Institutional Order Handling and Broker-Affiliated Trading Venues

Abstract

Institutional investors account for a majority of ownership of U.S. stocks and serve as the primary vehicle for household investments. An important driver of institutional performance is the ability to implement investment ideas at a low cost. Trading costs subtract from institutional performance, erode, or eliminate the value added by portfolio managers and lower the returns to research (Wagner 1993). Anand et al. (2012) show that institutional trading costs are economically large and that broker selection is an important decision for managing trading costs.¹

Brokers make a number of decisions on behalf of their clients, including splitting an institutional order into smaller pieces, selecting among execution venues, and sequencing the submission of the split orders across venues. To monitor and evaluate the quality of executions reported by brokers, institutional clients engage in transaction cost analysis. However, in recent years, industry participants and regulators have noted that the complexity of the U.S. equity market structure makes it difficult for institutions to assess broker performance, especially since routing data are either unavailable or difficult to decipher.²

Further, opaque reporting practices on the handling of institutional orders can obscure potential agency conflicts that influence broker routing decisions. A report by the International Organization of Securities Commissions (IOSCO 2017) identifies three incentives that may influence broker behavior: monetary benefits received from third parties; bundling of other client services with executions; and affiliated venues that have benefits for brokers.

Battalio, Corwin, and Jennings (2016) study the impact of monetary benefits related to exchange pricing models on broker routing. They note that fees and rebates to brokers from venues are not generally passed on to clients. The study finds that routing of nonmarketable retail orders to venues with higher rebates hurts execution quality. Bundling ancillary services (“soft dollars”), such as research and initial public offering (IPO) allocations, alongside order execution, raises the possibility that clients are less likely to select brokers based on the quality of executions. Consistent with this idea, Conrad, Johnson, and Wahal (2001) show that institutional orders routed to brokers with soft dollar arrangements are associated with higher trading costs.

In this study, we examine a largely unstudied potential source of agency conflict, namely, broker routing of client orders to alternative trading systems (ATSs) that they own (“affiliated ATS”). We study two questions: Do brokers show a preference for routing orders to an affiliated ATS? Is there a systematic association between affiliated ATS routing preference and execution quality? We recognize that routing decisions may reflect differences in broker clientele or a client’s preference for ATS venues and present several approaches to isolate the impact of a broker’s routing decisions on execution quality.

Venue ownership can influence the broker’s routing decisions in several ways. A broker pays fees to place and execute orders on trading venues, but, typically, these fees are not passed on to clients. A broker avoids paying these fees on affiliated venues. The fee savings create a conflict between obtaining the best outcome for the client and maximizing broker revenues.³ Brokers obtain additional benefits from higher volume on affiliated venues. Other participants (including liquidity providers) typically pay fees to the ATS operator for executing orders on the ATS. Higher ATS volume is a mark of success that attracts order flow, and can be a selling point to clients, thus, increasing market share and commissions earned. Higher ATS volumes can boost the brokerage firm’s rankings in a stock’s trading, which can help the investment banking arm win corporate finance engagements with the issuer. For these reasons, a broker may prefer an affiliated ATS over other venues even when such a preference is not optimal for the client.⁴

On the other hand, brokers have a duty of best execution to clients, a duty that generally requires brokers to seek as favorable of executions as possible under prevailing market conditions. Routing to an affiliated ATS can improve client outcomes in several ways. A broker can efficiently source liquidity from an affiliated ATS if the ATS order book information is visible to the broker’s smart order router. Competing brokers may have limited access to a broker’s affiliated ATS, thus, the affiliated ATS offers an additional venue to the client. Other often-noted benefits of ATSs over exchanges include the potential for price improvement, and the ability to hide orders, limit price impact, and interact with select participants (see IOSCO 2017). Given these potential benefits, it is unclear whether, and how, routing to affiliated ATSs affects execution quality of client orders.

To the best of our knowledge, no existing empirical work examines the relation between broker routing, venue ownership, and execution outcomes. The lack of research is likely due to insufficient suitable data. In this study, we use the FINRA Order Audit Trail System (OATS) database that provides detailed information on orders received by brokers from their clients (henceforth “top orders”), including the identity of the broker; the venue-specific routing decisions; and the venue-specific outcomes, such as executions, traded prices, and time stamps associated with each routing, modification, execution, or cancellation decision in an order’s life cycle. The OATS database, created for regulatory purposes, covers a comprehensive cross-section of brokers and does not suffer from attrition or sample selection bias issues that often affect data received from a subset of industry participants.⁵ We study a sample of over 350 million top orders that are received by 43 active institutional brokers for a size-stratified sample of 300 stocks in October 2016.

We find that not all brokers that own an ATS route more orders to the affiliated venue. We calculate the average deviation between a broker’s routing to an affiliated ATS and the aggregate comparable statistic for all brokers on a stock-day. Based on this measure, we place the brokers into three groups. For brokers in the highest tercile (T3), routing to ATSs accounts for 63% of the routed quantity, largely attributable to affiliated ATS routes (50% of the routed quantity), while for brokers in middle (T2) and lowest (T1) terciles, routing to ATSs (affiliated ATS) accounts for 26% (6%) and 9% (0%) of the routed quantity. The affiliated ATS accounts for 17% of the executed quantity for T3 brokers. For perspective, in the fourth quarter of 2016, an individual ATS typically executed less than 1%, and the combined market share of all ATSs is less than 14%.⁶ Thus, T3 brokers route and execute more orders on affiliated ATSs.

We find that higher affiliated ATS routing is associated with lower execution quality. These results are observed for the “fill rate” measure, defined as the ratio of executed shares to the submitted quantity for a top order, and execution cost measures based on the implementation shortfall (IS) approach, which accounts for fill rates, market impact, and price drift (see Perold 1988; Wagner and Edwards 1993). The OATS database does not include client identifiers or information on a client’s trade motive or certain order instructions. Consequently, we report several shortfall measures that reflect a range of assumptions about the client’s opportunity costs of unfilled orders.⁷

After controlling for stock characteristics, order attributes and market conditions, a one-standard-deviation increase in affiliated ATS routing is associated with an 11.6-percentage-point decline in fill rates. In similar regressions, higher affiliated ATS routing is associated with higher implementation shortfalls. We find robust evidence across the different approaches used to impute an execution price for the unfilled portion of the top order (“cleanup” costs). We find that the higher shortfall costs stem from higher delay costs (price drift and cleanup costs).

The theoretical model in Ye and Zhu (2019) predicts that large traders prefer ATSs when they have more information. Thus, the lower execution quality for T3 brokers can stem from a selection bias, where clients with more difficult or informed orders cluster with T3 brokers. The regression specifications include reasonable controls for order difficulty; however, as the data do not allow an explicit control for client identity, we implement several approaches to mitigate this concern.

We examine the effect of broker discretion in handling orders based on the “not-held” designation in OATS data.⁸ In a difference-in-differences specification, not-held orders (i.e., those with greater broker discretion) are associated with worse execution outcomes, such as lower fill rates and higher shortfalls, than held orders for T3 brokers relative to non-T3 brokers. These results suggest that execution quality differences for T3 brokers can be attributed, at least in part, to the brokers’ routing decisions.

To address the possibility that more difficult orders are sent to ATSs we implement a matched broker analysis where the broker pairs have a similar level of ATS routing (about 50% of routed volume) on the same stock-day. In the matched pairs, T3 brokers predominantly route (42% of routed volume) to their affiliated ATSs while non-T3 brokers predominantly route (43% of routed volume) to unaffiliated ATSs. The analysis controls for a selection bias where clients prefer ATS venues for difficult orders and helps isolate the impact of affiliated ATS routing on execution quality. In the matched regressions, T3 brokers have fill rates that are 5.4 percentage points lower and shortfall costs that are 2.2 basis points (bps) higher than their matched non-T3 brokers.

T3 brokers’ routes to affiliated ATSs may possibly occur in more difficult conditions than other ATS routes. To the extent that order difficulty is associated with market conditions (e.g., one-sided imbalances), do we observe higher ATS usage by T3 brokers when conditions are difficult? We find that T3-affiliated ATS executions are more likely in easier conditions than non-T3 ATS executions. The empirical determinants of ATS routing are markedly different than those of ATS executions: non-T3 brokers are more likely to route to ATSs under difficult conditions, while T3-affiliated ATS routes are largely insensitive to market conditions. These results do not suggest that T3 brokers route more orders to affiliated ATSs under relatively difficult conditions.

Zhu (2014) and Reed, Samadi, and Sokobin (2020) uncover significant variety in features across ATSs. We examine whether affiliated ATSs of T3 brokers exhibit design features that are less likely to be associated with information leakage, which is a concern for informed investors. For most ATSs, we rely on Reg ATS-N filings available on the SEC’s website.⁹ We create an ATS Opacity measure based on one/zero scores for characteristics that make an ATS less opaque (e.g., the presence of affiliated principal traders or external proprietary traders) and those that make an ATS more opaque (e.g., the ability to exclude counterparties). We do not observe significant differences in ATS Opacity measure between T3-affiliated ATSs and other ATSs. Thus, it does not appear that, all else equal, T3-affiliated ATSs offer design features that are more attractive in terms of opacity to informed investors.

We exploit the granularity of the data in order to examine route-level outcomes for commonly used order types across ATSs, which more precisely control for order type, market conditions and client intent. Specifically, we adapt the paired route “horse race” approach developed by Sofianos, Xiang, and Yousefi (2010) and Battalio, Corwin, and Jennings (2016) to our empirical setting. We pair individual routes made by T3 brokers to affiliated ATSs with other identically priced, concurrent routes to unaffiliated ATSs. The pairs are matched separately for two common order types: midpoint peg orders and primary peg orders. Midpoint pegs are limit orders that dynamically update their price to the NBBO midpoint, while primary pegs are tied to the passive side of the NBBO. The paired child orders are open at the same time, which controls for market conditions, and have the same order instructions, which accounts for client intent and the type of ATS venue. For both midpoint and primary peg orders, we find strong evidence that unaffiliated ATS routes have higher fill rates and obtain an earlier fill than matched T3-affiliated ATS routes. The horse race results point to a potential mechanism to understand the execution quality differences observed in the broader sample.

To further address the possibility that our results reflect clients placing more difficult orders with T3 brokers as opposed to agency conflicts, we study broker routing and execution quality surrounding a regulatory experiment likely to affect venue choice: the SEC’s implementation of the Tick Size Pilot (TSP) program in October 2016, which significantly constrained executions at ATSs that match orders at the quotes for a randomly assigned subset (treatment group 3) of stocks. We test the following competing hypotheses: if T3-affiliated ATS routing is associated with optimal routing, then, relative to non-T3 brokers, T3 brokers’ execution outcomes will worsen when venue choice is constrained. On the other hand, the TSP may help resolve agency conflicts as some ATSs become less viable destinations, implying that T3 brokers’ execution outcomes will improve relative to non-T3 brokers. Based on a triple difference analysis that compares T3 versus matched (based on overall ATS routing in the pre-TSP period) non-T3 brokers, in group 3 (G3) stocks versus control stocks, before and after the implementation of the TSP program, we find support for the agency conflict hypothesis in that the outcomes for T3 brokers for G3 stocks improve relative to non-T3 brokers after the TSP is implemented. Since we retain our brokers and their tercile classifications from the main analysis, the TSP analysis also provides a useful out-of-sample test of the main findings of the study.

The collective evidence from various approaches suggests that the execution quality differences for T3 brokers are consistent with agency conflicts that lead brokers to route more orders to their affiliated ATSs. Structural changes in equity markets have increased the challenges faced by institutions in measuring broker performance. The proliferation of trading venues has fragmented markets and decreased the transparency of brokers’ routing decisions.¹⁰ Our results indicate that institutions will benefit from recent regulatory and industry initiatives aimed at increasing transparency of order handling.

We caution that ownership of a venue alone does not constitute evidence of routing conflicts; indeed, not all brokers that own an ATS show a preference for their venues, and not all orders routed to an affiliated ATS exhibit worse execution quality. We recognize that institutions receive benefits from brokers that extend beyond execution services, and that commissions could be systematically lower for T3 brokers. In those cases, our study provides information on execution quality outcomes that institutions can compare with the value of bundled services or commission savings received from brokers.

1. Venue Selection, Broker Handling and Agency Conflicts

A typical institutional trading decision starts with a portfolio manager. These decisions flow to the institutional trading desk, where they are consolidated across portfolio managers. The trading desk decides how to slice the order and allocate the slices to brokers, increasingly through a broker algorithm. These allocated orders are defined in our analysis as a client-broker top orders. The broker (algorithm) typically splits the top order into smaller pieces and generates a mix of child orders over time. The child order generation depends on several decisions related to the child order size, timing, aggressiveness, the venues to access and whether venues are accessed sequentially or in parallel. The combination of these decisions makes order handling difficult to monitor and evaluate, especially since trading is fragmented across 13 exchanges and more than 40 ATSs during our sample period.¹¹ The institution typically pays a fixed commission to the broker for handling the order. According to Tabb group estimates, brokerage commissions for algorithmic executions have steadily declined from an average of |${\$}$|0.019 per share in 2005 to |${\$}$|0.0060 in 2017.¹²

One dimension of the routing decision is the choice between seeking executions on (relatively opaque) ATSs and (relatively transparent) exchanges. ATSs do not publicly display orders, typically use prices derived from displayed exchange quotes, and report trades through trade reporting facilities, which combine all nonexchange trading (ATS and wholesaler executions), thus obscuring venue identity.

The theoretical literature has modeled the strategic venue selection by informed traders between ATS venues and “lit” exchanges.¹³ Based on the Kyle (1985) framework, Ye (2011) and Ye and Zhu (2019) model the monopolist informed traders’ choice. In Ye and Zhu (2019), traders choose between the higher cost of paying for immediacy on an exchange (where a market maker provides liquidity) against the higher risk of nonexecution on an ATS. The model predicts that the informed traders prefer an ATS because trading on the exchange moves the price, which hurts the execution quality on the ATS, particularly when the trader wants to acquire a large position.¹⁴ In a setting with competition among informed traders, the model in Zhu (2014) highlights that informed traders with correlated information tend to cluster on the same side of the ATS and suffer from lower execution probability. The model predicts that informed traders avoid ATSs when they have information due to costly delays in execution. Given that our study examines institutional brokers, the Ye and Zhu (2019) framework has particular relevance, since institutions are sensitive to the price impact of their large orders.

Depending on model assumptions, theory offers opposing predictions on whether institutions prefer ATSs when they have more information. Regardless of the direction, since less informed trades have lower execution costs, the relation between information and ATS routing potentially introduces a selection bias, since a broker’s client mix may be associated with higher ATS routing. Our data constraints prohibit us from identifying the institutional trader; however, we present several empirical approaches to account for selection bias and isolate the impact of a broker’s ATS routing preference on execution quality.

Other theoretical papers model strategic venue selection in a setting without asymmetric information. Degryse, Van Achter, and Wuyts (2009) predict that traders prefer an ATS to a dealer market when the bid-ask spread is large, as ATSs provide a greater potential for price improvement. Buti, Rindi, and Werner (2017) offer the opposite prediction that traders prefer an ATS when limit order book is well populated, as the limit order queue at the best bid and offer is longer. Buti, Rindi, and Werner (2016) and Ready (2014) report empirical support for the latter model. Menkveld, Yueshen and Zhu (2017) offer a pecking order hypothesis where an investor prefers a lit exchange to ATSs when the trading needs become more urgent and report empirical support for the model. Based on the literature, it is likely that investors favor trading on ATSs over exchanges under certain conditions. For this reason, our empirical analysis includes several approaches to account for order difficulty and market conditions.

Finally, venue selection can be influenced by private benefits that accrue to brokers. In the presence of agency conflicts, a broker’s propensity to route orders to affiliated ATSs could lead to worse outcomes for clients. In this context, this study is related to the literature on brokers’ routing of order flow pursuant to monetary benefits from third parties.¹⁵

For retail orders, Battalio, Corwin, and Jennings (2016) show that order routing designed to maximize a broker’s liquidity rebates on exchanges does not maximize limit order execution quality. A number of recent studies examine the interactions between institutional orders, high frequency traders (HFT) and execution quality.¹⁶ This literature links the market impact of institutional orders to HFT “back-runners” who detect order flow “footprints” and trade ahead or alongside institutional investors (for theory, see Yang and Zhu 2020; for evidence, see Kirilenko et al. 2017; van Kervel and Menkveld 2019; Saglam 2018; and Korajczyk and Murphy 2019). Battalio, Hatch, and Saglam (2018) attribute higher institutional trading costs to information leakage that stems from early routes to high frequency liquidity providers.

2. Data and Sample Description

2.1 Data sources and sample

The primary data set used in the study is the FINRA OATS database for the month of October 2016. Almost every broker-dealer in the United States is required to report audit trail information on equity orders to FINRA.¹⁷ For each broker-level parent order (“top order”) received from a client, OATS provides information detailing how the broker handled the top order. The data set combines the identity of the broker handling the order, the beneficiary owner type, and the submitted quantity of the broker-level order, with the audit trail of routes, venues, executions, modifications, and cancellations associated with the order’s life cycle.¹⁸

The OATS data are distinct from transaction-level data, such as the consolidated tape, or Trade and Quote (TAQ) data in providing a complete audit trail of broker-level top orders. Our data are also distinct from institutional ticket academic data made available by Abel-Noser Solutions that were used by Puckett and Yan (2011) and Anand et al. (2012, 2013). A limitation of the OATS data is that, while we observe orders handled by a broker for an institution (i.e., top order), it is not possible to further stitch together orders split by an institution across multiple brokers or submitted to the same broker at a later time. A significant advantage of the OATS data is the detailed information on a comprehensive cross-section of brokers’ handling of top orders and the venue-level routes and execution outcomes, which are not available in other databases.¹⁹ Further, comprehensive administrative data help overcome concerns related to selection bias that often affect studies that examine data from a subset of market participants.

The appendix describes our selection of 43 large institutional brokers. We use the beneficiary owner classification field from OATS, along with institutional broker classification of Griffin et al. (2011) to identify institutional brokers. The beneficiary owner field indicates whether an order represents institutional, individual, market maker, proprietary interest, or is unknown. Griffin et al. (2011), classify institutional brokers based on “company web pages, news media, the NASD website, and conversations with NASDAQ officials.” We exclude brokers that are primarily associated with internalized flow or serve as conduits, sending 100% of received order flow to a single ATS. We focus on “active” institutional brokers that handle at least 10,000 top orders in October 2016. Using the FINRA broker reference file, which specifies the firm associated with each broker ID, we identify the brokers affiliated with firms that also own an ATS in the FINRA ATS transparency data.²⁰

For a part of the analysis, we use the “not-held” order handling code in OATS data to identify top orders from institutions where the broker has price and time discretion (within client-specified constraints) in handling the order. The not-held order code is well established and clearly understood within the industry. In response to SEC’s proposed order disclosure rule, several industry participants recommend the use of not-held designation to differentiate institutional from retail flow and the need for enhanced disclosure for these orders.²¹SEC (2018) incorporates this recommendation and focuses on the not-held versus held distinction for the revised order handling disclosures for brokers. On the other hand, we do not differentiate between “directed” and “nondirected” orders. Directed orders are those where the client specifies the venue, thus limiting the broker’s discretion on venue selection. At first glance, the nondirected orders seem to be obvious choice for our analysis. However, during our sample period, the Rule 606 definition of orders makes such an analysis difficult.²² Specifically, a broker can flag an order as directed if a client does not change the default venue choice in the broker’s algorithm, which makes it difficult to identify venue choices that are truly determined by the client.²³ With those caveats, we confirm that the main results are observed for the sample of nondirected orders in Internet Appendix Table A2.

Our sample consists of a size-stratified group of 300 stocks traded in October 2016, a recent month when the project was initiated.²⁴ In light of the size of OATS database, we form decile portfolios using CRSP data on market capitalization at the end of December 2015 and select the 30 largest stocks from each decile. We merge the initial sample with the OATS database and TAQ National Best Bid and Offer (NBBO) quotes. We apply a number of data filters to obtain a final sample as detailed in the appendix. We classify stocks from the bottom three CRSP deciles as “small,” the middle four deciles as “medium,” and the top three deciles as “large” stocks. The final sample consists of over 350 million life cycles of the top orders received by the 43 institutional brokers in the sample stocks.

The Tick Size Pilot (TSP) program was implemented by the SEC in October 2016. We obtain qualitatively similar results when we restrict the sample to stocks unaffected by the TSP. To assess the impact of the TSP program, we separately examine data from September and November of 2016 using group 3 stocks (as defined in the TSP) as the treated sample and group 2 stocks and the TSP control stocks as separate control samples. After applying the filters detailed in the appendix, the TSP sample includes 898 million total life cycles associated with the 43 brokers in 247 G3 stocks, 257 G2 stocks and 811 control stocks.

2.2 Measures of execution quality

2.3 Descriptive statistics

Table 1 describes the sample. For each broker-stock-day, we calculate the weighted average of the order characteristics and market quality measures across top orders, where the weights are the quantity of the top order. Table 1 presents equally weighted averages of the 151,716 broker-stock-day observations.

The size of the top order for our sample averages |${\$}$|59,807 and represents 0.48% of the average daily volume (ADV) in the stock during the month prior to our sample period (September 2016).²⁵ For mutual funds and pension funds that report to Abel Noser database, Hu et al. (2018) reports that an institutional ticket distributed across brokers averages |${\$}$|96,495 in 2011, and that ticket size has been declining over the 2000 – 2011 period. The average percentage quoted bid-ask (half) spread at the time of order arrival is 19 bps, and ranges from 4 bps for large stocks to 52 bps for small stocks. Given this cross-sectional variation in liquidity, we present univariate results separately for small, medium and large stocks, and include stock fixed effects in the regressions. The average fill rate across broker-stock-days is 30.4%. Fill rates range from 22.6% for small stocks to 37.5% for large stocks. Effective spread costs are smaller than quoted spreads indicating that, on average, institutional orders do not aggressively seek liquidity. Shortfall is 7.1 bps. In the cross-section, small stocks are more expensive to trade than large stocks with Shortfall of 16 bps for small stocks and 3 bps for large stocks.

Table 2 describes the average routing of our sample of brokers to three broad categories of venues: ATSs, exchanges, and firms. “Firms” refers to brokers that act as execution venues; for example, large wholesalers that purchase order flow fall in this category. Across all stocks, on the average broker-stock-day, exchanges receive about 59% (55%) of the routes (routed quantity) and account for 78% (74%) of executions (executed quantity). ATSs receive 30% (33%) of the routes (routed quantity) but account for only 13% (17%) of executions (executed quantity). This result is consistent with the Tuttle (2012) finding that ATSs have lower fill rates than exchanges. We observe similar patterns across all stock categories.

3. Results

3.1 Affiliated ATS routing

Table 3 examines whether brokers route more orders to affiliated ATSs by comparing a broker to the sample benchmark. Specifically, for each broker-stock-day, we calculate the %Affiliated ATS, which is the proportion of routed quantity sent by the individual broker to an affiliated ATS. The benchmark for each stock-day is the proportion of total routed quantity sent by all brokers as a group to their respective affiliated ATSs. Next, we calculate the average deviation from the stock-day benchmark for a broker across all stock-days. We assign the broker’s average deviation to each of their stock-day observations and then form terciles of broker-stock-days using the average deviation measure.²⁶ Importantly, each broker is assigned to a unique tercile, and by design, tercile 1 (T1) brokers route less, while tercile 3 (T3) brokers route more to affiliated ATSs venues, relative to the benchmark.

Table 3 describes venue choice for broker-stock-day observations in each tercile. Results in Table 3 indicate large differences in routing behavior across broker terciles. T1 and T2 brokers route 9% and 26% of shares to ATSs, with no affiliated ATS routing for T1 brokers and 6% for T2 brokers. T1 brokers route 73% of shares to exchanges, while T2 brokers route 60% to exchanges. Approximately 79% (73%) of the executed quantity for T1 (T2) brokers occurs on exchanges.

T3 brokers differ markedly from the other groups. T3 brokers route approximately 50% of shares to affiliated ATSs,²⁷ and 63% of shares to ATSs (affiliated and unaffiliated); however, only 17% of the executed quantity occurs on affiliated ATSs. Exchanges account for only 30% of T3 brokers’ routed quantity but account for over 70% of executed quantity. A notable statistic is the 17% average execution share of affiliated ATSs for T3 brokers. During our sample period, the combined market share of all ATS venues is 14% and the market share of the largest ATS is only around 2%. Thus, T3 brokers route and execute more orders on affiliated ATSs and exhibit a large discrepancy between routing quantity and execution quantity on affiliated ATSs.

In panel B, we study whether the routing behavior of brokers to affiliated ATS venues is persistent. Given that our sample period is a single month, we form broker terciles based on average deviation in the first week of October 2016 and report retention statistics for future weeks, where retention rate is the percentage of brokers classified in the same tercile in future weeks. Results suggest that broker ranks based on affiliated ATS routing are highly persistent. In all future weeks, the affiliated ATS routing statistic increases from tercile 1 to tercile 3, T3 routing to affiliated ATSs remains at approximately 50%, and the T3 retention percentage exceeds 80%.

3.2 Execution quality: Univariate statistics

Table 4, panel A, presents average execution outcomes for broker terciles formed on affiliated ATS routing. A notable result is that average fill rate of top orders handled by T3 brokers is significantly smaller than those handled by T1 or T2 brokers. Specifically, the fill rate across broker-stock-days is 46% for T1 brokers, 28% for T2 brokers but only 17% for T3 brokers. The differences are not simply an artifact of brokers receiving orders in different stocks, as similar patterns exist for small, medium, and large stocks.

In the overall sample, the effective spread costs of T1 brokers is 1.1 bps, the corresponding statistics for T2 brokers and T3 brokers are 3.3 bps and 3.32 bps. T3 brokers have higher costs than do T1 brokers for medium and large stocks, while the differences are not statistically significant for small stocks. Shortfall is higher for T3 brokers than T1 and T2 brokers in the overall sample and in each stock category. In terms of magnitude, the average Shortfall for T3 brokers is 10.7 bps, which is significantly larger than the 4.2 bps for T1 brokers and 6.6 bps for T2 brokers. We find that the higher Shortfall of T3 brokers is attributable to higher Delay.

One potential explanation for higher trading costs is that clients send orders to T3 brokers when conditions are difficult. To understand differences in market conditions, we examine the percentage quoted spreads that prevail at the time of order arrival. The patterns observed in panels B to D present mixed evidence, reflecting more difficult conditions for some stock groups and easier conditions for other stock groups. These patterns highlight the need to assess execution quality differences after controlling for stock characteristics, order difficulty, and market conditions.

3.3 Regression analysis of execution quality

Models 1 to 3 in Table 5 report specifications with the fill rate as the dependent variable. The negative coefficients for %Affiliated ATS indicate that brokers with higher affiliated ATS routing are associated with lower fill rates. In all specifications, %Affiliated ATS is highly significant at the 1% level. The stock fixed effects estimate (model 3) suggests that a one-standard-deviation increase in %Affiliated ATS is associated with an 11.6-percentage-point decline in fill rates. Thus, the T3 brokers’ decisions to route more orders to affiliated ATSs is economically significant. The coefficients for the control variables are of the expected sign. For example, fill rates are positively associated with market capitalization and negatively associated with arrival spreads.

Effective spread costs do not show a significant association with brokers’ affiliated ATS routing, indicating that a propensity to route to affiliated venues does not improve the execution prices received on the filled portion of a top order. On the other hand, brokers with high affiliated ATS routing are associated with larger shortfall. Based on model 3, a one-standard-deviation increase in %Affiliated ATS is associated with 1.8 bps larger shortfall. Delay shows similar patterns to shortfall, underscoring that the higher shortfalls stem from price drift and cleanup costs and not the prices received on executions.

3.4 Separating selection bias from broker conflicts

Ye and Zhu (2019) predict that large traders prefer ATSs when they have more information due to the higher opacity of such venues. Thus, T3 brokers’ higher routing to ATSs may simply reflect a more informed clientele. Since informed orders are associated with higher trading costs, clients’ information is potentially a source of selection bias, which could result in the lower execution quality associated with T3 brokers. When comparing model 1 to model 3 in Table 5, the shortfall coefficient for %Affiliated ATS declines from 6.91 to 5.84 with additional controls for stock and order attributes, which could potentially reflect such a bias. The OATS data do not include client identity, precluding the use of direct client-level controls. We present several approaches that account for the selection bias and attempt to isolate the impact of the brokers’ routing decisions on execution quality.

3.4.1 Broker discretion

Table 6 summarizes the results. We find that fill rates are lower and execution costs are higher for not-held orders relative to held orders for T3 brokers versus the corresponding difference in execution outcomes for non-T3 brokers. Specifically, the DID estimate indicates that shortfall is 3.2 bps higher for not-held orders than held orders of T3 brokers relative to non-T3 brokers. We find no significant difference in shortfall between not-held and held orders of non-T3 brokers. Overall, the results point to lower execution quality when T3 brokers have more discretion but we note that this analysis could potentially underestimate the true effect of broker discretion. This is because clients typically seek guidance from brokers on execution strategies, which is likely to affect the execution instructions associated with held orders. We also recognize that the factors affecting the not-held and held orders usage may differ between the clients of T3- and non-T3 brokers. Below, we will provide additional tests aimed at separating the selection bias explanation from agency conflicts affecting execution quality.

3.4.2 Matched brokers based on ATS routing

To control for the possibility that order difficulty is related to ATS routing, we implement a matched broker analysis where the broker-pairs exhibit a similar propensity to route orders to ATSs on the same stock-day. Specifically, using nearest-neighbor one-to-one propensity scores and a caliper of one quarter of a standard deviation, we obtain 42,208 matched broker-stock-days where average proportion of ATS routed quantity for T3 brokers is 51% and for non-T3 (control) brokers is 50%. Notably, while broker-pairs have a similar proportion of ATS routing on a stock-day, T3 brokers in matched-pairs predominantly route (42% of routed volume) to affiliated ATSs while non-T3 brokers in matched-pairs predominantly route (43% of routed volume) to unaffiliated ATSs. The matched analysis helps control for the client’s preference for ATSs and isolates the impact of affiliated ATS routing on trading outcomes.

Table 7 presents the results. Fill rates for T3 brokers are significantly lower (5.4 percentage points) than for matched brokers. The |$T3$| coefficient in Effective spread costs regressions is negative and marginally significant, suggesting that T3 brokers obtain lower realized costs on the filled portion of an order. The Shortfall coefficient indicates higher costs (2.2 bps) for T3 brokers than matched non-T3 brokers. As before, higher shortfall costs are driven by higher delay. Our results indicate that brokers who route more orders to affiliated venues offer lower execution quality than other brokers who route similar quantity of orders to unaffiliated ATSs.