Exploratory Trading

Adam Clark-Joseph∗

October 13, 2011

Abstract
Empirical research suggests that high-frequency traders (HFTs) tend to be better
informed in some respects than their lower-frequency counterparts, but the precise con-
nection between rapid trading and superior information remains unclear. Anecdotal
accounts suggest that at least some HFTs use their own trades to gather information,
but such “exploratory trading” is poorly understood. In this paper, I formalize the in-
tuitive concept of exploratory trading in a simple model, and I show that exploratory
trading and high-frequency trading bear a natural relationship to one another. My
model sheds light on the broad question of how HFTs could parlay their speed into
valuable information, and it provides explanations for a variety of empirical findings
about high-frequency trading. In addition, the exploratory trading model generates a
number of new testable predictions about HFT activity.

∗
Harvard University, E-mail: [email protected]. I thank Andrei Kirilenko and other seminar
participants at the Commodity Futures Trading Commission, as well as seminar participants at Harvard
University for their useful feedback, and I thank John Campbell, Andrei Shleifer, Alp Simsek and Jeremy
Stein for their invaluable advice and comments. I gratefully acknowledge the support of an NSF Graduate
Research Fellowship.

1
1 Introduction
1.1 High Frequency Trading
Over the past few decades, information technology has permeated and reshaped financial
markets. Several recent papers, such as Hendershott and Riordan (2009) [20] and Hender-
shott et al. (2011) [19] document the prevalence algorithmic trading in modern (electronic)
financial markets. The research of Hendershott et al. provides compelling evidence that
general algorithmic trading tends to improve liquidity and aid the price-discovery process.
However, algorithmic trading can potentially be used myriad ways, some of which are
socially desirable, and some of which may be deleterious to the public good.
Although the research of Hendershott et al. suggests that the positive effects of algo-
rithmic trading outweigh the negative effects in aggregate, this leaves open the possibility
that some subset of algorithmic traders is doing something harmful. The particular sub-
set of algorithmic traders that has attracted the greatest scrutiny in this regard are the
so-called “high-frequency traders” (“HFTs”).

1.1.1 Defining High-Frequency Trading

The term “high-frequency trader” lacks a precise definition, but the SEC [10] provides a
fairly standard characterization. Following the SEC’s terminology, HFTs are proprietary
trading entities1 that generate a large number of trades on a daily basis and are typically
attributed the following characteristics:
“(1) the use of extraordinarily high-speed and sophisticated computer programs
for generating, routing, and executing orders;
(2) use of co-location services and individual data feeds offered by exchanges
and others to minimize network and other types of latencies;
(3) very short time-frames for establishing and liquidating positions;
(4) the submission of numerous orders that are cancelled shortly after submis-
sion; and
(5) ending the trading day in as close to a flat position as possible (that is, not
carrying significant, unhedged positions over-night [when markets are closed])”
[10, pp. 45-46]
Hasbrouck and Saar (2011) characterize HFTs as a subset of proprietary traders who im-
plement “low-latency strategies”—strategies that respond to market events on a millisecond
timescale. Although not all low-latency strategies generate large numbers of trades2 , the
specific notion of speed that defines low-latency strategies is a quintessential element high-
frequency trading.
1
I use the term “entities,” because HFTs may be organized in a variety of ways. Although some HFTs
are organized as firms, others may be proprietary trading desks at some larger organization, and still others
might be organized as hedge funds. Cf. the SEC’s “Concept release on equity market structure,” Concept
Release No. 34-61358[10], page 45.
2
There are many ways to respond to a market event, such as revising a resting limit order, that do not
necessarily entail trading.

2
1.1.2 Empirical Results
High-frequency trading is notoriously difficult to study empirically, because data with suf-
ficient temporal resolution are typically anonymous. Although we can use 13-F forms to
track the behavior of institutional investors at a quarterly frequency, there is no general,
simple way to track the behavior of a HFT at second- or millisecond-frequency. Nev-
ertheless, there are at least three datasets3 in which HFT-activity can be distinguished
from non-HFT-activity, and analyses of these datasets have established some foundational
empirical results about HFTs.
Perhaps the most striking and consistent finding in this empirical literature is that
an enormous fraction of trading volume can be attributed to HFTs. Estimates vary, but
even the most conservative estimates suggest that HFTs account for more than 30% of
total trading volume in U.S. equities, and most estimates fall in the 50% − 70% range (see
[10, 3]). These trading volume figures suggest that HFT activity constitutes an important
facet of modern financial markets, which naturally raises the question of exactly how HFTs
affect markets.
Three recent studies—Hasbrouck and Saar (2011), Brogaard (2010), and Kirilenko et
al. (2010)—all examine the same general question of how HFT activity affects markets,
but each study takes a very different approach to answering this question. Hasbrouck and
Saar study order-level NASDAQ data, and they develop statistical techniques to identify
“strategic runs” of orders that they attribute to HFTs. In a more direct approach, Brogaard
uses novel dataset of order-level data for 120 stocks (60 listed on NASDAQ, 60 listed on
the NYSE); this dataset distinguishes messages from 26 firms that had been identified by
NASDAQ as engaging primarily in high frequency trading. Finally, Kirilenko et al. analyze
audit-trail, transaction-level data for the E-mini S&P 500 stock index futures market, from
May 3 to May 6, 2010. Kirilenko et al. use this data to sort over 15,000 trading accounts
into six categories on the basis of realized trading behavior (one category consisted of
HFTs, and the remaining categories consisted of different types of non-HFTs4 ).
Both Brogaard (2010) and Hasbrouck and Saar (2011) reach similar conclusions about
the effects of HFT activity. Hasbrouck and Saar conclude that increased HFT activity
tends to improve traditional measures of market quality such as short-term volatility and
spreads. Brogaard likewise finds evidence that HFTs may dampen intra-day volatility,
and that HFTs frequently provide inside quotes. Although Brogaard’s results suggest that
HFTs supply less additional book-depth than their inside-quote provision would typically
imply, he finds no evidence that HFTs flee the market in volatile times. Furthermore,
Brogaard finds that HFTs contribute significantly to the price-discovery process. Both
3
Specifically, the datasets are 1) the computerized trade reconstruction data from the Commodity
Futures Trading Commission (CFTC) that Kirilenko et al. (2010) [24] use to analyze the Flash Crash, 2)
Order-level data for 120 stocks (60 listed on NASDAQ, 60 listed on the NYSE) which flags the orders from
26 firms known by NASDAQ to primarily engage in high-frequency strategies, and 3) Transaction data
from the Deutsche Boerse that identifies whether or not each trade’s buyer and seller were participants in
the “Automated Trading Program”—i.e., whether the orders were generated by an algorithm. See [24, 11]
for details on (1), [3] for details on (2), and [14, 20] for details on (3).
4
Kirilenko et al. define “high-frequency trader” much more precisely than do either Hasbrouck and Saar
or Brogaard. The accounts that Kirilenko et al. classify as HFTs are archetypes of the SEC characteriza-
tion, but they are not necessarily close analogues of the “HFTs” that the other papers consider.

3
studies suggest that in aggregate, HFT activity tends to have a positive influence on
markets.
Whereas Brogaard (2010) and Hasbrouck and Saar (2011) examine HFT activity across
a large number of assets, over relatively long time spans, Kirilenko et al. focus on a single
asset during a short time span, because their primary interest is the role (if any) that HFTs
played in the Flash Crash of May 6, 2010. Kirilenko et al. conclude that HFT activity did
not actually trigger the Flash Crash, but HFTs did exacerbate market volatility through
their response to the unusually large selling pressure that day. These conclusions are
not incompatible with the broader findings of the other two studies, and some of the
more detailed results agree extremely well. For example, Brogaard’s analysis suggests that
rather than fleeing the market in volatile times, HFTs actually increase their participation
somewhat; Kirilenko et al. find that HFT trading volume increased dramatically in both
absolute and relative terms during short interval on May 6 when the largest price changes
occurred. Nevertheless, the discrepancy between the rosy conclusions of the two more
general studies and the darker conclusions of the more specific study highlight the need
for more thorough and detailed understanding HFT activities.

1.1.3 What HFTs are Actually Doing

Crucial to understanding high-frequency trading is the general question of how speed re-
lates to information. An important point of agreement across empirical studies is that
HFTs appear to possess some sort of informational advantage over other market partici-
pants. For example, Brogaard finds strong evidence that HFTs avoid providing liquidity to
informed traders, while Kirilenko et al. find that HFTs tend to trade in the same direction
as contemporaneous price changes—i.e., they can anticipate future prices. The popular
notion that “HFTs are better informed because they are fast” begs the question of why
speed should lead to superior information.
The standard explanation of the relationship between speed and information is that
HFTs are able to react to new public information faster than other market participants, so
that new public information briefly serves as private information for the HFTs [9, 21, 22, 2,
7]. However, as Hasbrouck and Saar (2011) establish in considerable detail, “at horizons of
extreme brevity, however, there is simply not sufficient time for an agent to be reacting to
anything except very local market information” [18, pp. 8]. Conceptually, the mechanism
underlying the standard explanation amounts to picking off stale quotes. Mcinish and
Upson (2011) estimate that HFTs’ direct gains from exploiting this kind of differential
in U.S. markets constitutes around 10% of their total annual profits; this figure is not
negligible, but it highlights the inadequacy of the standard explanation of the relationship
between speed and information.
Empirically, Brogaard finds that the aggregate trading of the 26 HFTs in his sam-
ple, viewed as the behavior of a single “representative HFT,” is consistent with a strategy
based on order-imbalance-driven price-reversals. Similarly, Kirilenko et al. conclude that
the HFTs in their sample exhibit trading patterns that are consistent with some notion of
market-making. However, this general resemblance to traditional market-making does not
resolve the mystery of exactly what HFTs are doing. First, various HFTs are believed to

4
employ a variety of different strategies5 , so these aggregate results potentially occlude im-
portant heterogeneity. More importantly, although traditional, formally-registered market-
makers typically have unique information about order-flow, almost no HFTs are formally
registered as market-makers [10]. Even if each HFT behaves exactly like a traditional
market-maker, the puzzle of how they use speed as a replacement for privileged order-flow
information would remain.
Many of the general techniques/strategies disclosed by and imputed6 to HFTs are
standard elements of the non-high-frequency realm. These include arbitrage, statistical
arbitrage, directional bets, and market-making. However, the high-frequency activities
commonly classified as “liquidity detection” lack low-frequency analogues. This classifica-
tion covers a number of slightly different techniques— “pinging,” “sniffing,” etc.—but all of
these techniques fundamentally entail the use of orders for the express purpose of gathering
information about the market. Obviously, such techniques point to a mechanism by which
HFTs might obtain superior information. Less obviously, but much more importantly, this
mechanism illuminates the connection between information and trading speed.

1.2 Trade-Revealed Information

I ultimately seek to address how trading speed could translate to superior information.
As a first step, I will examine the idea of “exploratory trading”—how an agent might
use his own trades to gather valuable information. A central objective of my analysis is
to establish that exploratory trading relates naturally and intimately to high-frequency
trading. Although exploratory trading may arise in a number of contexts, I frame my
model in a potentially high-frequency setting, and focus my analysis on the implications
of exploratory trading for understanding high-frequency trading.
The idea that the trading process can reveal new information about an asset dates
back at least to Romer (1993). Romer proposed this idea to explain why efficient asset
prices could change dramatically even in the absence of external news. Subsequent research
extended Romer’s basic idea to address a variety of other aspects of asset-price dynamics.
At a very general level, this literature explores how the information revealed by some set
of initial trades affects the subsequent trades of rational agents. In this paper, I consider
a converse issue, namely how optimal initial trades depend on the amount of information
that those trades are expected to reveal.
An agent who trades an asset ultimately cares about the prices at which she actually
executes trades. Conceptually, we can decompose the price at which an agent executes
a trade into a component that depends on the agent’s trade, and a component that does
not depend on her trade (I will refer to this latter component as the “asset-intrinsic”
component). More concretely, we might decompose a realized price into the price impact
of the agent’s actual trade, and the price that would have prevailed if the agent hadn’t
5
Brogaard notes that some of the 26 HFTs in his sample primarily take liquidity, while others primarily
provide liquidity. Although Brogaard cannot observe this heterogeneity directly in his data, the contacts
at NASDAQ who constructed the dataset conveyed this important fact. Publications by both the United
States Securities Exchange Commission and the Australian Securities and Investments Commission also
indicate considerable diversity in HFT strategies. [10, 30]
6
Cf. publications of various market regulatory bodies, [10, 30]

5
traded.
In principle, the trading process could reveal information about either the future asset-
intrinsic component of price, or the future price-impact of trades. However, essentially
all of research in the spirit of Romer (1993) focuses on the question of how past trading
can reveal information about the asset-intrinsic component of future prices. Although the
revelation of information about price impact is a topic of both practical and theoretical
importance, the mechanism by which past trading reveals information about future price
impact is relatively transparent. The sole paper on this topic, Hong and Rady (2002),
covers this basic mechanism fairly comprehensively.
The theoretical question that I address in the present paper—how optimal initial trades
depend on the amount of information that they are expected to reveal—is strictly more
complicated than the analogous, converse, Romer (1993)-style questions. In the simplest,
two-period setting, the Romer (1993)-style question must be addressed for period 2, but
then to answer my style of question, we must determine how optimal period-2 trading
profits depend on the amount of information that period-1 trade reveals. Next, we must
role back to period 1, solve a highly non-standard inference problem to find the relationship
between the period-1 trade and the amount of information the trade is expected to reveal,
use this to express the expected conditionally optimal period-2 trading profits in terms of
the period-1 trade, and then combine this with the expected direct trading profits from
the period-1 trade, and finally maximize the expected total profit with respect to the
period-1 trade. The very simplicity that makes learning about price-impact unattractive
for studying Romer (1993)-style issues makes “learning about price impact” an ideal setting
to develop a baseline model of exploratory trading.
Hong and Rady (2002) briefly discuss the topic that I address:

“In [a modification of our] set-up, the first-period action impacts on second-

period profit because it affects the informed investors’ second-period belief
about liquidity ... Informed traders may have an incentive to experiment and
sacrifice first-period trading profit for more precise information about liquidity,
which may improve the profitability of their second-period trades. Unfortu-
nately, even this simple set-up cannot be solved in closed form.” (pp. 427)

By using a slight variation on the Hong and Rady (2002) model (which is in turn a slightly
modified version of the Kyle (1985) model), I derive a closed-form solution for optimal
exploratory trading, I formalize Hong and Rady’s conjecture that informed traders may
have an incentive to engage in costly experiments in period 1 to increase their expected
profits in period 2, and I establish conditions under which the conjecture holds.
With the inner workings of exploratory trading laid bare, the connections between
exploratory and high-frequency trading follow easily. These connections shed light on the
general question of how trading speed relates to superior information, and they also suggest
explanations for a variety of existing empirical findings, as well as generating a number of
testable predictions.

6
2 A Model of Exploratory Trading
For purposes of tractability and expositional clarity, I initially consider exploratory trading
in the context of learning about price-impact. Although this is not necessarily the most
interesting application, the formal and conceptual results that I derive about exploratory
trading in this context extend to more general settings.

2.1 The Baseline Model

Let time be discrete, consisting of periods t = 1, 2. Consider the market for a single asset.
At the end of period 2, the asset pays a fixed terminal liquidation dividend of µ and the
world ends.
The net supply of the asset in period 1 is given by the random variable −s1 , which has
zero mean and finite variance σs2 > 0. For simplicity, assume that the net supply of the
asset in period 2 is exactly zero.
The respective market demand curves for the asset in periods 1 and 2 are

y1 = α − β −1 p1 (1)
−1
y2 = α − β p2 (2)

where pt denotes the price in period t7 . The parameter β is drawn from a distribution
with strictly positive support, bounded away from zero, with mean b and finite variance
σβ2 > 0; intuitively, β represents marginal price-impact of a market order. The parameter
α depends on β and follows a distribution such that
µ+ξ
α≡ (3)
β

for some zero-mean random variable ξ ∈ L1 that is independent of β and has finite variance
σξ2 > 0. The product αβ, which I will denote by p0 ≡ αβ, corresponds to the price at
which yt = E [st ]. Equation (3) implies that p0 ≡ µ + ξ, and the distributional assumptions
on ξ imply that E [p0 |β] = E [p0 ] = µ.
Consider a single agent—call him the “high-frequency trader,” or “HFT”—who submits
market orders8 in periods 1 and 2 with the objective of maximizing his expected aggregate
net profits. Denote by xt the number shares that the HFT purchases in period t. Assume
7
From a theoretical standpoint, it might be nicer to have a second-period demand curve of the form
y2 = α − β −1 p2 − y1 , but this would make the algebra much messier. To the extent that the interesting
features of this model revolve around the estimation of β, the algebraically simpler case still delivers the
same intuitive conclusions.
8
Nothing fundamental would change if the HFT used limit orders instead of market orders. If we assume
both that there is no supply noise in period 2, and that the HFT knows µ − p0 perfectly, then allowing
the HFT to use limit orders would lead to a trivial solution wherein the HFT offers to purchase/sell an
unlimited quantity at the price p0 + µ−p2
0
. However, as long as the HFT is uncertain about either the net
supply in period 2, or the true value of µ − p0 , allowing the HFT to submit limit orders would not produce
any pathologies. The HFT’s optimization problem would be less intuitive and far less tractable, but the
same basic concepts would apply.

7
that the HFT knows µ and the product p0 ≡ αβ 9 , but not α or β individually. The HFT
observes his own trades, as well as the market-clearing price in each period, but he observes
neither the market demand curves, nor the quantity s1 . When the HFT observes p1 he
can use his knowledge of x1 and p0 to estimate β, but his uncertainty about s1 prevents
him from learning β exactly.

2.2 How Chronological Time Enters the Model

Although the model is set in abstract discrete time, true chronological time enters the model
in an important, albeit implicit, way. The HFT’s inference problem basically amounts
to determining what portion of a price change was caused by his own trade, and what
portion was driven by something else. Random supply noise is a simple “something else,”
but many other factors could plausibly confound the relationship between the HFT’s trade
during some time interval and the price change in that interval. However, the prevalence
and comparative importance of such confounding factors must tend to decrease as time
intervals shorten10 .
As a concrete example, suppose that I sell 10,000 shares of AAPL sometime in October.
If I compare Apple’s opening price on the first trading day in October with its opening price
on the last trading day on October, this would not tell me very much about market-depth
around the time of my trade. By contrast, if I compared the price of AAPL 10 milliseconds
before and after I sold 10,000 shares, I would probably learn something meaningful about
market-depth around the time of my trade.
The role of true, chronological time in the model will ultimately illuminate the re-
lationship between exploratory and high-frequency trading. I address this deeper, more
intrinsically interesting issue after solving the baseline model. The formal framework that
emerges from solving the baseline model facilitates more precise discussion this central
issue.
9
The HFT’s trade at date 1 affects the quality of his information about the slope of the market demand
curve depends on his period-1 trades, the quality of his information about the intercept does not. Uncer-
tainty about the intercept does not raise any fundamentally new issues, so streamline my analysis, I will
essentially assume that the intercept is known. Rather than literally assuming that the HFT knows α, it
turns out to be more natural to suppose that he knows p0 ≡ αβ. Whereas the lone parameter α does not
have an obvious economic meaning, the product αβ corresponds to the market-clearing price that would
prevail in the absence of supply noise if the HFT did not participate in the market. The distributional
assumptions about β and α imply that β is independent of p0 , so the HFT’s knowledge of p0 removes the
extraneous uncertainty from the model without introducing additional information about β.
10
This claim can be made rigorous. Suppose that ´ the price change due to confounding factors during
some bounded interval I can be represented as I f (t) dt, where f : R �→ R is a measurable function
that is integrable on all subsets of R with finite measure. Let I1 , ..., In be non-overlapping sub-intervals
n
of I such that ∪n j=1 Ij = I (i.e. let the collection of intervals {Ij }j=1 be a partition of I). The av-
erage magnitude ˛´ of the price-change
˛ (due to confounding factors) on each sub-interval in this collec-
P ˛ ˛ Pn ´
tion is n1 n 1
j=1 Ij |f (t)| dt = n I |f (t)| dt. By assumption, I |f (t)| dt < ∞, so
1
´ ´
j=1 ˛ Ij
f (t) dt ˛ ≤ n
n´ o
1
|f (t)| dt → 0 as n → ∞. Note that if supj I 1dt ≤ ε I 1dt, it follows that n ≥ 1ε . Thus the
´ ´
n I j
average magnitude of the price-change on each element of a partition P of the interval I approaches zero
as the measure of the largest element of P approaches zero.

8
3 Solving the Model
Intuitively, the HFT’s basic strategy entails buying (selling) the asset when the price is
below (above) the terminal value µ. However, the HFT faces downward-sloping market
demand curves, so the market-clearing price pt depends on the HFT’s purchase xt . Con-
sequently, the HFT faces a trade-off between the number of shares that he buys, and the
spread µ − pt that he earns on each share. If the HFT knows the shape of the market
demand curve, then his optimization problem is isomorphic to that of a profit-maximizing
monopolist facing a downward-sloping demand curve.
If the HFT does not know the shape of the market demand curve, his task is more
complex than the standard monopolist’s problem. Although the HFT still faces the quan-
tity/spread trade-off, his ability to optimally balance this trade-off depends on the quality
of his information about the market demand curve. The price impact of the HFT’s trade
in the first period provides information about the slope of market demand curve that the
HFT can use to better choose his trade in period 2, but this information comes at the
expense of trading costs in the first period.
In the remainder of this section, I derive the HFT’s optimal trading strategy through
backward induction.

3.1 Date t = 2
The HFT pays x1 p1 at date 1, and x2 p2 at date 2, then he receives a payoff of µ (x1 + x2 )
at the end of period 2, so the HFT’s total realized profit is

Πtotal = −x1 p1 − x2 p2 + µ (x1 + x2 )

= x1 (µ − p1 ) + x2 (µ − p2 )

At date 2, the HFT chooses x2 to maximize the conditional expectation of his total
profit, E1 [x1 (µ − p1 ) + x2 (µ − p2 )]. Since x1 and p1 are determined before the second
period, the HFT’s choice of x2 depends only on his conditional expectation of his period-2
trading profits, E1 [x2 (µ − p2 )]. Thus the HFT solves

max E1 [x2 (µ − p2 )]
x2
s.t. p2 = p0 + βx2

The HFT’s feasible optimal trade at date 2, call it x̂2 , is given by

µ − p0
xˆ2 = (4)
2E1 [β]

Equation (4) confirms our basic intuitions about the HFT’s trading strategy; he trades
against perceived price dislocations, but he accounts for the anticipated price impact of
his trade.
Next, we wish to determine how the profit that the HFT earns from trading xˆ2 depends
on the quality of his information about the demand curve. Let x∗2 ≡ µ−p 2β denote the
0

infeasible optimal trade that the HFT would select if he knew the true value of β, and let

9
π2∗ ≡ β (x∗2 )2 denote the associated infeasible maximized profit. Intuitively, as xˆ2 deviates
from the infeasible optimum x∗2 , we should expect the HFT’s realized profits to decline
relative to π2∗ . Indeed, in the appendix I show that the true profit that the HFT would
earn from the optimal feasible trade x̂2 can be expressed as

πˆ2 = π2∗ − β (x̂2 − x∗2 )2 (5)

The HFT’s expected feasible profit at a given value of β can be expressed naturally in
terms of the mean-square error of E11[β] relative to β1 , and approximated in terms of the
conditional variance of E1 [β]:
�� � �
∗ β (µ − p0 )2 1 1 2
E [π̂2 |β, p0 ] = π2 − E − |β (6)
4 E1 [β] β
(µ − p0 )2
≈ π2∗ − V ar (E1 [β] |β) (7)
4β 3

Equation (6) illustrates how variability in the HFT’s estimate of the market demand curve
reduces the profits that the HFT earns from his optimal feasible trading strategy in period
2.

3.2 Date t = 1
In section (3.1), I related the HFT’s optimal period-2 trading strategy and associated
profits to the estimator E1 [β]. In particular, I showed that the HFT’s expected profits
from trading in period 2 increased as E1 [β] became a better estimator of β. As I will show
below, the quality of E1 [β] as an estimator of β depends on x1 , the HFT’s order in period
1. Consequently, the HFT’s optimal trading strategy in period 1 will depend not only on
the direct revenues associated with the trade, but also on the extent to which the trade is
expected to improve the HFT’s information about the market demand curve in the next
period.

3.2.1 Exploratory Trading at Date 1

I now consider the case in which the HFT can use the information from his period-1
trade to update his beliefs about β. The key idea is that the HFT’s purchase in period
1 induces predictable variation in the market-clearing price that the HFT can exploit to
better estimate the slope of the market demand curve.
The market-clearing price in the first period, p1 , can be expressed as

∆p = β (x1 + s1 ) (8)

where I define ∆p ≡ p1 − p0 .
Since the single observation (∆p, x1 ) constitutes the entirety of the HFT’s empiri-
cal data, β is underidentifed from the HFT’s perspective, regardless of the value of x1 .
However, the underidentification of β means simply that the HFT cannot perfectly (i.e.,
consistently) estimate β from a single, noisy observation. Although the traditional binary

10
identified/underidentified classification is well-suited for asymptotic analyses, the present
setting calls for finer distinctions.
No finite choice of x1 will allow the HFT to completely disentangle the effects of β from
those of supply noise on the basis of a single observation, but the value of x1 determines
how well the HFT can separate the effects of β from those of s1 . Intuitively, we might think
of x1 determining “how well” β is identified. We can make this intuitive notion precise by
using equation (6), and considering how x1 affects the HFT’s expected period-2 profits.
The exact effects of x1 will depend both on the distributions of β and s1 , and on the
HFT’s knowledge about these distributions, but the HFT will generally tend to learn more
about β the larger is the magnitude of x1 . Although I cannot invoke a central limit theorem
to sidestep these distributional details in the usual manner, I accomplish something similar
by considering the case in which |x1 | becomes large, and applying integrability/moment
conditions to characterize tail behavior. We can always bound the variance of E1 [β] by
� �
(σ2 +b2 )σ2
E (E1 [β] − β)2 ≤ β x2 s , and under mild regularity conditions, this bound becomes
1
tight as |x1 | becomes large (see mathematical appendix for details). To avoid a morass
of unenlightening algebra, I will appeal� to this tight
� bound and make the simplifying
2
assumption that V ar (E1 [β] |β) ≡ E (E1 [β] − β) |β is given by

β 2 σs2
V ar (E1 [β] |β) = K 2 (9)
x21

for x21 ≥ K 2 σs2 , where 1 ≥ K > 0 is some positive constant that depends on the uncondi-
tional distributions of β and of s1 .
We can combine (9) with the approximation (6) to characterize the relationship between
x1 and the HFT’s expected period-2 profits. Taking (6) to hold exactly11 , we obtain
� �� �
(µ − p0 )2 E β −1 K 2 σs2
E [π̂2 |p0 ] = 1− (10)
4 x21
Since the HFT could always choose not to trade at all in period 2, his expected period-2
profits must be non-negative. The right-hand side of equation 10 is negative for x21 < K 2 σs2 ,
so the model is not applicable in that region (this is why I only assume that (9) holds for
x21 ≥ K 2 σs2 ).
The expected direct trading profit from trading x1 is

E [π1 |p0 ] = (µ − p0 ) x1 − bx21 (11)

so the HFT’s total expected profit over both periods is

� �� �
(µ − p0 )2 E β −1 K 2 σs2
E [π1 + π̂2 |p0 ] = (µ − p0 ) x1 − bx21 + 1− (12)
4 x21
The optimal value of x1 depends on three factors. First, trading costs in the form price
impact tend to push the optimal value of x1 towards zero; in equation (12), trading costs
11
Since we could adjust the positive constant K to make the approximation good, ignoring the approx-
imation error is more innocuous in this case than it is in general.

11
enter through the “−bx21 ” term. Second, as the HFT trades more at date 1, he obtains
better information about β that he can use to trade more profitably (in expectation) at
2 2
date 2; in equation (12), the “− Kx2σs ” term reflects this informational benefit to trading.
1
This informational benefit pushes the optimal value of |x1 | away from zero. Finally, the
direct gains from trading in period 1, reflected by the “(µ − p0 ) x1 ” term in equation (12),
tend to push the optimal value of x1 away from zero.
Unlike the informational gains from trading, the direct gains from trading in period 1
are completely standard and are thoroughly understood. We can isolate the informational
motive for trading from the “direct gain” motive by supposing that the HFT only observes
µ − p0 after he has selected x1 . Recall that I assume E [µ − p0 ] = 0, so the HFT’s initial
expectation of his total profit is
� �
2 E β −1 � �
σ ξ K 2 σs2
E [π1 + π̂2 ] = −bx1 +
2
1− (13)
4 x21

Intuitively, (13) represents the profit that the HFT would expect to obtain if he had to
select the value of x1 before learning p0 . Since E [µ − p0 ] = 0, the “(µ − p0 ) x1 ” term from
(12) vanishes—in expectation, there is no direct gain from a “blind trade”.
In the mathematical appendix, I show that the x∗1 that maximizes the HFT’s uncondi-
tional expected profit, (13), is characterized by
 �
 1 σ σ K E[β −1 ] if σ 2 < E[β −1 ] σ 2
ξ s
(x∗1 )2 = 2 b s 16K 2 b ξ
E[β −1 ] 2
(14)
0 2
if σs ≥ 16K 2 b σξ

E[β −1 ]
The condition σs2 < 16K 2 b σξ2 ensures that the maximized expected profit is non-negative
(if the maximized expected profit is negative, HFT would simply not participate in the
E[β −1 ]
market). Also, the condition “σs2 < 16K 2 b σξ2 ” implies (x∗1 )2 > 2K 2 σs2 > K 2 σs2 , and
therefore guarantees that we are in a valid region of our model. Since I have removed the
“direct gain” motive for trading in period 1, the optimal trade characterized in equation
(14) is driven entirely by information-seeking concerns.
To rephrase this more bluntly, equation (14) validates the concept of exploratory trad-
ing by explicitly characterizing the phenomenon in a simple model. However, the concept
of exploratory trading is merely a tool to better unravel the mysteries of high-frequency
trading. Although equation (14) marks the end of my analysis of exploratory trading in the
abstract, it also marks the beginning of my analysis of the connection between exploratory
and high-frequency trading.

4 Exploratory and High-Frequency Trading

In section 2.2, I showed how chronological time implicitly appeared in the baseline model
through the distribution of supply noise, s1 . The discussion in 2.2 introduced the idea that
the HFT wants to determine the causal effect of his trade on prices, but that confounding
factors such as supply noise hinder his inference. The main point of section 2.2 is that the

12
“importance” (in some sense) of these confounding factors depends on the clock-time dura-
tion of the interval in which trade occurs. This point remains valid, but we can now state
it more precisely. As equation (10) reveals, the HFT’s expected period-2 profit depends
on σs2 , the variance of period-1 supply noise. The relevant measure of the “importance” of
period-1 supply noise is simply the variance σs2 .
The link between the parameter σs2 and the implicit duration of period 1 makes it pos-
sible to investigate chronological-time considerations elsewhere in the model. In particular,
we can use the results from section 3.2 to examine the relationship between exploratory
trading and chronological time in detail, which will in turn reveal the connections between
exploratory trading and high-frequency trading.

4.1 Speed is Necessary for Exploratory Trading

Equation (14) (reprinted below for convenience) suggests a simple but natural connection
between exploratory and high-frequency trading.
 �
 1 σ σ K E[β −1 ] if σ 2 < E[β −1 ] σ 2
ξ s
(x∗1 )2 = 2 b s 16K 2 b ξ
E[β −1 ] 2
0 2
if σ ≥ σ
s 16K 2 b ξ

E[β −1 ]
When the variance of supply noise exceeds some threshold (σs2 ≥ 16K 2 b σξ2 ), the optimal
level of exploratory trading drops to zero ((x∗1 )2 = 0). Exploratory trading only arises in
the model when the variance of supply noise is sufficiently small, or equivalently when the
duration of the first trading period is sufficiently short.
Although a trade reveals some amount of valuable information, trading is also costly
(due to price impact). Both the informational gain and the price-impact cost of a trade de-
pend on the trade’s magnitude, but only the informational gain depends on the variance of
supply noise. As the variance of supply noise increases, the informational gain from a given
magnitude of trade decreases. Up to a point, a trader can partially offset this reduction
in the informational gain by increasing the magnitude of his trade, but this also increases
his trading costs. Eventually, when σs2 becomes sufficiently large, the informational gains
become too small to justify the cost of any non-zero level of exploratory trade.
The relationship between exploratory trading and high-frequency trading can also be
2 2
understood in terms of equation (13). As noted earlier the “− Kx2σs ” term reflects the
1
informational gain from trading in period 1, expressed in terms of the HFT’s expected
2
total profit. There are two ways to make σxs2 small: make x21 large, or make σs2 small. Up to
1
this point, we have treated σs2 as fixed, and considered the optimal choice of x21 . However,
to the extent that σs2 depends of trading speed, there is another dimension along which
optimization is possible.
This raises two important points. First, the costs associated with reducing σs2 by in-
creasing trading speed are largely fixed; these include direct data feeds, colocation services,
some proprietary software development, and so on. These fixed costs of increasing speed
are considerable, and the cost reduction per trade would likely be small, increased trading
speed would be most valuable (from the standpoint of exploratory trading) for a trader

13
who intended to engage in a large number of trades. Hence the clear connection between
exploratory trade and low-latency trading also suggests a similar connection between ex-
ploratory trade and high-frequency trading per se.
The second important point that arises from the two possible approaches to making
σs2
x21
small is that in spite of the potential value associated with superior trading speed, this
does not necessarily imply that we should observe a Bertrand-competition-style latency
arms race (or at least not one driven by exploratory trading). An HFT could potentially
overcome some minute latency disadvantage by slightly increasing the magnitude of his
exploratory trades.

4.2 Speed is Basically Sufficient for Exploratory Trading

If the HFT trades the quantity x1 = �, the expected informational gain from this trade
depends on |�|, but the reasons why the HFT selected the quantity x1 = � are completely
irrelevant. Earlier, in order to establish and clarify the theoretical notion of exploratory
trading, I isolated the informational motive for trading from the standard “direct gain”
motive. The optimal exploratory trade that I characterize in equation (14) and analyze in
section 4.1 is the theoretically pure variety, which excludes any “direct gain” component.
In this subsection, I consider the potential informational value of trading in a more general
context.
An important preliminary result is that in the baseline model, the HFT’s information-
seeking motives for trading will never conflict with his direct-gain motives. Since trading
costs are directly proportional to −x21 , and the informational benefits of trading are in-
versely proportional to −x21 , these two factors can uniquely determine the optimal magni-
tude of x∗1 , but they do not determine the optimal sign of x∗1 —i.e., whether it is optimal
to sell |x∗1 | shares, or to buy |x∗1 | shares. If the HFT knew the sign of µ − p0 , he could
obtain direct gains from his first-period trade by choosing x∗1 to have the same sign. Given
this choice of sign, both the informational benefit factor and the direct gain factor push
the magnitude of x∗1 in the same direction. If the HFT considers both direct gains and
informational gains when he selects x1 , all of our earlier results remain qualitatively un-
changed12 .
Even if the HFT doesn’t consider informational gains when he selects x1 , he can still
extract information from the result of his trade. The trade generates some information,
regardless of whether the HFT chose x1 optimally or arbitrarily. Furthermore, the infor-
2
mational gain associated with x1 is still proportional to σxs2 , so increased trading speed
1
still translates to increased informational gains, exactly as discussed in section 4.1. The
key implication of the preceding results that novel information is a natural by-product
of high-frequency trades. Regardless of whether a HFT actually chooses his trades with
the intent of gathering information, those trades still generate non-public, non-negligible
information.
12
However, if we introduce direct gains from period-1 trading, the HFT’s optimization problem becomes
much less tractable, because the quartic equation that arises in the first-order condition no longer has a
simple biquadratic form.

14
5 Discussion
5.1 Relation to Empirical Findings
The preceding analysis of exploratory trading sheds light on a variety of empirical results
concerning HFT activity.

5.1.1 Cross-sectional Variation of HFT Participation in Stocks

Brogaard (2010) finds that fraction of total trading activity for which HFTs are responsi-
ble varies systematically across stocks. In particular, Brogaard finds that HFTs’ relative
fraction of market activity tends to be greatest in large market-cap stocks, and stocks with
greater market depth. To the extent that we can characterize such stocks as those for
which the price-impact of trading is small, the exploratory trading model helps to explain
Brogaard’s finding.
As equation (14) reveals, the optimal magnitude of exploratory trading increases as
the expected price-impact parameter (b ≡ E [β]) decreases. In markets where the ex-
pected price-impact is small, the expected cost of exploratory trading is also small, so
the optimal magnitude of exploratory trading is large, and HFTs obtain a greater amount
of information. Thus in addition to the direct increase in HFT activity associated with
exploratory trading, the improved information might also encourage HFT participation.
Consistent with such “improved information,” Brogaard finds that HFTs are best able to
avoid providing liquidity to informed traders in large-cap stocks.

5.1.2 Avoiding Informed Traders and Anticipating Price Changes

In absolute terms, Brogaard finds that HFTs avoid providing liquidity to informed traders
more successfully than do non-HFTs. As noted earlier, this finding suggests that HFTs
tend to know some things that non-HFTs do not. The baseline exploratory trading model
is not directly applicable to this setting, but conceptually simple extension of the baseline
model offers a natural explanation.
In section I of his (1993) paper, Romer presents a model in which (essentially) agents
know the precision of their own information about an asset, but they are unsure whether
their is more precise or less precise than that of other agents. Following some unexpected
but observable exogenous supply shock, agents observe the slope of the market demand
curve from which they can infer the relative precision of their signals. The problem of
determining whether other agents’ information is more or less precise than your own is
basically isomorphic to the problem of detecting informed traders, and Romer establishes
a framework in which the slope of the demand curve provides the information necessary to
solve this problem. The baseline exploratory trading model directly addresses the issue of
estimating the slope of the market demand curve. A natural explanation of how trading
speed would help HFTs detect informed traders requires little more13 than appending
Romer’s model to the baseline exploratory trading model.
13
There are a few uninteresting but potentially tedious technical details that arise in a strictly rigorous
concatenation of the two models, and I am actively working out these minutiae, but the intuition is
straightforward.

15
 The “exploratory trading + Romer” extension also helps to explain the finding of Kir-
ilenko et al. that HFTs appear to profitably trade in the same direction as contemporaneous
price changes (i.e., anticipate price changes). When agents deduce the relative precision
of their signal, they can use this information to infer whether the current market-price
properly reflects aggregate information that is appropriately weighted by its precision, and
thereby potentially uncover temporary mispricings.

5.1.3 Fleeting Orders

In their 2009 paper, “Technology and liquidity provision: The blurring of traditional defi-
nitions,” Hasbrouck and Saar (henceforth “HS”) investigate what they term “fleeting limit
orders,” that is, limit orders that are cancelled within two seconds of being placed. Among
the explanations of the existence of fleeting limit orders that HS consider is the hypothesis
that “fleeting orders are a byproduct of a strategy meant to ‘search’ for latent liquidity...the
search hypothesis implies that fleeting orders are intended to demand, rather than supply,
liquidity” (pp.154). HS find a variety of empirical evidence that supports their “search
hypothesis.”
Notwithstanding the limit order vs. market order difference, rapidly submitting liquidity-
demanding demanding orders for the purpose of learning about latent (i.e., hidden) liq-
uidity bears an obvious resemblance to engaging in exploratory trading for the purpose
of estimating the current slope of the market demand curve. While HS’s results do not
definitively establish the empirical importance of exploratory trading of precisely the form
I analyze in the present paper, their results clearly support the empirical validity of the
general exploratory trading concept.
In the opposite direction, the theory of exploratory trading that I develop in this paper
may prove useful for developing structural models of exactly how fleeting orders arise from
various trading strategies.

5.1.4 Competition Among HFTs

An interesting stylized fact about the HFT industry is that a relatively small number of
HFTs account for a vast majority of HFT activity. For example, Kirilenko et al. character-
ize only 16 of the 15,000+ accounts in their sample as HFTs. Likewise, only 26 HFTs are
identified in Brogaard’s sample. Although Brogaard notes that these group of 26 excludes
some entities that are often classified as HFTs, the 26 identified HFTs are involved in 68.5%
of the total dollar volume in Brogaard’s sample. All remotely plausible estimates of total
the total U.S. dollar volume in which HFTs are involved are below 80%, so omissions from
the 26 HFTs identified in Brogaard’s data are unlikely to be important in the particular
sample that Brogaard studies14 .
While there are some fixed costs to starting a HFT firm, (colocation, direct data-feeds,
etc.), these costs are on the order of a few hundred thousand dollars annually. Furthermore,
HFTs require relatively little operating capital—Brogaard estimates that the 26 HFTs in
his sample, together, require around $117 million to conduct their trades on the 120 stocks.
14
This does not necessarily mean that HFT entities excluded from Brogaard’s 26 are unimportant for
other stocks/assets, or even for the stocks that Brogaard considers at times not covered in his sample.

16
By comparison, Brogaard estimates the annual profit of the 26 HFTs from trading the 120
stocks is around $74 million15 .
Once again, the baseline exploratory trading model provides some insight. First, the
result that all high-frequency trades generate some form of potentially valuable information
suggests that there are natural economies of scale for HFTs. Roughly speaking, a HFT
who trades more gets better information.
We can take this analysis still further by noting that if there were more than one HFT
in the exploratory trading model, then one HFT’s trades would look like supply noise to
the other HFTs (and vice versa). If there is one incumbent HFT, then his trades would
look like noise to a potential entrant, and at least reduce the incumbent’s prospective
profits. As the number of HFTs in a given market increases, the apparent supply noise
from the perspective of a potential entrant also increases. In other words, high-frequency
trading inherently imposes barriers to entry. An interesting possibility that arises from
this conclusion is that HFTs might engage in excessive trading for purely anti-competitive
purposes16 .

5.2 Testable Predictions

The baseline exploratory trading model generates a number of testable predictions.
Although the exploratory trading model provides a possible explanation for the system-
atic cross-sectional variation in HFT participation, it also makes much crisper predictions
about how HFT profits should depend on model parameters. In particular, in the cross-
section, HFT profits should decrease in σs2 , decrease in E [β], and increase in σξ2 . HFT
profits should also tend to be greatest in stocks with higher relative levels of undisplayed
liquidity. All of these cross-sectional predictions above have direct time-series analogues.
A particularly interesting issues is how changes in σs2 over time affect HFT profits. While
distinguishing σβ2 from σs2 may require some subtlety, HFT behavior in response to the two
types of change would help to isolate the importance/value of the information that HFTs
may derive from their own trades.
The audit-level CTR data that Kirilenko et al. analyze would be particularly useful
for empirical analyses of these predictions.

5.3 Discussion of Modeling Assumptions

Three assumptions of the baseline exploratory trading model merit some discussion.

5.3.1 Exogenous Private Information

For purposes of tractability, the baseline exploratory trading model in this paper analyzes
the optimal manner in which an agent can use his own trades to learn about the price-
impact of his trades. To motivate the value of such price-impact information, I consider
15
Note that these figures ($117 million and $74 million) correspond to trading in the specific stocks in
Brogaard’s sample. Brogaard extrapolates that the required capital and profits for the entire U.S. equities
market are $4.68 billion and $2.8 billion, respectively.
16
I thank John Campbell for pointing out this intriguing issue.

17
an agent who has some private knowledge about future prices. In terms of the model, I
examine how the HFT can best make use of his knowledge of µ − p0 during periods 1 and
2, but I take the HFT’s knowledge of µ − p0 to be exogenously given. I assume that the
HFT exogenously knows µ − p0 purely to simplify the exposition, and to focus on the novel
aspect of my model. In other words, this assumption is not crucial !
In earlier drafts of this paper, the HFT inferred both the intercept and the slope, but
inference about the intercept turns out to be a standard type of problem, and it complicates
the more interesting and unusual inference about the slope. Another, potentially more
attractive way to remove the “exogenous private information” assumption is to tack a
slight variation of the model in section 1 of Romer (1993) onto the basic exploratory
trading model. I am currently working on the details of this extension.
Alternatively, the baseline exploratory trading model can be viewed as an extension of
the Kyle (1985) model to a setting in which the informed trader is uncertain about market
depth, and the market-maker doesn’t act quickly enough to alter his initial strategy. This
is not necessarily an attractive option if our ultimate goal is to use exploratory trading to
illuminate the inner workings of high-frequency trading, but it is is reasonable if we simple
want to think about exploratory trading for its own sake.

5.3.2 Limited Information

Limitations on what the HFT can directly observe are an indispensable component of the
exploratory trading model. Clearly, if the HFT could observe the market demand curve
directly, he would have no reason to engage in exploratory trading. Similarly, if the HFT
could directly observe s1 , the net supply in period 1, he could perfectly infer the value of
β without recourse to exploratory trading.
While the rationale for imposing these limitations on what the HFT can observe is
obvious from a modeling standpoint, such limitations are slightly harder to motivate from
an empirical perspective. In many markets, traders can observe the limit-order book,
which seems rather similar to observing the market demand curve. However, while the
order book certainly contains some information about the market demand curve, it does
not perfectly reveal the true market demand curve. In the context of a real market, it
might be more precise to think of exploratory trading revealing information about “the
residual component of market demand not explained by the order book” rather than “the
market demand curve” per se, but ideas are isomorphic. In fact, if we assume in the model
that the HFT conditions all of his beliefs β on some not-perfectly-informative order book,
our earlier analyses still hold exactly, without any modification17 .
Limiting the HFT’s ability to observe s1 might also be implausible in some markets.
However, “supply noise” was just a convenient way to represent the problem of distinguish-
ing between the component of price impact that depended on the HFT’s order, and the
component of price impact that did not. Alternatively, we could have assumed that the
HFT was uncertain about the size of the transitory and permanent components of the
price change between dates −1 and 0, so that he could not be sure how much of the price
17
The HFT’s prior for β would be different, but since we never specified the particular form of this prior
to begin with, this difference would not change any of our results.

18
change between dates 0 and 1 could be attributed to the price impact of his trade, and how
much could be attributed to some persistent component of the preceding price change.

5.3.3 Slowly-Varying Demand Parameters

Stable demand curves constitute the second vital component of the exploratory trading
model. When the HFT trades in period 1, he learns something about the shape of the
demand curve in that period. Since the HFT does not obtain this information until after
period 1 ends, the information is only valuable to him if it reveals something about the
demand curve that he will face in period 2. The demand parameters need not actually
remain constant over both periods18 , but the period 2 parameters cannot be independent
of the period-1 parameters.
In addition to this temporal stability, the demand curves must be sufficiently well-
behaved that their local shape provides meaningful information about their global shape.
Since an exploratory order and the subsequent order that it informs will not generally be
identical, the HFT must extrapolate. If demand curves are linear, then their local shape
is perfectly informative about their global shape, but under more general assumptions,
extrapolation error will limit the scale (in terms of order size) at which exploratory trading
could be profitable.
Both of the stability requirements above become easier to satisfy in a high-frequency
context, where we replace large, infrequent trades with smaller, more frequent ones.

6 Conclusion
In this paper, I address a central puzzle about high-frequency trading, namely how trad-
ing speed could translate to superior information. To this end, I analyze the idea of
“exploratory trading”—how an agent might use his own trades to gather valuable informa-
tion. A central result of my analysis is that exploratory trading bears a natural connection
to high-frequency trading, and conversely, that high-frequency trading inherently raises
issues analogous to those of exploratory trading. This connection between exploratory
and high-frequency trading ultimately illuminates the issue of how trading speed could
translate to superior information. Beyond these general results, the model of exploratory
trading also helps to explain a variety of specific empirical findings and generates a number
of testable implications.

References
[1] Terrence Hendershott Alex Boulatov and Dmitry Livdan. Informed trading and port-
folio returns. June 2011.
[2] Bruno Biais, Thierry Foucault, and Sophie Moinas. Equilibrium algorithmic trading.
Working Paper, October 2010.
18
Although I abstract away from any sort of strategic or adaptive behavior by the agents who (in
aggregate) submit the market demand curve, we could relax this simplifying assumption somewhat without
dramatically altering the qualitative behavior of the model.

19
 [3] Jonathan A. Brogaard. High frequency trading and its impact on market quality.
Northwestern University Kellogg SOM Working Paper, November 2010.

[4] Miguel Sousa Lobo Bruce Ian Carlin and S. VIiswanathan. Episodic liquidity crises:
Cooperative and predatory trading. The Journal of Finance, LXII(5), OCTOBER
2007.

[5] Álvaro Cartea and José Penalva. Where is the value in high frequency trading? Banco
De España, Documentos de Trabajo N.Âo 1111, June 2011.

[6] Jeff Castura, Robert Litzenberger, Richard Gorelick, and Yogesh Dwivedi. Market
efficiency and microstructure evolution in u.s. equity markets: A high-frequency per-
spective. Working Paper from RGM Advisors, LLC, October 2010.

[7] Giovanni Cespa and Thierry Foucault. Insiders-outsiders, transparency and the value
of the ticker. Queen Mary University Dept. of Economics Working Paper No. 628,
April 2008.

[8] Tarun Chordia, Richard Roll, and Avanidhar Subrahmanyam. Recent trends in trading
activity. UCLA Working Paper, January 2010.

[9] Adam Clark-Joseph and Brock Mendel. A model and analysis of high-frequency trad-
ing. Harvard University Working Paper, February 2011.

[10] Securities Exchange Commission. Concept release on equity market structure, concept
release no. 34-61358. FileNo. 17 CFR Part 242 [Release No. 34-61358; File No. S7-02-
10] RIN 3235-AK47, January 2010.

[11] U.S. Securities Exchange Commission. Findings regarding the market events of may
6, 2010, September 2010.

[12] Jaksa Cvitanic and Andrei Kirilenko. High frequency traders and asset prices. Cal-
Tech/CFTC Working Paper, March 2011.

[13] David Easley, Marcos M. López de Prado, and Maureen O’Hara. The microstructure
of the ‘flash crash’ flow toxicity, liquidity crashes and the probability of informed
trading. Journal of Portfolio Management, Forthcoming.

[14] Peter Gomber and Markus Gsell. Algorithmic trading engines versus human traders
– do they behave different [sic] in securities markets? Published by the Center for
Financial Studies at Goethe-University Frankfurt„ April 2009.

[15] Sanford J. Grossman and Merton H. Miller. Liquidity and market structure. The
Journal of Finance, 43(3), July 1988.

[16] L.E. Harris. Optimal dynamic order submission strategies in some stylized trading
problems. Financial Markets, Institutions and Instruments, 7:1–76, 1998.

[17] Joel Hasbrouck and Gideon Saar. Technology and liquidity provision: The blurring
of traditional definitions. Journal of Financial Markets, 12:143–172, 2009.

20
[18] Joel Hasbrouck and Gideon Saar. Low-latency trading. NYU Stern/ Cornell GSM
Working Paper, May 2011.

[19] Terrence Hendershott, Charles M. Jones, and Albert J. Menkveld. Does algorithmic
trading improve liquidity? The Journal of Finance, 66(1), February 2011.

[20] Terrence Hendershott and Ryan Riordan. Algorithmic trading and information. NET
Institute Working Paper No. 09-08, September 2009.

[21] Robert Jarrow and Philip Protter. A dysfunctional role of high frequency trading in
electronic markets. Johnson School Research Paper Series No. 08-2011, March 2011.

[22] Boyan Jovanovic and Albert J. Menkveld. Middlemen in limit-order markets. Working
Paper, June 2010.

[23] Michael Kearns, Alex Kulesza, and Yuriy Nevmyvaka. Empirical limitations on high
frequency trading profitability. University of Pennsylvania CIS Working Paper, 2010.

[24] Andrei Kirilenko, Mehrdad Samadi, Albert S. Kyle, and Tugkan Tuzun. The flash
crash: The impact of high frequency trading on an electronic market. October 2010.

[25] Thomas McInish and James Upson. Strategic liquidity supply in a market with fast
and slow traders. September 2011.

[26] Ciamac C. Moallemi, Beomsoo Park, and Benjamin Van Roy. Strategic execution in
the presence of an uninformed arbitrageur. Columbia GSB/ Stanford EE Working
Paper, December 2010.

[27] Ciamac C. Moallemi and Mehmet Saglam. The cost of latency. Columbia GSB
Working Paper, June 2010.

[28] Marco Lutat Tim Uhle Peter Gomber, Bjorn Arndt. High-frequency trading. Pubished
by Deutsche Borse AG Market Policy and European Public Affairs, March 2011.

[29] David Romer. Rational asset-price movements without news. The American Economic
Review, 83(5):1112–1130, December 1993.

[30] Australian Securities and Investments Commission. Australian equity market struc-
ture. REPORT 215, November 2010.

[31] X. Frank Zhang. High-frequency trading, stock volatility, and price discovery. Yale
SOM Working Paper, December 2010.

A Mathematical Appendix
A.1 Calculations Related to x2 and π2
Solving for x∗2 :

21
 ∂
(−β (α + x2 ) x2 + µx2 )
∂x2
= −βα − 2βx2 + µ
0 ≡ −βα − 2βx∗2 + µ
µ − βα
x∗2 =
2β
µ − p0
=
2β
Solving for π2∗ :
� �
µ − βα µ − βα µ − βα
π2∗ = −β α + +µ
2β 2β 2β
� �
βα + µ µ − βα µ − βα
= −β +µ
2β 2β 2β
(βα)2 − µ2 2µ2 − 2µβα
= +
4β 4β
2
(βα) + µ − 2µβα
2
=
4β
� �
µ − βα 2
= β
2β
= β (x∗2 )2

Solving for πˆ2 :

π̂2 = −β (α + x̂) x̂ + µx̂

= (−βα + µ) x̂ − β (x̂)2
= (µ − βα) (x̂ − x∗ + x∗ ) − β (x̂ − x∗ + x∗ )2
= (µ − βα) x∗ + (µ − βα) (x̂ − x∗ )
−β (x̂ − x∗ )2 − 2β (x̂ − x∗ ) x∗ − β (x∗ )2
= (µ − βα) x∗ − β (x∗ )2 − β (x̂ − x∗ )2
+ (µ − βα) (x̂ − x∗ ) − 2β (x̂ − x∗ ) x∗
= π2∗ − β (x̂ − x∗ )2 + (µ − βα − 2βx∗ ) (x̂ − x∗ )
= π2∗ − β (x̂ − x∗ )2

The final equality uses the fact that

� �
∗ µ − βα
µ − βα − 2βx ≡ µ − βα − 2β
2β
≡ 0

22
Expected Feasible Period-2 Profit Taylor Approximation:
� �2 � � � �� �
1 1 1 1 2 1 1 −1
− = − +2 − |z=β (z − β)
z β β β z β z2
� � � �
1 −2 3 2 2 6 12
+ 2 + |z=β (z − β) + − 5 (z − β)3
2 βz 3 z 4 6 βζ 4 ζ
� �
1 ζ − 2β
= 4
(z − β)2 + 2ζ −5 (z − β)3
β β
1 32
≤ (z − β)2 + (z − β)3
β4 3125β 5
z 2 − 2zβ + β 2 z 3 − 3z 2 β + 3zβ 2 − β 3
= + 3125 5
β4 32 β
� �2 � � � �
1 1 1 2 −5 β − 2β 3 2 10 4
− = (z − β) + 2β (z − β) + − 5 (z − β)4
z β β4 β 24 η 6 βη
� �
1 2 η −6 5β − 2η
= (z − β)2 − 5 (z − β)3 + (z − β)4
β4 β 6 β
1 2 1
≥ (z − β)2 − 5 (z − β)3 − (z − β)4
β 4 β 4374β 6

A.2 Calculations Related to the HFT’s Inference Problem

Bounding the variance of E1 [β]
� �2 � �2
∆p s1
−β = β 2
x1 x1
�� � �
�2 � σ 2 + b2 σ 2
∆p β s
E −β = 2
x1 x1
� �
≥ E (E1 [β] − β)2

The last line uses the fact that the conditional expectation of β will be at least as good
an estimator (in the L2 -sense) as the expectation of β conditioned upon the value of ∆p x1 ,
∆p
which in turn will be at least as good an estimator of β (in the L -sense) as x1 itself.
2

Next, since x1 is known (and we shall assume x1 �= 0) we can normalize equation (8)
by dividing through by x1 :
∆p
= β + β�x
x1
s1
where we define �x ≡ x1 .
Suppose that for a given fixed value of x1 , the conditional expec-
� �
tation E1 [β] is a smooth function of ∆p
x1 , say E1 [β] = g ∆p
x1 , with continuous derivative
g � (·) that is not identically zero. By standard delta-method-type Taylor approximation

23
arguments, it is easy to show that
� � � �
(x1 )2 E (E1 [β] − β)2 → E g � (β)2 V ar (∆p)
� �� �
= E g � (β)2 σβ2 + b2 σs2

� g (·)
as |x1 | → ∞ . Since we assume that the derivative � �is not identically zero, the term
�
�
2
on the right-hand side of the equation above, E g (β)
� σβ + b σs , is strictly positive.
2 2 2

Therefore we can choose J > 0 such that for all x1 satisfying (x1 )2 > J, we have
� � 1 � �� �
(x1 )2 E (E1 [β] − β)2 > E g � (β)2 σβ2 + b2 σs2
2
� � 2
Hence as |x1 | → ∞ , we can bound E (E1 [β] − β)2 below by a function of the form Aσ x21
s

, where A is some strictly positive constant. �

The restrictive assumption in the argument above was taking the conditional expecta-
tion E1 [β] to be a smooth function of ∆p x1 . Although this condition is easy to satisfy if we
∆p
restrict attention to x1 > 0, the condition becomes more restrictive if we permit ∆p x1 ≤ 0.
Because we assume that the support of β is strictly positive, a negative value of ∆p x1 im-
∆p
plies that s1 < −x1 , whereas a positive value of x1 does not imply this. (More precisely,
the break-point for ∆px1 is the infimum of the support of β, which is strictly positive, but
the same ideas apply.) Although E1 [β] will typically be a continuous function of ∆p x1 in a
neighborhood of zero, it is not obvious that it is likely to be everywhere-differentiable in
that neighborhood—see the expressions below.

� ���
∆p
ps x1
θ − θ f (θ)
x1
f (θ|∆p = c; x1 ) = ´ � x1 � ∆p ��
ps t x1 − t f (t) dt
� � � �
p�x 1θ ∆p
x1 − 1 f (θ)
= � � � �
p�x 1t ∆p
x1 − 1 f (t) dt
´
� � � �
θp�x 1θ ∆p 1 f (θ) dθ
´
x1 −
E [β|∆p = c; x1 ] = � � ��
p�x 1t ∆p f (t) dt
´
x1 − t

Of course, the probability of encountering this break-point decreases as |x1 | increases, and
Chebychev’s inequality implies that this probability is bounded above by a function of the
form xH2 for some strictly positive constant H. Since we are only interested in deriving a
1 � �
lower bound for E (E1 [β] − β)2 , the basic result above should probably hold under more
general conditions, but the proof would be more delicate and more involved.

24
A.3 Maximizing Total Expected Profits
Expected Period-2 Profits as a Function of x1 : Using equation (9), we can re-
2 2
place V ar (E1 [β] |β) with K 2 βxσ2 s in equation(6). Then, taking the approximation to hold
1
exactly for simplicity, we get

(µ − p0 )2 σs2
E [π̂2 |β, p0 ] = π2∗ − K 2
4β x21

By taking expectations of both sides of the above equation with respect to β, we obtain
an expression for the unconditional expectation of π̂2 as a function of x1 :
� �
∗ K 2 E β −1 σs2 ξ 2
E [π̂2 |p0 ] = E [π2 ] −
4x21
� � � �
ξ 2 E β −1 K 2 E β −1 σs2 ξ 2
= −
4 4x21
� � � �
ξ 2 E β −1 K 2 σs2
= 1−
4 x21

Recall that we have defined ξ ≡ βα − µ and p0 ≡ βα, so ξ ≡ p0 − µ.

The direct trading profit from trading x1

π1 = x1 (µ − p0 − βx1 )
= (µ − p0 ) x1 − βx21
= ξx1 − βx21

The optimal choice of x1 in the absence of direct trading gains:

� �
2 E β −1 � �
σ ξ K 2 σs2
E [π1 + π̂2 ] = −bx1 +
2
1−
4 x21

� � �� ��
σξ2 E β −1 K 2σ2
s
max −bx21 + 1−
x1 4 x21
� �
∗
σξ2 E β −1 K 2 σs2
F OC : −2bx1 + 2 ≡ 0
4 (x∗1 )3
� �
σξ2 E β −1 K 2 σs2
⇒ = b (x∗1 )4
� 4
E [β −1 ]
⇒ σξ σs K = (x∗1 )2
4b

25
 The maximized value of the objective function is therefore given by:
�
2 E β −1
�� �
σ ξ K 2σ2
max E [π1 + π̂2 ] = −b (x∗1 )2 + 1 − ∗ 2s
x1 4 (x1 )
�
2 E β −1
� �
2 E β −1
�
σ ξ σ ξ K 2 σs2
= −b (x∗1 )2 + −
4 4 (x∗1 )2
� � �
σξ2 E β −1 E [β −1 ]
= − bσξ σs K
4� � 2 2 4b
−1 √
σξ E β
2 K σs 2 b
− �
4 σξ σs K E [β −1 ]
� � � �
σξ2 E β −1 σξ σs K E [β −1 ] b σξ σs K E [β −1 ] b
= − −
4
� � 2 2
σξ2 E β −1 �
= − σξ σs K E [β −1 ] b
�4
σξ E [β −1 ] � � √ �
= σξ E [β −1 ] − 4σs K b
4
Hence the maximized value of the expected total profit (in the absence of direct trading
gains) will be positive when σξ2 is sufficiently larger than σs2 , specifically, when

16K 2 b
σξ2 > σs2
E [β −1 ]

If σξ2 does not satisfy the condition above, the HFT would obtain a higher expected
profit (zero) by not participating in the market than he would by participating in an
otherwise optimal manner.

26