∗

Where is the Value in High Frequency Trading?

Álvaro Cartea† and José Penalva‡

First Version: November 21, 2010, This version: February 17, 2012

Abstract

We analyze the impact of high frequency (HF) trading in financial markets based on a
model with three types of traders: liquidity traders (LTs), professional traders (PTs), and
high frequency traders (HFTs). Our four main findings are: i) The price impact of liquidity
trades is higher in the presence of the HFTs and is increasing with the size of the trade. In
particular, we show that HFTs reduce (increase) the prices that LTs receive when selling
(buying) their equity holdings. ii) Although PTs lose revenue in every trade intermediated
by HFTs, they are compensated with a higher liquidity discount in the market price. iii)
HF trading increases the microstructure noise of prices. iv) The volume of trades increases
as the HFTs intermediate trades between the LTs and PTs. This additional volume is a
consequence of trades which are carefully tailored for surplus extraction and are neither
driven by fundamentals nor is it noise trading. In equilibrium, HF trading and PTs coexist
as competition drives down the profits for new HFTs while the presence of HFTs does not
drive out traditional PTs.
∗
We would like to thank Harrison Hong for his valuable comments and discussions. We are also grateful to
Andrés Almazán, Gene Amromin, Michael Brennan, Max Bruche, Celso Brunetti, Guillermo Caruana, Terrence
Hendershott, Andrei Kirilenko, Pete Kyle, Albert Menkveld, Rafael Repullo, Eduardo Schwartz and Enrique
Sentana for their comments. We also thank seminar participants at the SEC, CFTC, Federal Reserve Board,
CEMFI, University of Toronto, and Universidad Carlos III. The usual caveat applies. This article benefited
from financial support from the Comunidad de Madrid-UC3M 2011/00093/00, 2009/00138/00, and Ministerio
de Ciencia e Innovación 2010/00047/001. We welcome comments, including references we have inadvertently
missed.
†
alvaro.cartea@uc3m.es, Universidad Carlos III de Madrid.
‡
jpenalva@emp.uc3m.es, Universidad Carlos III de Madrid.

Electronic copy available at: http://ssrn.com/abstract=1712765

 Keywords: High frequency traders, high frequency trading, flash trading, liquidity
traders, institutional investors, market microstructure, microstructure volatility, execution
costs, market quality

1 Introduction

Around 1970, Carver Mead coined the term “Moore’s law” in reference to Moore’s statement
that transistor counts would double every year. There is some debate over whether this “law”
is empirically valid but there is no discussion that the last forty years have seen an explosive
growth in the power and performance of computers. Financial markets have not been immune
to this technological advance, it may even be one of the places where the limits of computing
power are tested every day. This computing power is harnessed to spot trends and exploit profit
opportunities in and across financial markets. Its influence is so large that it has given rise to a
new class of trading strategies sometimes called algorithmic trading and others high frequency
trading. We prefer to use algorithmic trading (AT) as the generic term that refers to strategies
that use computers to automate trading decisions, and restrict the term high frequency (HF)
trading to refer to the subset of AT trading strategies that are characterized by their reliance
on speed differences relative to other traders to make profits based on short-term predictions
and also by the objective to hold essentially no asset inventories for more than a very short
period of time.1

The advent of AT has changed the trading landscape and the impact of their activities is at
the core of many regulatory and financial discussions. The explosion in volume of transactions
we have witnessed in the last decade, and the speed at which trades are taking place, is highly
suggestive that AT is very much in use and that these strategies are not being driven out of the
market as a result of losses in their trading activities. Indeed, different sources estimate that
annual profits from AT trading are between $3 and $21 billion (Brogaard [2010] and Kearns
1
This definition is consistent with the one used in Kyle [Kirilenko et al. [2010] p3]: “We find that on May
6, the 16 trading accounts that we classify as HFTs traded over 1,455,000 contracts, accounting for almost a
third of total trading volume on that day. Yet, net holdings of HFTs fluctuated around zero so rapidly that
they rarely held more than 3,000 contracts long or short on that day.” See also the definition of HFT provided
by CFTC Comissioner O’Malia [2011].

Electronic copy available at: http://ssrn.com/abstract=1712765

et al. [2010]). These strategies have supporters and detractors: on one side we find trading
houses and hedge funds who vigorously defend their great social value, whilst being elusive
about the profits they make from their use; and on the other hand there are trading houses
that denounce high frequency traders (HFTs) as a threat to the financial system (and their
bottom line). Although AT in general and HF trading in particular have been in the market
supervisors’ spotlight for quite some time and efforts to understand the consequences of HF
trading have stepped up since the ‘Flash Crash’ in May 6 2010 (SEC [2010], Commission et al.
[2010], Kirilenko et al. [2010], and Easley et al. [2011]) there is little (but increasing) academic
work that addresses the role of these trading strategies.

In this paper, we provide a model of HFTs, where HFT profits come from extracting rents
from other traders. The objective of the paper is to understand the impact of incorporating
HFTs’ rent extraction behavior into a standard model of the stock market. In particular we want
to ask the following questions: How is the burden of HFT rent extraction distributed amongst
different market participants? How does HFT rent extraction affect market aggregates such as
liquidity provision, price volatility, volume? What are the incentives for market participants
to become HFTs? What factors enter into the long-term profitability of HFTs and under what
conditions HFTs and other traditional market participants can coexist?

To answer these questions we develop a model with three types of traders: liquidity traders
(LTs), professional traders (PTs), and HFTs. In this model LTs experience a liquidity shock
and come to the market to unwind their positions which are temporarily held by the PTs in
exchange for a liquidity discount. HFTs are introduced as traders who, thanks to their rapid
information processing and quick execution ability, can position themselves in the order book
so as to profit from intermediating between the equity investor (LT) and the PTs. HFTs work
in a much finer time scale than other traders. Their strategy of keeping inventories close to zero
together with their speed leads them to act essentially as “instantaneous” intermediaries, whose
economic profits must come at the expense of those of others, but whose actions have additional
effects on market conditions. Although we discuss how HFTs could be extracting surplus in
practice, HFT behavior is modeled in a highly stylized way to allow us to focus on studying the
impact of HFTs on other market participants and market aggregates. Our abstraction from the

3
precise details of HFT trading strategies allows us to provide a parsimonious and analytically
tractable treatment of their overall impact, and evaluate their effects relative to the benchmark
provided by the same model without HFTs, which corresponds to that of Grossman and Miller
[1988].

We highlight five main findings. First, HF trading strategies introduce “microstructure

noise”: in order to profit from intermediation HFTs buy shares from one trader at a cheap price
and sell it more dearly to another trader, generating price dispersion where before there was
only a single price. Moreover, HFTs’ additional intermediation increases the volume of trade.
The additional volume is neither driven by fundamentals (only the original trades, without the
HFTs, are driven by fundamentals) nor is it noise trading. Far from it, the extra volume is
a consequence of trades which are carefully tailored for surplus extraction. These properties,
which are built into the model, closely correspond to behavior reported in Kirilenko et al. [2010]
and are also found by Brunetti et al. [2011].

Second, market participants react to the presence of HFTs by adjusting asset demand and
supply functions, leading to an additional market distortion in the form of an exacerbated price
impact of initial liquidity trades (the ones that generate a temporary order imbalance). Thus
the burden of HFT rent extraction is primarily borne by liquidity demanders who face a double
burden: the direct cost from the trading surplus extracted by the HFT, and the indirect cost
of a greater price impact. Furthermore, this effect is increasing in the size of the liquidity need,
and is consistent with the results in Zhang [2010].

Third, we find that the effect of HFTs on the market’s liquidity providers, the PTs, is
ambiguous. On the one hand they suffer increased trading costs from HFT surplus extraction.
On the other hand, the increased market impact of the initial liquidity trades implies increased
profits from higher liquidity discounts. The paper focuses on the knife-edge case where these
two effects cancel each other, leaving expected profits for PTs unchanged. Nevertheless, despite
the a-priori negative definition of HFTs as pure rent extractors, it is possible for the second effect
to be dominant so that HFTs lead to an increase in overall liquidity provision in the market
(as considered in the Appendix). We also find, when we introduce competition between HFTs,

4
that competition generates additional uncertainty for PTs, which, in equilibrium, distorts total
market liquidity.

Fourth, we find that standard measures of market liquidity, specially volume-based ones,
may lead to erroneous conclusions: in our model, HFTs may not increase liquidity and yet we
obtain substantially higher trading volumes. In fact, LTs face overall lower sales revenue and
higher costs of purchase, suggesting that liquidity is better measured through total cost of trade
execution.

Finally, we consider competition between HFTs and the possibility that a PT may react to
the presence of HFTs by deciding to become one as well. We find that profits from HF trading
attract entrants (e.g. PTs) who are willing to make the necessary investments to acquire the
skills necessary to compete for these profits. In a context where price competition is limited,
for example by minimum tick size, we find that as the number of HFTs increases, the expected
profits of HF trading falls until the expected skills of an entrant (relative to those of existing
HFTs) are insufficient to generate enough profits to cover the initial investments required to
become an HFT. Thus, in equilibrium PTs with low expected skills as HFTs will continue
in their traditional role, coexisting with others acting as skilled and profitable HFTs, to the
point that any one of the actual market participants could be in effect playing both roles
simultaneously.

The model also incorporates skill differences between HFTs and other market participants
(LTs and PTs). This allows us to discuss how other market participants may react to the
presence of HFTs by improving their trading technology–using AT or subcontracting their
trade execution to (additional, unmodelled) financial intermediaries–and thereby limiting the
impact of HFT rent extraction.

Our analysis focuses on the effect of HFTs’ surplus extraction on trades initiated by liq-
uidity needs that generate temporary trading imbalances. We interpret a liquidity need very
generally, in that it encompasses trades initiated by mutual fund managers, hedge funds, insur-
ance companies and other large investors, as well as trading motivated not only by immediate
liquidity needs, but also trading to build up or unwind an asset position, for hedging, etc.

5
 Our analysis starts from the position that HF trading is pure rent-extraction and does
not describe how this occurs in order to allow us to study the effect of the presence of HFTs
on overall market aggregates. The empirical literature provides conflicting views on whether
such description of HFTs is consistent with reality. Different empirical studies employ different
approaches and look at different markets, and a key difference between them is the depth
and breadth of the data they employ. Our starting point for describing HFTs is consistent
with the subset of empirical studies that have access to audit-trail data which includes trader
identity. The works of Kirilenko et al. [2010], SEC [2010] and, very recently, Brunetti et al.
[2011] study the role of HFTs in the E-mini S&P 500 futures contracts. The fact that they can
identify traders individually allows them to study trading patterns that are characterized by
high speed, and inventories that are not significant (relative to the trader’s overall volume).

Kirilenko et al. [2010] study the impact of such HF trading during the Flash Crash on
May 6 2010. They find that HFTs have the highest price impact among all types of traders
and that “HFTs are able to buy right as the prices are about to increase. HFTs then turn
around and begin selling 10 to 20 seconds after a price increase.” Moreover, they find that
“The Intermediaries [our Professional Traders] sell when the immediate prices are rising, and
buy if the prices 3-9 seconds before were rising. These regression results suggest that, possibly
due to their slower speed or inability to anticipate possible changes in prices, Intermediaries
buy when the prices are already falling and sell when the prices are already rising.” The way
we characterize HFTs in our model embeds HFTs’ higher trading speed by allowing them to
make many trades as the LT transactions enter the market and are eventually transferred to
intermediaries’ (PTs’) books. We also abstract from the dynamics of the order book that allows
HFTs to anticipate price movements (as described by Kirilenko et al. [2010]) and lead to PTs
buying when prices are falling and sell when they are rising, in order to focus on the net effect
of HFTs on the final transaction, as the asset is transferred from LTs to PTs. Thus, the HFT
behavior is summarized as extracting surplus and ending up with zero inventory.

Recent empirical work by Brunetti et al. [2011] find that HFTs are responsible for a large
volume of trades in the E-Mini S&P 500 futures contracts, but this “high volume does not
necessarily imply a substantive flow of information. To the contrary, high volume may con-

6
tribute to the noise, at least, at high frequency sampling.” These results are consistent with
our theoretical model where HFTs do not have information on the fundamentals of the asset,
and yet they increase microstructure noise and volume.

Zhang [2010] studies changes in volatility over time and concludes that HF trading increases
stock price volatility and that this positive correlation between volatility and HF trading “is
stronger for stocks with high institutional holdings, a result consistent with the view that high-
frequency traders often take advantage of large trades by institutional investors”. This, again, is
consistent with our theoretical result that the cost of execution (liquidity discount) is increasing
in the size of the liquidity imbalance.

Further empirical papers find that HFTs improve market quality, may dampen volatility
and reduce price impact, Brogaard [2010], Hendershott et al. [2011], Hendershott and Riordan
[2009], and Hendershott and Riordan [2011]. Most of these studies find that AT and HFTs
have a positive impact on market quality but the data employed are not as comprehensive as
audit-trail data. For example, Brogaard [2010] and Hendershott and Riordan [2011] use data
provided by NASDAQ with an HFT identifier, instead of trader identity, but the identifier is
narrowly defined for specialized HFT trading operations and explicitly excludes HFT activity
from major traders, such as Goldman Sachs, Morgan Stanley, and other large integrated firms.

Finally, one of the early papers that looks at the impact of AT in financial markets is that
of Chaboud et al. [2009]. The authors study the impact of AT in the foreign exchange market.
One of their findings is that the presence of more AT is associated with lower volatility of the
‘fundamental value’ of exchange rates. This interesting finding is not in contradiction with
our theoretical prediction that HFTs increase microstructure noise of the fundamental value of
the asset because in their study they employ minute-by-minute observations which dilutes the
contribution of microstructure noise to the variance of the exchange rate.

The rest of the paper proceeds as follows. In Section 2 we develop our framework and
analysis that describes how the HFTs compete to win trades and how the winner of the trade
mediates between LTs and PTs. Within this framework, we establish the properties of equilib-
rium prices, quantities, and optimal HFT behavior. In Section 3 we develop the general model,

7
which allows for competition between HFTs and entry–in the form of an PT who considers
setting up an HF trading desk. Section 4 discusses the effects of HFTs on market aggregates,
and Section 5 concludes with a discussion of some key features about HFTs that require further
research (and quality data).

2 The Model

We study the role of HF trading in a normative setting, where the stock market has social
value because it facilitates the financing of economic activity, adding value to equity holders
by providing a way for them to convert their equity into cash (and viceversa) quickly and
at a reasonable price. Grossman and Miller [1988] (GM) provide a highly stylized model of
such a stock market. In the GM model, equity investors or LTs quickly find counterparties for
their trading needs.2 These counterparties are PTs who are willing to take the other side of
investor trades and hold those assets temporarily–until another investor enters the market to
eliminate the temporary order imbalance.3 Holding these assets entails price risk, which PTs
are willing to bear in exchange for a small discount on the asset’s price (the liquidity discount).
We introduce HFTs into the GM model, where an HFT is a trader who, thanks to her rapid
information processing and quick execution ability, can position herself in the order book in
such a way that she can profit from intermediating between the equity investor (LT) and the
PTs.
2
The other interpretation of the model is intended for the futures market. Everything we say for the stock
market is also equally valid for trading in the futures market in exactly the same terms. We refer the interested
reader to the original article for details of how to reinterpret the model.
3
In GM, our PTs are called “market makers”. This is because the GM model was designed with specialist
markets in mind. Nevertheless, the same idea of markets providing liquidity for a premium is valid in more
general settings, such as those with an open limit order book, where liquidity is not provided by “specialists”
but by a group of PTs that are in the business of providing liquidity by exploiting temporary price deviations
from “fundamental” prices. A model where PTs arise endogenously can be obtained by, for example, including a
cost for monitoring the order book. In such a model, PTs are traders who are permanently monitoring the order
book and who, despite having no privileged information, are willing to assume trading positions for a possibly
significant period of time when the price deviates from its fundamental value.

8
2.1 The Basic Structure

The setting for our model is a world with three dates, t = 1, 2, 3 in which a temporary order
imbalance of size i affects conditions in a stock market with a cash asset (with a return normal-
ized to zero) and a risky asset. There are two “outside investors”, LT1 and LT2. At time 1, LT1
is endowed with cash and an amount i of shares which due to a liquidity shock he would like to
sell (that is, he wants to trade −i shares). On the other hand, the other trader, LT2, wants to
buy i > 0 units (trade i shares) but only reaches the market at date 2.4 If both traders met at
the same time they would exchange the asset at the current “fundamental” price, which is the
expected future cash value of the asset. But, as LT1 arrives to the market much earlier than
LT2, his only option is to sell to the rest of the market, represented by M professional traders
(PTs). These act as intermediaries and will accept LT1’s order at date 1 at the “right price”
and hold the shares until LT2 enters the market at date 2.

The “right price”, which we call the market price of the asset, is determined by the funda-
mental price minus a discount commanded by the PTs for holding the asset until date 2. PTs
bear the risk of having to sell the asset at a loss after the release of public news that might ad-
versely change the fundamental price of the asset, and they require the discount to compensate
them for assuming this risk. The “liquidity discount” is calculated as the difference between
the “fundamental” value of the asset and the price at which the transaction takes place, the
market price.

All these traders (LT1, LT2, and PTs) are price-taking and risk averse. They have expected
utility from wealth at date 3 and they choose their asset positions so as to maximize E [U (W3 )],
where U (W ) = − exp (−aW ) is the utility function, W is wealth and a is the risk aversion
parameter. The future cash value of the (non-dividend paying) asset is P3 = µ+2 +3 , where µ
is constant, and 2 and 3 are normally and independently distributed with mean 0 and variance
σ 2 . The random variables, 2 and 3 , represent public information that is announced between
dates t = 1 and t = 2, and between t = 2 and t = 3 respectively and σ 2 represents the variance
of the shocks of the fundamental value of the asset. Let µ2 = E [P3 |F2 ] = µ + 2 denote the
4
We make the assumption that LT1 wants to sell and LT2 wants to buy to streamline the presentation. The
analysis is equally valid if LT1 wants to buy i shares and LT2 wants to sell them.

9
expectation of P3 conditional on information available at t = 2. We use the notation θtx (Pt ) to
describe the asset holdings of trader x after trading at date t given price Pt . A positive asset
position, θtx > 0, implies holding a positive (long) number of shares.

2.2 High Frequency Trading

The GM model identifies the relevant “liquidity” time scale as the time between two liquidity-
driven and mutually balancing trades, those by LT1 and LT2, and this time period is sufficiently
long that news can produce significant changes in the fundamental value of the asset (hence
the compensation for risk demanded by PTs). HFTs, on the other hand, are characterized by
having very short holding periods–so short that the probability of substantive changes in the
fundamental value of the asset during the period is insignificant. Thus, a single HFT cannot, by
definition, provide “true liquidity” in the GM model: her holding period is not long enough for
her to be exposed to uncertainty arising from changes in the fundamental value of the asset, and
hence does not bridge the time period between LT1 entering the market and the arrival of LT2.
This does not mean that market prices are constant during the HFTs holding period. In fact,
they cannot be constant, otherwise the HFT could not profit from buying and selling the asset
(they could even lose money from paying the spread net of exchange fees). Our assumption,
which is consistent with the evidence described in the contemporaneous empirical studies by
Kirilenko et al. [2010] and Brunetti et al. [2011], is that HFTs can make use of the microstructure
details of the price setting process (the dynamics of the order book) together with their ability
to transact at millisecond speeds to make profits by intermediating between the LT and the rest
of the market (the PTs). Thus we model HFTs as an additional (and profitable) intermediation
layer between LT and PTs.5 The LT cannot avoid this additional intermediation layer and trade
directly with the PT in the sense that trades are matched through an anonymous matching
engine which is populated by HFTs, and they will be there whenever the LT brings his order
to the exchange. Although in practice PTs’ orders are also posted in the exchange’s book and
some trades will bypass HFT intermediation, focusing on the case where HFTs intermediate all
5
As pointed out by Kirilenko and Tuzun (personal communication), the way we capture HFT behavior may
also be useful to model other market frictions. In particular, they suggest that our model may help understand
the (historical) figure of “scalpers” on the trading floor.

10
trades allows us to obtain closed-form solutions and characterize the overall effect of HFTs. As
we will see in Section 3, allowing for some trades to bypass HFTs will not change the results.

To limit the negative role of HFT behavior, HFT profit making is done by extracting existing
trading surplus in a way that is frictionless. At t = 2, the presence of HFTs does not alter the
quantities traders are willing to supply (and demand) at any price–they are the same as without
HFTs. However, at t = 1, traders anticipate future surplus extraction by HFTs and adjust their
demand and supply curves, which, as we will see, generates a higher liquidity discount for LT1
as PTs move their positions away from the bid and ask.

That HFTs intermediate profitably is evidenced by their activity and the fact that they
report positive profits. Only HFTs know the exact details of the strategies they use, but from
our discussions with market participants we gather that HF trading is implemented using a
variety of strategies. A component of these strategies, which we believe drives the profitability
of HF trading, is that HFTs can exploit their speed to alter market conditions (their positions
in the order book) in a way that makes buyers accept a slightly higher price and sellers a slightly
lower one than they would have done otherwise. The simplest such strategies are associated
with the term “price anticipation”, and the idea is relatively simple and works in a setting where
LTs split their trades in small packages and PTs are outcompeted by HFTs who are able to
advantageously position themselves ahead in the limit order book.

To illustrate how this could work consider the following stylized numerical example: Suppose
a trader (LT) needs liquidity and wants to sell a block of shares, splitting them in several
packages. As the first package comes into the system (say at the best buy price of $5.50 per
share), the HFT identifies an increase in selling pressure and cancels her outstanding posted
buy offers that have not been filled. She then posts additional sell offers, adding to the increased
selling pressure, in order to help clear the remaining posted buy orders in the book. Once the
book is clear, the HFT quickly reposts a significant number of offers at lower prices (say $5.47)
so that she is first in the buying queue. This is only possible if she can move quickly enough
that by the time the PT reacts to the increased selling pressure, new offers posted by the PT
sit behind those posted by the HFT at $5.47 per share. The LT finds that the market around
$5.50 has dried up and can only sell at $5.47. The following share packages are bought by the

11
HFT, who is at the front of the queue. Being at the front of the queue allows the HFT not only
to be the first to buy at this low price, but also to be the first to notice when the selling pressure
eases. At that moment, the HFT cancels her posted buy offers and starts posting sell offers at
slightly higher prices (say $5.49 first and then at $5.48). The PT sees the selling pressure shift
and the price rebound. Although the PT is sorry to see the lower prices disappear, he is still
willing to buy at $5.49, which allows the HFT to sell her earlier purchases and end up with a
zero net position. During this process, the HFT has been able to make profits on the shares
bought at $5.47 and sold at $5.48 or $5.49, while taking a loss on the shares executed at the
beginning, bought at $5.50 and sold at $5.48 or $5.49. An HFT who is fast enough will have
bought many more shares at $5.47 than at $5.50 by selectively canceling and reposting her
offers to make the strategy profitable. As for the other traders: The PT bought some shares
below the initial best buy price so he is satisfied; the other LT is also satisfied from having been
able to sell his shares even though for some of them he received a cent less than what the PT
paid for them.

Note that the HFT in this example has used her speed to cancel and repost at different
levels of the order book so as to (a) allow the selling pressure to drive prices lower than they
would have been in the absence of HFTs, and (b) position herself at the front of the queue in
the order book so as to be the first to exploit the temporary drop in prices, buy at low prices
and repost so that the rest of the market absorbs her trades at slightly higher prices.

In order to keep the model tractable we do not introduce the dynamics of HFT strategies
into the model. Instead, we focus on the key outcomes of HFT activity: profitable trading
at a much finer timescale, and zero net positions. To do this, we allow several actions to
take place at liquidity trading times, t = 1, 2, which in reality they take place sequentially
at several points in time around, but very close to, those liquidity trading times. Thus, by
executing different actions in a very short time span, HFTs can exploit the microstructure of
trade execution and the dynamics of the order book, both at t = 1, as the price moves from the
pre-trade fundamental level (µ) to its post-trade market price (P1 ), and at t = 2–as it moves
from P1 + 2 to P2 .

12
 Instead of modeling the details of how HFTs profitably exploit their speed advantage, we
use the following abstract characterization: [(i)] HFTs use their speed to position themselves
between LTs and PTs. In the benchmark model (this section) we start by assuming that a single
HFT intermediates all trades. This is profitable because in the process of intermediation, HFTs
can extract trading surplus. In particular, [(ii)] we assume the HFT takes over other traders’
positions in two blocks: she obtains a first block of shares at an advantageous price: the market
price Pt plus or minus a haircut ∆ (depending on whether the HFT is selling or buying), and
obtains the second block of shares at the market price, Pt . Thus, the HFT profits from each
trade twice: once, when acquiring the position from one side of the market, and a second time,
when transferring the position to the other side of the market.

As we discuss below (in Section 3) the ability of HFTs to profit from their speed and
knowledge of the dynamics of trade execution depends on a number of factors and in general
HFTs will not be able to intermediate all trades, and even those they intermediate may not
be profitable. It could also happen that LTs break their trades into blocks, some of which are
matched with one or more HFTs during the trading process, while others are matched to PTs
directly, or the trade may become a “hot potato” moving from one HFT to another. As adding
these extra properties to the model makes it less tractable, we introduce them later, when we
include competition amongst HFTs in the model (Section 3).

The assumption that the HFT captures the trade in two blocks is mathematically convenient
for two reasons: one, it allows us to avoid having to model the time interval around liquidity
trades, and two, it allows us to capture the notion that HFTs exploit the microstructure of
price setting without having to model the exact (and unknown) details of how they actually do
it.

2.3 The GM model with an HFT: a numerical example

The following numerical example helps illustrate the mechanics of HFT intermediation profits
in our stylized, two-block trading model, and how those profits are determined by market

13
conditions. The example is immediately followed by the description of the general benchmark
model.

Suppose that an LT wants to sell 1,000 shares, the expected price of the asset is $5.50
and there are nine PTs. Suppose further that all traders are risk averse and they have linear
demands for assets given by

θT (P ) = θLT (P ) = 5,000 (5.50 − P ) .

Hence, they will only buy shares if the price is below $5.50.

In equilibrium the holdings of all traders (one LT and nine PTs) have to add up to the
supply of LT (1,000 shares) so that

9θT (P ) + θLT (P ) = 1,000

⇒ 9 × 5,000 (5.50 − P ) + 5,000 (5.50 − P ) = 1,000

⇐⇒ P = 5.48 .

The price drop to $5.48 per share reflects how the market reacts to the LT’s selling pressure.
At the lower price each trader (including the LT) will hold 100 shares. This implies that the
(price sensitive) LT who initially wanted to sell 1,000 shares (at $5.50) ends up only selling
900 shares at a price of $5.48 (the fundamental price $5.50 minus a 2 cent liquidity discount).
Thus, the liquidity trade involves fundamental traders (LT and PTs) exchanging assets at the
market price, $5.48.

In a market with HFTs, the bid stack is populated with HFT bids that compete to capture
the trade and unwind it at a profit. Buying bids disappear as soon as the market feels the
increased selling pressure. In our simplified benchmark model, the increased selling pressure is
entirely absorbed by a single HFT who successfully captures the trade at a discount, say she
takes a 1 cent per share haircut and buys at $5.47. Having bought the shares from LT, the
HFT unloads it on the market (PTs), and again takes a haircut, selling to PTs at the slightly
higher price of $5.49.

14
 However, the HFT is an intermediary and her profits are determined by the quantities she
trades, which are determined by “fundamental” traders’ asset demands. At a price of $5.47, LT
will want to hold 5,000 (5.50 − 5.47) = 150 shares, thus selling only 850 shares. At a price of
$5.49, PTs will hold 5,000 (5.50 − 5.49) = 50 shares each, which in total amounts to 450 shares.
The trades from LT and the PTs do not cancel so that the HFT is left holding 400 shares. This
is not satisfactory for an HFT who wants to hold zero inventory at the end of trading. Thus,
the HFT makes use of her ability to acquire blocks at different prices and modifies her trading
strategy slightly: After offering to buy 850 shares from LT at $5.47, she offers to buy another
batch of 50 shares from the LT at $5.48. The LT finds these trades acceptable and executes
them. Having acquired all LT’s shares, the HFT turns around and first offers to sell 50 shares
to each of the PTs at $5.49. Then, she offers to sell them another 50 shares each at $5.48.
Again, this combination of offers is acceptable to the PTs and they agree to execute them. At
the end of these trades, the HFT has made 2 cents on each of 450 shares sold at $5.49 (and
bought at $5.47) and another cent on 400 shares bought at $5.47 and sold at $5.48 for a total
of $13.00. She also has sold 50 shares at $5.48 at no profit to clear her inventory, and is left
with no shares, only profits. The market has seen a doubling of the volume of shares traded,
plus some transactions taking place at prices above and below the “market price” of the asset
($5.48)–the fundamental price ($5.50) minus the equilibrium liquidity discount ($0.02).

This numerical example captures how we model the HFT’s profit extraction strategy, and
how the HFT behavior affects market variables, such as prices and trades. Furthermore, note
that the HFT’s strategy is frictionless in the sense that the quantity demanded by the market
and supplied by the LT (900 shares) is the same that would have been traded without the HFT
(at the constant market price of $5.48). We now proceed formally with the benchmark model.

2.4 The benchmark model

The timeline is as follows (Figure 1): at date 1, the price sensitive trader LT1 comes to the
market wanting to sell i shares. He is price sensitive so that his date 1 excess demand is given
by θ1LT 1 (P1 ) − i, where we recall that θtx (Pt ) is the asset holding after trading at date t. The
HFT trades with LT1 in two blocks, she buys the first block at P1 − ∆ and the other at P1 .

15
Immediately, and still at date 1, the HFT turns around and sells to the PTs in two blocks: one
block at P1 + ∆ and the other at P1 . As time passes between dates 1 and 2, there are public
announcements of news about the future value of the asset, P3 , which generates fluctuations
in the expected value of the asset. At date 2, the other price sensitive liquidity trader LT2
decides to enter the market to acquire i shares. His excess demand is θ2LT 2 (P2 ) + i. The HFT
quickly purchases all the shares that the PTs had bought in date 1, plus those that LT1 could
not sell in date 1, in two blocks: one block at P2 − ∆ and the other at P2 . Then the HFT turns
around and sells all the shares to LT2 in two blocks: one at P2 + ∆ and the other at P2 . This
translates to a slightly worse (average) buy price for LT1 and also a slightly worse (average)
sell price by PTs.6

Figure 1: Timeline of trading in the presence of the HFT

We now proceed by solving the general model by backward induction, starting at date 2
and then moving one period back to the problem at date 1.
6
The time sequence of trades is not important, as the HFT can acquire or sell short shares in anticipation
of, or in reaction to liquidity trades. The key is that the HFT’s net position coming in and out of the liquidity
shocks is essentially zero.

16
Trading at t = 2

The second liquidity trader, LT2, enters at t = 2 and given price P2 , this price sensitive trader’s
excess demand is θ2LT 2 (P2 ) + i shares, but he knows that there is an HFT who will extract
some surplus. In the benchmark model, any trader who wants to buy θ (P ) + i units at price P
will have to go through the HFT who will split the order in two blocks at quantity θ̃: the HFT
will sell θ̃ + i units at price P + ∆ and the remaining θ (P ) − θ̃ units at P . To ensure trades are
mutually acceptable at all prices the quantity θ̃ corresponds to the trader’s excess demand at
P + ∆, that is θ̃ = θ (P + ∆). We assume traders accept trading in two blocks and paying the
haircut as part of the cost of doing business, and that they take the quantity θ̃ as exogenously
given.7 We use the notation θ̃tx to describe the quantity that the HFT uses to split trader x’s
desired trades at date t into the two blocks. The consistency condition, θ̃ = θ (P + ∆), coupled
with the fact that the HFT splits the orders into only two blocks, limits the amount of surplus
the HFT can extract.

We also limit the advantage gained by the HFT in terms of the information she can use to set
haircuts. In principle, the HFT could distinguish between different traders and different periods
of time. But, being able to make such fine grained distinctions between traders would require
extracting an amount of information from the order flow that we consider unrepresentative
of what goes on in stock markets today. Nevertheless, we do believe that, depending on the
amount of liquidity available, HFTs may be able to extract sufficient information from the
order flow at a sufficient speed to allow them to distinguish a large liquidity trade coming in
(one that will move the market price substantially) from regular trades posted by other market
participants. The more liquid the market, the harder it will be to make this distinction. In the
basic analysis, we allow the HFT to distinguish “large” trades (the initial trade by LT1 at date
1 and LT2’s trade at date 2) from “small” trades (trades by PTs and LT1 at t = 2). Thus, the
HFT sets two haircuts, one for large trades ∆L , and one for small trades ∆S .8
7
It is possible to introduce a strategic element whereby the trader alters his bidding behavior in anticipation
of the HFT. This can greatly complicate the model. Basically, it affects how much of the trading surplus is
extracted by the HFT and imposes an additional distortion on trade.
8
In the Appendix we have included more general formulas and additional results for the case where the
haircut is constant across dates and traders.

17
 Since LT2’s excess demand is θ2LT 2 (P2 ) + i, the HFT applies the large haircut to him and
sells i + θ̃2LT 2 shares at the (higher) price, P2 + ∆L , and θ2LT 2 (P2 ) − θ̃2LT 2 , at the market price
P2 . Thus, given an initial wealth of W2LT 2 , LT2’s final wealth is

 
W3LT 2 = W2LT 2 + (P3 − P2 ) θ2LT 2 + i − P3 i − θ̃2LT 2 + i ∆L ,

where we are omitting the functional dependence of asset holding on prices, using θ2LT 2 instead
of θ2LT 2 (P2 ).

Substituting for wealth in the utility function, we obtain:

     1 
LT 2

LT 2 LT 2 2 2 2
E [U (W3 ) |F2 ] = − exp −a θ2 + i (µ2 − P2 ) − iµ2 − W2 − θ̃2 + i ∆L + a σ (θ2 ) .
2
(1)
LT2’s final excess demand, θ2LT 2 + i, is such that it maximizes (1). Hence,

µ2 − P2
θ2LT 2 + i = + i. (2)
aσ 2

Similarly, LT1’s and PTs’ excess asset demands at t = 2 are given by

µ2 − P2 µ 2 − P2
θ2LT 1 − i = −i and θ2T = .
aσ 2 aσ 2

These three expressions reflect the frictionless introduction of HFTs, as asset demand sched-
ules (at t = 2) are unaffected by the known presence of HFTs (relative to the original GM
model).

Market clearing requires that the excess demand of LT1 (who arrived at date 1), plus PTs’,
plus LT2’s sum to zero:
θ2LT 1 − i + M θ2T + θ2LT 2 + i = 0 ,

so that the equilibrium price is P2 = E [P3 |F2 ] = µ2 , and the asset holdings after trade in
date 2 are θ2LT 2 = θ2LT 1 = θ2T = 0. The quantities used by the HFT for extracting surplus are

18
obtained from the corresponding demand functions θ̃2LT 1 = θ̃2T = θ2LT 1 (P − sign (i) ∆S ) and
θ̃2LT 2 = θ2LT 2 (P + sign (i) ∆L ):

∆S ∆S ∆L
θ̃2LT 1 = sign (i) , θ̃2M = sign (i) , θ̃2LT 2 = −sign (i) .
aσ 2 aσ 2 aσ 2

The HFT will extract trading surplus equal to9

     
Π2 (i) = θ̃2LT 2 + i ∆L + θ̃2LT 1 − θ1LT 1  ∆S + M θ̃2T − θ1T  ∆S . (3)
     

The benchmark model: Trading at t = 1

Traders at date t = 1 anticipate what will happen at t = 2. They also know that there will be
public information revealed prior to date 2 trading and that P2 = µ2 = µ + 2 is random and
normally distributed with mean µ and variance σ 2 .

Traders’ future wealth can be written as

   
W3LT 1 = W0LT 1 +(P2 − P1 ) θ1LT 1 − i +(P3 − P2 ) θ2LT 1 − i +P3 i− θ1LT 1 − θ̃2LT 1 ∆S − i − θ̃1LT 1 ∆L

and
   
W3T = W0T + (P2 − P1 ) θ1T + (P3 − P2 ) θ2T − θ1T − θ̃2T  ∆S − θ̃1T  ∆S ,
   

where θtLT 1 − i and θtT are LT1’s and PT’s excess demands respectively in dates 1 and 2.

Using E [P2 |F1 ] = E [µ + 2 ] = µ, simplifies the expression for traders’ wealth, and it is
straightforward to derive optimal excess demands:

1 1
θ1LT 1 − i = (µ − P1 − sign (i) ∆S ) − i and θ1T = (µ − P1 − sign (i) ∆S ) .
aσ 2 aσ 2

Notice that traders anticipate that their current asset demand (and hence positions at the end
of trading at t = 1) will affect future trading and hence the (date 2) haircuts they will have to
9
We are assuming that the PTs and LT1 are net sellers of the asset. This is confirmed as in equilibrium PTs
will be net buyers in the first period and have zero net final holdings (and θ2LT 1 − θ1LT 1 has the same sign as
PTs’ changes in asset holdings at date 2).

19
pay. As they must pay these haircuts whether they are buying or selling, they move their asset
demand and supply a small step away from the no trade price (P1 = µ), essentially introducing
an additional spread. Thus, the HFT’s strategy, though statically frictionless, does generate a
dynamic distortion, as future haircuts affect current asset demands.

The date 1 market clearing condition is

θ1LT 1 − i + M θ1T = 0 .

From this we obtain: (i) the market clearing price

iaσ 2
P1 = µ − sign (i) ∆S − , (4)
M +1

(ii) traders’ asset holdings after trading in date 1,

i i
θ1LT 1 = and θ1T = ,
M +1 M +1

and, using θ̃1LT 1 = θ1LT 1 (P − sign (i) ∆L ) and θ̃1T = θ1T (P + sign (i) ∆S ), (iii) the quantities
that are subject to haircuts

i ∆S i ∆S
i − θ̃1LT 1 = i − − sign (i) 2 and θ̃1T = − sign (i) 2 .
M +1 aσ M +1 aσ

Again, the HFT is able to extract trading surplus and generate profits of:

   
Π1 (i) = i − θ̃1LT 1  ∆L + M θ̃1T  ∆S . (5)
   

In summary, from the above analysis we extract the following conclusions:

Theorem 2.1. For a given order imbalance of magnitude i > 0:

20
 1. Market clearing prices are

iaσ 2
P1 = E [P2 |F1 ] − − sign (i) ∆S , (6)
M +1
P2 = E [P3 |F2 ] = µ2 ,

and the liquidity discount at t = 1 is

iaσ 2
E [P2 |F1 ] − P1 = + sign (i) ∆S . (7)
M +1

2. Asset trading (for i > 0) is described on Table 1:

Table 1: Prices and Volume of Trades

Price LT1 LT2 PT (total)

∆L
P1 − ∆L − MM+1 i + aσ 2

∆L ∆S
P1 − aσ 2 M aσ 2

M ∆S
P1 + ∆ S M +1
i − M aσ 2

∆L
P2 + ∆L i− aσ 2

∆S ∆L ∆S
P2 − aσ 2 aσ 2
−M aσ 2

∆S ∆S
P2 − ∆S − M1+1 i+ aσ 2 − MM+1 i+M aσ 2

3. The HFT acts as counterparty to all these trades and makes profits equal to

Π (i) = Π1 (i) + Π2 (i) , (8)

where Π1 (i) and Π2 (i) are described in Equations (5) and (3) respectively.

The presence of HFT intermediation affects the market clearing price at t = 1, when the
order imbalance is initiated. Traders, anticipating future haircuts imposed by HFTs, amplify
the selling pressure of the liquidity trader initiating the order imbalance (LT1), driving the date
1 market clearing price away from what it would have been without HF trading (that is, with

21
∆L = ∆S = 0), and increasing the liquidity discount paid by LT1 to encourage PTs to buy the
asset from him (or increasing the liquidity premium paid to buy it, see Equation (7)).

Table 1 illustrates a number of the properties and implications of our stylized model. For
example, our model exhibits (additional) “microstructure volatility” induced by HFT interme-
diation. We can observe how the HFT’s surplus extraction, which involves buying and selling
at different prices, introduces additional prices at which transactions take place. These price
movements hide the increase in the liquidity discount, as prices close to the “fundamental”
price are executed between PTs and the HFT, to the HFT’s advantage, while the LTs suffer
price execution away from the “fundamental” price. An outside observer could conclude that
there were many transactions with small price differences (and narrow spreads) when in fact
there was only one “true” liquidity transaction, suffering a wide spread.10 In addition, the
price movements around the market clearing price have no informational content and may lead
a casual outside observer to erroneous conclusions about what the “true” market price of the
asset should be at each date. These properties and implications are supported by the empirical
findings in Brunetti et al. [2011].

Naturally, having allowed the HFT to intermediate all trades, the volume of trade doubles.
This assumption is weakened in Section 3. Nevertheless, these additional trades (and hence
HFTs) do not add liquidity–in fact, the presence of HFTs results in an increase in the liquidity
discount, thus liquidity is reduced. Hence, rebate schemes based on volume are likely to dis-
proportionately benefit HFTs who are not providing liquidity, as the following simplistic rebate
scheme illustrates:

Corollary 2.2. If there is a rebate of c cents per share, the HFT gets half of all rebates, and
the M professional traders split a quarter of all rebates between them.
10
In this context a spread corresponds to the price difference from the starting point, E [P3 ] = µ, and we
can think of it as a half-spread or movement away from the midpoint, even though, as we discuss below, the
appropriate way to measure the effects we are addressing is via the overall cost of execution. An outside observer
(following i > 0) may observe µ followed by P1 + ∆S , P1 , P1 − ∆L and think that these where three transactions
with small price increments (half-spreads) while in fact the actual change in the market price, to P1 , includes a
greater liquidity discount.

22
2.5 Optimal Haircuts

Having captured a trade, the HFT realizes that her profits depend on the size of the haircut and
the constraint imposed by the minimum tick size. We discuss this possibility and its impact
on the our results below. Without this constraint, the HFT faces a trade-off between a larger
haircut, and a smaller number of assets traded using that haircut. Thus, she will adjust her
behavior by setting haircuts that maximize the profits of her trading activity. As the HFT
distinguishes the haircut for large trades (∆L ) from the one for the small trades (∆S ), the
HFT’s profits described in Equation (9) can be expressed as:11
   
M ∆L M ∆S
Π (i) = ∆L |i| − 2 + ∆S |i| − M 2 , (9)
M +1 aσ M +1 aσ
   
∆L (M + 1) ∆S
+∆L |i| − 2 + ∆S |i| −
aσ aσ 2

and by maximizing (9) with respect to ∆L and ∆S we obtain:

Lemma 2.3. The HFT maximizes profits by choosing

1 1
∆S = |i| aσ 2 ,
2 (M + 1)
 
1 2M + 1 1
∆L = |i| aσ 2 = ∆S M + .
4 M +1 2

The size of the haircuts is increasing in i, the size of the liquidity need, because of the
pressure from LTs. This implies that the corresponding additional microstructure noise and
liquidity discounts will increase for larger order executions. If we interpret the number of PTs,
M , as a proxy for the amount of “real liquidity” available for the traded asset, then the haircut,
microstructure noise and liquidity discount decrease with available liquidity.

Furthermore, since the HFT can distinguish between large and small trades, she will impose
bigger haircuts on large trades than on small trades, and the increase in haircut for large
trades is proportional to the amount of “true” liquidity. This seems counterintuitive, but as we
mentioned above and will see in more detail below, in the competition section, “true” liquidity
11
In the Appendix we discuss another case (where the HFTs make no distinctions and apply the same haircut
to all trades).

23
has the additional effect of making it easier for LTs to remain undetected, and hence making it
harder for the HFT to distinguish large from small trades thereby counterbalancing this effect
on haircuts.12

Figure 2 below illustrates how the haircut for the large (small) trades increases (decreases)
with M , and Figure 3 shows the optimal asset holdings, θtx and θ˜tx , for the liquidity traders
and PTs.

Figure 2: Optimal Haircuts for large and small trades

0.5
∆LT1
1 = ∆LT2
2
∆T1 = ∆T2 = ∆LT1
2
0.45

0.4
Demands assuming i = a = σ2 = 1

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
1 2 3 4 5 6 7 8 9 10
Number of PTs

Moreover, we see that in a market with more PTs, the HFT increases the haircut imposed
on large trades from liquidity seekers, LT1 at date 1 and LT2 at date 2, and lowers it on small
trades. When an HFT can identify large trades, in the limit as the number of PTs goes to
infinity,
1
lim ∆S = 0 and lim ∆L = iaσ 2 ,
M →∞ M →∞ 2
12
Note also that by making the haircut smaller for small liquidity trades, the HFT ensures that PTs will
participate and hence she can make the roundtrip trade and profits. As we will see below, with a fixed tick size
there may not be sufficient “true liquidity” to absorb a haircut and the HFT may abstain from intermediating.

24
Figure 3: Optimal asset holdings when the HFT sets a haircut for large trades and another
haircut for small trades.
1
θ 1T1 = θ 1T
θ̃ 1LT1
θ 1T = θ 1LT 1
θ̃ 1T
θ̃ 2LT1 = θ̃ 2T
θ̃ 2T = θ̃ 2LT1
Demands assuming i = a = σ2 = 1

θ̃ 2LT2
0.5

−0.5
1 2 3 4 5 6 7 8 9 10
Number of PTs

we see that the haircut on small trades disappears and all the trading surplus extracted by the
HFT comes from large trades. In addition, as the number of PTs goes to infinity, the liquidity
discount (Equation (7)) also goes to zero, so that LT1 and LT2 behave in the same way: their
liquidity needs are immediately satisfied (LT1 sells i in date 1), but they trade half their shares
at a price of µ ± 12 iaσ 2 and the other half at the asset’s fundamental value, µ. Then, overall
the HFT then makes profits equal to 12 i2 aσ 2 .

2.6 Entry and Exit of PTs

The equilibrium level of liquidity in the market is determined by the number of PTs, M .
These traders have to forgo alternative investment opportunities in order to tie their capital
so as to provide liquidity. We incorporate this opportunity cost using a fixed entry cost c. In
equilibrium, capital will be diverted to providing liquidity in the market as long as the gains
from this activity cover the entry cost. Below we see that, as expected, the presence of HFTs
affects the gains from providing liquidity.

25
 We assume that the liquidity shock is normally distributed (i ∼ N (0, 1)) and independent
of the other (fundamental) shocks affecting prices. The liquidity provider must decide whether
to sink the cost c to enter the market before he observes the size and direction of the liquidity
shock i. At time t = 0 the expected utility of an individual PT is E U W3T − c  F0 , where

 

W3T is given by (24). Entry of PTs will occur until

E U W3T − c  F0 ; q = 1 = E U W0T  F0 ; q = 1 ,

  
 
(10)

where, as we saw above, θ2T = 0. To calculate the expectations in equation (10) we require the
expected value and variance of wealth in period three.

1


In the benchmark model, using ∆L = ∆S M + 2 , the expected value of wealth can be
decomposed into a liquidity discount and a loss of surplus:

E W3T  F1 ; q = 1 = W0T − c
  

iaσ 2
+θ1T + θ1T ∆S , liquidity discount, (11)
M +1
1  
−θ1T 2∆S + 2 2 (∆S )2 . loss of surplus. (12)
aσ

From equation (11) we observe that the PT receives a liquidity discount for buying θ1T shares
from period 1 to 2. This discount consists of two terms. The first term iaσ 2 /(M + 1) is the
liquidity discount (per share) that a PT receives in the absence of an HFT; and the second term
(θ1T ∆S ) is the extra discount per share that a PT receives if there is an HFT intermediating
all trades. On the other hand, the term in equation (12) identifies the loss of wealth from the
HFT’s surplus extraction strategies.13
13
Note that there are two terms in equation (12). The first term is surplus extraction if the HFT took a
haircut in periods 1 and 2 for all shares that the PT buys and sells, but recall from above that in every period
the HFT breaks PT’s order in two batches and only applies a haircut to one batch. Therefore, the second term
in equation (12) is a correction term that accounts for the quantities which are traded in both periods without
paying a haircut.

26
 As for the variance of wealth, it depends on the fundamental volatility of the asset which
is unaffected by the presence of the (faultless) HFT:

2
Var W3T  F1 ; q = 1 = σ 2 θ1T .
  
(13)

The extra liquidity discount in the presence of an HFT (second term in equation (11)) exactly
offsets the haircuts (12) paid to the HFT in the round-trip trade (buy from LT1 in period 1
and sell to LT2 in period 2), and the number of PTs will be the same with or without HFT.

This implies that HFTs have no effect on equilibrium market liquidity in the benchmark case
with optimal haircuts. There are several factors which will move us from this knife-edge case:
limitations on the possibility to filter information from incoming trades and distinguish liquidity
shocks from less aggressive trades from LTs, competition from other HFTs, or minimum tick
size.

The simplest case is that of a constant, but optimal haircut. This case is analyzed in detail
in the Appendix. What we find is that the optimal haircut implies a substantial liquidity
discount which overcompensates PTs for the haircut and leads to increased entry.

Alternatively, we can consider the case of an exogenously fixed and constant haircut ∆, such
as the one that could occur if there is a minimum tick size. One of the problems that arises in
this case is whether the HFT will always be able to impose the haircut regardless of the size of
the liquidity shock i. Recall that the size of the liquidity discount required by the PTs depends
on the size of the liquidity shock – the smaller is the liquidity shock, the smaller is the liquidity
discount. If the liquidity shock is small relative to the haircut imposed by the HFT, PTs will
not want to provide liquidity. Then, the HFT must either provide liquidity or withdraw. We
assume that in such circumstances the HFT withdraws completely. To incorporate this into the
model we need to introduce a significant variation in the model: assume there is a threshold
level of i, ξ, such that if |i| < ξ the HFT does not intermediate the trade, and that at any time
all market participants know whether |i| is above or below ξ.

27
 This assumption alters the main expressions derived above, but the qualitative nature of
the analysis remains the same. Let 1|i|≥ξ denote the indicator function for the event {|i| ≥ ξ}.
Then, LT1’s demand function at date 1 is:

1
θ1T (P1 ) =


2
µ − P1 − ∆1|i|≥ξ .
aσ

The equilibrium prices and demands have a similar structure as in the previous case.

Proposition 2.4. The equilibrium price and PTs’ optimal asset demands are

iaσ 2
P1 = µ − sign (i) ∆1|i|≥ξ − ,
M +1
i
θ1T (P1 ) = .
M +1

Under these circumstances, at date one, conditional on the size of i, the expected wealth of
a PT is given by
2
∆2
  
i i
W3T  F1 W0T 2
  
E = − c + aσ − ∆−2 2 1|i|≥ξ .
M +1 M +1 aσ

We can now revisit the PTs entry and exit decision. The optimal entry decision can be reduced
to a straight-forward analytical expression involving the error function, and we find two effects:
a negative effect from the presence of the constant haircut, and a tempering effect from the
limited participation by HFTs (through) ξ. Our numerical calculations show that the second
effect does not eliminate the first, so that overall the presence of a constant haircut reduces
liquidity.

Thus, the overall effect of HFTs on liquidity is ambiguous and depends on several factors.
In particular, it depends on the ability of HFTs both in processing the information to identify
the liquidity pressure in the market, and on the possibility of tailoring the haircut to market
conditions.

28
3 Competition in the expanded model

In the previous section we analyzed the benchmark model with a single HFT who intermediated
all trades and set the haircut. We saw how the haircuts affected trading behavior and market
prices. In particular, we saw how it affected trades, prices, and the liquidity discount. In this
section we expand the model to incorporate additional interactions between HFTs and between
HFTs and LTs, so as to capture the tempering effect of competition on the ability of HFTs to
intermediate “fundamental” trades and profit from doing so.

3.1 Modeling competition amongst HFTs and entry decisions

Our discussions with regulators and practitioners on both sides of the HFT fence identified a
number of issues regarding the way in which HFTs profit from financial intermediation (and
incorporate issues raised in the previous section): (A) Becoming an HFT, whether it is a new
market entrant or a PT that decides to upgrade, requires high initial sunk costs in hardware,
co-location, software and data, but, most importantly, it demands skills specific to HF trading,
all of which pose barriers to entry. (B) Traders, both LTs and PTs, have reacted to changes in
market conditions. One of these changes is that they break up their trades into small blocks
to try to mitigate their price impact when they take significant positions; a proportion of these
trades are captured directly by PTs (bypassing HFTs). (C) An HFT’s ability to position herself
at the front of the queue (and capturing an incoming trade) is no guarantee that the HFT will
be able to close the trade at a profit. (D) An HFT cannot know whether an incoming trade is
a liquidity trade or if it is coming from another trader, specially another HFT.

In order to incorporate competition into our model we make two assumptions: that there
are entry barriers, and that competition centers around intermediating trades and not around
price discounts.

The existence of entry barriers can take two forms as mentioned in (A), both of which
we include. The first is entry barriers in terms of initial costs of entering the HFT market.
The publicity surrounding HF trading profits attracts competition, specially from PTs who

29
understand the usefulness of the additional speed and processing advantages. These would
be the first to be able to value the costs and benefits from becoming HFTs, and whether the
investments required for entry are worthwhile or not. The second is skill. There is a steep
learning curve to become a profitable HFT, and it is not clear that differences with early
entrants can be reduced. The joint effect of these barriers is to limit the number of entrants,
something which is consistent with available evidence: In the literature and discussions with
market participants we find that the number of HFTs has not increased in some markets and
increased in others, but the potential profits for the latest and future entrants have dwindled
to the point of being essentially negative. Furthermore, despite the increase in the number
of HFTs, profits from early entrants seem to remain stable. Empirically, the precise number
of HFTs operating in different markets and venues, and how this figure has evolved over the
last few years, is largely unknown. Exceptions are the study by Kirilenko et al. [2010] which
identifies 16 HFTs, out of the almost 12,000 active traders, during the Flash Crash, Brunetti
et al. [2011] identify 11 HFTs, out of approximately 15,000 active traders, and that of Menkveld
[2011] which discusses the role of a single monopolist HFT that operates in Chi-X. What we
conclude from this is that competition amongst HFTs is present though limited.

The assumption on the nature of competition is based on three arguments. First, price
competition is naturally restricted by the minimum tick size. Second, price competition should
have wiped out HFT profits, something that is inconsistent with the information received from
financial markets. A standard Bertrand price competition model tells us that you only need
two competitors to drive HFT profits to zero, and that is not happening. And third, HFTs’
relative advantage is in exploiting their edge in speed and information processing. It is more
about skill and technology (in capturing and closing a trade, and positioning trades in the order
book) than about offering the best execution price.

We implement our first assumption by imposing an entry cost and incorporating skill dif-
ferences into our model of competition. The second assumption is incorporated by fixing prices
(we keep the haircuts used above, ∆S and ∆L , as given), and focusing on the effects of compe-
tition between HFTs on quantities (trades intermediated). In particular, we assume that HFTs
(as a block) do not profitably intermediate all trades between LT and PTs, but only a fraction

30
of those, denoted by ϕ. Equivalently, ϕ is the probability that HFTs capture the whole trade,
and does so profitably. Skill differences is what determines how trades intermediated by HFTs
are distributed amongst competing HFTs. In particular, conditional on the HFTs capturing an
incoming liquidity trade, we assume that HFT j = 1, . . . , n, obtains expected profit pj Π (i),
where pj is the probability that HFT j is the one to capture the trade.14

We focus on two factors that determine how skills and competition affect ϕ and p: the
number of competing HFTs, n, and each HFT’s (relative) skill which is parametrized by ηj .
The skill parameter, ηj , determines the success of HFT j in capturing a round-trip trade. In
our model a high ηj means that HFT j’s probability of capturing a trade that yields profit Π (i)
is higher; thus we let pj = p (ηj ) be an increasing function of ηj . The individual skill parameter
ηj allows us to capture differences between early and late entrants, and in particular, to be
able to have successful early HFTs coexisting with marginally profitable new entrants. It also
captures the idea that entrants can continue to extract profits as long as they can retain their
skill advantage at capturing rents from HF trading both relative to other HFTs, and other
traders in general.

The probability of HFT intermediation, ϕ, is affected by two factors. First, there is a direct
effect of competition: a higher number of competitors, n, leads to a lower ϕ. This captures
the effect that more HFTs will generate noisier trading (related to issues (C) and (D) above),
and hence reduce the overall effectiveness of HFT profit extraction strategies. As ϕ is also the
fraction of trades that are profitably intermediated, it reflects the effect that with a greater
number of HFTs a trade, even if captured, may be intermediated at a zero profit (excluding
rebates).

A second key factor is the relative technological gap between HFTs and other traders, as
suggested by (B) above. The individual skill parameter allows us to calculate an average skill
14
This assumption is consistent with the following anecdotal description of HF trading behavior: HFTs
position orders in the queue so that their bids and offers are hit or lifted whilst maximizing the probability
that they close the position at a profit over a very short-term horizon. These strategies require the HFTs to
constantly cancel and send orders to both sides of the book as well as orders that could be passive or aggressive
to jockey for position in the limit order book. Thus, once a market order hits a bid or lifts an offer, it is not
surprising to see a large number of orders disappear from both sides of the order book. The HFTs who are
unable to capture the trade or are not adequately positioned, cancel their orders and withdraw, leaving the best
positioned HFT to intermediate the incoming trade.

31
 1 Pn
parameter: η̄ = n j=1 ηj . A greater technological gap between HFTs and other professional
traders (a higher η̄) increases the proportion of trades that are profitably captured by HFTs.
These two factors are incorporated into the probability that HFTs (as a block) profitably
intermediate trades, as described by the function ϕ (n, η̄), where ϕ (n, η̄) is decreasing in n and
increasing in η̄.

3.2 The model, expanded to include competition

We incorporate competition in HF activities in two stages. First, we describe the outcome of

competition when there is a fixed number, n, of HFTs who compete to extract surplus when a
liquidity trade of size i is brought to the market. Second, we consider the entry decision of a
potential new HFT—for example a PT who considers whether or not to become an HFT—who
must make an initial investment (sunk cost) to compete with other HFTs.

Suppose that an LT enters the market with a liquidity need of size i. Now, market partic-
ipants face uncertain HFT intermediation. Overall, this increase in realism comes at the cost
of analytical tractability. In particular, asset demands at t = 1 are obtained as the solution to

1
θ1 − q (θ1 ) ∆S = (µ − P1 ) ,
aσ 2

where
1
q (θ1 ) =    . (14)
1 + 1−ϕ
ϕ exp a θ 1 − θ̃ 2 ∆ S

These demands correspond to an intermediate case between the GM model (where q = 0)

and the benchmark model (where q = 1) and, as expected q ∈ [0, 1] is increasing in ϕ. We
lose analytical tractability because q is not a constant, but a function of θ1 . The additional
uncertainty requires a relatively greater liquidity discount, nevertheless, the qualitative effects
remain unchanged.

A trade captured by an HFT represents a potential profit of Π (i), as approximated by those

described in Equation (8) above. As a block the HFTs compete for the expected profit ϕΠ (i),

32
but the individual success of HFT j in profiting from a particular trade is determined by the
probability that once captured, the profits from intermediation go her, pj .

To make the effect of relative skill differences on the profitability of being HFT analytically
tractable we introduce additional assumptions. For each trade and each HFT j, there is an
i.i.d. realization θj of a random variable Θj . We assume that Θj is exponentially distributed
with parameter ηj (the HFT’s relative skill), i.e. Θj ∼ exp (ηj ). Given the realizations of Θ for
all HFTs, (θj )nj=1 , the trader with the smallest realization “wins the trade”. Thus, given ϕ, the
probability that trader j extracts profits from the trade is given by

ηj
ϕ (n, η̄) p (ηj ) := ϕ Pr {Θj = min {Θ1 , . . . , Θn }} = ϕ Pn .
u=1 ηu

Hence, the expected profit for trader j from a trade of size i is ϕp (ηj ) Π (i). The reader
can readily verify that if one were to introduce the parameters ϕp (ηj ) in the optimal haircut
problem in the benchmark model of Section 2 the results do not change.15

Now we analyze the decision of a potential entrant. Suppose a PT observes the activities of
n HFTs and considers becoming an HFT himself. The PT identifies the irreversible investment
costs K of setting up an HF trading desk (hardware, software, human capital, co-location,
etc.), including the opportunity cost of abandoning current activities such as traditional market
making. The potential entrant does his research and determines that if he were to obtain a skill
level η from his investment, the expected profits would depend on p (η) and on the expected
number of liquidity trades N (i) of size i. Thus PT’s expected utility from future HFT is
ϕp (η) E [U (Π (i) ; N (i))]. Then, a PT with skill η compares the utility from wealth K, U (K),
with the expected utility from future profits, and becomes an HFT if

k ≤ ϕp (η) ,
15
With the possible exception of the size of the optimal haircut. Nevertheless, the optimal haircut as described
in the previous section could still be valid if the aggregate HFTs’ price setting behavior emulates that of a
monopolist–maximizing the sum of profits from both the incoming liquidity shock and the subsequent one,
canceling the first (by LT2). This could occur in a context with a small number of major players who are able to
reach tacit collusion agreements. Alternatively, the minimum tick size could act as an effective barrier to price
competition. In the latter case, the haircut would not be optimal, but fixed endogenously, and hence unaffected
by competition.

33
where k = U (K) /E [U (Π (i) ; N (i))] denotes the relative “utility” cost of a trade.

The entrant has a reasonable estimate of k, but faces uncertainty with respect to the effec-
tiveness of the investment, η and ϕ, though uncertainty on the latter is relatively insignificant
for large n (where the relative importance of ηj on η̄ is small). For simplicity we assume that the
skill of the entrant, η ∈ (1, ∞), is random and with a distribution such that its inverse, 1/η, is
1
uniformly distributed, that is η ∼ U (0, 1), and the effect on η̄ is negligible. As discussed above,
the probability of becoming the HFT that intermediates a liquidity trade depends not only on
the HFT’s skill but more importantly, on her skill relative to those of existing HFTs. Thus,
any potential entrant needs to take into account that there are n HFTs with skills described by
parameters {ηu }nu=1 . Then, an entrant with skill η becomes the monopolist with probability

 Pn −1
η i=1 ηi
p (η) = Pn = 1+n .
η+ i=1 ηi η
Pn
Let αj = i=1 ηi > 1 denote the total skill pool of other existing HFTs (excluding j)–we
drop the j subscript until the final part of this section when it becomes useful. Then, the
cdf and pdf of p (η) are Gp (z) = α1 1 + α − z1 and gp (z) = α1 z12 respectively, and E [p (η)] =

1
α log (1 + α).16

Given a relative cost of entry of k, and letting α̂ denote the total skills of incumbent HFTs,
there will be entry until

ϕ
k ≥ log (1 + α̂) .
α̂

This condition establishes the relationship between entry costs, relative skill levels, and
the probability of successful intermediation by HFTs. But, as our model does not incorporate
positive effects from HFT, we cannot use it to make a reasonable welfare analysis. Nevertheless,
allowing for positive effects from increased execution speed, or fewer welfare-enhancing trades
being lost, as in Foucault et al. [2011], we would find externalities in the entry decision by a
16
See the Appendix for details.

34
PT. A PT’s entry calculation of the value of becoming an HFT does not account for the fact
that it affects the total amount of HFT intermediated trades (unmodeled, but the effect could
be positive) and the amount of trades profitably intermediated by HFTs. The latter effect is
ambiguous, as an increase in n reduces ϕ, but the skill of the entrant can be above or below η̄,
and a sufficiently high η entrant could have a positive net effect on ϕ. Overall, as discussed in
Foucault et al. [2011], the role for regulation depends on whether these externalities bias PTs
towards excessive or insufficient entry.

Beyond the externalities of becoming an HFT, our model also allows for an analysis of the
(unmodeled) dynamics of skill differences. In particular, if skill differences between early and
later entrants are not eroded over time, differences in profitability can become permanent. This
would lead to a situation in which, as HFTs’ relative skills stabilize, the market settles around
a core set of players who extract profits from a large fraction of trades, while less skilled, later
entrants compete for the remaining trades and scrape enough to cover their operating (and
opportunity) costs. An additional dynamic factor to consider is that, as other non-HF traders
upgrade and adapt their trading strategies and technology to the presence of HFTs, the overall
technological gap (η̄) decreases and the overall profitability of HF trading is eroded.

4 Measuring the impact of HFTs on financial markets

Empirical studies of HFT offer seemingly conflicting conclusions on the effect of HFTs. On
the one hand, Hendershott et al. [2011] find that AT improves liquidity and narrows effective
spreads and Brogaard [2010] finds that HFTs contribute to price discovery and reduce volatil-
ity. Although the study of Brogaard [2010] identifies a subset of HFTs, his data set cannot
differentiate how much of the activity of the 26 firms in his study is AT and how much is HF
trading and there might be large market participants that use HF strategies which are not
identified in the data as HFTs; the same applies to the results of Hendershott et al. [2011].
On the other hand, there are the studies that employ proprietary audit-trail data, SEC [2010],
Kirilenko et al. [2010], and Brunetti et al. [2011]. They find evidence that HFTs increase price
impact, increase volume of trading without increasing the flow of information, and increase

35
noise around the fundamental value of the asset. Our abstract model allows us to identify
the effect of HFTs and how they may be affecting different aggregate financial metrics such as
volume, liquidity, price impact, and price volatility. This we do here, and find that our model
is more consistent with the results from the latter line of empirical research with data that
includes trader identity.

Throughout this section we focus on the two extreme cases where the HFTs capture all
incoming trades and the benchmark GM model where there are no HFTs; i.e. we assume
that q = 1 and q = 0 in (14) respectively. The general case, with competition and imperfect
intermediation, falls in between these two.

4.1 Liquidity, Volume and Price Impact

Our stylized model is designed to capture the value of a stock market as a forum where equity
holders can convert their equity into cash (and viceversa) quickly and at a reasonable price. In
practice it is usually hard for an outside observer to determine both the time and the overall cost
of execution of significant blocks of shares, as share blocks are split in several (and increasingly
many) parts executed separately, and the identity of traders is not available for empirical
analysis.17 Thus, most studies, by necessity, focus on indirect measures of liquidity such as
bid-ask spreads, execution speed, and fill rates, for small trades (less than 10,000 shares).

In our stylized model, liquidity (speed and cost of execution) is essentially measured in two
ways: (i) how many shares can the liquidity trader sell when he originates a trade imbalance,
that is, how many shares does LT1 sell at date t = 1, relative to his desired total sale of i units;
and (ii) what is the difference between the revenue obtained by liquidity traders when selling
(the cost paid when buying) their total trading blocks (−i and i) relative to executing them at
the asset’s “fundamental” value.
17
This is recognized by the SEC in their report [SEC concept release Jan10] “Measuring the transaction costs
of institutional investors that need to trade in large size can be extremely complex. These large orders often are
broken up into smaller child orders and executed in a series of transactions. Metrics that apply to small order
executions may miss how well or poorly the large order traded overall.” (p38-39)

36
 In our model we are able to identify two ways in which the HFT affects the amount of
shares the liquidity trader can sell at t = 1: a direct and an indirect way. Although, having
identified them we find that in both ways the net effect of HFTs on quantities is zero.

The direct way is the effect of the HFT’s surplus extraction strategy on trader’s demand for
assets (both on the shape of the demand function and on the equilibrium quantity demanded).
Direct effects are limited. First, by assumption we have ruled out strategic effects on the part
of traders vis-a-vis the HFT, which would be a natural source of distortions of asset demand
functions and reduced trading (and hence liquidity). Second, the HFT’s surplus extraction
process is faultless. This, together with her extreme aversion to holding any inventory leads
to no distortions in the final quantity of assets demanded: LT1 continues to carry over θ1LT 1 =
i/(M + 1) to date 2—the same amount as if there were no HFT.

The indirect way is through the effect of the HFT on the number of PTs, which in turn
affects liquidity (in terms of the number of shares sold at t = 1). Below (and in more detail
in Section ??) we find that HFTs have both a negative and a positive effect on PT’s revenue,
which, when q = 1, exactly offset each other. Thus, in the benchmark model HFTs have neither
a direct nor an indirect effect on the amount of LT1’s initial liquidity need i that is immediately
executed–leaving “fill rates” (the percentage of the desired trade executed) and “trade delays”
(the time to completion of desired trade) unaltered.

Liquidity can also be measured in terms of the cost of executing trades. We first look at
average price paid and received; and then we look at the total revenue and total cost for each
type of trader when selling and buying shares. As benchmark we use market clearing prices in
the absence of an HFT. In that case the market clearing prices (which are the same for buyers
and sellers) are:

iaσ 2
P1∆=0 = µ − , P2∆=0 = µ2 .
M +1

This implies that without the HFT, LT2 faces infinite liquidity (buys at the asset’s fundamental
value), while LT1, the originator of the trade imbalance, faces a liquidity discount at date t = 1
iaσ 2
of M +1 (given to PTs).

37
 When the HFT is present (q = 1) we observe two effects on average execution price: traders
pay the HFT a haircut on a fraction of their trades, and the liquidity discount increases. We
start with LT2’s simpler problem, as he buys shares at their fundamental value. LT2 wants to
buy i units (trade i) and ends up buying θ̃2LT 2 at P2 + ∆L per share, followed by i + θ2LT 2 − θ̃2LT 2
at P2 per share. Then, using the equilibrium values of P2 , θ2LT 2 and θ̃2LT 2 , the average price
paid by LT2 is:

θ̃2LT 2 ∆L (∆L )2
P̄2LT 2 = P2 − = µ2 + ,
i iaσ 2
  2
∆=0 1 1
= P2 + 1− iaσ 2 . (15)
4 M +1

From Equation (15), we can see that LT2 pays more to acquire his position than when there is
no HFT. The price uplift payed by LT2 is given by the second term on the right-hand side of
equation (15) which is the surplus per share lost to the HFT.

PTs trade twice. At date 1 they buy the asset, receive the liquidity discount and pay the
haircut to the HFT; at date 2 they sell and pay the haircut. The average purchase price at
date 1 is:

θ̃1T iaσ 2 1 iaσ 2

P̄1T = P1 + ∆ S = µ − −
θ1T M +1 4M +1
1 iaσ 2
= P1∆=0 − , (16)
4M +1

from which we see that PTs receive an even better price than without HFT—they manage a 25%
higher liquidity discount, even after accounting for the haircut paid to the HFT. Nevertheless,
at date 2, the average sale price they obtain is:

T
θ̃2 1 iaσ 2
P̄2T = µ − ∆S + ∆ S = P ∆=0
2 − , (17)
θ1T 4M +1

so they end up selling back at a lower price (relative to the case without HFT), and losing the
initial extra liquidity discount garnered from the first transaction.

38
 Finally, LT1 offers a liquidity discount and pays a haircut at date 1. In exchange he receives
P̄2LT 1 when he completes his liquidity trade at t = 2. The average price received by LT1 at
date 1 is:

i − θ̃1LT 1
P̄1LT 1 = P1 − ∆L
i − θ1LT 1
!
1 M 2 + 5M + 54
= P1∆=0 − iaσ 2 , (18)
4 M 2 + 2M + 1

and at date 2:
 LT 1

µ2 θ1LT 1 − θ1LT 1 − θ̃2 ∆S
P̄2LT 1 = ,
θ1T 1
1 iaσ 2
= P2∆=0 − . (19)
4M +1

This implies that LT1 receives less money for liquidating his position at both dates (relative
to the case without HFT) as can be seen in Figure 4.

From equations (15-19), it is clear that LTs, LT1 and LT2, bear the brunt of the haircut
imposed by the HFTs, and that this impact is increasing in i. Although PTs suffer the loss of
consumer (date 1) and supplier (date 2) surpluses, the liquidity discount received by the PTs
for shares at date 1 is greater (the equilibrium price is lower) than in the absence of the HFT,
and the net effect on PT’s expected wealth is zero.

39
 Figure 4: Average price for LT1 in period 1

2.5
With HFT
No HFT
Average price for LT1 in period 1, i = a = 1 and µ = 2

1.5

1
1 2 3 4 5 6 7 8 9 10
Number of PTs

2
With HFT
No HFT

1.9
Average price for PT in period 1, i = a = 1 and µ = 2

1.8

1.7

1.6

1.5

1.4

1.3
1 2 3 4 5 6 7 8 9 10
Number of PTs

40
 Figure 5: Average price for LT1 (sells) and LT2 (buys) in period 2

2.5
LT1 (sells) with HFT
No HFT
Average price for LT1 and LT2 in period 2, i = a = 1 and µ = 2

2.4 LT2 (buys) with HFT

2.3

2.2

2.1

1.9

1.8

1.7

1.6

1.5
1 2 3 4 5 6 7 8 9 10
Number of PTs

2.1
With HFT
No HFT
Average price for PT in period 2, i = a = 1 and µ = 2

2.05

1.95

1.9

1.85

1.8
1 2 3 4 5 6 7 8 9 10
Number of PTs

41
 We now consider the effect on total revenues and costs. Clearly, if PTs face no average price
impact, PTs’ total revenue from intermediation is unaffected by the presence of HFTs. On the
other hand, LT1’s total revenues decrease in the presence of the HFT due to an increase in the
liquidity discount and the loss of surplus. Similarly, LT2 faces a higher cost of acquiring his
desired position as a result of the loss of surplus. But, if we compare which of the two liquidity
traders is worse off we find that the effect of HFTs on LT1’s revenue from selling i is the same
as that on LT2’s costs from buying i.

From Equations (18) and (19) we can obtain LT1’s revenues with and without HFT:

i2 aσ 2 1 2

∆=0 M 2 2
RLT 1 = µi − 4M + 24M + 3 and RLT 1 = µi − 2 i aσ .
(M + 1)2 16 (M + 1)

Then, the presence of the HFT reduces his revenue from selling i shares by the amount

∆=0 i2 aσ 2 1 2


RLT 1 − RLT 1 = 2 16 4M + 8M + 3 , (20)
(M + 1)

which increases, at an increasing rate, with the size of i and also increases in M .

Similar calculations give us LT2’s costs with and without HFT:

2M + 1 2 2 i2 aσ 2 1 2

∆=0
CLT 2 = µi + i aσ − 4M + 4M + 1 and CLT 2 = µi .
4(M + 1) (M + 1)2 16

Interestingly, not only do LT2’s costs of buying i shares increases in M , and in the size of
the liquidity need, i (at an increasing rate), but the increase in cost from the HFT for LT2 is
exactly the same as the reduction in revenue for LT1:

∆=0 i2 aσ 2 1
4M 2 + 8M + 3 = RLT
∆=0


CLT 2 − CLT 2 = 2 1 − RLT 1 .
(M + 1) 16

Finally, with a large number of PTs (as M tends to infinity), and using the fact that the
liquidity discount (Equation (32)) becomes zero, we can characterize what happens to costs
and revenues: CLT 2 → µi + 14 i2 aσ 2 and RLT 1 → µi − 14 i2 aσ 2 as M → ∞, as well as what
happens to HFT profits, which go to 21 i2 aσ 2 .

42
4.2 Price volatility and microstructure noise

We recall that in the GM model the volatility of the fundamental value of the asset is σ
and pricing pressures, due to a temporary liquidity imbalance, push market prices away from
their fundamental value but do not affect the ‘long-term’ volatility of prices. Put differently,
temporary pricing pressures increase the volatility of microstructure noise but do not affect
the (true) volatility of the value of the asset. Therefore, in a market with no HFTs, as that
described in the GM model, microstructure noise is caused by temporary selling and buying
pressures and this effect is exacerbated if HFTs that intermediate trades between LTs and PTs
are introduced. In our benchmark model we find that the HFT increases microstructure noise
in two ways. First, the price impact (i.e. selling pressure) of LT1 is larger than in the absence
of the HFT, and second, the instantaneous intermediation of the HFT around periods 1 and 2,
sees transactions fluctuate about the market prices P1 and P2 (see Table 1). In this section we
provide numerical examples where we show how the (short-term) volatility of prices increases
due to the microstructure noise induced by the HFT.

When there is no HFT intermediating transactions between LTs and PTs there is one trans-
action price at date 1, P1∆=0 , and another transaction price at date 2, µ2 . However, when HFTs
intermediate all transactions the tape will record four prices in date 1, {P1 − ∆L , P1 , P1 + ∆S , P1 },
and four in date 2, {µ2 − ∆S , µ2 , µ2 + ∆L , µ2 }. Therefore, it is inevitable to observe an in-
crease in microstructure noise within dates 1 and 2 which is solely caused by the HFT’s presence.
Figure 6 shows this microstructure volatility for prices in dates 1 and 2 that result from the HFT
intermediating all trades.18 Furthermore, Figure 7 depicts the standard deviation of prices at
both dates. We see that price volatility is much higher when an HFT operates in the markets.

Note however that in our model this increased volatility does not generate additional risk
for traders. Traders, in their evaluation of price risk (in their objective functions) recognize that
the microstructure noise generated by the HFT translates into a deterministic effect on their
18
To calculate the mean and variance of prices we also use the price at time t = 0 and assume it is the
fundamental value µ, thus the mean and variance of the price in date 1 are
µ + 4P1 1n
2
2
2
2 o
P̄1 = and V [P1 ] = µ − P̄1 + P1 − ∆L − P̄1 + 2 P1 − P̄1 + P1 + ∆S − P̄1 .
5 5
Similarly, to calculate the mean and variance of prices with date 2 we assume that the first observation is µ.

43
 Figure 6: Microstructure volatility of prices in dates 1 and 2 induced by HF trading

0.38
Price volatility period 2
Price volatility period 1
0.36
Standard Dev of prices in period 2, i = a = 1 and µ = 2

0.34

0.32

0.3

0.28

0.26

0.24

0.22

0.2

0.18
1 2 3 4 5 6 7 8 9 10
Number of PTs

execution costs while leaving their final holdings unaffected. In the case of random execution, as
studied in Section 3, the additional uncertainty on future execution costs introduces an element
of risk which results in an increase in the liquidity discount.

5 Conclusions

We use the GM model as benchmark to introduce an HFT who due to her rapid execution
and information processing ability is able to intermediate trades between LTs and PTs. The
strategies used by HFTs to exploit the microstructure of price setting allows them to extract
trading surplus from market participants and hold essentially zero inventories over time.

Our model departs from this premise and shows that the presence of HFTs increases the
price impact of liquidity trades (in proportion to the size of the trade), and when HFTs suc-
cessfully intermediate all trades they have no effect on the number of PTs, although they

44
 Figure 7: Standard deviation of prices in dates 1 and 2 with and without HF trading

0.5
With HFT
No HFT
0.45
Standard Dev of prices in both periods, i = a = 1 and µ = 2

0.4

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
1 2 3 4 5 6 7 8 9 10
Number of PTs

increase microstructure volatility, and increases trading volume. Thus, the presence of HFTs
distorts market conditions, not through the amount of shares traded, but through prices. By
exacerbating price impact, the HFTs induce additional market impact costs on participants,
especially the LT. This cost is proportional to the size of the trade which implies that large LTs,
such as institutional investors trading to change the composition of a portfolio, are the most
affected by the presence of HFTs–an effect that is consistent with Zhang [2010]’s findings. It is
also consistent with a growing trend amongst such investors to move execution to off-exchange
HFT-protected venues such as crossing networks and dark pools.

We show that the introduction of HFTs increases the microstructure noise around the
fundamental value of the asset in two ways. First, the price impact of LTs is larger than
in the absence of the HFT, and second, the instantaneous intermediation of the HFTs sees
transactions fluctuate around market prices; this result is supported by the empirical work of
Brunetti et al. [2011].

45
 In the particular case we have analyzed, it is the LTs who bear all the costs from the
presence of HFTs. LTs lose on two accounts: they lose trading surplus to the HFT, and they
must offer a higher liquidity discount to PTs to get them to buy their shares. Moreover, in the
extreme case where HTFs successfully intermediate all trades PTs find themselves unaffected
because the revenue they lose to the HFTs is compensated by a higher liquidity discount from
LTs. This implies that overall, the effect of HFTs is to reduce the value of the stock market
as a forum for providing a way for investors to convert their equity into cash (and vice-versa)
quickly and at a reasonable price. This is due primarily to the adverse effect of HFTs on prices
(costs of execution). This value reduction would, in a more general framework, be passed on
to the firms raising capital in equity markets. As equity buyers recognize the increased trading
execution costs from HFTs, they will require greater discounts from IPO issuers, resulting in
greater IPO underpricing, specially for shares sold to large institutional investors.

By taking an extreme position on the role of HFTs, our model has allowed us to identify ways
in which an additional intermediation layer negatively affects trading. Two factors reduce this
negative effect: the first is competition. As discussed above, competition between HFTs helps
reduce unnecessary intermediation directly, though it may also lead to increased skill differences
between HFTs and PTs, and a corresponding increase in unnecessary intermediation. The
second, is the positive effects from HFT intermediation which we have not explicitly included
in the model.

Public discussions of HFT identifies several of these benefits. One is that HFTs have reduced
the time to execution faced by LTs. Our model does not explicitly include this as execution
around liquidity dates is collapsed to a single date. Nevertheless, if the HFT is to intermediate
between LTs and PTs, it must act before PTs and hence execution time for LTs must be
lower. SEC [2010] reports a reduction in average execution time from 10.1 seconds in January
2005 to 0.7 seconds in October 2009. Whether the increased execution speed compensates
for the additional trading costs is something that requires a more detailed analysis, but also
it is something that traders are facing little choice on. A second benefit is that increased
trading speed (exploited by cross-market and index arbitrage HFT strategies) has increased
the speed of adjustment of prices across markets and between similar assets, such as index

46
derivatives and ETFs. The advantage of such rapid adjustments and the implied increase in
price informativeness and efficiency has to be weighed against the increased microstructure
volatility. In order to provide a more complete evaluation of HFTs, one would include these
(unmodeled) benefits, but also other (unmodeled) costs, such as the possibility that increased
trading speeds increase the cost of regulation and the well-functioning of market institutions,
and the many challenges it imposes on market integrity.

An additional issue that arises from our analysis is the question of how to measure (socially
valuable) liquidity when the analyst, as an outside observer, has no access to the identity
of traders and hence cannot determine directly the market impact costs of trades. A clear
conclusion from our model is that the total number of trades seems like a poor measure of
the ability of the market to provide prompt and fair value to investors, in a market with rent-
seeking hyperfast algorithms whose asset positions are essentially zero most of the time. Also,
the presence of HFTs who generate additional microstructure noise at smaller intervals and
accelerate market transactions raises several questions regarding how to measure price impact,
whether we should adjust current measures to account for the increase in the speed of execution,
and whether these measures adequately capture an investor’s cost of executing a trade.

Finally, in our knife-edge benchmark case HFTs had no effect on the number of liquidity
providing PTs. Alternatively, in the Appendix we consider an another strategy where HFTs
always charge the same haircut per trade (regardless of the size of the trade, type of trader,
and date). The effect in that case is that for PTs, the increase in the liquidity discount is
greater that the loss of surplus to HFTs, and in equilibrium, the number of PTs increases. This
suggests that the overall effect of HFTs on the “true liquidity providers” in not at all clear,
and we need good empirical work to (a) determine the speed-cost effect on outside investors
and LTs, (b) determine if HFTs are raising the cost of business for PTs, and (c) if they do,
whether the liquidity they provide is a good substitute for the one that is being driven out.
HFTs clearly generate costs, but they also generate benefits, and the net effect requires more
precise empirical analysis.

47
References

Jonathan Brogaard. High frequency trading and its impact on market quality. SSRN Working
Paper, 2010.

Celso Brunetti, Andrei Kirilenko, and Shawn Mankad. Identifying high-frequency traders in
electronic markets: Properties and forecasting. 2011.

Alain Chaboud, Erik Hjalmarsson, Clara Vega, and Ben Chiquoine. Rise of the Machines:
Algorithmic Trading in the Foreign Exchange Market. SSRN eLibrary, 2009.

U.S. Commodity Futures Trading Commission, the U.S. Securities, and Exchange Commission.
Findings regarding the market events of may 6, 2010. Report, SEC, September 2010.

David Easley, Marcos Mailoc Lopez de Prado, and Maureen O’Hara. The microstructure of
the ‘Flash Crash’: Flow toxicity, liquidity crashes and the probability of informed trading.
The Journal of Portfolio Management, 37(2):118–128, November 2011.

Thierry Foucault, Sophie Moinas, and Bruno Biais. Equilibrium High Frequency Trading.
International Conference of the French Finance Association (AFFI), May 2011, 2011.

Sanford J. Grossman and Merton H. Miller. Liquidity and market structure. The Journal of
Finance, 43(3):617–37, July 1988.

Terrence Hendershott, Charles M. Jones, and Albert J. Menkveld. Does algorithmic trading
improve liquidity? The Journal of Finance, 66(1):1–33, 2011.

Terrence J. Hendershott and Ryan Riordan. Algorithmic trading and information. SSRN
eLibrary, 2009.

Terrence J. Hendershott and Ryan Riordan. High Frequency Trading and Price Discovery.
SSRN eLibrary, 2011.

Michael Kearns, Alex Kulesza, and Yuriy Nevmyvaka. Empirical limitations on high frequency
trading profitability. SSRN worjing papers, 2010.

48
Andrei A. Kirilenko, Albert (Pete) S. Kyle, Mehrdad Samadi, and Tugkan Tuzun. The Flash
Crash: The Impact of High Frequency Trading on an Electronic Market. SSRN eLibrary,
2010.

Albert J. Menkveld. High Frequency Trading and The New-Market Makers. SSRN eLibrary,
2011.

Scott O’Malia. Letter to the members of the technology advisory committee. CFTC, 2011.

SEC. Concept release on equity market structure. Concept Release No. 34-61358; File No.
S7-02-10, SEC, January 2010. 17 CFR PART 242.

Frank Zhang. The Effect of High-Frequency Trading on Stock Volatility and Price Discovery.
SSRN eLibrary, 2010.

49
A Appendix

Above we argued that the HFT can discriminate across order size when trades come to the
market which enables her to apply haircuts for large and small trades. If the HFT is not
able to discriminate by size or type of trader she can still apply one haircut to all trades she
intermediates. Hence we can repeat the analysis for the benchmark case above while setting
the following haircuts:
∆ = ∆T1 1 = ∆T2 1 = ∆T1 = ∆T2 = ∆T2 2 .

Then, the optimal ∆ set by the HFT is:

(2M + 1)
∆ = iaσ 2 . (21)
(M + 1) (3 + 2M )

As in the case with two haircuts discussed in the body of the paper, LT1’s optimal holding is
θ1T 1 = i/(M + 1), but as we show below, the equilibrium number of PTs in this case is higher
than without HFT when q = 1.

Other results are similar to those discussed above: First, the microstructure volatility of
prices increases. Second, the price impact of the liquidity trades in both periods is substantial:
equilibrium sell (buy) prices are lower (higher) than the competitive price in the absence of
the HFT . Third, the expected returns that the PT face from buying in period 1 and selling in
period 2 increase (at the expense of traders), and the equilibrium number of PTs present in the
market increases when compared to the number of PTs that are present in a market without a
HFT. Fourth, the total volume of trades doubles relative to the number of trades observed in
the absence of the HFT.

In the interest of space we only discuss the equilibrium number of PTs. We proceed as
above where the entry condition (10) becomes

12M 2 + 12M + 7
 2

ac − 21 a2 σ 2 ( Mi+1 ) β(M )
e E e = 1 where β(M ) = . (22)
(3 + 2M )2

By inspecting (22) we know that the number of PTs with and without an HFT will be the same
only when β(M ) = 1 and this occurs for M = 1/2. When M > 1/2 we have more PTs in the

50
presence of HFT. If we denote the number of PTs by M and the number of PTs in the absence
of HFT by M∆=0 we can show that

p
M= β (M ) (M∆=0 + 1) − 1 ,

hence we have that M > M∆=0 . (The function β(M ) is increasing in M and we are interested
in values M > 1).

A.1 General Haircuts

Here we assume that the HFT applies different haircuts to every trader at every date.

       
 LT 2
Π2 (i) = θ̃2 P2 + ∆LT 2
+ i ∆LT 2
+ θ̃2LT 1 P2 − ∆LT 1
− θ1LT 1  ∆LT 1
  
2 2 2 2
   
+M θ̃2T P2 − ∆T − θ1T  ∆T2. (23)
 
2

     
W3LT 1 = W0LT 1 + θ2LT 1 P3 + θ1LT 1 − θ2LT 1 P2 − θ1LT 1 − θ̃2LT 1 ∆LT
2
1
− θ1LT 1 P1 + i − θ̃1LT 1 ∆LT
1
1
,
   
W3T = W0T + θ2T P3 + θ1T − θ2T P2 − θ1T − θ̃2T ∆T T T T
2 − θ1 P1 − θ̃1 ∆1 . (24)

1  
θ1LT 1 = 2
µ − P1 − ∆LT
2
1
, (25)
aσ
1  
θ1T = µ − P1 − ∆T
2 . (26)
aσ 2

∆LT
2
1 + M ∆T + iaσ 2
2
P1 = µ − , (27)
M +1

i 1 M  
θ1LT 1 = + ∆T LT 1
2 − ∆2 , (28)
M +1 aσ 2 M + 1
i 1 1  
θ1T = − ∆T LT 1
2 − ∆2 , (29)
M +1 aσ 2 M + 1

51
 i 1 M   ∆LT 1
θ̃1LT 1 = + ∆T LT 1
2 − ∆2 + 12 ,
M +1 aσ 2 M + 1 aσ
i 1 1   ∆T
θ̃1T = − T LT 1
∆2 − ∆2 − 12 .
M +1 aσ 2 M + 1 aσ

     
Π1 (i) = i − θ̃1LT 1 P1 − ∆LT 1  LT 1
+ M θ̃1T P1 + ∆T
 T
 ∆1 1  ∆1 . (30)
 
1

For a given order imbalance of magnitude i > 0:

1. Market clearing prices are

iaσ 2 ∆LT 1 + M ∆T
P1 = E [P2 |F1 ] − − 2 2
, (31)
M +1 M +1
P2 = E [P3 |F2 ] = µ2 ,

and the magnitude of the liquidity discount at t = 1 is

iaσ 2 ∆LT 1 + M ∆T
E [P2 |F1 ] − P1 = + 2 2
. (32)
M +1 M +1

(a) Asset trading is described on Table 2:

52
 Table 2: Prices and Volume of Trades

Price LT1 LT2 PT (total)

∆T LT 1
∆LT 1
 
2 −∆2
P1 − ∆LT
1
1
− MM+1 i − aσ 2
+ 1
aσ 2

∆LT 1
∆T
P1 − 1
aσ 2
M aσ12

∆T LT 1
∆T
 
2 −∆2
P1 + ∆T1 M
M +1
i− aσ 2
− M aσ12

∆LT 2
P2 + ∆LT
2
2
i− 2
aσ 2

∆LT 1
∆LT 2
∆T
P2 − 2
aσ 2
2
aσ 2
−M aσ22

∆T −∆LT 1 ∆LT 1
 
P2 − ∆LT
2
1
− M1+1 i + M 2 aσ22 2
+ aσ 2

∆T LT 1
∆T
 
2 −∆2
P2 − ∆T2 − MM+1 i − aσ 2
+M aσ22

(b) The HFT acts as counterparty to all these trades and makes profits equal to

Π (i) = Π1 (i) + Π2 (i) , (33)

where Π1 (i) and Π2 (i) are described in Equations (30) and (23) respectively.

A.2 Competition

1 1 1 1


Result: The distribution of p (η) has cdf Gp (z) = α 1+α− z and pdf gp (z) = α z2 .
1
Furthermore, E [p (η)] = α log (1 + α).

Proof: As 1/η ∼ U [0, 1] then 1/p (η) ∼ U [1, 1 + α], and the support of the distribution of
h i
1
p (η) is 1+α , 1 . To compute the cdf of p (η) we use the result that for a random variable X

53
with cdf FX (z) the cdf of the random variable Y = g(X), where g is a deterministic decreasing
function, is given by
Y ∼ G (z) where G (z) = 1 − FX g −1 (z) .

Thus, letting g(x) = 1/x and using FX (k) = (k − 1) /α, the cdf and pdf of p(η), denoted by
Gp and gp respectively, are given by
 
1
Gp (z) = 1 − FX
z
z −1 − 1
 
1 1
= 1− = 1+α− , (34)
α α z

and by differentiating (34) with respect to z we obtain

1 1
gp (z) = .
α z2

Finally, it is straightforward to calculate

1
E [p (η)] = log (1 + α) .
α

54