A Buy-side Handbook

A buy-side handbook

Algorithmic trading

Published by The TRADE in association with:

Published by

1 Sekforde Street London EC1R 0BE UK www.thetrade.ltd.uk Tel: +44 (0) 20 7075 6115

Graphic design by Tina Eldred [email protected]

©The Trade Ltd. London 2005. Although The Trade has made every effort to ensure the accuracy of this publication, neither it nor any contributor can accept any legal responsibility whatsoever for consequences that may arise from errors or omissions or any opinions or advice given. This publication is not a substitute for professional advice on a specific transaction. No reproduction allowed without prior permission.

■ A buy-side handbook Algorithmic trading

Foreword ifferentiating between the algorithmic trading offerings of brokers D remains a problem for the buy-side. At the same time, brokers are searching for ways to achieve competitive edge and raise the profile of their algorithmic trading capabilities. These issues have to be overcome to realise the exponential growth that is forecast for algorithmic trading. The TRADE in association with leading industry participants drawn from the brokerage and vendor communities has set out to bring clarity and thought-leadership to the issues that are driving developments in the algorithmic space by publishing ‘A buy-side handbook on algorithmic trading’. Part 1, ‘Market and mechanics’, examines what is driving the growth of algorithmic trading, focusing on the rapidly evolving shape of the market. Insights are offered into how algorithms work and the relative merits of broker-driven versus broker-neutral algorithms are quantified. Part 2, ‘Honing an algorithmic trading strategy’, highlights the issues that buy-side traders must address once the decision has been taken to adopt an algorithmic strategy. Selecting an appropriate trading benchmark, the importance of anonymity to stem information leakage, applying stealth through sophisticated gaming theory, and customisation of broker algorithms are all addressed here. Part 3, ‘Quantifying and enhancing value’, focuses on measuring and interpreting the performance of disparate broker algorithms, the value added through independent third-party transaction cost analysis and the role of technology in enhancing market access. Part 4, ‘Emerging trends and future direction’, covers ‘next generation’ algorithms, focusing on implementation strategies for basket trading and the shape of the market going forward, when competition and increased buy-side demand will call for a higher order of intelligence in engineering algorithms. The handbook is completed with a guide to broker algorithms, containing details of individual broker offerings and including information on the range of benchmarks available, levels of customisation, performance measurement and connectivity options. ■ John Lee Editor & Publisher The TRADE ■ ALGORITHMIC TRADING

■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

3

■ A buy-side handbook – Algorithmic trading Contents

Part 2: Honing an algorithmic trading strategy

Part 1: Market and mechanics page 9 Chapter 1: Algorithmic trading – Upping the ante in a more competitive marketplace

page 41 Chapter 4: Choosing the right algorithm for your trading strategy

Wendy Garcia, analyst, TABB Group Tracy Black, executive director, European Sales Trading, UBS Investment Bank

4 page 21 Chapter 2: Understanding how algorithms work

Owain Self, executive director – Equities, UBS Investment Bank

Dr Tom Middleton, head of European Algorithmic Trading, Citigroup

page 51 Chapter 5: Anonymity and stealth Richard Balarkas, global head of AES™ Sales, CSFB

page 29 Chapter 3: Build or buy? Allen Zaydlin, CEO, InfoReach

■ ALGORITHMIC TRADING


page 59 Chapter 6: Customising the broker’s algorithms Richard Balarkas, global head of AES™ Sales, CSFB

■ THE TRADE 2005

■ A buy-side handbook – Algorithmic trading Contents

Part 3: Quantifying and enhancing value page 67 Chapter 7: Measuring and interpreting the performance of broker algorithms

Henry Yegerman, director, ITG Inc.

Ian Domowitz, managing director, ITG Inc.

page 79 Chapter 8: Making the most of third-party transaction analysis: the why, when, what and how? Robert Kay, managing director, GSCS Information Services

Part 4: Emerging trends and future direction page 97 Chapter 10: Basket algorithms – The next generation

Richard Johnson, senior vice president in charge of Product Sales, Miletus Trading

Anna Bystrik, PhD, research analyst, Miletus Trading

page 107 Chapter 11: The future of algorithmic trading

Carl Carrie, head of Algorithmic Trading, USA, JP Morgan

Andrew Freyre-Sanders, head of Algorithmic Trading, EMEA , JP Morgan

Robert L Kissell, vice president, Global Execution Services, JP Morgan

page 89 Chapter 9: Enhancing market access

Appendix

Mark Muñoz, senior vice president, Corporate Development, Nexa Technologies

Mark Ponthier, director – Engineering, Automated Trading Systems, Nexa Technologies


page 115 The TRADE guide to broker algorithms page 130 Contact information ■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

5

part 1:

Market and mechanics

9

Chapter 1 Algorithmic trading – Upping the ante in a more competitive marketplace

21

Chapter 2 Understanding how algorithms work

29

Chapter 3 Build or buy?

■ Chapter 1 Market and mechanics

Algorithmic trading – Upping the ante in a more competitive marketplace What will fuel the growth in algorithmic trading, and what impact will the widespread adoption of algorithms have on the direction of order flow? Wendy Garcia*

t is undeniable that the popularItinues ity of algorithmic trading conto grow. Market participants, especially those on the buyside, have willingly folded algorithmic trading models into their everyday methodologies in efforts to effectively and anonymously find sources of liquidity for their large orders, while simultaneously minimising market impact. Over the past year, adoption of algorithms has grown faster than any other trading tool. This rapid acceptance is attributable primarily to changes in market structure, cost, efficiency, and the need to achieve best execution. The questions begin to develop when examining the market conditions that are expected to support the further growth of this particular ■ ALGORITHMIC TRADING

electronic trading mechanism, and how its evolution and even deeper acceptance within the marketplace will impact the direction of order flow in the future. Firms are significantly reallocating the way they route their orders to the market in response to a number of interdependent and unique forces, one of which is the increasingly difficult struggle to provide access to liquidity centres. Brokers and technology providers are offering better and more integrated technologies to both access and utilise low- and no-touch trading technologies, making trading easier and more efficient. This year, TABB Group has seen tremendous swings in the way buy-side traders route their order flow. For example, ■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

9

* Wendy Garcia, analyst, TABB Group


■

“The compound annual growth rate for algorithm use from 2004 through 2007 is projected at 34%.”

10

overall, buy-side firms routed 17% less order flow over the phone than they did only one year ago, and if this trend were to be projected out to 2007, buy-side traders will only route 20% of their flow by phone. At the same time, we have seen a sizable increase in FIX-based order flow from 2004 to 2005, from 7% to 38% of flow. However, FIX traffic will level out, remaining steady through 2007. On a relative basis, we still see algorithms making the greatest advances over that same time period, as the compound annual growth rate for algorithm use from 2004 through 2007 is projected at 34%. Drivers of algorithmic use There are a number of drivers behind the increased use of algorithms, including the changing dynamics of the relationship between the buy-side and sell-side, the more sophisticated needs of the buy-side trader, and the increasing presence of order management system (OMS) vendors. ■ ALGORITHMIC TRADING


The buy-side/sell-side relationship: One key factor that has and will continue to play a significant role in the way in which algorithms are developed is the changing relationship the buy-side has with its brokers. As algorithms have become more widely accepted and usage has increased, the buyside is anxious to trim the number of broker relationships it maintains, in part to help reduce costs as they look to brokers for execution-only services. It is typical in today’s markets that, unless a broker has an improved algorithmic trading model to offer for order execution, buy-side traders are not interested in developing new relationships or even furthering existing relationships. This reality has significantly strengthened the competition among brokers, as well as redefined the basis on which partnerships are developed and maintained. What used to be seen as key in holding on to clients, namely a long-standing trusted client/broker relationship, is no longer considered a primary foothold against the competition. Currently, a small number of ‘bulge-bracket’ brokers lead the way in algorithmic trade model development, with a coterie of firms vying for valuable scraps. While in 2004 there were a mere ■ THE TRADE 2005


handful of brokers providing algorithms, today there is a much wider selection of broker-based algorithms on firms’ desks. Competition in this market will only become fiercer as more firms enter the market, additional strategies are created, buy-side adoption grows and high commission business becomes more challenging. In order to compete more effectively, many sell-side firms have developed targeted marketing strategies that align their models with specific customer segments. While those brokers who were first off the line with algorithmic trade models have succeeded at landing more clients, others more recently have focused on model customisation as a way to gain buy-side favour in an increasingly competitive environment. Still other firms have targeted size, focusing on either large or small firms, and design their strategies according to the particular needs of their targeted firm size.

buy-side makes its needs known and seeks to address them. Trading models also enable traders to better align their execution strategies against their goals. As algorithms become more sophisticated and more widely adopted, buy-side traders have an even more compelling reason for leveraging automated trading strategies. In satisfying perceived and actual buy-side needs for further algorithmic trade development, brokers need to look at the desires clients have for increased customisation, including the ability of algorithmic trade orders to react – appropriately – to changes in market conditions and updates. In addition, strategy security is a concern. If the perceived value of algorithms is derived from anonymity, then any sign that traders can ‘reverse-engineer’ these algorithms would endanger the entire field. A lesser concern is the potential of a firm’s proprietary desk to illegitimately access the electronic flow.

The needs of the sophisticated buy-side trader: As the buy-side continues to take a more directive approach to order flow, the format and direction of order flows will continue to shift in accordance to the speed and fluidity with which the

The increasing presence of the OMS: Also changing and impacting the trading environment and order flow is the relationship between clients, brokers and order management system vendors. One obstacle to the increased use of



■ THE TRADE 2005

11


■

“The buy-side continues to look for additional ways to more efficiently integrate algorithms into their processes, including linking them to their transaction cost analysis tools and their order management systems.”

12

algorithmic trade models is brokers’ use of tags as they continually search for a competitive edge. Tags, which are how different algorithms are electronically marked as they are sent through the electronic trade process, give brokers the ability to offer customisable algorithmic trades to clients. However, although the generic tag for, say, a VWAP or Market-On-Close algorithmic trade strategy may be similar across brokers, it is not yet possible to standardise the systems with the availability of algorithmic customisation. Buy-side and sell-side traders are in favour of standardising systems for consistency in parameters such as aggressiveness and time constraints. In these instances the tags differ from trade to trade across brokers, and brokers are looking to the OMS providers to ■ ALGORITHMIC TRADING


normalise the data or provide an internal matching system in order to make the trade process more efficient. The buy-side continues to look for additional ways to more efficiently integrate algorithms into their processes, including linking them to their transaction cost analysis tools and their order management systems. The recent acquisition of Macgregor by ITG illustrates the value of a more seamless integration of OMS and advanced trading tools. As traders seek a more holistic trade management system, TABB Group sees this integration becoming a more important driver of algorithms’ growth. However, there also is a danger in overloading the existing technology to the point where its proper functionality is hindered, creating inefficiencies that could lead to a loss of the ease of implementation and integration that initially drew a buy-side trading desk to use a particular OMS to begin with. As usage increases, buy-side firms are aiming to integrate algorithms more tightly into the trading process by linking their OMS directly to the algorithmic server. The advantages of tight integration are twofold. If traders enter orders through a separate webservice or desktop application, fills must be either manually ■ THE TRADE 2005


entered into the OMS, or imported into Excel and then uploaded to the OMS. In addition, thirdparty applications may not offer all the variables of a direct connection. Large firms, defined as those with over $50 billion in assets under management, are twice as likely to have OMS algorithm links as smaller firms, or those with fewer than $10 billion in assets under management. Tighter integration between the buy-side and sell-side trading platforms continues at breakneck speed. As the number of relationships decreases, the percentage of brokers connected to the investment manager’s order management system is rising. Indeed, connecting to the OMS is a requirement for doing business. For all the benefits of electronic trading, the downside for brokers is that the OMS/FIX infrastructure is the stepping-stone to alternative execution vendors and has been an integral cause of the liquidity shift. TABB Group has found that quantitative firms are deeply engaged in optimising the trading process, not surprising considering they traditionally have a broader knowledge base about how the algorithmic trade systems function, thereby increasing their usage comfort level significantly. Buy-side quant firms on average have five OMS algo■ ALGORITHMIC TRADING

rithm links, double the number of fundamental and mixed firms. The automation of almost every process is a key component to the quantitative business model. Algorithmic trade strategies In a short time period, the buyside has graduated from basic users of algorithms to fickle clients, growing more selective of the algorithms they deploy and even building strategies around them. The buy-side now is questioning with more frequency where and how algorithms can add value to the trade process. As more options are made available, such as increases in algorithmic trade options and crossing networks, the buy-side trader is maintaining growth in control over its order flow. Indeed, it is noteworthy that the buy-side is actually using a method to choose which algorithmic model to use for particular trades at this point, given the use of trial and error a year ago (see Exhibit 1 overleaf). TABB Group can cite several reasons for this progression. With some measurable time under their belts in using algorithms, traders are now better equipped to use historical information from preand post-trade analysis with enough confidence to develop trade methodologies based on past performance. It follows logically ■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

13


Exhibit 1: How algorithms are selected – two-year comparison Trader’s PM discretion

18% 16%

Order objectives Stock characteristics Market conditions

15% Analysis

17%

Trading strategy

17%

13% 11% 11%

Liquidity As any destination Flexibililty

57%

Experiment

TCA

9% 7%

Simple orders 9%

2005

2004 Response: 65%

14 Source: TABB Group study ‘Institutional Equity Trading 2005: A Buy-Side Perspective’

that the more the industry learns about different algorithms the more often we will see implementation of strategies incorporating algorithmic trade models. Secondly, as algorithm use grows more pervasive throughout the industry, the race to the top will be based on differentiation and ability to disguise intent to prevent gaming. This high level of competition raises the bar for all algorithm providers and can propel the ones that are first to offer this to a role of market leader. The number of strategies employed by a firm is a good proxy for its level of algorithmic ■ ALGORITHMIC TRADING


sophistication. Each strategy has its own pros and cons, and each one must be measured against its own benchmark. As firms become more sophisticated about algorithms, their demands for more flexible, customised products will increase. Quantitative shops and some large firms are increasingly building in-house technologies in an attempt to develop customised proprietary algorithms that are better suited to their direct needs, rather than relying on their brokers. However, most firms still are unable to break entirely free of their brokers for algorithmic trad■ THE TRADE 2005


ing. Until the majority of market participants develop a more complete understanding of and a higher comfort level with the use of algorithmic models, Volume Weighted Average Price (VWAP) executed trades will continue to be a staple in the algorithmic trader’s portfolio. It remains one of the simplest algorithms, and simultaneously the most used and the most hated of them. Fewer traders will be accepting of its ‘at the market’ benchmarking strategy and instead will seek algorithms that encourage swinging for the fences. As brokers compete for algorithmic order flow, new strategies are being created. Newer strategies, such as Guerilla and Liquidity Forecast will help attract the more aggressive traders. As more customisable and client-specific alternatives to VWAP algorithms – such as Arrival Price, which has 24% of the market share for algorithmic trade strategies and is a close second to the 27% market share held by VWAP trading – continue to gain ground in the algorithmic trading space, TABB Group expects VWAP to lose more of its lustre and become one among many strategies employed. Factors for growth Algorithms will grow at a 34% rate through 2007, intensifying the competition among providers, who ■ ALGORITHMIC TRADING

■

“As algorithm use grows more pervasive throughout the industry, the race to the top will be based on differentiation and ability to disguise intent to prevent gaming.”

are racing to offer more unique and differentiated models and strategies. The move to broker algorithms is based on four major factors: reduced market impact, increased trading efficiency, better alignment between strategy and execution, and lower cost (see Exhibit 2 overleaf). The perception is held, especially by large firms that are looking for ways to execute their large orders while minimising the impact they have on the marketplace, that using algorithms to break down large trade orders results in reduced market impact. As a result, firms with large order trades to settle look to algorithms to break apart the original trade into many smaller fills, or orders that are executed, thereby hitting the markets with many small orders that find liquidity in different places rather than filling one large order from only one source. Not only is the order more likely to be filled using this ■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

15


Exhibit 2: The advantages of algorithms Anonymity

18%

Lower cost of trading

16%

Automate easy orders

14%

Decrease market impact

14%

Ease of use

9%

Best execution

9%

Control

7%

Solves fragmentation

7%

Hit benchmarks

7%

Response: 65%

16 Source: TABB Group study ‘Institutional Equity Trading 2005: A Buy-Side Perspective’

method, but the impact as the smaller trades hit the market is decreased. Another reason behind the increase in the use of broker algorithms is the idea that they increase the efficiency of trading by automating the execution of less complicated orders and freeing traders to concentrate on their most difficult trades. In addition, the anonymity of algorithms offers a significant advantage over the human element and solves the buy-sides’ biggest complaint about brokers, which is information leakage. However, algorithms ■ ALGORITHMIC TRADING


are not a means by which to achieve alpha. Rather, they are tools to help increase efficiency, remain anonymous and trade effectively in an environment where large blocks of shares are unavailable. There also is the desire of the buy-side to achieve better alignment between strategy and execution, which it can achieve to some degree as brokers increasingly make their algorithmic trades more customisable in order for clients to put specific limitations or definitions in place for particular trades. Incredibly, algorithmic ■ THE TRADE 2005


offerings are so critical to firms’ market positions that the quantitative engineers who are developing algorithms are more highly cherished than many traders. Many firms also believe that the success of their low- and notouch offerings is critical to their business. Among the distinct advantages offered by broker algorithms is the ability to maintain a lower cost structure than a humanbased trading floor, as broker algorithms allow firms to retain overall market share while eliminating fixed costs. In addition, commissions will decrease with the increased use of algorithms since they offer a relatively lowcost means of efficient execution. Since broker algorithms are far less expensive and risky than developing and implementing proprietary ones, buy-side firms will continue to prefer that option. Hence, the increase in the use of broker algorithms is rising at a pace faster than that of proprietary algorithms – or even third party algorithmic providers – despite the buy-side’s discomfort with its reliance on brokers. In addition, it is easier for a broker to shut down an inefficient algorithm than it is to fire a person. Another advantage to using algorithmic strategies is their ■ ALGORITHMIC TRADING

ability to satisfy regulatory compliance. Due to the methodological documentation that occurs during the algorithmic trading process, this mechanism offers a regulatory solution through the trail of information that is a result of the transactions, unlike manual trades that are carried out on the floor. Hence, it offers the buy-side trader improved monitoring capabilities over the impact of their trades on the marketplace, as well as a superior view of the execution quality of their trades. In its short history, broker algorithm market share has been dominated by those with the quickest time to market. The ‘first-mover’ advantage is valuable, but as the algorithmic space becomes more crowded, the game is changing. Buy-side firms are now placing increased importance on how brokers package additional trading tools into the most compelling electronic trading value proposition. The brokerages with the keenest vision, the best tools, the most comprehensive support and the sharpest pencil will win. It will no longer be the case that the first to market will get the spoils. Instead, the winning firms will be those that effectively help their clients navigate the challenging markets, in the least complex manner and at the lowest possible cost. ■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

17


■

“Until the buy-side has the knowledge and the capital of the brokers in order to develop and maintain their own systems, there will always be a place for those brokers who stay on the leading edge of technology and provide the marketplace with more advanced, customisable and efficient algorithmic trade models.”

18 The future The buy-side will continue to grasp more control over its order flow as algorithms become more enhanced and readily available, even if it means routing their orders through broker algorithms in increasing numbers as they look for ways to reduce costs. This will be the case especially as brokers increasingly make attempts to supply the buy-side with updated and additionally customisable models with which to trade. The recent passage of Reg NMS also will have its own impact on the order flow as, with the increasing electronic capabilities of the marketplace, use of algorithmic and other electronic ■ ALGORITHMIC TRADING


forms of order routing will increase. As participants in the marketplace realise the market structure itself is not to blame for such market conditions as increased fragmentation, rather it is the increased use of electronic trading strategies and technologies to effectively hide and seek liquidity within the marketplace, traders will continue to adopt algorithmic trading models out of competitive necessity. Order flow will maintain its surge – not only in equities, but across other asset classes as well, as we see algorithmic trading penetrate further into the marketplace to incorporate foreign exchange, derivatives, and even fixed income, as players in these markets look for anonymous trading and search for hidden liquidity – providing vendors with opportunities to shine as increased demands for more sophisticated technology are made by brokers and the buy-side alike. The buy-side’s lack of trust for the brokers will not subside, even as they decrease the number of brokers with whom they do business and develop more integrated relationships with the ones they do maintain. However, until the buy-side has the knowledge and the capital of the brokers in order to develop and maintain their own systems, there will always be a place for ■ THE TRADE 2005


those brokers who stay on the leading edge of technology and provide the marketplace with more advanced, customisable and efficient algorithmic trade models. In addition, the order management system vendors, in order to remain competitive, need to unfailingly develop further ways for the brokers to integrate more effectively their particular trade methodologies at an always more efficient rate of trade processing. Transaction cost analysis will grow in use and importance as the marketplace will look to historical data to determine the most effective trade methods and to develop better ones. Also defining the usage patterns of algorithms will be the linkage between algorithms and direct market access (DMA) platforms as a holistic execution management system is sought. Since DMA platforms will likely be the vehicle that wraps and delivers all trading tools to the buy-side trader, algorithms will increasingly seek more effective associations with DMA providers. We are already seeing this become more evident in platforms such as Goldman Sachs’ REDIPlus® and Morgan Stanley’s Passport, which are being used to increase their own growth potential. Since algorithmic trading is rapidly gaining ■ ALGORITHMIC TRADING

■

“Defining the usage patterns of algorithms will be the linkage between algorithms and direct market access (DMA) platforms as a holistic execution management system is sought.”

traction, the importance of a link to DMA will become critical to DMA’s acceptance and growth, as buy-side firms strive for that competitive edge. Algorithmic trading is still young and in its developmental phases. As it grows and becomes more widely used, understood and accepted, the marketplace and the way it functions will transform into what we expect will be a more efficient and effective place to trade. ■


■ THE TRADE 2005

19


Understanding how algorithms work Where does time slicing and smart order routing end and randomising your orders through complex algorithms begin? Dr Tom Middleton* re we witnessing a revolution in A trading or are algorithms little more than a novelty that can be readily outperformed by the average human trader? To answer this question, we need to determine which algorithms are ‘smart’, what makes them smart, and how they can optimise the performance of a trading desk. Choices abound in algorithmic trading, even in its current nascent state. Any broker worth its salt now offers a diverse suite of algorithms and parameters, and third-party providers are now stepping in to offer various ‘customisable’ and niche products. With such a multitude of choices, it is important to differentiate between algorithmic trading engines that are essentially enhanced Direct Market Access (DMA), and algorithmic trading products based on robust statistical models of market microstructure, that have been found to increase trader productivity in both buyside and sell-side firms. ■ ALGORITHMIC TRADING

Enhanced DMA strategies (see Table 1, overleaf), such as pegging, ‘iceberging’, and smart order routing require minimal quantitative input, and the trader does not delegate any real decision-making to the algorithm. While these facilities add value to trading desk capabilities and performance, their behaviour is fairly easy to understand. Algorithms that require quantitative input are generally designed to minimise execution risk against a user-specified benchmark, typically Volume Weighted Average Price (VWAP), Time Weighted Average Price (TWAP), Implementation Shortfall (IS), Participate, or Market on Close (MOC). This chapter aims to shed some light on the relatively opaque world of these algorithms, clarifying the thought process of their designers and the inputs of the models. Risk and reward Central to any investment process is the trade off between risk and ■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

21

*Dr Tom Middleton, head of European Algorithmic Trading, Citigroup


Table 1: Examples of common algorithmic trading strategies, emphasising the difference between algorithms that require quantitative market microstructure modelling and simpler enhanced DMA strategies. Enhanced DMA strategies Iceberging

Pegging Smart order routing

Simple time slicing Simple Market on Close (MOC) Quantitative algorithms VWAP

22

TWAP Participate MOC Implementation Shortfall (aka Execution Shortfall or Arrival Price)

A large order can be partially hidden from other market participants by specifying a maximum number of shares to be shown. An order is sent out at the best bid (ask) if buying (selling) and if the price moves the order is modified accordingly. Mainly a US phenomenon – liquidity from many different sources is aggregated and orders are sent out to the destination offering the best price or liquidity. The order is split up and market orders are sent at regular time intervals. The order is sent into the closing auction.

Attempts to minimise tracking error while maximising performance versus the Volume Weighted Average Price traded in the market. Aims to match the Time Weighted Average Price. Similar to simple time slicing, but aims to minimise spread and impact costs. Also known as Inline, Follow, With Volume, POV. Aims to be a user-specified fraction of the volume traded in the market. Enhanced MOC strategy that optimises risk and impact, possibly starting trading before the closing auction. Manages the trade off between impact and risk to execute as close as possible to the mid-point when the order is entered.

reward. Naturally, this applies to algorithmic trading – a smart algorithm will maximise performance for a given level of tracking error against a chosen benchmark. Thus, the standard deviation of slippage against the benchmark is as important as the average when comparing the performance of a selection of algorithms. The choice of algorithm and provider depends on the individual user’s benchmark, style and urgency of trading. Just as in portfolio man■ ALGORITHMIC TRADING


agement, the appropriate place on the efficient trading frontier – how much risk a trader is prepared to take in exchange for improved performance – will depend on the nature of their alpha, how often they trade and against which benchmark their performance is evaluated. It makes sense that a small hedge fund with a short time horizon and high trading rate will tolerate higher risk on each individual execution in exchange for enhanced performance, than a fund that ■ THE TRADE 2005


trades much less frequently with a longer time horizon. Thus, balance between average execution performance and its variance is essential to the design of algorithmic trading models. Whether implicitly or explicitly, a well-designed strategy will optimise this trade off to deliver the best performance possible for a given level of risk. This methodology pervades every aspect of model development. Ever-smaller slices? The first component of an algorithm is the trading schedule – the rate at which the model aims to execute the order as a selection of smaller slices. It is difficult to pick up an article or attend a conference without seeing a graph illustrating the fall of the average trade size in recent years. Will we all be trading in one-lots in 10 years’ time? Clearly not – there must be a level at which transaction costs become prohibitive. Furthermore, every trade represents a piece of information leaked to the market, defeating one of the principal objectives of algorithmic trading in the first place. The trading schedule is essentially dictated by the benchmark describing the trading strategy – it is fairly obvious that a VWAP algorithm will execute volume according to a historical volume profile, possibly with some dynamic ■ ALGORITHMIC TRADING

adjustment; Participate will track current volume; Implementation Shortfall will involve some sort of impact/risk optimisation. Managing the trading schedule is particularly important in Participate algorithms. One could envisage a ‘not-very-smart’ algorithm, that when aiming to trade one third of the volume, fired out a market order for one share for every two that printed elsewhere. Not only would this generate a huge number of tickets but would be incredibly easy for a predatory proprietary algorithm to ‘sniff out’ and front run. As an order traded in this way would typically create a significant amount of impact, this would result in poor execution performance for the user. Thus, in designing a Participate algorithm, it is important to allow the user to specify the urgency of the order – a stop-loss 1/3 of volume order should be executed differently to an order where the trader is prepared to tolerate much more tracking error against volume in exchange for better performance through less impact and more intelligent, opportunistic type trading. Implementation Shortfall – the midpoint when the trade is started – is perhaps the most logical benchmark to use, but perhaps opinions about the optimal trading strategy are the most diverse. Some providers have apparently repackaged VWAP ■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

23


Figure 1: Example trading schedules for VWAP, a passive IS strategy and an active IS strategy. The IS profiles are calculated in ‘Volume Time’, taking into account both the need to unwind more rapidly than VWAP and the likely liquidity in the market. 0.12 VWAP

IS active

IS passive

0.10

0.08

0.06

0.04

0.02

8: 25 8: 50 9: 15 9: 40 10 :0 5 10 :3 0 10 :5 5 11 :2 0 11 :4 5 12 :1 0 12 :3 5 13 :0 13 0 :2 5 13 :5 0 14 :1 5 14 :4 0 15 :0 5 15 :3 0 15 :5 5 16 :2 0

0.00

8: 00

24

or Participate algorithms with a participation rate or end-time determined by a model rather than by the user, while others have built entirely different algorithms that react to price and liquidity in an opportunistic manner. The danger with opportunistic strategies is that one can end up cutting winners and letting losers run – buying falling stocks aggressively, while buying rising stocks more gradually. Ultimately, it ■ ALGORITHMIC TRADING


depends on the underlying investment style and the trader’s objectives which Implementation Shortfall strategy is the most appropriate. Ideally, an IS algorithm should obtain an end-time and trading schedule statically from an impact cost model, consistent with any pretrade employed: but still allow the user the flexibility to specify any dynamic constraints or parameters to take advantage of price or liquidi■ THE TRADE 2005


ty if appropriate. At Citigroup we use an impact cost model1 to determine an optimal end-time and optimal trading schedule based on the user’s risk aversion – typically an accelerated initial trading rate compared to a VWAP profile (see Fig. 1). However, this static strategy can be enhanced as required with parameters that constrain the order or allow it to take liquidity or price-related opportunities – if these modify the schedule we re-optimise the profile in real time. The trade schedule for an enhanced MOC is obtained in a similar way to Implementation Shortfall – but in reverse. The strategy with the least tracking risk against the close price places the whole order in the closing auction, where appropriate, while that with the minimum impact is a VWAP trade, starting as early as possible. Obviously, most users will choose a schedule intermediate between these two strategies, according to their own risk and impact tolerances. Managing each slice The second major component of an algorithm is how each slice of an order is managed. A dumb algorithm might send out market orders every time it wants to trade. This strategy would have very low tracking error against the target volume, but would pay the spread irrespective of whether it was necessary to do so. ■ ALGORITHMIC TRADING

The question that we are attempting to answer in a logical, quantitative and statistical fashion is what a successful human trader deals with instinctively hundreds of times a day, namely whether to (if buying): (i) Pay the offer. (ii) Wait on the bid side of the order book. (iii) Wait outside the order book and wait for a tight spread or a liquidity opportunity. In a risk-reward framework, options (ii) and (iii) both increase our execution risk relative to option (i). The reward that we expect for taking this risk is decreased spread and impact cost. In the formulation of our decision logic, the variables that we need to consider here are spread, volatility and liquidity on the electronic order book. The trading behaviour of diverse stocks across the world can be simplistically classified according to two simple ratios – the ratio of tick size2 to volatility, and the ratio of median3 bid/ask spread to volatility. The method used to estimate volatility is not important here, as we are simply interested in classifying the stocks according to their rankings by these two ratios. Essentially, there are three categories of stock, illustrated in Table 2 (overleaf). A well-researched trading model will adapt its behaviour according ■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

25 1 R. Almgren, C. Thum, E. Hauptmann and H. Li, Equity Market Impact, Risk, July 2005. 2 The tick size is the minimum price increment allowed, and so is also the minimum possible bid/ask spread during continuous trading 3 The median is a method of estimating the average by choosing the middle value of the sorted dataset. Its advantage over using the mean is that it is affected less by very large or small ‘outlying’ values.


Table 2: Three-way classification of stocks, with examples, according to spread and volatility behaviour. Tick size/volatility High Low Low

Median spread/volatility High High Low

to the type of stock that it is attempting to trade.

26

High tick size/volatility, high median spread/volatility This type of stock is well represented by Ericsson. Although it is reasonably volatile, this is swamped on short timescales by its enormous tick size – approximately 36 bps as I write. Stocks whose behaviour is constrained by the tick size in this way characteristically have large amounts of liquidity at the best bid and ask, relative to average trade size or average daily volume. Thus, the optimal spread capture strategy in this case is to be patient – waiting on the order book, sometimes for hours, in order to stand a chance of capturing spread. Algorithmic trading adds huge value for this class of stock, as a computer is able to constantly monitor liquidity on the order book, using the supply/demand imbalance as an indicator of future price movement. If an adverse move is predicted on this basis, then the model should pay the spread – otherwise patience is key. In terms of a slicing strategy, small is not beautiful for this type of stock. ■ ALGORITHMIC TRADING


Example Ericsson EMI Deutsche Telekom

As queuing time is generally long, it is important to get liquidity on the order book as soon as possible to stand a chance of trading on the right side of the spread. The large amounts of stock on the best bid and ask also mean that the effect of adding more liquidity is unlikely to significantly affect the supply/ demand imbalance, and so there is unlikely to be adverse price movement as a consequence of our adding to the order book. Low tick size/volatility, high median spread/volatility These stocks are typically mid-capitalisation, less liquid stocks with volatile spreads. Many providers do not recommend that this type of stock should be traded on the engines, owing to the difficulty of achieving high quality performance. The counter-argument to this is that these are also the stocks where information leakage can be the most damaging and so while trading them electronically can be problematic at least anonymity is maintained. The correct strategy here is to trade opportunistically and take advantage of spread and liquidity ■ THE TRADE 2005


opportunities within the blink of an eye. A tight spread with good liquidity on this type of stock suggests that the stock is fairly priced and the spread represents ‘good value’. Thus, the risk involved in waiting at the best bid or offer is unlikely to be rewarded sufficiently to justify passing on the opportunity of executing immediately on the other side of the spread. As time passes, a wider spread has to be accepted in order to keep up with the required trading rate. The taking of opportunities does not preclude adding liquidity to the order book – but care has to be taken to avoid leaking information to the market by creating a supply/demand imbalance that may create an adverse price movement. Low tick size/volatility, low median spread/volatility The final class of stock tend to be liquid, blue-chip stocks with low tick sizes and median spreads relative to volatility, that are efficiently priced and easy to trade by both human beings and algorithms – orders in these stocks tend to be referred to as ‘low-touch’ or commoditised. These orders should be sliced relatively finely to avoid excessive execution risk, and the waiting time on the order book is relatively short, as the reward for capturing the spread is relatively small compared to the volatility ■ ALGORITHMIC TRADING

risk to which the trader is exposed waiting on the bid or ask. Algorithms add value in the trading of these stocks as they allow traders to focus on more difficult trades in less efficiently priced stocks. Balancing risk and return Algorithmic trading models may remain somewhat opaque to many users – and many brokers are unwilling to disclose too much information about what is ‘under the hood’. Furthermore, the mathematics of how they operate is often rather intractable and poorly explained. I hope in this chapter I have managed to communicate at an intuitive level how a successful algorithm should help the user achieve the best possible outcome in executing their trades. Algorithms must be designed to manage the risk-return trade off in an optimal way, both in the way that trading is scheduled and the way in which each slice of an order is managed. While this should be apparent in performance statistics, algorithms should be transparent in the way that they are designed to balance risk and return so that the user can choose the appropriate strategy and provider that best suits their trading and investment styles. ■


■ THE TRADE 2005

27


Build or buy? What are the relative merits of broker-driven versus broker-neutral algorithms? Understanding the trade off between cost and performance Allen Zaydlin*

ince the introduction of elecS tronic trading, the landscape of trade execution, risk management and market access has undergone significant and rapid changes. Being interdependent in nature, the advancements in one area force other areas to evolve. To that extent the automation of electronic trade execution can be seen as the ‘first violin’, because the increasing demand in execution speed, throughput and low latency on one hand requires more efficient market access and risk management on the other. What are algorithms? Automation of trading processes can be grouped into two general categories – automation of trading (what to trade) and automation of execution decisions (how to trade). While the first one deals primarily with investment decisions, the second one focuses on their implementations. There is no ■ ALGORITHMIC TRADING

market consensus on the precise definition of each category, but for the purpose of this chapter we will define them as alpha models and execution algorithms or simply algorithms. Algorithms – when used appropriately – can improve multiple facets of the investment cycle. They give traders the ability to handle larger sets of securities. At the same time traders are able to focus their attention on the instruments that are difficult to trade, for instance illiquid stocks, while letting algorithms take care of more liquid names. Algorithms benefit from the speed of computers and therefore can follow the market more closely than traders. They also eliminate the emotional aspect of the trading process, resulting in more consistent performance. In addition, algorithms help reduce information leakage. All of the above can be summarised in a single statement: ■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

29

*Allen Zaydlin, CEO, InfoReach


■

“The increase in algorithmic offerings has made the process of selecting the right algorithm for your trading desk a more exacting task.” “Algorithms are a more efficient way to trade in an environment where cost, speed, consistency and the prevention of information leakage are crucial in attaining alpha”

30

Survival of the fittest In recent years, execution algorithms have become more prevalent. Initially, algorithms were offered by only a handful of brokers to the larger buy-side firms that were willing to pay higher commissions. Since then, however, algorithms have grown to become a common service available from the majority of brokers, as well as trading system vendors. Furthermore, as algorithmic execution becomes increasingly popular, more third-party software vendors are building trading platforms that serve as the foundation for the rapid development of custom algorithms. The increase in algorithmic offerings has made the process of selecting the right algorithm for your trading desk a more exacting task. Nevertheless, ■ ALGORITHMIC TRADING


the dilemma is quite simple to resolve and comes down to three options: 1. Utilise algorithms provided by brokers. 2. Utilise ‘broker-neutral’ algorithms provided by third-party vendors. 3. Develop proprietary algorithms. The first thing you will need to do in your quest to select an algorithmic provider is to think clearly and objectively. Simple analysis and adequate understanding of the goals that you need to reach, as well as the means available can yield surprisingly unexpected results. This will help you to avoid some common misconceptions about algorithms’ functionality and performance, and will also assist in separating the marketing hype from the algorithms. This will help you avoid the following scenario: Q: “Does your VWAP algorithm guarantee to beat the VWAP? No! The broker I am talking to guarantees their algorithm to beat VWAP every time. A: “Great! Then you should be able to send any buy and sell orders of the same symbol to this VWAP algorithm simultaneously and enjoy positive P&L every day. ■ THE TRADE 2005


Q: How good is your VWAP? A: What are the average sizes you need to trade? How liquid are your instruments? Q: Liquid, about 2000 shares. A: Rest assured our VWAP will do a good job. First of all, this thinking process should include a better understanding of how different algorithms function and the problems they solve. Secondly, try to decide which algorithm fits your trading cycle objective. And, lastly, pick providers that offer algorithms most efficiently. Comparing the options Confusion on behalf of buy-side portfolio managers and traders comes in part from the lack of consolidated and systematic analysis of algorithmic execution performance and cost ratios. It is also partly from the absence of a single scale that can be applied to compare different options. Execution performance attribution is also somewhat obscure and lacks standardisation. Nevertheless, a certain structured quantitative comparison of different options can be accomplished. Let’s start off by identifying the pros and cons of using broker-provided, ‘off-the-shelf ’ and custom-built algorithms. ■ ALGORITHMIC TRADING

■

“Confusion on behalf of buy-side portfolio managers and traders comes in part from the lack of consolidated and systematic analysis of algorithmic execution performance and cost ratios.”

Broker-provided Pros ■ Require minimum technological infrastructure on the client side to access execution models. ■ Provide a wider range of advanced algorithms that rely on significant research, infrastructure and maintenance cost. This includes quantitative study of historical data, computer hardware and network infrastructure to deal with vast calculation of considerable amounts of real-time market and execution data. There is also the ongoing expense in improving the performance of existing models. ■ Ability to pay only for usage. Cons ■ Higher commission rates. ■ Higher risk of information leakage. ■ Fewer algorithmic parameters are exposed to end-users. ■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

31


■

“The emergence of various ‘off-the-shelf’ algorithms serves as an indication of the level of commoditisation that has occurred with certain execution models and the depreciation in their relative value.” Off-the-shelf

32

Pros ■ Lower commission rate by taking advantage of DMA rates. ■ Tighter control over parameterisation of the algorithms. ■ Reduced risk of information leakage. ■ Ability to work with multiple brokers simultaneously, while keeping trading data consolidated within a single system. ■ Broker neutrality. ■ Anonymity. Cons ■ Require more infrastructure on a client side to run. ■ Increased financial commitment regardless of usage. ■ Tendency to lack more sophisticated algorithms that require a more elaborate infrastructure. ■ ALGORITHMIC TRADING


Custom-build Pros ■ Customised functionality not currently available from other sources. ■ Quicker ability to modify. ■ Tighter control. Cons ■ Bear unscalable expense of infrastructure, development and enhancement effort. ■ Risk of failure achieving the execution performance objectives. Broker-driven vs. off-the-shelf The emergence of various off-theshelf algorithms serves as an indication of the level of commoditisation that has occurred with certain execution models and the depreciation in their relative value. Such off-the-shelf algorithms will have a cost advantage in comparison to their broker-driven siblings. A simple comparison of VWAP algorithms provided by brokers and off-the-shelf providers will serve as a guideline on how to compare algorithms in general. Assuming that the ultimate goal is selecting an algorithm with maximum efficiency – meaning the algorithm that combines best execution performance with lowest cost – one can create a common scale that combines cost and performance into a single unit of mea■ THE TRADE 2005


sure. The cost basis is vital because if two different algorithms perform alike then cost becomes the main differentiating factor. At the same time we need to factor in performance, because poor execution performance can quickly eliminate any advantage of a lower cost algorithm. In determining the cost basis, it is useful to start with commission rates. From what we observe in the market today, it should not be unreasonable to accept $0.0075 per share as a common rate for brokerprovided algorithms. At the same time, brokers’ DMA can offer a very compelling commission rate, where $0.0015 is not unheard of. However, we also need to factor in the cost of the third-party vendor technology which delivers the broker-neutral algorithms that allow you to take advantage of low DMA rates in the first place. We will use an average of $10,000 per month for such technology and assume that it is a fixed cost that does not increase with trading volumes. Taking the example of a firm trading one million shares per month, we arrive at a commission cost of $7,500 per month when using a broker-provided algorithm, compared with $11,500 when using the broker-neutral equivalent. Clearly, it does not add up – particularly if information leakage is not a major concern for the firm. ■ ALGORITHMIC TRADING

■

“Assuming that the ultimate goal is selecting an algorithm with maximum efficiency – meaning the algorithm that combines best execution performance with lowest cost – one can create a common scale that combines cost and performance into a single unit of measure.”

This hypothesis changes however, once the number of shares traded via the algorithm increases to two million shares per month. Now we are looking at $15,000 with broker-driven versus $13,000 with broker-neutral algorithms. Thus, the breaking-point lies somewhere between one million and two million shares per month, or between 50,000 and 100,000 shares per day. It does not really matter exactly where the breaking-point lies, what is important is that in every particular case it can be calculated. Applying the same commissions’ rates for 20 million shares per month (one million per day) we will get over $100,000 of savings per month (over $1 million per year) – a level of cost saving that most firms would regard as signifi■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

33


■

“The cost advantage of an off-the-shelf algorithm over a broker-driven model will count for far less if it posts an inferior performance. A quick analysis of the numbers can help us determine where the breakingpoint is in terms of performance.”

34

cant. While for the purpose of this exercise the numbers have been rounded off, by using this approach and calculating your own numbers you can easily find the point at which the cost benefit of introducing broker-neutral algorithms is going to start making sense to your business. (Go to http://www.in4reach.com/ compare.html to run a quick comparison using your own numbers as input.) The chart opposite provides a good indication of the projected differences in execution costs between the two venues relevant to the number of shares traded. For the sake of objectivity, we also need to factor in the performance of the algorithms we are comparing. This is key. The cost advantage of an off-the-shelf algorithm over a broker-driven model ■ ALGORITHMIC TRADING


will count for far less if it posts an inferior performance. A quick analysis of the numbers can help us determine where the breakingpoint is in terms of performance. Using VWAP as the example once more, and assuming that a brokerdriven algorithm slipped 0.03 of a basis point on 10 million shares traded over a month, and using $30 per share as an average stock price, the value of slippage is $90,000. As outlined in the Chart 1, the cost advantage of the off-theshelf algorithms was around $50,000 based on 10 million shares traded per month. To have this advantage nullified due to the better performance of a broker-driven VWAP model, the off-the-shelf algorithm has to slip 0.047 of a basis point or, in terms of the math, under-perform by 156%. That said, in ‘real-world’ trading, it is well known that various VWAPs produce results only marginally different from one another. If this approach is calculated using different monthly trading volumes and average stock prices only, the core finding remains the same: that the performance ratio is driven solely by the commission rates ratio. In order to be on par with a broker-neutral algorithm, a broker-provided algorithm has to outperform it to a level that can compensate for the difference in their respective commission rates. ■ THE TRADE 2005


Chart 1: Commissions comparison 300,000 250,000

Commissions per month ($)

200,000 150,000 100,000 50,000 0

00 00

,0

00 ,0 40

35

,0

00

,0

00 30

,0

00

,0

00 25

,0

00

,0

00 ,0

00

,0

00 ,0 00 ,0 15

20

0 10 ,0 00 ,0 00

0, 00 5,

1,0

00

,0

00

00

-50,000

Shares per month Broker algorithms rate

It is simply incorrect to assume that something that costs twice as much performs twice as well. Most broker-provided algorithms reveal only a limited number of parameters for buy-side traders to manipulate. Commonly, they are: algorithm start and end times, aggressiveness, percentage of the volume, plus three or five others. In contrast, off-the-shelf algorithms provide significantly more options – in some cases as many as 30 para■ ALGORITHMIC TRADING

Broker DMA rate

Commissions difference

meters – fine-tuned for each individual order. Furthermore, most broker-provided algorithms are not geared to work with lists of orders that have correlations or share constrains. For example, if you need to trade a sector neutral basket of stocks there is no easy way to make broker-provided TWAP understand your criteria of sectors – which can be totally specific to you – and the threshold of intra-day imbalance between sectors. ■ A BUY-SIDE HANDBOOK

■ THE TRADE 2005

35


■

“The fact that more brokerdriven algorithms are being made available from third-party vendors, wrapped within a brokerneutral model, can be taken as a sign of increasing efficiency in the market place.”

36

Reducing information leakage is more a question of trust than an item that can be measured through quantitative analysis. However, one clear advantage broker-neutral algorithm will have is that by spreading your trades in smaller sizes across multiple DMA pipes you reduce the risk of information leakage regardless of your level of trust. Proprietary algorithms There are several reasons for building proprietary algorithms. First among these is control over the functionality and performance of the algorithm, followed by lower commission expense by making the algorithm broker-neutral. As outlined earlier in the chapter, some algorithms are fairly basic and require minimal infrastructure to run, while others call for realtime analysis of market depth and extensive statistical analysis of ■ ALGORITHMIC TRADING


historical data. The cost of developing more complex algorithms in-house in terms of the network and computer hardware that is required can be prohibitive. In some instances, it would be hard to justify such a high level of investment in infrastructure, since buyside demand for highly sophisticated algorithms varies depending on the trading strategy being deployed and the volumes involved. Meanwhile, the ongoing operating costs remain constant. It is advisable, therefore, that before deciding to build proprietary execution algorithms, the cost and performance of existing offerings is fully understood. Having considered the relative merits of execution algorithms – broker-driven, off-the-shelf and proprietary – there appears to be a breaking-point, both in terms of cost and performance, at which broker-neutral algorithms deliver clear advantages. The fact that more broker-driven algorithms are being made available from thirdparty vendors, wrapped within a broker-neutral model, can be taken as a sign of increasing efficiency in the market place. ■

■ THE TRADE 2005

A Buy-side Handbook

Recommend Documents