Accessing Single Dealer Platforms (SDPS) in Execution Algorithms: Penny-Wise and Pound-Foolish?

What is a Single Dealer Platform (SDP)?

Execution algorithm designers have three options for executing market or marketable limit order slices of large orders being traded; they can access exchanges, ATSs or Single Dealer Platforms (SDPs). An SDP is a market maker that accepts small Immediate-or-Cancel (IOC) market or marketable limit orders and can choose whether to fill the orders at the time of receipt (as opposed to wholesalers who guarantee execution when they receive such orders from retail brokers).

SDPs typically provide their services to algorithmic execution brokers, aiming to receive small slices of algorithmic order flow from Smart Order Routers (SORs); they are not block desks that provide liquidity for an entire block order all at once. Basically, SDPs are market makers that compete with market makers on exchanges by offering liquidity at the same execution price. Exchanges charge a fee for taking liquidity of up to $.0030 per share whereas SDPs provide liquidity for “free” (and may even offer a rebate), hence providing an opportunity to reduce costs for algorithmic execution brokers.

When the earliest SDPs appeared around 2005, they were known as Electronic Liquidity Providers (ELPs). Currently, Citadel, Virtu, and Jane Street are some of the largest SDPs. Although SDPs have attempted to brand themselves as dark pools, there are critical differences between SDPs and traditional dark pools registered as Alternative Trading Systems (ATSs) with the SEC. While this article focuses on US equity markets, similar venues exist in European equity markets known as “systematic internalizers” (SIs) and are operated by large banks that also offer execution algorithms.

The most critical difference between SDPs and ATSs or exchanges is the counterparty an algorithmic broker faces when accessing these venues. Orders sent to an SDP trade only against the market maker operating the SDP, whereas orders sent to a dark pool can trade against a variety of counterparties accessing that ATS or exchange.

A second major difference between SDPs and ATSs lies in the amount of information leakage associated with orders sent to these venues. There is more information leakage for an algorithm sending a market order/marketable limit order to an SDP than in other ATSs or exchanges. When sending a marketable order to an exchange or ATS, the order is anonymous to the counterparty trading against it. Since a liquidity provider at an exchange or ATS does not have information about the identity of its counterparty, they have no ability to reconstruct the counterparty’s trading patterns. Contrarily, an SDP trading against an arriving market order has complete information about the identity of the algorithmic broker in real time and information about the “segment” of flow an order belongs to. On one hand, knowing that the broker’s flow is institutional or belongs to a certain category (segment) of institutional order flow indicated by the broker to the SDP allows the SDP to tailor its liquidity based on flow toxicity (e.g., providing better fill rate for non- toxic flow). But on the other hand, it creates information leakage concerns for the party sending the orders.

Also, at an exchange or ATS, the counterparty finds out about the order only after the trade occurs, rather than prior to the trade. If a match occurs, the counterparty receives a fill for the matching quantity and has no opportunity to react to the order before the trade occurs. Nor do they receive any information about the residual quantity of the market order left unfilled after the order. Contrarily, an SDP market maker knows the order size before they agree to a trade and can deny a fill. SDPs have the opportunity for a “last look”, allowing them to check current quotes at exchanges before providing a fill, giving them a significant advantage over market makers providing liquidity at exchanges. And if they decide not to trade against the order or to fill only part of the order, they have information about the residuals that are likely to appear on exchanges soon after. More importantly, since they also know who is sending the order (and the segment that this order flow belongs to), they can establish order patterns from corresponding brokers in real time and historically. Although frequently advertised as dark pools, there is little “darkness” in an SDP.

One last—but critical—difference between registered ATSs and SDPs is that they are not subject to the same regulations. ATSs must be registered with the SEC in accordance with RegATS and must provide complete transparency around their operation procedures and the ATS-related activities of the broker-dealer operator and its affiliates (via Form ATS-N). SDPs enjoy what is known as the “single market maker exception” and are not required to register with the SEC as an exchange or ATS and thus their users are left without the transparency ATS-N filings provide. We did find a seemingly voluntary disclosure on one SDP’s website. We include some of the disclosure below, and it underscores the amount of flexibility an SDP has in using the information obtained from counterparties.

“[SDP Name] generally may use information about current and historical client orders, cancellations, and fills (both full and partial), internal positions, real time market data, and commercially available data, information that clients provide about their customers (if applicable), information about clients’ and, if applicable, their customers’ trading activities and patterns, and any other information clients transmit or otherwise provide to [SDP Name], in determining whether to fill a client order (in whole or in part), the price at which to fill an order (including whether to provide price improvement), and, more generally, in connection with any other trading or business purpose of [SDP Name], including its market making and nonmarket making activities.”

Concern #1: Information Leakage at the Parent Order Level

When clients send orders to algorithms provided by execution brokers, the full size of an order—referred to as the “parent¹ order”—is broken into tiny slices to be traded over a period of time and sent to a variety of execution venues by the execution broker’s algorithm. By far the largest concern when trading with an SDP is the information leakage of the parent order that comes from trading just one small slice—not the information about the slice itself. While every trade is eventually printed to the consolidated tape (SIP) regardless of its execution venue, that information is not enough to reconstruct the parent order behind a single trade, because trading on exchanges and in ATSs is anonymous. Matching counterparties do not know who sent the market order and whether it was a smaller piece of a larger parent order. The fact that a trade went off at the offer, and hence the instigating order was a buy market order is not sufficient to indicate whether there is a large buy order behind it.

However, in the case of an execution sent to an SDP, the SDP knows the following aspects of the order that are not available to a liquidity provider at an exchange or ATS:

• The order originated from an institutional broker-dealer’s execution algorithm and hence is likely a small slice of a larger parent order, with more slices likely coming soon
• The name of the institutional broker-dealer sending the order; and SDPs encourage brokers to further segment flow into different categories, with narrower segments revealing more information about the order flow
• Ability to detect patterns about the algorithm, such as when it started slicing a parent order and when it ended
• The size and side of the child order that will be sent to exchange(s) if the SDP did not fill it or did not fill it completely

Given this information, even though an SDP gets information about only the child orders at any given point, they can easily detect parent order information. For example, if an algorithmic broker sends the first couple of slices for a stock to an SDP at 10:00 AM, the SDP can detect that it is the beginning of a parent order, and the likelihood that the broker’s client will also be buying the same stock at 10:01 or 10:02 or 10:05 is extremely high. Similarly, after receiving many slices, when the SDP stops receiving additional slices of an order from an algorithmic broker, they can deduce that the parent order has likely completed. SDPs can also statistically compare new information to historical patterns in broker (or segment) flow, optimizing their market making or non-market making algorithms. Often the three most important pieces of information needed to predict price changes due to order flow are 1) when an institutional block order began, 2) when the order ended, and 3) the pace of execution. Though SDPs may not have complete information, they may have enough to be able to use the information they do have to their advantage.

A common pattern in institutional algorithmic trading is the concave shape of the market impact curve. As an algorithm starts trading an order, the first slices of the order create the largest market impact and subsequent slices create less and less additional market impact. The temporary impact from the earlier slices begins to revert while later slices are entering the market and creating their own impact. When the order is fully executed there are no new slices creating impact, but temporary impact from previous slices is still reverting, creating the well-known “reversion” effect following order completion.

This effect is illustrated in Emmanuel Bacry², Adrian Iuga, Matthieu Lasnier, and Charles-Albert Lehalle’s Market Impacts and the Life Cycle of Investors Orders, which studies 400,000 institutional algorithm orders. We have included their illustration of the life cycle of market impact in the figure below, where the horizontal axis represents the elapsed time in a parent order (and 1 represents the completion of the entire parent order). The image shows the cumulative market impact in price movement and its partial reversion after the entire parent order is finished trading.

Liquidity providers on exchanges and ATSs do not have access to information that could help link the tiny slices of larger parent orders and indicate whether additional price movement is likely. Their ability to establish a pattern for the parent order is close to impossible because your child orders are anonymously interleaved with other participants’ child orders. It is the ability to identify the sender of the orders that allows a market maker to estimate where in the life cycle of market impact they are and what change in prices can be expected next as a result.

It is an open question whether SDPs’ informational advantage has any adverse impact on the implementation shortfall of the parent orders a broker algorithm is working. SDPs often claim that they hardly ever remove liquidity from exchanges and thus the impact they create is likely to be low. But a market participant does not need to remove liquidity from exchanges in order to have an adverse effect on algorithm performance. In anticipation that a large institutional buy order has just started executing, for example, an SDP can predict a price increase and become a better buyer of the stock than seller, increasing the prices at which they are willing to buy or sell. In doing so, they may create additional impact, increasing the cost of trading the large parent order providing the information. It is a well-established fact that price impact is created through not just removing liquidity but also through order book events; improving bid prices, increasing bid sizes, and reducing offer sizes, for example, can lead to price increases just as sending a buy market order can³. The fact that SDP operators are also some of the largest market makers on exchanges doesn’t help.

Parent order analysis requires a huge sample

The most effective way to analyze the impact of SDPs on the implementation shortfall of parent orders is to run a controlled experiment where a large sample of parent orders from a fixed set of clients are randomly distributed to two equal algorithms—the only difference being that one routes to SDPs and the other does not. By statistically evaluating the difference in implementation shortfall of the two algorithms, the [positive or negative] impact of trading with SDPs could be uncovered.

However, the issue with controlled experimentation is the high signal-to-noise ratio associated with parent order performance metrics. While implementation shortfall costs for US equities commonly range between 5 and 30 basis points, the standard deviation of these costs is on the order of 200 basis points (taking, for example, the median volatility of 45% for a US stock trading over the entire day). Hence, a very large number of observations is required to see a statistically significant effect of trading with SDPs on overall implementation shortfall costs—whether there is, in fact, a difference or simply enough evidence to declare that no difference is present. For example, if trading with SDPs adds one basis point of additional implementation shortfall, we estimate it would take over 50,000 full-day orders to see the effect become statistically significant. For another example, for baskets that are dollar-neutral with daily volatility of approximately 0.5%, it would take roughly 800 days of data to see a statistically significant effect. Of course, if the impact on performance were higher or the duration of orders were shorter, it would take less time to gather enough observations; but the point remains that the amount of data needed is quite large.

Why hasn’t anyone studied this yet?

Noise at the parent order level makes it very difficult for institutional asset managers to evaluate how effectively their brokers navigate existing market structure. Large investment banks that trade hundreds of millions of shares a day through their algorithms have more than enough data to run controlled experiments and demonstrate the impact of using SDPs (whether positive or negative) within execution algorithms. But SDPs offer execution more affordably than exchanges (and even rebates), which reduces motivation to explore their value to clients as brokers and banks work to reduce their own costs. In addition, large banks run their own SDPs in the US, and if not, they may not want to preclude the possibility of running an SDP in the future with a study revealing [potentially] negative effects. Also, while in the US, the largest SDPs are not operated by banks, their European counterparts run their own “systematic internalizers” (SIs) that have similar implications for information leakage. Such conflicts likely keep banks from running these A/B experiments.

Concern #2: Information Leakage Related to Child Orders

In addition to information leakage as it relates to parent orders—which we believe is a bigger concern—there is also the issue of information leakage for child orders. Because of the opportunity for a “last look” described above, where the SDP’s market maker can check the quotes on exchanges one more time before choosing to execute an order, there is also an opportunity for other actions, like canceling their limit orders to sell on exchanges upon receiving a child order to buy.

It is much easier and requires far less data to measure the harm [or benefit] of trading with SDPs at the child order level than the parent order level, as long as the right metric is being used. For example, for child orders, if no execution is received when an order is sent to an SDP, an algorithm would simply route to an exchange for execution afterwards. If by that time the exchange quote has faded—whether by chance or because the SDP that received the original order has faded their quote—the algorithm would pay a higher price to buy that stock. This effect can be measured directly by comparing the previous market price with the new market price in the instances when the algorithm did not receive a fill (or received a fill smaller than the size available at exchanges) when routing to an SDP. This measurement, however, is in no way a replacement for the measurement of parent order information leakage.

Misconceptions & Marketing

In addition to sharing our concerns related to trading with SDPs in our algorithms, we also wanted to clear some of the common misunderstandings we hear in discussion.

MYTH: IOIs eliminate information leakage

Most SDPs offer two types of interaction: 1) client algorithms can send “blind” Immediate-or-Cancel orders without knowing whether the SDP will provide a fill, or 2) client algorithms can consume and respond to SDP Indications of Interest (IOIs) (though a fill is still not guaranteed in this case, but the likelihood of no-fill is reduced significantly).

In our conversations with SDPs, we often hear the pitch that IOIs “essentially eliminate” information leakage. While IOIs will certainly mitigate the adverse effect on the execution of child order slices, the far larger concern of the client should be information leakage regarding the parent order from observing the patterns of incoming child orders received from a broker as described above.

MYTH: Markout analysis shows there is no harm in trading with SDPs

Another misconception related to trading with SDPs is that evaluating fills via markout analysis is a reasonable methodology for determining that these fills are non-toxic. Markouts are useful for measuring the adverse selection costs of passive fills. Markouts compare the fill price to the midpoint price at some future point following a fill, for example, 10 seconds after execution. Passive orders are more likely to receive a fill when prices are about to improve (passive orders to buy are more likely to receive fills when the price is about to go down) and less likely to receive a fill when prices will soon be worse. If indeed passive orders are getting adversely selected, markouts will reflect that.

But for aggressive orders, adverse selection is not the primary concern because the execution algorithm is almost guaranteed to get that fill from either the SDP or the exchange, and hence, markouts are not useful in understanding the impact of SDPs on child order performance. If an SDP does not provide a fill when the price is about to get worse, the algorithm can trade unfilled shares on an exchange. Here again, the bigger issue is the information leakage of the parent order through the pattern of child order slices, and this can only be analyzed at the parent order level rather than through the markouts of executions.

MYTH: There is no harm in my broker segmenting my flow

SDPs make a concerted effort to encourage their broker clients to segment their order flow as much as possible, as discussed above. SDPs offer broker-dealers better fill rates and rebates in exchange for segmentation. A trader executing a large block order would prefer to blend it among many other client orders to best disguise its signals and preserve information about its potential market impact, but the more segmented the flow from a broker is, the easier it is for SDPs to establish patterns and predict prices based on order flow.

For example, a commonly requested segmentation by SDPs is that brokers indicate whether child order slices they send are part of the “last X%” of a parent order traded in an algorithm. SDPs benefit from the knowledge that they are trading against the last part of an order. If an institutional execution algorithm is buying and the end of the order is known (or known to be coming soon), the SDP knows the price will likely go down after the order completes. In that case, the SDP benefits from earning the spread and from the price action associated with the reversion after the order completes. As a result, SDPs are likely to provide even better fill rates and rebates for the last part of a parent order. While such outcomes sound like a win-win, they may have an unintended consequence.

Because there tends to be less temporary market impact created by later child order slices than earlier slices of a parent order (as illustrated above), it might be a reasonable hypothesis that there is less harm in sending the last slices to SDPs. As an order nears its end, the price impact of additional slices should be of far less concern. But the more segmentation a broker creates in their order flow, the more information is provided to an SDP. By segmenting the last part of an order, the broker is implicitly giving even more information about other flows not tagged as the last portion of an order, leaking information about untagged parent orders being executed.

Of course, a broker never really knows whether an algorithm is trading the last portion of a parent order. Many institutional firms break up their orders before submitting them to brokers, and they may have more to trade after a “parent order” sent to a broker algorithm is complete.

Another assumption the sending broker could make incorrectly is that prices will revert after the order completion. While that is likely to be the pattern for investment managers with a long-term investment horizon, it is not true for high-turnover investment managers (for example, investment managers engaged in statistical arbitrage) that have shorter investment horizons spanning only hours or days. For such firms, the cost of information leakage can be even higher if some of their alpha is re-engineered by SDPs.

Brokers Access SDPs to Reduce Explicit Costs, But Implicit Costs Dwarf Commissions

Algo brokers have been under immense pressure to reduce commissions, and SDPs help mitigate that. Like the payment for order flow (PFOF) practice in retail market structure, the interests of the algo brokers and institutional SDPs are aligned. SDPs provide liquidity at zero cost, sometimes even offering rebates to algo brokers. This is in direct contrast to the liquidity available at exchanges, where algo brokers pay as much as 30 mills per share to take liquidity. Assuming an SDP offers a rebate of 5 mills, the difference in cost between taking liquidity at an exchange versus an SDP can be as high as 35 mills per share—significant savings given the downward trajectory of commission rates for algo execution. For an algo broker, taking liquidity from SDPs instead of exchanges can convert an unprofitable account to a profitable account.

The incentives are big enough that algorithmic execution brokers are more likely to recycle the marketing pitches of SDPs in conversation with their institutional clients than to raise information leakage concerns. And SDPs do not only target flow from algorithmic brokers, they also source it by offering their own execution algorithms at a hefty discount (often at a rebate) to other brokers who do not operate algorithms in-house. In these cases, information leakage is likely to be even higher.

All that said, an important consideration that is often marginalized is that commission savings are generally negligible compared to the implicit execution costs (implementation shortfall) most investment managers incur. For example, if a broker saves 30 mills on a trade by sending an order to an SDP over an exchange or ATS and can get 20% of a parent order filled there, the result is 6 mills of savings per share. Even if broker passes the entire savings to its investment manager client, with the average price of an S&P 500 stock at approximately $60, this equates to .00001% (1/10^th of a basis point) in savings. But the implicit costs of execution can range from 5 basis points to 30 basis points per share for most investment managers. Setting conflicts of interest aside, if the use of SDPs leads to higher implicit trading costs, an increase of more than 0.1 basis point is simply not justifiable.

Conclusion

SDPs are just another source of liquidity alongside ATSs and exchanges, but they operate very differently. They are not registered as ATSs and receive much more information in their order flow than is typical for a liquidity provider in an ATS or on an exchange. While choosing only to act on indications of interest (IOIs) may reduce the amount of child order information sent to an SDP, it does not alleviate the issue of information leakage completely because child orders are serially correlated with other slices from their parent order.

Institutional investors should be aware of the differences in the way SDPs operate compared to traditional dark pools in order to best evaluate whether to and how to interact with SDPs in their execution algorithms. We recommend investors and brokers conduct controlled experiments to evaluate performance differences in parent orders rather than relying on misleading metrics around child order performance produced by SDPs or algorithmic execution brokers.

Finally, we sincerely hope that more focus is applied to the implicit costs of trading rather than explicit costs, given the difference in magnitude between the two. A focus on reducing implicit costs rather than explicit costs could be far more fruitful and should include considerations of market impact, adverse selection, and of course, information leakage.

‍

‍

End Notes

¹ Here, when we refer to a “parent order”, we mean the entirety of the order planned for trading by its institutional owner. Trading just 100 shares at an exchange is what we call a “child order” of the algorithm; the parent order could be 800,000 shares, for example.

² The full version of Emmanuel Bacry, Adrian Iuga, Matthieu Lasnier, and Charles-Albert Lehalle’s Market Impacts and the Life Cycle of Investors Orders can be found here.

³ The Price Impact of Order Book Events by Rama Cont, Arseniy Kukanov, and Sasha Stoikov compares the market impact of removing liquidity to the impact of other order book events and finds they have similar effects. The full version can be found here.

‍

‍

At BestEx Research, your trading costs keep us up at night. We know from experience that systematic, quantitative decision-making around order placement contributes to globally optimal execution and results in significantly reduced execution costs. We’re constantly refining our algorithms and our technology and recently implemented A/B testing for comparing strategies in just a few clicks from our cloud-based AMS, which could be used to conduct the controlled experiments discussed above. Whether you’re a customer of ours or not, we hope you’re asking questions and driving your brokers to design algorithm behaviors based on quantitative assessments with appropriately applied metrics. Reach out to us with questions at research@bestexresearch.com or learn more about us at bestexresearch.com.

©2022 BestEx Research. All rights reserved.

‍