Statistical Models 179

Chapter 17. Statistical Models of Order-Price Dynamics

This chapter develops a small number of statistical models that describe how security prices
evolve and how order flow affects these dynamics. From one perspective these models are
short-term predictive tools. Entire books have been written about stock price forecasting, and it
is probably a safe bet that the diversity of techniques and paucity of reliable results are un-
matched by any other field of scientific endeavor.
The question guiding the present analysis, though, is not, “Which way are stock prices
headed?” The concern is instead, “How will my orders and trades affect prices?” The reason this
second question is interesting is that many of the trading strategies that we’ll discuss in the next
chapter involve sequences of trades. In a sequence, early trades may affect the prices of later
ones. Clarifying this interaction and accounting for it will be an important guiding principle.
The present analysis doesn’t take a position on any prediction tool – including your current
favorite. The results can peacefully coexist with most prediction models. Along the way I’ll point
out how the predictions can be incorporated.
The models here tend to be descriptive and statistical, rather than theoretical and eco-
nomic. The distinction is driven by practicalities. We earlier explored a simple model of in-
formed trading that exhibited order impacts and information-dependent bid-ask spreads. It is
difficult, however, to map the model directly to a real stock. (It is 9am. What will be today’s High
and Low values for Microsoft?) Instead, we’ll take the intuition that orders move prices and add
it to a statistical model. The statistical model, in turn, will be estimated from recent data that
presumably reflect how the stock behaves.
The development proceeds in stages, starting with simple models, and moving to more
complex ones.

17.1. The random walk.

The simplest model of security prices is the random walk:

𝑝𝑡 = 𝑝𝑡−1 + 𝑢𝑡

© Joel Hasbrouck, 2022. STPPms13c.docx; 2022-11-21; Part IV. Chapter 17. Statistical Models of Order-Price Dynamics; §24 p. 179
 Statistical Models 180

where 𝑡 = 1,2, … might index minutes, seconds, or some similarly brief time interval; 𝑝𝑡 is the
stock price at the close of interval t, and 𝑢𝑡 is a random perturbation, disturbance, or error. ut
reflects new information that arrives over the interval. The news might be good or bad. More
formally, 𝐸𝑡−1 𝑢𝑡 = 0: as of the end of last interval, we expected this period’s news to be neutral.
We’re often concerned with the volatility of the price dynamics, characterized by the stand-
ard deviation of ut, 𝜎𝑢 . If pt is redefined, for the moment, as the logarithm of the stock price,
then 𝑝𝑡 − 𝑝𝑡−1 is approximately the proportional (“percentage”) change in the stock price over
the interval, and u is (approximately) a proportional disturbance. In log price form, an annual
volatility of, say, 𝜎𝑢 ≈ 40% corresponds to a monthly volatility of √0.402⁄12 ≈ 11.5%, or a
daily volatility of of √0.402⁄250 ≈ 2.5%.1
We now consider estimation. The model has one parameter, the volatility 𝜎𝑢 . Suppose that
we have a set of T prices {𝑝1 , 𝑝2 , … , 𝑝𝑇 }. Then 𝑢𝑡 = 𝑝𝑡 − 𝑝𝑡−1 for 𝑡 = 2, … 𝑇 (we lose u1 because
we don’t have p0). The usual estimate for the variance would be
𝑇
1
𝜎̂𝑢2 = ∑ 𝑢𝑡2 .
𝑇−1
𝑖=2

This is the mean sum-of-squares. There are many ways of estimating volatility, however. In op-
tion pricing the above estimate is called the historical volatility. From the market prices of op-
tions, one can compute an implied volatility, which may be a better forecast of the volatility that
will arise in the future.
Stock prices lie on a discrete grid. Bids and offers in US equity markets cannot normally
specify an increment finer than $0.01, that is, there can’t be more than two places to the right of
the decimal. Other markets adopt similar conventions. While accepting this discreteness as a
feature of the data, however, we typically don’t incorporate it as a feature of the statistical
model. This simplifies estimation. It doesn’t usually cause problems, but we must remember
some of our statistical intuitions implicitly invoke continuous models, and well-behaved ones at
that. With a normally-distributed (Gaussian) random variable, for example, the intervals de-
fined by the mean ±𝜎, ±2𝜎, ±3𝜎 contain on average 68.2%, 95.4%, and 99.7% of the outcomes.
This generally won’t be true for most of the discrete distributions encountered in security pric-
ing.

17.2. Drift and alpha

An obvious shortcoming of the model is that it doesn’t allow for any expected capital gains. So
we introduce a drift term:

𝑝𝑡 = 𝜇 + 𝑝𝑡−1 + 𝑢𝑡

If pt is redefined as the logarithm of the stock price, then 𝑝𝑡 − 𝑝𝑡−1 is approximately the propor-
tional (“percentage”) change in the stock price over the month, and  is the expected percentage
price appreciation. An annual return of 20% per year works out to about 𝜇 = 0.017 (1.7%)per
month. The implied daily expected return is about  ≈ 0.0008 (0.08%) per day (using 250 trad-
ing days per year).
The random-walk model is important due to:

1Note: there are about 250 trading days in a year. Most of the cumulative volatility in financial
markets arises on trading days. Some people read this as “trading causes volatility.”

© Joel Hasbrouck, 2022. STPPms13c.docx; 2022-11-21; Part IV. Chapter 17. Statistical Models of Order-Price Dynamics; §24 p. 180
 Statistical Models 181

• The concept of informational efficiency. This is an economic argument that a rational

market incorporates all known information into the security price, so all future move-
ments must arise from information not previously known. This implies that price move-
ments should be unpredictable. The random-walk model is a simple construct with un-
predictable movements.
• Statistical tests generally show that the model is “difficult to reject.” There aren’t solid
grounds for preferring any simple alternative model.
In monthly, weekly or daily data, the model is sufficiently close to reality that it is useful for
many purposes, notably option pricing. For trading (intraday) horizons, the model needs en-
hancement to allow for drift (alpha) and order impact. We turn now to the specifics of these
modification.
When we’re using the random-walk model applied to, say, monthly log prices, the drift term
 is the expected price change over the month. Ignoring dividends this is the expected return
over the month. Expected returns over monthly and longer horizons are determined by eco-
nomic fundamentals, risk, and investors’ beliefs about these things. At intraday horizons, we
take a broader view of the drift.
The drift corresponds to predictable short-term price changes. The extent of the predictabil-
ity depends on the information, the model, and the faith placed in the model. It may be different
for different market participants
In finance, predictable over- or under-performance (net of compensation for risk) is often
called “alpha” (𝛼). In a trading context, to distinguish this predictability from effects associated
with long-run fundamentals, it is sometimes called “short-term alpha”. This parameter, then, in-
corporates the output of any supplemental prediction model. The present analysis assumes that
it is constant over the horizon of the trading problem, but with a little additional effort it can be
viewed as time-varying (𝛼𝑡 ).
The random walk with drift has two parameters, 𝜇 and 𝜎𝑢 . The first difference of the price
is Δ𝑝𝑡 = 𝑝𝑡 − 𝑝𝑡−1 = 𝜇 + 𝑢𝑡 . So we can estimate 𝜇 and 𝜎𝑢2 as the mean and variance of the Δ𝑝𝑡 ,
something that might easily be accomplished with a calculator or spreadsheet.
In practice things are bit more complicated. If the drift is attributed to the expected long-run
return, its mean is likely to be quite close to zero. As a result, while useful estimates of 𝜎𝑢 can be
obtained from a sample of 30 days, important changes in  require a few years of data to dis-
cern.

17.3. Permanent order impacts

The discussion of private information in Chapter 13 noted that it can lead to order price im-
pacts. An incoming customer buy order, for example, that lifts the market offer price might have
originated from an informed buyer, someone who holds positive information not publicly
known. This possibility causes other traders to revise upwards their beliefs about the security
value. Buying drives the price up; selling drives the price down.
We can modify the random-walk model to incorporate short-term alpha and an order price
impact. Let St = net buy volume = (# shares actively purchased – # shares actively sold), the ac-
tively bought or sold shares are understood to be those that lift the offer or hit the bid. Then:

𝑝𝑡 = 𝑝𝑡−1 + 𝛼 + 𝜆𝑆𝑡 + 𝑢𝑡 (17.1)

where  is the price impact parameter. We’d expect 𝜆 > 0. Later, when we use this model to
make predictions, 𝑆𝑡 for future times will include our own trades. But for the moment, think of
the model as describing what’s happened in the past.
By rearranging, we obtain

© Joel Hasbrouck, 2022. STPPms13c.docx; 2022-11-21; Part IV. Chapter 17. Statistical Models of Order-Price Dynamics; §24 p. 181
 Statistical Models 182

Δ𝑝𝑡 = 𝑝𝑡 − 𝑝𝑡−1 = 𝛼 + 𝜆𝑆𝑡 + 𝑢𝑡 (17.2)

This looks like a simple linear regression of Δ𝑝𝑡 on 𝑆𝑡 . The model is typically estimated over
short time intervals (like one minute). By aligning the trade prices with the record of the bids
and asks, we can compute 𝑆𝑡 as the number of executed shares priced at the ask less the num-
ber of shares priced at the bid. The model parameters (𝛼, 𝜆, and 𝜎𝑢2 ) can then be estimated using
a spreadsheet (or calculator) regression command. Not only is the model convenient to esti-
mate, but it will also turn out to be convenient for solving the order splitting problem.
Of course, models chosen for convenience aren’t necessarily the most realistic. The intui-
tion of the private information model only suggested a generally positive relation between or-
der flow and quote changes. Why should the order flow variable be net buy volume? Why
should the relation be linear?
This basic model admits many variations. The extent to which a regression fits the data is
usually summarized in its coefficient of determination, 𝑅 2 . This quantity varies between zero
(no fit at all) and one (perfect fit). We can investigate alternative specifications and see which
changes seem to improve the fit.
The basic specification is linear in signed order flow. This linearity implies that the price
change caused by a 10,000 share purchase should be 100 times the impact of a 100 share pur-
chase. In practice it is usually much less.
One way of lessening the impact of large trades is to simply ignore size altogether. Net
buys, defined as 𝑆𝑡′ = (#of active buys − #of active sells), is based on counts. A buy counts as
“+1” no matter how many shares are involved.
The use of count-based variables goes against the intuition that a large order should move
the price at least somewhat more than a small order. The desired property is an incremental
impact that diminishes with size: the additional price impact associated with adding one share
to a 10-share order should be less than if one share were added to a 10,000 share order. The de-
clining incremental impact implies a concave relation. At present, the preferred specification
involves a signed square root:

𝑆𝑡′′ = 𝑆𝑖𝑔𝑛(𝑆𝑡 )√|𝑆𝑡 |.

The Sign function is defined by 𝑆𝑖𝑔𝑛(𝑥) = +1 if 𝑥 > 0; −1 if 𝑥 < 0; and 0 if 𝑥 = 0. For example,
if 𝑆𝑡 = −10,000 𝑠ℎ𝑎𝑟𝑒𝑠, then 𝑆𝑡′′ = (−1)√10,000 = −100. This form is suggested by supportive
evidence from various studies in different markets. It is sometimes called the “square root law”
due to its apparent universality (Bouchaud, Farmer and Lillo, 2008).
Sometimes 𝑆𝑡 , 𝑆𝑡′ , and 𝑆𝑡′′ are used together in the specification (which then becomes a mul-
tiple regression). If the goal is fitting a set of data, after all, why not use all reasonable varia-
tions? The problem is that the ultimate goal is not explaining the past, but forecasting the fu-
ture. Market models are notoriously prone to a phenomenon known as “over-fitting”. (The ana-
lyst carefully crafts a statistical model that achieves an “in-sample” 𝑅 2 of 0.99, but then falls
apart when applied to a different sample or used for out-of-sample prediction.)

17.4. Temporary order impacts

All of the action in the model (or the variants discussed) takes place within interval t, suggesting
that the effect of an order on price is instantaneous. There are several reasons why we might
want to allow for more flexibility in the timing.
Different market participants receive information in differing timeliness, process infor-
mation at different speeds, and trade over different horizons. The business-cycle investor lies at

© Joel Hasbrouck, 2022. STPPms13c.docx; 2022-11-21; Part IV. Chapter 17. Statistical Models of Order-Price Dynamics; §24 p. 182
 Statistical Models 183

one extreme, the high-frequency trader at the other. The existence of slow participants opens
the possibility that some of the reaction to new information might involve a gradual price ad-
justment over time.
The market might also be subject to temporary price movements due to liquidity needs of
large investors. In contrast to the lagged information adjustment effect, though, these effects
might well involve “overshooting” and reversion. A large mutual fund, for example, that experi-
ences a large unexpected investor outflow might need to trade in size quickly. As the fund’s
large sell orders hit the market, the price is pushed down. Bidders initially suspect negative pri-
vate information and react accordingly. But once the selling order flow abates, and it appears
that there are no pending news announcements that could explain the drop, bidders will return
and the price will rise.
A simple specification that allows for a temporary impact is:

𝑝𝑡 = 𝑝𝑡−1 + 𝜆0 𝑆𝑡 + 𝜆1 𝑆𝑡−1 + 𝑢𝑡

The total impact coefficient for an order is 𝜆0 + 𝜆1 > 0. If the process is dominated by lagged
adjustment, then typically 𝜆0 > 𝜆1 > 0. If there is overshooting, then 𝜆1 < 0.

17.5. Estimating the trade price impact using daily data

The statistical models here are properly estimated with intraday, high-frequency data. These
datasets, however, are large and complex. With some simplifying assumptions, we can form an
estimate from the more common daily price and volume data.
Consider the model given in equation (17.1). On average, 𝛼 ≈ 0 and 𝑢𝑡 ≈ 0. If we assume
that these approximations hold period-by-period, we can write

Δ𝑝𝑡 = 𝜆𝑆𝑡

and, taking the absolute value of each side,

|Δ𝑝𝑡 | = 𝜆|𝑆𝑡 |

Now 𝑆𝑡 is the signed order flow over the interval. If the interval is short, we can approximate its
absolute value by the trading volume, 𝑉𝑡 ≈ |𝑆𝑡 |. Rearranging, the implied value of 𝜆 is:

|Δ𝑝𝑡 |
𝜆=
𝑉𝑡

This suggests estimating 𝜆 by

̅̅̅̅̅̅̅̅̅̅
|Δ𝑝𝑡 |
𝜆̂ = ( )
𝑉𝑡

where the overbar indicates an average taken over a sample of intervals. Since price and vol-
ume data are widely reported at a daily frequency, it is convenient to use a sample of daily data.
This estimate is a variant of the Amihud illiquidity measure, often denoted I (Amihud, 2002).

© Joel Hasbrouck, 2022. STPPms13c.docx; 2022-11-21; Part IV. Chapter 17. Statistical Models of Order-Price Dynamics; §24 p. 183
 Statistical Models 184

Summary of terms and concepts

Three models (random-walk; random-walk plus drift; random-walk plus drift plus order im-
pact).

Problems and sample exam questions

Consider the pure random-walk model, 𝑝𝑡 = 𝑝𝑡−1 + 𝑢𝑡 . The 𝑢𝑡 in this model cap-
tures
a. New public information.
b. Drift or price trend.
c. New private (asymmetric) information.
d. The probability of a takeover bid.

In the random-walk-plus-drift model, 𝑝𝑡 = 𝑝𝑡−1 + 𝛼 + 𝑢𝑡 , we’d expect a positive

drift in a situation with
a. A positive earnings surprise announced last month.
b. A belief that a large investment fund was trying to sell a large position in the stock.
c. A dividend payment made yesterday.
d. An immediate positive earnings surprise.

A trading strategy is using a random-walk-plus-drift model, 𝑝𝑡 = 𝑝𝑡−1 + 𝛼 + 𝑢𝑡

with parameter values 𝛼 = −0.01 and a standard deviation 𝑆𝐷(𝑢) = 𝜎𝑢 = 0.02. The cur-
rent stock price is 𝑝0 = $20.00. Set up the recursions that define 𝑝3 . What are the ex-
pected value and standard deviation of the stock price three periods later, that is,
𝐸𝑝3 and 𝑆𝐷(𝑝3 )

© Joel Hasbrouck, 2022. STPPms13c.docx; 2022-11-21; Part IV. Chapter 17. Statistical Models of Order-Price Dynamics; §24 p. 184
 Statistical Models 185

Answers
Answer to problem 17.1 a.

Answer to problem 17.2 d.

Answer to problem 17.3

𝑝1 = 20 − 0.01 + 𝑢1
𝑝2 = 𝑝1 − 0.01 + 𝑢2 = 20 − 2 × 0.01 + 𝑢1 + 𝑢2
𝑝3 = 𝑝2 − 0.01 + 𝑢3 = 20 − 3 × 0.01 + 𝑢1 + 𝑢2 + 𝑢3

The expectation of each u is zero, so 𝐸𝑝3 = 20 − 3 × 0.01 = 19.97;

The variance of 𝑝3 is 𝑉𝑎𝑟(𝑝3 ) = 3 × 𝜎𝑢2 = 3 × 0.022 ;
The standard deviation is 𝑆𝐷(𝑝3 ) = √𝑉𝑎𝑟(𝑝3 ) = √3 × 0.02

References
Amihud, Yakov, 2002, Illiquidity and stock returns: Cross-section and time-series effects,
Journal of Financial Markets 5, 31-56.
Bouchaud, Jean-Philippe, J. Doyne Farmer, and Fabrizio Lillo, 2008, How markets slowly digest
changes in supply and demand, in Thorsten Hens, and Klaus Schenk-Hoppe, eds.:
Handbook of Financial Markets: Dynamics and Evolution (Elsevier: Academic Press).

© Joel Hasbrouck, 2022. STPPms13c.docx; 2022-11-21; Part IV. Chapter 17. Statistical Models of Order-Price Dynamics; §24 p. 185
 Order Splitting 186

Chapter 18. Order Splitting

A $1 Billion mutual fund might hold 1% of its assets ($10 Million) in a single stock. At a repre-
sentative share price of $50, this holding consists of 200,000 shares. The median trade size in
US equity markets is about 200 shares. It is therefore almost certain that should the fund seek
to sell or reallocate the holding, the shift will have to be accomplished in multiple trades. A
larger order, like “sell 200,000 XYZ” will typically be divided among many smaller “child” (or
“daughter”) orders that are fed to the market over time. This is called order splitting.
For each child order, we face complex problems of how to peg, how much to show, what
discretionary conditions should be used, whether to make our orders visible or dark, and so on.
The larger problem, though, involving the number, size and timing of the child orders is more
tractable.
Order splitting results from minimizing the cost of buying (or maximizing the receipts from
selling) relative to some benchmark price. (This is a special case of the implementation shortfall
approach discussed in Chapter 16.) There are two important special cases. In first instance, the
benchmark is an average price computed over the working period or interval of the parent or-
der. In the second instance, we’re using a pre-trade benchmark like the bid-ask midpoint at the
time the parent order is given to the trading desk. We’ll consider each in turn.

18.1. Order splitting with interval benchmarks (TWAP and VWAP)

Among the candidates for benchmarks, we encountered the time-weighted average price
(TWAP) and the volume-weighted average price (VWAP). Now if one is trying to buy as low as
possible, or sell as high as possible, why should either of these, (or indeed, any benchmark) be
considered as a goal? The answer is that sometimes investors are more concerned with moni-
toring and measuring the performance of their broker.
For example, suppose that at 9:30 on Monday, the NBBO midpoint stands at $10, the man-
ager of the Essex fund sends a large buy order to a broker to be worked over three days, and
that Essex ends up paying an average price of $10.20 per share. The implementation shortfall
cost is $0.20 per share. Essex confronts the broker with this large trading cost, but broker

© Joel Hasbrouck, 2022. STPPms13c.docx; 2022-11-21; Part IV. Chapter 18. Order Splitting; §25 p. 186
 Order Splitting 187

simply replies, “Sorry, but the market was generally tending up over those three days.” But if
the volume-weighted average price over the period was $10.15, Essex can come back with, “Al-
right, the market trend can explain $0.15 of the implementation shortfall, but that still leaves
$0.05. Other brokers working for other customers over this same period somehow managed to
buy their shares $0.05 cheaper than ours.”
Now perhaps Essex should be evaluating the original instruction. Why was the horizon
three days? If it had been one day, the overall cost would probably have been lower. What was
the manager thinking? There will be answers to these questions, but they will be qualitative, im-
precise, and probably constructed after the fact. The implementation shortfall relative to VWAP,
on the other hand, is easily measured, interpreted, and assigned to a particular agent (the bro-
ker). The measurement issue also arises at higher levels of delegation. It is likely that the Essex
manager is herself bearing a fiduciary responsibility to manage and report trading costs. TWAP
and VWAP are relatively easy to justify and to achieve.
TWAP is the simpler of the two. Because the passage of time is perfectly predictable, one
simply trades at a constant rate over the working horizon. If 80,000 shares are to be purchased
over four hours, then 20,000 must be purchased in each hour, 5,000 over each fifteen minutes,
or 1,000 every three minutes. This does not, of course, guarantee that TWAP will be achieved,
because it depends on the trades of others as well as our own. There also remains the question
of how to manage the child orders. But the strategy of using a constant rate of trade is certainly
the place to begin.
VWAP is more complicated. In aiming at TWAP we know the total duration of the trading
day. If we knew the total volume on the day would be, say, 𝑉, we could aim at a fixed fraction of
this volume. If we seek to buy 𝑣 shares (assumed to be included in 𝑉, for simplicity), we’d want
to buy a fraction (𝑣/𝑉) of each trade. For example, if we want to buy 𝑣 = 5,000 shares on total
volume 𝑉 = 100,000 shares, we’d want to be buying 5 of every 100 shares traded. This would
require participating in each trade. Although market makers are sometimes allowed rights of
participation, this ability is not usually extended to other traders.
If the trading rate is constant over the day, then VWAP and TWAP are identical. They differ
because the trading rate (volume per unit time) changes throughout the day. Some of this varia-
tion is regular. Volume tends to be “U”-shaped: elevated at and immediately after the open, de-
clining and leveling off during mid-day, and rising again toward the end of regular closing
hours. (Madhavan, 2002) provides an example (Microsoft).
We can estimate the average trading rate by examination of behavior on previous days. Ta-
ble 18.1 illustrates the approach for a planned VWAP trade of 5,000 shares. We divide the day
into thirteen half-hour intervals starting at 9:30 AM (columns A and B). In column (C) we enter
the average volume in each interval, estimated over a sample of previous trading days. Column
(D) reports the proportion of volume for each interval. That proportion is applied to the
planned order size (5,000 shares) to get the planned trading volume in each interval (column
E). This calculation distributes the trades over time in a way matches the average volume pro-
file of the stocks throughout the day.
TWAP and VWAP strategies simply distribute orders over time. They do not explicitly take
into account the total size of the trade relative to overall market activity. If a stock has an aver-
age daily volume of 1,000 shares, then a sell order for 50,000 shares being worked over the day
will roil the market if the strategy is TWAP, VWAP or just about anything else. To avoid extreme
market impacts, one can impose participation constraints. For example, “sell 50,000 VWAP, but
our own trades should not exceed ten percent of the total volume.”

© Joel Hasbrouck, 2022. STPPms13c.docx; 2022-11-21; Part IV. Chapter 18. Order Splitting; §25 p. 187
 Order Splitting 188

Table 18.1 Planning a VWAP trade for 5,000 shares

(A) (B) (C) (D) (E)

Interval Start Avg Vol Proportion Planned Trade Volume
1 9:30 AM 11,900 9.8% 492
2 10:00 AM 10,180 8.4% 421
3 10:30 AM 9,440 7.8% 390
4 11:00 AM 8,500 7.0% 352
5 11:30 AM 7,820 6.5% 323
6 12:00 PM 6,380 5.3% 264
7 12:30 PM 6,540 5.4% 270
8 1:00 PM 7,480 6.2% 309
9 1:30 PM 8,740 7.2% 361
10 2:00 PM 8,720 7.2% 361
11 2:30 PM 10,320 8.5% 427
12 3:00 PM 12,000 9.9% 496
13 3:30 PM 12,880 10.7% 533
Total 120,900 100.0% 5,000

Participation targets, though, can have unexpected consequences. In the May 6, 2010 “flash
crash”, the S&P 500 index futures contract fell about 6% over the course of a few minutes. The
precipitating event was a large trade. According to the joint CFTC-SEC report (U.S. Commodity
Futures Trading Commission and Commission, 2010):
A large fundamental trader (a mutual fund complex) initiated a sell program to sell a total
of 75,000 E-Mini contracts (valued at approximately $4.1 billion) as a hedge to an existing
equity position. … This large fundamental trader chose to execute this sell program via an
automated execution algorithm (“Sell Algorithm”) that was programmed to feed orders
into the June 2010 E-Mini market to target an execution rate set to 9% of the trading vol-
ume calculated over the previous minute, but without regard to price or time.
An algorithm that was more aware of the impact of its orders might not have behaved so wildly.
(Madhavan, 2002) discusses alternative ways of pursuing VWAP. Instinet runs a crossing
session in which buyers and sellers are matched in advance, for a given quantity, before the
start of the trading session. At the end of the day, VWAP is determined and the price is set. This
option is only available, of course, if a matching counterparty can be found.

18.2. Optimal order splitting [No changes]

Note: This section parallels the class session on order splitting and Powerpoint slides or-
derSplitting.pptx. The Powerpoint slides are more up to date than the notes in this section.

A typical pre-trade benchmark is the bid-ask midpoint at the time the parent order was re-
leased to the trading desk, denoted 𝐵𝐴𝑀𝑝𝑟𝑒 . The implementation shortfall for, say, a buy order
is then 𝐼𝑆 = (𝑎𝑣𝑒𝑟𝑎𝑔𝑒 𝑒𝑥𝑒𝑐𝑢𝑡𝑖𝑜𝑛 𝑝𝑟𝑖𝑐𝑒) − 𝐵𝐴𝑀𝑝𝑟𝑒 . Since the 𝐵𝐴𝑀𝑝𝑟𝑒 doesn’t change over the

© Joel Hasbrouck, 2022. STPPms13c.docx; 2022-11-21; Part IV. Chapter 18. Order Splitting; §25 p. 188