Algorithmic Trading A market order is a demand for an immedi-

ate execution of a certain number of shares at the

best possible price. To get the best possible price, a
Technology continues to have an increasingly sig- market order sweeps through one side of the limit
nificant impact on how securities are traded in order book—starting with the best price—matching
today’s markets. Many trading floors have been against resting orders until the full quantity of the
replaced by electronic trading platforms and more market order is filled or the book is completely
than a third of the trading volume in the United depleted.
States can be attributed to algorithmic trading. Every Unlike a market order, a marketable limit order
large broker–dealer provides algorithmic trading ser- can be executed only at a specified price or better.
vices to their institutional clients in order to assist For example, a marketable limit order to buy 100
their trading. The algorithms used by institutional shares at $90.01 can match with a resting limit order
investors, hedge funds, and many other market par- to sell 200 shares at $90.00. The trade print—the
ticipants are used to make trading decisions about the price at which the trade would take place—would be
timing, price, and size of trades, with the objective $90.00.
of reducing risk-adjusted costs. The following examples illustrate how market
In a broad sense, the term algorithmic trading orders to sell interact with resting limit orders to buy.
is used to describe trading in an automated man- Figure 1 shows the idealized market impact of a
ner according to a set of rules. It is often used 200-share market order to sell. The x- and y-axes
interchangeably with statistical trading or statistical display the time and price, respectively.
arbitrage, which may or may not be automated, but The bid side of the limit order book contains
is based on signals derived from statistical analyses bids to buy a certain number of shares of stock at
or models. Smart order routing, program trading, and a certain price. Resting limit orders—orders that sit
rules-based trading are some of the other terms asso- in the order book—are said to provide liquidity by
ciated with algorithmic trading. More recently, the mitigating the market impact of orders that must be
range of functions and activities associated with algo- filled immediately. The state of the book establishes
rithmic trading has grown to include market impact a pretrade equilibrium (1), which is disturbed by a
modeling, execution risk analytics, cost aware portfo- market order to sell 200 shares (2). Market orders
lio construction, and the use of market microstructure must be filled immediately, and therefore represent a
effects. demand for liquidity.
In this article, we first explain the basic ideas As the sell order depletes the bid book by match-
of market impact and optimal execution from both ing with limit orders to buy, it obtains an increasingly
the sell- and buy-side perspectives. We then provide less favorable (lower) trade price, resulting in the
an overview of the most popular algorithmic trading trade print (3). Assuming no other trading activity,
strategies. over time, liquidity providers replenish the bid book
to (4), which is the posttrade equilibrium.
The difference between (4) and (1) is an infor-
Market Impact and the Order Book mation-based effect called permanent market impact.
It is the market’s response to information that a
The limit order book contains resting limit orders. market participant has decided not to own 200 shares
These orders rest in the book and provide liquidity as of this stock. This effect is typically modeled as
they wait to be matched with nonresting orders, which immediate and linear in the total number of shares
represent a demand for liquidity. The three most executed. Huberman and Stanzl [9] show that, if
common types of nonresting orders are marketable the effect were not linear and immediate, buying
limit orders, market orders, and fill-or-kill orders. and selling at two different rates could produce an
The bid side of the limit book contains resting arbitrage profit.
bids to buy a certain number of shares of stock at a The difference between (4) and (3) is called tem-
certain price. The offer side contains resting offers to porary market impact. The trader who initiated the
sell a certain number of shares of stock at a certain trade is willing to obtain a less favorable fill price
price. (3) to get his/her trade done immediately. This cost
2 Algorithmic Trading

1 2
Posttrade equilibrium Permanent
Pretrade equilibrium impact
4

200 shares sold

Price
Temporary
impact

Trade print
3

Time

Figure 1 Idealized market impact model showing sell of 200 shares

of immediacy is typically modeled as linear or square than the pretrade equilibrium because it incorporates
root. Under the assumption of square root impact, the information of the executed market order.
with all other factors held constant, a trade of 200 Our trader then places another market sell order
shares executed over the same period of time as a for 100 shares (6) and obtains a trade print (7). Over
trade of 100 shares would have square root of two time, the temporary impact—(8) minus (7)—decays
times more temporary impact per share. and results in a new posttrade equilibrium (8). As
Figure 2 shows what would happen if the same the permanent impact is assumed to be linear and
trader were willing to wait some time between trades. immediate, the posttrade equilibrium is shown to be
The trade print from the previous figure is shown the same for one order of 200 shares as it is for two
as a reference point (1). As in Figure 1, a pretrade orders of 100 shares each.
equilibrium (2) is disturbed by a 100-share market
order to sell (3). As the market order depletes the
bid book by matching with limit orders to buy, Optimal Execution
it obtains a fill price (4). Over time (5), liquidity
providers refill the bid book with limit orders to buy. While our trader waits between trades (5), he/she
However, the new posttrade equilibrium (6) is lower incurs price risk—the risk that his/her execution will

2 3
6 Permanent
Pretrade equilibrium Posttrade equilibrium
impact
100 shares 8
100 shares sold sold Temporary
4 impacts
Trade print 7
Trade print
Price

5
Source of risk and
exposure to
expected drift
1

Time

Figure 2 Idealized market impact model showing two sells of 100 shares each
 Algorithmic Trading 3

be less favorable due to the random movement of the losses associated with this expectation of price
prices. In this context, a shortfall is the difference change.
between the effective execution price and the arrival For example, a trader has positive alpha if he/she
price—the prevailing price at the start of the exe- expects prices to move lower while he/she is exe-
cution period. If we use the variance of shortfalls cuting his/her sell orders. He/she may choose to
as a proxy for risk, a trader’s aversion to risk estab- front-weight his/her trade schedule—execute more
lishes a risk/cost trade-off. In the first scenario, he/she rapidly at the beginning of the execution period—to
pays a higher cost—the difference between (8) and obtain better execution prices. Similarly, a seller
(1)—to eliminate risk. In the second scenario, he/she who believes that prices are moving higher may
pays a lower cost—the average of the differences back-weight his/her trade schedule or delay the
between (8) and (4), and (8) and (7)—but takes on execution.
a greater dispersion of shortfalls associated with the The general form of the optimal execution prob-
waiting time between trades (5). This is the trade- lem involves finding of the best trade-off between
off considered in the seminal paper of Almgren and the effects of risk, market impact, and alpha by
Chriss [1]. minimizing risk-adjusted costs relative to a prespec-
Risk aversion increases a trader’s sense of urgency ified benchmark. Common benchmarks are volume-
and makes it attractive to pay some premium to weighted average price (VWAP) and arrival price (the
reduce risk. The premium the trader pays is in the price prevailing at the beginning of the execution
form of higher temporary market impact. All other period).
factors held constant, a higher expected temporary The first formulations of this problem go back
market impact encourages slower trading, while a to the seminal papers of Bertsimas and Lo [4] and
higher expected risk or risk aversion encourages Almgren and Chriss [1]. Assuming a quadratic utility
faster trading. function, a general formulation of this problem takes
Risk aversion embodies the notion that people dis- the form
like risk. For a risk-averse agent, the utility of a fair
game, u(G), is less than the utility of having the min[E(C(xt )) + λV ar(C(xt ))] (1)
xt
expected value of the game, E(u(G)), with certainty.
The degree of risk aversion may be captured by the where C(xt ) is the cost of deviating from the bench-
risk aversion parameter λ, which is used to trans- mark. The solution is given by the trade schedule xt
late risk into a certain dollar cost equivalent—the that represents the number of shares that remains to
smallest certain dollar amount that would be accepted buy/sell at time t. The trader’s optimal trade schedule
instead of the uncertain payoff from the fair game. is a function of his/her level of risk aversion, λ ≥ 0,
For an agent with quadratic utility, the certain dollar which determines his/her urgency to trade, and dic-
cost equivalent is given by E(G) − λV ar(G). Hence, tates the preferred trade-off between execution cost
his/her degree of risk aversion is characterized by and risk.
the family of risk/return pairs with the same con- In the following two subsections, we describe the
stant trade-off between the expected return and risk. sell-and buy-side perspectives of the typical arrival
An annualized target return and standard deviation price optimal execution models.
imply a risk aversion, and may be translated to a risk
aversion parameter of the type used in some optimal The Sell-side Perspective
execution algorithms.
Another factor that influences the decision to trade The typical optimal execution model uses arrival
more quickly or more slowly is the expectation of price as a benchmark and balances the trade-off
price change. For the purpose of execution, a positive between market impact, price risk, and opportunity
alpha is an expectation of profits per share per cost. Alpha is assumed to be greater than or equal to
unit time for unexecuted shares. A faster execution zero, which means that delaying execution may carry
captures more of the profits associated with this an associated opportunity cost, but does not carry an
expectation of price change. A negative alpha is expectation of profit. The optimal strategy lies some-
the expectation of losses per share per unit time for where between two extremes: (i) trade everything
unexecuted shares. A slower execution incurs less of immediately at a known cost or (ii) reduce market
4 Algorithmic Trading

impact by spreading the order into smaller trades over Impact Models
a longer horizon at the expense of increased price risk
and opportunity cost. An impact model is used to predict changes in price
Bertsimas and Lo [4] proposed an algorithm due to trading activity. This expectation of price
for the optimal execution problem that finds the change may be used to inform execution and portfolio
minimum expected cost of trading over a fixed period construction decisions. Several well-known models
of time for a risk neutral trader, λ = 0, facing an have been proposed. For example, see Hasbrouck [8],
environment where price movements are assumed to Lillo et al. [12], and Almgren et al. [3].
be serially uncorrelated. Almgren et al. use a proprietary data set obtained
Almgren and Chriss [1] extended this concept from Citigroup’s equity trading desk in which a
using quadratic utility to embody the trade-off trade’s direction (buyer or seller initiated) is known.
between the expected cost and price risk. The more Note that for most public data sets, trade direction
aggressive (passive) trade schedules incur higher is not available and has to be estimated by a classi-
(lower) market impact costs and lower (higher) fication algorithm. Classification errors in algorithms
price risk. Similar to classical portfolio theory, as such as Lee and Ready [11], and Ellis et al. [5] intro-
λ varies the resulting set of points (V ar(λ), E(λ)) duce a bias that produces an overestimate of the true
trading cost.
traces out the efficient frontier of optimal trad-
In Almgren et al., trades serve as a proxy for
ing strategies. The two extreme cases λ = 0 and
trading imbalance. The authors assume that, some
λ → ∞ correspond to the minimum impact strat-
time after the complete execution of a parent order,
egy—trading at a constant rate throughout the execu-
only permanent impact remains. This allows them
tion period—and the minimum variance strategy—a
to separate impact into its temporary and permanent
single execution of the entire target quantity at the
components.
start of the execution period.
The model parameters can then be calculated
Let us consider selling X shares, that is, we
from a regression, giving the following results. First,
want x0 = X and xT = 0. Under the assumptions that permanent impact cost is linear in trade size and
asset prices follow an arithmetic Brownian motion, volatility. Second, temporary impact cost is linear
permanent impact is immediate and linear in total in volatility and roughly proportional to the square
shares executed, and temporary impact is linear in root—Almgren et al. find a power 3/5—of the
the rate of trading, the solution of the Almgren and fraction of volume represented by one’s own trading
Chriss model is during the period of execution. Hence, for a given
rate of trading, a less volatile stock with large
sinh(κ(T − t)) average daily volume has the lowest temporary
xt = X (2)
sinh(κT ) impact costs.

2
where κ = λση . Here, σ and η represent stock The Buy-side Perspective
volatility and linear temporary market impact cost. Optimal execution algorithms have less value to a
Note that the solution is effectively a decaying typical portfolio manager if analyzed separately from
exponential X exp(−κt) adjusted such that xT = 0. the corresponding returns earned by his/her trading
It does not depend on the permanent market impact, strategy. In fact, high transaction costs are not bad
consistent with the discussion in the previous section. per se —they could simply prove to be necessary for
The urgency of trading is embodied in κ. This param- generating superior returns. At present, the typical
eter determines the speed of liquidation independent sell-side perspective of algorithmic trading does not
of the order size X. For a higher risk aversion parame- take expectation of profits or the client’s portfolio
ter or volatility—for example, representing increased objectives into account. Needless to say, this is an
perceived risk—the speed of trading increases as important component of execution.
well. We also see that for a higher expected tem- The decisions of the trader and the portfolio man-
porary market impact cost, the speed of trading ager are based on different objectives. The trader
decreases. decides on the timing of the execution, breaking
 Algorithmic Trading 5

large parent orders into a series of child orders (selling) orders, which implies that risk is reduced
that, when executed over time, represent the correct (increased). If this term is ignored, which occurs
trade-off between opportunity cost, market impact, when portfolio allocation and optimal execution are
and risk. The trader sees only the trading assets, performed separately, then the measurement of total
whereas the portfolio manager sees the entire port- risk is biased.
folio, which includes both, the trading assets and the
static—nontrading—positions.
The portfolio manager’s task is to construct a Popular Algorithmic Trading Strategies
portfolio by optimizing the trade-off between the
opportunity cost, market impact, and risk for the A small number of execution strategies have become
full set of trading and nontrading assets. In general, de facto standards and are offered by most technology
the optimal execution framework described by Alm- providers, banks, and institutional broker/dealers.
gren and Chriss is not appropriate for the portfolio However, even among these standards, the large
manager. number of input parameters makes it difficult to
Engle and Ferstenberg [6] proposed a framework compare execution strategies directly.
that unites these objectives by combining optimal Typically, a strategy is motivated by a theme or
execution and classical mean–variance optimization style of trading. The objective is to minimize either
models. In their model, trading takes place at discrete absolute or risk-adjusted costs relative to a bench-
time intervals as the portfolio manager rebalances mark. For strategies with mathematically defined
his/her portfolio holdings wt at times t = 0, 1, . . . , T objectives, an optimization is performed to deter-
subject to changing expected returns, µt , and risk mine how to best use the strategy to maximize a
(as measured by the covariance matrix of returns), trader’s or portfolio manager’s utility. A trade sched-
t , until he/she reaches the portfolio that reflects ule —or trajectory —is planned for strategies with
his/her final view wT = 2λ1
−1
T µT . The joint dynamic
a target quantity of shares to execute. The order
optimization problem has the form placement engine—sometimes called the micro-
 T trader —translates from a strategy’s broad objec-
  tives to individual orders. User-defined input param-
max wTT µT − λwTT T wT eters control the trade schedule and order placement
{wt }
t=1 strategy.

T
 In this section, we review some of the most
− wtT τt + (wT − wt−1 )µt common algorithmic trading strategies.
t=1

+ λ(wT − wt−1 )t (wT − wt−1 ) Volume-weighted Average Price


T
Six or seven years ago, the VWAP execution strategy
+ 2λ (wT − wt−1 )t wT (3)
represented the bulk of algorithmic trading activity.
t=1
Currently, it is second in popularity only to arrival
where τt = τt (wt ) is the temporary market impact price. The appeal of benchmarking to VWAP is that
function (for simplicity of exposition, we ignore the benchmark is easy to compute and intuitively
permanent impacts). This is a dynamic program- accessible.
ming problem that has to be solved by numerical The typical parameters of a VWAP execution are
techniques. the start time, the end time, and the number of shares
Each one of the three terms in the objective to execute. Additionally, optimized forms of this
function above has an intuitive interpretation. The strategy require a choice of risk aversion.
first term represents the standard mean–variance The most basic form of VWAP trading uses a
optimization problem. The second term corresponds model of the fractional daily volume pattern over
to the optimal execution problem. The third term the execution period. A trade schedule is calculated
is the covariance between the remaining shares to to match this volume pattern. For example, if the
be traded and the final position. In the single asset execution period is one day, and 20% of a day’s
case, the third term is positive (negative) for buying volume is expected to be transacted in the first hour,
6 Algorithmic Trading

a trader using this basic strategy would trade 20% but has a higher expected cost. For a particular choice
of his/her target accumulation or liquidation in the of risk aversion, somewhere between the highest
first hour of the day. Since the daily volume pattern and lowest risk strategies, is a compromise optimal
has a U shape—with more trading in the morning strategy that perfectly balances risk and costs.
and afternoon and less in the middle of the day—the For example, a risk-averse VWAP strategy might
volume distribution of shares executed in a VWAP place one market order of 100 shares every 20 s,
pattern would also have this U shape. whereas a less risk-averse strategy might place a limit
VWAP is an ideal strategy for a trader who meets order of 200 shares, and, 40 s later, place a market
all of the following criteria: order for the difference between the desired fill of 200
and the actual fill (which may have been smaller). The
• his/her trading has little or no alpha during the choice of the average time between market orders in
execution period; a VWAP execution implies a particular risk aversion.
• he/she is benchmarked against the VWAP; For market participants with a positive alpha, a
• he/she believes that market impact is minimized frequently used rule-of-thumb optimization is com-
when his/her own rate of trading represents pressing trading into a shorter execution period. For
the smallest possible fraction of all trading example, a market participant may try to capture more
activity; and profits by doing all of his/her VWAP trading in the
• he/she has a set number of shares to buy or sell. first half of the day instead of taking the entire day
to execute.
Deviation from these criteria may make VWAP In another variant of VWAP—guaranteed
strategies less attractive. For example, market partic- VWAP—a broker commits capital to guarantee
ipants who trade over the course of a day and have his/her client the VWAP price in return for a pre-
strong positive alpha may prefer a front-weighted tra- determined fee. The broker takes on a risk that the
jectory, such as those that are produced by an arrival difference between his/her execution and VWAP will
price strategy. be greater than the fee he/she collects. If institu-
The period of a VWAP execution is most typi- tional trading volume and individual stock returns
cally a day or a large fraction of a day. Basic VWAP were uncorrelated, the risk of guaranteed VWAP trad-
models predict the daily volume pattern using a sim- ing could be diversified away across many clients
ple historical average of fractional volume. Several and many stocks. In practice, managing a guaranteed
weeks to several months of data are commonly used. VWAP book requires some complex risk calculations
However, this forecast is noisy. On any given day, the that include modeling the correlations of institutional
actual volume pattern deviates substantially from its trading volume.
historical average, complicating the strategy’s objec-
tive of minimizing its risk-adjusted cost relative to
Time-weighted Average Price
the VWAP benchmark. Some models of fractional
volume attempt to increase the accuracy of volume The time-weighted average price (TWAP) execution
pattern prediction by making dynamic adjustments strategy attempts to minimize market impact costs
to the prediction based on observed trading results by maintaining an approximately constant rate of
throughout the day. trading over the execution period. With only a
Several variations of the basic VWAP strategy are few parameters—start time, end time, and target
common. The ideal VWAP user (as defined earlier) quantity—TWAP has the advantage of being the
can lower his/her expected costs by increasing his/her simplest execution strategy to implement. As with
exposure to risk relative to the VWAP benchmark. VWAP, optimized forms of TWAP may require a
For example, assuming an alpha of zero, placing limit choice of risk aversion. Typically, the VWAP or
orders throughout the execution period and catching arrival price benchmarks are used to gauge the quality
up to a target quantity with a market order at the end of a TWAP execution. TWAP is hardly ever used as
of the execution period will lower the expected cost its own benchmark.
while increasing risk. This is the highest risk strategy. The most basic form of TWAP breaks a parent
Continuously placing small market orders in the order into small child orders and executes these child
fractional volume pattern is the lowest risk strategy, orders at a constant rate. For example, a parent order
 Algorithmic Trading 7

of 300 shares with an execution period of 10 min trading volume to show up on the tape, and follows
could be divided into three child orders of 100 shares this volume with market orders. For example, if the
each. The child orders would be executed at the 3:20, target fractional participation rate is 10%, and an
6:40, and 10:00 min marks. Between market orders, execution of 10 000 shares is shown to have been
the strategy may place limit orders in an attempt to transacted by other market participants, a participa-
improve execution quality. tion strategy would execute 1000 shares in response.
An ideal TWAP user has almost the same charac- Unlike a VWAP trading strategy, which for a
teristics as an ideal VWAP user, except that he/she given execution may experience large deviations from
believes that the lowest trading rate —not the lowest an execution period’s actual volume pattern, partic-
participation rate —incurs the lowest market impact ipation strategies can closely track the actual—as
costs. opposed to the predicted—volume pattern. However,
TWAP users can benefit from the same type close tracking has a price. In the earlier example,
of optimization as VWAP users by placing market placing a market order of 1000 shares has a larger
orders less frequently, and using resting limit orders expected market impact than slowly following the
to attempt to improve execution quality. market’s trading volume with smaller orders. An
optimized form of the participation strategy amor-
Participation tizes the trading shortfall over some period of time.
Specifically, if an execution of 10 000 shares is shown
The participation strategy attempts to maintain a to have been transacted by other market participants,
constant fractional trading rate. That is, its own instead of placing 1000 shares all at once, a 10% par-
trading rate as a fraction of the market’s total trading ticipation strategy might place 100 share orders over
rate should be constant throughout the execution some period of time to amortize the shortfall of 1000
period. If the fractional trading rate is maintained shares. The result is a lower expected shortfall, but a
exactly, participation strategies cannot guarantee a higher dispersion of shortfalls.
target fill quantity.
The parameters of a participation strategy are Market-on-close
the start time, end time, fraction of market volume
the strategy should represent, and max number of The market-on-close strategy is popular with mar-
shares to execute. If the max number of shares ket participants who either want to minimize risk-
is specified, the strategy may complete execution adjusted costs relative to the closing price of the day
before the end time. Along with VWAP and TWAP, or want to manipulate—game —the close to create
participation is a popular form of nonoptimized the perception of a good execution. The ideal market-
strategies, though some improvements are possible on-close user is benchmarked to the close of the day
with optimization. and has low or negative alpha. The parameters of
VWAP and arrival price benchmarks are often a market-on-close execution are the start time, the
used to gauge the quality of a participation strat- end time, and the number of shares to execute. Opti-
egy execution. The VWAP benchmark is particularly mized forms of this strategy require a risk-aversion
appropriate because the volume pattern of a perfectly parameter.
executed participation strategy is the market’s volume When market-on-close is used as an optimized
pattern during the period of execution. An ideal user strategy, it is similar in its formulation to an arrival
of participation strategies has all of the same charac- price strategy. However, with market-on-close, a
teristics as an ideal user of VWAP strategies, except back-weighted trade schedule incurs less risk than a
that he/she is willing to forego certain execution to front-weighted one. With arrival price, an infinitely
maintain the lowest possible fractional participation risk averse trader would execute everything in the
rate. opening seconds of the execution period. With
Participation strategies do not use a trade schedule. market-on-close, an infinitely risk averse trader would
The strategy’s objective is to participate in volume execute everything at the closing seconds of the
as it arises. Without a trade schedule, a participation day. For typical levels of risk aversion, some trading
strategy cannot guarantee a target fill quantity. The would take place throughout the execution period.
most basic form of participation strategies waits for As with arrival price optimization, positive alpha
8 Algorithmic Trading

increases urgency to trade and negative alpha encour- For traders with positive alpha, the feasible region
ages delayed execution. of trade schedules lies between the immediate exe-
In the past, market-on-close strategies were used cution of total target quantity and a constant rate of
to manipulate—or game —the close, but this has trading throughout the execution period.
become less popular as the use of VWAP and arrival A more general form of arrival price optimiza-
price benchmarks have increased. Gaming the close tion allows for both the buyers and sellers to have
is achieved by executing rapidly near the close of either positive or negative alpha. For example, under
the day. The trade print becomes the closing price the assumption of negative alpha, shares held long
or very close to it, and hence shows little or no and scheduled for liquidation are—without con-
shortfall from the closing price benchmark. The true sidering one’s own trading—expected to go up
cost of the execution is hidden until the next day in price over the execution period. This would
when temporary impact dissipates and prices return encourage a trader to delay execution or stretch
to a new equilibrium. out trading. Hence, the feasible region of solutions
that account for both positive and negative alpha
Arrival Price includes back-weighted as well as front-weighted
trade schedules.
The arrival price strategy—also called the imple- Other factors that necessitate back-weighted trade
mentation shortfall strategy—attempts to minimize schedules in an arrival price optimization are ex-
risk-adjusted costs using the arrival price benchmark. pected changes in liquidity and expected crossing
Arrival price optimization is the most sophisticated opportunities. For example, an expectation of a cross
and popular of the commonly used algorithmic trad- later in the execution period may provide enough cost
ing strategies. savings to warrant taking on some price risk and the
The ideal user of arrival price strategies has the possibility of a compressed execution if the cross fails
following characteristics: to materialize. Similarly, if market impact costs are
expected to be lower later in the execution period, a
• he/she is benchmarked to the arrival price; rational trader may take on some risk to obtain this
• he/she is risk averse and knows his/her risk cost savings.
aversion parameter; A variant of the basic arrival price strategy is
• he/she has high positive or high negative alpha; adaptive arrival price. A favorable execution may
and result in a windfall in which an accumulation of
• he/she believes that market impact is minimized a large number of shares takes place at a price
by maintaining a constant rate of trading over the significantly below the arrival price. This can happen
maximum execution period while keeping trade by random chance alone. Almgren and Lorenz [2]
size small. demonstrated that a risk-averse trader should use
Most implementations are based on some form some of this windfall to reduce the risk of the
of the risk-adjusted cost minimization introduced by remaining shares. He/she does this by trading faster
Almgren and Chriss [1]. In the most general terms, and thus incurring a higher market impact. Hence,
an arrival price strategy evaluates a series of trade the strategy is adaptive in that it changes its behavior
schedules to determine which one minimizes risk- based on how well it is performing.
adjusted costs relative to the arrival price benchmark.
As discussed in the section on optimal execution, Crossing
under certain assumptions, this problem has a closed
form solution. Though crossing networks have been around for some
The parameters in an arrival price optimization time, their use in algorithmic trading strategies is
are alpha, number of shares to execute, start time, a relatively recent development. The idea behind
end time, and a risk aversion parameter. For buyers crossing networks is that large limit orders—the
(sellers), positive (negative) alpha encourages faster kind of orders that may be placed by large insti-
trading. For both buyers and sellers, risk encourages tutional traders—are not adequately protected in a
faster trading, while market impact costs encourage public exchange. Simply displaying large limit orders
slower trading. in the open book of an electronic exchange may
 Algorithmic Trading 9

leak too much information about institutional traders’ to see if their behavior is more consistent with normal
intentions. This information is used by prospective trading than with gaming.
counterparties to trade more passively in the expecta- There are several different kinds of crossing net-
tion that time constraints will force traders to replace works. A continuous crossing network constantly
some or all of the large limit orders with market sweeps through its book in an attempt to match buy
orders. In other words, information leakage encour- orders with sell orders. A discrete crossing network
ages gaming of large limit orders. Crossing networks specifies points in time when a cross will take place,
are designed to limit information leakage by making say every half hour. This allows market participants
their limit books opaque to both their clients and the to queue up in the crossing network just prior to a
general public. cross instead of committing resting orders to the net-
A popular form of cross is the midquote cross, in work for extended periods of time. Some crossing
which two counterparties obtain a midquote fill price. networks allow scraping—a one time sweep to see
The midquote is obtained from a reference exchange, if a single order can find a counterparty in the cross-
such as the NYSE or other public exchange. Reg- ing network’s book—while others allow only resting
ulations require that the trade is then printed to a orders.
public exchange to alert other market participants that In automated crossing networks, resting orders are
it has taken place. The cross has no market impact matched according to a set of rules, without direct
but both the counterparties pay a fee to the cross- interaction between the counterparties. In negotiated
ing network. These fees are typically higher than the crossing networks, the counterparties first exchange
fees for other types of algorithmic trading because indications of interest, then negotiate price and size
the market impact savings are significant while the via tools provided by the system.
Some traditional exchanges now allow the use
fee is contingent on a successful cross.
of invisible orders, resting orders that sit in their
More recently, crossing networks have offered
order books but are not visible to market participants.
their clients the ability to place limit orders in the
These orders are also referred to as dark liquidity. The
crossing networks’ dark books. Placing a limit order
difference between these orders and those placed in
in a crossing network allows a cross to occur only at
a crossing network is that traditional exchanges offer
a certain price. This makes crossing networks much
no special antigaming protection.
more like traditional exchanges, with the important
Private dark pools are collections of orders that
difference that their books are opaque to market are not directly available to the public. For example,
participants. a bank or pension manager might have enough order
To protect their clients from price manipulation, flow to maintain an internal order book that, under
crossing networks implement antigaming logic. As special circumstances, is exposed to external scraping
previously explained, opaqueness is itself a form by a crossing network or crossing aggregator.
of antigaming, but there are other strategies. For A crossing aggregator charges a fee for man-
example, some crossing networks require orders to aging a single large order across multiple crossing
be above a minimum size or to remain in the networks. Order placement and antigaming rules dif-
network longer than a prespecified minimum time. fer across networks, making this task fairly complex.
Other networks will cross only orders of similar size. A crossing aggregator may also use information about
This prevents traders from pinging—sending small historical and real-time fills to direct orders. For
orders to the network to determine which side of the example, failure to fill a small resting buy order in a
network’s book has an order imbalance. crossing network may betray information of a much
Another approach to antigaming prevents crosses larger imbalance in the network’s book. This makes
from taking place during periods of unusual market the network a more attractive destination for future
activity. The assumption is that some of this unusual sell orders. In general, the management of informa-
activity is caused by traders trying to manipulate the tion across crossing networks should give crossing
spread in the open markets to get a better fill in a aggregators higher fill rates than exposure to any indi-
crossing network. vidual network.
Some networks also attempt to limit participation Crossing lends itself to several optimization strate-
by active traders, monitoring their clients’ activities gies. Longer exposure to a crossing network not only
10 Algorithmic Trading

increases the chances of an impact-free fill, but also current holdings. When executed—often relatively
increases the risk of a large and compressed execution quickly—these deviations result in significant market
if an order fails to obtain a fill. Finding an optimal impact costs.
exposure time is one type of crossing optimization. Engle and Ferstenberg [6] show that to correctly
A more sophisticated version of this approach is solv- measure risk, we must take both the existing positions
ing for a trade-out, a schedule for trading shares out and unexecuted shares into account. This idea unites
of the crossing network into the open markets. As execution risk with portfolio risk. Portfolio construc-
time passes and a cross is not obtained, the strategy tion and optimal execution are similarly united by
mitigates the risk of a large, compressed execution incorporating market impact costs directly into the
by slowly trading parts of the order into the open portfolio construction process.
markets. Ideally, the portfolio manager would like to
solve a problem similar in nature to the multi-
Other Algorithms period consumption-investment problem [13], that,
in addition, takes market impact costs and changing
Two other algorithms are typically included in the probability distributions for a large universe of secu-
mix of standard algorithmic trading offerings. The rities into account. This dynamic portfolio or small
first is liquidity seeking where the objective is to soak delta continuous trading problem represents the next
up available liquidity. As the order book is depleted, step in the evolution of institutional money manage-
trading slows down. As the order book is replenished, ment. However, it presents some mathematical and
trading speeds up. computational challenges. As has been pointed out
The second algorithm is financed trading. The idea by Sneddon [14], dynamic portfolio problem dif-
behind this strategy is to use a sale to finance the fers in several important ways from the classical
purchase of a buy with the objective of obtaining multiperiod consumption-investment problem. First,
some form of hedge. This problem has all of the the return probability distributions change through-
components of a full optimization. For example, if, out time. Second, the objective functions for active
after a sell, a buy is executed too quickly, it will portfolio management do not depend on the pre-
obtain a less favorable fill price. On the other hand, dicted alpha/risk, but rather on realized return/risk.
executing a buy leg too slowly increases the tracking Finally, the dynamics of the model may be far more
error between the two components of the hedge and complex. Grinold [7] provides an elegant and analyti-
increases the dispersion of costs required to complete cally tractable but greatly simplified model. Kolm and
the hedge. Maclin [10] describe a full-scale simulation-based
framework that incorporates realistic constraints and
What is Next? a transaction cost model.
Other efforts of ongoing research in algorithmic
The average trade size for IBM, as reported in the trading are extending market microstructure and opti-
Trade and Quote (TAQ) database, declined from 650 mal execution models to futures, options, and fixed
shares in 2004 to 240 shares in 2007. Falling trade income products. These initiatives follow the domi-
sizes are evidence of the impact of algorithmic tra- nant theme of algorithmic trading, the creation of a
ding. Large, infrequent portfolio rebalancing and unified view, an all-encompassing framework for the
trading are being replaced by small delta continuous entire trading process, including modeling, portfolio
trading. construction, risk analytics, and execution across all
The antithesis of the small delta continuous trading tradeable asset classes.
approach is embodied in the idea of lazy port-
folios, in which portfolios are rebalanced infre- References
quently to reduce market impact costs. The first
argument against lazy portfolios is that as time passes,
[1] Almgren, R. & Chriss, N. (2000). Optimal execution of
the weights drift further away from optimal target portfolio transactions, Journal of Risk 3(2), 5–39.
holdings, in both the alpha and risk dimensions. [2] Almgren, R. & Lorenz, J. (2007). Adaptive arrival price,
Second, the use of an optimizer after long hold- in Algorithmic Trading III: Precision, Control, Execu-
ing periods tends to produce large deviations from tion, B.R. Bruce, ed, Institutional Investor Journals.
 Algorithmic Trading 11

[3] Almgren, R., Thum, C., Hauptmann, E. & Li, H. (2005). [10] Kolm, P. & Maclin, L. (2009). Small Delta Continuous
Equity market impact, Risk 18(7), 57–62. Trading for Dynamic Portfolios, Courant Institute, New
[4] Bertsimas, D. & Lo, A.W. (1998). Optimal control of York University.
liquidation costs, Journal of Financial Markets 1, 1–50. [11] Lee, C. & Ready, M. (1991). Inferring trade direction
[5] Ellis, K., Michaely, R. & O’Hara, M. (2000). The from intraday data, Journal of Finance 46, 733–746.
accuracy of trade classification rules: Evidence from [12] Lillo, F., Farmer, J.D. & Mantegna, R.N. (2003). Master
Nasdaq, Journal of Financial and Quantitative Analysis curve for price-impact function, Nature 421, 129–130.
35(4), 529–551. [13] Merton, R. (1969). Lifetime portfolio selection under
[6] Engle, R.F. & Ferstenberg, R. (2007). Execution risk, uncertainty: The continuous-time case, Review of Eco-
Journal Portfolio Management 33, 34–44. nomics and Statistics 51, 241–257.
[7] Grinold, R.C. (2007). Dynamic portfolio analysis, Jour- [14] Sneddon, L. (2005). The Dynamics of Active Portfolios,
nal of Portfolio Management 34(1), 16–26. Westpeak Global Advisors.
[8] Hasbrouck, J. (1991). Measuring the information content
of stock trades, Journal of Finance 46(1), 179–207. PETTER N. KOLM & LEE MACLIN
[9] Huberman, G. & Stanzl, W. (2004). Price manipulation
and quasi-arbitrage, Econometrica 72(4), 1247–1275.