Equity Execution Strategies. Issue 35 October 16, When the going gets tough, the algos get going

Transcription

1 Equity Execution Strategies Issue 5 October 6, 8 Mark Gurliacci mark.gurliacci@gs.com NY: David Jeria david.jeria@gs.com NY: George Sofianos george.sofianos@gs.com NY: --57 Related analysis: Gurliacci, Jeria & Sofianos, The short-sell ban and quoted spreads, Issue, October, 8 Sofianos & Abrokwah, Market turmoil and trading costs, Issue 7, August 7, 7 Sofianos & Cai, Algo performance in market turmoil, Issue 8, September, 7 Rakhlin & Sofianos, The impact of an increase in volatility on trading costs,, Journal of Trading, Spring 6 When the going gets tough, the algos get going In this issue of, we examine how the current market turmoil is affecting trading costs in general and algo performance in particular. In our analysis we compare the September 8 market turmoil period to a normal benchmark period, May 8. We find that in September: Quoted spreads on the S&P 5 stocks increased 5 percent and quoted depth decreased 5 percent Algos made up 5 percent of the total value clients executed through Goldman Sachs compared to 6 percent in May, so despite the market turmoil clients did not shift away from algos Algo clients, however, shifted away from the relatively passive VWAP algo to more aggressive algos and within each algo chose more aggressive settings The liquidity impact of the average,-share order in our sample more than doubled to 5 ( in May); part of this increase was caused by the more aggressive executions Alpha-to-close was higher, so despite the higher liquidity impact the average Goldman Sachs algo client captured some alpha-to-close Exhibit. Liquidity impact of algo executions a May to Sep to Extreme Friday (Sep ) Sep to Volatility b.6%.8% 7.%.% 5.% Alpha-to-close c Alpha-to-close capture d Execution half-life e Participation rate f Benchmark period: May 8 +5 min 7% Market turmoil period: September 8 Sharp increase in liquidity impact +5 5 min 6 +6 min 6% 8 Quad-witch, bail-out plan announced, short-sell ban on financials, NASDAQ erroneous prints. -57 min % +6 5 min % Sharp increase in execution aggressiveness Goldman Sachs Equity Execution Strategies This material has been prepared by Goldman Sachs Equity Execution Strategies and is not a product of Global Investment Research or Fixed Income Research. Important disclosures appear at the end of this material. a. GSEC algo flow, external clients only (excludes GSCO flow) ; $ value-weighted averages, market orders greater than, shrs; each bucket excludes lowest and highest shortfall order. b. S&P5 stocks; for each stock, daily high minus daily low as percent of average high plus low; we average across stocks using the S&P5 index weights. c. For buy orders, closing price minus arrival price as percent of arrival price; for sell orders, arrival price minus execution price as percent of arrival price. d. For buy orders, closing price minus execution price as percent of arrival price; for sell orders, execution price minus closing price as percent of arrival price. e. For each parent algo order, the value-weighted mean of each child order's execution time. f. Executed quantity as percent of consolidated market volume over execution horizon. Exhibit summarizes our main finding on algo performance. In our analysis, we divide the market turmoil period into four sub-periods:. September to ; modest increase in volatility and liquidity impact. September 5 to 8; sharp increase in volatility and sharp increase in liquidity impact. September ; an extreme, outlier day, liquidity impact spikes and, adding insult to injury, alpha-to-close is negative. September to ; volatility remains sharply higher, algo clients switch to more aggressive executions and liquidity impact remains high We next discuss our findings in more detail.

2 The volatility environment Exhibit shows that average daily volatility in the S&P 5 stocks was.6 percent in May and more than doubled to 5.5 percent in September. From September to volatility increased slightly to.8 percent. Beginning September 5, the volatility increase started accelerating, averaging 7. percent through September 8. On September, extreme Friday, volatility reached. percent. Extreme Friday was a quadruple expiration ( quad-witch ) and on that same day the US Treasury announced its bail-out plan, the SEC imposed a short-sell ban on Exhibit. The volatility environment, S&P 5 stocks a % May to Sep to Sep Sep to %.8% 7.%.% 5.% % Sep to / 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/7 5/8 5/ 5/ % Sharp increase in volatility, S&P 5 stocks / / / /5 /6 /7 /8 / / / / / / /5 /6 /7 /8 / / / / / / /5 /6 /7 /8 / / a. For each S&P 5 stock, daily high minus daily low as percent of average high plus low; we average across stocks using the S&P5 index weights. financial stocks and NASDAQ reported a large number of erroneous prints. For all these reasons, we treat September as an extreme outlier and examine it separately. For the rest of the month (Sep to ), volatility remained high, averaging 5. percent with another spike on September when the US House of Representatives voted against the bail-out plan. Quoted spreads and depth To put algo performance in context, we next examine the effect of the market turmoil on the basic execution building blocks: quoted spreads and depth. We focus on the S&P5 stocks. Exhibit. The basic building blocks: quoted half-spreads, S&P 5 stocks a May to Sep to Sep Sep to Sharp increase in quoted half-spreads, S&P 5 stocks / 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/7 5/8 5/ 5/ / / / /5 /6 /7 /8 / / / / / / /5 /6 /7 /8 / / / / / / /5 /6 /7 /8 / / a. For each S&P 5 stocks, the intra-day time-weighted NBBO; we average across stocks using the S&P5 index weights; we use half spreads to be consistent with the executions shortfall measure (one-way cost). Exhibit shows that in September quoted spreads on the S&P 5 stocks widened substantially. In May, the average quoted half-spread was.7. 5 From September to half-spreads widened slightly to. and from September 5 to half-spreads widened sharply to.6, 5 percent higher than in May. The two most volatile days in our sample had the highest half-spreads:. on September and. on September. 6 Goldman Sachs Equity Execution Strategies

3 Exhibit. The basic building blocks: quoted depth, S&P 5 stocks a shares 6, 5,,,,,68,865,7,,8,67,55,7,6 5,5,66 6,7 5, 5,68,75,,86,,5,78, shares 6 5,76,,88,8,55,,,76, May to Sep to Sep Sep to,6,78,7,876,7,,,,6,, shrs, shrs, shrs, shrs, shrs,6,8, Quoted depth dropped 5 percent, S&P 5 stocks 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/7 5/8 5/ 5/ / / / /5 /6 /7 /8 / / / / / / /5 /6 /7 /8 / / / / / / /5 /6 /7 /8 / / a. For each S&P 5 stock, the intra-day tine-weighted size at best (NBBO) bid plus size at best ask divided by two; we average across stocks using the S&P 5 index weights. As spreads widened, quoted depth declined. Exhibit shows that the average quoted depth in May was, shares, dropped to, shares from September to, and averaged, shares for the rest of September. In summary, Exhibits and suggest that the cost of executing small (less than quoted depth) market orders in the average S&P 5 stock increased by more than 5 percent. We next examine how the market turmoil is affecting the cost of executing much larger algo orders. Algo usage and performance Exhibit 5. Algo choice a Cast % Particip ate 8% Cast % Particip ate % Sonar % TWAP PortX TWAP 6% PortX % Other Other 8% May 8 VWAP 8% VWAP 8% September 8 Sonar 5% a. $ value executed; GSEC algo flow, external clients only (excludes GSCO flow) In September, clients self-directed through the various Goldman Sachs algos 5 percent of the total dollar value they executed using Goldman Sachs high-touch and low-touch channels. 7 In May, the algo share was 6 percent. So despite the market turmoil, clients did not shift away from algos. Similarly, the mix of algo clients (hedge funds, traditional asset managers, etc.) did not change during the market turmoil. This is important because it suggests that the May-to-September differences in algo usage were caused by changes in the way the average client used the algos and not by changes in the client mix. 8 In September, for example, the average client shifted away from the passive VWAP to the more aggressive Sonar and Participate algos (Exhibit 5). In our analysis we focus on algo market orders greater than, shares. After some additional filtering, our final sample consists of 7,6 orders in May and 5, orders in September. The much larger September sample reflects the market-wide volume surge during the market turmoil. As usual, our main execution quality measure is execution shortfall. An important determinant of execution shortfall is the underlying stock price move over the execution horizon. We measure this underlying price move by the alpha-to-close. Exhibit 6 shows that alpha-to-close was higher during the market turmoil. From September 5 to 8, for example, alpha-to-close was 7 compared to 7 in May. Exhibit 6 makes clear that alpha-to-close is an important factor in understanding algo performance during the market turmoil. We therefore introduce an expected shortfall benchmark adjusted to incorporate the alpha-to-close effect. We define expected shortfall as the pre-trade expected liquidity impact plus an estimate of the execution horizon (EH) alpha loss. We get the expected liquidity impact from the Goldman Sachs t-cost model and estimate the EH-alpha loss from the actual alpha-to-close (see appendix). For each executed order, our expected shortfall measure answers the following question: given the actual alphato-close, what should the shortfall have been on this particular order? Goldman Sachs Equity Execution Strategies

4 Exhibit 6. Alpha-to-close: algo market orders a May to Sep to Sep Sep to Increased alpha-to-close opportunities? / 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/ 5/7 5/8 5/ 5/ -5 / / / /5 /6 /7 /8 / / / / / / /5 /6 /7 /8 / / / / / / /5 /6 /7 /8 / / -5 a. GSEC algo flow, external clients only (excludes GSCO flow) ; $ value-weighted averages, market orders greater than, shrs Exhibit 7 compares algo performance in May (normal times) and September (market turmoil). In May, the average order size was 8,5 shares (three percent of same day s volume). 5 With the exception of extreme Friday, the average order size in September was similar to May making comparisons easier. In May, the average execution was moderately aggressive, with seven percent participation rate and minutes execution half-life. 6 Expected shortfall was 8 and actual shortfall was. During May, therefore, the Goldman Sachs algos performed 5 better than expected. Alpha-to-close was 8, so algo clients also captured 5 of alpha-to-close. Exhibit 7. Market turmoil and algo performance a RECEIVED EXECUTED Order size % # Volume Shares Arrival time Fill rate Execution half-life (min) Expected shortfall Actual shortfall Participation rate Alpha-toclose Alpha-toclose capture May to.% 8,5 : 88% 7% Sep to.8% 5,688 :5 5% %, :8 88% 6% Friday Sep.7%,77 :5 7% % Sep to %,865 : 6% % We next examine algo performance in September: a. GSEC algo flow, external clients only (excludes GSCO flow) ; $ value-weighted averages, market orders greater than, shrs; each bucket excludes lowest and highest shortfall order. September to, volatility was slightly higher and algo performance did not change much. Actual shortfall was, better than expected ( ) and like in May, clients captured 5 of alpha-to-close ( ). September 5 to 8, volatility surged, quoted spreads widened and depth dropped. Actual and expected shortfall rose sharply to 56 and 5, respectively. One factor explaining this rise in actual and expected shortfall was the big jump in alpha-to-close (7 ). Actual shortfall was higher than expected because of the more difficult environment, but also because pre-trade cost models adjust slowly to rapid changes in the environment so expected shortfall does not fully reflect the market turmoil. 7 On the bright side, despite the high shortfall, algo clients captured 6 of alpha-to-close. September to, volatility and spreads remained high and algo clients sharply increased their execution aggressiveness. The average participation rate rose to percent compared to seven percent in May and six percent from September 5 to 8. The average execution half-life dropped to 5 minutes from minutes in May and minutes from September 5 to 8. 8 We see this increase in execution aggressiveness both in the shift away from the passive VWAP algo (Exhibit 5) and with clients executing at higher participation rates within each algo. Actual shortfall over this period was, 8 Goldman Sachs Equity Execution Strategies

5 higher than expected (5 ). Algo clients, nevertheless, managed to capture 6 of alpha-to-close ( ). Extreme Friday again stands out because clients were executing aggressively while the alpha-to-close was on average negative. Actual shortfall was but expected shortfall was only 6. Expected shortfall was low because the alpha to close was negative (selling in a rising market) and this reduced expected shortfall. On extreme Friday, therefore, actual shortfall was 8 higher than expected. The average alpha-to-close was minus, so the average algo client lost 57 relative to the closing price. Our discussion so far focused on execution shortfall, but execution shortfall has two components, liquidity impact and the EH-alpha loss. The large variability of alpha-to-close we see in Exhibits 6 and 7 introduces large variability in the alpha-loss component of shortfall and makes it difficult to isolate the effect of market turmoil on liquidity impact. In the next section we estimate the liquidity impact component of execution shortfall and show how it was affected by the market turmoil. Isolating liquidity impact Exhibit 8. Backing out liquidity impact from shortfall Benchmark period Arrival price = Liquidity impact = execution shortfall execution horizon alpha loss Actual shortfall Actual impact : Execution half-life min 6: 87 min EH-alpha loss Alpha-to-close 8 Exhibit 8 shows how we estimate liquidity impact. Liquidity impact equals execution shortfall minus EH-alpha loss. We know the actual shortfall ( in May) and we estimate the EHalpha loss from the actual alpha-to-close, 8. This 8 alpha move took place over 87 minutes from order arrival (:) to the close. The execution half life was minutes, so we allocate 6 percent (/87) of the 8 alpha-to-close as the alpha loss ( ). In May, therefore, the shortfall consisted of liquidity impact and alpha loss. Back to Exhibit, we see the dramatic effect the market turmoil is having on liquidity impact. From September to, with volatility moderately higher, liquidity impact was, up from in May. Excluding extreme Friday, from September 5 to, with volatility surging, liquidity impact averaged 5,.5 times the May level. On extreme Friday, liquidity impact was 8, times the May level. Part of the increase in liquidity impact reflects the higher execution aggressiveness during the market turmoil and the higher aggressiveness is an optimizing response to sharply higher execution risk and alpha-to-close. A large part of the liquidity impact increase, however, reflects liquidity scarcity and the need to pay up to attract counterparties in a very uncertain environment. Is the end of turmoil near? High volatility continues and is likely to continue in the months ahead. Our sample ends on September. From October to, average daily volatility in the S&P 5 stocks was even higher,.5 percent compared to 5.5 percent in September. We will continue to monitor the effects of the ongoing market turmoil on trading costs in general and algo performance in particular, and we will provide updates. Goldman Sachs Equity Execution Strategies 5

6 Exhibit A. Derivation of expected shortfall a May to Arrival price = Arrival Half-life Expected impact time (min) : Goldman Sachs model 5 EH-alpha loss Expected shortfall 8 Expected shortfall = Expected liquidity impact + Execution horizon alpha loss Expected shortfall Expected impact 5 The two components of shortfall 8 : Execution half-life min 6: 87 min EH-alpha loss Alpha-toclose 8 Alpha-to-close 8 a. GSEC algo flow, external clients only (excludes GSCO flow) ; $ value-weighted averages, market orders greater than, shrs; each bucket excludes lowest and highest shortfall order. APPENDIX: Derivation of expected shortfall Exhibit A shows how we estimate expected shortfall and mirrors the way we back-out liquidity impact from actual shortfall in Exhibit 8. In both cases we calculate the EH-alpha loss in the same way by allocating the actual alpha-to-close (8 in May) to the actual execution horizon. The allocation factor is 6 percent, the ratio of the average execution half-life ( minutes) to the time from order arrival to close (87 minutes). To estimate expected shortfall we add the EH-alpha loss to the expected liquidity impact which we get from the Goldman Sachs t-cost model (5 ). In May, therefore, expected shortfall was 8, 5 expected impact and expected alpha loss. Expected shortfall as we define it here is not a pre-trade measure since it requires the actual alpha-to-close and is best interpreted as a post-trade attribution measure. The average time-weighted quoted half-spread on the S&P 5 stocks increased from.7 in May to.6 from September 5 to ; average quoted depth decreased from, shares to, shares. The algo percentage reflects flow clients self-directed to Goldman Sachs algorithms; the total value executed with Goldman Sachs includes both high-touch GSCO flow and low-touch GSEC flow. Average daily volatility in the S&P 5 stocks was only. percent in July 7 right before the August 7 market turmoil. During the August 7 market turmoil, volatility rose to. percent. See Sofianos & Abrokwah Market turmoil and trading costs,, Issue 7, August 7, 7. In calculating the average daily volatility we average across the S&P 5 stocks using the index weights. For each stock we define daily volatility as high minus low as percent of high plus low divided by two. Quad witch is a quarterly event in which stock index futures and options, and individual stock futures and options all expire at the same time. The NASDAQ prints issue happened between :5 and :. The SEC announced the short-sell ban on financial stocks the evening of September 8 effective at start of trading on September. 5 The full spread measures the round-trip cost of liquidity. We prefer using the half-spread (the one-way cost) to align it with our execution shortfall measure which measures the one-way cost of trading. 6 On September the SEC imposed a short-sell ban on financial stocks. Beginning on September and throughout the period of the ban that ended on October, quoted spreads on the short-sell ban stocks were on average percent wider than spreads on non-ban stocks, see Gurliacci, Jeria and Sofianos The short-sell ban and quoted spreads, Issue, October, 8. 7 The other Goldman Sachs execution channels are: (a) high-touch single-stock trading, (b) high-touch portfolio trading, (c) DMA SIGMA smart-router, (d) DMA SIGMA X crossing network, and (e) DMA client-directed to specific execution venues. 8 We repeated the analysis using a constant client universe for May and September; our results do not change. Market orders were 6 percent of algo executed value in May and 6 percent in September. We exclude limit orders from our analysis because limit orders have low fill rates and to properly evaluate we need to estimate the cost of he clean-up trade and natural adverse selection (see Jeria and Sofianos, Passive orders and natural adverse selection,, Issue, September, 8). For example, we focus on common stocks, drop Pink Sheets, drop high-price (>$5) and low-price stocks (<$). September also includes the quad witch with its traditional volume surge. For buy orders, execution shortfall equals execution price minus midquote at arrival as percent of arrival midquote. For sell orders, execution shortfall equals arrival midquote minus execution price as percent of arrival midquote. For buy orders, alpha-to-close equals closing price minus midquote at arrival as percent of arrival midquote. For sell orders, alpha-to-close equals arrival midquote minus closing price. For an empirical analysis of the importance of alpha-to-close see Abrokwah and Sofianos, Shortfall Surprises, Journal of Trading, Summer 7. Exhibit 6 also shows that extreme Friday was unusual in one more way: alpha-to-close was negative, a rare event in normal times (happened again September 5). On extreme Friday the DJIA closed 6 points up from the previous day close, so negative ST-alpha suggests the average client in our sample was selling in a rising market. 5 Average arrival time was : and the average fill rate 88 percent. Since we exclude limit orders from our sample, non-filled orders are either market orders that clients canceled or orders with volume participation limits. 6 We use the following rule of thumb: participation rates less than percent indicate passive executions, more than percent aggressive executions, and in-between moderate aggressive executions. 7 Model sluggishness arises in two ways: from the estimated model coefficients and from the factor input values. We re-estimate the Goldman Sachs t-cost model coefficients every month using a nine-month rolling window, so a new regime affects the model coefficients gradually. Typically the model coefficients do not change much even with dramatic regime changes. The model inputs are more important for sluggishness. Among the inputs used in the Goldman Sachs model are volatility, quoted spreads and volume. The model has two calibrations: one uses the prior -day median values for these inputs and the other the prior 5-day median values. We tested the model s response to the market turmoil and found that the -day medians lead to a sluggish response but the 5-day medians catch up quickly to the new regime. 8 Part of the reduction in half-life and increase in participation rate was caused by later order arrivals. The average arrival time from September to was : compared to : in May and :8 from September 5 to 8. Fill rates also dropped: 6 percent from September to compared to 88 percent in May and from September 5 to 8. The lower fill rates suggest that clients were canceling their orders more frequently or faster. Also see Exhibit in Abrokwah and Sofianos, Shortfall Surprises, Journal of Trading, Summer 7. The critical assumption behind this calculation is that the alpha move from arrival to close is linear. Although this is unlikely to be the case for each individual execution, this is a reasonable assumption when averaging over a large number of executions as we do here. 6 Goldman Sachs Equity Execution Strategies

7 This material has been prepared by personnel in Securities Division Sales & Trading area of one or more affiliates of The Goldman Sachs Group, Inc. ( Goldman Sachs ) and is not the product of Global Investment Research or Fixed Income Research. It is not a research report and is not intended as such. Non-Reliance and Risk Disclosure: This material should not be construed as an offer to sell or the solicitation of an offer to buy any security in any jurisdiction where such an offer or solicitation would be illegal. We are not soliciting any action based on this material. It is for the general information of our clients. It does not constitute a recommendation or take into account the particular investment objectives, financial conditions, or needs of individual clients. Before acting on any advice or recommendation in this material, you should consider whether it is suitable for your particular circumstances and, if necessary, seek professional advice. Certain transactions - including those involving futures, options, equity swaps, and other derivatives as well as non-investment-grade securities, foreign-denominated securities and securities, such as ADRs, whose value is influenced by foreign currencies - give rise to substantial risk and may not be available to or suitable for all investors. This material is not for distribution to private customers, as that term is defined under the rules of the Financial Services Authority in the United Kingdom; and any investments, including derivatives, mentioned in this material will not be made available by us to any such private customer. The material is based on information that we consider reliable, but we do not represent that it is accurate, complete or up to date, and it should not be relied on as such. Opinions expressed are our current opinions as of the date appearing on this material and only represent the views of the author and not those of Goldman Sachs, unless otherwise expressly noted. Legal Entities Disseminating this Material: This material is disseminated in Australia by Goldman Sachs JBWere Pty Ltd (ABN ) on behalf of Goldman Sachs; in Canada by Goldman Sachs Canada Inc. regarding Canadian equities and jointly by Goldman, Sachs & Co. and Goldman Sachs Execution & Clearing, L.P. regarding all other materials; in Hong Kong by Goldman Sachs (Asia) L.L.C.; in Japan by Goldman Sachs Japan Co., Ltd.; in the Republic of Korea by Goldman Sachs (Asia) L.L.C., Seoul Branch; in New Zealand by Goldman Sachs JBWere (NZ) Limited on behalf of Goldman Sachs; in Singapore by Goldman Sachs (Singapore) Pte. (Company Number: 8665W); in Europe by Goldman Sachs International (unless stated otherwise); in France by Goldman Sachs Paris Inc. et Cie and/or Goldman Sachs International; in Germany by Goldman Sachs International and/or Goldman, Sachs & Co. ohg; in Brazil by Goldman Sachs do Brasil Banco Múltiplo S.A.; and in the of America, unless otherwise noted, by each of Goldman, Sachs & Co. and Goldman Sachs Execution & Clearing, L.P. (both of which are members NASD, NYSE and SIPC). Goldman Sachs International, which is authorized and regulated by the Financial Services Authority, has approved this material in connection with its distribution in the United Kingdom and European Union. Unless governing law permits otherwise, you must contact a Goldman Sachs entity in your home jurisdiction if you want to use our services in effecting a transaction in the securities mentioned in this material. Reproduction and Re-Distribution: No part of this material may be (i) copied, photocopied or duplicated in any form by any means or (ii) redistributed without our prior written consent. 8, The Goldman Sachs Group, Inc. All rights reserved. Goldman Sachs Equity Execution Strategies 7