Cul de Sacs Revisited: Dark Aggregation, Lit Markets, and the Cost of Duration

Transcription

1 AUTHOR Ian Domowitz Managing Director ITG Analytics Kristi Reitnauer Assistant Vice President ITG Analytics Colleen Ruane Director ITG Analytics CONTACT Asia Pacific Canada EMEA United States Cul de Sacs Revisited: Dark Aggregation, Lit Markets, and the Cost of Duration ABSTRACT We demonstrate links between order duration, dark pool performance, and orders filled in both lit and dark venues. Five questions are addressed. Does exposure of an order to lit markets affect trading performance in dark aggregation schemes? Do market conditions affect the degree of dark order exposure to lit markets? Do market conditions affect trading performance in the case of mixed dark and lit child orders? Is there a connection between market conditions and the duration of an order? Do trading costs increase with the number of venues visited by an order? In answering these questions, we find that dark pool aggregation is beneficial. The fragmentation of liquidity enabled through dark aggregation algorithms has not reduced transaction costs relative to directly accessing a single block crossing venue, however. INTRODUCTION In 008, we used the analogy of cul de sacs and highways to illustrate the potential of information leakage through dark liquidity aggregation. 1 Cul de sacs are dark; the highway from one to another has street lights. Traditional trading footprints may be seen on the highway, even if each cul de sac lacks a light. Information leakage is not strictly provable; rather, we observed a sharp increase in transaction costs for orders that spend any time on the highway, relative to direct access into a single crossing system. Information leakage remains a plausible explanation, however, and there are a few possibilities. For example, within the world of transaction cost analysis, formulas often are used to aggregate trades into orders, when the explicit order size is not made available. The same techniques can be used to examine the tape in real time to infer trading interest. A multiplicity of prints from disparate venues is a source of information. Gaming in dark pool executions has proved to be an issue in the context of liquidity aggregators, to the point of producing explicit safeguards. Gaming is all about a search for information in settings that are otherwise dark. 1 Ian Domowitz, Ilya Finkelstein, and Henry Yegerman, Cul de Sacs and Highways: An Optical Tour of Dark Pool Trading Performance, Journal of Trading, 008. See, for example, Hitesh Mittal, Are You Playing in a Toxic Dark Pool? A Guide to Preventing Information Leakage, Journal of Trading, 008.

2 Alternatively, traders are sending some part of the flow through the liquidity aggregator, simply to shop, and pieces to other venues at the same time. Leakage possibilities abound in that scenario. In our 008 work, we suggest a link between order duration, dark pool performance, and orders filled in both lit and dark venues. Given the current interest in venue analysis, we revisit this theme, and add an emphasis on volatility and volume conditions. THE QUESTIONS We ask and answer five questions in this paper: Does exposure of an order to lit markets affect trading performance in dark aggregation schemes? The answer is yes, and increasing lit market executions can affect transactions costs by orders of magnitude. Do market conditions affect the degree of dark order exposure to lit markets? Although changes in long-run expectations may have an effect, surprises in volatility and volume conditions do not. In what way, if at all, do market conditions affect trading performance in the case of mixed dark and lit child orders? We find that there is a tipping point in the percentage of lit market fills, after which market conditions have a significant impact, despite their negligible contribution to the dark/lit split percentage in terms of routing behavior. Is there a connection between market conditions and the duration of an order, measured by time or number of venues visited for completion? Except for extreme values, the connection is lacking. Although the correlation between the number of venues and duration is very high, the impact of number of venues visited on performance is not mitigated or exacerbated by market conditions. Do trading costs increase with the number of venues visited by an order? As one moves from a single venue to up to fifteen, costs increase sharply, by roughly a factor of five. This result is robust with respect to changes in market conditions. We limit the universe of data for this analysis. Trading activity is for the U.S. only, and restricted to large cap stocks. High frequency trading is concentrated in this capitalization subset. All activity took place over the 013 calendar year. Small orders are analyzed, up to one percent of median daily volume. This last restriction is not terribly binding, since the vast majority of orders submitted for algorithmic trading fall into this category. 3 On the other hand, the data are not limited to ITG executions; all buy-side participants represented in the sample trade with a variety of brokers. We also cover virtually the entire universe of execution venues in the U.S. during 013. This includes 40 dark pools and 50 venues classified as lit. 4 Finally, we focus only on algorithmic trading orders, which are classified as dark aggregation strategies. In other work, we have found that consideration of trading strategy is an essential component in assessing venue performance. 5 Other strategies may access a combination of lit and dark markets, but venues exhibit different characteristics for disparate strategies. We limit this analysis to the single strategy with the greatest affinity for dark pool executions. 3 We also exclude extreme cost outliers, defined by plus or minus 100 basis points (bps) versus the order decision price, as well as any orders for which an implementation shortfall calculation is not possible for any reason; orders with limits greater than 100 bps away from next price post decision price are excluded. 4 We have monikers for each venue and venue type. If a complex like Nasdaq hosts several different systems, we analyze each of those separately. A granular dissection of so-called grey markets is not supportable with our data, however. 5 Ian Domowitz, Kristi Reitnauer, and Colleen Ruane, Garbage In, Garbage Out: An Optical Tour of the Role of Strategy in Venue Analysis, ITG, August 014, at

3 3 MIXING LIT AND DARK EXECUTIONS At the time of writing the 008 paper on information transmission, we did not have data to confirm an obvious conjecture: execution characteristics in dark pools are adversely affected by shopping the order, not only to other dark pools, but also to lit markets. We fill that gap here, beginning with Figure 1. The chart illustrates implementation shortfall for dark aggregator orders, broken down by lit venue usage for fills distributed between dark and lit venues. The clock starts ticking for the implementation shortfall calculation at the time the order hits the dark aggregator algorithm. FIGURE 1: Implementation Shortfall Performance for Dark Aggregator Orders by Lit Venue Usage IS Performance (bps) A)80-100% Lit B)60-80% Lit C)40-60% Lit D)0-40% Lit E)0-0% Lit The trend is clear to the naked eye. Orders filled primarily in the dark, regardless of order duration, exhibit costs which are a fraction of those executed for the most part in lit markets. 6 Any increase in lit-execution percentage results in an increase in implementation shortfall. Pre-trade information leakage is one likely culprit, consistent with several possible explanations of how the lit/dark split is achieved by the router. Another explanation could lie in volatility conditions determining the extent of the lit/dark split in practice. 6 Average duration for orders in each of the dark/lit groups is similar. Those with the highest percentage filled in lit markets tend to have lower duration and higher costs. The difference between predominately lit and predominately dark is about six minutes.

4 4 This possibility is first explored in Figure, which illustrates venue usage characterized by the distribution of volatility conditions relative to expectations. FIGURE : Lit Venue Usage by Volatility Profile 100% 90% 80% 70% 60% 50% 40% 30% 0% 10% 0% E)0-0% Lit D)0-40% Lit C)40-60% Lit B)60-80% Lit A)80-100% Lit Although there is some variation in the chart, it is very small. Not only is the overall lit/dark percentage constant over volatility conditions, but the distribution of that split is constant as well. A similar chart for volume conditions shows the same result. As with the volatility graph, volume conditions are measured with respect to deviations from expectations. FIGURE 3: Lit Venue Usage by Volume Profile 100% 90% 80% 70% 60% 50% 40% 30% 0% 10% 0% E)0-0% Lit D)0-40% Lit C)40-60% Lit B)60-80% Lit A)80-100% Lit

5 5 In short, dark aggregation engines, and associated routing, are not responsive to deviations in expected volume and volatility. To the extent that the details of the usage of routing mechanisms are tweaked through instructions to the broker, the effect is negligible. A natural conclusion is that the end user is not tuning venue strategy to market conditions. The same result is observed in the case of algorithm selection, where despite the sensitivity of certain strategies to volume and volatility conditions, users tend not to change strategy once initially decided upon. 7 We investigate this further in the context of trading performance, illustrated in Figure 4. FIGURE 4: Implementation Shortfall with Varying Lit Venue Usage by Volatility Condition 4 0 IS Cost (bps) A)80-100% Lit B)60-80% Lit C)40-60% Lit D)0-40% Lit E)0-0% Lit Three conclusions emerge, despite a bit of noise in the low volatility category. The effects of changes in volatility are exacerbated for orders with over forty percent of completions in lit markets. Dark executions appear to be more insulated from shifts in volatility than those in the lit venues. Finally, there is a trend downwards in transaction costs, for every volatility condition, as orders are more fully completed in dark markets. 7 Ian Domowitz and Henry Yegerman, Algorithmic Trading Usage Patterns and their Costs, Journal of Trading, Summer 011. We note that the conclusions on routing are based on the history of fills, as opposed to detailed routing tables, to which we do not have access for all brokers.

6 6 NUMBER OF VENUES AND ORDER DURATION The duration of an order within a dark aggregator setting may also correlate with market conditions. Broadly speaking, minimization of opportunity cost entails keeping duration short, and opportunity cost increases with volatility. Longer duration in today s routing environment also should have a link to the number of venues within which executions occur. This is illustrated in Figure 5. FIGURE 5: Average Order Duration and Number of Venues Visited by an Order Duration (minutes) A) 1 B) -3 C) 4-5 D) 5-10 E) F) 15-0 G)>0 We will examine cost effects due to the number of venues and the associated duration of an order. First, we revisit the idea that order routing usage may not be influenced by market conditions. An analysis of volatility is contained in Figure 6. FIGURE 6: Distribution of Venues Visited as a Function of Volatility Conditions 100% 90% 80% 70% 60% 50% 40% 30% 0% 10% 0% A) 1 B) -3 C) 4-5 D) 5-10 E) F) 15-0 G)>0

7 7 There are three regimes represented in the chart. The trend in orders visiting over fifteen venues is down, as volatility increases relative to expectations. On the other hand, the percentage of orders touching five venues or less is fundamentally flat across volatility conditions. For the middle group (5 to 15 venues), a sharp eye may notice a bit of a trend upwards with volatility, but the differences are quite small in magnitude. Except for extreme values, there is little to no change in the number of venues used as a function of volatility surprises. Even less movement, in terms of number of venues visited, is observed in the case of volume surprises. FIGURE 7: Distribution of Venues Visited as a Function of Volume Conditions 100% 90% 80% 70% 60% 50% 40% 30% 0% 10% 0% A) 1 B) -3 C) 4-5 D) 5-10 E) F) 15-0 G)>0 In terms of usable information, this brings us to implementation shortfall as a function of routing behavior within dark aggregators. Aggregate statistics appear in Figure 8. FIGURE 8: Transaction Cost by Number of Venues Visited by an Order 3 1 IS Cost (bps) A) 1 B) -3 C) 4-5 D) 5-10 E) F) 15-0 G)>0

8 8 As one moves from a single venue to up to fifteen, costs increase sharply, by roughly a factor of five. These data from 013 mimic those of 007 in this regard. Combining this chart with Figure 5, costs rise rapidly as the average duration of the order climbs from under twenty minutes to an hour. Although duration naturally increases as the number of venues increases beyond fifteen, the number of observations from which we can calculate cost drops dramatically. The small sample size prevents us from interpreting what appears to be a decline in cost as the number of venues visited by an order becomes very large. As a check on the results, we compute implementation shortfall as a function of venue count, by volatility regime. This exercise is illustrated in Figure 9. FIGURE 9: Implementation Shortfall by Number of Venues and Volatility Conditions IS Cost (bps) A) 1 B) -3 C) 4-5 D) 5-10 E) F) 15-0 G)>0 As in Figure 8, the narrative is relatively clear until one reaches an extreme number of venues. Controlling for volatility, costs increase as the number visited goes up. Since routing behavior appears to be relatively unchanged across volatility regimes, the information in this chart essentially replicates that of Figure 4. The same graph for volume conditions, exhibits the same qualitative behavior. GENERALIZATIONS AND OTHER MARKETS Detailed conclusions were presented in the introduction, but the overall message is clear: dark pool execution and associated dark aggregation schemes are beneficial. The fragmentation of liquidity enabled through dark aggregation algorithms has not reduced transaction costs relative to directly accessing a single block crossing venue, however. Costs are more adversely affected by the amount of exposure to lit markets in aggregation schemes than they are impacted by market conditions alone. Market conditions themselves do not have much of an effect on lit market exposure, nor on the number of venues visited by an order, or on order duration. This should focus attention on the nature of the dark pool cul de sac and the lit highways which connect them, as opposed to whether it is raining on the road.

9 9 The analysis here is limited to trading in the U.S. A smaller, but similar sample in terms of order size and capitalization from European trading exhibits the same sort of results. In particular, as the percent of an order done in the dark rises, transaction costs fall. 8 Further, longer order duration is highly correlated with increases in the observed cost of trading. Nevertheless, orders with less than 0 percent of shares filled in dark venues represent over 50 percent of total value. There is more work to be done along these lines, especially given the move to limit dark venues in Europe. The qualitative nature of all conclusions echoes results based on 007 data. Comparisons of the current findings with other authors work also changes little. For example, George Sofianos and David Jeria find that the use of a dark pool in the context of VWAP strategies reduces execution shortfall by 1 percent, relative to order arrival price. 9 This leads us to a much broader generalization: despite all the publicity these days, the interaction between market structure and behavior has not changed much in the last six years. 8 Source, ITG, based on averages over George Sofianos and David Jeria, Quantifying the SIGMA X Crossing Benefit, in Street Smart, Goldman Sachs, March 31, Investment Technology Group, Inc. All rights reserved. Not to be reproduced or retransmitted without permission These materials are for informational purposes only and are not intended to be used for trading or investment purposes or as an offer to sell or the solicitation of an offer to buy any security or financial product. The information contained herein has been taken from trade and statistical services and other sources we deem reliable but we do not represent that such information is accurate or complete and it should not be relied upon as such. No guarantee or warranty is made as to the reasonableness of the assumptions or the accuracy of the models or market data used by ITG. These materials do not provide any form of advice (investment, tax or legal). All trademarks, service marks, and trade names not owned by ITG are the property of their respective owners. The positions taken in this document reflect the judgement of the individual author(s) and are not necessarily those of ITG. Broker-dealer products and services are offered by: in the U.S., ITG Inc., member FINRA, SIPC; in Canada, ITG Canada Corp., member Canadian Investor Protection Fund ( CIPF ) and Investment Industry Regulatory Organization of Canada ( IIROC ); in Europe, Investment Technology Group Limited, registered in Ireland No ( ITGL ) and/or Investment Technology Group Europe Limited, registered in Ireland No ( ITGEL ) (the registered office of ITGL and ITGEL is First Floor, Block A, Georges Quay, Dublin, Ireland and ITGL is a member of the London Stock Exchange, Euronext and Deutsche Börse). ITGL and ITGEL are authorised and regulated by the Central Bank of Ireland; in Asia, ITG Hong Kong Limited, licensed with the SFC (License No. AHD810), ITG Singapore Pte Limited (CMS Licence No ), and ITG Australia Limited (AFS License No. 1958). All of the above entities are subsidiaries of Investment Technology Group, Inc. MATCH NowSM is a product offering of TriAct Canada Marketplace LP ( TriAct ), member CIPF and IIROC. TriAct is a wholly owned subsidiary of ITG Canada Corp.