middlemen in limit-order markets1 - semantic scholar · 2015-07-28 · control sample as they had...
TRANSCRIPT
Electronic copy available at: http://ssrn.com/abstract=1624329
(preliminary and incomplete)
Middlemen in Limit-Order Markets1
Boyan Jovanovic and Albert J. Menkveld
first version: June 30, 2010
this version: May 19, 2011
1Boyan Jovanovic, Department of Economics, New York University, 19 West 4th Street, New York, NY, USA,tel +1 212 998 8900, fax +1 212 995 4186, [email protected]. Albert J. Menkveld, VU University Amsterdam,FEWEB, De Boelelaan 1105, 1081 HV, Amsterdam, Netherlands, tel +31 20 598 6130, fax +31 20 598 6020,[email protected] and TI Duisenberg School of Finance. We thank Robert Engle, Thierry Foucault,Phil Haile, Joel Hasbrouck, Terrence Hendershott, Johannes Horner, Charles Jones, Lasse Pedersen, Emiliano Pag-notta, and seminar or conference participants at the Autoriteit Financiele markten (AFM), Autorite Marches Financiers,Deutsche Borse, City U London, Dutch Central Bank, European Securities and Markets Authority (ESMA), KC FED,Universita Cattolica del Sacro Cuore, U Paris Dauphine, Queen Mary U, SED Meeting Montreal, St. Louis FED,and Yale for comments. We are grateful to both Euronext and Chi-X for acting as data sponsor and to Sean Flynnand Bernard Hosman for invaluable research assistance. Jovanovic thanks the Kauffman Foundation for support andMenkveld gratefully acknowledges Robert Engle for sponsoring him NYU visit in 08/11, VU University Amsterdamfor a VU talent grant, and NWO for a VIDI grant.
Electronic copy available at: http://ssrn.com/abstract=1624329
Abstract
Middlemen in Limit-Order Markets
A limit-order market enables an early seller to trade with a late buyer by leaving a price quote. But, public
news in the interarrival period creates adverse selection for the seller and therefore hampers trade. High-
frequency traders/machines (HFT) might restore trade by bringing the capacity to quickly update quotes
on news. A theoretical model shows that HFT entry can indeed increase welfare, but it might just as well
destroy it. Empirically HFT entry coincided with a 13% reduction in the effective bid-ask spread and an
unchanged trade frequency.
(for slides click http://goo.gl/RjzDt)
(for the internet appendix click http://goo.gl/xvQuN)
[insert Figure 1 here]
Successful entry of new limit-order markets has changed the industry of securities trading. Figure 1 illus-
trates how the NYSE had an 80% market share in its listed stocks in 2003, yet only 25% in 2011. Technology
facilitated the entry of cheap and fast electronic limit-order markets. BATS is an example of a successful
entrant market. It started in June 2005, picked up significant market share in 2007, and gradually grew to
a 10% share in 2011. Europe showed a parallel trend. Chi-X started in April 2007 and captured a 15%
market share in European equity trading. The London Stock Exchange, an incumbent, saw its market share
plummet from 30% in 2008 to 20% in 2011.
Coincidentally, high-frequency traders (HFTs) have become the most active participants in equity markets.
The Securities and Exchange Commission (SEC) published a policy document that refers to HFT as “one of
the most significant market structure developments in recent years. . . it typically is used to refer to profes-
sional traders acting in a proprietary capacity. . . characteristics often attributed to proprietary firms engaged
in HFT are. . . the use of extraordinarily high-speed and sophisticated computer programs for generating,
routing, and executing orders. . . very short time- frames for establishing and liquidating positions. . . ending
the trading day in as close to a flat position as possible. . . estimates of HFT volume in the equity markets
vary widely, though they typically are 50% of total volume or higher (SEC (2010, p.45)).”1 High-frequency
traders are therefore naturally thought of as middlemen in particular due to the proprietary nature of their
trading and the short holding period (milliseconds, seconds, minutes, or at most hours).
This paper studies how high-frequency trading might affect investor welfare in limit-order markets both
theoretically and empirically. A limit order is a pre-commitment to buy or sell a given amount of shares at
a pre-specified price. Limit orders arrive at the market and, if possible, are either matched with an existing
limit order or added to the stack of limit orders known as the limit order book (see Parlour and Seppi (2008)
for a survey). An important reason for the success of this type of electronic markets is that they provide a
cost-efficient way to match investors who are spaced out in time. In classic human-intermediated markets
such matching requires middlemen who need compensation (see, e.g., Grossman and Miller (1988)). In
limit-order markets matching seems possible without middlemen.
1The Committee of European Securities Regulators (CESR) published a similar discussion document where HFT is a focusarea; it solicits research as the HFT chapter’s title is: “The impact of High Frequency Trading (HFT) on market structure is still notfully understood (CESR (2010, p.31)).
1
Trade in limit-order markets without middlemen is hampered by a winner’s curse if the late investor is likely
to be better-informed (inherent to being the late investor) on the (common) value of the asset than the early
investor. For example, there might have been a press release on the stock in the interarrival period. The
early investor’s limit order is adversely selected; it is more likely to execute if the information was good,
less likely if it was bad. A rational early investor is fully aware of such risk and prices in an ‘adverse
selection cost’ (see Glosten and Milgrom (1985)) which reduces the likelihood of trade and destroys social
value as some natural re-allocations from low to high private value investors do not happen.
A competitive sector of middlemen — high-frequency traders — might reduce the informational friction,
and thus improve welfare, as the nature of their operation is information technology driven. To the extent that
the information between two investor arrivals is hard, machine-processable information the early investor,
say a seller, might instead of submitting a limit sell order — thus effectively post a price — pass off him asset
to a middleman who can post a price and quickly update it on incoming hard information thus reducing the
adverse selection risk. Examples of hard information are price changes in the market index, same-industry
stocks, oil etc.2 The extent to which this alternative route to the late investor improves welfare is the central
objective of our study.
Middlemen might increase the informational friction if their hard information was unknown to the late
investor and there was therefore no winner’s curse to begin with (at least with respect to this type of infor-
mation). In this case middlemen destroy welfare. The model section of the paper provides a single unified
analysis of this ugly face of the middleman’s along with its friendly face discussed in the previous paragraph.
The model fills the empty intersection of the literature on middlemen in human-intermediated markets and
the literature on limit-order order markets without middlemen. The middlemen literature assumes an in-
formation asymmetry where, rather than better informed, middlemen are uninformed which was natural for
markets where humans intermediate (see, e.g., Glosten and Milgrom (1985) and Kyle (1985)). Foucault,
Roell, and Sandas (2003) introduce costly monitoring ability for liquidity suppliers and analyze optimal
intensity. The most related limit-order models analyze how adverse selection risk affects the choice of an
arriving agent to choose between a limit and a market order (see, e.g., Hollifield, Miller, Sandås, and Slive
(2006) and Goettler, Parlour, and Rajan (2009)). They do not consider the entry of middlemen.
The empirical part analyzes the advent of middlemen by exploiting the introduction of an HFT-friendly
2See Petersen (2004) for the distinction between hard and soft information.
2
trading venue as an instrument. Chi-X started trading Dutch index stocks on April 16, 2007. Unlike the
incumbent market, Euronext, it did not charge limit order modifications, nor executions. Quite the contrary,
limit orders that led to execution received a rebate. The first 77 trading days of 2007 and 2008 are compared
to establish a ‘treatment’ effect. To control for time effects, Belgian index stocks serve as an ‘untreated’
control sample as they had not yet been introduced in Chi-X, yet were trading in Euronext under the same
rules as Dutch stocks.
Post Chi-X, a middleman has entered for Dutch stocks who trades primarily passive (i.e., through limit
orders). One new broker ID (not present in the pre-entry sample) trades very frequently — it is present
in roughly every third trade in Chi-X and in every twelfth trade in Euronext. It is particularly active on
days when most of a stock’s price movement is explained by the index — the R2 of a single-factor CAPM
explains 45% of time variation in middleman participation. What makes this broker an HFT, though, is
that its net position over the trading day is zero almost half of the sample days. A further finding is that
75-80% of its trades were passive. These observations provide empirical support for our theory which
characterizes middlemen as cost-efficient intermediaries. They minimize adverse selection cost by quickly
updating quotes on incoming hard information.
A diff-in-diff analysis (post- minus pre-entry, treated minus untreated) shows that middlemen entry is ac-
companied by an increase in liquidity supply, yet it did not change the number of trades. Post-entry realized
volatility is 64% higher for (treated) Dutch stocks, which is the same order of magnitude as the 69% higher
volatility for the (untreated) Belgian stocks. This is a reassuring result as it is unlikely that low-frequency
volatility is affected by HFT entry. The diff-in-diff analysis finds a significant 13% reduction in the effective
bid-ask spread, i.e., it increased 13% less for Dutch stocks as compared to Belgian stocks. The number of
trades appears to be unaffected by middlemen entry as it increased by a modest and insignificant 3% for
Dutch stocks relative to Belgian stocks.
The diff-in-diff results should be interpreted with care due to potential endogenous timing and selection
bias. Chi-X might have decided to start the system on April 16, 2007, and select Dutch stocks because
these stocks at that time showed particularly high activity and large spreads. If these were transitory effects,
the subsequent volume decline and spread reduction are to be expected irrespective of HFT entry. We
plan to study the size of transitory effects to verify whether (i) timing was indeed endogenous (spread and
volume were temporarily high on April 16, 2007) and (ii) whether the size of these effects could explain the
3
magnitude of the reported treatment effect.
The empirical findings add to a growing literature on high-frequency trading. Brogaard (2010) studies
trading of 26 NASDAQ-labeled HFT firms in the NASDAQ market in 2008-2010. In the aggregate, he
concludes that HFT tends to improve market quality. Kirilenko, Kyle, Samadi, and Tuzun (2010) study the
behavior of high-frequency traders in the E-mini S&P 500 stock index futures on May 6, the day of the flash
crash. Menkveld (2011) provides an in-depth empirical analysis of a high-frequency trader’s profitability
through a decomposition of its trading revenue and an analysis of the capital tied up in the operation through
margin calls.
The paper also fits into a broader literature on algorithmic or automated trading. Foucault and Menkveld
(2008) study smart routers that investors use to benefit from liquidity supply in multiple markets. Hen-
dershott, Jones, and Menkveld (2011) show that algorithmic trading (AT) causally improves liquidity and
makes quotes more informative. Chaboud, Chiquoine, Hjalmarsson, and Vega (2009) relate AT to volatil-
ity and find little relation. Hendershott and Riordan (2009) find that both AT demanding liquidity and AT
supplying liquidity makes prices more efficient. Hasbrouck and Saar (2010) study low-latency trading or
“market activity in the millisecond environment” in NASDAQ’s electronic limit order book ‘Single Book’
in 2007 and 2008 and find that increased low-latency trading is associated with improved market quality.
More broadly, our findings contribute to the literature on middlemen in search markets. In Rubinstein
and Wolinsky (1987) and Masters (2007) middlemen may improve allocations because they speed up the
meeting process. Rents are divided via Nash bargaining. Hosios condition and optimality obtain when
numbers are endogenous. In our model, however, market participation is exogenous. Our paper is closer to
Li (1996) in which where middlemen have an advantage in terms of information in a lemons model. The
contrast is (i) that the quality of the traded good is exogenous, and (ii) that the information is not about
independent values but, rather, common values, which we believe to be at the heart of evaluating welfare
effects of HFTs who specialized in speed to minimize being adversely selected on public information. A
unique feature of limit order markets is that late investors cherry pick take-it-or-leave-it offers (i.e., limit
orders) of early investors, leaving the latter with a winner’s curse problem. Closer to home are models on
search in decentralized markets, e.g., Duffie, Garleanu, and Pedersen (2005), Lagos, Rocheteau, and Weill
(2009), and Duffie, Malamud, and Manso (2009).
The remainder of the manuscript is structured as follows. Section 1 develops the model. Section 2 presents
4
empirical results. Section 3 concludes.
1 Model
The model features two investors who are spaced out in time: an early-arriving (potential) seller (S) who
owns the asset and a late-arriving (potential) buyer (B) who does not own the asset (cf. Grossman and Miller
(1988)). Eventually middlemen or machines (M) are introduced which, by their very nature, can stay in the
market indefinitely (i.e., at no cost). The agents (S, B, and M) have available to them a limit order market to
potentially effectuate a trade and therefore a transfer of the asset. All agents in the model are risk-neutral. An
overview of all notation and what the various variables are a modeling device for is included as Appendix I.
S values the asset at x+ z, and B values it at y+ z. The private values x and y are drawn independently from a
common distribution with CDF F. The common value (innovation) z is drawn from a distribution with CDF
G.
The common value z is itself composed of two independent components:
z = zh + zs
The superscripts denote machine-processable ‘hard’ information and its complement, ‘soft’ information (cf.
Petersen (2004)).
The game may also include a number of middlemen or machines (M) that all have a zero private value for
the asset. In other words, each M’s valuation of the asset is just z. M always know zh because they can track
this machine-processable information at high frequencies. They do not know zs which, by its very nature,
requires human skill to process. It is important to note that machines have zero opportunity cost and are
therefore assumed to always be in the market (unlike S or B for whom it is costly to stay in the market).
The two components are, by definition, orthogonal which implies that σ2z = σ
2h + σ
2s . Let
α Bσ2
h
σ2h + σ
2s
(1)
denote the relative importance of hard information.
5
[insert Figure 2 here]
The informational structure of the game is depicted in Panel A of Figure 2. Private values are observed
only by the agents themselves. The hard common value zh is always observed by the middlemen, but only
observed with probability λ by the buyer. The buyer is more likely to be informed on common values than
the seller by the mere virtue of arriving late. The soft common value zs is observed only by the buyer.
1.1 The game without middlemen
This section develops the structure of the game in which M are absent. The game consists of two periods:
1. S observes (only) x, posts an ask ps (x), and leaves.
2. B arrives, observes y and zs, and observes zh with probability λ. This probability of observing zh is
independent of zh and zs.
Note that a bid by B is excluded as by the time that B can post it, S is gone.
As S can only condition him ask price p on him private value x, this price is uncorrelated with z and both of
its components. B accepts iff y + z > p with probability λ
y + zs > p with probability 1 − λ(2)
In the event of a sale, S receives p. On the other hand, if z ≤ p − y, S receives x + z. Conditioning on y, S’s
expected return is:
λ
(p[1 −G (p − y)
]+
∫ p−y
−∞(x + z) g (z) dz
)+ (1 − λ)
(p[1 −Gs (p − y)
]+
∫ p−y
−∞(x + z) gs (z) dz
)(3)
where
g (z) =∫ ∞
−∞gh (
z − z′)
gs (z′) dz′ (4)
is the density of z, gh is the density of zh, and gs the density of zs.
Note that in the process B conditions on z and which creates a winner’s curse for S. That is, if him ask quote
is ‘successful’ and wins a trade he is cursed as, conditioning on the event of a trade the expectation of z is
6
positive. She is more likely to sell on good news (positive z) than on bad news (negative z). The seller is
adversely selected by the buyer.
Evidently, (3) makes S optimally choose p as follows:
v(x, gλ
)≡ max
p
∫ ∞
−∞
[p[1 −Gλ (p − y)
]+
∫ p−y
−∞(x + z) gλ (z) dz
]dF (y) (5)
where, for all z, Gλ B λG + (1 − λ) Gs and gλ B λg + (1 − λ) gs.
When is the winner’s curse absent? The winner’s curse is absent in two situations which differ in
what they imply for the role of an informed M. First, there may be no common value of either type, i.e.,
σ2z = σ
2h + σ
2s = 0. In this straightforward case M neither help nor hurt – they are irrelevant and would not
be able to make money by trading, as will be shown in the next subsection.
More significant, however, is the case
σ2s = λ = 0, coupled with σ2
h > 0 (6)
In this case, only zh matters and neither S nor B is aware of it. There is therefore no adverse selection
problem in the market to begin with; the decision rule (2) reduces to ‘ buy iff y > p’, and (5) reduces to
maxp
{p[1 − F (p)
]+ xF (p)
}(7)
If M is not in the market, the ask price that maximized this expression would be an equilibrium.
How M might introduce a winner’s curse. Potential entry by a single M dramatically changes the
equilibrium, provided that S and B are aware of M’s presence in the wings. When σ2h > 0, any equilibrium
– and particularly the equilibrium implicit in (7) – is vulnerable to entry of M who, knowing zh, would wish
to buy the security whenever zh > p, leaving S or B with the security in states where zh < p and creating,
in other words, a winner’s curse problem for both S and B. An opportunistic M can thus reduce welfare by
placing market orders in states that favor him.
7
1.2 The game with middlemen
High-frequency traders are most realistically modeled as a sector of competitive middlemen. This setting is
therefore very different from what was described in the previous paragraph. Competition forces M to place
limit orders to buy (i.e., bid prices) of their own. These bids are not vulnerable to a winner’s curse caused
by zh because M all know zh and can incorporate its realization into their bids. Indeed, it will turn out that
competitive M will help most when σ2h is large relative to σ2
s , which is to say when α in equation (1) is large.
The game is set up to reflect the fact that S is (i) slow and (ii) cannot stay in the market (at zero cost).
Therefore S can only take one action while he is in the market, i.e., he can either post an ask or accept a bid.
Panel B of Figure 2 graphically illustrates the following sequential structure of the game:
Period 1. S arrives and can take only one action: either post an ask or accept a bid.
1. S observes x and (maybe) posts an ask pS . Simultaneously, M post competitive bids.
2. If S did not post an ask he can accept a M bid. If he rejects it, the game ends.
3. If S did post an ask (and therefore foregoes being able accept an M’s bid), the M accept or reject it.
4. S leaves
Period 2. B arrives.
1. B observes zh that is revealed through M’s bids. B also observes y and zs.
2. If M bought the asset from S, he now tries to sell it to B.
3. If the M rejected the ask that S posted, the ask is now available for B, and B takes it iff y + z > pS .
Under these assumptions, we shall now show that M compete each other out of all rents that might be
derived from knowing zh and from being fast. Competitive bids by M in period 1.1 allow both S and B to
infer zh and, as a result, unless S has posted an ask, the adverse selection on zh disappears.
One consequence of this is that, regardless of λ, B is able to get information on zh and therefore gets to know
both components of z. The parameter λ thus affects only the game in which M are absent.
8
1.2.1 The second period sub-game between M and B
As B infers zh from competitive bidding by the middlemen in period 1, this hard information is common
knowledge between them in the period 2 subgame when M owns the asset and posts an ask price for B. B
takes this ask price pM(zh), shortened to pM , and trade occurs iff
y + zs + zh > pM (8)
Now let
p B pM − zh (9)
Then trade takes place if
zs > p − y (10)
So M’s value of holding the security is
zh + vM
where
vM ≡ maxp
∫ ∞
−∞
[p[1 −Gs (p − y)
]+
∫ p−y
−∞zgs (z) dz
]dF (y) (11)
That is, M faces a winner’s curse itself induced by B’s privately knowledge of zs. As the size of soft
information (σ2s) grows, M’s value of holding the asset (vM) tends to zero.
Lemma 1
vM = v(0, gs)
Proof. The RHS of (5) and the RHS of (11) differ only with respect to the distribution gλ (z) on the one hand
and gs (z) on the other. It is, in other words, the same function of g.
1.2.2 The first period choice for S to post an ask price or not
S knows neither zs nor zh. Posting an ask price in period 1 therefore exposes him to a winner’s curse on
common value components (as will be shown below in equation (16)). Therefore he may prefer not to post
9
an ask as it avoids the winner’s curse. She cannot avoid it altogether as even instead of posting accepting
M’s bid exposes him indirectly to a winner’s curse over zs because M are exposed to it and their bid will
reflect that exposure. M’s exposure reduces vM in the formulas below, when the size of soft information
(σ2s) increases. Let us compare him expected payoff under the two strategies.
If S does not post an ask. In this case S essentially avoids a zh winner’s curse on him ask price (as there
effectively always is a zs curse that he is exposed to). Instead, he has the option to accept a middleman’s bid
which we expect to be a particularly attractive option if the size of hard information (σ2h) is relatively large.
If S does not post an ask, he rationally anticipates the competitive M’s bid to be:
pM = zh + vM
which reveals zh to him.3
Now S has the option to accept or reject. Now aware of zh he accepts the bid iff
x + zh + E(zs) < zh + vM ⇔ x < vM (12)
as E (zs) = 0. M therefore obtains the asset iff (12) holds and thus with probability:
F(vM
)(13)
When deciding whether or not to post an ask S does not know zh. If he decides not to post an ask, S’s
expected utility conditional on x is:
U (x) ≡ max(x, vM
)(14)
Differentiating U (x) for x , vM yields:
U′ (x) = I{x>vM} = Pr (no trade | x) (15)
3Note that S observing M’s bid, turning around, and posting an ask price himself is explicitly ruled out. The reason is S is tooslow for two actions. This is modeling device to capture a machine’s ability to process vast amounts of information quickly andupdate ask quotes to avoid the winner’s curse. If S willingly passes off the asset to the middleman he essentially buys its ability toavoid the curse on hard information.
10
where I{.} is the indicator function.
[insert Figure 3 here]
Figure 3 illustrates U (x), the value to S of not posting.
If S does post an ask. In this case S fully exposes himself to the winner’s curse. We expect this option to be
particularly attractive when the size of hard information (σ2h) is small. The value of this option is developed
in this subsection and graphed as V (x) in Figure 3.
Let us first discuss the special case of no hard information (σ2h = 0). S is now able to extract the same rents
from B as M could in the second-period subgame; he collects vM in expectation. The conjectured value
graph of posting an optimal ask is depicted in Figure 3 as V(x | σ2h = 0). It crosses the vertical axis exactly
at vM. We therefore conjecture that the strategy post is better for S than not-post for private values x ≥ 0.
As for the general case with hard information, if S posts an ask pS (x) it will be accepted if either a M or the
B values the security more, i.e., if:
pS (x) < max(zh + vM, y + zs + zh
)= zh +max
(vM, y + zs
)≡ zh + ϕ
(y + zs) (16)
M can act before B arrives and they therefore have the first take on the offer. When pS < zh + vM, any M
that is lucky enough to snap up the offer makes positive expected rents. Thus if pS is in the money for M,
one of them will snap it up. If, however, pS > zh + vM then the ask pS remains in the order book and is
available for B when he shows up. B will accept the offer if pS < y + zs + zh.
Leading up to S’s optimization problem, condition (16) is more easily expressed as:
zh > p −max(vM, y + zs
)
11
Anticipating these optimal responses by M and B, S chooses pS to solve:
V (x) ≡ maxp
∫ p[1 −Gh
(p −max
(vM, y + zs
))]+∫ p−max(vM ,y+zs)
−∞
(x + zs + zh
)dGh
(zh
) dGs (zs) dF (y) (17)
Differentiating (17), using the envelope theorem, and simplifying yields:
V ′ (x) = Pr(zh + ϕ
(y + zs) ≤ pS (x)
)> 0 (18)
where pS (x) is the optimal policy for S and V is convex.
This expression is completely analogous to (15); the derivative V ′ (x) is equal to the probability of no trade
conditional on x. More explicitly:
V ′ (x) =∫
Gh (ps (x) − ϕ (y + zs)) dGs (zs) dF (y) (19)
Moreover, limx→∞ V ′ (x) = 1 and V ′ (x) < 1 as illustrated in Figure 3.
If pS (x) is differentiable, we may use the chain rule to conclude that
V ′′ (x) =∂pS (x)∂x
·∫
gh (ps (x) − ϕ (y + zs)) dGs (zs) dF (y) > 0
because (still to be shown), pS is strictly increasing in x.
S’s decision whether or not to post. This choice problem is illustrated in Figure 3. The optimal decision
depends on x. S posts iff
V (x) ≥ U (x) (20)
Let
A = {x | (20) holds} (21)
be the set of private values x for which S posts an ask price.
As V (x) > x, (20) holds for x = vM . Therefore the intersection of V and U occurs at a value of x below vM.
Moreover, (19) shows that V ′ > 0 any intersection of U and V therefore must:
12
1. be unique and
2. occur at a value of x, say a, that is smaller than vM in which case it satisfies
vM = V (a) (22)
The set A therefore is:
A = {x | (20) holds} = [a,∞)
Finally, since V (x) > x, (22) implies that
x < A⇒ x < vM
which proves
Proposition 1 If S does not post then S accepts M’s bid with probability 1.
This proposition effectively removes the lowest branch of the tree depicted in Panel B of Figure 2.
1.3 The model’s implications for how middlemen presence affects the trading game
This section develops the implications of the presence of middlemen in the trading game between S and B.
It is focused on the number of trades, the distance between transacted bid and ask prices (referred to as the
‘effective bid-ask spread’), and most importantly welfare. The trade and price implications are developed
in order to calibrate the model to the data. Also, the conjectured relationship of more active middlemen on
relatively more hard information as opposed to soft information (i.e., a higher α) is developed as it will be
tested in the data (see, e.g., Figure 8).
1.3.1 The game without middlemen
Trades. The expected number of trades in the case without middlemen is:
TNM =
∫ ∞
−∞
∫ ∞
−∞
(1 −Gλ
[pS (x) − y
])dF (x) dF (y)
13
Effective bid-ask spread. In the game without M there are no bids, there is only an ask pS (x) which solves
(5). A bid can be obtained, however, if the arrivals of S and B are reversed where the late investor (now S)
retains the informational advantages. Let pB (y) denote B’s bid then the bid-ask spread is naturally defined
as:
Spread =∫ ∞
−∞pS (x) dF (x) −
∫ ∞
−∞pB (y) dF (y) (23)
In an internet appendix — the http address in on the abstract page — it is shown that with symmetry of the
densities f , gh, and gs one obtains the following result:
Proposition 2 For (x, y) satisfying F (x) = 1 − F (y):
pS (x) − x = y − pB (y) (24)
Spread = 2∫ ∞
−∞
[pS (x) − x
]dF (x) (25)
A standard measure in data analysis is the effective spread which effectively samples the bid-ask spread only
at the time of trade. The model equivalent for this ex-post measure interacts the bid or ask price quote with
a trade dummy and sums over (x, y, z):
Effective spread =1
TNM
∫ ∞
−∞
∫ ∞
−∞
[pS (x) − x
] (1 −Gλ
[pS (x) − y
])dF (x) dF (y) .
Welfare. Welfare in this case is the expectation of the sum of each investors private value interacted with a
dummy that is one if the investor is the final owner, zero otherwise:
WNM = E(xI{S owns} + yI{B owns}
)(26)
=
∫ ∞
−∞
(∫ ∞
−∞
[xGλ
[pS (x) − y
]+ y
(1 −Gλ
[pS (x) − y
])]dF (x) dF (y)
)
The common value z drops out as it is zero in expectation. It is not as innocuous as it seems as it drives who
ultimately ends up holding the asset.
14
1.3.2 The game with middlemen
Trades. If S posts an ask price (i.e., x ∈ A), M has the first take on it (as M is fast) and gets the asset iff the
price is below M’s reservation value, i.e., iff:
pS < zh + vM (27)
As pS (x) is strictly increasing in x and thus invertible, the probability that M gets the asset is:
F[(
pS)−1 (
zh + vM)]≡ qS M (28)
where(pS
)−1denotes the inverse function of pS (x). This expression mirrors (13).
If S posts an ask (x ∈ A), the number of trades is:
TA(x, y, zh, zs
)=
0 if pS (x) > max(zh + vM , y + zh + zs
)1 if
pS (x) > zh + vM and pS (x) < y + zh + zs or
pS (x) < zh + vM and pM > y + zh + zs
2 if pS (x) < zh + vM and pM < y + zh + zs
(29)
If S does not post an ask (x < A), the number of trades is:
T˜A(x, y, zh, zs
)=
0 if x > vM
1 if x < vM and pM > y + zh + zs
2 if x < vM and pM < y + zh + zs
(30)
The expected total number trades therefore is:
TM = E (TA + T˜A)
=
∫(TA + T˜A) dF (x) dF (y) dGh
(zh
)dGs (zs)
which is the (unconditional) expectation over the four mutually independent shocks. The only trades that M
do not participate in are those for zh + vM < pS (x) < y + zh + zs.
15
Effective bid-ask spread. The only bids in the game with M are those by M in the first period of the game.
They all bid zh + vM where vM is given by the LHS of (11). The average ‘effective’ bid is therefore:
pbid = vM (31)
as trade does not depend on zh.
The game features three potential ask prices:
1. pask = pM if x < A, M’s ask price solves (11),
2. pask = pS (x) if x ∈ A, S’s ask price solves (17),
3. pask = pM if x ∈ A and the first ask price is snapped up by M which then posts the ask pM for B.
The set A does not depend on zh; the selection therefore is only on x. The effective ask is now obtained by
averaging these ask prices multiplied by a trade event dummy over(x, zh
):
pask =1
TM{pF (a)
[1 − F (p)
]+
∫ [∫ ∞
apS (x) I{pS(x)<zh+max(vM ,y+zs)}dF (x)
]dF (y) dGh
(zh
)dGs (zs)
+
∫ [∫ ∞
a
(pM + zh
)I{pS (x)<zh+vM ∩ pM<y+zh+zs}dF (x)
]dF (y) dGh
(zh
)dGs (zs) } (32)
where p is defined in (9) for E(pM
)= p. The top line of (32) captures sales by S to either M or B; the
bottom line captures sales by M to B.
The effective bid-ask spread in the game with M therefore is:
Effective Spread = pask − vM
Welfare. The common value z drops out again from the welfare equation but its two components critically
determine who ends up owning the asset. Welfare itself is a weighted average of welfare in the event that S
posts (x ∈ A) and welfare in the event that S does not (x < A). The second case is easier, so we start with it.
16
The contribution to expected welfare of x < A is:
WNP ≡∫{∼A∩(zh+vM ,∞)}
xdF (x) +∫{∼A∩(−∞,zh+vM)}
dF (x)∫ ∞
−∞
(∫pM−zs
ydF (y))
dGs (zs) (33)
where pM is defined in (11). If M ends up with the asset welfare is zero.
In the case that S posts an ask (x ∈ A) welfare is:
WP ≡∫{A∩(pS (x)>max(zh+vM ,y+zs+zh))}
xdF (x) +∫{A∩(pS (x)<max(zh+vM ,y+zs+zh))}
dF (x) D
where D is the welfare obtained if S parts with the asset. Once again, if M ends up with the asset welfare is
zero. In fact, welfare y is collected in one of following two mutually exclusive events:
(pS (x) ≤ zh + vM and y + zs + zh ≥ pM
)≡ D1
(x, y, zh, zs
)with p defined in (9) or:
zh + vM < pS (x) ≤ y + zs + zh ≡ D2(x, y, zh, zs
)Therefore:
D =∫
D1∪D2
ydF (x) dF (y) dGh(zh
)dGs (zs)
The welfare for the case with middlemen is becomes:
WM = WP +WN P
1.4 Parameter choices and the welfare effect of middlemen
This section first calibrates parameters to fit salient features of the data and uses these to discuss the welfare
effect of middlemen.
1.4.1 Parameter choices
Private values are chosen to be normally distributed, i.e., x, y ∼ N(µ, σ2).
17
The private-value dispersion parameter σ. Private value variance is fixed at π to fix the first-best welfare
gains from trade to one. That is, relative to autarky, the expected gains from trade among S and B equal one
unit. Therefore all variables will be measured in ‘gain-from-trade’ units.
The variance parameter π is obtained as follows (cf. Clark (1961, equation (2))):
E max (x, y) − E (x) = µ +σ√π− µ (34)
The mean private value parameter µ. The welfare analysis compares the regime with M to the one
without M. The introduction of M might ‘artificially’ raise welfare as a zero-private-value agent is added
to the trade economy. That is, the social planner would assign the asset to M in case both S and B are on a
negative private value draw. To suppress this effect µ is chosen to make this a small probability event, i.e.,:
Pr {max (x, y) < 0} = F (0)2 =
[Φ
(− µσ
)]2= 0.01 (35)
where Φ is the standard normal CDF. As the private value variance was set to π this implies: µ = 2.27.
The relative size of hard information parameter α. The total common value innovation is measured in
the data as the long-term price change in between trades. A proxy for how much of this is hard information
is ‘hard’ is obtained by identifying the part that is correlated with a change in the index futures midquote in
the same time interval (although, admittedly, this is a lower bound as M might exploit other correlated price
series). Essentially, the relative size of hard information is the R2 of a regression of stock return on the index
return (for details we refer to the discussion of Table 4. Panel B of Table 4 implies: α = 4.02/9.47 = 0.42.
The relative size of common value versus private value θ (and therefore σ2z ). As private value variance
σ2 was fixed at π the common value variance is best described as its size relative to private value variance
which is captured by the parameter θ:
θ :=σ2
z
π
The parameter θ is estimated in two ways. One way is based on one of the strongest empirical results which
is that middlemen particicpation increases with the relative size of hard information (cf. Figure 8). The
18
slope of the fitted least-squares line is: 0.10. The model equivalent of middlemen participation τ is:
τ :=TM −
∫ [∫A I{pS (x)>zh+vM and pS (x)<y+zh+zs}dF (x)
]dF (y) dGh
(zh
)dGs (zs) ,
TM(36)
The parameter θ is estimated as the solution to:
θ1 = arg minθ
(τ (θ, α = 1) − τ (θ, α = 0) − 0.10)2
The model-implied relationship between M participation and θ is plotted as the top graph of Figure 4.
A second way to estimate θ is to fit the markup charged by S relative to adverse selection component in
the regime without M. In the data, the period without M identifies this markup as the ratio of the effective
spread and its adverse selection component which is, on average, 2.61/1.89=1.38 in the pre-event period (cf.
Table 6). The parameters (θ,λ) for which the model-implied relative markup equals the observed markup
are established as the solution to:
(θ2, λ) = arg minθ,λ
E(pS (x) − x | y + zλ > pS (x)
)E
(zλ | y + zλ > pS (x)
) − 1.38
2
(37)
subject to zλ having the CDF
Gλ = λN(0, σ2
z
)+ (1 − λ) N
(0, ασ2
z
)which is a mixture of two normal distributions where α = 0.42 (cf. earlier discussion).
The complete list of parameter choices is summarized below.
pre-set estimated
parameter µ σ2 α θ1 (θ2,λ)
value 2.27 3.14 0.42 0.27 (2.63,0)
1.4.2 How middlemen affect welfare
This section discusses how the entry of middlemen affects welfare in the trading game between S and B.
It first discusses how they might improve welfare by assuming maximum information asymmetry between
19
S and B to begin with, i.e., setting λ = 1. It then analyzes how they might destroy welfare by potentially
creating adverse selection rather than solving it, i.e., λ < 1.
How might M improve welfare? The effect of middlemen entry is analyzed by plotting welfare for the
regime with middlemen against welfare in the regime without middlemen. It is plotted as a function of
either the relative size of hard information in total common value information, α, and as a function of the
relative size of common value variation vs. private value variation, θ. Generally, we find that the entry of M
increases welfare in both these parameters (as expected), α and θ, but increasing θ raises the welfare increase
more than increasing α.
Let us now discuss the effect of each parameters in turn.
[insert Figure 4 here]
Raising α. The bottom panel of Figure 4 graphs the realized gains from trade (welfare) as a function of
α where θ is kept fixed at its estimate of 0.27(= θ1). The green dashed line shows the first-best gains from
trade which is slightly above one (which is to be expected given the introduction of M as a third agent,
see discussion of (34)). A comparison of the top and bottom panels of Figure 4 shows that M participation
is correlated with welfare, as is to be expected. The gains are positive but small – about two percent of
maximum welfare — and they rise only slightly with α.
[insert Figure 5 here]
Raising θ. Figure 5 shows that M participation (top panel) and the welfare gains due to M entry (bottom
panel) both increase with θ.
How might M reduce welfare? Figures 4 and 5 show a uniformly positive effect of M on welfare, which
begs the question of whether there are parameter values for which the entry of M destroys welfare. Prima
facie, the model might generate a welfare-reducing effect of M only when λ is close to zero, which is to say,
only when B is unaware of zh in which case there is no adverse selection unless M is in the game.
When θ rises, the markup (pS (x) − x) rises in order to undo the increase in adverse selection. We generated
a version of Figure 5 in which, when increasing θ, we simultaneously reduce λ so as to keep the absolute
20
informational advantage of B constant in the no-M regime. Figure 5 conditions on hypothetical values of
θ and evaluates welfare in the absence of M under the value of λ that produces the closest model fit to the
criterion in the RHS of (37). Let welfare without M be defined as w (θ, λ) , and let welfare with M be defined
as w (θ). Note that the parameter λ does not affect outcomes in the M regime because B can reverse-engineer
zh from the bids by M in the first period. Figure 6 plots w (θ, λ (θ)) , w (θ) where
λ (θ) := arg minλ
(RHS of (37))
s.t. z ∼ λN (0, θπ) + (1 − λ) N (0, αθπ)
This calculation answers the question: Given a hypothetical amount of total adverse selection (θπ), what
information structure (λ) among buyers matches best the markup charged relative to the adverse selection
that an uninformed seller faces in the game without M?
[insert Figure 6 here]
Figure 6 shows that θ has to be high before M actually harm welfare. For these θ’s we find that λ(θ) = 0
which means that if M is absent there is no adverse selection. As a result, welfare does indeed become
higher for the no M regime. However, the estimated θs or, rather, the range of θs that our estimates imply
as being reasonable all suggest a positive role of M. The two estimates of θ implied by the values of σ2z (1)
and σ2z (2) in the above Table are θ1 = 0.27 and θ2 = 2.63 respectively. For θs in that range, the effect of M
on welfare is positive as Figure 6 shows.
2 Empirical results
2.1 Background
The European Union aimed to create a level playing field in investment services when it introduced the
Markets in Financial Instruments Directive (MIFID) on November 1, 2007. For markets, MIFID created
competition between national exchanges and it allowed new markets to enter.
21
Instinet pre-empted MIFID when it launched the trading platform Chi-X (Chi-X) on April 16, 2007, for
Dutch and German index stocks.4 At the end of 2007, it allowed a consortium of the world largest brokers to
participate in equity through minority stakes.5 Before Chi-X Instinet had successfully introduced the product
as ‘Island’ in the U.S. which distinguished itself from others through fast-execution and subsidization of
passive orders (see fee discussion below). Eventually Instinet sold the U.S. license to NASDAQ but kept the
international license which led to Chi-X.
In the first 77 trading days of 2008, our sample period, Chi-X traded British, Dutch, French, German, and
Swiss local index stocks. It had captured 4.7% of all trades and was particularly strong in Dutch stocks with
a share of 13.6%. Volume-wise, Chi-X overall market share was 3.1% and its Dutch share was 8.4%. Chi-X
started off particularly strong for the stocks studied in this manuscript.
Prior to Chi-X entry, Euronext was by far the main venue for trade in Dutch stocks. Its trading platform ran
in much the same way as the Chi-X platform and competition focused on fees and speed (see discussion
below). Dutch stocks also traded as ADRs in the U.S. and in the Xetra system run by the German Stock
Exchange. They did not yet trade in NASDAQ OMX, Turquoise, or BATS-Europe which entered later on a
business model similar to Chi-X: subsidies on passive orders and a fast system.
The broker identified as a middleman in this study was a substantial participant in Chi-X. In our sample
period, it participated in 10.8 million of the 99.2 million Chi-X trades. It was particularly active in Dutch
stocks with participation in 1.7 million out of 8.6 million Chi-X trades.
Fee structure. In our sample period, Chi-X did not charge for limit order submissions and cancellations.
Quite the opposite, it paid 0.2 basis points if the order becomes the passive side of a trade. If however
the order is ‘marketable’ and executes against a standing limit order upon arrival it gets charged 0.3 basis
points. For example, in the stock of limit sell orders in the book, the one with the lowest price becomes
the prevailing ‘ask’ quote. If a limit buy order arrives with a price (weakly) higher than this ask price, it
immediately executes against this limit sell and a transaction is recorded. For simplicity, in this example
it is assumed that orders are of the same size. For a detailed description of the generic limit-order market
mechanism we refer to Biais, Hillion, and Spatt (1995).
4“Chi-X Successfully Begins Full Equity Trading, Clearing and Settlement,” Chi-X press release, April 16, 2007.5These brokers were: BNP Paribas, Citadel, Citi, Credit Suisse, Fortis, Getco Europe, Ltd, Goldman Sachs, Lehman Brothers,
Merrill Lynch, Morgan Stanley, Optiver, Societe Generale and UBS (op. cit. footnote 7).
22
Euronext on the other hand charges a fixed fee of e1.20 per trade which for an average size trade of
∼e25,000 (see Table 1) is effectively 0.48 basis points. Highly active brokers benefit from volume dis-
counts that can bring the fixed fee down to e0.99 per trade (∼0.40 basis points). In addition, Euronext
charges an ‘ad valorem’ fee of 0.05 basis points. The act of submitting an order or cancelling it is not
charged (i.e., only executions get charged without an aggressive/passive distinction). But, if on a daily basis
the cancellation-to-trade ratio exceeds 5, all orders above the threshold get charged a e0.10 fee (∼0.04 basis
points).
In terms of post-trade costs, Chi-X clears and settles through EMCF which claims to be over 50% cheaper
than other European clearing houses including the ones used by Euronext.6
System speed. In a April 7, 2008 press release Chi-X celebrates its first anniversary. It claims to run one
of the fastest platforms in the industry with a system response time (often referred to as ‘latency’) of two
milliseconds. This is “up to 10 times faster than the fastest European primary exchange.”7
Overall, Chi-X appears to be particularly friendly venue for the middleman type that is central to the theory.
Its fees are lowest across the board and particularly low for a strategy that relies on passive orders that get
cancelled and resubmitted upon the arrival of public news. Its speed advantage allows one to do so quick
enough in order not to be picked off.
2.2 Data, Approach, and Summary statistics
Data. The main sample consists of trade and quote data on Dutch index stocks for both Chi-X and Euronext
from January 1 through April 23, 2008. The quote data consist of best bid and ask price and the associated
depth. The trade data contain transaction price, size, and an anonymized broker ID for both sides of the
transaction. The broker ID anonymization is done for each market separately and broker IDs can therefore
not be matched across markets — say the first market uses 1,2,3,. . . and the second one uses a,b,c,. . . . The
time stamp is to the second in Euronext and to the millisecond in Chi-X. In the analysis, Chi-X data is
aggregated to the second in order to create a fair comparison across markets.
In a final analysis we aim to identify the net effect of middlemen introduction through a difference-in-
6See, “EMCF cuts clearing fees”, The Trade News, 4/24/08.7“Chi-X Europe Celebrates First Anniversary,” Chi-X press release, April 7, 2008.
23
difference analysis (see Section 2.6). The instrument is essentially the introduction of Chi-X and the advent
of the identified middleman to the market. A first step in identifying such effect is to collect and analyze
the exact same data for Euronext in the first 77 trading days in 2007 when there was no Chi-X, nor was the
middleman broker ID active in the Euronext data (comparing the same period in 2007 and 2008 avoids the
impact of calendar effects).8 This is the ‘treated’ sample. To control for all that changed comparing 2007
with 2008, Belgian index stocks are analyzed as the ‘untreated’ sample as Chi-X had not yet been introduced
(as a matter of fact, Chi-X introduced Belgian stocks on April 24, 2008 which motivates the choice of our
main sample period). Other than the absence of Chi-X, Belgian stocks traded in the same way as Dutch
stocks as Euronext operated its trading system across its four markets (Belgium, France, the Netherlands,
and Portugal). In terms of the actual data, the Belgian sample only differs from the Dutch sample in that it
lacks broker IDs on transactions.
The list of all stocks that are analyzed in this study is included as an appendix. It contains security name,
isin code, and weight in the local index.
Finally, a dataset with quotes from the highly active index futures market is used to track changes in both
the Dutch AEX index AEX and the Belgian BEL20 index.
Approach. The analysis was done in two steps (i) to make it feasible (the entire dataset contains roughly
100 million event records) and (ii) to do proper statistical inference. The first step calculates all variables of
interest for each stock-day. For example, it calculates the time-weighted quoted half spread (ask minus bid
divided by two) for Heineken on January 2, 2008. To make activity measures comparable across stocks, we
convert, e.g., the number of shares traded to aeamount by multiplying it with the average transaction price in
the sample period. The second step is a panel data analysis on the ‘box’ with all stock-day results. Standard
errors are based on residuals that are clustered by day so as to avoid ‘double counting’ in the presence of
commonality and to explicitly recognize heteroskedasticity. The results are presented as weighted-averages
across stocks where the weight corresponds to the stock’s weight in the local index (see appendix with the
list of all stocks). Also, results are reported separately for large and small stocks where the cutoff is the
median index weight.
[insert Table 1 here]8This makes the pre-entry period run through April 20, 2007. This appears at odds with the Chi-X effective date of April 16,
2007, but in the data no trade materialized in the Chi-X system until after April 20, 2007.
24
Summary statistics. Table 1 presents summary statistics to illustrate Chi-X role in the trading of Dutch
stocks. It leads to a couple of observations.
1. Chi-X managed to obtain a nontrivial market share within one year of its existence. Chi-X’ share of
overall volume is 8.4% its share of trades is 13.6%. Its performance is particularly strong for large
stocks.
2. The use of Chi-X in addition to Euronext leads to a substantial improvement of liquidity supply. The
average quoted half spread is 3.70 basis points for Euronext and 5.09 basis points for Chi-X. Average
depth is e108,100 for Euronext and e48,500 for Chi-X. The wider spread and lower depth in Chi-X
does not necessarily imply that investors who demand liquidity/immediacy only focus on Euronext.
The results are, for example, consistent with Chi-X always having a strictly better price on one side of
the market, i.e., a strictly lower ask or a strictly higher bid. To assess Chi-X’ contribution to liquidity
supply one might calculate the lowest ask across markets minus the highest bid — the inside spread
— and compare it to the Euronext spread. The inside spread is, effectively, what investors with ‘smart
routers’ pay (see Foucault and Menkveld (2008)).
The inside spread however might lead one to overestimates liquidity supply improvement as the av-
erage Chi-X depth is lower than Euronext depth. To control for depth, we define the ‘generalized’
inside spread which adjusts both the bid and the ask quote for potentially better prices in Chi-X. For
example, the adjusted ask is e29.995 if the Euronext ask is e30.00 with depth e100,000 and the
Chi-X’ ask is e29.99 with depth e50,000. The generalized inside spread is a conservative measure
as it calculated for transaction sizes that consume full Euronext depth; smaller transaction sizes imply
an even tighter spread.
The generalized inside spread is 2.86 basis points, which is a significant 1.02 basis points (-23%)
lower than the Euronext-only spread of 3.70 basis points. A conservative t-value of this differential
assumes perfect correlation and therefore equals 3.92. The differential is statistically significant for
both large stocks (-24%) and small stocks (-16%).
3. A standard effective spread decomposition shows that, by far, its largest component is adverse selec-
tion (91%) which thus supports the focus on information asymmetry in the theoretical model. The
effective spread is defined as the transaction price minus the midquote — the average of the bid and
ask quote — at the time of trade. It can be decomposed into a component that is compensation for
25
being adversely selected and the orthogonal component which is gross profit to the passive order
submitter. A standard decomposition relies on ‘waiting out’ the time it takes until prices reflect the
long-term information in the trade (see, e.g., Glosten (1987)) which we set to 30 minutes. Overall,
we find that the average effective spread is 3.04 basis points which is the sum of a significant 2.78 ad-
verse selection (91%) and an insignificant 0.26 basis points gross profit (9%). The adverse-selection
component is higher for small stocks.
2.3 Caught on tape! a middleman
The anonymization scheme by market makes simple matching of broker IDs across markets impossible.
Instead, we match pairs of broker IDs and find that one combination (say broker 7 and broker d) achieves
mean-reversion in net position within the day. Also, both broker IDs trade very frequently. It thus matches
the SEC’s defition of ‘high frequency trader’ and appears to fit the profile of the middleman in our main
theory.
[insert Figure 7 here]
Figure 7 illustrates the middleman’s trading by plotting inventory in ING stock throughout January 30, 2008
assuming he starts off at zero. This inventory at any particular point in time is thus defined as the number of
shares bought minus the number of shares sold since the start of the trading day. The two top graphs plot this
inventory by market and show that the middleman is quite active and runs into positions of almost 40,000
shares. Yet, these time series also appear nonstationary which, by market, would not qualify these broker
IDs as middlemen. But, if summed across markets, which is the bottom graph, the pattern does exhibit
high-frequency mean-reversion and it seems that we did catch a middleman. Although hard to see from the
graph, the inventory at the end of the day is exactly zero shares.
[insert Table 2 here]
In Table 2, Panels A and B generate statistics in support of the conjecture that the broker IDs underlying
Figure 7 represent a middleman. Panel A reports that on almost half (0.46) of the stock-days in the sample
the middleman’s change in inventory across the day is exactly zero. On average, this daily inventory change
26
is e-56,000. Panel B shows that the middleman trades 1.40 times per minute in Euronext and 0.96 times per
minute in Chi-X. She trades an average e32,000 per minute in Euronext and e21,000 per minute in Chi-X.
The average open-to-close inventory change is therefore of the same magnitude as the amount he trades in a
minute which supports strong inventory mean-reversion. She is a substantial market participant as, based on
these numbers, he is a counterparty in every third trade in Chi-X (35.6%) and every fourteenth trade (7.7%)
in Euronext. Disaggregating across small and large stocks, it appears him participation rate is the same in
Chi-X but substantially lower in Euronext. She also seems less eager to carry positions overnight as the
fraction of days with a zero inventory change is 0.60 as opposed to 0.33 for large stocks.
Panel C of Table 2 reports that most often the middleman is at the passive side of a transaction. In Chi-X,
78.8% of him transactions was another broker’s aggressive order executing against him limit orders waiting
in the book. In Euronext, it was slightly lower, 74.0%. The standard errors, 0.7% and 1.0% respectively,
show that this is a structural pattern as the distance to 50% is statistically significant. This predominant use
of passive orders in addition to the inventory mean-reversion is consistent with dynamic inventory control
models that have been proposed for market makers (see, e.g., Ho and Stoll (1981), Amihud and Mendelson
(1980), and Hendershott and Menkveld (2010)). These models predict an intermediary to skew him quotes
opposite to the direction of him inventory in order to mean-revert. This could explain why Chi-X quotes,
where our middleman is in every third trade, often features strictly better prices on one side of the market
only (see the discussion on the ‘generalized inside spread’ in Section 2.2). Interestingly, disaggregation
across large and small stocks shows that the middleman is less passive for small stocks in Euronext (52.7%)
which, along with him high fraction of days that he ‘goes home flat’, indicates that he often willingly pays
the half spread to keep inventory close to zero.
Panel D of Table 2 reveals adverse selection is a relatively smaller component of the effective spread when
the middleman is on the passive side of the trade relative to when he is not. The panel conditions the
effective spread decomposition on whether or not the middleman was on the passive side of the trade.
The middleman effective spread is 3.47% on average, which is significantly higher than the nonmiddleman
effective spread of 2.96$ (conservative t-value is 3.92). The adverse-selection cost, however, is 2.75 basis
points for middleman trades which is not significantly different from the 2.81 basis points for nonmiddleman
trades. The remainder is gross profits to the passive side of the trade or so-call realized spread. It is 0.72
basis points for middleman trades which is significantly higher than the 0.15 basis points for nonmiddleman
trades (t value is 1.97). Testing against zero reveals that gross profits are only significantly positive if the
27
middleman is on the passive side of the trade. In relative terms, adverse selection is 79% of the effective
spread for middleman trades vs. 95% for nonmiddleman trades. We note that this result should not be
interpreted as evidence that the middleman is bad for the market as he appears to have significantly positive
gross profits on him passive orders. First off, trading conditions might be worse when the middleman is on
the passive side, e.g., volatility might be higher — Section 2.5 will provide evidence for this conjecture.
But, any judgement on whether the middleman is good or bad for liquidity supply or trading in general is
deferred until the diff-in-diff analysis of Section 2.6.
2.4 Information-processing speed
This section studies the theory’s conjecture that the middleman has superior information-processing speed
which makes him a cost-efficient limit-order submitter. The idea is that a fast computer with enormous
processing power can follow all relevant public news and quickly cancel and resubmit limit orders to reflect
it. This reduces the adverse-selection cost associated with limit orders. The theory captures this idea by
allowing the middleman to condition him price on the common value innovation z.
The middleman’s ability to avoid picking-off risk in real markets is not perfect as not all (public) information
is ‘hard’. We believe the middleman is particularly well-positioned to quickly do the ‘statistics’ and infer-
ence a security’s change in fundamental value by tracking price series that correlate with it, e.g., the index
level, same industry stocks, foreign exchange rate, etc. We label such information ‘hard’ information. The
middleman is at a disadvantage for ‘soft’ information which is, for example, an assessment of the quality of
a new management team, the value of a new patent, etc.
The econometric challenge is to test the prediction that the middleman’s quotes reveal hard information on
a stock’s fundamental value before anyone can pick him off. A data limitation is that him quotes are not
observed. Instead, we conjecture that Chi-X quotes are more revealing of middleman quotes than Euronext
quotes as Chi-X is the most friendly venue for their activity (see passive order and speed discussion in
Section 2.1). The identified middleman’s higher participation rate in Chi-X supports it. Midquotes in the
highly active index futures market are used to trace an important piece of hard information that matters for
a security’s fundamental value.
We believe the most appropriate analysis of whether Chi-X (read: middlemen) quotes are more likely to
28
reflect hard information inbetween trades requires a cointegration model that identifies the information in
the trade. But, leading up to such model we first perform two analyses on the raw data to measure (i) the
speeed with which both markets’ quotes reveal index futures information and (ii) to what extent their quote
updates inbetween trades correlate with the long-term information in the trade.
[insert Table 3 here]
Simple raw data analysis. Table 3 shows that Chi-X quotes are more responsive to changes in the Dutch
index futures market than are Euronext quotes. A natural and simple approach is to consider all events where
the index future midquote changes and count how often a stock midquote adjusts in the same second. This
comparison, however, includes ‘mechanical’ stock midquote changes that are the result of executions that
knock off stale quotes rather than the result of quote updates. It is for this reason that the analysis further
conditions down on index changes that are not accompanied by stock transactions one second before, during,
or one second after the change. The results show that same-second Chi-X stock quote updates happen three
times more often than Euronext quote updates and their correlation with the index change is significantly
more positive. Both these observations are statistically significant.
[insert Table 4 here]
Panel A of Table 4 shows that Chi-X quote updates appear more informed than Euronext quote updates.
First, midquote changes strictly inbetween trades are calculated9, i.e., the log midquote one second after
trade (t-1) is subtracted from the log midquote one second prior to trade t (prices are expressed as log prices
whenever price changes are analyzed throughout the study). These are then correlated with the long-term
information revealed in the trade interval which is proxied by transaction price (t+10) minus transaction
price (t-1). This correlation is 0.050 for Chi-X quote updates which is significantly higher than the Euronext
correlation of 0.013. A drawback of this approach is that it does not recognize and strip out transient effects
in quotes due to, e.g., dynamic inventory control. And, it cannot assess to what extent an informational
advantage reflects ‘hard’ information. This is why we turn to a cointegration approach.
A cointegration model. A cointegration model is proposed to gauge quote update informativeness in
both the Euronext and the Chi-X market. The approach extends Hasbrouck (1995) to include the market9The transaction clock used in this table aggregates across markets; it does not distinguish between Chi-X and Euronext trades.
29
index so that quote informativeness can be decomposed into an index-correlated part (hard information)
and a remainder part (which arguably is a mix of hard and soft information). This enables us to quantify
Chi-X quote informativeness and compare it to Euronext quote informativeness. Moreover, it allows for to
decompose any such differential into an index and a nonindex differential to test the conjecture that Chi-X
quote informativeness is particularly strong for hard information.
The cointegration model is defined as:
∆pt B [indext midquote euronextt− midquote chi xt− trade pricet]′ (38)
where t runs over the transaction clock, t− indicates that the quote snapshot is taken one second prior to the
transaction, index is the midquote price in the local index futures, trade pricet is the transaction price, and
midquote Xt− indicates the midquote price in market X.
pt = φ1∆pt−1 + φ2∆pt−2 + · · · + β(A′pt−1) + εt (39)
β′ =
0 β22 β32 β42
0 β21 β31 β41
A′ =
0 1 −1 0
0 1 0 −1
The vector error correction term βA′pt−1 in equation (39) reflects the presence of two random walks, one
associated with the market index and the other with the security’s ‘efficient price’. This common efficient
price disciplines differentials across both midquote price series and the trade price series to be stationary
with mean zero. Price changes are assumed to be covariance stationary which implies that they can be
expressed as a vector moving average (VMA):
∆pt = θ1εt−1 + θ2εt−2 + · · · = θ(L)εt (40)
where L is the lag operator. The two random walks now show up in the coefficient polynomial θ(L) evaluated
at 1 which reflects the long-term response of prices to an error term impulse. This matrix has rank 2, i.e.,
the second, third, and fourth row are equal as all three are security prices that in the long-term agree on what
the current shock’s impact is on the efficient price. A useful econometric proxy for this efficient price is the
best long-term linear forecast of prices conditional on all historical price information up until and including
30
time t:
ft B limt→∞
E∗[pt+k|pt, pt−1, . . . ] (41)
where the asterisk indicates that it is the best linear forecast. Hasbrouck (2007, Ch.8) shows that the forecast
innovation from (t-1) to t for the efficient price is:
∆ ft = [θ(1)]2εt (42)
where [.]i indicates ‘the ith row of the matrix in brackets’.
The forecast innovations ∆ ft are a natural measure for the theory’s common value changes in between trades
and linear projections allows for further analysis of these common-value innovations. For ease of exposition,
let Px(y) be the best linear projection of the random variable y on random variable x. In regression terms,
Px(y) B x′β (43)
where β is the coefficient of a standard linear regression of y on x. The projections allow for the following
analysis.
1. Quote informativeness inbetween is naturally measured by the variance of the efficient price projection
onto the quote innovation, e.g., var(Pe(∆ ft)) where Pe projects onto the Euronext quote innovation
[ε]2. In other words, how much information in the intertrade interval can be learned from obtaining
a market’s quote update? This measure is zero in a Glosten and Milgrom (1985) type of model and
equal to the full ∆ ft variance if liquidity demanders are uninformed.
2. The variance of efficient price changes (as measured by ∆ ft) and its projection onto quote innovations
can be decomposed into an index-correlated component (hard information) and an orthogonal com-
ponent. For example, var(Pm ◦Pe(∆ ft)) where m corresponds to [ε]1 indicates how much of Euronext
quote informativeness reflects the index innovation. If Euronext quote updates are uncorrelated with
index innovations this measure is zero. P−m is defined to be the orthogonal part, i.e., y = (Pm+P−m)(y)
by construction.
Panel B of Table 4 provide empirical support for Chi-X quotes being significantly more informative on hard
information. The proxy for such information is the index-correlated part of efficient price changes which
31
amounts to 42% (=100%*4.02/9.48). By Roll’s standards these firms are large and this percentage cor-
responds to him finding that most large U.S. firm are in the 40% to 50% range (see Roll (1988, p.545)).
Projecting efficient price changes onto Euronext and Chi-X quotes reveals that Chi-X quotes are more infor-
mative (3.83 vs 3.54 basis points squared) but this difference is not significant. If, however, this differential
is decomposed into index and nonindex components, Chi-X is significantly more informative on the index
component (0.30 vs. 0.05 basis points squared with a t value of 5.0 for the differential).
Disaggregating according to large and small stocks reveals a cross-sectional heterogeneity as Chi-X quotes
are significantly more informative for large stocks, but significantly less informative for small stocks. For
large stocks, Chi-X quote innovations reveal 3.70 basis points squared of the intertrade innovation which
is a significant 28% higher than the 2.90 basis points squared for Euronext. For small stocks on the other
hand, Chi-X quote updates reveal 4.60 basis points squared which a significant 37% lower than Euronext
the informativeness of Euronext quote updates. Decomposing according to index and nonindex information
reveals that Chi-X quote appear significantly more informative on the index part for both types of stocks,
but are significantly less informative on the nonindex for small stocks. For large stocks the nonindex part is
not significantly different across markets (although Chi-X informativeness is slightly higher for this part).
2.5 Middleman participation: which stocks and when?
[insert Table 5 here]
This section studies middleman participation in both the cross-section (through ‘between’ correlations) and
in the time dimension (through ‘within’ correlations). Table 5 presents these correlations of various trading
variables which lead to the following observations.
[insert Figures 8 here]
The relative size of hard information as proxied by the ‘R2 of a single factor CAPM’ correlates positively
with middleman trade participation (0.67 between and 0.46 within) and Chi-X share of trades (0.64 between
and 0.23 within). Figure 8 illustrates the 0.46 within correlation by plotting middleman participation against
the size of hard information. It reveals that this strong correlation does not appear to be driven by outliers
32
and is based on considerable time variation in both variables, double-digit percentages. The increase in
Chi-X share of trades is not surprising given that it is a middleman’s preferred habitat (low fees and a
fast system). The table also shows that middleman participation correlates positively with Chi-X trade
share (0.89 between and 0.41 within) which is further evidence that the relationship between Chi-X and
middlemen appears symbiotic.
[insert Figure 9 here]
The differential between Chi-X and Euronext price-quote informativeness correlates positively with the
relative amount of hard information (0.64 between and 0.23 within). A market’s price-quote informativeness
is defined as the predictive power of midquote price changes for the amount of information that is revealed
inbetween trades. The methodology is based on linear projections (Pe − Pi)(∆ f ) and a cointegration model
(see Section 2.4). Figure 9 illustrates the cross-sectional relationship by plotting the size of hard information
against both market’s price-quote informativeness for each stock. The black dots represent Chi-X, the white
dots represent Euronext, and dot size corresponds to stock’s overall volume. The graph leads to the following
two observations. First, black dots appear to be above white dots on the right-hand side of the graph, below
them on the left-hand side. This illustrates the positive between-correlation of hard information size and
Chi-X minus Euronext price-quote informativeness. Second, dot size is generally larger on the right-hand
side of the graph. Hard information is relatively more important for active stocks.
[insert Figure 10 here]
Figure 10 replots Figure 9 but decomposes price-quote informativeness into an index-related part (top graph)
and an orthogonal part (bottom graph). It illustrates that Chi-X price-quote informativeness is higher than
Euronext quote informativeness across all stocks for the hard information (black dots are generally above
with dots in the top graph) and results are mixed for the nonindex part which is likely to be a combination of
hard and soft information. The latter result illustrates that it is not only index information that drives Chi-X
quote efficiency.
33
2.6 The counterfactual: what if middlemen had not been introduced?
As we cannot rerun trading in the first 77 days of 2008 without Chi-X and the identified middleman, we
revert to a diff-in-diff approach as a second best. As discussed in Section 2.2, additional data have been
collected to create a pre-event sample of the first 77 trading days of 2007 to compare before and after Chi-X
introduction. This difference for the ‘treated’ sample is then compared to the difference for an untreated
sample of Belgian stocks. Belgium is a natural choice as it a neighboring country whose stocks also trade in
the Euronext system. The difference in difference (‘∆ Dutch - ∆ Belgian’) identifies a treatment effect.
[insert Table 6 here]
The diff-in-diff results of Table 6 (Panel A) reveal that Chi-X introduction raises liquidity supply, does not
affect the number of trades, and lowers trading volume. Before discussing these results, we like to point out
that the comparison across Dutch and Belgian stocks is not perfect as the latter typically trade less actively,
3.71 vs. 11.05 trades per minute in the post-event period. Yet, volatility increase is comparable across
markets; realized volatility increases slightly less for Dutch stocks (64% vs. 69%) whereas the increase of
intertrade volatility based on the cointegration model is not significantly different across markets (41% vs
39%). Bearing this in mind, the table leads to the following observations:
1. The generalized inside spread which equals the Euronext spread in the absence of Chi-X has in-
creased by 35% in Belgian stocks (reflecting the higher volatility), yet stayed put for Dutch stocks
(0%). The treatment effect is a significant 35%. Depth at this quote has declined by 13% more
for Dutch stocks but we consider this effect to be second order as this decrease does not undo the
large price discount. If one were to transact the 13% at one tick behind the best price quote this
implies a 175% worse price10 and the overall effect is thus still a spread improvement of 100%*(1-
(0.87*0.65+0.13*0.65*1.75))=29%. The significant treatment effect of minus 13% for effective spread
further supports a general increase in liquidity supply on the introduction of Chi-X/ middlemen.
2. The number of trades in Dutch stock increases by 53% which is not significantly different from the in-
crease of 55% witnessed for Belgian stocks. The number of trades in Dutch stocks has been corrected
10This calculation is based on a one cent tick size and an average share price of ∼e20 which imply a 5 basis points worse priceon a one tick move. This is a 175% increase relative to the 2.86 basis points average half spread (see Table 1).
34
for double-counting of trades due to the presence of the new middleman that has been identified (see
Table 2). In other words, if a security that in the past traded directly between investor A and B now
travels via the middleman this would artificially inflate trade activity.
3. Volume increases by only 5% for Dutch stocks, which is a significant 15% less than the 21% increase
in Belgian stocks. Again, this volume was corrected for double-counting.
3 Conclusion
We model high-frequency traders in electronic markets. We base this conclusion on the introduction of
Chi-X, an HFT-hospitable market, and on being able to compare the post Chi-X entry change in the trading
of Dutch stocks which do trade on Chi-X, and similar Belgian stocks that do not. We showed evidence that
middlemen are better informed about recent news than the average investor, in that their reaction times were
faster and in the right direction.
Being better informed, middlemen still can make a positive or a negative contribution to welfare. On the
one hand, they can raise welfare by solving a pre-existing adverse-selection problem. In that case their entry
should be accompanied by a rise in trade and a fall in bid-ask spreads. On the negative side, they can create
or exacerbate a pre-existing adverse-selection problem in which case bid-ask spreads should rise and trade
declines.
Our evidence on the welfare contribution of middlemen is mixed. On the one hand, middlemen’s participa-
tion lowers the effective bid-ask spreads but, on the other, it does not affect the number of trades. The net
effect is therefore uncertain.
Our theoretical analysis and the mixed evidence on welfare suggest that there is room for optimal market
design. Regulators and market operators should think carefully about how adverse-selection risk affects the
various participants. For example, the speed privilege that HFTs can buy into, co-location, might require a
differentiated order-fee schedule. Passive orders submitted through this pipe might optimally be rewarded
more whereas aggressive orders might have to be charged more. The reason is that passive orders come with
the positive externality of liquidity supply to others whereas aggressive orders have a negative externality of
creating adverse selection for non-co-located participants. The latter is, in spirit, similar to the old NYSE
35
market structure where specialists were not allowed to have a live data feed of market-index information
into their system. Also, markets might want to enable limit-order submitters to peg their order to, e.g., the
market index (cf. Black (1995)).
36
Appendix I: Variable/parameter description and modeling device discussion
Variables:
variable meaningM middleman/machine, a representation of high-frequency traders
S seller, an investor who owns the asset and arrives early
B buyer, an investor like S except that he does not own the asset and arrives late (no overlapwith S)
x,y private value of asset ownership for S and B, respectively
z=zh+zs common value innovation of asset ownership produced in between arrival of S and B
zh, zs zh is hard information, i.e., “quantitative, easy to store and transmit in impersonal ways,” ;zs is soft information; zh is known by M and potentially partially by B; zs is known by Bonly
Parameters:
parameter description modeling device for. . .σ2
x=σ2y=σ
2=π
µx=µy=µ
mean and variance of indepen-dent normally distributed pri-vate valus of ownership for Sand B, respectively
M private value of ownership is set to zero; vari-ance of private value is set to π so that expectedsocial value of optimal re-allocation is standard-ized to one (E max(x, y) − µ = σ√
π); µ captures
expected social loss of M ownership relative toinvestor ownership (by construction expressed inunits of social value of re-allocation as the latterwas set to equal 1)
σ2z variance of normally distributed
common value innovation zsize of winner’s curse (adverse selection) frictionin the absence of M
λ probability that B observes hardinformation zh
extent to which M participation produces socialvalue (might be negative for small λ)
α relative size of hard informationin z
relative size on how informed M is on z
θ relative size of common valuevariance (σ2
z ) relative to privatevalue variance (σ2
x)
relative size of the informational friction
37
App
endi
xII
:Lis
tofa
llst
ocks
Thi
sta
ble
lists
alls
tock
sth
atha
vebe
enan
alyz
edin
this
man
uscr
ipt.
Itre
port
sth
eoffi
cial
isin
code
,the
com
pany
’sna
me,
and
the
inde
xw
eigh
tw
hich
has
been
used
thro
ugho
utth
est
udy
toca
lcul
ate
(wei
ghte
d)av
erag
es.
Dut
chin
dex
stoc
ks/
‘tre
ated
’sam
ple
Bel
gium
inde
xst
ocks/
‘unt
reat
ed’s
ampl
eis
inco
dese
curi
tyna
me
inde
xw
eigh
tais
inco
dese
curi
tyna
me
inde
xw
eigh
ta
NL
0000
3036
00in
ggr
oep
22.3
%B
E00
0380
1181
fort
is17
.5%
NL
0000
0094
70ro
yald
utch
petr
ol20
.1%
BE
0003
5657
37kb
c15
.8%
NL
0000
0095
38ko
nph
ilips
elec
tr12
.1%
BE
0003
7961
34de
xia
13.5
%N
L00
0000
9355
unile
ver
11.3
%B
E00
0379
3107
inte
rbre
w9.
5%N
L00
0030
3709
aego
n7.
5%B
E00
0347
0755
solv
ay6.
5%N
L00
0000
9082
koni
nklij
kekp
n7.
4%B
E00
0379
7140
gpe
brux
el.la
mbe
rt5.
7%N
L00
0000
9066
tnt
4.8%
BE
0003
5627
00de
lhai
zegr
oup
5.4%
NL
0000
0091
32ak
zono
bel
4.2%
BE
0003
8102
73be
lgac
om4.
4%N
L00
0000
9165
hein
eken
3.0%
BE
0003
7395
30uc
b4.
1%N
L00
0000
9827
dsm
2.4%
BE
0003
8456
26cn
p2.
9%N
L00
0039
5903
wol
ters
kluw
er2.
2%B
E00
0377
5898
colr
uyt
2.3%
NL
0000
3606
18sb
moff
shor
e1.
2%B
E00
0359
3044
cofin
imm
o2.
2%N
L00
0037
9121
rand
stad
1.1%
BE
0003
7647
85ac
kerm
ans
and
van
haar
en2.
2%N
L00
0038
7058
tom
tom
0.6%
BE
0003
6788
94be
fimm
o-si
cafi
2.1%
BE
0003
8264
36te
lene
t2.
1%B
E00
0373
5496
mob
ista
r1.
5%B
E00
0378
0948
beka
ert
1.1%
BE
0003
7850
20om
ega
phar
ma
1.0%
a :The
inde
xw
eigh
tsar
eba
sed
onth
etr
uein
dex
wei
ghts
ofD
ecem
ber3
1,20
07.T
hew
eigh
tsar
ere
scal
edto
sum
upto
100%
ason
lyst
ocks
are
reta
ined
that
wer
ea
mem
bero
fthe
inde
xth
roug
hout
the
sam
ple
peri
od.T
his
allo
ws
forf
airc
ompa
riso
nsth
roug
htim
e.
38
References
Amihud, Y. and H. Mendelson (1980). Dealership market: Market-making with inventory. Journal ofFinancial Economics 8, 31–53.
Biais, B., P. Hillion, and C. Spatt (1995). An empirical analysis of the limit order book and the order flowin the Paris Bourse. Journal of Finance 50, 1655–1689.
Black, F. (1995). Equilibrium exchanges. Financial Analysts Journal 51, 23–29.
Brogaard, J. A. (2010). High-frequency trading and its impact on market quality. Manuscript, KelloggSchool of Management.
CESR (2010). Trends, risks, and vulnerabilities in financial markets. Release No. 10-697.
Chaboud, A., B. Chiquoine, E. Hjalmarsson, and C. Vega (2009). Rise of the machines: Algorithmictrading in the foreign exchange market. Manuscript, Federal Reserve Board.
Clark, C. E. (1961). The greatest of a finite set of random variables. Operations Research 9, 145–162.
Duffie, D., N. Garleanu, and L. Pedersen (2005). Over-the-counter markets. Econometrica 73, 1815–1847.
Duffie, D., S. Malamud, and G. Manso (2009). Information percolation with equilibrium search dynam-ics. Econometrica 77, 1513–1574.
Foucault, T. and A. J. Menkveld (2008). Competition for order flow and smart order routing systems.Journal of Finance 63, 119–158.
Foucault, T., A. Roell, and P. Sandas (2003). Market making with costly monitoring: An analysis of thesoes controversy. Review of Financial Studies 16, 345–384.
Glosten, L. (1987). Components of the bid ask spread and the statistical properties of transaction prices.Journal of Finance 42, 1293–1307.
Glosten, L. and P. Milgrom (1985). Bid, ask, and transaction prices in a specialist market with heteroge-neously informed agents. Journal of Financial Economics 14, 71–100.
Goettler, R., C. Parlour, and U. Rajan (2009). Informed traders in limit order markets. Journal of Finan-cial Economics 93, 67–87.
Grossman, S. J. and M. H. Miller (1988). Liquidity and market structure. Journal of Finance 43, 617–633.
Hasbrouck, J. (1995). One security, many markets: Determining the contributions to price discovery.Journal of Finance 50, 1175–1199.
Hasbrouck, J. (2007). Empirical Market Microstructure. New York: Oxford University Press.
Hasbrouck, J. and G. Saar (2010). Low-latency trading. Manuscript, Cornell University.
Hendershott, T., C. M. Jones, and A. J. Menkveld (2011). Does algorithmic trading improve liquidity?Journal of Finance 66, 1–33.
Hendershott, T. and A. J. Menkveld (2010). Price pressures. Manuscript, VU University Amsterdam.
Hendershott, T. and R. Riordan (2009). Algorithmic trading and information. Manuscript, University ofCalifornia, Berkeley.
Ho, T. and H. R. Stoll (1981). Optimal dealer pricing under transaction cost and return uncertainty.Journal of Financial Economics 9, 47–73.
39
Hollifield, B., R. A. Miller, P. Sandås, and J. Slive (2006). Estimating the gains from trade in limit ordermarkets. Journal of Finance 61, 2753–2804.
Kirilenko, A., A. Kyle, M. Samadi, and T. Tuzun (2010). The flash crash: The impact of high frequencytrading on an electronic market. Manuscript, U of Maryland.
Kyle, A. (1985). Continuous auctions and insider trading. Econometrica 53, 1315–1335.
Lagos, R., G. Rocheteau, and P.-O. Weill (2009). Crashes and recoveries in illiquid markets. Manuscript,UCLA.
Li, Y. (1996). Middlemen and private information. Journal of Monetary Economics 42, 131–159.
Masters, A. (2007). Middlemen in search equilibrium. International Economic Review 48, 343–362.
Menkveld, A. J. (2011). High frequency trading and the new-market makers. Manuscript, VU UniversityAmsterdam.
Parlour, C. A. and D. J. Seppi (2008). Limit order markets: A survey. In A. Boot and A. Thakor (Eds.),Handbook of Financial Intermediation and Banking, Amsterdam, Netherlands. Elsevier Publishing.
Petersen, M. A. (2004). Information: Hard and soft. Manuscript, Northwestern University.
Roll, R. (1988). R2. Journal of Finance 43, 541–566.
Rubinstein, A. and A. Wolinsky (1987). Middlemen. Quarterly Journal of Economics 102, 581–593.
SEC (2010). Concept release on equity market structure. Release No. 34-61358; File No. S7-02-10.
40
Table 1: Summary statistics
This table provides summary statistics on a sample of 14 Dutch index stocks that trade both in theincumbent market Euronext and in the entrant Chi-X (Chi-X). The sample runs from Jan 1, 2008 throughApril 23, 2008. The table reports weighted averages where weights are based on a stock’s local indexweight. Time-clustered standard errors account for commonality and heteroskedasticity and are reported inparentheses.
variable (units) large small allEuronext volume (e1000/min) 524.4
(19.9)128.1
(4.3)466.6(17.4)
Chi-X volume (e1000/min) 49.4(2.0)
4.5(0.2)
42.8(1.8)
Chi-X share volume (%) 8.6 3.4 8.4
Euronext #trades (/min) 19.03(0.62)
9.30(0.21)
17.61(0.55)
Chi-X #trades (/min) 3.16(0.12)
0.57(0.03)
2.78(0.10)
Chi-X share #trades (%) 14.2 5.7 13.6
Euronext time-weighted quoted half spread (basis points) 3.47(0.07)
5.00(0.14)
3.70(0.08)
Chi-X time-weighted quoted half spread (basis points) 3.44(0.10)
14.76(0.98)
5.09(0.20)
time-weighted generalized inside half spreada (basis points) 2.63(0.18)
4.20(0.24)
2.86(0.18)
Euronext time-weighted quoted depth (e1000) 121.4(2.4)
30.6(0.3)
108.1(2.0)
Chi-X time-weighted quoted depth (e1000) 53.3(1.2)
21.0(1.1)
48.5(1.0)
trade-weighted effective half spread (basis points) 2.89(0.06)
3.90(0.10)
3.04(0.06)
trade-weighted adverse selection, 30 min (basis points) 2.62(0.14)
3.74(0.21)
2.78(0.14)
trade-weighted realized spread, 30 min (basis points) 0.28(0.14)
0.16(0.17)
0.26(0.13)
N=1078 (14 stocks, 77 days)a: defined as a Chi-X adjusted Euronext half spread in the sense that it is the cost of demanding the full Euronext depthbut re-routing (part of) the order to Chi-X if Chi-X has strictly better prices; it is a ‘generalized’ inside spread since itcontrols for depth; for example, the adjusted ask is e29.995 if the Euronext ask is e30.00 with depth e100,000 and theChi-X’ ask is e29.99 with depth e50,000
41
Table 2: Caught on tape! a middleman
This table produces statistics on middleman trading for Dutch index stocks from January 1 through April 23,2008. The middleman is discovered as a combination of an anonymous Chi-X broker ID and an anonymousEuronext broker ID that achieves high-frequency trading and mean-reversion in inventory across markets.Figure 7 plots middleman inventory for a single stock on a representative day to illustrate him trading. Time-clustered standard errors account for commonality and heteroskedasticity and are reported in parentheses.
large small allPanel A: middleman inventoryaverage net change in middleman inventory (e1000) −75.6
(93.3)55.1(19.6)
−56.5(79.8)
standard deviation net change in middleman inventory (e1000) 1, 417.2(186.7)
298.4(33.5)
1, 314.6(172.3)
fraction of days with zero net change in inventory 0.33 0.60 0.46
Panel B: middleman activitymiddleman Euronext volume (e1000/min) 36.8
(1.8)4.3(0.3)
32.0(1.6)
middleman Chi-X volume (e1000/min) 24.2(1.8)
1.9(0.2)
21.0(1.6)
middleman Euronext #trades (/min) 1.56(0.07)
0.43(0.03)
1.40(0.06)
middleman Chi-X #trades (/min) 1.09(0.08)
0.19(0.02)
0.96(0.07)
middleman participation rate Euronext trades (%) 8.2(0.3)
4.8(0.3)
7.7(0.3)
middleman participation rate Chi-X trades (%) 35.7(1.8)
35.2(2.3)
35.6(1.8)
Panel C: middleman order typesmiddleman relative use of passive orders in Euronext (%) 78.6
(0.8)53.5(2.7)
74.9(0.9)
middleman relative use of passive orders in Chi-X (%) 76.9(0.5)
84.8(0.8)
78.0(0.5)
Panel D: middleman vs nonmiddleman effective spread and its decompositionmiddleman trade-weighted effective half spread (basis points) 3.25
(0.06)4.72(0.11)
3.47(0.07)
nonmiddleman trade-weighted effective half spread (basis points) 2.81(0.06)
3.81(0.10)
2.96(0.06)
middleman trade-weighted adverse selection, 30 min (basis points) 2.54(0.17)
3.98(0.66)
2.75(0.17)
nonmiddleman trade-weighted adverse selection, 30 min (basis points) 2.64(0.15)
3.78(0.21)
2.81(0.14)
middleman trade-weighted realized spread, 30 min (basis points) 0.72(0.17)
0.74(0.62)
0.72(0.16)
nonmiddleman trade-weighted realized spread, 30 min (basis points) 0.17(0.14)
0.03(0.17)
0.15(0.13)
N=1078 (14 stocks, 77 days)
42
Table 3: Speed comparison across markets
This table compares Euronext and Chi-X in terms of how quickly their price quotes for stocks reflectschanges in the Dutch local index (AEX) future. A natural approach is to consider all events where the indexfuture midquote (the average of the bid and ask quote) changes and count how often the stock midquote isadjusted in the same second. This comparison, however, includes ‘mechanical’ stock midquote changes thatare the result of executions that knock off stale quotes rather than quote updates. It is for this reason thatthe analysis further conditions down on index future midquote updates that show no stock transactions onesecond before, during, or one second after the index futures quote update. The sample Dutch index consistsof Dutch index stocks from January 1 through April 23, 2008. Time-clustered standard errors account forcommonality and heteroskedasticity and are reported in parentheses.
large small allcount of Euronext and index futures quote change in same second (/day) 426
(24)637(40)
457(25)
count of Chi-X and index futures quote change in same second (/day) 1256(52)
1257(58)
1256(50)
correlation Euronext quote change and index futures quote change 0.33(0.01)
0.16(0.01)
0.31(0.01)
correlation Chi-X quote change and index futures quote change 0.46(0.02)
0.15(0.01)
0.41(0.01)
14 stocks, 77 days
43
Table 4: Euronext and Chi-X quote informativeness
This table analyzes to what extent midquote (the average of the bid and ask quote) updates inbetween tradesreveal the information that arrives in the intertrade intervals. The observation clock runs in transaction time.Panel A correlates log midquote changes strictly inbetween trades with the information revealed in the tradeinterval which is proxied by the the log trade price (t+10) minus log trade price (t-1). Panel B is based oncointegration model which stacks all price series of interest into a single price vector
∆pt B [indext midquote euronextt− midquote chi xt− trade pricet]′
where t runs over the transaction clock, t− indicates that the quote snapshot is taken one second prior tothe transaction, index is the midquote in the local index futures, trade pricet is the transaction price, andmidquote Xt− indicates the midquote in market X. The price series is modeled as a vector error correctionmodel (VECM) to capture cointegration:
pt = φ1∆pt−1 + φ2∆pt−2 + · · · + β(A′pt−1) + εt
β′ =
(0 β22 β32 β420 β21 β31 β41
)A′ =
(0 1 −1 00 1 0 −1
)The βA′pt−1 reflects the presence of two random walks, one associated with the market index and the otherwith the security’s ‘efficient price’ which is naturally defined as
ft B limt→∞
E∗[pt+k|pt, pt−1, . . . ]
where the asterisk indicates that it is the best linear forecast (see Hasbrouck (2007, Ch.8)). The extent towhich Chi-X and Euronext quotes reveal efficient prices and whether it is the index or the nonindex compo-nent is established through linear projection of ∆ ft on the price innovation vector εt where Pm denotes a pro-jection on the first element which captures the market-index innovation (and P−m is its residual), Pe projectsonto the second element which is the Euronext quote innovation, and Pc projects onto the third elementwhich is the Chi-X quote innovation. The sample Dutch index consists of Dutch index stocks from January1 through April 23, 2008. Time-clustered standard errors account for commonality and heteroskedasticityand are reported in parentheses.
large small allPanel A: correlations based on raw dataa
correlation Euronext midquote return and long-term price impact of (signed) trade 0.007(0.003)
0.046(0.003)
0.013(0.003)
correlation Chi-X midquote return and long-term price impact of (signed) trade 0.052(0.004)
0.039(0.003)
0.050(0.003)
Panel B: cointegration analysisa
overall efficient price innovationvariance efficient price innovation inbetween trades, ∆ f 8.19
(0.49)16.98(1.16)
9.47(0.58)
variance efficient price innovation correlated with market index, Pm(∆ f ) 3.74(0.28)
5.66(0.51)
4.02(0.30)
variance efficient price orthogonal to market index, P−m(∆ f ) 4.45(0.28)
11.32(0.75)
5.45(0.33)
- continued on next page -
44
- continued from previous page -large small all
efficient price innovation correlated with the Euronext midquote returnvariance efficient price innovation inbetween trades, Pe(∆ f ) 2.90
(0.17)7.25(0.45)
3.54(0.21)
variance efficient price innovation correlated with market index, Pm ◦ Pe(∆ f ) 0.05(0.00)
0.08(0.01)
0.05(0.00)
variance efficient price orthogonal to market index, P−m ◦ Pe(∆ f ) 2.85(0.17)
7.18(0.45)
3.48(0.21)
efficient price innovation correlated with the Chi-X midquote returnvariance efficient price innovation inbetween trades, Pc(∆ f ) 3.70
(0.17)4.60(0.44)
3.83(0.20)
variance efficient price innovation correlated with market index, Pm ◦ Pc(∆ f ) 0.33(0.02)
0.14(0.02)
0.30(0.02)
variance efficient price orthogonal to market index, P−m ◦ Pc(∆ f ) 3.37(0.16)
4.47(0.43)
3.53(0.19)
14 stocks, 77 daysa: based on midquotes that are sampled strictly inbetween trades (i.e. one second after the last trade and one second ahead ofthe next trade) in order to avoid spurious correlation due a trade knocking off the best bid or ask quote
45
Tabl
e5:
Cor
rela
tion
daily
vari
able
s
Thi
sta
ble
pres
ents
corr
elat
ions
fora
pane
ldat
aset
oftr
adin
gva
riab
les
that
are
calc
ulat
edby
stoc
k-da
y.It
dist
ingu
ishe
d‘b
etw
een’
and
‘with
in’
corr
elat
ions
whi
chst
udy
vari
able
inte
rdep
ende
nce
inth
ecr
oss-
sect
ion
and
thro
ugh
time,
resp
ectiv
ely.
The
with
inco
rrel
atio
nis
base
don
time
mea
ns:
x i=
1 T∑ T t=
1x i
t.T
hebe
twee
nco
rrel
atio
nis
base
don
day
t’sde
viat
ion
rela
tive
toth
etim
em
ean:
x∗ it=
x it−
x i.T
hesa
mpl
eD
utch
inde
xco
nsis
tsof
Dut
chin
dex
stoc
ksfr
omJa
nuar
y1
thro
ugh
Apr
il23
,200
8.St
anda
rder
rors
are
repo
rted
inpa
rent
hese
s(w
ithin
corr
elat
ion
stan
dard
erro
rsac
coun
tfor
com
mon
ality
and
hete
rosk
edas
ticity
).
vari
able
(uni
ts)
corr
type
vari
ance
effpr
ice
inno
va-
tion
mid
dle-
man
part
ic-
ipat
ion
rate
#tra
des
Chi
-Xsh
are
#tra
des
mid
dle-
man
rela
tive
use
ofpa
ssiv
eor
ders
Chi
-Xm
inus
Eu-
rone
xtqu
ote
info
rma-
tiven
ess,
(Pc−
Pe)
(∆f)
rela
tive
size
inde
xco
mpo
nent
a(%
)be
twee
nb−0.5
0(0.2
7)0.
67∗
(0.2
7)0.
55∗
(0.2
7)0.
75∗∗
(0.2
7)0.
07(0.2
7)0.
64∗
(0.2
7)w
ithin
c0.
13(0.1
0)0.
46∗∗
(0.0
5)−0.0
9(0.0
8)0.
13∗∗
(0.0
4)0.
04(0.0
7)0.
23∗∗
(0.0
8)
vari
ance
effpr
ice
inno
vatio
nin
betw
een
trad
es,∆
fbe
twee
nb−0.7
3∗∗
(0.2
7)−0.6
2∗(0.2
7)−0.6
1∗(0.2
7)−0.6
4∗(0.2
7)−0.8
8∗∗
(0.2
7)w
ithin
c0.
15∗∗
(0.0
5)0.
55∗∗
(0.1
2)−0.1
5∗(0.0
7)0.
03(0.0
6)−0.0
4(0.0
8)
mid
dlem
anpa
rtic
ipat
ion
rate
(%)
betw
eenb
0.64∗
(0.2
7)0.
89∗∗
(0.2
7)0.
53∗
(0.2
7)0.
78∗∗
(0.2
7)
with
inc
0.11∗
(0.0
5)0.
41∗∗
(0.0
5)0.
08(0.0
5)0.
20∗∗
(0.0
5)
#tra
des
(/m
in)
betw
eenb
0.73∗∗
(0.2
7)0.
34(0.2
7)0.
40(0.2
7)w
ithin
c0.
01(0.0
6)−0.0
5(0.0
7)−0.2
9∗∗
(0.0
6)
Chi
-Xsh
are
#tra
des
(%)
betw
eenb
0.23
(0.2
7)0.
67∗
(0.2
7)w
ithin
c−0.1
5∗∗
(0.0
5)0.
05(0.0
4)
mid
dlem
anre
lativ
eus
eof
pass
ive
orde
rs(%
)be
twee
nb0.
52∗
(0.2
7)w
ithin
c−0.0
2(0.0
4)
14st
ocks
,77
days
a :siz
eof
the
inde
xco
mpo
nent
inth
eeffi
cien
tpri
cein
nova
tion
b :bas
edon
the
time
mea
ns:
x i=
1 T
∑ T t=1
x it
c :bas
edon
day
t’sde
viat
ion
rela
tive
toth
etim
em
ean:
x∗ it=
x it−
x i∗ /∗∗
:sig
nific
anta
ta95/9
9%le
vel
46
Tabl
e6:
Diff
-in-d
iffan
alys
istr
adin
gva
riab
les
Thi
sta
ble
com
pare
str
adin
gva
riab
les
base
don
Dut
chin
dex
stoc
ksbe
fore
and
afte
rth
ein
trod
uctio
nof
Chi
-Xan
dth
ead
vent
ofm
iddl
emen
(e.g
.,th
eID
ofth
em
iddl
eman
we
iden
tified
inTa
ble
2di
dap
pear
inth
epr
e-ev
ent
peri
od).
Itco
mpa
res
the
first
77tr
adin
gda
ysof
2007
(thr
ough
Apr
il20
)an
d20
08(t
hrou
ghA
pril
23).
Thi
sse
rves
asth
e‘t
reat
ed’
sam
ple.
The
sam
eco
mpa
riso
nis
done
for
Bel
gian
inde
xst
ocks
whi
chcr
eate
san
‘unt
reat
ed’
sam
ple
asC
hi-X
had
not
yet
been
intr
oduc
ed.
Bel
gium
isa
natu
ral
choi
ceas
ita
neig
hbor
ing
coun
try
who
sest
ocks
also
trad
ein
the
Eur
onex
tsys
tem
.A
diff
-in-
diff
anal
ysis
then
iden
tifies
the
trea
tmen
teff
ect.
The
perc
enta
gech
ange
was
dete
rmin
edba
sed
onth
elo
gse
ries
.Ti
me-
clus
tere
dst
anda
rder
rors
acco
untf
orco
mm
onal
ityan
dhe
tero
sked
astic
ityan
dar
ere
port
edin
pare
nthe
ses.
vari
able
(uni
ts)
Net
herl
ands/‘
trea
ted’
Bel
gium/‘
untr
eate
d’diff
-in-
diff
a
pre
post
∆pr
epo
st∆
∆∆
Pane
lA:F
ulls
ampl
e20
-min
real
ized
vola
tility
(bp/
min
)3.
9(0.1
)7.
6(0.3
)64
%∗∗
(4%
)4.
1(0.1
)8.
3(0.4
)69
%∗∗
(4%
)−4
%∗
(2%
)
vola
tility
effici
entp
rice
inno
vatio
nin
betw
een
trad
es,∆
f(b
p)1.
9(0.0
)2.
9(0.1
)41
%∗∗
(3%
)3.
5(0.1
)5.
0(0.1
)39
%∗∗
(3%
)2% (2
%)
time-
wei
ghte
dge
nera
lized
insi
deha
lfsp
read
a(b
asis
poin
ts)
2.93
(0.0
3)2.
86(0.1
8)0% (2
%)
5.00
(0.0
8)6.
86(0.1
6)35
%∗∗
(2%
)−3
5%∗∗
(2%
)
time-
wei
ghte
dqu
oted
dept
h(e
1000
)21
3(2
)10
8(2
)−6
2%∗∗
(2%
)67 (1
)46 (0
)−4
9%∗∗
(2%
)−1
3%∗∗
(2%
)
trad
e-w
eigh
ted
effec
tive
half
spre
ad(b
asis
poin
ts)
2.61
(0.0
2)3.
04(0.0
6)13
%∗∗
(2%
)3.
99(0.0
7)5.
07(0.1
2)27
%∗∗
(2%
)−1
3%∗∗
(1%
)
trad
e-w
eigh
ted
adve
rse
sele
ctio
n,30
min
(bas
ispo
ints
)1.
89(0.0
9)2.
78(0.1
4)32
%∗∗
(6%
)3.
16(0.1
1)5.
23(0.2
2)54
%∗∗
(5%
)−2
1%∗∗
(7%
)
#tra
des
(/m
in)
11.0
5(0.3
4)20.3
9(0.6
3)59
%∗∗
(3%
)3.
71(0.1
2)6.
89(0.2
1)55
%∗∗
(3%
)3% (2
%)
#tra
des
afte
rrem
ovin
gm
iddl
eman
’str
ades
(/m
in)
11.0
5(0.3
4)19.8
0(0.6
1)56
%∗∗
(3%
)3.
71(0.1
2)6.
89(0.2
1)55
%∗∗
(3%
)0% (2
%)
volu
me
(e10
00/m
in)
446
(16)
509
(18)
10%∗
(4%
)72 (2
)10
0(4
)21
%∗∗
(4%
)−1
0%∗∗
(2%
)
volu
me
afte
rrem
ovin
gm
iddl
eman
’svo
lum
e(e
1000/m
in)
446
(16)
496
(18)
8% (4%
)72 (2
)10
0(4
)21
%∗∗
(4%
)−1
3%∗∗
(2%
)
#obs
erva
tions
4746
,14+
18=
32st
ocks
,77+
77=
154
days
a :defi
ned
asth
eN
ethe
rlan
dsdiff
eren
tial(
post
min
uspr
e)m
inus
the
Bel
gium
diff
eren
tial
∗ /∗∗
:sig
nific
anta
ta95/9
9%le
vel(
only
appl
ied
todiff
eren
tials
)
47
Figure 1: Market fragmentation
These graphs illustrate the entry of new markets and the subsequent fragmentation of trading. The topgraph presents the total entrant share of EUR volume in Europe-listed stocks as well as the largest entrant(Chi-X) share and the largest incumbent (London Stock Exchange) share. The bottom graph presents thetotal entrant share of share volume in NYSE-listed stocks as well as a large entrant (BATS) share and theincumbent (NYSE) share. The data sources for these graphs are Barclays Capital Equity Research andFederation of European Securities Exchanges.
Jan-03 Jan-04 Jan-05 Jan-06 Jan-07 Jan-08 Jan-09 Jan-10 Jan-110
10
20
30
40
50
60
70
80
90
%
U.S. trading in NYSE-listed equities
total entrant share
BATS (entrant) share
NYSE (incumbent) share
Jan-08 May-08 Sep-08 Jan-09 May-09 Sep-09 Jan-10 May-10 Sep-10 Jan-110
5
10
15
20
25
30
35
%
European trading in Europe-listed equities
total entrant share
Chi-X (entrant) share
LSE (incumbent) share
48
Figure 2: Structure of the game
This figure illustrates the game between the seller S, the middleman M, and the buyer B. Panel A illustratesthe informational structure. Panel B depicts the various steps in the game.
panel A: information structure
private
value
x y
zh
SB
private
value
common
value
zs
MM
hard
common
value
soft
z
with
prob
λ
panel B: steps in the game
B rejects
S owns
S owns
S posts pS(x)
M bids pM(zh)
S does not post
M bids pM(zh)
M rejects
M posts pM(zh)
B rejects
S owns
M acceptsB accepts
B owns
B rejects
M ownsS accepts
S rejects S owns
S owns
S owns
49
Figure 3: The posting decision for the seller
This figure depicts the value function for S when he posts and for when he does not post a price, U (x) andV (x) respectively.
private value seller (x)
total value seller
U(x)
V (x |σ2
h>> 0)
V (x |σ2
h= 0)
vM
vM
a
50
Figure 4: Middlemen participation and welfare vs. the relative size of hard information
These figures illustrate how middlemen participation (top) and welfare (bottom) change with the relativesize of hard information. The alpha parameter measures how much of the common value variance is hardinformation. The remaining parameters are set to: λ = 1, µ = 2.27, θ = 0.27.
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0alpha, relative size common value innovation
0.48
0.50
0.52
0.54
0.56
0.58
0.60
mid
dle
man p
art
icip
ati
on in t
rades
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0alpha, relative size hard information
0.65
0.70
0.75
0.80
0.85
0.90
0.95
1.00
1.05
welfare no middleman
with middleman
first best
51
Figure 5: Middlemen participation and welfare vs. the size of the informational friction
These figures illustrate how middlemen participation (top) and welfare (bottom) change with the size of theinformational friction. The theta parameter describes how large the common value variance is relative toprivate value variance. The remaining parameters are set to: λ = 1, µ = 2.27, α = 0.42.
0 1 2 3 4 5theta, relative size common value innovation
0.45
0.50
0.55
0.60
0.65
0.70
0.75
0.80
0.85
0.90
mid
dle
man p
art
icip
ati
on in t
rades
0 1 2 3 4 5theta, relative size common value innovation
0.2
0.4
0.6
0.8
1.0
welfare no middleman
with middleman
first best
52
Figure 6: Middlemen participation and welfare vs. the size of the informational friction (2)
This figure illustrates how welfare changes with the size of the informational friction while keeping theinformational asymmetry fixed in the no middleman case. The theta parameter describes how large thecommon value variance is relative to private value variance. The remaining parameters are set to: µ =2.27, α = 0.42.
0 1 2 3 4 5theta, relative size common value innovation
0.2
0.4
0.6
0.8
1.0
welfare
(advers
e s
ele
ctio
n n
o m
iddle
man c
ase
unch
anged)
no middleman
with middleman
first best
53
Figure 7: Middleman inventory
These graphs plot the middleman’s intraday inventory for ING stock on January 30 starting him offwith zeroshares (by assumption). They plot inventory both by market (top two graphs) and aggregated across markets(bottom graph). The middleman is discovered as a combination of an anonymous Chi-X broker ID andan anonymous Euronext broker ID that achieves high-frequency trading and mean-reversion in inventoryacross markets. The graphs illustrate this trading pattern for one stock-day and Table 2 produces middlemanstatistics for the full sample.
-40000
-35000
-30000
-25000
-20000
-15000
-10000
-5000
0
5000
09:0
0:00
10:0
0:00
11:0
0:00
12:0
0:00
13:0
0:00
14:0
0:00
15:0
0:00
16:0
0:00
17:0
0:00
18:0
0:00
units
:sha
res
middleman’s Euronext inventory
-5000
0
5000
10000
15000
20000
25000
30000
35000
40000
09:0
0:00
10:0
0:00
11:0
0:00
12:0
0:00
13:0
0:00
14:0
0:00
15:0
0:00
16:0
0:00
17:0
0:00
18:0
0:00
units
:sha
res
middleman’s Chi-X inventory
-15000
-10000
-5000
0
5000
10000
15000
09:0
0:00
10:0
0:00
11:0
0:00
12:0
0:00
13:0
0:00
14:0
0:00
15:0
0:00
16:0
0:00
17:0
0:00
18:0
0:00
units
:sha
res
middleman’s inventory aggregated across markets
54
Figu
re8:
Scat
ter
plot
ofm
iddl
eman
activ
ityve
rsus
amou
ntof
hard
info
rmat
ion
Thi
sfigu
reco
ntai
nsa
scat
terp
loto
fthe
trad
epa
rtic
ipat
ion
byth
eid
entifi
edm
iddl
eman
(see
Tabl
e2)
agai
nsta
prox
yfo
rthe
rela
tive
impo
rtan
ceof
‘har
d’in
form
atio
nde
fined
asan
ypu
blic
info
rmat
ion
that
can
bepr
oces
sed
bym
achi
nes
(e.g
.,pr
ice
chan
ges
inth
ein
dex
futu
res,
sam
ein
dust
ryst
ocks
,for
eign
exch
ange
rate
).T
heco
njec
ture
isth
atm
iddl
emen
oper
atin
gw
ithfa
stm
achi
nes
have
aned
gew
hen
such
info
rmat
ion
isa
larg
erpa
rtof
tota
linf
orm
atio
n.A
prox
yfo
rthe
rela
tive
size
ofha
rdin
form
atio
nis
the
‘R2 ’o
fare
gres
sion
ofst
ock
retu
rnon
inde
xre
turn
.T
his
prox
yis
cons
truc
ted
fore
ach
stoc
k-da
yin
the
sam
ple
base
don
the
coin
tegr
atio
nm
odel
resu
ltsof
Tabl
e4.
Mid
dlem
antr
ade
part
icip
atio
nis
also
calc
ulat
edfo
rea
chst
ock-
day.
Bot
hva
riab
les
are
dem
eane
dby
stoc
kto
only
focu
son
the
time
vari
atio
n(a
ndno
thav
eth
ere
sults
bedr
iven
byun
obse
rved
hete
roge
neity
acro
ssst
ocks
).T
hesc
atte
rplo
toft
hese
seri
esth
usill
ustr
ates
the
with
inco
rrel
atio
npr
esen
ted
inTa
ble
5.
-10-5 0 5 10
15
20 -5
0-4
0-3
0-2
0-1
0 0
10
20
30
40
50
60
middleman participation in trades, demeaned by stock (%)
rela
tive
size
inde
x co
mpo
nent
in e
ffici
ent p
rice
inno
vatio
n, d
emea
ned
by s
tock
(%
)
55
Figu
re9:
Chi
-Xan
dE
uron
extp
rice
-quo
tein
form
ativ
enes
sin
the
cros
s-se
ctio
n
Thi
sfig
ure
grap
hsC
hi-X
and
Eur
onex
tpri
ce-q
uote
info
rmat
iven
ess
agai
nsta
prox
yfo
rth
ere
lativ
eim
port
ance
of‘h
ard’
info
rmat
ion
defin
edas
any
publ
icin
form
atio
nth
atca
nbe
proc
esse
dby
mac
hine
s(e
.g.,
pric
ech
ange
sin
the
inde
xfu
ture
s,sa
me
indu
stry
stoc
ks,f
orei
gnex
chan
gera
te).
Pric
equ
ote
info
rmat
iven
ess
isde
fined
asth
epr
edic
tive
pow
erof
mid
quot
e(a
vera
geof
the
bid
and
ask
quot
e)pr
ice
chan
ges
for
the
amou
ntof
info
rmat
ion
that
isre
veal
edin
betw
een
trad
es.
The
met
hodo
logy
isba
sed
onlin
ear
proj
ectio
ns(P
e−
Pi)(∆
f)an
da
coin
tegr
atio
nm
odel
(see
Sect
ion
2.4)
.T
heva
riab
les
are
calc
ulat
edas
aver
ages
per
stoc
kan
dth
egr
aphs
repr
esen
tdis
pers
ion
inth
ecr
oss-
sect
ion.
The
size
ofth
edo
tcor
resp
onds
toav
erag
est
ock
volu
me.
0 1 2 3 4 5
15
20
25
30
35
40
45
50
55
60
(bp)
rela
tive
size
of i
ndex
com
pone
nt in
effi
cien
t pric
e in
nova
tion
(%)
Eur
onex
t mid
quot
e pr
ice
upda
te in
form
ativ
enes
sC
hi-X
mid
quot
e pr
ice
upda
te in
form
ativ
enes
s
56
Figure 10: Price quote informativeness: index vs nonindex component
This figure decomposes the bottom graph of Figure 9 into the price-quote informativeness of the index-correlated component of the predictability and an orthogonal component.
0
1
2
3
4
5
15 20 25 30 35 40 45 50 55 60
(bp)
relative size of index component in efficient price innovation (%)
Euronext midquote price update informativeness on index componentChi-X midquote price update informativeness on index component
0
1
2
3
4
5
15 20 25 30 35 40 45 50 55 60
(bp)
relative size of nonindex component in efficient price innovation (%)
Euronext midquote price update informativeness on nonindex componentChi-X midquote price update informativeness on nonindex component
57