69 57.20 76.57 37.31 13.57 12.00 10.56 9.34 6,241 -0.60 -1.86 52.98 53.72 70.64 36.05 14.05 11.90 11.25 9.15 8,343 -0.47 -1.60 13.09 13.36 21.43 4.93 0.42 0.42 0.31 0.28 360 -0.21 -0.37 44.43 44.61 61.79 27.25 9.60 7.72 8.16 6.31 5,467 -0.36 -1.40 28.26 29.27 46.01 11.52 4.65 3.44 3.30 2.33 2,849 0.24 1.72 14.62 15.14 25.18 4.58 1.92 1.41 1.62 1.17 1,279 0.17 1.79 38.59 39.42 55.64 22.37 7.26 7.00 6.05 5.64 3,727 -0.70 -2.50 21.85 22.32 38.53 5.64 0.71 0.51 0.61 0.43 381 0.85 5.59

Bid-Ask Spreads All Trades: Total Trades (000) 5,789 Realized Spread () -0.32 Realized Spread (bps) -0.47 Bid-Ask Spreads - No Sweep Trades: Total Trades (000) 2,999 Effective Spread () 2.01 Price Impact () 2.11 Realized Spread () -0.10 Effective Spread (bps) 7.68 Price Impact (bps) 7.42 Realized Spread (bps) 0.27

749 1.88 1.20 0.68 9.57 5.50 4.07

36 1.89 2.46 -0.57 7.60 10.41 -2.81

2,642 1.85 2.35 -0.50 7.29 8.99 -1.70

3,826 1.62 1.94 -0.32 6.84 8.12 -1.29

153 1.36 1.49 -0.13 7.85 8.02 -0.16

2,631 1.55 1.78 -0.23 6.76 7.81 -1.05

1,205 1.53 1.20 0.33 8.57 6.20 2.37

675 1.41 1.14 0.27 8.73 6.28 2.45

1,937 1.48 2.02 -0.55 7.16 9.44 -2.28

204 1.64 0.76 0.88 8.42 3.16 5.26

42

Table VII (continued) % Time at Inside/ Market Share Time at the Inside: At Bid At Ask At One Side At Both Sides Market Share: % All Trades % All Share Volume % No Sweep Trades % No Sweep Volume NYSE AMEX BSX NSX CHX ARCA NDAQ PSX BZX BYX EDGA EDGX CBOE

Panel B Nasdaq Stocks (N=1960) 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 2.22 2.32 2.97 1.57 0.03 0.02 0.01 0.01 43 -0.25 -1.45 14.62 15.03 25.00 4.66 3.06 2.19 2.65 1.84 1,010 0.57 4.81 9.12 9.13 15.14 3.11 0.64 0.59 0.24 0.22 88 -0.45 -4.49 0.14 0.07 0.20 0.01 0.07 0.13 0.02 0.05 42.69 42.77 62.50 22.97 12.23 10.76 9.83 8.68 3,328 -0.37 -1.89 68.24 69.08 87.52 49.81 29.48 27.42 20.68 18.17 5,993 -0.55 -2.45 7.14 7.65 11.48 3.31 0.32 0.33 0.22 0.20 253 -0.12 -0.50 32.15 33.37 48.87 16.65 8.79 6.96 7.42 5.65 3,388 -0.33 -1.55 20.40 21.82 34.45 7.78 3.70 2.64 2.86 1.97 1,430 0.32 2.31 9.85 10.48 17.33 3.00 1.76 1.25 1.71 1.19 646 0.31 3.14 35.62 36.24 54.21 17.66 10.77 11.77 8.05 8.38 2,155 -0.72 -3.66 18.05 18.67 32.20 4.52 0.89 0.64 0.59 0.41 209 0.98 9.14

Bid-Ask Spreads - All Trades: Total Trades (000) Realized Spread () Realized Spread (bps) -

Bid-Ask Spreads - No Sweep Trades: Total Trades (000) 6 470 22 1,605 2,838 96 1,739 629 356 1,124 104 Effective Spread () 1.09 2.56 2.47 2.33 2.17 1.46 1.85 1.93 1.96 2.01 1.83 Price Impact () 1.32 1.76 2.58 2.54 2.56 1.48 2.05 1.52 1.40 2.61 0.88 Realized Spread () -0.22 0.80 -1.06 -0.21 -0.39 -0.02 -0.20 0.40 0.56 -0.60 0.94 Effective Spread (bps) 15.64 15.07 21.13 11.38 11.52 12.35 10.01 13.13 15.10 13.04 13.61 Price Impact (bps) 17.22 9.02 18.55 12.92 13.50 12.22 10.84 9.99 10.67 16.51 5.17 Realized Spread (bps) -1.58 6.05 2.57 -1.54 -1.97 0.14 -0.83 3.15 4.44 -3.47 8.44 Notes: The tables provides summary statistics for time at the inside quote, market share, and effective spread decomposition by trading venue. Time at the inside is computed for each stock-day as the proportion of seconds during the day for which the venue is equal to either the NBBO bid, the NBBO ask, or both. Market share is computed for each stock-day as the proportion of trades (volume) executed on each venue, both including and excluding intermarket sweep orders. For each stock, time at the inside and market share are averaged across all trading days during the sample period. The table then lists the cross-sectional average of these stock-specific means. For each venue, Total Quoted Depth is a trade-weighted average of bid plus ask depth across all non-sweep trades and spread components are calculated as an equal-weighted average across all nonsweep trades executed on that venue. Trade direction is determined based on the Lee-Ready algorithm. For buy (sell) trades, the Effective Spread
43

Table VII (continued) is defined as two (negative two) times the difference between the trade price and the NBBO midpoint at the time of the trade, Price Impact is defined as two (negative two) times the difference between the NBBO midpoint five minutes after the trade and the NBBO midpoint at the time of the trade, and the Realized Spread is defined as two (negative two) times the difference between the trade price and the NBBO midpoint five minutes after the trade. Percentage spread components are defined by dividing each spread component by the NBBO midpoint at the time of the trade. Panels A and B provide separate results for stocks listed on the NYSE and Nasdaq, respectively. Market share for the Nasdaq TRF are not reported. Based on all trades, TRF market share of trades (volume) is 28.25% (35.30%) for Nasdaq stocks and 24.32% (30.28%) for NYSE stocks. Based on no sweep trades, TRF market share of trades (volume) is 45.72% (53.22%) for Nasdaq stocks and 38.08% (44.23%) for NYSE stocks.

44

Table VIII Market Share of at-the-quote Trades when all Venues are at the Inside
NYSE Stocks Qtd Realized Depth Spread Own Realized Net of Trades Spread Fee (000) (bps) (bps) Nasdaq Stocks Qtd Realized Depth Spread Own Realized Net of Trades Spread Fee (000) (bps) (bps)

Take Fee ($)

Market Share (%)

Qtd Depth (000)

Good Fill Ratio (%)

Market Share (%)

Qtd Depth (000)

Good Fill Ratio (%)

BSX (B) EDGA (J) BYX (Y) NYSE (N) BZX (Z) NDAQ (T/Q) ARCA (P) EDGX (K)

-0.14 -0.04 -0.02 0.23 0.29 0.30 0.30 0.30

9.04 6.67 7.11 6.43 6.92 11.23 7.17 7.30

2.32 2.34 2.73 11.93 8.28 12.15 11.08 9.86

2.91 3.42 2.91 10.21 6.02 7.99 7.82 8.81

6.18 3.70 3.93 4.33 -3.32 -3.08 -3.25 -1.26

4.19 3.07 3.59 5.81 -1.06 -1.43 -1.42 1.00

59.16 56.31 55.49 55.48 46.85 46.52 46.39 49.15

9.16 6.03 6.83 NA 9.34 13.85 8.22 6.37

2.57 2.63 2.87 NA 11.08 21.63 11.28 11.06

3.21 5.17 2.82 NA 8.37 13.07 9.00 7.88

7.37 4.52 3.38 NA -2.81 -3.75 -2.43 -3.52

4.71 3.63 2.97 NA -0.09 -1.58 -0.06 -0.57

60.68 57.81 55.81 NA 47.90 46.49 47.50 46.78

Number of Stocks 850 831 Avg. Trades per 4,254 2,586 Stock Notes: The table provides summary statistics for market share of at-the-quote trades and trade outcomes when all venues are simultaneously at the inside quote. The analysis is limited to trades that occur at the NBBO bid (ask) when all relevant venues are simultaneously posting quotes equal to the NBBO bid (ask). The relevant exchanges considered include Nasdaq BSX, EDGE-A, BATS-Y, NYSE, BATS-Z, Nasdaq, NYSE ARCA, and EDGE-X. Amex, NSX, Chicago, PSX, and CBOE have less than one percent average market share and are excluded from the analysis. Intermarket sweep trades are excluded. Market Share is defined across all trades on all venues and Quoted Depth is defined as an equal weighted average across all trades. Quoted Depth on Own Trades, Realized Spread, and Good Fill Ratio are defined as equal weighted averages across all relevant trades on a particular venue. Quoted depth is a venues posted depth on the relevant side of the market. Good Fill ratio is defined as the proportion of trades with realized spread greater than zero. For trades at the NBO (NBB), the percentage Realized Spread is defined as two (negative two) times the difference between the trade price and the NBBO midpoint five minutes after the trade, divided by the NBBO midpoint at the time of the trade. Realized Spread Net of Fee is defined as the Realized Spread net of any Take Fee or Rebate. Results are reported separately for stocks listed on the NYSE and Nasdaq.
45

Table IX Realized Spread Regressions

Panel A NYSE Stocks At-the-Quote Trades Intercept Ln(Avg. Volume) Ln(Trade Price) Ln(Trade Size) Ln(Daily High/Low) Take Fee Take Fee * Min_Fee Take Fee * Low_Price Take Fee * High_Depth Inverted Dummy Take Fee =-$0.04 Take Fee =-$0.02 Take Fee =+$0.23 Take Fee =+$0.29 Take Fee =+$0.30 Intraday Period Dummies N Adjusted R2 F (p-value) 0.0538 (0.000) 0.0038 (0.000) -0.0083 (0.000) -0.0112 (0.000) -0.3632 (0.000) -0.1327 (0.000) 0.0159 (0.000) Yes 13.5M 0.0017 1272.19 (0.000) 0.0654 (0.000) 0.0033 (0.000) -0.0108 (0.000) -0.0108 (0.000) -0.3582 (0.000) -0.1212 (0.000) -0.0974 (0.000) Yes 13.5M 0.0018 1325.22 (0.000) 0.0238 (0.000) 0.0031 (0.000) -0.0081 (0.000) -0.0105 (0.000) -0.3644 (0.000) 0.0406 (0.000) Yes 13.5M 0.0015 1187.33 (0.000) 0.0706 (0.000) 0.0043 (0.000) -0.0084 (0.000) -0.0127 (0.000) -0.3569 (0.000) -0.0172 (0.000) -0.0187 (0.000) -0.0342 (0.000) -0.0552 (0.000) -0.0591 (0.000) Yes 13.5M 0.0018 1160.50 (0.000) At-the-Quote Trades with All Venues at the Inside 0.1153 0.0613 0.0506 0.1106 (0.000) (0.000) (0.000) (0.000) 0.0046 (0.000) -0.0356 (0.000) -0.0141 (0.000) -0.2208 (0.000) -0.1382 (0.000) -0.1794 (0.000) Yes 2.0M 0.0059 670.08 (0.000) 0.0073 (0.000) -0.0324 (0.000) -0.0134 (0.000) -0.2718 (0.000) -0.1268 (0.000) -0.0767 (0.000) Yes 2.0M 0.0056 637.37 (0.000) 0.0049 (0.000) -0.0290 (0.000) -0.0141 (0.000) -0.2447 (0.000) 0.0546 (0.000) Yes 2.0M 0.0049 591.29 (0.000) 0.0067 (0.000) -0.0291 (0.000) -0.0183 (0.000) -0.2228 (0.000) -0.0242 (0.000) -0.0205 (0.000) -0.0010 (0.574) -0.0835 (0.000) -0.0771 (0.000) Yes 2.0M 0.0066 634.92 (0.000)

46

Table IX (Continued)
Panel B Nasdaq Stocks At-the-Quote Trades Intercept Ln(Avg. Volume) Ln(Trade Price) Ln(Trade Size) Ln(Daily High/Low) Take Fee Take Fee * Min_Fee Take Fee * Low_Price Take Fee * High_Depth Inverted Dummy Take Fee =-$0.04 Take Fee =-$0.02 Take Fee =+$0.29 Take Fee =+$0.30 Intraday Period Dummies N Adjusted R F (p-value)
2

0.1465 (0.000) -0.0008 (0.000) -0.0058 (0.000) -0.0153 (0.000) -0.5087 (0.000) -0.1902 (0.000) -0.0053 (0.679) Yes 6.9M 0.0022 843.55 (0.000)

0.1708 (0.000) -0.0010 (0.000) -0.0159 (0.000) -0.0146 (0.000) -0.4735 (0.000) -0.1556 (0.000) -0.1973 (0.000) Yes 6.9M 0.0026 1009.58 (0.000)

0.0901 (0.000) -0.0008 (0.000) -0.0059 (0.000) -0.0154 (0.000) -0.5098 (0.000) 0.0682 (0.000) Yes 6.9M 0.0021 869.77 (0.000)

0.1768 (0.000) -0.0011 (0.000) -0.0056 (0.000) -0.0152 (0.000) -0.5043 (0.000) -0.0138 (0.000) -0.0278 (0.000) -0.0719 (0.000) -0.0876 (0.000) Yes 6.9M 0.0023 779.09 (0.000)

At-the-Quote Trades with All Venues at the Inside 0.0834 0.0194 -0.0285 0.0695 (0.000) (0.000) (0.000) (0.000) 0.0060 (0.000) -0.0366 (0.000) -0.0133 (0.000) 0.1116 (0.000) -0.1227 (0.000) -0.3289 (0.000) Yes 1.2M 0.0096 627.13 (0.000) 0.0073 (0.000) -0.0182 (0.000) -0.0126 (0.000) 0.0503 (0.051) -0.2014 (0.000) -0.0212 (0.000) Yes 1.2M 0.0069 449.70 (0.000) 0.0064 (0.000) -0.0177 (0.000) -0.0133 (0.000) 0.0609 (0.017) 0.0781 (0.000) Yes 1.2M 0.0066 455.57 (0.000) 0.0063 (0.000) -0.0175 (0.000) -0.0129 (0.000) 0.0591 (0.021) -0.0275 (0.000) -0.0371 (0.000) -0.0933 (0.000) -0.0992 (0.000) Yes 1.2M 0.0069 405.67 (0.000)

Notes: The table describes results from a pooled regression of realized spreads on stock and trade characteristics. The initial sample includes all trades executed during October and November 2012 on a venue with a make-take fee schedule for which the realized spread could be calculated, excluding intermarket sweep trades. We then restrict the subsample to the set of trades that take place at the quotes and the subset of at-thequote trades that occur when all relevant venues are at the inside quote. Trade direction is determined based on the Lee-Ready algorithm. For buy (sell) trades, the Realized Spread is defined as two (negative two) times the difference between the trade price and the NBBO midpoint five minutes after the trade. The percentage realized spread is then defined by dividing by the NBBO midpoint at the time of the trade. Average Volume is

47

Table IX (Continued) the average daily share volume computed over the six months prior to the sample period. Trade Price and Trade Size are the price and number of shares in the current trade. Daily High/Low is the ratio of the highest and lowest trade price for the stock associated with the current trade on the day of the current trade. Take Fee is the fee charged to liquidity takers by the venue associated with the current trade, where rebates are defined as negative fees, and take fee dummy variables identify each level of take fee. Inverted is a dummy variable equal to one if the trade occurred on a venue with an inverted fee schedule. Min_Fee is a dummy variable equal to one if the trade venue has the lowest take fee of all venues at the inside quote at the time of the trade. High_Depth is a dummy variable equal to one if the aggregate depth at the relevant inside quote is at least 20,000 shares. Low_Price is a dummy variable equal to one if the trade price is less than $6. Intraday dummy variables are used to define 30-minute intervals within the trading day. Results for NYSE and Nasdaq-listed stocks are listed in Panels A and B, respectively.

48

Figure 1 Horseraces: All Order Pairs Panel A. Percentage of order pairs in which both orders fill. For each pair of venues, the venue with the lower take fee is listed first.
100.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00%

64.73% 54.75% 34.69% 22.61% 11.13% 9.34% 39.22% 35.27%

BZX vs ARCA Fee Diff = $0.01 N = 1,698

BZX vs NDAQ Fee Diff = $0.01 N = 13,559

BZX vs EDGX Fee Diff = $0.01 N = 32,300

NYSE vs BZX Fee Diff = $0.06 N = 3,368

NYSE vs ARCA Fee Diff = $0.07 N = 743

NYSE vs NDAQ Fee Diff = $0.07 N = 79,157

NYSE vs EDGX Fee Diff = $0.07 N = 719

BSX vs NDAQ Fee Diff = $0.44 N = 3,374

Panel B. Percentage of horseraces involving all orders won by venue with lower fee (excluding ties). For each pair of venues, the venue with the lower take fee is listed first.
100% 90% 80% 70% 60% 50% 40% 30% 23.59% 68.35% 64.02% 54.45% 68.78% 60.01% 93.79% 99.74%

20%
10% 0% BZX vs ARCA Fee Diff = $0.01 N = 1,433 BZX vs NDAQ Fee Diff = $0.01 N = 9,811 BZX vs EDGX Fee Diff = $0.01 N = 24,309 NYSE vs BZX Fee Diff = $0.06 N = 2,637 NYSE vs ARCA Fee Diff = $0.07 N = 550 NYSE vs NDAQ Fee Diff = $0.07 N = 51,240 NYSE vs EDGX Fee Diff = $0.07 N = 676 BSX vs NDAQ Fee Diff = $0.44 N = 3,128

49

Figure 1 (continued) Panel C. Good fill ratios for orders executed in horseraces. 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0%
BZX $0.29 ARCA $0.30 BZX $0.29 NDAQ $0.30 BZX $0.29 EDGX $0.30 NYSE $0.23 BZX $0.29 NYSE $0.23 ARCA $0.30 NYSE $0.23 NDAQ $0.30 NYSE $0.23 EDGX $0.30 BSX -$0.14 NDAQ $0.30
55.91% 56.30% 52.13% 50.98% 53.24% 57.15% 52.98% 52.57% 49.65% 54.42% 49.83% 47.11% 41.88% 63.23% 54.08% 52.91%

Notes: An Order Pair (e.g., a horserace) involves a pair of limit orders that have the same stock symbol, same order date, same order side (buy or sell), same limit price, same order time (to within one ms), but different destination venues. Take fee is the cost of accessing 100 shares of liquidity on the given venue. The Good Fill Ratio is the percentage of executed limit orders with positive realized spreads. Take fees are expressed in cents per 100 shares.

50

Figure 2 Horseraces: Displayed Order Pairs Panel A. Percentage of displayed order pairs in which both orders fill. For each pair of venues, the venue with the lower take fee is listed first.
100.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00%

63.96% 51.24% 44.90% 46.68%

57.78%

33.80%

BZX vs ARCA Fee Diff = $0.01 N = 213

BZX vs NDAQ Fee Diff = $0.01 N = 564

BZX vs EDGX Fee Diff = $0.01 N = 4,493

NYSE vs BZX Fee Diff = $0.06 N = 407

NYSE vs ARCA Fee Diff = $0.07 N = 197

NYSE vs NDAQ Fee Diff = $0.07 N = 2,307

Panel B. Percentage of horseraces involving displayed orders won by venue with lower fee (excluding ties). For each pair of venues, the venue with the lower take fee is listed first.
100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% BZX vs ARCA Fee Diff = $0.01 N = 170 BZX vs NDAQ Fee Diff = $0.01 N = 327 BZX vs EDGX Fee Diff = $0.01 N = 2,934 NYSE vs BZX Fee Diff = $0.06 N = 310 NYSE vs ARCA Fee Diff = $0.07 N = 149 NYSE vs NDAQ Fee Diff = $0.07 N = 1,604 51.27% 85.30% 81.04% 71.95% 65.77%

69.88%

51

Figure 2 (continued) Notes: An Order Pair (e.g., a horserace) involves a pair of limit orders that have the same stock symbol, same order date, same order side (buy or sell), same limit price, same order time (to within one ms), but different destination venues. Take fee is the cost of accessing 100 shares of liquidity on the given venue. The Good Fill Ratio is the percentage of executed limit orders with positive realized spreads. Take fees are expressed in cents per 100 shares.

52

Appendix: Details on the horserace analysis. In the first two columns, we denote the venue pair, with the ordering consistent with the ordering in the data. That is, ARCA-NYSE indicates that the first identical order in these order pairs went to ARCA and the second order in these order pairs went to the NYSE. Conversely, NYSE-ARCA indicates that the first order went to the NYSE and the second went to ARCA. The third column indicates the number of pairs we find for the indicated venue pair. We see that for some venue pairs, the ordering appears to be roughly equal (e.g., there are 336 pairs in which ARCA receives the order before the NYSE and 407 pairs in which the NYSE receives the order before ARCA) while in others one venue is clearly favored (e.g., there are 30,637 pairs in which EDGX receives the order before BZX and only 1,633 pairs in which BZX receives the order before EDGX). In the fourth column, we report the percentage of order pairs in which both orders execute. In the fifth column, for the order pairs where both orders execute, we present the difference in times to execution for each of the two orders in the order pair. If the order resting at the first venue executes before (after) the order on the second venues order book fills, this difference is negative (positive). We present the outcomes of our horseraces in columns 6, 7 and 8. We declare the outcome of a horserace a tie if both orders in an order pair execute within 500 s of one another. Otherwise, the order filling first is declared the winner. Finally, in columns 9 (column 10), we document the percentage of orders executed on the first venue (second venue) that generated positive realized spreads. We present results for both displayed and hidden orders in Panel A of Appendix Table I, while Panel B contains results for only displayed orders. In each case we only consider comparisons between venues with different take fees. The first two rows present results for order-pairs involving the ARCA, which has a take fee of $0.003/share, and the NYSE, which has a take fee of $0.0023/share. Row 1 (row 2) contains order-pairs in which ARCA (the NYSE) received the first order. For over 62% of the order pairs, both orders execute, suggesting similar execution likelihood. However, for these orders the NYSEs time to exec ution was, on average, 12 to 13 seconds faster than ARCAs. In about one-quarter of the horse races between ARCA and the NYSE, both orders in an order pair execute within 500 s of one another. When ARCA receives the first order, the NYSE wins 43.75% of the horseraces compared to 31.85% for ARCA. When the NYSE receives the first
53

order, it wins 44.96% of the horseraces compared to 28.01% for ARCA. For limit orders that execute in contests between the NYSE and ARCA, the NYSEs good fill ratio exceeds ARCAs good fill ratio. Together, we interpret the data in the first two rows of Panel A as suggesting that, all else equal, our sample limit orders receive better execution quality on the NYSE than on ARCA. Rows three through eight of Panel A contain the remainder of the contests involving the NYSE. Over 79,000 pairs of identical orders are routed to the NYSE and NDAQ, which has a take fee of $0.30 per hundred shares. Both orders execute in 63% of the order pairs when NDAQ receives the first order. When the NYSE receives the first order, both orders execute only 39% of the time. Regardless of who receives the first order, the average time to execution is shorter on the NYSE. Nearly 40% of the 51,474 order pairs in which NDAQ receives the first order execute on both the NYSE and on NDAQ within 500 s of each another. The NYSE wins 64.41% (38.83/(38.83+21.46)) of the remaining horseraces and it has a higher good fill ratio for its executed limit orders. Excluding ties, the NYSEs advantage over NDAQ improves to 75. 48% when it receives the identical order prior to NDAQ. Of the horseraces between the NYSE and EDGX in which there is a clear winner, the NYSE wins 89.12% of the time when EDGX receives the order first and the NYSE wins 96.34% of the time when the NYSE receives the first order. In addition winning a large majority of the horseraces, the NYSE also has a good fill ratio that is 799 basis points higher than EDGXs when EDGX receives the first order and is 1,036 basis points higher than EDGXs when the NYSE rec eives the first order. The difference in execution quality for our sample orders on the BZX and NYSE venue pair is much smaller than the differences in execution quality on the NYSE and the three venues with take fees of $0.30 per hundred shares. The NYSE, wins 53.76% of its horseraces with the BZX (excluding ties) when it receives the second order and it wins 57.61% of the non-tied horseraces when it receives the first order. The NYSE also has a slightly higher proportion of good fills than the BZX. Together, the fact that the NYSE wins a higher percentage of its horseraces with the four venues that charge higher take fees suggests that take fees negatively impact limit order execution quality. Next, in rows nine through fourteen, we present the outcome of order pairs involving BZX and three venues with take fees that are $0.01 per hundred shares higher than the BZX take fee. As might be expected
54

given the closeness of the take fees, the execution quality differences here are much less definitive. The lowerfee BZX provides better executions (as we measure execution quality) when compared to EDGX. However, ARCA wins more of the horseraces than does the lower-fee BZX and it is difficult to pick a winner comparing BZX and NDAQ. Finally, in the only set of contests pairing an inverted and a traditional venue, BSX wins nearly all of the NDAQ contests. In the 3,256 cases in which BSX receives the first order, the BSX wins 99.77% contests that do not result in a tie. In Panel B of Appendix Table I we require both members of the order pair to display at least part of the orders size. There are six different horseraces presented in Panel B. In five of the horseraces, independent of who receives the first order, the venue with the lowest take fee wins. Only in the comparison of order pairs routed to the BZX and the NYSE, with the BZX receiving the first order, does the venue with the higher take fee prevail. However, in this instance the margin of victory is small BZX wins 50.21% of the horseraces.

55

Appendix Table I Limit Order Horseraces Panel A. All order pairs when venues have different take fees. Order Pair 1 Venue 2nd Venue
st

(take fee)

(take fee)

# of Order Pairs 336 407 51,474 27,683 258 461 2,205 1,163 386 1,312 30,637 1,663 4,530 9,029 118 3,256

% Both Fill 62.50 66.58 63.10 39.21 14.34 9.33 32.47 40.58 20.98 23.09 39.33 37.16 49.69 27.16 11.02 9.28

Difference in Time to 1st Fill 13.24 -12.40 0.86 -3.98 1.01 57.91 -3.78 -1.06 -105.14 -7.56 12.19 -20.67 -2.34 0.27 16.53 -3.18

% 1st Venue Trades First 31.85 44.96 21.46 55.11 10.08 91.32 37.19 44.71 68.13 20.96 27.13 50.03 34.22 60.44 0.85 92.44

% 2nd Venue Trades First 43.75 28.01 38.83 17.90 82.56 3.47 41.54 32.76 16.32 63.41 48.08 26.16 27.55 17.13 93.22 0.21

Good Fill Ratio % Tie 1st Venue 58.62 53.42 47.86 49.87 51.79 65.72 49.33 51.00 56.18 56.65 53.08 51.98 50.93 52.83 60.00 52.68 2nd Venue 61.67 50.96 49.81 45.72 58.78 55.36 53.40 50.26 53.38 56.33 53.31 51.05 50.82 50.40 59.32 41.22

Arca ($0.30) NYSE ($0.23) NDAQ ($0.30) NYSE ($0.23) EDGX ($0.30) NYSE ($0.23) BZX ($0.29) NYSE ($0.23) Arca ($0.30) BZX ($0.29) EDGX ($0.30) BZX ($0.29) NDAQ ($0.30) BZX ($0.29) NDAQ ($0.30) BSX ($0.14)

NYSE ($0.23) Arca ($0.30) NYSE ($0.23) NDAQ ($0.30) NYSE ($0.23) EDGX ($0.30) NYSE ($0.23) BZX ($0.29) BZX ($0.29) Arca ($0.30) BZX ($0.29) EDGX ($0.30) BZX ($0.29) NDAQ ($0.30) BSX (-$0.14) NDAQ ($0.30)

24.40 27.30 39.72 26.99 7.36 5.21 21.27 22.53 15.54 15.63 24.79 23.81 38.04 22.43 5.93 7.34

56

Appendix Table I (Continued) Panel B. Displayed order pairs when venues have different take fees. Order Pair 1 Venue 2nd Venue
st

(take fee)

(take fee)

# of Order Pairs 73 124 1,715 592 336 71 9 204 4,318 175 185

% Both Fill 61.64 65.32 56.27 62.16 45.24 53.52 44.44 33.33 44.91 44.57 69.19

Difference in Time to 1st Fill 0.47 -19.37 -8.31 -35.17 -7.48 -9.53 24.79 -52.09 0.05 0.11 61.93

% 1st Venue Trades First 26.03 47.58 20.47 43.92 38.10 45.07 22.22 68.63 18.32 46.86 12.43

% 2nd Venue Trades First 53.42 25.81 50.20 22.30 37.78 32.39 55.56 11.27 46.99 18.29 27.03

Good Fill Ratio % Tie 1st Venue 2nd Venue

Arca ($0.30) NYSE ($0.23) NDAQ ($0.30) NYSE ($0.23) BZX ($0.29) NYSE ($0.23) Arca ($0.30) BZX ($0.29) EDGX ($0.30) BZX ($0.29) NDAQ ($0.30)

NYSE ($0.23) Arca ($0.30) NYSE ($0.23) NDAQ ($0.30) NYSE ($0.23) BZX ($0.29) BZX ($0.29) Arca ($0.30) BZX ($0.29) EDGX ($0.30) BZX ($0.29)

20.55 26.61 29.33 33.78 24.11 22.54 22.22 20.10 34.69 34.86 60.54

NDAQ ($0.30) 379 42.48 -12.43 10.29 32.98 BZX ($0.29) 56.73 Notes: An Order Pair (e.g., a horserace) involves a pair of limit orders that have the same stock symbol, same order date, same order side (buy or sell), same limit price, same order time (to within one ms), but different destination venues. The 1st (2nd) venue receives the first (second) order of the order pair. Take fee is the cost of accessing 100 shares of liquidity on the given venue. % Both Fill refers to the percentage of order pairs in which both orders receive at least a partial execution. For the order pairs where both orders execute, Difference in Time to 1st Fill is the difference in times to execution for each of the two orders in the order pair. If the order resting at the first venue executes before (after) the order on the second venues order book fills, this difference is negative (positive). % 1st Venue Trades First (% 2nd Venue Trades First) is the percentage of horseraces won by the first (second) venue. We declare a horserace a tie if both orders execute within 500 s of one another. The Good Fill Ratio is the percentage of executed limit orders with positive realized spreads. Take fees are expressed in cents per 100 shares.
57