Professional Documents
Culture Documents
97 Chapter 10
Basket algorithms – The next generation
107 Chapter 11
The future of algorithmic trading
■ Chapter 10
Basket algorithms –
The next generation
Implementation Shortfall strategies for baskets of stocks
more than half of the original posi- for the second strategy
tion is still exposed to market U = 40 + 0.5 70 = 75 bps,
■
volatility. The timing risk (under- and for the third strategy
stood as standard deviation of the U= 48 + 0.5 60 = 78 bps.
■
average trade price) depends pri- The results are close, but the penalty
marily on the volatility of the stock is the smallest for the second strategy.
and is thus easy to estimate. If daily Thus, for = 0.5, the second strategy
volatility for XYZ is 170 bps, then is preferable. The VWAP strategy is
the timing risk for a VWAP strategy too passive, while the third strategy is
will be around 100 bps. overly aggressive. Note that for =
Is VWAP the optimal strategy if 1.5 the third, more aggressive, strate-
the benchmark is Arrival Price? It gy is preferable due to a higher risk-
may be, but only for a trader who is aversion level.
not worried about market risk. A In their work, Almgren and
risk-averse trader may prefer a more Chriss outline how to build a
aggressive strategy, which starts to unique optimal trading strategy for
trade at a higher participation rate each level of risk-aversion; the set 99
(e.g. 20% in the morning), reducing of these optimal strategies defines
the position faster. This strategy may an efficient trading frontier. Most
cause 40 bps of market impact, but modern pre-trade tools are capable
reduce timing risk to 70 bps. An even of calculating these optimal strate-
more aggressive strategy will lead to gies and assessing the difficulty of a
the estimates of 48 bps of impact and trade schedule. However, these cal-
60 bps of risk. Which is preferable? culations are of limited value when
The correct way to approach this it comes to designing and imple-
question is to construct a ‘utility menting a real-world trading strat-
function’ U = I + R, where I is the egy. There are many high frequency
expected impact, R is the timing risk variables in the data which are not
of the strategy and is a risk-aver- going to conform to their historic,
sion coefficient which reflects the pre-trade estimates. Factors such as
trader’s preferences. Since the opti- intra-day volatility, spread and pre-
mal trading strategy seeks to min- dicted volume may vary with each
imise this function, it may be more tick; hence the static execution
appropriate to refer to it as a ‘penalty schedule will no longer be optimal.
function’. Take = 0.5 and compute A workable high frequency
the penalty for each of the three sam- trading strategy necessitates a more
ple strategies (see Fig. 1): for the flexible application of ETF con-
VWAP strategy cepts. Rather than calculate a static
VWAP
Medium
Aggressive
Remaining shares
100
30
0
:0
:3
:0
:3
:0
:3
:0
:3
:0
:3
:0
:3
:0
9:
10
11
12
13
14
15
10
11
12
13
14
15
16
egy delivers better execution qual- involved. There are two possible
ity by reacting to changing market approaches here, although both
conditions. have their limitations.
The first approach to risk estima-
Optimal trading strategies for tion is based upon the use of historic
baskets correlation coefficients among the
The calculations are more complex stocks in the portfolio. One problem
in the design and implementation is that it is difficult to obtain a reli-
of an IS portfolio trading strategy. able estimate of a correlation matrix
At this point, let’s consider the fol- – and virtually impossible to accu-
lowing questions: rately estimate all of the correlation
■ How does one estimate impact coefficients in large portfolios. Even
and timing risk for the basket? smaller portfolios require a sample
And what is the correct way to size of several months in order to
construct an efficient trading obtain a meaningful estimate – and
frontier in this case? there is no guarantee that the
■ How does one take advantage of obtained values are suitable for
market opportunities and differ- intra-day timing risk calculation. 101
ent sources of liquidity while The second approach involves
maintaining the prescribed con- selecting and using risk factors to
straints and urgency of execution? construct and utilise a model for
■ How does one accommodate the timing risk estimates. There are
assorted constraints on the bas- some drawbacks here as well:
ket (dollar balance/ratio, sector which risk factors to select besides
constraints) without sacrificing the standard triple (market, size,
performance? value); and how to filter out noise
A good portfolio trading strategy in this model. However, with the
should address these problems; sim- proper choice of risk factors and
ply combining single stock IS strate- sampling intervals, this approach is
gies will not solve any of them. preferable to correlation matrices.
Once both expected input and
Efficient trading frontier – timing risk are calculated, it becomes
portfolios possible to compute the optimal 4 G. Iori, M. G.
The market impact estimate for a trading strategy for a given . Since Daniels, J. D.
Farmer, L.
portfolio can be calculated by sim- the market impact is usually mod- Gillemot et al., ‘An
ply adding the market impact esti- elled as a nonlinear function of par- analysis of price
impact function in
mates of individual trades. ticipation rate (cf. 4), this calculation order-driven
However, aggregating timing risk immediately leads to a large nonlin- markets’, Physica
A, 324 (2003),
estimates is considerably more ear optimisation problem. For a 146-151.
portfolio of 300 names the optimisa- If x(k) is the size of the position
tion problem for one day will involve k in the basket, then the MCR of
at least 300 13 = 3900 variables (if
●
this position can be estimated as
the trading day is split into thirteen the partial derivative q(k) =
30-minute bins). This is pushing the ∂R/∂x(k). Knowing the MCR of the
upper limit of capabilities for mod- position allows us to estimate how
ern optimisers. Because of this, many the portfolio risk changes if the
existing pre-trade engines settle for position is reduced by some small
an approximate solution in the case increment ∆x(k).
of large portfolios. It is crucial to
remember that the static optimal Moreover, in terms of short
schedule has very limited value – time interval ∆T, the overall change
more so for a basket of stocks as in risk can be approximated as
there are many more variables which n
can change throughout a given day. ∆R=∑q(k)∆x(k)
k=1
Dynamic portfolio trading
102 strategies and the overall change in utility
A dynamic portfolio strategy should then can be represented as
satisfy the same requirements as a n
single stock IS algorithm; being ∆U=∑∆U(k)
capable of reacting to changing mar- k=1
ket conditions and consequent
updates in statistical estimates. where
Let ∆T be a relatively short time ∆U(k) = ∆I(k) + q(k)∆x(k).
interval (on the order of a few
minutes). The strategy should aim The advantage of this approach is
to minimise the change in the utili- that now the interaction between
ty function (∆U = ∆I + ∆R) over stocks in the portfolio is eliminat-
this interval. This problem may ed, at least for very short time-
seem intractable, primarily due to frames. Then it becomes possible
the complicated nature of portfolio to find the optimal participation
risk estimates; nevertheless, given a rate for each stock in the basket –
good factor risk model, ∆R can be that is the trading rate which min-
estimated in a relatively simple imises each individual utility incre-
manner. The key here is to use a ment ∆U(k). If the strategy then
well-known concept of MCR (mar- executes at this rate over the period
ginal contribution to risk) for each ∆T, the result will be optimal for
security in the basket. the entire portfolio.
Zero MCR
Positive MCR
30
0
:0
:3
:0
:3
:0
:3
:0
:3
:0
:3
:0
:3
:0
9:
10
11
12
13
14
15
10
11
12
13
14
15
16
trading schedule for each side is from the static schedules derived
optimal. If the value of is the from historical U-curves that
same for each sub-basket, the dollar underpinned the first generation of
ratio constraint is easy to satisfy, algorithms. These original algo-
and only minor adjustments to the rithms helped traders manage their
participation rates will be required. workload by allowing them to send
In special situations, transition large lists of stocks for automated
trading may require having separate execution; the trading logic, howev-
benchmarks for each side of a bal- er, was applied individually to each
anced basket. For example, if a plan stock in the list. Now traders can
sponsor wishes to reduce his large- access advanced algorithms that
cap holdings and replace them with will trade each stock in the portfo-
small-cap stocks, it makes sense for lio according to how every other
him to consider a different bench- stock is behaving and adapt to con-
mark for each side of a transition tinually changing market condi-
portfolio. Once again, the portfolio tions. The benefits of incorporating
is split into two smaller baskets, and real-time risk and market impact
each is traded separately versus a analytics will be immediately 105
different benchmark – such as S&P apparent to index fund managers,
500 and Russell 2000, respectively. hedge funds and transition traders,
The strategy performance is then who desire a risk neutral way to
judged based on the tracking error move from one portfolio to anoth-
for each side. er. After these early adopters, usage
Other varieties and customisa- of basket algorithms will spread to
tions of portfolio trading strategy the wider investment community,
may be demanded by transition especially if delivered through an
traders. In fact, some transition intuitive trading application that
managers now apply the ETF allows them to monitor risk and
approach to planning the transition, performance in real time and adjust
viewing it as a multi-period optimi- constraints to align the execution
sation problem5. Since a dynamic strategy with their trading goals. ■
portfolio-trading strategy is also
5 C. Blake, D.
based on the ETF approach, its use Petrich, A.Ulitsky,
allows the manager a better control ‘The right tool for
the job: using
over the process of transition. multi-period
optimization in
transitions’,
Next generation Transition
The next generation of trading Management,
Institutional
algorithms has come a long way Investor 2003
The future of
algorithmic trading
Ialgorithms
n the last few years, we have wit-
nessed the rapid adoption of
to trade single stocks.
firms to trade stealthily to reduce
both the explicit and implicit trad-
ing costs by lowering commissions
107
Future pundits might call 2005 the and reducing impact costs.
‘year of the algorithm for the insti-
tutional equities trading business’. Fast forward to 2006
As the institutional trading envi- In 2006, the battle for market share
ronment has become more com- in the algorithmic space will extend
petitive, traders have turned to effi- across the European, Latin
cient algorithmic execution. American and Asian markets. In the
Algorithms like VWAP, TWAP, Americas we will likely see more
POV, PEG, SMARKET, and creative algorithmic deal making as
Implementation Shortfall are all broker/dealers will struggle to *Robert L Kissell,
part of the traders arsenal when remain competitive in the ‘low- vice president,
Global Execution
executing single stock orders. A touch’ segment. As buy-side firms Services,
recent survey of buy-side traders continue to reduce the number of JP Morgan
indicates that the drivers behind execution partners in their efforts to **Andrew
the trend of algorithmic adoption increase cost-efficiencies, many Freyre-Sanders,
head of Algorithmic
are: (1) control over the trading small broker/dealers will not be able Trading, EMEA,
process, (2) ability to focus on to commit the required financial JP Morgan
value added activities, and (3) cost resources needed to remain com- ***Carl Carrie,
control. In addition to these gains, petitive in the low-touch DMA and head of Algorithmic
Trading, USA,
trading algorithms have allowed algorithmic segment of the market. JP Morgan
■ “Sustaining algorithmic
performance will require
new investment in low-latency
some broker/dealer algorithmic
providers to creatively partner
with vendors and other
broker/dealers, while some clients
will look to outsource dealing
market data and order and/or partner to create unique
connectivity.” competitive advantages in capabili-
ties and cost structure.
Figure 1
The information contained herein is provided for Ltd., J.P. Morgan Europe Limited and J.P. Morgan plc,
information only, and any views or opinions members of the London Stock Exchange and
expressed herein are solely those of the individual regulated by the Financial Services Authority. Issued
authors and may differ from the views and opinions and distributed in Australia by J.P. Morgan Australia
expressed by other departments or divisions of Limited and J.P. Morgan Markets Australia Pte.
JPMorgan and its affiliates. Securities Limited which accept responsibility for its
JPMorgan is the marketing name for JPMorgan contents and are regulated by the Australian
Chase & Co. and its subsidiaries and affiliates Securities and Investments Commission. J.P. Morgan
worldwide. J.P. Morgan Securities Inc. is a member Markets Australia Pty. Ltd. is a licensed investment
of NASD, NYSE and SIPC. The JPMorgan Chase Bank advisor and a futures broker, and it is a member of
is a member of the FDIC. J.P. Morgan Futures Inc. is the Sydney Futures Exchange. J.P. Morgan Securities
a member of the NFA. J.P. Morgan Securities Ltd. is (Far East) Limited Seoul Branch is a member of the
authorized by the FSA. J.P. Morgan Securities Asia Korean Stock Exchange and J.P. Morgan Futures
Pte Ltd., (JPMSA) and J.P. Morgan Securities (Asia (Korea) Limited is a member of the Korean Futures
Pacific) Limited, are regulated by the Hong Kong Exchange. In the UK and other EEA countries, this
Securities & Futures Commission. JPMSA is commentary is not available for distribution to
regulated by the Monetary Authority of Singapore persons regarded as private customers (or
and the Financial Services Agency in Japan. Issued equivalent) in their home jurisdiction.
and approved for distribution in the UK and the Copyright 2005 JPMorgan Chase & Co. All rights
European Economic Area by J.P. Morgan Securities reserved.
Trading benchmarks
ETS has both agency and principal
strategies with both single stock and list
capabilities. Eight strategies are currently
offered to clients: VWAP, TWAP, TVOL,
Razor, Market on Close, Arrival Price,
Market Call and Premier Block Trading™
(PBT), an electronic and anonymous block
liquidity utility for orders up to $20
million for the entire Russell 1000 stock
universe.
BNY Brokerage
Direct Execution Services (DExSM)
Trading benchmarks
BNY does not place benchmark
constraints on its algorithmic offering.
Algorithms can be executed across a
‘myriad’ of customer-driven benchmarks.
Citigroup
Alternative Execution
CSFB
Advanced Execution Services (AES™)
Goldman Sachs
Goldman Sachs Algorithmic Trading (GSATsm)
Trading benchmarks
The prevailing benchmarks of choice used Connectivity options
by clients are VWAP and Implementation REDIPlus provides clients with access to
Shortfall. Other commonly available GSAT’s algorithms. The models are also
benchmarks include Piccolo (Small Order accessible via third party OMS vendors,
Spread Capture algorithm) and TWAP. FIX connections or Bloomberg.
GSAT’s newest algorithm, 4CAST,
explicitly balances market impact against
opportunity cost.
A number of customised algorithms
have been created to match unique,
customer-driven requirements and GSAT
intends to widen its focus to meet other
benchmarks as identified by clients.
Instinet
Algorithmic Trading Group
JPMorgan
Electronic Execution Services
Lehman Brothers
LMX™
Merrill Lynch
Merrill Lynch Execution via Algorithm and
Computer-based Trading (ML X-ACTSM)
Morgan Stanley
Benchmark Execution Strategies (BXS)
Trading benchmarks
Algorithms are constrained to meet a
number of benchmarks, including VWAP,
Arrival Price (Implementation Shortfall),
Close and Target Percentage of Volume.
Piper Jaffray
Contacts
1 Sekforde Street
London
EC1R 0BE
UK
www.thetrade.ltd.uk
Tel: +44 (0) 20 7075 6115