part 2:

Honing an algorithmic trading strategy

41 Chapter 4
Choosing the right algorithm for your trading strategy

51 Chapter 5
Anonymity and stealth

59 Chapter 6
Customising the broker’s algorithms
 ■ Chapter 4

Honing an algorithmic trading strategy

Choosing the right algorithm

for your trading strategy
What are the options buy-side traders need to consider in selecting an
algorithm best suited to a particular investment style?

Tracy Black* and Owain Self**

Ia nvestors trading styles and

benchmarks can be different for
number of reasons. Index track-
algorithms without going into the
complex mathematics behind
their construction.
ers and fund managers may have
specific price goals such as the Algorithmic choice 41
close price, whilst others may be When deciding if you can utilise
more constrained to the price at an algorithm to execute a particu-
which the investment decision lar order a number of questions
was made. Some investors may be need to be answered. Besides the
investing into 3-4 days volume of obvious question of ‘what is my
a stock and others in smaller benchmark?’ other factors will
more frequent trades. Reactions ultimately dictate whether algo-
to changes in prices and volumes rithmic trading is an option and,
over the course of the implemen- if so, what type of algorithm and
tation of their trade and, general- which parameters to apply.
ly, the degree of impact they are First, you need to assess
willing to have in order to com- whether the stock is suitable. Blue
plete, will differ from investment chip liquid names that trade a
case to investment case. Whether large percentage of their volume
or not an investor’s trading style on the order book will be good **Owain Self ,
executive director –
involves large slow money orders candidates. Small/mid cap stocks Equities,
or smaller high frequency trades, that trade a very small percentage UBS Investment Bank
there is a place for algorithms if on the order book are only suit- *Tracy Black,
used appropriately. In this chapter able with correct parameterisa- executive director –
European Sales
we will explore the attributes of tion. The reasons for this are obvi- Trading,
the more commonly available ous, a computer can only react to UBS Investment Bank

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 4

Honing an algorithmic trading strategy

■ “VWAP algorithms do not

generally take into account
absolute volume levels and will
using them will a client be able to
know which algorithms at which
firms best suit their trading style.

Undetermined benchmarks
still try to complete the order even Algorithms can generally be split
if this would cause additional into two types of benchmark, pre-
determined and undetermined.
market impact.” First generation algorithms try to
obtain a yet undetermined bench-
mark, such as VWAP, where the
information that is electronically benchmark will be determined
fed to it, it cannot participate in over the life of the order. The more
off market prints and it cannot recently developed
make phone calls to negotiate Implementation Shortfall algo-
block trades. rithms will be measured against a
Secondly, you have to decide benchmark predetermined at
42 what proportion of the order you order creation.
want to execute via an algorithm. VWAP (Volume Weighted
You might want to put your Average Price) has been the most
entire order into an algorithm if commonly used algorithm histor-
it suits your benchmark or you ically. As things have evolved,
may want to combine algorithms VWAP has gained its critics but it
with more of a traditional trading still has its uses. Ultimately used
service, such as Block Trading or with the aim to minimise market
DMA. impact, VWAP is useful for exe-
Making the process even more cuting trades where you don’t
complex for the client is the fact necessarily have a view on a stock
that not all algorithms are equal. and want to obtain a fair price by
Brokers have different names for sampling market levels over a
algorithms that have similar trad- specified period. Due to its sensi-
ing styles and sometimes two tivity to changes in volume distri-
firms will offer an algorithm bution, VWAP will participate
under the same name, which will relative to liquidity. However,
execute very differently. This cre- VWAP algorithms do not gener-
ates a minefield for the client. ally take into account absolute
Only through education from bro- volume levels and will still try to
kers on what to expect of their complete the order even if this
algorithms and through actually would cause additional market

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 4

Honing an algorithmic trading strategy

impact. It is therefore important Figure 1: GlaxoSmithKline PLC

that traders apply price and vol-
ume caps on larger trades to 8.00 16.35
2004/07/12 2004/07/12
minimise such an impact, 2004/10/1 2004/10/1

te
although this potentially means 2005/07/29 2005/07/29

Da
the order will not complete. 2,800,000 2,800,000
The expectation of VWAP is

Volume
2,400,000 2,400,000
not just about mean perfor-
2,000,000 2,000,000
mance; we know that if you exe-
1,600,000 1,600,000
cuted an order each and everyday
1,200,000 1,200,000
of the year in the same stock, the
800,000 800,000
mean performance would be
400,000 400,000
acceptable, however this is an 0 0
unlikely scenario. Therefore, the 8.00 16.35
Time
risk-adjusted performance and
the standard deviation of the
returns become important. Due
to VWAP’s sensitivity to volume Figure 2: LogicaCMG PLC 43
distributions, the performance
risk is affected by its ability to 8.00 16.35
2004/07/12 2004/07/12
predict changes in these distribu- 2005/01/11 20,050,111
2005/01/11
tions. The majority of trading
te

2005/07/25 2005/07/25
Da

engines are based on historical 800,000 800,000

data. This can often mean that
Volume

you have to choose your stocks

carefully. Some will have a rela-
tively stable historical trading 400,000 400,000
pattern – e.g. GlaxoSmithKline
(Fig. 1) – and therefore a more
predictable outcome. Others,
0 0
however, can have a more volatile 8.00 16.35
Time
trading pattern historically – e.g.
LogicaCMG (Fig. 2). In these
cases, using an average historical
curve will not deliver acceptable patterns based upon both histori-
performance, as your standard cal and real-time data analysis.
deviation would be too large. It is This results in an improved stan-
important that the algorithm you dard deviation of returns even
are using can predict trading when trading less suitable stocks.

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 4

Honing an algorithmic trading strategy

■ “Neither VWAP nor TWAP have

any macro level price
sensitivity – the overall profile of
stock had not been bought bid
side the algorithm would pay the
offer. Besides, the obvious foot-
print left behind by doing the
same trade in the same size
execution is not affected by repeatedly, paying the offer only
movements in the price of the because time dictates, will lead to
poor execution quality. Advanced
stock. Their aim is to get the order TWAP algorithms trade in a more
executed by the end time, sophisticated manner, deciding the
prices they pay in the market on
irrespective of price.” the basis of how much they are
ahead/behind and whether it is
the right price, whilst still trying
VWAP’s sensitivity to the vol- to achieve an even average.
ume distribution over the life of Important to note here is that
the order may not always suit neither VWAP nor TWAP have
44 your trading style. For example, if any macro level price sensitivity –
you need to execute an order over the overall profile of execution is
the remaining hour of the day. If not affected by movements in the
you use a VWAP algorithm to price of the stock. Their aim is to
trade, then based on the volume get the order executed by the end
distribution you could execute time, irrespective of price. This
25%-50% of your order in the macro price sensitivity needs to be
closing auction. This ties your added by way of price limits. The
execution price to the closing inbuilt price sensitivity of these
price and samples fewer market algorithms will be at a micro level
prices during continuous trading. – i.e. the part of the algorithm
If your aim is to minimise market that makes the decisions on the
impact by sampling over a period individual executions of the order.
of time but you have a lower sen- The price sensitivity here results
sitivity to changes in volume pro- in the algorithm deciding how
files, then TWAP (Time Weighted much risk it can take in order to
Average Price) can be a useful take advantage of favourable
algorithm. prices, the degree of sensitivity
Early versions of a TWAP algo- can often be set by a risk aversion
rithm simply split the order into or aggression level. This will dic-
portions of equal quantity and tate to the algorithm how far it
time. At the end of each portion, if can fall behind before needing to

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 4

Honing an algorithmic trading strategy

pay the offer and how far ahead it

can get when buying bid side.
Higher risk aversion generally
results in a tighter standard devia-
■ “Generally, Inline algorithms
do not have any macro price
sensitivity besides price limits.
tion of returns at the expense of
lower mean performance. The The micro sensitivity is the same
opposite is true with a lower risk as in VWAP/TWAP, deciding what
aversion level.
Inline/Percentage of Volume price to pay based on your risk
algorithms also target an undeter- (amount ahead/behind).”
mined benchmark. The aim is
purely to participate with market
volume at a rate specified by the you, you wouldn’t need to buy
user. The Inline algorithm is sen- 3,300 to catch up, you would need
sitive to absolute changes in vol- to buy 5,000. This is because you
ume levels. This results in it need to be 33% of the total vol-
actively participating when vol- ume of 15,000 once you have
ume trades and scaling back if traded. Participating at a rate of 45
volume does not permit. The level 33% means you have to trade
of participation sets the aggres- 50% of volume that trades away
sion level of this strategy – i.e. and this is amplified as the target
how much impact you are willing percentages get higher. For exam-
to have in order to get the trade ple, take two buyers at a rate of
completed more quickly. An 33%, combined they would need
investor who wants to trade if liq- to be 200% of any volume missed.
uidity permits but does not want Due to this compounding
to have significant impact will nature of Inline algorithms, stocks
choose a low level, e.g. 5%. are often seen spiralling out of
Someone who wants to execute control, with algorithms partici-
more quickly at the cost of added pating at unfavourable levels.
impact will choose an aggressive Generally, Inline algorithms do
level, e.g. 33%. One major con- not have any macro price sensitiv-
cern for this style of execution is ity besides price limits. The micro
that 33% has become the market sensitivity is the same as in
default, and the market impact of VWAP/TWAP, deciding what price
such a high percentage is often to pay based on your risk (amount
underestimated. For example, ahead/behind). Efficient Inline
when buying at a rate of 33% and algorithms will manage risk in
10,000 shares trade away from terms of liquidity and not simply

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 4

Honing an algorithmic trading strategy

■ “If you are executing a high

frequency of smaller orders,
the generation of which are often
Implementation Shortfall, and
ones with high macro price sensi-
tivity, which we will call Arrival
Price.
Implementation Shortfall style
triggered by the price of the stock, algorithms are designed to min-
Implementation Shortfall imise the average shortfall over a
number of trades. This shortfall is
algorithms are the most suitable.” measured as the difference between
the execution price and the price at
initiation. To minimise this we
in terms of the actual percentage need to find the optimal level
it has traded. For example, it is between how much we move the
better to target 30% +/- a normal price (market impact) and how
trading size in the stock versus long we work the order (risk). We
being 30% +/-5%. The latter will know that if we didn’t execute in
result in inconsistent risk the market we wouldn’t have mar-
46 throughout the life of the order. ket impact but we are exposed to
There will be little risk at the movements in the stock.
beginning and therefore no Conversely, if we bought the entire
favourable prices, but high risk amount immediately, we would no
towards the end of the order. longer be exposed to future move-
ments but could have extremely
Predetermined benchmarks large market impact.
Predetermined benchmarks have In order to optimise the execu-
been a more recent trend, using, tion, the algorithm will determine
for example, the mid price at initi- when and how much to trade by
ation or the previous nights close. taking into account a number of
The algorithms to use in these situ- factors, primarily the size of the
ations are Implementation order, the stock’s liquidity, volatility
Shortfall and Arrival Price style and the time remaining. Generally
algorithms (unfortunately, these this will involve being more active
names are sometimes used to in the market initially, as the stock
describe the same or different types price will be at your benchmark
of algorithm). What we can do is and becoming less active as a high-
split these into two distinct trading er proportion of your order is
styles, ones with low macro price completed. This is not a new con-
sensitivity, which for the purpose cept; clients have been using this
of this chapter we will call trading style for many years.

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 4

Honing an algorithmic trading strategy

Typically a percentage of the order

was traded on risk to start, then
worked relatively aggressively in
the market. Once the majority had
■ “If a stock is trading on the
favourable side of your
benchmark they [Arrival Price
been executed more passive execu-
tion would follow. This ultimately algorithms] become very
is what the Implementation aggressive until the order is
Shortfall algorithms do, they
decide how aggressively to trade complete. If the stock is moving
based on how much risk they are away from the benchmark they
offsetting. If the stock had low
volatility then this reduces how far become much more passive.”
it is expected to move so you can
afford to be less aggressive and
have less impact. However, with a these can sub-optimise the strategy.
volatile stock you can afford more A volume cap might mean you are
impact to reduce the larger risk. unable to be as aggressive at the
For the algorithm to optimise this beginning of the order when the 47
trade off it needs to know the full stock is at the initiation level. A
extent of the order, otherwise it crucial parameter is some kind of
miscalculates how much risk it is risk aversion (aggression) setting.
actually offsetting. Algorithms have been built with a
One thing to note about this risk level in mind. However, your
type of algorithm is that it will not appetite might be very different.
necessarily be more aggressive You may believe you have more
below the initiation level. alpha and therefore are willing to
Supporting a stock at a certain take more impact in order to get
level in the market does not min- the trade executed quickly and
imise market impact. Market reduce risk; in this case you should
impact isn’t just judged as the choose a higher risk aversion level.
amount you move a stock against Arrival Price algorithms are the
you, it is also a measure of how other style of algorithm in this
much you restrict the stock from space. These work similarly to that
moving for you. of Implementation Shortfall, but
When using an Implementation have inbuilt macro price sensitivity
Shortfall algorithm, you need the to your benchmark (usually mid at
ability to set macro sensitivities, initiation or a set level such as pre-
such as volume and price caps, but vious close). If a stock is trading
you also need to understand that on the favourable side of your

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 4

Honing an algorithmic trading strategy

according to the price of the stock

Figure 3
will deliver similar results.

Order execution
Implementation
The proportion of your order exe-
Probability density

Arrival
Price Shortfall cuted via an algorithm is also
important. If you are executing a
high frequency of smaller orders,
the generation of which are often
triggered by the price of the stock,
Implementation Shortfall algo-
rithms are the most suitable. Given
Price achieved the entirety of the order the algo-
rithm can work out how best to
optimise execution and minimise
benchmark they become very the shortfall on average. However,
aggressive until the order is com- if you had a large order that was
48 plete. If the stock is moving away also measured relative to the price
from the benchmark they become at which the investment decision
much more passive. was made, a standard
This results in a different distri- Implementation Shortfall algo-
bution of returns to that given by rithm may not be suitable; one rea-
the Implementation Shortfall algo- son being optimal execution may
rithm (Fig. 3). In Implementation take several days. The algorithm
Shortfall we get a relatively sym- will need to know completion is
metrical distribution of returns. not required by the end of day one
However, with Arrival Price we get and each day following, it would
a skewed distribution. This is need to know all the details of the
because we often complete our previous algorithms.
order before we get the chance to In this scenario we don’t neces-
participate at more favourable sarily have to rule out algorithmic
prices and as we slow down when execution, it just means more con-
the stock moves away we poten- trol will need to be taken. You can
tially participate at very still utilise a combination of algo-
unfavourable levels. This skewed rithms to achieve the desired
distribution of returns can be seen results. Many people will start
in any algorithm with macro price trading aggressively with a small
sensitivity. Inline algorithms that part of their order using an Arrival
vary their participation rate Price or Inline algorithm. As and

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 4

Honing an algorithmic trading strategy

when they have completed the

majority of the order or the stock
has moved significantly, they will
start to use the more passive algo-
■ “It is becoming more common
to see algorithms that are
sensitive to additional factors
rithms such as VWAP/TWAP for
remaining portions of the order; such as momentum indicators,
ultimately replicating the trading mean reversion and relative
pattern of the Implementation
Shortfall strategy. performance.”
The point to note here is that
an algorithm is not the be all and
end all for a particular order. For better mean but at the expense of a
example, if you had an order that higher standard deviation, and a
was benchmarked against VWAP higher risk aversion will tighten
and you felt you could add value deviation but at the expense of the
to the execution, you may put half mean.
of the order into an algorithm and The depths to which sensitivi-
choose levels to execute the ties to external factors can be 49
remaining via other algorithms, a introduced are endless. It is
block desk or DMA. becoming more common to see
In summary, when assessing algorithms that are sensitive to
which algorithm to use you have to additional factors such as momen-
take into account which type best tum indicators, mean reversion
suits your benchmark. Then you and relative performance. In order
need to decide on your sensitivity to gain access to the best possible
to changes in price and volume algorithms to suit their trading
levels. Algorithms with low sensi- styles, investors will need to retain
tivity to price movements will give close relationships with sell-side
you a more symmetrical distribu- brokers who can deliver customis-
tion in returns for both rising and able algorithmic solutions as trad-
falling markets, highly sensitive ing styles evolve. ■
ones will give you a skewed distrib- © UBS 2005. All rights reserved.
ution. Sensitivity to volume
changes will ultimately result in
you trading with the crowd and
less independently. Additionally,
you need to decide on your
appetite for risk. Algorithms with
lower risk aversion will give you a

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 5

Honing an algorithmic trading strategy

Anonymity and
stealth

What assurances can the sell-side offer to safeguard the client’s alpha
capture and minimise information leakage?

Richard Balarkas*

Fareirst let’s be clear what we mean

by anonymity and stealth as they
two quite different things:
who is behind the order, is itself
capable of moving the price.
Stealth is the act of completing
any order in a manner which 51
Anonymity – Refers to the reveals as little as possible to the
expectation that information wider market in the hope of min-
relating to the identity of the client, imising impact.
information which the client must
of necessity give to the broker, is Are anonymity and stealth
not divulged in the trading process important?
or at any subsequent point post- In 1997 the silver market was in
execution. the doldrums. From July ’97 until
early ’98 its price rose 25% (at
Stealth – Refers to the act of one point it was up 50%). In
moving, proceeding, or acting in February ’98, Berkshire
a covert way. In conflict and Hathaway, the investment vehicle
game play this denotes achieving of Warren Buffet, famous for
ones objective without being building large if not controlling
detected, uncovering what others stakes in corporate stocks with
are attempting to conceal or long term value and a strong
obscure, and otherwise avoiding brand image, announced that it
conflict. had been buying silver. (Buffet
Putting it another way, anonymity already had a 32-year investment
is important in instances where history during which time *Richard Balarkas,
global head of
not the order, but the identity of Berkshire Hathaway had risen by AES™ Sales, CSFB

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 5

Honing an algorithmic trading strategy

■ “Anonymity isn’t restricted to

active managers. Passive
managers also need to take care
next move in order to beat the
market to the punch. Any public
statement on his next hunch
would be a self-fulfilling prophecy
as the investment herd try to
that repeated habitual portfolio anticipate his move. It is not sur-
slices are not sending signals that prising that the last place to look
for his ideas is the Berkshire
others can learn to anticipate.” Hathaway home page.
Buffet appreciates the value
that can be lost through informa-
an average 33% each year.) With tion leakage. So did we at CSFB
regard to silver he had the same when we constructed our
information as everyone else, Advanced Execution Services
essentially that demand was run- (AES™) algorithmic trading
ning ahead of supply and capability. Like many of the fea-
appeared likely to continue for tures that are at the heart of our
52 the foreseeable future. Over a algorithmic trading service, the
seven-month period he bought principles of protecting client
silver through a single broker anonymity and stealth trading
without taking any position in were already embodied on our
futures or options. He amassed trading floor and in our trading
what amounted to more than practices, algorithms simply gave
25% of the world’s supply. Years us a new medium in which to
later commentators were still take anonymity and stealth to the
debating whether he had actually next level.
taken delivery of the silver, or still
owned it, or leased it out… Valuing anonymity
Buffet understands the value of From a user perspective, the
anonymity and stealth. process of selecting whose algo-
He needs to. Regarded by many rithms to use should be based pri-
as the oracle of the investment marily on performance. Piles of
world, Buffet’s every move is colourful marketing literature may
watched via scores of internet sites give some vague insight into how
selling Buffet-related books and different broker services are con-
software, hosting chat room structed and delivered, and the
threads, Buffet-dedicated sites, fan similarities that are present in
clubs etc. There is a whole indus- high-level marketing descriptions
try out there trying to guess his of tactic objectives may create the

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 5

Honing an algorithmic trading strategy

impression that all broker algo-

rithms are much the same thing Anonymity
and achieve very similar results.
This is far from true. The con-
struction of trading algorithms
Early attempts at
and their further refinement anonymous trading
through practical use is a highly
quantitative process.
were not entirely
Clients clearly believe successful
anonymity and stealth are
extremely important, and in
seeking to continuously improve
the performance of CSFB’s algo- move in anticipation. And the
rithms it would be ideal to disag- benefit of trading with full
gregate the performance in order anonymity isn’t restricted to
to focus on those components active managers. Passive managers
where the potential value-add is also need to take care that repeat-
the highest. However, the reten- ed habitual portfolio slices are 53
tion of alpha gained through not sending signals that others
anonymity and stealth are hard can learn to anticipate.
components for a broker to
measure. Science of stealth
The benefits of anonymity Whereas anonymity has been
will be readily understood by enshrined in CSFB’s AES™ ser-
buy-side traders, many of whom vice from the start, stealth tactics
are regularly handling orders can always be improved and is
that are on average multiples of the area our AES™ developers
ADV where revealing the size find the most exciting. Many
alone would be sufficient to beginners think that playing
move the price. poker online will prove to be
However, knowing who is completely different than playing
behind a trade has additional offline and they are sometimes
informational value. The better a

■
money manager is perceived to be
at stock selection and timing, the “The retention of alpha gained
greater the informational compo- through anonymity and
nent of the trade, and the greater
the likelihood that if this infor- stealth are hard components for a
mation leaks out the market will broker to measure.”
■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005
 ■ Chapter 5

Honing an algorithmic trading strategy

■ “Even if anonymity is assured,

an algorithm will not perform
well unless it uses stealth in order
can, when necessary, pay inside
the spread or even cross the
spread without being so aggres-
sive as to send out signals, keep-
ing trades ‘information-less’. The
to take advantage of other traders winner is the one who can spot
and other less sophisticated reversion, whose participation is
overweight on the dips when buy-
algorithms.” ing and on the highs when selling.
So even if anonymity is
assured, an algorithm will not per-
correct, though usually for the form well unless it uses stealth in
wrong reasons. There is a com- order to take advantage of other
mon but misguided notion that it traders and other less sophisticat-
is much harder to ‘read’ your ed algorithms. The use of stealth is
opponents when you do not see also defensive, as there are plenty
them sitting at the table. What of trading models out there
54 most fail to recognise is that the designed to make money from
majority of available information reading signals generated by less
when making a decision comes sophisticated traders and black
from a variety of factors other boxes. At the market micro-struc-
than ‘reading’ your opponents ture level stealth is important, and
faces. Most of the required infor- CSFB’s AES™ incorporates
mation comes from patterns, advanced probability and game
position at the table and the theory tactics in order to outwit
hands your opponents play. the opposition. ■
It’s the same when you are
trading on a public limit order
book – you cannot see your oppo-
nents but if you can read their
signals they may, often unwitting-
ly, reveal their intentions. At the
same time you must be careful
that none of your actions are giv-
ing your game away. The winner
is the one who can coax traders
on the other side of the touch to
cross the spread and pay the pre-
mium. The winner is the one who

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 5

Honing an algorithmic trading strategy

Game theory
Although game theory has been studied since the 1940s, it has only recently been
applied to the world of finance. Game theory champions garnered the 1994 Nobel
Prize in Economics, and, today, this theory is used to analyse everything from the
baseball strike to auctions. Increasingly, game theory is making its mark as a potent
tool for traders.
In simple terms, game theory is the study of conflict based on a formal approach
to decision-making that views decisions as choices made in a game. Whether
playing individually or in a group, each player in a conflict has more than one course
of action available to him, and the outcome of the ‘game’ depends on the interaction
of the strategies pursued by each party. Algorithms can take advantage of the fact
that game theory and probability often have the edge over human intuition. To
illustrate this, here are some problems where the answer does not appear to be
intuitive, and in one case is actually counter-intuitive (for answers and explanations,
see pages 56 and 57):

Problems 55
Example 1.
If you throw a die until the running total exceeds 12, what is the most likely final
total?

Example 2.
This is a demonstration of the power of faith in random decision-making over simple
logic and probability. It was inspired by the format of an old USA TV gameshow ‘Let’s
Make A Deal’, hosted by Monty Hall.
The conundrum is that you are on a game show and given the choice of three
doors: Behind one door is £1million, behind the others nothing. You are invited to
pick a door. The host, who knows what’s behind the doors, opens one of the two
remaining doors to reveal there is nothing behind it. He then invites you to pick again
between the two remaining doors. Is it to your advantage to switch your choice?

Example 3.
You are in a game of Russian roulette, but this time the gun (a six-shooter revolver)
has three bullets in sequence in three of the chambers. The barrel is spun only once.
The two players then take it in turn to pull the trigger. If they live, the gun is passed
to the other player who then pulls the trigger, etc. Would you rather be first or
second to shoot? Continues overleaf ➧

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 5

Honing an algorithmic trading strategy

Game theory (continued)

Answers & explanations

Example 1.
Answer: 13
The way to get a final total of 13 is to build up some total between 7 and 12
inclusive, then make a single throw of the appropriate value.
The way to get a final total of 14 is to build up some total between 8 and 12
inclusive, then make a single throw of the appropriate value.
Thus if we take the list of sequences producing 14, then subtract 1 from the final
throw of each sequence, we will have part but not all of the list of sequences
producing 13. Moreover, corresponding sequences are equally likely to occur,
because they contain the same number of throws. Thus 13 is strictly more likely than
14. A similar argument shows that 14 is strictly more likely than 15, and so on. Hence
13 is the most likely total

56 Example 2.
Answer: You should change your choice.
The problem is called ‘counter-intuitive’, because the answer seems for many to defy
instinct and logic, even after it’s been explained several times. Most contestants on
Monty Hall’s show were apparently reluctant to change their original choice for fear
that it was right, or because intuitively they felt that probability could not be altered
by revealing one of the ‘losing’ doors.
The door you originally chose was a 1-in-3 chance – i.e., the likelihood of your
guessing the winning door was 1-in-3. The ‘other’ door is now a 1-in-2 chance, and
the likelihood of your guessing the ‘other’ door to be the winning door is 1-in-2. You
are 50% more likely to correctly guess a 1-in-2 chance than a 1-in-3 chance, so pick
the other door in preference to your original choice of door.
If you’re still in doubt, imagine there are 20 doors – one has the money, the
others nothing. You pick a door. Then 18 doors are opened revealing nothing, leaving
your choice and the one other door. Would you change your choice now? By
switching doors you’d improve your chances from 1-in-20, to 50:50 evens, or
(depending on how you look at it) arguably 19-in-20. Still sceptical? How about 100
doors? Pick a door. Open 98 revealing nothing, leaving two doors, one a winner and
the other a loser. Would you still prefer your original 99-to-1 shot compared to the
alternative that is at worst 50:50, and arguably a massive 99% chance?

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 5

Honing an algorithmic trading strategy

Game theory
Answers & explanations

Example 3.
Answer: Player 2 is preferable.
All you need to consider are the six possible bullet configurations:

BBBEEE ➠ player 1 dies

EBBBEE ➠ player 2 dies
EEBBBE ➠ player 1 dies
EEEBBB ➠ player 2 dies
BEEEBB ➠ player 1 dies
BBEEEB ➠ player 1 dies

57

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 6

Honing an algorithmic trading strategy

Customising the broker’s

algorithms

How much flexibility does the buy-side trader require to adjust and
fine-tune the broker’s algorithmic models?

Richard Balarkas*

A s conference organisers try to

squeeze more and more ‘algo’
food hamburger outlet. Not so – if
you want your algorithm on organ-
events into an increasingly crowd- ic bread with the gherkin removed 59
ed space, the topics to be discussed from the pickle, your initials
by their expert panels seem spelled out in caviar on top with
increasingly innovative. So whilst strips of spring onion laying strict-
the majority of money managers ly north to south (Tuesday and
have yet to enjoy their first algo- Fridays only) – it would be our
rithm experience, many conference pleasure.
organisers are already filling their It is important to recognise that
bills with debates headlined – the term ‘algorithm’ has, unfortu-
‘Algorithms: does one size fit all?’ – nately, been stretched to include
‘Algorithms: is customisation the not only the most complex mathe-
future?’ – and, ‘Algorithms: canned matical trading models but also
or customised?’ very mechanical and simplistic
What I want to do in this short trading techniques such as ‘ice-
chapter is explain how customisa- berging’ (the simple drip feeding of
tion is not a future trend, but a fea- an order into the market in pre-
ture that has been around since day defined clip sizes). In some cases
one. At the same time I want to even ‘stop loss’ orders have been
show how it is incorrect to cate- defined as algorithmic tactics.
gorise all broker-provided algo- There is nothing disreputable
rithms as ‘canned’, as if they were about this – good results can be *Richard Balarkas,
global head of AES™
the trading equivalent of a fast achieved using these tactics if you Sales, CSFB

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 6

Honing an algorithmic trading strategy

■ “The ability to tweak the

parameters means there is
significant scope for customising
after all, efficiency tools. Over and
above deciding which order is suit-
able for trading through which
strategy and at what point in time
to execute, many traders do not
each algorithm on an order by necessarily want to have to consid-
er too many other factors – it may
order basis.” be counter productive. As a result,
we aim to ensure that our algo-
rithms are optimised to deliver the
pick the right spots, but it is best performance without any
important to recognise that some additional input from the end user.
algorithms are more… well, ‘algo- For many traders this approach
rithmic’ than others. For example, works perfectly well.
those designed to anticipate vol-
ume curves, react dynamically to Customising to order
complex signals, and trade with Perhaps the first step towards cus-
60 stealth to minimise impact are far tomisation happens when a trader
more advanced than their more decides to adjust one of the para-
mechanical stable mates and, as a meters available with each tactic –
result, are offered by fewer brokers. typically, start and end times,
It is worthwhile making this dis- price limit, aggression level, min
tinction between the more complex and max percentage volumes. The
algorithms and simplistic mechani- ability to tweak the parameters
cal options. ‘Customising’ the latter means there is significant scope
is not particularly challenging, and for customising each algorithm
it is understandable that users on an order by order basis, even
might perceive such tactics as all to the extent that different tactics
being the same regardless of the can be forced to perform like oth-
provider. ers, or combinations of others.
Leading on from the above, if For example:
the serious algorithms are such
complex beasts in which highly ■ A trader who wanted to trade
qualified knowledge engineers at ‘volume in line 20%’ but didn’t
the major brokerages have embod- want the tactic to rigidly stick to
ied the firms trading skills, the first the 20% target irrespective of
issue worth exploring is why a price opportunities, might
client might wish to customise an instead use ‘price in line with a
algorithm at all. Algorithms are, 15% min and 25% max’ – which

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005

 ■ Chapter 6

Honing an algorithmic trading strategy

would then aim to be on average start of the trading period and

20% participation, but could less aggressive towards the end.
speed up or slow down within
the 15-25% range to respond to Beyond such requests, clients also
pricing fluctuations. approach CSFB with ideas for
custom strategies, usually varia-
The further step towards cus- tions on the menu tactics we pro-
tomisation is when a trader finds vide, examples of which it would
that his personal preference is be inappropriate for us to reveal
leading him to consistently use as they offer real competitive
the same tactic with the same advantage to the client. These
parameter adjustments for cer- ideas represent the desire of
tain sectors or markets. In this traders to further automate their
instance a request can be made to own trading style; in effect when
adjust the default settings for that they come to us with these
client so that the revised parame- requests they have developed their
ters are always used. The revised own strategy and are simply ask-
tactic can be re-named if the ing CSFB to put them into prac- 61
client also wishes to continue tice. For example:
using the default version.
Examples of this are: ■ When the stock gaps I like to…

■ A trader who always wants to ■ Cross asset correlations – when

finish by 2pm GMT ahead of trading mining stocks I want
US opening. All selected tactics participation curves that
are defaulted to finish at that respond dynamically to com-
time. modity prices…

■ A trader in French mid-caps ■ I have buy and sell baskets, I

who is less interested in poten- want to maintain any natural
tial impact and more interested hedges as it progresses and keep
in grabbing available liquidity both sides dollar neutral…

■
might ask for the TEX strategy
to default to very aggressive
mode for these stocks.
“Customisation has been
■ A trader who trades VWAP in around from the start. Indeed,
the morning period but wants
the curve skewed to be more
it is hard to see how the product
aggressive/overweight at the could have worked had it not.”
■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005
 ■ Chapter 6

Honing an algorithmic trading strategy

One other area of customised algorithms, the buy-side needs to

automation offered on AES™ is assess the merits or otherwise of
our ‘Storyboard’ product. insourcing versus outsourcing.
Storyboard automatically sends Hopefully, this chapter gives those
clients messages triggered by who have yet to adopt algorithms a
price movements, news, volume better understanding of the current
spikes etc. in the stocks they are scope of customisation and flexi-
trading in AES. Here again, all the bility that is already available. ■
trigger limits are configurable by
the client.

FAQ
Customisation has been around

■
from the start. Indeed, it is hard to
see how the product could have “There is a view,
worked had it not. There is a view,
inaccurate in our opinion, that bro- inaccurate in our
62 ker algorithms are canned and opinion, that broker
therefore inflexible. Hopefully, the
examples that have been outlined algorithms are canned
prove otherwise. There are also and therefore
views expressed that all broker algo-
rithms deliver the same perfor- inflexible.”
mance, that they are ‘commoditised’.
We have not been presented with
evidence that shows this to be the
case, but it is understandable how
this viewpoint might add weight to
the argument that the only valuable
algorithm is a customised algo-
rithm. In our experience, even using
the ‘plain vanilla’ versions of our
algorithms, different clients achieve
different results – from good to
excellent!
As with all aspects of the buy-
side/sell-side relationship, be it
research, trading or the develop-
ment and use of trading

■ ALGORITHMIC TRADING ■ A BUY-SIDE HANDBOOK ■ THE TRADE 2005