THE

MICROSTRUCTURE
APPROACH
TO EXCHANGE
RATES

RICHARD K.

LYONS

The Microstructure
Approach to Exchange
Rates

This Page Intentionally Left Blank

The Microstructure
Approach to Exchange
Rates

Richard K. Lyons

The MIT Press

Cambridge, Massachusetts
London, England

( 2001 Massachusetts Institute of Technology

All rights reserved. No part of this book may be reproduced in any form by any
electronic or mechanical means (including photocopying, recording, or information
storage and retrieval) without permission in writing from the publisher.
This book was set in Palatino in 3B2 by Asco Typesetters, Hong Kong, and was printed
and bound in the United States of America.
Library of Congress Cataloging-in-Publication Data
Lyons, Richard K.
The microstructure approach to exchange rates / Richard K. Lyons.
p. cm.
Includes bibliographical references and index.
ISBN 0-262-12243-X (hc.)
1. Foreign exchange ratesMathematical models. I. Title.
HG3823 .L96 2001
2001044337
332.4 0 5 0 015118dc21

I dedicate this book to my wonderful wife, Jen.

This Page Intentionally Left Blank

Contents

Preface

ix

Overview of the Microstructure Approach

The Economics of Order Flow Information

19

The Institutional Setting

37

Theoretical Frameworks

63

Empirical Frameworks

Exchange Rate Models: Macro to Microfoundations

Macro Puzzles: The Challenge for Microstructure

Microstructure and Central Bank Intervention

Customers: Underlying Demand in the Economy

10 Looking Forward
Notes
279
References
305
Index
325

263

113
151
171

221
243

This Page Intentionally Left Blank

Preface

This book addresses a growing interaction between two elds

exchange rate economics and microstructure nance. Historically,
these elds have progressed independently. More recently, however,
they have begun to interact, and that interaction has stimulated
new perspective within both. In exchange rate economics, that new
perspective has given rise to a distinct approacha microstructure
approach.
A natural audience for this book includes people with expertise in
only one of these two elds who are interested in learning more
about the other. In the past this has not been easythese areas have
different intellectual traditions, so bridging them required a large
investment of time. This book should facilitate that investment. To
that end, I have tried to present material on both elds in an accessible way, so that readers lacking specialization in either one should
nd the material within reach. Practitioners, too, should nd most of
the material accessible (the empirical work in chapters 7 through 9
is most relevant to that audience).
A notable feature of this book is its treatment of both theoretical
and empirical work (c.f. O'Hara's 1995 theory text). Indeed, the book
is organized to highlight their interplay. Chapter 4 is dedicated to
the canonical frameworks in microstructure theory. Some of these
frameworks were originally developed to address the New York
Stock Exchange, a market with a single dealer in each stock. The
circumstances under which these frameworks are appropriate for
foreign exchange markets (FX) is an important topic addressed in
that chapter. Chapter 5 summarizes empirical microstructure frameworks, with emphasis on those employed in FX. These empirical
frameworks draw heavily on chapter four's theory. The last four
chapters, chapters 710, are applications of tools presented in the

Preface

rst six. They bring the theoretical/empirical interplay into sharp

focus.
A second notable feature of this book is its emphasis on information economics rather than institutions. People unfamiliar with
microstructure nance typically believe its focus is institutional (e.g.,
the consequences of different trading mechanisms). This is an important part of the eld. But in terms of this book, it is not the most
important part. The focus of this book is the economics of nancial
information and how microstructure tools clarify the types of information relevant to exchange rates. In keeping with this focus, I move
immediately in chapter 2 to the economics of nancial information,
saving detailed institutional material on FX trading until chapter 3.
This book also has some notable features in terms of pedagogy.
They are concentrated in chapters 46, the three survey chapters:
microstructure theory, microstructure empirical frameworks, and
exchange rate theory, respectively. In chapter 4, I begin the survey of
microstructure models with the ``rational expectations'' model. The
chapter proceeds from that model's implicit auctioneer to the explicit
auctioneer of the Kyle model. Conceptually, I consider this a more
natural progression than the typical sequencing in microstructure
books. To maintain focus on essentials, chapter 4 also presents the
microstructure models in a common format. I close chapter 4 with
an appendix that collects several important technical tools for easy
reference. On the empirical side, chapter 5 opens with a survey of
available data sets and information on accessing them via the Internet (several are available from my web site: hwww.haas.berkeley.
edu/@lyonsi). Data quality has increased substantially in recent
years (in large part due to the shift to electronic trading), and this
survey provides an easy reference for familiarizing oneself. Chapter
6 is the transition from the earlier, more micro-oriented chapters, to
the later, more macro-oriented chapters. That chapter begins its survey of exchange rate theory with a treatment of ``valuation,'' which
includes traditional denitions of exchange rate ``fundamentals.''
This should help readers whose perspective on valuation comes primarily from the dividend-discount models used in equity markets.
Chapter 6 also highlights the parallels between exchange rate models
and microstructure models (such as the conceptual link between
portfolio balance models and inventory models).
This book also contains new work. For example, some of the
material in chapters 7 and 8 is new to the literature, and most of

Preface

xi

the material in chapter 9 is new. Many of the arguments and points

made throughout other chapters are new as well (e.g., perspectives
on the microstructure approach's potential applications). Much of
the material in chapters 7 to 9 comes from work I have done
jointly with three co-authors: Martin Evans, Mintao Fan, and Michael
Moore.
For teaching, there are three types of courses for which this book
would be appropriate. The rst is a Ph.D. course in international
macroeconomics. Of the ten chapters, the best ones for this type of
course are chapters 1, 4, 6, 7, 8, and 9. A second course for which the
book would be appropriate is a Ph.D. course in market microstructure. The best chapters for this type of course are 1, 2, 3, 4, 5,
and 10. A third course for which the book would be appropriate is a
Master's course in international nance. The best chapters for a
course at this level include 1, 3, 6, 7, and 9.
Some caveats are in order. Like microstructure theory in general,
my treatment of the microstructure approach emphasizes the role
of order ow (signed transaction quantities) in price determination.
This emphasis aligns closely with my own interests and my own
work. Though much of the material I present is not my own, my
work plays a more prominent role than it would in a balanced survey. To borrow the words of one reviewer, this book is more of a
personal synthesis than a professional consensus. I apologize in advance to those whose work is underrepresented. (For those interested in a survey, let me recommend Sarno and Taylor 2000a.)
A second caveat is my use of the term ``microstructure approach.''
Though there is no consensus denition of what constitutes a
microstructure model, it is safe to say that the denition I adopt
is rather broad. For me, the microstructure approach is not just a rich
set of tools for addressing the issues, but also a way of framing
those issues. Indeed, the framing per se is an important aspect of the
approach's value. If I have done my job, the chapters that follow
should make this point increasingly clear.
To these previous caveats I must add another. The tone of parts
of this book may to some readers seem a bit missionary. This is particularly true of chapter 1, where I address the book's overarching
themes. (Exchange rate economists will nd these themes more provocative than will people in microstructure nance.) At the risk of
appearing to over-sell, I felt it best to address these themes early.
Most of the evidence that supports these themes appears in later
chapters.

xii

Preface

I owe thanks to many people. The support and encouragement of

Victoria Richardson and Elizabeth Murry at MIT Press are much
appreciated. I also want to thank several of my colleagues who, as
anonymous reviewers of early drafts, provided valuable comments.
I am indebted to them. I am also indebted to Alain Chaboud,
Mike Dennis, Petra Geraats, Philipp Hartmann, Harold Hau, Jen
Lyons, Michael Melvin, Michael Moore, Carol Osler, Richard Portes,
Helene Rey, Dagnn Rime, Andrew Rose, Patrik Safvenblad, Sergio
Schmukler, Mark Taranto, and Paolo Vitale for extensive comments
and suggestions. A friend and colleague, Michael Klein, planted the
seed for this book and also provided many insights along the
way. I thank the University of Toulouse and Stockholm University
for providing generous visiting opportunities while I was writing.
Financial support from the National Science Foundation is gratefully
acknowledged; this book (and much of the research underlying it) is
in large part the product of their investment. Finally, I'd like to thank
Jim McCarthy, a long-time buddy and FX trader who introduced me
to the world of FX trading.

Overview of the
Microstructure Approach

Ten years ago, a friend of mine who trades spot foreign exchange for
a large bank invited me to spend a few days at his side. At the time, I
considered myself an expert, having written my thesis on exchange
rates. I thought I had a handle on how it worked. I thought wrong.
As I sat there, my friend traded furiously, all day long, racking up
over $1 billion in trades each day (USD). This was a world where the
standard trade was $10 million, and a $1 million trade was a ``skinny
one.'' Despite my belief that exchange rates depend on macroeconomics, only rarely was news of this type his primary concern. Most
of the time my friend was reading tea leaves that were, at least to me,
not so clear. The pace was furiousa quote every ve or ten seconds, a trade every minute or two, and continual decisions about
what position to hold. Needless to say, there was little time for chat.
It was clear my understanding was incomplete when he looked
over in the midst of his fury and asked me, ``What should I do?'' I
laughed. Nervously.
This book is an outgrowth of my subsequent interest in this area. It
is principally concerned with the gap between what I knew before I
sat down with my friend and what I saw when I got there. In effect,
this gap is the space between two elds of scholarship: exchange rate
economics on the one hand, and microstructure nance on the other.
Exchange rate economists use models of exchange rate determination that are macroeconomic (i.e., rates are determined as a function of macro variables such as ination, output, interest rates, etc.).
These same exchange rate economists are largely unfamiliar with
microstructure models. Most microstructure scholars, in contrast,
view foreign exchange as the purview of international economists,
and are unfamiliar with macroeconomic exchange rate models. Their

Chapter 1

traditional focus is the microeconomics of equity markets, particularly the New York Stock Exchange.
Though this book has several objectives, two deserve mention at
the outset. The rst is to lower entry barriers faced by scholars
interested in this burgeoning area. Lowering barriers on both sides
exchange rate economics and microstructure nancewill help this
research domain realize its potential. A second objective is to channel
past work into a more unied approacha microstructure approach.
In the 1990s, many authors applied microstructure tools to foreign
exchange (FX) markets, but existing work is still largely fragmented.
Does exchange rate economics need a new approach? Yes. It is
in crisis. It is in crisis in the sense that current macroeconomic approaches to exchange rates are empirical failures. In their recent survey in the Handbook of International Economics, Jeffrey Frankel and
Andrew Rose (1995, 1709) put it this way: ``To repeat a central fact
of life, there is remarkably little evidence that macroeconomic variables have consistent strong effects on oating exchange rates, except
during extraordinary circumstances such as hyperinations. Such
negative ndings have led the profession to a certain degree of pessimism vis-a-vis exchange rate research.''
In the end, it is my hope that this book might rouse a little
optimism.
1.1

Three Approaches to FX: Goods, Assets, and Microstructure

Before the 1970s, the dominant approach to exchange rate determination was the goods market approach. According to this approach,
demand for currencies comes primarily from purchases and sales of
goods. For example, an increase in exports increases foreign demand
for domestic currency to pay for those exported goods. In this simple form, the implication is rather intuitive: countries with trade
surpluses experience appreciation (which comes from the currency
demand created by the surplus). Despite the approach's intuitive
appeal, however, it fails miserably when one looks at the data: trade
balances are virtually uncorrelated with exchange rate movements in
major-currency FX markets. This negative result is perhaps not surprising given that trade in goods and services accounts for only a
small fraction of currency tradingless than 5 percent of the average
$1.5 trillion of FX traded daily.

Overview of the Microstructure Approach

In the 1970s, the asset market approach emerged. It built on the

earlier approach by recognizing that currency demand comes not
only from purchases and sales of goods, but also from purchases and
sales of assets. For example, in order to purchase a Japanese government bond, a U.S. investor rst purchases the necessary yen. In
addition, the investor's dollar return will depend on movements in
the yen, so his demand for the bond depends in part on his desire to
speculate on those currency movements. This shift in perspective
brought a shift in modeling strategy. Models began to conform to the
notion of speculative ``efciency'': exchange rates were modeled as
efcient in that they incorporate all publicly available information,
making public information useless for producing excess returns. This
is a feature the goods market approach did not share.1
Disconcertingly, empirical work does not support the asset market
approach either. The macroeconomic variables that underlie the
approach do not move exchange rates as predicted. The classic
reference is Meese and Rogoff 1983a; they show that asset approach
models fail to explain major-currency exchange rates better than a
simple ``no change'' model. Thus, asset approach models are not even
consistently getting the direction right. In his later survey, Meese
(1990) summarizes by writing that ``the proportion of (monthly or
quarterly) exchange rate changes that current models can explain is
essentially zero.'' (The literature documenting this poor empirical
performance is vast; for surveys see Frankel and Rose 1995; Isard
1995; and Taylor 1995.)
The FX market's enormous trading volume is also problematic for
the asset approach. Explaining volume is difcult because actual
transactions are awarded no role in mapping macroeconomic variables into exchange rate behavior. Rather, because all macroeconomic
news is publicly available, when news occurs, the exchange rate is
presumed to jump to the new consensus level; the change in expectations that causes the jump does not require any trading. Differing
beliefs is not a driver of trading under this approach either, because
the approach assumes homogeneous beliefs.
These negative observations do not imply that the asset market
approach is ``wrong''; indeed, most agree that it is, in broad terms,
appropriate. Rather, it appears the approach is missing some key
featuresfeatures that matter for how exchange rates are actually
determined.

Chapter 1

This book presents a new approach to exchange rates, the microstructure approach. Under this approach, like the asset market
approach, the demand for currencies comes from purchases and
sales of assets. In this sense these approaches are complementary,
not competing. What distinguishes the microstructure approach is
that it relaxes three of the asset approach's most uncomfortable
assumptions:2
1. Information: microstructure models recognize that some information relevant to exchange rates is not publicly available.
2. Players: microstructure models recognize that market participants
differ in ways that affect prices.
3. Institutions: microstructure models recognize that trading mechanisms differ in ways that affect prices.
People unfamiliar with microstructure believe its focus is on the
third of thesethe consequences of different trading mechanisms.
The focus of this book is resolutely on the rstthe information economics.
(In keeping with this focus, the next chapter moves immediately to
the economics of nancial information; material on trader heterogeneity and trading mechanismspoints 2 and 3is in chapter 3.)3
Empirically, it is simply not true that all information used to determine market-clearing exchange rates is publicly available. We can
analyze the consequences of thistheoretically and empirically
using tools within the microstructure approach. The resulting analysis shows that the public-information assumption is not a good one:
it misses much of exchange rate determination.
Consider some examples that suggest that the microstructure
approach is on target with respect to these three assumptions. Regarding non-public information, FX traders at banks regularly see
trades that are not publicly observable. As I show in later chapters,
this information forecasts subsequent exchange rates (e.g., seeing
the demands of private participants or central banks before the rest
of the market). Regarding differences across market participants,
traders with common information regularly interpret it differently.
Another example of differences across participants is motives for
trade: some traders are primarily hedgers, whereas others are primarily speculators (and even among the latter, speculative horizons
can differ dramatically). Regarding trading mechanisms that affect
prices, consider a market where transparency is low (e.g., where in-

Overview of the Microstructure Approach

dividual transaction sizes and prices are not generally observable).

Low transparency can slow the updating of beliefs about appropriate
prices, thereby altering the path of realized prices.
From these examples one might describe the microstructure
approach as taking an in-the-trenches, trading-room perspective.
Given that exchange rates are actually determined in the trading
room, this would seem a reasonable starting point. But can relaxing
the three assumptions above help us understand exchange rates?
Relaxing the corresponding assumptions for other asset classes has
certainly deepened our understanding of these other markets. The
nal judgment, though, will be based on specic applications of
microstructure tools. The latter half of this book presents a number
of applications.
In advance of those applications, I urge the reader to bear in mind
the overarching fact that traditional approaches are not consistent
with the data. Indeed, this fact induces Flood and Taylor (1996, 286)
to conclude that
Given the exhaustive interrogation of the macro fundamentals in this respect
over the last twenty years, it would seem that our understanding of the
short-run behavior of exchange rates is unlikely to be further enhanced by
further examination of the macro fundamentals. And it is in this context that
new work on the microstructure of the foreign exchange market seems both
warranted and promising.

1.2

Hallmarks of the Microstructure Approach

The previous section introduces microstructure in the context of

exchange rates but does not dene the term as used in domestic
nance. Maureen O'Hara (1995, 1) denes market microstructure
as ``the process and outcomes of exchanging assets under explicit
trading rules.'' The denition I adopt here is consistent with hers.
Because her denition is so broad, though, it may be helpful to clarify further.
When one moves from a macro approach to a micro approach, two
variables that play no role in the macro approach take center stage.
These variables are hallmarks of the micro approach, and as hallmarks, they help to dene microstructure. These variables are
1. order ow, and
2. spreads (bid-ask).

Chapter 1

If I labeled these ``quantity'' and ``price,'' it would be clear that they

are the old mainsprings of economics after all. These labels are a
bit facile, though. Describing them as quantity and price viewed
through a magnifying glass is nearer the truth. Let me clarify by
touching on each.
Order Flow
Understanding order ow is essential for appreciating how the
microstructure approach to exchange rates departs from earlier approaches. First, it is important to recognize that transaction volume
and order ow are not the same. Order ow is transaction volume
that is signed. For example, if you approach a dealer (marketmaker)
and you decide to sell the dealer 10 units (shares, euros, etc.), then
transaction volume is 10, but order ow is 10: You as the initiator
of this transaction are on the sell side, so order ow takes on a negative sign. The quoting dealer is on the passive side of the trade. The
trade is signed according to the active, or initiating side. Over time,
order ow can then be measured as the sum of the signed buyerinitiated and seller-initiated orders. A negative sum means net selling pressure over the period. Thus, despite the immutable fact that
all trades involve a buyer and a seller, microstructure theory provides a rigorous way of attaching a sign to individual transactions
when measuring order ow.
This denition needs to be adjusted slightly for markets that do
not have dealers. Some nancial markets replace dealers with something known as a ``limit order book.''4 Here is an example of a limit
order: ``Buy 10 units for me when the market reaches a price of 50.''
Limit orders are collected together in an electronic ``book.'' The most
competitive orders in the book dene the best available bid and offer
prices. For example, the limit order to buy with the highest buying
price becomes the best bid in the market. If you entered the market
and wanted to sell 10 units immediately, you could sell at this best
bid price, but no higher. (Think of these best limit orders as analogous to dealer bid and offer quotes in markets that have dealers.) The
limit orders are the passive side of any transaction, just as the quoting dealer is always on the passive side in the previous example.
When orders arrive that require immediate execution (e.g., an order
to ``sell 10 units now at the best available price), these orderscalled
market ordersgenerate the signed order ow. In the example

Overview of the Microstructure Approach

above, as in the earlier case, executing the market order to sell 10

units produces an order ow of 10.
Order ow, as used in microstructure nance, is a variant of a key
term in economics, ``excess demand.'' It is a variant of excess demand
rather than a synonym for two reasons, the rst relating to the excess
part and the second relating to the demand part. For the former, note
that excess demand equals zero in equilibrium by denitionthere
are two sides to every transaction. This is not true of order ow: in
markets organized like foreign exchange, orders are initiated against
a marketmaker, who if properly compensated, stands ready to absorb imbalances between buyers and sellers. These ``uninitiated''
trades of the marketmaker drive a wedge between the two concepts,
excess demand and order ow. The second reason the concepts differ
is that order ow is in fact distinct from demand itself. Order ow
measures actual transactions, whereas demand shifts need not induce transactions. For example, the demand shifts that move price in
traditional exchange rate models (e.g., the monetary models reviewed
in chapter 6) are caused by the ow of public information, which
moves exchange rates without transactions ever needing to occur.
A distinctive feature of microstructure models, across the board, is
the central role order ow plays. This across the board property
deserves emphasis because it expands the applicability of microstructure enormously. Recall that order ow plays no role in the
asset approach; in the asset approach, order ow is not a variable
that helps us understand exchange rates. That microstructure models
of all types tell us this variable is important expands microstructure
from the narrow concept of ``institutional structures with price
effects'' to the broader concept of ``a new lens for viewing markets.''
It instructs us that order ow deserves our attention. The question
of order ow's importance in FX is distinct fromand in my judgment much larger thanthe question of how specic FX institutions
affect prices.
Consider a simple diagram that illustrates an important feature of
microstructure models that relates directly to order ow. The diagram shows that information processing has two stages. The rst
stage is the analysis or observation of fundamentals by nondealer
market participants (mutual funds, hedge funds, individuals with
special information, etc.). The second stage is the dealer'sthe price
setter'sinterpretation of the rst-stage analysis. The dealer's interpretation comes from reading the order ow. Dealers set price on the
basis of this interpretation.5

Chapter 1

Figure 1.1
Two stages of information processing.

Order ow conveys information about fundamentals because it

contains the trades of those who analyze fundamentals. In this sense,
it is a transmission mechanism. Naturally, though, these informative
trades may be mixed with uninformative trades, making the task of
``vote counting'' rather complex. In standard microstructure models,
the dealer learns nothing about fundamentals that he or she does not
learn from order ow. As a practical matter, this is clearly too strong.
The dealer's dependence on learning from order ow arises in these
models because the information being learned is not publicly known.
When information is publicly known, dealers do not need to learn
from order ow. In practice, though some information relevant to FX
is publicly known, some is not, so learning from order ow can be
important.
Spreads
Spreadsthe second hallmark variable of the micro approach
receive a lot of attention within the eld of microstructure. There are
three reasons for this, one scientic, one practical, and one historical.
The scientic reason relates to data: spread data are a core element of
most data sets, and as such, are a ready target for testable hypotheses. This stands in contrast to other features of the trading environment that are important but not as readily measurable (such as
information ow, belief dispersion, and liquidity-motivated order
ow). The second reason spreads receive so much attention is practical. Practitioners are intensely concerned with managing trading
costs. This concern, and the resources devoted to it, has naturally
inuenced the course of research within microstructure. The third
reason spreads receive so much attention is historical. From the
beginning, the eld of microstructure sought to separate itself
from the literature on trading under rational expectations. Rational

Overview of the Microstructure Approach

expectations models abstract from trading mechanisms completely,

the premise being that trading mechanisms have little effect on the
relationship between underlying fundamentals and price. A natural
means of distinguishing microstructure research was to orient it
toward the following question: How does altering the trading mechanism alter price? This orientation led to a focus on the determination
of real-world transaction pricesspreads. 6
Though spreads receive a lot of attention, the subeld of spread
determination is but one branch of the broader eld of microstructure. Many microstructure models, for example, do not even include
a spread (such as Kyle models, presented in chapter 4). I raise the
issue early because for some people the association between spreads
and microstructure is very tight. That many microstructure models
have no spread should loosen this association.
Of the two hallmarksorder ow and spreadsthis book focuses
much more on order ow. In those instances where I do address
spreads, my focus is on their information-theoretic implications, which
can be substantial. For example, in chapter 2, I present evidence
gleaned from spread behavior alonethat implies that order ow
forecasts price movements. This is important because it violates the
premise of traditional models that all FX market participants are
equally well informed: Some participants are better informed because they observe more order ow. Thus, on the basis of spread
behavior alone, one reaches a rather profound conclusion: Contrary
to the asset approach, exchange rate determination is not wholly a
function of public news.7
1.3

Overarching Themes

Readers familiar with exchange rate economics are unlikely to be

familiar with microstructure. For them, it may be helpful at the outset to address some of this book's overarching themes. Introducing
them early provides valuable advance perspective on applications in
later chapters. There are four themes in particular that exchange rate
economists may nd provocative.
Theme 1:

Order Flow Matters for Exchange Rate Determination

Let me offer some perspective on this assertion, as it is one I will

revisit. Consider an example, one that claries how economist and

10

Chapter 1

practitioner worldviews differ. The example is the timeworn reasoning used by practitioners to account for price movements. In the case
of a price increase, practitioners will assert that ``there were more
buyers than sellers.'' Like other economists, I smile when I hear this. I
smile because in my mind the expression is tantamount to ``price had
to rise to balance demand and supply.'' These phrases may not be
equivalent, however. For economists, the phrase ``price had to rise to
balance demand and supply'' calls to mind the Walrasian auctioneer,
which is an abstract way to think about how price adjusts to a
market-clearing level. The Walrasian auctioneer collects ``preliminary'' orders, which he uses to nd the market-clearing price. All actual
trades occur at this priceno trading occurs in the process of nding
it. (Readers familiar with the rational expectations model of trading
will recognize that in that model, this property is manifested by all
orders being conditioned on a market-clearing price.8)
Practitioners seem to have a different model in mind. In the practitioners' model there is a dealer instead of an abstract auctioneer.
The dealer acts as a buffer between buyers and sellers. The orders the
dealer collects are actual orders, rather than preliminary orders, so
trading does occur in the transition to the new price.9 Crucially, the
dealer then determines new prices from information about demand
and supply that is embedded in the order ow (as suggested in the
``two-stage processing'' diagram above).
Can the practitioner model be rationalized? Not at rst blush, because it appears that trades are occurring at disequilibrium prices
(prices at which the Walrasian auctioneer would not allow trading).
This suggests irrational behavior. But this interpretation misses an
important piece of the puzzle: Whether these trades are out of equilibrium depends on the information available to the dealer. If the
dealer knows at the outset that there are more buyers than sellers,
eventually pushing price up, then it is unclear why that dealer would
sell at a low interim price. If the buyer/seller imbalance is not
known, however, then rational trades can occur through the transition. (Put differently, in setting prices the dealer cannot condition on
all the information available to the Walrasian auctioneer.) This is
precisely the story developed in standard microstructure models.
Trading that would be irrational if the dealer knew as much as
the Walrasian auctioneer can be rationalized in models with more
limitedand more realisticconditioning information.

Overview of the Microstructure Approach

Theme 2:

11

Microstructure Implications Can Be Long-Lived

It is common to associate ``microstructure'' with ``high frequency.''

The association is natural, but deceptive. It is true that empirical
work in microstructure is in general high frequency, but this does not
imply that microstructure tools are irrelevant to lower frequency,
resource relevant phenomena. Indeed, there are ample tools within
the microstructure approach for addressing lower frequency phenomena, and new tools continue to emerge, thanks in part to recognition within the broader microstructure literature that resource
allocation warrants greater attention. The later chapters of this book
are dedicated to examples of lower frequency relevance, particularly
chapter 7.
Regarding long-lived effects, the most important point to recognize is that when order ow conveys information, its effect on price
should be long-lived. Indeed, a common empirical assumption for
distinguishing information from pricing errors is that information's
effects on price are permanent, whereas pricing errors are transitory
(French and Roll 1986; Hasbrouck 1991a; chapter 2 provides details).
The data in equity markets, bond markets, and FX markets bear out
these long-lived effects. In FX, for example, Evans (1997), Evans and
Lyons (1999), Payne (1999), and Rime (2000) show that order ow
has signicant, persisting effects on exchange rates. Indeed, statistically these effects appear to be permanent. Among microstructure's
long-lived implications, this ``information'' channel is denitely the
most fundamental.
Let me touch on another source of lower frequency relevance:
multiple equilibria that depend on microstructure parameters. Certain parameters' values can determine whether multiple equilibria
within a given model are possible, and if so, which equilibrium is
more likely to be selected (e.g., Portes and Rey 1998; Hau 1998;
Hartmann 1998a; Rey 2001). These different equilibria apply in some
models to the exchange rate's level, and in other models to the exchange rate's volatility (multiple volatility states are the focus of
Jeanne and Rose 1999 and Spiegel 1998). Either way, multiple equilibria that depend on microstructure parameters open another door
through which price effects are long-lived (long-lived because a
given equilibrium is by nature persistent).10
An analogy may be helpful. Microstructure can speak to longer
horizon exchange rates in much the same way that microscopes

12

Chapter 1

speak to pathologies with macro impact. In medicine, microscopes

provide resolution at the appropriate levelthe level at which the
phenomenon emerges. This is true irrespective of whether the phenomenon also has macro impact. Resolution at this level is the key to
our understanding. Similarly, microstructure tools provide resolution at the level where its ``phenomenon'' emergesthe level where
price is determined. What information do dealers have available
to them, and what are the forces that inuence their decisions?
(Whether we like it or not, it is a stubborn fact that in the major currency markets, there is no exchange rate other than the price these people
set.) Answering these questions does indeed help explain exchange
rates over longer horizons. I provide evidence of this in section 7.1
and elsewhere.
Theme 3:

Microstructure Is Relevant to Macroeconomists

In 1987, stock markets crashed around the world, an event that

most people consider macro relevant. What analytical tools did the
profession use to address the crash? The tools were microstructure
tools (see, e.g., Grossman 1988; Gennotte and Leland 1990; Jacklin,
Kleidon, and Peiderer 1992; Romer 1993). These leading papers on
the crash are explicit about relaxing all three of the asset approach
assumptions noted above. In particular, the richness of these models
comes from (1) information structure: which participants knew what;
(2) heterogeneity: what types of participant were active and what
were their motives for trading; and (3) institutions: what role did
each participant play in the trading process and what trading information did each have available. The microstructure approach certainly bore fruit in this case.
Macroeconomists have also applied microstructure tools to understand exchange rate collapses during the 1990s nancial crises in
Asia. These papers also introduce information structures, heterogeneity, and institutional factors that are not in general present within
the traditional macro approach (see Chen 1997; Calvo 1999; Corsetti,
Morris, and Shin 1999; Carrera 1999).
Theme 4: Exchange Rate Economics Merits an InformationTheoretic Perspective
In many ways, this theme follows from the rst three. The information economics of traditional exchange rate models is rather simple.

Overview of the Microstructure Approach

13

Macroeconomic news is announced publicly, so everybody learns

new information at the same time. This news can then be impounded
directly in price. The aggregation of dispersed bits of information
that are not publicly known is presumed to play no role. The question is, of course, whether this captures the essence of price determination, or whether it neglects something important. I address this
question extensively in the next chapter.
1.4

Applying Microstructure Tools to Exchange Rate Puzzles

I turn now to puzzles within exchange rate economics and how

microstructure helps to resolve them. Though later chapters address
this in detail, let me offer some initial thoughts. Consider rst the
puzzle of the FX market's enormous trading volume ($1.5 trillion per
day, by far the largest of any nancial market). In fact, the microstructure approach has made considerable progress on this puzzle.
I have in mind here recent ndings on the volume-amplication
effects of the so-called hot potato. Hot potato trading is the passing
of unwanted positions from dealer to dealer following an initial
customer trade. In the words of Burnham (1991, 135): ``[When hit
with an incoming order, a currency dealer] seeks to restore his own
equilibrium by going to another marketmaker or the broker market
for a two-way price. A game of `hot potato' has begun. . . . It is this
search process for a counterparty who is willing to accept a new
currency position that accounts for a good deal of the volume in the
foreign exchange market.'' Thus, the passing of unwanted positions
is a consequence of dealer risk management. Under the asset approach, in contrast, volume is attributed to speculation.
Understanding the causes of high volume is not as important as
understanding price determination, but it is still important. Three
reasons come immediately to mind. First, misunderstanding the
causes of high volume can lead to bad policy. Consider the issue of
transaction taxesan issue that has attracted much attention among
exchange rate economists. Proponents of levying transaction taxes
tend to associate high volume with excessive speculation. If instead
much of this volume reects dealer risk management, then a transaction tax wouldunintentionallyimpede that risk management.
Second, high volume can impede order ow's information role. As
suggested above and detailed in the next chapter, order ow conveys
information. The precision of that information is a function of the

14

Chapter 1

underlying causes of order ow. It is important to understand

whether those causes contribute to, or detract from, informational
precision. Third, misunderstanding the causes of order ow can lead
to bad theory. The mere existence of the volume puzzle is an indication that the asset approach is not addressing key features of the FX
market. The features it is missing may not be important, but to be
condent that they are unimportant requires a tremendous leap of
faith.
What about the big puzzles in exchange rate economics? Chapter 7
addresses this question directly. The three biggest puzzles are11
1. the determination puzzle: Exchange rate movements are virtually
unrelated to macroeconomic fundamentals (at least over periods of
less than about two years);
2. the excess volatility puzzle: Exchange rates are excessively volatile relative to our best measures of fundamentals; and
3. the forward bias puzzle: Excess returns to speculating in foreign
exchange are predictable and inexplicable.
The microstructure approach links these puzzles to one another
through expectations formationthat is, how market participants
form their expectations of future fundamentals. It makes this link
without departing from the asset approach convention of rational
expectations. Rather, the microstructure approach grounds expectations formation more directly in a richer, information-economic
setting. The focus is on information types (such as public versus private) and how information maps into expectations (e.g., whether the
aggregation of order ow ``votes'' is efcient).12 The issues of information type and mapping to expectations are precisely where
microstructure tools provide resolving power.
Chapter 7 addresses the three big puzzles and shows that the
microstructure approach has already made empirical progress. Section 7.1 addresses the rst puzzle; it reviews the work of Evans
and Lyons (1999), who nd that exchange rate movements can be
explainedthey are largely a function of order ow. Section 7.3
addresses the excess volatility puzzle by focusing on recent work on
the sources of FX volatility (e.g., Killeen, Lyons, and Moore 2000a;
Hau 1998; Jeanne and Rose 1999). Section 7.4 offers a microstructurebased explanation of the third puzzleforward discount bias. It
would be wrong for me to suggest that these three big puzzles have

Overview of the Microstructure Approach

15

been put to rest by the above-mentioned work, but progress is being

made.13
1.5

Spanning the Micro-Macro Divide

As noted above, a core distinction between the microstructure

approach and the asset approach is the information role of trades.
Under the asset approach, trades play no role (macroeconomic information is publicly announced), whereas in microstructure models
trades are the driving force. It is instructive to frame this distinction
with a bird's-eye view of structural models used by empiricists
within these two approaches.
Structural Models: Asset Approach
Equations of exchange rate determination within the asset approach
take the form:
DPt f i; m; z et ;

1:1

where DPt is the change in the nominal exchange rate over the
period, typically one month. The driving variables in the function
f i; m; z include current and past values of home and foreign nominal interest rates i, money supply m, and other macro determinants,
denoted here by z.14 Changes in these public-information variables
are presumed to drive price without any role for order ow (though
there is of course a role for demand; recall the distinction between
order ow and demand in section 1.2). If any price effects from
order ow should arise, they would be subsumed in the residual et .
Though logically coherent and intuitively appealing, a long literature
documents that these macro determinants account for only a small
portion (less than 10 percent) of the variation in oating exchange
rates (see the surveys by Frankel and Rose 1995; Isard 1995; Taylor
1995).
Structural Models: Microstructure Approach
Within the microstructure approach, equations of exchange rate
determination are derived from the optimization problem faced by
the actual price setters (the dealers).15 These models are variations on

16

Chapter 1

the following specication:

DPt gX; I; Z et ;

1:2

where now DPt is the change in the nominal exchange rate between
two transactions, versus the monthly frequency of the macro model
in equation (1.1). The driving variables in the function gX; I; Z include order ow X (signed so as to indicate direction), a measure of
dealer net positions (or inventory) I, and other micro determinants,
denoted by Z. It is interesting to note that the residual in this case is
the mirror image of the residual in equation (1.1) in that it subsumes
any price changes due to the public-information variables of the
asset approach.
The key to spanning the micro-macro divide is the role of signed
order ow X. Microstructure models predict a positive relation between signed order ow and price because order ow communicates
nonpublic information, which can then be impounded in price. Empirical estimates of this relation between DP and X are uniformly
positive and signicant in securities markets generally (including
stocks, bonds, and foreign exchange). It is noteworthy that these
empirical estimates have been possible only a relatively short time:
the switch to electronic trading means that we now have detailed
records of order ows. What used to be a black box is no longer.
A Hybrid Approach
To establish the link between the micro and macro approaches, I
investigate in chapter 7 equations with components from both
approaches:
DPt f i; m; z gX; I; Z et :

1:3

These equations are estimable at frequencies corresponding to the

asset approach through the use of time-aggregated measures of
order ow X. The time-aggregated measures of X span much longer
periods than those typically employed in empirical microstructure.
Estimates of this equation show that time-aggregated order ow
has much more explanatory power than macro variables. In fact,
chapter 7 shows that regressing daily changes in log DM/$ rates on
daily order ow produces an R 2 statistic greater than 60 percent.16
Figure 1.2 below provides a convenient summary of this explanatory
power. The solid lines represent the spot rates of the DM and yen

Overview of the Microstructure Approach

17

Figure 1.2
Four months of exchange rates (solid) and order ow (dashed). May 1August 31,
1996.

against the dollar over the four-month sample of the Evans (1997)
dataset (described in chapter 5). The dashed lines represent cumulative order ow for the respective currencies over the same period.
Order ow, denoted by X, is the sum of signed trades over the sample period between foreign exchange dealers worldwide.17 Cumulative order ow and nominal exchange rate levels are strongly
positively correlated (price increases with buying pressure). This result is intriguing. Order ow appears to matter for exchange rate
determination, and the effect appears to be persistent (otherwise the
exchange rate's level would reect only concurrent or very recent
order ow and not cumulative order ow). This persistence is an important property, one that I examine more closely in later chapters.
For order ow to be helpful in resolving big exchange rate puzzles,
its effects have to persist over horizons that match those puzzles
(monthly, at a minimum).18
That order ow matters for exchange rate determination does not
imply that order ow is the underlying cause of exchange rate
movements. Order ow is a proximate cause. The underlying cause
is information. How, specically, can one identify the information
that determines order ow? The notion of order ow as an intermediate link between information and price suggests several strategies
for answering this question, which I touch on now and address further in later chapters (particularly in chapters 7 and 9; readers may
nd a quick look at gure 7.1 helpful at this juncture).

18

Chapter 1

One strategy for linking order ow to underlying determinants

starts by decomposing order ow. (That it can be decomposed is one
of its nice properties.) In chapter 9, I test whether all parts of the aggregate order ow have the same price impact. They do not: the
price impact of FX orders from nancial institutions (e.g., mutual
funds and hedge funds) is signicantly higher than the price impact
of orders from nonnancial corporations. This suggests that order
ow is not just undifferentiated demand. Rather, the orders of some
participants are more informative than the orders of others. Analyzing order ow's parts gives us a view of this market's underlying
information structure.
A second strategy for linking order ow to underlying determinants is based on the idea that order ow measures individuals'
changing expectations. As a measure of expectations, it reects a
willingness to back one's beliefs with moneythe backed-by-money
expectational votes, if you will. This strategy corresponds to the following variation on equation (1.3):
e
et ;
DPt f zt ; zt1

1:4

where zt denotes current macro fundamentals (interest rates, money

e
denotes expected future macro fundamentals.
supplies, etc.) and zt1
These expected fundamentals are not well captured by macroeconometric techniques: estimates are slow moving and imprecise. If
order ow is serving as an expectation proxy, then it should forecast
surprises in important macroeconomic variables (like interest rates).
New order ow data sets that cover up to six years of FX trading
provide enough statistical power to test this.

The Economics of Order

Flow Information

As noted in chapter 1, a focus of this book is the economics of order

ow information. The objective of this chapter is to provide a foundation for that focus. How do we know order ow conveys information? What types of information does order ow convey? I
answer both of these questions here.
In the section 2.1, I provide background on the information economics of traditional asset approach models. In section 2.2, I review
recent empirical evidence on order ow information. The evidence
supporting an information role for order ow comes from many
sources, here documented using several different methodologies. In
section 2.3, I provide a working denition of the term private information; if order ow is going to convey information that is not public, we need a clear sense of what that information is. The concluding
section provides a more theoretical view of the types of information
that order ow conveys. Specic examples that are grouped by type
in that section include information about differential interpretation of
macroeconomic announcements (ination, money supply, real output, etc.), about changing institutional risk tolerance, about hedging
demands and about portfolio rebalancing, among others. Although
these types share certain characteristics, they also have important
differences, which have implications for theory and empirical work.
This chapter does not extend to the mechanics of how order ow is
impounded in price. I address that topic in chapter 4, which presents
the key theoretical frameworks. It is only in the context of the theory
that the details become meaningful. The goal of this chapter is more
modest: to establish a comfort level with the idea that order ow is a
transmission mechanism, a means of transmitting information to
price.

20

2.1

Chapter 2

Background

To many macroeconomists, the idea that order ow conveys incremental information relevant to exchange rates is controversial. In
traditional models, macroeconomic news is announced publicly, and
can therefore be impounded in price directly, without any role for
order ow.1 In light of this common belief, let me provide a bit more
background before examining evidence that order ow does indeed
play an information role.
As noted in chapter 1, models of exchange rate determination
within the asset approach take the form:
DPt f i; m; z et ;
where DPt is the change in the nominal exchange rate over the
period, typically one month. The driving variables in the function
f i; m; z include current and past values of home and foreign nominal interest rates i, money supplies m, and other macro determinants
(e.g., home and foreign real output), denoted here by z.
Now consider the possibility of private information. It is difcult
to imagine circumstances in which agents would have private information about interest rates. Perhaps somebody had an enlightening
conversation with the chairman of the Federal Reserve Board (Fed)
that morning. Doubtful. Maybe somebody has inside information
about the next Fed vote regarding monetary policy. Again, doubtful.2 The natural presumption is that all agents have the same information about currentand futureinterest rates. How about
private information on money supplies, or real output? For these
variables, too, it is natural to presume that agents are symmetrically informed. When money supply or real-output data are publicly
announced, all agents learn new information at the same time. This is
not a recipe for speculative activity based on information advantage.
With a slight shift in perspective, however, an information role for
order ow emerges. This shift in perspective is perfectly consistent
with the ideas of the last paragraph: even if exchange rate determination is based wholly on public information, this is not sufcient
to rule out an information role for order ow. To understand why,
recognize that there are in fact two crucial assumptions in these
macro-asset models that disconnect order ow from price. These two
assumptions are as follows:

The Economics of Order Flow Information

21

1. all information relevant for exchange rates is publicly known;

and
2. the mapping from that information to prices is also publicly
known.
If either assumption is relaxed, then order ow can convey information that is relevant for prices. As an example of relaxing the second,
suppose you and I are FX traders for large banks and we have both
seen the same macro announcement. If I do not know how you will
interpret the announcement's price implications, then I need to
watch your trade to learn about your interpretation. Your trade
order owwill both inuence price and teach me something.
As an empirical matter, relaxing the assumption that everybody
knows the mapping from public information to price should not be
controversial. There is no existing consensus on the ``right'' macroeconomic model: extant macro-asset models t the data poorly, as
noted in chapter 1. In his book Exchange Rate Economics, Isard (1995,
182) puts it this way:
Thus, economists today still have very limited information about the relationship between equilibrium exchange rates and macroeconomic fundamentals.
Accordingly, it is hardly conceivable that rational market participants with
complete information about macroeconomic fundamentals could use that
information to form precise expectations about the future market-clearing
levels of exchange rates.

This fact has important implications for the role that order ow plays
in mapping information to price.
Relaxing the all-agents-know-the-mapping assumption is not, however, the only way to restore a role for order ow. There are many
types of information that do not conform to the rst of the two
macro-asset model assumptionsthat all information relevant for
exchange rates is publicly known. Section 2.3 reviews these types of
information. In anticipation of that material, I offer a recent quotation from Rubinstein (2000, 17) that presages some of these information types: ``Perhaps the most important missing generalization in
almost all work on asset prices thus far has been uncertainty about
the demand curves (via uncertainty about endowments or preferences) of other investors. This injects a form of endogenous
uncertainty into the economy that may be on a par with exogenous
uncertainty about fundamentals.'' Before reviewing these informa-

22

Chapter 2

tion types, let me rst provide some empirical evidence on order

ow's information role.
2.2

Empirical Evidence that Order Flow Is Informative

Four distinct empirical methodologies have been used to generate

evidence of an information role for order ow; that these different
methodologies produce similar results is an indicator of the evidence's robustness. The methodologies look in different places for
indications of order ow's information effects: (1) persistent effects of
order ow on price; (2) adverse selection components of bid-offer
spreads; (3) volatility responses to trading halts; and (4) survey data
from FX dealers. In this section, I briey review some of the papers
that t in each of these four categories. I keep it brief here because it
is too early to present these empirical methodologies in detail. The
detailed material appears in chapter 5, which is dedicated to empirical frameworks.3
Methodology 1:

Persistent Order Flow Effects on Price

One methodology used to show that order ow is informative focuses on order ow's price effects and their persistence. It is common in the literature to distinguish between order ow that has
transitory effects on price and order ow that has permanent effects
on price. When order ow has transitory effects on pricesometimes
called ``indigestion'' or ``inventory'' effectsthese effects are often
referred to as pricing errors. When order ow has permanent effects
on price, however, these effects are taken to reect underlying fundamental information. (French and Roll 1986 use this identication
scheme, for example, in their celebrated paper on information arrival
and stock return volatility.)
In empirical microstructure, the standard way to implement this
idea is to estimate vector auto-regression models (VAR) and test
whether innovations in order ow have long-run effects on price
(Hasbrouck 1991a, b). When applied to data from major FX markets,
one nds that order ow innovations do indeed have long-run
effects on price, indicating that they convey bona de information.
Examples of ndings along these lines include Evans 2001 and Payne
1999.

The Economics of Order Flow Information

23

A second method for testing whether order ow effects are persistent uses time-aggregated order ow to explain price movements in
FX markets. That is, rather than asking whether a single trade has an
impact on price, this asks whether trades aggregated over time (say a
day) have an impact on price. The idea is that if single trades have
only eeting effects on price, then order ow aggregated over the
day will not be closely related to daily price movements. When applied to the FX market, one nds that daily order ow does remain
strongly positively related to daily price changes. Examples of ndings along these lines include Evans and Lyons 1999 and Rime 2000.4
Methodology 2:

Spreads and Informative Order Flow

Though not obvious to people new to microstructure, bid-offer

spreads provide a means of testing whether order ow is informative. This is rather striking. It implies that on the basis of spread
behavior alone, one can learn something quite important about the
FX market's information structure. To wit, the information structure
is not one of public information with a publicly known mapping of
that information to price, but one where individual orders convey
nonpublic information.
To understand why spreads provide a test, it is important to
understand how spreads are determined. I show in chapter 5 that
spreads exist because of three costs faced by dealers. One of these
coststypically referred to as an adverse selection costresults
from asymmetric information. Dealers know that when they trade
with someone who is better informed, they can expect to lose money
on the trade. If one could identify better-informed traders before
trading, then this would not be a problemdealers could choose not
to trade, or could adjust price appropriatelybut dealers typically
cannot identify those who are better informed. Given this, one way
to protect against losses is to increase the width of the quoted spread
to all potential counterpartiesinformed and uninformed alike. That
way, there is more room for prices to move against the dealer before
he begins to lose money. When a dealer protects himself this way, we
say that he has included an adverse selection effect (or component) in
his spread. If we nd this adverse selection component of the spread
empirically, this indicates that dealers believe that some traders'
orders reect better information (despite the dealers' inability to
identify exactly who those traders are).

24

Chapter 2

Empirical ndings show that an adverse selection effect is indeed

present: dealers increase spreads to protect themselves against informative incoming order ow (Lyons 1995; Yao 1998b; Naranjo and
Nimalendran 2000).5 In Lyons 1995, for example, I nd that the FX
dealer I track protects himself from adverse selection by increasing
the width of his spread by about one pip (or 0.0001 DM/$) for every
$5 million increase in the size of the incoming order. (This estimate is
a pure adverse selection effect on the spread; i.e., the model controls
for other factors that affect spread width.) Yao (1998b) also nds an
adverse selection effect on the spread. In addition, he nds that the
dealer he tracks prots from trades executed after observing order
ows. Unlike the Lyons and Yao papers, which look at market trading in general, Naranjo and Nimalendran (2000) focus on central
bank tradingthat is, intervention. They nd that the adverse selection effect on the spread is positively related to the variance of unexpected intervention trades.
Methodology 3:

Volatility Responses to Trading Halts

The third methodology used to show that order ow is informative

focuses on price volatility over periods that contain a subperiod
during which trading is halted. (Think, for example, of the volatility
of prices over weeks in which trading is halted on Wednesdays.) The
trick in this case is to identify trading halts that are unrelated to the
ow of public information. If the trading halt is related to the ow of
public informationholidays, for examplethen changes in volatility that occur because of the halt could easily be due to changes in
that public information ow. If one is condent that the ow of
public information has not changed, however, then changes in volatility that occur because trading is halted must be attributed to something else. In particular, a nding of lower volatility over periods
that contain trading halts indicates that either (1) informative order
ow is not reaching the market during the halt, (2) the lack of trading during the halt reduces pricing errors, or (3) some combination
of both.
French and Roll (1986) were the rst to use this methodology,
which they applied to stock trading. They identied a set of days
mostly Wednesdayson which the New York Stock Exchange
(NYSE) was closed due to order processing backlogs. These were
regular weekdays, not economy-wide holidays, so the closure had no

The Economics of Order Flow Information

25

impact on the underlying rms nor on the economy as a whole; the

only difference was that the NYSE was not trading. Thus, there is no
reason to believe the ow of public information changed (save, perhaps, rms' choosing not to release information due to the closure).
French and Roll then measured the volatility of returns from Tuesday to Thursday in weeks when there was trading on Wednesdays,
and compared this to the volatility of returns from Tuesday to
Thursday in weeks with no trading on these special Wednesdays.
They found that volatility was signicantly lower in weeks with the
Wednesday closures. This, together with other evidence they provide, leads them to conclude that the main source of the volatility
reduction is possibility (1) above: informative order ow was not
reaching the market during the Wednesday halts.
Ito, Lyons, and Melvin (1998) analyze an analogous experiment in
the FX market. Their experiment was made possible by a change in
the trading rules in the Tokyo FX market. Until December 1994,
banks were restricted from trading in Tokyo over a lunch period
(from 12:00 to 1:30 local time). In December 1994, the restriction was
lifted. There were no other changes in policy or other shifts in the
ow of macroeconomic information (e.g., announcement dates and
times remained unchanged), so the ow of public information was
not affected. Ito et al. nd that after lifting the restriction, volatility in
the $/yen market over the lunch period doubled.6 Was this due to
informative order ow or pricing errors? The increase in volatility
does not, in itself, allow one to distinguish between them, but because it is unlikely that the increase is due entirely to pricing errors,
this result also supports an information role for order ow.
Methodology 4:

Surveys of Foreign Exchange Dealers

If we think dealers might believe order ow is informative, why

not just ask them? This is the tack taken by Cheung and Yuk-Pang
(1999, 2000), Cheung and Chinn (1999a, b) and Cheung, Chinn, and
Marsh (1999). These papers survey foreign exchange dealers in major
trading centers around the world (London, New York, and Tokyo).
Fifty percent of the dealers surveyed by Cheung and Yuk-Pang believe that large players in the FX market have a competitive advantage that derives from ``better information'' and ``a large customer
base.'' The latter is often described as a source of advantage for
dealers because it gives them privileged information about their

26

Chapter 2

customers' orders, and they can base their speculative trades on this
information (see, e.g., Yao 1998b).
Other authors report results from discussions with dealers that are
similar, though these results are less formal than the above-noted
surveys. For example, Goodhart (1988, 456) writes, ``A further
source of informational advantage to the traders is their access to,
and trained interpretation of, the information contained in the order
ow.'' Similarly, based on interviews with nine FX dealers in London, Heere (1999) reports that the dealers emphasize the importance
of asymmetric information. The dealers she interviews state that information asymmetry is based on both order ows and the identities
of the institutions behind those order ows.
2.3

Dening Private Information

In this book, information is private information if it

1. is not known by all people; and
2. produces a better price forecast than public information alone.7
This is a natural denition. I should say, though, that it is a bit
broader than some people have in mind. For example, under this
denition, if a dealer has privileged order ow information and that
information aids him or her in forecasting prices, then this constitutes private information. (The dealer could be expected to take
speculative positions based on it.) For understanding motives for
speculative trade, this denition is quite useful.
To add concreteness to the denition, let me identify two subcategories of private information and provide some examples from
FX. Consider a simple two-period trading model in which trading
occurs initially at a price P0 , again at t 1 at a price P1 , and then a
terminal payoff value V (e.g., an unknown nal dividend) is realized
at t 2. The rst type of private information that can arise in this
setting is private information about the size of the terminal payoff V.
Information that alters expectations of the payoff V will clearly be
relevant for the prices P0 and P1 . Let me offer two examples of private information in the FX market that are arguably of this type (the
next section presents more examples). The rst example is private
information about central bank intervention: the dealer who receives
a central bank's order has also received private information (Peiers

The Economics of Order Flow Information

27

1997). The second example is dispersed private information about

particular components of exchange rate fundamentals, for example,
the trade balance: real trade in goods and services generates FX
orders that provide information to dealers about trade balances long
before published statistics are available (Lyons 1997a).
A second type of private information is unrelated to the payoff
value V (in contrast to the rst type), but is related to interim prices
P0 and P1 . P0 and P1 depend on many variables beyond expectations
of the payoff V; in essence, this includes any variable that determines
the risk premium in these prices. An example is traders' risk preferences. Other examples include traders' trading constraints, the
supply/distribution of the risky asset, and other features of the
trading environment. Insofar as these affect P0 or P1 without altering
expectations of V, superior knowledge of them is private information
of this second type.
Consider two examples from the FX markets of private non-payoff
information. The rst involves information about transitory risk
premia, whereas the second involves information about persistent
risk premia. In the rst case, because the transparency of order ow
in FX is low, dealers have better knowledge of their own inventory
and that of other dealers than the general public. If inventory risk
earns a risk premium, as many microstructure models predict, then
superior knowledge of this kind allows a dealer to forecast interim
price more accurately than the market at large.8 As aggregate inventory across dealers becomes known, this induces a change in the risk
premium and an attendant change in price, even though terminal
payoff expectations remain unchanged.9 (Discussions with FX dealers
indicate that this type of private information is indeed relevant.) The
second example of private non-payoff information, which involves
persistent risk premia, is an extension of the rst. Specically, if the
dealers' positions in aggregate are large enough that absorption by
the rest of the market requires a sustained risk premium, then the
dealers' superior information will forecast these price effects as well.
(The following section provides more detail on this mechanism; see
also Cao and Lyons 1998 and Saar 1999.)
This book tends to highlight the second of these two broad types
of private informationinformation about interim prices. There are
several reasons for this. First, in my judgment, this type of private
information is especially relevant for the FX market. Second, people

28

Chapter 2

who believe that private information does not exist in the FX market
typically have in mind only the rst of the two typesprivate information about the payoff, V (which in the FX market translates to
private information about future monetary fundamentals like interest differentials). Highlighting the second of the two information types
broadens that perspective. Third, previous literature tends to neglect
this category of private information. Information-theoretic models
of trading are specied with private information about terminal
payoffs. Empirical models follow suit. But this makes interpretation
of empirical results difcult: should one interpret evidence of private information as reecting the rst type or the second type? The
answer is not clear.
A Comment on the Term ``Fundamentals''
The term fundamental means different things to different people.
For example, one might be tempted to consider the second of my
two private-information types as nonfundamental. The quote from
Rubinstein (2000) that closed section 2.1 is suggestive of this narrower denition of fundamentals. In that quote, he distinguishes
uncertainty about fundamentals (i.e., payoffs) from uncertainty
about agents' preferences and endowments. But all of these factors
are fundamental to asset pricing. My choice to put the two broad
types of private information on equal footing recognizes the joint,
complementary nature of these two categories of fundamentals. The
issue is more than semantic; it affects the way we frame our thinking
about price determination.
My use of the term ``fundamental'' to refer to information of both
types is not so broad that it is no longer meaningful. The examples of
private information above are all bona de determinants of price in
optimizing, well-specied models. None of the examples presented
here require ``bubbles,'' ``greater fool'' behavior, or irrationality.10
2.4

Extending the Taxonomy of Information Types

Let us shift to a more theoretical approach to identifying the types of

information that order ow might convey. This section extends the
last section by adding more granularity to the earlier two-category
breakdown of private information. The danger in adding more

The Economics of Order Flow Information

29

granularityand grounding it in theoryis that readers new to

microstructure will nd this section tougher going. Readers who do
nd this section more difcult should rest assured that the previous
section is an ample introduction to the basic information-theoretic
issues that arise later in the book.
Let us begin by recognizing an important connection within microstructure models between order ow and information. As we will
see in chapter 4 (which reviews microstructure theory), order ow
is the proximate determinant of price in all the standard models.
This ``all models'' property is important: It ensures that the causal
role played by order ow is not dependent on the specication of
market structure.
To understand the specic types of information that order ow
can convey, one needs to understand the specic channels through
which order ow has price impact. At the broadest level, the information conveyed by order ow can be any information that currency
markets need to aggregate (e.g., differential interpretation of macro
announcements, changing institutional risk tolerance, shocks to
hedging demands, etc.). Within this broad class, theory provides a
taxonomy of different information types.
To set the stage for those different types, consider the following
simplied view of asset pricing. Let us write the initial price of the
risky asset in last section's two-period setting as
P0

EV
;
1d

2:1

where P0 is the price at t 0, EV is the expected value of the risky

asset's payoff, and d is the market-clearing discount rate (twoperiod). In the case of a stock, where V is a dividend, this equation is
the familiar dividend discount model. (The whole stream of discounted dividends would be included in a setting with multiple
payoffs.) In the background of such a pricing equation is a marketclearing condition, which equates market demand with market supply (supply is typically assumed xed). Any factorother than the
numerator EVthat affects market demand/supply will affect price
through the market-clearing discount rate d.
Now we are ready to outline the types of information that order
ow can convey. There are three key types that arise within microstructure theory:

30

Chapter 2

1. Payoff information
2. Discount-rate informationinventory effects
3. Discount-rate informationportfolio balance effects
For a stock, payoff information refers to information about future
dividendsthe numerator EV in our dividend discount model. For
a bond, payoffs take the form of coupons and principal (which are
publicly known as long as the bond is default free). For foreign exchange, payoffs include future short-term interest differentials (foreign
minus domestic; see section 6.1 for more detail). These represent the
net cash ows that accrue to holders of money market instruments
denominated in foreign exchangeakin to the dividends that accrue
to holders of a stock. (FX speculators who buy foreign exchange do
not hold actual currency, which bears no interest, but instead invest
their holdings in short-term, interest-bearing instruments.) We will
see that private information about payoffs is the basis for a class of
microstructure models known as information models (reviewed in
chapter 4).
Let me provide examples of how order ow, per se, might convey
private information about payoffs. The simplest examplethough
not so common in the major FX marketsis information about
future interest rates conveyed in the orders of a central bank (intervention). A second example likely to operate on a more regular basis
is information about people's expectations of future interest differentials (as noted in chapter 1).11 To understand this example, recognize that in reality, people do not all share the same EV. Instead,
each of us has our own expectation about the direction of future interest rates, based on the millions of bits of information we use to
form this view. This can be described by expressing the numerator as
EV j Wi , where Wi denotes the information that market participant i
uses to form expectations. Because participant i's orders depend on
EV j Wi , observing his or her orders conveys information about that
expectation. Thus, order ow serves as a proxy for people's expectations about future payoffs, and the information embedded therein.
These orders are the backed-by-money expectational votes that the
market ``counts'' when determining price.12
Turning to discount-rate information, microstructure theory emphasizes two distinct causes of time variation in discount rates.13
Both causes involve changing risk premia and rely on order ow to

The Economics of Order Flow Information

31

play the central role. The rst of these causes is inventory effects.
(These arise in the class of microstructure models known as inventory models; see chapters 4 and 5 for details.) The idea here is that
risk-averse dealers will require compensation for absorbing transitory mismatches in supply and demand over time. The larger the
mismatch, the greater the risk the dealer must assume and the
greater the compensation the dealer requires. Suppose, for example,
that the mismatch is such that the dealer needs to absorb market sell
orders (i.e., the dealer needs to buy). The dealer may be willing to
absorb a small amount at only a slightly lowered price, but he would
require a signicantly lowered price to absorb a large amount.
Dealers thus earn a transitory risk premium for providing liquidity.
These effects on price last only as long, on average, as the mismatches in market supply and demand. Once the supply-demand
mismatch is remedied, the dealer no longer holds a position (inventory), so the effect on price dissipatesthe risk is diversiable. (In
terms of equation 2.1, the inventory effect would alter the discount
rate d that establishes p0 , but not the discount rate that later establishes p1 .) This type of order ow effect on price is what people have
in mind when they assert that ``microstructure effects zzle quickly.''
To summarize, these effects arise because risk is not perfectly and
instantaneously spread throughout the whole market; instead, dealers
bear disproportionate risk in the short run, and this affects price in
the short run.14
The second cause of time variation in discount ratesthe third of
our three information categories aboveis what macroeconomists
call portfolio balance effects. To distinguish these effects from inventory effects, the idea in this case is that even after risky positions are
spread through the economy as a whole, order ow's effect on price
will not disappear completely. (In terms of equation 2.1, the discount
rates that establish both p0 and p1 are affected. With an innite
number of periods, all discount rates will in general be affected.) In
other words, the risk that drives the portfolio balance effect is undiversiable (unlike the risk that drives the inventory effect). Of course,
to distinguish this from the rst of our information typesinformation about payoffsit must be shown that order ow is not conveying information about EV.
Let me offer two types of FX order ow that are unrelated to
EV, but may be large enough to have persistent portfolio balance

32

Chapter 2

effects. These are ordersor the aggregation of ordersthat arise

from liquidity demand (e.g., importing and exporting) or from
hedging demand. Consider a specic example. Suppose IBM sells
equipment worth $3 billion in Great Britain, is paid in pound sterling, and then dumps those pounds on the spot market to exchange
them for dollars. (Suppose also that the transaction is unrelated to
the future monetary fundamentals V.) Somebody must be willing to
step up and hold those pounds for the indenite future. (The word
``indenite'' is important here: to keep the example simple I am
assuming that IBM is not going to reverse its decision to sell the
pounds sometime in the futureit is a permanent portfolio shift.
Also, I am assuming that the equipment buyer makes no adjustment
in its portfolio other than the pound payment.) If elasticity of market
demand for those pounds is less than innitecalled imperfect substitutability in the macro literaturethen the dollar price of a pound
(the $/ rate) must fall to induce other market participants to purchase the pounds. Algebraically, for the market to clear we need
S D1 EV P D2 IBM;
where I will assume that the aggregate supply of pounds, S, is xed
by the bank of England, and the aggregate demand for pounds has
two components: a speculative component D1 and a nonspeculative
component D2 (here driven by IBM's liquidity demand for pounds).
The speculative component depends negatively on the current spot
rate, P (with less than innite elasticity), and the nonspeculative
component demands positively on the liquidity demand for pounds.
Though quite simple, this example is in the same spirit as the portfolio balance effects from sterilized intervention by central banks
an example familiar to international economists. The only difference
is that instead of the central bank forcing a (payoff-unrelated) position on the public, in my example it is a subset of the public forcing
a (payoff-unrelated) position on the rest of the public. (See chapter
8 for more on central bank intervention, including the distinction
between sterilized and unsterilized intervention.)
Let me relate the above discussion to the previous section's introduction of private information. In effect, the discussion above splits
one of the earlier private information types into two parts. The previous section introduced a type of private information unrelated to

The Economics of Order Flow Information

33

Figure 2.1
Supply curves with only transitory inventory effects. The dotted region represents the
transitory inventory effects. The effective spread faced by a customer for a 10-unit
order is the difference in price along the short-run net supply curve S SR between 10
and 10. If a customer wants to buy 10 British pounds from the dealeran order of
10then he must pay the higher dollar price. If the customer wants to sell 10 pounds
to the dealeran order of 10then he will receive the lower dollar price. Over the
longer run, however, the dealer unloads his position on the rest of the market at a price
that does not include the transitory inventory effects. The market's net supply is perfectly elastic, by assumption, which corresponds to a longer-run supply curve S LR
slope of zero. The linear relationship shown along S SR is a special case, which I adopt
for simplicity.

the payoff value V but relevant to interim prices P0 and P1 so-called

nonpayoff information. This is what I am now calling private information about discount rates, of which I have established two distinct
categories, inventory effects and portfolio balance effects. By splitting
the earlier type in two, this section provides more granularity and
links the earlier description more tightly to microstructure theory. It
is helpful to distinguish between the two subcategories because their
different properties allow us to isolate them empirically: inventory
effects are transitory, but portfolio balance effects persist.
A Graphic View of These Information Types
To clarify these three types of order ow information further, I turn
to a graphic interpretation. Consider, for example, the second type of
information outlined aboveinformation about transitory inventory
effects on discount rates. These effects are presumed to dissipate
quickly because dealers are not holding these positions for long; the
spreading of risky positions to nondealer participants occurs rapidly
(within a day in foreign exchange). Figure 2.1 provides a qualitative
illustration. The short-run market (net) supply curve, denoted S SR ,

34

Chapter 2

Figure 2.2
Supply curves with inventory and portfolio balance effects. The light-gray region represents the transitory inventory effects. The darker gray region represents persistent
portfolio balance effects. Due to inventory effects, the short-run price impact of an incoming order is larger than the long-run impact. But the long-run impact is nonzero,
due to imperfect substitutability; that is, the long-run net supply curve S LR now slopes
upward. The linear relationships shown are a special case, which I adopt for simplicity.

slopes up. (Think of this supply curve as the dealers' willingness to

sell to accommodate incoming order ow, rather than as changes in
the physical supply of foreign exchange.) If order ow is not conveying either of the other two information types whose effect on
price persists beyond the short run, then longer run (net) supply
curve S LR is at.15
Now let us add persistent portfolio balance effects.16 Doing so
implies that the market as a wholebeing risk averseneeds to be
compensated for holding a position it would not otherwise hold.
This requires an enduring risk premium, which takes the form of a
price-level adjustment (per the IBM example). This price adjustment
is not temporary because the risk premium for holding this position
must be sustained. Figure 2.2 illustrates this. The short-run market
(net) supply curve still slopes up, but now it reects both the transitory effect of inventory at the dealer level and the longer run effect
from imperfect substitutability. It is, as a result, more steeply sloped
than the short-run supply curve from inventory effects alone illustrated in gure 2.1. To understand why this short-run effect goes in
the same direction as the longer run effect, think of the underlying
dealer behavior. An individual dealer will buy pound sterling over
the short run only at a relatively discounted dollar price. The market
as a whole will take those pounds off the dealer's hands at a slightly
discounted price, but not as discounted as was required by the dealer

The Economics of Order Flow Information

35

Figure 2.3
Supply curves when order ow conveys information about payoffs and discount rates.
The light-gray region represents the transitory inventory effects. The dark-gray region
represents persistent payoff-information effects. The medium-gray region represents
persistent portfolio balance effects. The gure therefore reects all three of the infor
mation types that arise in microstructure theory. The long-run supply curve S LR
LR
reects the long-run effects from portfolio balance (S ), plus an additional long-run
effect due to the payoff information conveyed by order ow. The linear relationships
shown are a special case, which I adopt for simplicity.

in doing the trade initially with the customer. The dealer, knowing
this cost of laying off his inventory has increased, will pass this on to
the customer in his initial quotes.
Now I will allow order ow to convey information about expected
future payoffs. Like in the case of portfolio balance effects, order ow
effects on price from this channel will persist (see, e.g., French and
Roll 1986; Hasbrouck 1991a, b). Mapped into the market supply
curve, this channel adds additional slope to the long-run schedule
shown in gure 2.2. Figure 2.3 provides an illustration of these longrun supply curves. Note that the short-run supply curve is more
steeply sloped than either of the two long-run curves.

In gure 2.3 there is now a new long-run supply curve, SLR . This
new long-run supply curve reects both the long-run portfolio balance effects S LR , as in gure 2.2, plus an additional long-run effect
due to the payoff information conveyed by order ow.
In later chapters, I address the slopes of these net supply curves
empirically. At this stage, it is worth bearing one point in mind:
order ows in the FX market are enormous relative to other asset
markets. In the gures, this corresponds to being far to the left or far
to the right of the order-ow-equals-zero point. Thus, even if the
slopes of these supply curves are nearly zero, large order ow can
still produce substantial price impact.

36

Chapter 2

Concluding Thoughts
To conclude this chapter, it is worth stepping back to reect on an
important, overarching point. The microstructure tools applied in
this chapter are useful for addressing a rather deep question:
What is the nature of the information this market is processing?
By focusing attention on order ow, these tools help to characterize
which types of information are relevant, and how this information is
aggregated.
In terms of nancial markets' economic role, the aggregation of
dispersed information is of profound conceptual importance. Nobel
laureate Friedrich Hayek (1945, 519) provides an early and powerful
articulation of this point. He writes:
The ``data'' from which the economic calculus starts are never for the whole
society ``given'' to a single mind which could work out the implications, and
can never be so given. The peculiar character of the problem of rational economic order is determined precisely by the fact that the knowledge of the
circumstances of which we must make use never exists in concentrated or
integrated form, but solely as dispersed bits of incomplete and frequently
contradictory knowledge which all the separate individuals possess. The
economic problem of society is thus . . . a problem of the utilization of
knowledge not given to anyone in its totality.

Relative to traditional exchange rate approaches, the informationtheoretic perspective offered here is qualitatively different. As we
shall see in chapter 6 (where I survey macro exchange rate models),
exchange rate economics may warrant a richer information-theoretic
perspective.

The Institutional Setting

Chapter 1 began with an overview of the microstructure approach,

making the point that the approach relaxes three of the asset approach's most uncomfortable assumptions. Those assumptions are that
(1) all FX-relevant information is publicly available, (2) all market participants are alike, and (3) trading mechanisms are inconsequential for
prices. (To be fair, these are not assumptions that macroeconomists
believe are literally true; they are employed as useful abstractions.)
Chapter 2 addressed the rst of these by providing frameworks for
thinking about information that is both relevant and dispersed
throughout the economy. This chapter addresses the other two assumptions: trader heterogeneity and the role of trading mechanisms.
In section 3.1, I provide an overview of the market participants
and how they differ from one another. The rst section also describes
the trading mechanisms used in major FX markets and includes
comparisons with other nancial markets (see also Luca 2000 for a
great deal of institutional information, including an historical account of FX market development). Section 3.2 introduces an important source of institutional informationthe triennial central bank
surveys summarized by the Bank for International Settlements (BIS
1999a, 2002). These survey data provide institutional perspective not
available from any other source. Section 3.3 addresses market transparency, which is crucial to understanding how order ow information is conveyed. In markets that are highly transparent, order ow
is observed by all participants, thereby affecting expectations rapidly
and precisely. In opaque markets (the FX market is relatively opaque),
order ow is not widely observed, so any information it conveys is
impounded into prices more slowly. The concluding section, section
3.4, provides reections on a common and powerful association
that between ``institutions'' and the eld of microstructure.

38

Chapter 3

One should be more careful than I have been thus far when referring to ``the'' FX market. Many people understand this term to mean
spot markets in the major oating exchange rates, such as $/euro
and $/yen, the two largest spot markets. In its broadest sense,
though, the term includes markets other than spot and rates other
than the major oaters. FX markets other than spot include the
full array of derivative instruments (forwards, futures, options, and
swaps). Rates other than the major oaters include those in smaller
markets, such as emerging markets, and those in pegged regimes,
such as Western Europe before the euro. When people quote the
daily trading volume in FX at $1.5 trillion, that statistic applies to the
broadest denition of the market.1
Nevertheless, the essence of the FX market is the spot market. In
1998, the spot market accounted for 40 percent of total turnover
across all FX instrument categories ($600 billion out of $1.5 trillion).
Though this share has been trending downwardit was 59 percent
in the BIS survey of 1989the falling spot share is not due to lower
spot turnover in absolute terms; rather, the derivatives markets have
grown up around the spot market.
For the purposes of this book, however, a vital point must be
understood about the previous paragraph's market share gures: Of
the $900 billion of daily volume that is not from the spot market,
$734 billion of this is FX swaps,2 and FX swaps have no order ow
consequences in the FX market. To understand why, one rst needs to
know what these swaps are. An FX swap bundles two FX transactions that go in opposing directions. For example, I agree to buy
100 million euros today for dollars (spot), and at the same time I
agree to sell 100 million euros for dollars for settlement in one month
(forward). This example is called a spot-forward swap. (One can also
do forward-forward swaps, in which case the rst of the two transactions is a nearer dated forward transaction than the other.) Note
that the two orders in this example are of equal size, but opposite
sign, so the net order ow impact is zero. Readers familiar with
covered interest parity will recognize that this contract is a means of
locking in an interest differential, and market participants use them
for this purpose (whether hedging or speculating).3 The net demand
impact is therefore mainly on relative short-term interest rates, not
on the FX market. This point is borne out in the behavior of banks:
bankers tell me that when they design in-house models for forecasting exchange rates using order ow, they exclude FX-swap trans-

The Institutional Setting

39

actions from their order ow measures. The bottom line is that the
spot market accounts for about $600 billion of $766 billion, or 78
percent, of the transaction activity that corresponds to the order ow
models of this book.
Nevertheless, 78 percent is not 100 percent, so these statistics still
highlight a tension in dening the scope of this book. On the one
hand, dening the FX market broadly to include derivative instruments is consistent with ofcial denitions like that of the BIS;
moreover, arbitrage relationships link these submarkets tightly, suggestive of one market rather than many. On the other hand, these
submarkets do not share the same market structure, particularly in
the case of futures, which are often traded in physical pits using faceto-face open outcry.
To avoid these difculties, henceforth I focus attention explicitly
on the spot market, in particular the major oating-rate spot markets. (Fixed-rate spot markets are much smaller in terms of trading
volume than oating-rate spot markets.) Unless otherwise noted, my
use of the term ``the FX market'' corresponds to spot markets like
$/euro and $/yen. Work thus far on FX microstructure is heavily
concentrated on spot markets; broadening the scope to derivatives
would bring us into uncharted terrain. I do not, however, completely
exclude work on the derivatives segmentthere are many notes and
references to these related areas.4
3.1

Features of Spot FX Trading

Before digging deeper into institutional detail, it will be helpful to

start by reviewing the basic institutional forms. Though in practice,
most market structures are hybrids of these forms (or involve concurrent use of more than one form), appreciation of this fact requires
familiarity with those forms. The three basic forms of market structure are
1. auction markets
2. single-dealer markets
3. multiple-dealer markets
Naturally, within each of these forms there are further renements.
Because this is not intended as a survey, I offer a few words about
each and provide some examples.

40

Chapter 3

In an auction market (in particular, a ``two-sided'' auction market),

a participant can submit a buy order, a sell order, or both. These
orders can be market orders (buy X units now at the best available
price) or limit orders (buy X units when the market reaches a price of
Y ). In a pure auction market, there is no explicit dealer, so the most
competitive limit orders dene the best available bid and offer prices.
Examples of auction markets include the Paris Bourse and the Hong
Kong Stock Exchange, both of which operate electronically.5
In a single-dealer market, a lone dealer stands ready to buy at
his bid quote and sell at his offer quote, thereby dening the best
available price.6 In this setting, incoming orders from customers are
necessarily market ordersa customer either buys at the dealer's
offer, sells at the dealer's bid, or chooses not to trade. Though some
consider the ``specialist system'' of the New York Stock Exchange a
single-dealer market, this is not accurate. In reality, the NYSE is a
hybrid system, with both auction and single-dealer features. Each
specialist (marketmaker) maintains a collection of customer limit
ordersthe limit order book. If a market order to buy arrives, the
specialist can either match it with the best (i.e., lowest priced) sell
limit order, or, if he offers an even lower price himself, then he can
take the other side. (In the parlance, he can step in front of the limit
order book.) Thus, the specialist must work within the parameters of
the best buy and sell limit orders when trading for his or her own
account. In this sense, the specialist is forced to compete against
the limit order book. This constrains the specialist's ability to exercise monopoly power. Pure examples of single-dealer markets are
rare. Examples include FX markets in some developing countries
where all orders must be routed through a single dealerthe central
bank (developing country markets of this type tend to be xed-rate
markets).
Multiple-dealer markets come in two main varietiescentralized
and decentralized. In both cases, competition is provided via multiple competing dealers, rather than via limit orders as is the case
in auction and hybrid specialist systems. In a centralized market,
quotes from many dealers are available in a consolidated format,
such as on a single screen (like the United States's NASDAQ), or in a
single physical place (like a futures trading pit). In a decentralized
market, there is generally some degree of fragmentation because not
all dealer quotes are observable. One result of this fragmentation is
that simultaneous transactions can occur at different prices.

The Institutional Setting

41

The spot foreign exchange market is best described as a decentralized multiple-dealer market. (This is also true of forwards, options,
and swaps markets in major currencies worldwide.) There is no physical locationor exchangewhere dealers meet with customers, nor
is there a screen that consolidates all executable dealer quotes in the
market.7 In this way, it is quite different from most equity and futures
markets. In its structure, the spot FX market is perhaps most similar to
the U.S. government bond market (bond markets have only recently
attracted attention in the microstructure literature).8
Three characteristics in particular distinguish trading in FX from
that in other markets:
1. trading volume is enormous;
2. trades between dealers account for most of this volume;
3. trade transparency is low.
Volume in the spot $/euro market alone is about $150 billion per
day, dwarng that of any other single nancial instrument. Remarkably, interdealer trading currently accounts for roughly two-thirds
of this volume, a much higher share than in other multiple-dealer
markets (the remaining one-third is between dealers and nondealer
customers).9 Finally, the FX market has an uncommon information
structure. Specically, order ow in FX is not as transparent as in
other multiple-dealer markets: in most national marketswhether
equity or bondsby law, trades must be disclosed within minutes.
FX trades have no disclosure requirement, so trades in this market are not generally observable. From a theoretical perspective, this
feature is quite important because order ow can convey information
about fundamentals. If order ow is not generally observed, then the
trading process will be less informative and the information reected
in prices will be reduced.
Let me clarify the players in the spot foreign exchange market. In
addition to providing context, this will help to classify trades into
types depending on who the counterparties are. (The classication of
trades into types is relevant to the material in later chapters.) The
three main categories are
1. dealers
2. customers
3. brokers (strictly interdealer)

42

Chapter 3

Dealers provide two-way prices to both customers and other dealers.

In major spot markets (like $/euro and $/yen), most dealers trade
only a single currency pair. Though the number of banks that have a
dealer in any major spot market is large (greater than 100 worldwide), the top ten banks handle the lion's share of the order ow,
and concentration is increasing over time: Over the past 10 years, the
combined market share of the top ten dealers has risen from around
40 percent to around 50 percent (BIS 1999a, 15; see also Financial
Times, Survey: Foreign Exchange, 5 June 1998).
The customer category includes many institution types, such as
nonnancial corporations, nancial rms/managers, and central
banks. Chapter 9 provides some more detailed analysis of the trades
of these individual customer categories. Also, chapter 8 provides
analysis of the central bank category and the role of these trades
in intervention. (Because it is natural to consider central banks as
having superior information, this customer type receives special
attention in the macro literature.)
The term ``broker'' is confusing to people more familiar with
equity markets than the FX market. Brokers in equity markets trade
for both their customers and for themselves. FX brokers do not trade
for themselves; they only facilitate trades between dealers. This
facilitation role is important in the spot market. To understand the
FX brokers' role, note that there are two methods for dealers to trade
with one another. One way is for a dealer to call another for a quote
and either sell at the quoted bid or buy at the quoted offer. This is
referred to as a direct interdealer trade. The other method is to trade
indirectly through a broker. (In 1998, about half of all interdealer
trades in the largest spot markets were direct, which implies that the
total volume pie split rather cleanly into one-third customer-dealer,
one-third direct interdealer, and one-third brokered interdealer. By
the end of 2000, only about 10 percent of interdealer trades were
direct; see BIS 2001.)
Think of brokers as a bulletin board. Brokers do not make prices
themselves. They gather rm prices from dealers and then communicate those prices back to dealers. A dealer might want to post a
price through a broker because he prefers not to reveal his identity
before the trade is executed (revealing one's identity before the
trade is a necessary consequence of trading directly). For example,
one dealer may post with the broker a limit order to buy $5 million at
a price of 10. Another dealer may post a limit order to sell $3 million

The Institutional Setting

43

at a price of 12. If these are the best prices the broker has received on
either side, then the broker will advertise a two-way price of 1012,
and will do so without identifying the dealers posting those prices. A
third dealer can choose to trade at one of those prices through the
broker. (If so, after the transaction the broker reveals the counterparty, and settlement occurs directly between the counterparties;
both pay the broker a small commission.) Thus, brokers are pure
matchmakersthey do not take positions of their own, they only
connect dealers that might not otherwise nd each other. In the parlance of the three basic market structures above, brokers are running
an interdealer auction market that operates concurrently with the
multiple-dealer market. In this way, brokers provide a degree of
centralization in an otherwise decentralized FX market.10
These three categories of market participants give rise to three
basic types of trades. We can illustrate these trade types using three
concentric rings, shown in gure 3.1.
The inner ring represents direct interdealer trading, the most liquid part of the market. In the $/euro market, current spreads in this

Figure 3.1
Three types of trades.

44

Chapter 3

inner ring are one to two basis points (one basis point equals one
one-hundredth of 1 percent) for $10 million trades (the standard size)
between large banks during active trading hours (the full London
trading day plus the morning hours in New York). Historically,
dealers have chosen direct trading for larger interdealer trades
(above $10 million). The second ring represents brokered interdealer
trading. The effective spread for a $10 million trade in this ring is
roughly 23 basis points during active trading, though this continues
to fall as brokers take market share away from direct dealing. I add
the word ``effective'' here because the insidethat is, the lowest
spread in the brokered part of the market can be less than 23 basis
points, but that inside spread may apply to trade sizes less than $10
million; any remainder must be executed at less attractive prices,
such that a $10 million trade will have price impact beyond the initial inside spread. The third ring represents customer-dealer trading.
Dealers tell me that current spreads for a $10 million trade are in the
37 basis point range for ``good'' customers. (``Good'' to most dealers
means high volume.)
Visually, gure 3.1 reects a common metaphor used for the foreign exchange market, namely that the market is like a pool of water,
with stones being thrown in the center, where the action is most intense. The stones are the customer orders. Direct interdealer trading
lies at the center. Stones landing in that center send ripple effects
through the brokered interdealer trading, and, ultimately, back to
the customers themselves. Why back to customers themselves? Because dealers tend not to hold positions for very long in this market,
as we shall see below. This metaphor also claries the typical order's
``life cycle.''11
One might ask why a dealer would use a broker if direct prices are
tighter (brokers also charge a commission). Part of the answer is that
smaller banks often do not have access to the tighter direct spreads
that large banks extend to one another. Large banks, too, have
incentives to use brokers. From the large bank perspective, providing
a broker with a limit order provides a wider advertisement of a
willingness to trade than bilateral direct quoting provides. (Keep in
mind that a large bank that provides a broker with a limit order to
buy, for example, is still buying at the bid price, which is below the
offer price. If a second, smaller bank hits that bid, it is the second
bank that sells at this lower price.) Another reason bankslarge and
small alikemay choose to trade via brokers is that they provide
pretrade anonymity (as noted above). In a direct interdealer trade,

The Institutional Setting

45

the dealer providing the quote knows the identity of the other dealer.
(For more detail on incentives to use particular FX trading systems,
see Luca 2000.)
Features for Modeling
With this more complete picture of FX market institutions, let us now
consider other features of FX microstructure that inuence modeling
strategies. Among many cited in the literature, three in particular
deserve note:
Dealers Receive Information from Their Customer Orders
As Citibank's head of FX in Europe said, ``if you don't have access
to the end user, your view of the market will be severely limited''
(Financial Times, 29 April 1991). In a similar spirit, Goodhart (1988,
456) writes: ``A further source of informational advantage to the
traders is their access to, and trained interpretation of, the information contained in the order ow. . . . Each bank will also know what
their own customer enquiries and orders have been in the course of
the day, and will try to deduce from that the positions of others in
the market, and overall market developments as they unfold.'' Note
that banks have little information regarding the customer orders of
other banks. Consequently, insofar as this order ow information
helps forecast prices, it is private information (by the denition of
chapter 2).
Dealers Learn about Marketwide Order Flow Largely from
Brokered Interdealer Trades
Because dealers do not observe one another's customer orders, they
need to gather order ow information from interdealer trading. As
noted above, though, direct interdealer trading does not provide
order ow information to anyone other than the counterparties.
Brokered interdealer trading, on the other hand, does provide order
ow information beyond the counterparties. This is important: of the
three trade types (customer-dealer, direct interdealer, and brokered
interdealer), the brokered interdealer trades are the only order ow
information communicated to all dealers. The broker systems, which
are now electronic, typically communicate this information by indicating whether incoming market orders are executed at the bid or
offer side (indicated with the words ``given'' for a trade at the bid and
``paid'' for a trade at the offer) and by providing information on how

46

Chapter 3

the incoming market order has changed the quantity available on the
bid or offer side (information on the size of the order ow). Though
there is noise in this order ow measure, on a marketwide basis it is
the best measure that is available to dealers.12
Large Dealer Positions are Frequent and Nontrivial
They are a natural consequence of marketmaking in a fast-paced
market with tight spreads (less than 2 basis points in the interdealer
$/euro market).13 FX dealers manage these large positions intensively. The large bank dealer in the $/DM market that I tracked in
1992 (see Lyons 1995) nished his trading day with no net position
each of the ve days in the sample, despite trading over $1 billion
each day. Within the day, the half-life of the gap between his current
position and zero was only ten minutes (Lyons 1998). From the plot
of that dealer's net position in gure 3.2, the strong reversion toward
zero is readily apparent.14
Though the three features noted above are the most important
from a modeling strategy perspective, let me provide a bit more
perspective on what the life of a dealer in a major FX market is like.15

Figure 3.2
Dealer's net position (in $ millions) over one trading week. The vertical lines represent
the overnight periods over which this dealer was not trading. The horizontal distance
between those vertical lines is scaled by the number of transactions made by this
dealer each trading day.

The Institutional Setting

47

Two aspects of a dealer's life are particularly telling: his position

sheets and his protability. Tables 3.1 and 3.2 below provide some
perspective on each of these two aspects. The position sheet is the
source of the dealer position shown in gure 3.2.
Table 3.1 provides a representative position sheet from the dollarmark dealer I tracked through one week of trading in August 1992
(see Lyons 1998 for details). The position sheets provide the dealer
with a running record of his net position and the approximate cost of
that position. The dealer lls the sheets in by hand as he trades
(though these days it is mostly electronic and automatic). Each sheet
(page) covers about fteen transactions. The ``Position'' column
accumulates the individual trades in the ``Trade'' column. Quantities
are in millions of dollars. A positive quantity in the Trade column
corresponds to a purchase of dollars. A positive quantity in the
Position column corresponds to a net long position in dollars. The
``Trade Rate'' column records the exchange rate for the trade, in DM
per dollar (which is the way dealers quote prices in this market). The
``Position Rate'' column records the dealer's estimate of the average
rate at which he acquired his position. The Position and Position
Rate are not calculated after every trade due to time constraints
(the average intertransaction time of this dealer is 1.8 minutes over
this trading week, and during especially active periods it is much
shorter). The ``Source'' column reports whether the trade is direct
over the Reuters Dealing 2000-1 system16 (r Reuters) or brokered
(b Broker). One dimension of actual position sheets not shown in
the diagram is the names of the counterparties. Most banks are reluctant to provide these names, as they are considered condential.
There are several take-aways. First, this dealer's trading day
started around 8:30 a.m., New York time. (The New York market as
a whole begins picking up liquidity around 7:30 a.m.) Second, the
distinction between direct and brokered trades is evident in the way
this dealer structures his sheet, with ``r'' capturing his direct trades
and ``b'' capturing his brokered trades. (Beyond tracking position,
one needs to track brokered trades because they involve a commission payment.) Third, trading is hectic enough that this dealer does
not have time to update all information following every trade. (His
average daily volume is over $1 billion, as we shall see in gure 3.3.)
Fourth, each entry does not record the time, though the dealer does
record the time at the beginning of every card (to the minute), and
occasionally within the card, as was the case here.

48

Chapter 3

Table 3.1
Diagram of Position Sheet Structure
Trade date: 8/3
Value date: 8/5
Position

Position
rate

Trade
rate

Source

Time

1
2

1.4794
1.4797

r
r

8:30

1.4796
28

1.4795

10

1.4797

10

1.4797

10

1.4797

1.4797

1.4797

1.4791

1.4797

1.4797

Trade

0.5

1.4794

0.75

1.4790

1.4791

10

1.4797

1.4799

1.4805

1.4810

8:38

1.4808

The ``Position'' column accumulates the individual trades in the ``Trade'' column.
Quantities are in millions of dollars. A positive quantity in the Trade column corresponds to a purchase of dollars. A positive quantity in the Position column corresponds to a net long dollar position. The ``Trade Rate'' column records the exchange
rate for the trade, in deutschemarks per dollar. The ``Position Rate'' column records the
dealer's estimate of the average rate at which he acquired his position. The Position
and Position Rate are not calculated after every trade due to time constraints. The
``Source'' column reports whether the trade is direct over the Reuters Dealing 2000-1
system (r Reuters) or brokered (b Broker). All trades on this position sheet are
interdealer. First fourteen trades on Monday, August 3, 1992.

The Institutional Setting

49

Perhaps most important, one needs to remember that this is but

one dealer. In many respects, he appears to be representative; in
other respects, he clearly is not. One important way in which he is
not representative is that roughly 95 percent of his trades are interdealer (taking all his position cards together), compared to only
about 80 percent of marketwide volume at the time (1992the
interdealer share has fallen to about two-thirds since then). Thus,
the only types of trades that appear on this position sheet are either
direct interdealer (r) or brokered interdealer (b) tradesno customer
trades are listed (see the detailed discussion in chapter 5 of how
these data compare to data on other dealers). A better understanding
of heterogeneity across dealers is an important frontier for analysis
in this area (see, e.g., Bjonnes and Rime 2000; Vitale 1998).
The second aspect of a dealer's livelihood that is particularly telling is protability. It is illuminating to get a sense for the source and
size of a dealer's prots (particularly given the common, and not
unreasonable view, that FX trading is a zero-sum game). Table 3.2
provides some perspective on these prots, based on the same dealer
whose position sheet appears in table 3.1. From the ``Prot: Actual''
column, we see that this dealer averages about $100,000 prot per
day (on volume of about $1 billion per day). By comparison, equity
dealers average about $10,000 prot per day (on volume of roughly
$10 million per day).17 So, even though this FX dealer's prot as a
percentage of his volume is only one-tenth that of the average equity
dealer, because his volume is hundred times as high, each day this
dealer earns ten times as much.
To determine where these prots come from, rst recognize that
there are two possible sources: speculation and intermediation.
Speculative prots come from being long on dollars, on average,
when the DM price of dollars goes up (and vice versa). Intermediation prots come from the bid-offer spread: buying at the lower
price (bid) and selling at the higher price (offer). The column ``Prot:
Spread'' is an attempt to impute the prot the dealer would have
earned each day purely from intermediation. It is calculated under
the assumption that the dealer earns one-third of his spread on every
transaction. Specically, it is his daily dollar volume times his median quoted spread (0.0003 DM/$) times one-third, divided by the
average DM/$ rate over the sample (1.475 DM/$).18
Let me explain the rationale behind this assumption that the dealer
earns one-third of his spread on every transaction. Suppose the

50

Chapter 3

dealer has no net position, and quotes bid and offer prices of 1.4750
and 1.4753 DM/$, respectively (quotes apply to a standardized
amount in this market, at the time $10 million). If the counterparty
chooses to sell $10 million at 1.4750, then the dealer is long $10
million after the transaction. If the market has not moved, and
another potential counterparty calls for a quote, this particular dealer
will typically shade the price to induce the counterparty to buy
relieving the dealer of his or her long position. For example, this
dealer would probably quote bid and offer prices to the next potential counterparty of 1.4749 and 1.4752 DM/$ (versus the original
1.4750 and 1.4753). Relative to the rst pair of quotes, the new quote
is attractive on the offer sidethe 1.4752but unattractive on the
bid side. If the caller goes for the attractive offer quote and buys, then
the dealer will have sold the $10 million position at 1.4752 DM/$.
The net result of both transactions is that the dealer cleared twothirds of the spread (two ticks) on two transactionsor one-third of
the spread on each transaction.
From the last line of table 3.2 we can see that, under this assumption, this dealer makes most all of his prot from the spread. Of the
$507,929 he made over the week, our estimate of the amount that
came from intermediation is $472,496. (If I had assumed that the
dealer makes half his spread from intermediation, then the prot
from intermediation would have been higher than his total prot,
Table 3.2
Summary of DM/$ Dealer's Trading and Prots
Transactions

Volume
(mil)

Prot:
Actual

Prot:
Spread

Monday

333

$1,403

$124,253

$ 95,101

Tuesday

301

$1,105

$ 39,273

$ 74,933

Wednesday

300

$1,157

$ 78,575

$ 78,447

Thursday
Friday

328
458

$1,338
$1,966

$ 67,316
$198,512

$ 90,717
$133,298

1,720

$6,969

$507,929

$472,496

Total

The ``Prot: Spread'' column reports the prot the dealer would have realized if he
had cleared one-third of his spread on every transaction. It is calculated as the dollar
volume times one-third the median spread he quoted in the sample (median spread
0.0003 DM/$), divided by the average DM/$ rate over the sample (1.475 DM/$)
from Monday, August 3, to Friday, August 7, 1992.

The Institutional Setting

51

indicating that he suffered speculative losses.) This is broadly consistent with the idea that, in terms of speculative prots, this market
is a zero-sum game. However, the market need not be a zero-sum
game in terms of intermediation prots. In that case, customers in
this market are paying the dealers, on average, for the liquidity that
the dealers provide.19 Is that compensation inordinate? The data in
table 3.2 do not allow us to answer that question because these
trades are mostly interdealer trades. A hundred thousand dollars per
day is not a bad day's work, though, at least not where I come
from.20
3.2

Descriptive Statistics: The BIS Surveys

Unlike equity markets, because FX trades are not reported in most

countries, marketwide volume in foreign exchange is not generally
available. Every three years, however, individual central banks survey their nancial institutions regarding FX trading activity (for a
single month, typically April), creating a snapshot. Though the latest
triennial survey was in April 2001, results are not yet available for
the purposes of this book (see BIS 2002). It is useful, nevertheless, to
review and interpret key results from the 1998 survey, in part as
guidance for readers of future surveys (see the summary in BIS
1999aavailable at hwww.bis.orgiand also individual banks'
ndings, e.g., Bank of England 1998 and Federal Reserve Bank of
New York 1998).21 This 1998 survey was the fth triennial survey.
Forty-three countries' central banks participated. Because these data
provide institutional perspective not available from any other source,
they warrant attention here.
Let us begin with the rst table in the BIS (1999a) report, reproduced here as table 3.3. This table shows that the $1.5 trillion in daily
volume in April 1998 is composed of $600 billion in spot trading and
$900 billion in trading of outright forwards and forex swaps. However, as I noted earlier in this chapter, forex swaps have no order
ow consequences in the FX market; the net demand impact from
this category of FX trading is on relative short-term interest rates, not
on the FX market. Of the transaction activity that corresponds to the
order ow models of this book, the spot market accounts for $600
billion of $766 billion, or 78 percent. This is an important point that is
easily missed by readers of BIS reports.

52

Chapter 3

Table 3.3
Foreign Exchange Market Turnover
Category

April
1989

April
1992

April
1995

April
1998

Spot transactions1

350

400

520

600

Outright forwards and forex swaps1

Total ``traditional'' turnover

240
590

420
820

670
1,190

900
1,500

Memorandum item:
Turnover at April 1998 exchange rates

600

800

1,030

1,500

Includes estimates for gaps in reporting.

Source: BIS 1999a, table A-1, adjusted for local and cross-border double counting.
Daily averages in billions of U.S. dollars.
1

When comparing data across surveys, it is important to keep in

mind that the coverage of the survey has changed quite a bit since
the rst survey in 1986. Only four countries participated in that rst
survey. In 1989, the number rose to twenty-one (but some countries
did not provide all types of information). In 1992, twenty-six countries participated, including all countries with larger FX markets. In
1995, the number of countries remained the same, but the coverage
of nancial activity was expanded signicantly to include FX-related
nancial derivatives. In 1998, the number of reporting countries
increased to forty-three and the coverage of derivatives activity was
further expanded.22
Subject to the caveat of the previous paragraph, table A-1 of the
BIS (1999a) report (not reproduced here) shows that spot turnover
increased by 14 percent from 19891992, 30 percent from 19921995,
and 15 percent from 19951998. Notice, though, how much the total
turnover statistics change when measured at the constant (April
1998) exchange rates. The big change is in the 1995 total turnover,
which falls considerably when measured at the April 1998 rates
(from $1,190 to $1,030 billion). This is because the dollar was quite
weak in April of 1995, particularly against the Japanese yen. Indeed,
at the end of April 1995, the yen/$ rate fell to an unprecedented low
of about eighty. The translation to April 1998 exchange rates means
that each dollar of 1995 turnover is scaled down to reect that a
dollar in April 1995 was worth less than a dollar in April 1998 (when
measured against other currencies). If one applies the same constantrate correction to the spot growth statistics, one nds that the three
growth rates cited above14 percent, 30 percent, and 15 percent

The Institutional Setting

53

change to 10 percent, 15 percent, and 33 percent, respectively. 23 This

changes the picture from slowing growth to accelerating growth. It
will be interesting to see whether the statistics for spot turnover in
the April 2001 survey continue this trend of accelerating growth.
Currently, many market participants are predicting a slowing of
spot-market growth, due to two factors: (1) the collapsing of many
European cross markets into the euro, and (2) the more efcient inventory management that is resulting from the dominance of electronic interdealer brokers (which can dampen the hot potato process
described in chapter 1).
Another key table in the BIS report is table B-4, which breaks
down the turnover statistics by currency pair (reproduced here in
table 3.4). As noted earlier in this chapter, the $/DM and $/yen spot
markets are a good deal larger than any other. Although the dollar
versus all other EMS currencies is listed third, most of this trading is
in foreign exchange swaps, not spot. (These foreign exchange swap
tradesbeing in fact trades on interest differentialswere probably
unusually high in April 1998 due to speculation on the convergence
of interest rates in the run-up to the January 1999 launch of the euro.)
Note that markets in currencies against the dollar are the largest.24
These major dollar exchange rates are oating rates (i.e., are market
determined, with little intervention on the part of central banks). 25
Not until the DEM/othEMS line does one nd rates that are ofcially
pegged (though not rigidly sothese rates were allowed to vary
within pre-set bands). Now that trading in the euro has been introduced, there is no need for a DEM/othEMS line in the 2001 survey
table (the largest market, USD/DEM, is replaced by the USD/euro).
Table E-1 from the Statistical Annex of BIS 1999a, not reproduced
here, provides statistics on spot counterparty types. The table shows
an interdealer share of 60 percent (347,689/577,737). Recall that in
section 3.1 I described the interdealer share as roughly two-thirds.
The central bank surveys upon which the BIS draws tend to underestimate the interdealer share in total trading because the category
``other nancial institutions'' includes some nonreporting investment
banks, some of which are important in dealing. (Central banks' role
in supervision and regulation applies mainly to commercial banks,
so commercial banks are more thoroughly represented in the surveys than investment banks.) The evidence that dealers are included in this ``other nancial institutions'' comes from the fact
that this category includes signicant brokered trading; FX brokers

54

Chapter 3

Table 3.4
Reported Foreign Exchange Market Turnover by Currency Pair
April 1998

April 1995

Percentage share

Percentage share
Total
amount

Spot

Outright
forwards

Foreign
exchange
swaps

Total
amount

Spot

Outright
forwards

Foreign
exchange
swaps

USD/DEM
USD/JPY

253.9
242.0

56
36

7
9

37
55

USD/DEM
USD/JPY

290.5
266.6

49
45

8
10

43
44

USD/othEMS

104.3

19

73

USD/othEMS

175.8

14

79

USD/GBP

77.6

33

60

USD/GBP

117.7

33

59

USD/CHF

60.5

37

55

USD/CHF

78.6

30

62

USD/FRF

60.0

17

74

USD/FRF

57.9

16

76

DEM/othEMS

38.2

74

17

USD/CAD

50.0

25

68

USD/CAD

38.2

32

11

57

USD/AUD

42.2

33

59

DEM/FRF
USD/AUD

34.4
28.7

86
31

4
7

9
63

DEM/othEMS
DEM/GBP

35.1
30.7

75
79

12
10

13
11
9

DEM/JPY

24.0

79

12

DEM/JPY

24.2

77

14

DEM/GBP

21.3

84

10

DEM/CHF

18.4

85

DEM/CHF

18.4

86

USD/XEU

16.6

89

USD/XEU

17.9

11

82

USD/SGD

17.2

71

27

1,136.9

43

48

All currency pairs

1,441.5

40

51

All currency pairs

USD U.S. dollar, DEM Deutsche mark, JPY Japanese yen, othEMS other EMS (European Monetary System) currencies, GBP British
pound, CHF Swiss franc, FRF French franc, CAD Canadian dollar, AUD Australian dollar, XEU European currency unit (a basket
currency that includes all European Union members), and SGD Singapore dollar.
Source: BIS, 1999a, table B-4. Daily averages in billions of U.S. dollars and percentage shares.

The Institutional Setting

55

are strictly interdealer, though, so these trades belong in the interdealer category. It is difcult to know how much this biases the
survey-measured interdealer share downward; my adjustment from
60 percent to two-thirds is an educated guesstimate.
Section 7 of the BIS (1999a) report provides some information on
the share of interdealer trading that is brokered and the degree to
which these brokered trades are handled by electronic brokers,
rather than the traditional voice-based brokers.26 Because this section
is not linked to specic tables, one needs to be cautious in interpreting the data. For example, one needs to take care to distinguish statistics that apply to total FX turnover, as opposed to spot turnover.
(This is important throughout the BIS report.) One particularly useful sentence in that section is the following: ``Electronic brokers
now handle almost one quarter of total spot transactions in the UK
market.''27 That represents nearly one-half of all interdealer spot
transactions (because, per above, the survey nds that interdealer
transactions are roughly 60 percent of total spot transactions). To
arrive at a more complete picture of the share and type of brokered
tradingelectronic versus voice-basedone needs to piece together
data from the individual central bank reports.
3.3

Transparency of Order Flow

Any model that includes order ow as a proximate determinant of

price must also specify who observes that order ow. In microstructure research, this is called transparency. The term transparency
is broader than just the observability of order ow, however. It is
dened to encompass the full array of information types that the
trading process might transmit. The three primary categories include
1. pre-trade versus post-trade information
2. price versus order ow (quantity) information
3. public versus dealer information
Applying this three-part taxonomy to chapter two's discussion of
order ow information, we see that the material of that chapter
relates most directly to the (1) post-trade (2) order ow information
that is (3) available to dealers. In equity markets, this issuewhat
post-trade order ow information is seen, and by whomis at the
heart of current policy debates.28 Theoretically, too, post-trade order

56

Chapter 3

ow information is the most relevant because it is the main communicator of shifts in asset demand. When interpreting this information
as shifts in asset demand, however, one needs to be precise. One
cannot infer the sign of a shift in demand from the information that,
say, ten units just traded. One needs to know whether the trade
represents buying or selling pressure. The trade needs to be signed
it needs to be converted from trading volume to order ow.
Actual markets differ radically in terms of order ow transparency. In equity markets, the transparency regime is typically imposed (e.g., regimes are imposed on the London Stock Exchange, the
NYSE, and NASDAQ). On the London Stock Exchange, for example,
the price and size of smaller trades must be disclosed within three
minutes, whereas disclosure of the largest trades can be delayed up
to ve business days. FX markets, in contrast, have no disclosure
requirements. For this reason, FX is particularly interesting because
its degree of transparency has arisen without regulatory inuence. 29
With no disclosure requirements, it is perhaps not surprising that
most FX trades do not generate public order ow information. But,
as described earlier in this chapter, some trades do generate widely
available order ow information. Interestingly, these tradesthe
brokered interdealer tradesproduce a level of transparency that
arises as a by-product of dealer's selective use of this trading
method. The FX market is therefore not an example of purposeful
transparency regime design, as is true for most equity markets.
The only other nancial markets similar to FX in terms of low
transparency are other nonequity OTC markets. (OTC, or over-thecounter, simply means not traded on a centralized exchange.) These
include the U.S. bond markets and much of trading in derivatives.
With the advent of centralized electronic trading in these other markets, however, they are on the way to becoming more transparent
than the FX market.30
Now that we have a better sense from section 3.1 for how transparency arises in the FX market, we can examine the impact of
this transparency on price determination. In markets that are highly
transparent, all participants observe order ow, thereby affecting
expectationsand pricesrapidly and precisely. In opaque markets, order ow is not widely observed, so the information it conveys
may be impounded in price more slowly.31 The FX market is opaque
with respect to customer-dealer order ow. As noted, however,
interdealer FX transactions are not completely opaque. One of the

The Institutional Setting

57

models I present in chapter 4 captures this differential dissemination

of order ow information, depending on order ow type.
For market design, determining which participants see what, and
when, is central. At a broad level, the key tradeoff that concerns
policymakers is the following: Though greater transparency can accelerate revelation of information in price, it can also impede dealers'
ability to manage risk. Full transparency may therefore not be optimal, which must be considered when designing a transparency
regime. Board and Sutcliffe (1995, 2) make the point this way: ``The
purpose of a transparency regime is to allow marketmakers to offset
inventory risk by trading before the market as a whole is aware of
the large trade.'' It is unclear, however, whether the current low level
of transparency of the FX market is optimal from the social perspective. Because low transparency has arisen without regulatory inuence, a reasonable premise is that low transparency serves the
interests of dealers. (Lyons 1996a provides a model in which this is
true; dealers prefer low transparency because it slows the pace at
which price reects information, enabling dealers to better manage
riskin keeping with the logic of the Board and Sutcliffe passage
above.)
But if low transparency is an ``equilibrium'' outcome in the FX
market, how is that equilibrium maintained? It is unlikely to result
from collusion, which is difcult to maintain in competitive, decentralized markets. More likely is that equilibrium low transparency
arises as a kind of externalitya by-product of dealers' individual
decisions to trade using brokers (i.e., to trade using a transparent
method). A fact broadly consistent with this view is that actual
transparency levels produced by brokered trading are quite similar
across the world's trading centers: the share of total trading that is
brokered has historically remained in the 2040 percent range in all
the major trading centers (BIS 1996). Though it is true that dealers
have other reasons for using brokers, such as anonymity, brokered
trading does determine transparency, so dealers are choosing transparency de facto. From a welfare perspective this is an important
issue in institution design.
3.4

Moving Beyond Institutions

In many people's minds, there is a powerful association between the

eld of microstructure and the study of institutions. The association

58

Chapter 3

is natural, but also a bit deceptive. It is true that institution design is

one of the ``poles'' within the microstructure literature; the material
presented in this chapter is suggestive of this pole. But there is a
second polethe economics of nancial market information. This
book is aligned more with the second pole. To emphasize this, I
presented the information framework in chapter 2, before this chapter on institutions. Later chapters, too, reect primarily this second
pole: they present models and methods that characterize how, in
reality, the FX market aggregates dispersed information. The questions addressed within this second pole are of a broader nature than
institution design.
Let me provide some more concrete examples of how microstructure's ``information'' pole extends beyond its ``institutions'' pole.
The rst example relates to order ow and the sense in which order
ow's role is not about institutions. Within microstructure, order
ow is an information transmission mechanism, and, crucially, it
operates regardless of market structure type. Given this rather general property of order ow, it would be a mistake to attribute its
information role to a specic institutional conguration. Pushing
further, for reasons apart from institutions per se, order ow's information role has the potential to realign modeling strategyat
least within exchange rate economics. To understand why, recall
from chapter 1 that within traditional exchange rate economics,
order ow's role in transmitting information was not considered.
This omission is evident from surveys of macro-style empirical work.
Consider, for example, Meese's (1990, 130) survey, where he writes:
``Omitted variables is another possible explanation for the lack of
explanatory power in asset market models. However, empirical
researchers have shown considerable imagination in their specication searches, so it is not easy to think of variables that have escaped
consideration in an exchange rate equation.''
It is hard to argue with Meese's point from a macro perspective.
But the macro perspective considers only variables within the traditional macroeconomic set. From the microstructure perspective,
there is indeed a variable that escaped macro considerationorder
ow. Microstructure has opened macroeconomists' minds to the idea
that order ow can serve as a real-time measure of dispersed information about changing fundamentals.
Let me illustrate further why order ow's role has little to do with
institutions per se. Consider the following thought experiment: sup-

The Institutional Setting

59

pose spot FX markets began to trade in a completely centralized

auction format (akin to the Paris Bourse or Hong Kong Stock Exchange, as described above).32 Would the microstructure approach
still be useful? Yes, because order ow would still be a determinant
of prices. True, one would need to measure order ow in a different
way, because the passive (noninitiating) side of each trade would
now be a limit order rather than a dealer's quote (see chapter 1). But
with this changed measure of order ow, one could then produce the
same analysis of how order ow determines exchange rates that I
present in later chapters. It is unlikely that this would change the
main results in a qualitative way: the underlying information structure of this market has more to do with the properties of the asset
being tradedforeign exchangethan it does with the market
structure per se, particularly at lower frequencies.
Analysis of crashes and collapses is a second example of how
microstructure addresses questions of a broader nature than institution design. The global stock market crash of 1987 attracted tremendous research attention, much of it set within microstructure models
(see, e.g., Gennotte and Leland 1990; Jacklin, Kleiden, and Peiderer
1992). Note, however, that equity prices collapsed across a host of
different market structure types (e.g., the specialist-market NYSE, the
dealer-market NASDAQ, and several auction-type stock markets
around the world as well). Because the crash was common to markets with different structures, one might argue that microstructure
cannot help us understand the crash. But this would be too extreme:
by providing a disciplined approach to complex information problems, microstructure models provide a useful way to understand
crashes and collapses. This kind of analysis moves beyond narrow
institutional concerns such as how auction and dealer markets differ.
Microstructure Effects versus Microstructure Approach
The term microstructure effects is commonly used, particularly
among people who do not work in microstructure. The concept
deserves attention in this chapter on institutions because its connection to institutions is quite close. People typically use the term to refer to temporary effects on prices that arise from specic institutional
features. As such, the term pertains more to the institutions pole of
microstructure than the information pole. I place emphasis on the
word ``temporary'' because the term microstructure effects is most

60

Chapter 3

Figure 3.3
Microstructure effects question: Does trading process affect mapping?

often used to describe eeting effects on asset prices that, at lower

frequencies, are unlikely to be signicant.
Figure 3.3 provides a graphic representation of microstructure
effects and where they arise in the process of price determination.
The lefthand column lists certain fundamental variables that might
drive equilibrium prices. These variables do not, however, translate
directly into price. Rather, they are inputs to the trading process,
represented by the gray box. The microstructure effects ``question'' is
whether the trading process alters the mapping from fundamental
variables to price, and if so, for how long. The presumption is that it
probably does, but that these effects are short-lived.
What is it that people really have in mind when they use the term
microstructure effects? Though there is no explicit denition, in my
judgment people use the term to refer to two particular types of
effects. The rst type is the temporary inventory effects described in
chapter 2 (and diagrammed in gure 2.1). As an empirical matter,
inventory effects in FX are indeed likely to be short-lived: As noted
in section 3.1, the half-life of a dealer's inventory in FX is very short
(ten minutes for the dealer whose inventory appears in gure 3.2).
The second type of microstructure effect that people often have
in mind is the price effect from order ow as it accelerates the
impounding of payoff information. The word ``accelerate'' is important here for understanding why this effect, too, might be only temporary. Consider the example of a pending earnings release by a rm
and the possibility that insider trading prior to the announcement
might signal the size of those earnings. Figure 3.4 provides an illus-

The Institutional Setting

61

Figure 3.4
Accelerationist view of order ow information. The solid line shows a hypothetical
price path for a stock under the assumption that price responds to a higher than
expected public earnings announcement at time t. The dotted line shows the price path
under the assumption that insider trading (buying) in advance of the announcement is
pushing price up.

tration. The solid line is the price path under the assumption that the
market responds to the (positive) public earnings announcement
only. The dotted line shows how the price path would look if, in
addition, the market were to respond to informative order ow
occurring prior to the public announcement. The difference in the
two paths is only temporary here because the announcement reveals
all the information contained in the prior order ow (and then some).
If, as an empirical matter, order ow were conveying only information that is on the verge of public announcement, then one would
be justied in treating its effects as temporary, in the sense portrayed
in gure 3.4. However, this is certainly not the case for the FX market. Order ow's important role in determining exchange rates
documented in later chaptersis virtually unrelated to macroeconomic news that arrives within the subsequent year. The tight relation
between order ow and exchange rates is not, therefore, simply the
result of short-run acceleration of public information ow. Over the
longer run, whether order ow conveys information about more
distant macro policies has yet to be determined. This is an active
topic of ongoing research, one that I return to in chapter 7.
I raise the notion of microstructure effects because depicting
the eld of microstructure in this limited way can affect research

62

Chapter 3

strategies. Let me provide an example that is embedded in the discussion above about analyzing crashes and collapses. In that discussion, I made the point that although the crash occurred across several
different market structure types, microstructure analysis was still
fruitful for specifying information problems that can lead to a crash.
More generally, people less familiar with microstructure are prone to
assert that microstructure cannot resolve any puzzle that (1) is common to markets with different structures or (2) is not common to
markets with the same structure. As an example of the former, it
might be argued that although (apparent) excess volatility is a property of both equity and FX markets, because the NYSE and FX markets have different structures, microstructure cannot help to resolve
the excess volatility puzzle. This reasoning, in my judgment, is too
oriented toward the institutions pole of microstructure. Effective use
of information models within microstructure may indeed help to resolve puzzles in the FX market, even if the same puzzles occur in
other, differently structured markets. Chapters 7 through 9 make this
case.

Theoretical Frameworks

This chapter provides an overview of microstructure theory. It is

only an overview because, in a single chapter, there simply is not
enough room to present the main models in depth. Instead, my objective is to design a road map that people unfamiliar with microstructure can use for navigation. To make the map easy to follow, I
present each model beginning with that model's most valuable
insights. Also, I present simpleand less generalversions of each
model in order to communicate the underlying economics as efciently as possible. Clarity on the underlying economics is important
for understanding the later chapters' applications.1
There are four distinct models in this chapter:
1. Rational expectations auction model
2. Kyle auction model
3. Sequential-trade model
4. Simultaneous-trade model
Together, these four models span the three categories of markets
introduced in chapter 3: auction, single dealer, and multiple dealer.
The rst two models use an auction structure, whereas the third and
fourth use a dealer structure. (Recall that in a dealer market, the best
available price is dened by dealer quotes, in contrast to an auction
market, where best available price is dened by submitted orders.
Based on these differing sources of best price, dealer markets are
sometimes referred to as ``quote-driven'' markets, whereas auction
markets are sometimes referred to as ``order-driven'' markets.2) The
third modelthe sequential-trade modelis a single-dealer market:
only one dealer sets prices. The fourth modelthe simultaneoustrade modelis a multiple-dealer market: many dealers participate

64

Chapter 4

Figure 4.1
A bird's-eye view of microstructure models.

in price setting. We can organize the four models as shown in gure

4.1.
The rational expectations model of securities trading is a natural
rst model for this chapter, even though it is not typically classied
as a microstructure model. Opening with this model claries how
its shortcomings spurred the development of later microstructure
models. Among these shortcomings is that the act of setting prices
is not addressed in the model; there are no participants whose job
it is to set prices. When pressed about where these prices actually
come from, people who work with these models typically refer to a
Walrasian auctioneer, a hypothetical agent outside the model who
collects orders, sets price based on these orders, and executes the
orders at the market-clearing price he sets.3 The other models I
present in this chapter, in contrast, are explicit about who sets price
and what information is available for doing so (i.e., they address
what really happens when the rubber meets the road). Another reason I include the rational expectations trading model within this
theoretical overview is that distinguishing it from microstructure
models is becoming increasingly difcult. For example, in the introduction to his paper on bilateral trading, Wolinsky (1990, 3) refers to
a recent literature that ``looks at the microstructure of rationalexpectations equilibrium.'' This phrasing signals a growing connection between these literatures. Recent papers on central bank
intervention in FX markets provide more evidence of connection;
these papers' use of the rational expectations model ts comfortably
within the microstructure approach (see chapter 8).
The Kyle (1985) model is a natural follow-up to the rational
expectations model. As noted, both models have an auction structure. The key difference is that the Kyle model addresses the act of

Theoretical Frameworks

65

setting price explicitly. This is achieved by introducing an actual

auctioneer to replace the hypothetical auctioneer of the rational
expectations model. Kyle's auctioneer has privileged information
about order ow and uses that information to determine the marketclearing price. In addition to his auctioneer role, he also takes
speculative positions. Because the protocol that governs trading is
specied in detail in the Kyle model (in contrast to the rational
expectations model), the Kyle auctioneer's price setting and speculative decisions are fully specied. This produces an intimate link
between trading protocol and price determination, a hallmark of
microstructure modeling.
The Kyle model is not classied as a dealership market because
dealer quotes do not dene the best available price. The Kyle
auctioneer therefore does not share a characteristic that is true of
dealersthat they rst provide other individuals with a quote, and
then orders are submitted conditional on that quote. Rather, orders
are submitted to the Kyle auctioneer before price is determined, and
the Kyle auctioneer then determines a market-clearing price based on
those orders.
The sequential-trade modelour third modelis the rst of the
two dealership models. A single optimizing dealer determines the
best price before any orders are submitted. Thus, the trading protocol of this model is more ``dealer-like'' than the auction market
models. Also, because orders are executed sequentially (one by one),
this model provides a framework for analyzing individual orders.
This contrasts with the Kyle model, where all orders are batched and
executed simultaneously at a single market-clearing price.
The simultaneous-trade modelour fourth modelrecognizes
that the institutional settings of the previous three models are quite
different from that in FX. All three previous models adopt a centralized market structure, which contrasts sharply with the FX market's
decentralized, multiple-dealer structure. The simultaneous-trade
model is designed to t this FX market structure. It also has features
that are consistent with empirical results on FX markets, such as
dealer aversion to risk. (The sequential-trade and Kyle models assume the dealer/auctioneer is risk neutral.) Empirically, there is
strong evidence that FX dealers behave in a risk-averse way, as we
shall see in chapter 5. Introducing this feature, and the attendant
management of risk by dealers, has implications for trading volume
and price efciency.

66

Chapter 4

I should be clear from the outset that all four of the models I
present below qualify as information models. Information models
constitute one of two broad modeling approaches within microstructure theory. The other is inventory models. The purpose of information models is to explain permanent price adjustment toward a
changed expected future payoff (payoff information was dened in
chapter 2). Order ow is what induces this price adjustmentit
conveys information about these future payoffs. The focus of inventory models, on the other hand, is transitory price variation around a
xed expected future payoff (chapter two's inventory effects). Order
ow is central to inventory models as well, though in this case it
affects price by inuencing dealer inventories.4 Maintaining inventories in these models is costly because it exposes the dealer to risk;
the need to compensate dealers for these costs is what drives price
adjustment.
Though all four models I present are information models, two are
pure information models and two have both information-model and
inventory-model features. The common thread, the information component, is in my judgment integral to the link between microstructure
and exchange rate economics. Though both approaches are relevant,
if pure inventory models were all there were to the microstructure
approachthat is, transitory price variation due to changing dealer
inventoriesthen there would be little hope of resolving the big
exchange rate puzzles.5 Consider, for example, the puzzle of what
determines exchange rates at lower frequencies. Lower frequency
exchange rates, by their nature, are a function of persistent variation,
not transitory variation, so inventory effects alone are not sufcient
for resolving this puzzle.
4.1

An Implicit Auctioneer: The Rational Expectations Model

For perspective on the rational expectations auction model, it is

helpful at the outset to distinguish between two types of equilibria.
These equilibrium types are
1. fully revealing equilibrium
2. partially revealing equilibrium
In a fully revealing equilibrium, all information is embedded in
price, including private information (so-called strong-form ef-

Theoretical Frameworks

67

ciency). More formally, in a fully revealing equilibrium, price is a

``sufcient statistic'' for the underlying fundamental, making private
signals redundant in an information sense.6 Conversely, in a partially revealing equilibrium, price reects a combination of private
information and extraneous noise. The early literature on rational
expectations trading focuses primarily on fully revealing equilibria.
Later papers focus more on partially revealing equilibria. The secret
to producing partially revealing equilibria, which are more true to
life, is to add sources of noise to the trading process that make it
hard to disentangle the causes of price movements. For example,
consider adding noise related to asset supply to a model with private
information. With the additional source of noise, when price rises
people cannot tell whether the cause is more positive private information or smaller asset supply, because both would push price up.
These partially revealing equilibria have important theoretical implications, as we see below.
Insights
With these equilibrium categories claried, here is a summary of
some of the key insights from the rational expectations model.
1. Price plays two roles: it clears markets and conveys information.
2. In a fully revealing equilibrium, an individual's asset demand
depends only on price, not on any private signals received by the
individual.
3. In a fully revealing equilibrium, there is no incentive to invest in
costly information; this incentive is restored in a partially revealing
equilibrium.
Let me comment briey on each of these before moving to the
model's specics. Insight (1) is the most fundamental of the three.
The traditional role of price in determining equilibrium is a simple
market-clearing role: price changes to eliminate excess demand/supply. To clarify this traditional role, consider a nancial market equilibrium based wholly on this role, that is, an equilibrium in which
individuals neglect price's information role.7 Because individuals
neglect the information in price, price has no effect on expectations,
so each individual's demand is a function of his own information only. The market-clearing price that would result would be
a function of all these individual bits of information. That is the in-

68

Chapter 4

consistency: the equilibrium price would reect everybody's information, but individuals are acting as though price is uninformative.
Neglecting the information in price this way would be patently irrational. By bringing the information role of prices to center stage, the
rational expectations model provides a valuable framework for analyzing price as an information aggregator.
Insights (2) and (3) are related. They are often called the two
``paradoxes'' of the rational expectations model. Insight (2) is a consequence of each individual's private signal being redundant when
price is fully revealing. The paradox arises when one recognizes that,
although the equilibrium price fully aggregates all private information, individual demands are conditioned wholly on that equilibrium
priceindividuals neglect their own private information. But private information cannot get into price if individuals are not using
their private information in determining their demands. In a partially revealing equilibrium, this paradox does not arise because individual demands are conditioned on price and private information.
The second paradox, associated with insight (3), is that in a fully
revealing equilibrium there is no incentive for individuals to acquire
their own private informationit would be reected in price before
a protable position could be opened. So if private information can
only be acquired at a cost (e.g., by investing in better research), then
nobody will acquire it, and there will be no private information
to aggregate. This paradox, too, does not arise in a partially revealing equilibrium because when one acquires private information, a
protable position can be opened before price can fully reect that
information.
The Model
The following Grossman-Stiglitz (1980) version of the rational
expectations auction model is widely used. Because it includes both
private information and noise in asset supply, its equilibrium only
partially reveals private information. The model has two traders,
both of whom are risk averse and nonstrategic (nonstrategic meaning that they act as perfect competitors and take market prices as
given). There is a single risky asset and a single trading period. 8
Within the single trading period, there are three events, shown in
gure 4.2.
The value of the risky asset's end-of-period payoff is denoted here
as V, a Normally distributed random variable with mean zero and

Theoretical Frameworks

69

Figure 4.2
Timing of rational expectations model.

variance sV2 .9 Before V is paid or observed, the risky asset is traded at

price P. Initially, before trading at price P, one of the two traders is
``informed,'' meaning that he receives private information about V
in the form of a signal S. Though only the informed trader observes
the signal S, both traders know that S is Normally distributed, with
mean V and variance sS2 . We specify this signal as
S V e;
where the noise in the signal S, denoted e, has a mean of zero and a
variance denoted as sS2 . The other trader is ``uninformed,'' meaning
that he has not observed S and therefore does not have an information motive for trading.
The uninformed trader does, however, have a hedging motive for
trading. In fact, both traders have a hedging motive for trading. Each
trader initially receives a random endowment in units of the risky
assetfor example, shares or euroswhich we denote XI and XU
(I and U denote the informed and uninformed trader, respectively).
Each endowment is Normally distributed, with mean zero and variance sX2 .10 We denote the aggregate supply of the risky asset by X,
where X XI XU . (If XI > XU , then based purely on hedging
motives, one would expect the informed trader to be selling to the
uninformed trader, other things equal.) XI and XU are distributed
independently of one another and independently of the signal S and
the payoff V.
The exponential utility function we use here is used throughout
microstructure theory. This utility function is so prominent in the
literature that a good portion of the appendix to this chapter is
dedicated to explaining its properties and appeal. It is dened over
end-of-period dollar wealth W:11
UW expW:

4:1

This utility function has two convenient properties. First, the riskyasset demands that it implies do not depend on wealth, so realloca-

70

Chapter 4

Figure 4.3
Summary of rational expectations auction model.

tion of wealth in the trading process does not affect equilibrium. This
obviates the need to keep track of individuals' trading gains/losses.
Second, when coupled with the assumption that returns are Normally distributed, this exponential utility function produces a demand function for the risky asset that takes a simple linear form.
At the center of rational expectations models is the pricing rule,
which describes how the model's random variables determine equilibrium price. All traders know the pricing rule. (An implication
of knowing the rule is that the uninformed trader can use the rule,
together with the market price, to back out information about the
informed trader's signal.) In rational expectations equilibrium, the
pricing rule must meet two conditions: the rst is the ``rational
expectations'' part and the second is the ``equilibrium'' part:
Conditions for Rational Expectations Equilibrium
1. Expectations of the payoff V are consistent with the equilibrium
pricing rule.
2. Markets always clear, that is, excess demand equals zero for all
random variable realizations.
When solving for equilibrium below, I will show that the proposed
equilibrium conforms to these rational expectations conditions. (See
gure 4.3 for a summary of the model's key features.)
Solving for Equilibrium
We solve for equilibrium in this type of model ``by construction,''
that is, by proposing a pricing rule and then verifying that it meets

Theoretical Frameworks

71

the two conditions above. Our assumptions about utility and Normal
distributions give us a basis for proposing a linear pricing rule. As
usual, though, it is not a priori clear what the content of that linear
rule should be (or that an equilibrium rule need even be linear). The
initial conjecture is a matter of judgment and experience. The following conjectured rule does in the end meet the two equilibrium
conditions, but this is far from obvious at this stage:
P aS bX

4:2

The key components are the realized signal S and the realized riskyasset supply X. These are natural choices for the proposed rule: they
are the random variables on which asset demands are based. The
remaining random variable V is not a candidate because the payoff V
is not observable at the time of trading. Values for the constants a
and b are determined at the end of the solution process in a manner
that makes them consistent with optimizing behavior of both traders.
There are three additional steps to solving for equilibrium. First,
we need expressions for each trader's expectation of the payoff V;
these must be consistent with the equilibrium pricing rule. Second,
based on these expectations of V from step one, we need expressions
for each trader's risky-asset demand. Finally, we use those demands
to nd a market-clearing price that matches the proposed pricing
rule in equation (4.2). Then we will have our rational expectations
equilibrium, because it conforms to equilibrium condition (2). In that
equilibrium, expectations are formed using the correct pricing rule,
conforming to condition (1).
Expectations
Expressions for traders' expectations are not difcult to produce in
this setting. This is particularly true in the case of the informed
trader because the informed trader learns only from his own signal
he knows the other trader is uninformed. In the appendix to this
chapter, I show why we can write the informed trader's posterior
beliefs about the payoff V conditional on his signal S as Normally
distributed with




sS2
1
S and VarV j S
:
EV j S
sS2 sV2
sS2 sV2
These expressions make intuitive sense. As sS2 the variance of the
signal S about Vgoes to innity (a weaker signal), EV j S goes to

72

Chapter 4

the unconditional expectation of V, or EV 0, and VarV j S goes to

the unconditional variance of V, or sV2 . These are the unconditional
mean and variance of V. As sS2 goes to zero (a stronger signal),
EV j S goes to S and VarV j S goes to zero.
The uninformed trader's expectation is, by denition, not based on
any private signal. Rather, the uninformed trader extracts all his information from price, which, in equilibrium, will embed information
from the informed trader's trades.12 Equilibrium price thus becomes
a kind of ``signal'' for the uninformed. What the uninformed trader
would like to know is the additional information observed by the
informed trader, the signal S. To use price to make inferences about
S, the uninformed trader can use the proposed pricing rulein particular its parameters a and bto transform price P into a second
signal that is distributed about S. Specically, starting from the
pricing rule
P aS bX;
the uninformed trader can divide the price P that he observes by a to
yield:13
P=a S b=aX:
This variable P=a is distributed around a mean of Sthe information
the uninformed trader wants to know. (Neither the informed nor
the uninformed know the value of X, a point I return to below.) For
notational convenience, I will use Z to denote this second signal:
Z P=a S b=aX:
Because S @ NV; sS2 , X @ N0; 2sX2 , and S and X are independent,
we know that Z is distributed Normally about V with variance
sZ2 sS2 2b=a 2 sX2 . (I have introduced the notation ``@'' to denote
``distributed.'') With this value for sZ2 , the uninformed trader's posterior distribution is Normally distributed about V with




sZ2
1
Z and VarV j P; a; b
:
EV j P; a; b
sZ2 sV2
sZ2 sV2
Knowledge of the pricing-rule coefcients a and b is clearly vital
to the uninformed trader's inference. I recognize them explicitly as
conditioning information to highlight that fact. Note too that neither
the informed trader nor the uninformed trader conditions on the re-

Theoretical Frameworks

73

alization of his own endowmentXI and XU , respectivelybecause

they are assumed to be nonstrategic (i.e., to take price as given).
(Below I address this assumption of nonstrategic behavior in more
detail.)14
Demand
Expressions for traders' demand are not difcult to produce as long
as expected returns conditional on available information are still
Normally distributed. From the analysis above, we know that both
trader's posterior distributions are Normal. Given this, and our exponential utility specication, the demand functions for the informed
trader D I and the uninformed trader D U in units of the risky asset,
for example, shares or eurostake the following form (see appendix
for details):
DI

EV j S P
VarV j S

EV j P; a; b P
:
D
VarV j P; a; b

4:3

Note the information role that price plays in the demand of the
uninformed trader (it enters in the conditional expectation and conditional variance).
Inserting the values above for EV j S and VarV j S into this expression for D I and D U yields the following:
D I sS2 S sS2 sV2 P
D U sZ2 Z sZ2 sV2 P:

4:4

Market-Clearing Price
Market-clearing price is determined by equating demand with supply so that excess demand is zero:
D I D U X:
Inserting our expressions from equation (4.4) for D I and D U in this
market-clearing condition yields a price of
P aS bX;
where

4:5

74

Chapter 4


sZ2 sS2
a
sZ2 sS2 2sV2


1
:
b
sZ2 1 a1 sS2 2sV2
(Recall that sZ2 was dened above as sS2 2b=a 2 sX2 .) These values
for a and b insure that excess demand equals zero for all randomvariable realizations, which fullls condition (2) above for rational
expectations equilibrium. Further, we imposed in our derivation of
these coefcient values that the pricing rule used to form expectations is the actual rule used to determine price. This fullls equilibrium condition (1) above. Thus, we have veried what we set out to
verify: that the conjectured pricing rule in equation (4.2) describes a
rational expectations equilibrium.
This equilibrium is partially revealing, a fact evident from the
uninformed trader's expectation. Specically, the uninformed trader
does not know as much in equilibrium as the informed trader, as
shown by the distributions of posterior expectations. Recall that the
variance of the informed trader's posterior expectation is


1
;
VarV j S
sS2 sV2
and the variance of the uninformed trader's posterior expectation is


1
:
VarV j P; a; b
sZ2 sV2
The only difference is the replacement of sS2 with sZ2 , where sZ2 has a
value of sS2 2b=a 2 sX2 . Because 2b=a 2 sX2 must be positive, sZ2 must
be larger than sS2 , so the variance of the uninformed trader's posterior expectation is larger.
The Implicit Auctioneer
I began this chapter by framing the two auction-market models
rational expectations and Kyleas the difference between an implicit auctioneer and an explicit auctioneer. Yet the model description above makes no reference to an implicit auctioneer, so let me
clarify. Strictly speaking, the rational expectations model does not
require an auctioneer, which is why the specication above contains
no reference to one. Nevertheless, the ction of an implicit, Walra-

Theoretical Frameworks

75

sian auctioneer is the traditional way to envision how prices are

actually set in rational expectations models. The implicit auctioneer
collects the ``preliminary orders'' (described in chapter 1), and uses
them to nd the model's market-clearing price. Without a story like
this in the background, there is no way to understand how price is
actually determined in real timethe model requires only that price
clears the market and is consistent with expectations. By introducing an explicit auctioneer, the Kyle auction model brings this process of price determination from a background abstraction to the
foreground.
Discussion
I close this section with four drawbacks of standard rational expectations modeling.15
Generality
A common concern in this literature is that examples like the one
used above cannot be generalized to more complex model settings.
The problem is that for a given model, a rational expectations equilibrium may not exist (i.e., the two conditions for rational expectations equilibrium cannot be met). This is important: models without
an equilibrium are not very useful. Whether an equilibrium exists,
and what properties it has if it does exist, depends on a particular
feature of the model's specication: the number of signals relative to
the number of assets. In the simple specication I present here, the
number of signals and the number of assets are the same, namely
one. When these numbers are not the same, the existence of an equilibrium becomes more fragile. Accordingly, some concern about
generality is warranted.
Nonstrategic Behavior
Notice from the demand function in equation (4.3) that the informed
trader takes the current price as given. Thus, he does not exploit the
fact that his trade has a direct effect on that price (though, by denition, expectations are validated in equilibrium). This is often referred
to as the ``schizophrenia problem'' inherent in these models: the effect that individuals' trades have on price is not negligible, but these
traders behave like perfect competitors (price takers) nonetheless.

76

Chapter 4

Knowledge of Pricing Rule

As noted above, the uninformed trader needs to condition his demand on the pricing rulein this case the values of a and b. As a
practical matter, it is not clear how he acquires such knowledge; this
is not specied within this modeling approach. Individuals may be
able to learn the rule over time, in which case the rational expectations equilibrium might be considered a long-run steady state. Research on learning to form such pricing rules shows this can be
problematic, however.
Order Flow
The pricing rule at the center of the rational expectations model
treats demands (orders) symmetrically. Consequently, trades cannot
be split into an active (or initiating) side and a passive side (recall the
denition of order ow in section 1.2). The model therefore provides
no guidance for empirical work on the links among order ow, information, and price. It is this feature, more than any other, that
makes this model less of a ``microstructure'' model than the other
models of this chapter.
The four drawbacks above are at least partially addressed in the
models to come. There is no question that the models to come have
their own generality problems, so I have little to add on that count.
The second and third drawbacks, however, are largely assuaged by
the other models. For example, they incorporate strategic behavior.
Also, the models below do not rely on a pricing-rule conjecture.
Rather, they introduce an explicit price setter, who optimizes subject
to the constraints imposed by the model. The fourth drawback is a
nonissue in the other models because they provide a clear means of
signing order ow according to which counterparty initiates the
trade.
4.2

An Explicit Auctioneer: The Kyle Model

The Kyle (1985) model and the rational expectations model are close
cousins. Both have an auction market structure, and at the heart of
both is an expectations-consistent pricing rule. The key conceptual
difference is that the Kyle model includes an explicit auctioneer
rather than an implicit one. This changes the nature of the pricing
rule because the act of price setting is now assigned to a player
within the model. Introducing an explicit auctioneer also introduces

Theoretical Frameworks

77

an information dimension that is missing in the rational expectations

model. Specically, the auctioneer can only use available information
for determining price, but the information available to the auctioneer
is determined by the trading protocol itself. This produces an intimate link between trading protocol and price.
Before presenting the model, let me clarify my use of the term
auctioneer. The presence of an auctioneer in the Kyle model provides
a link to the Walrasian auctioneer that is implicit in the rational
expectations model. The term also reminds us that both of these
models are auction models. However, the Kyle auctioneer does more
than just set prices: he also takes trading positions and has privileged
access to order ow information. In many ways, then, he shares features commonly associated with dealers (even though this is not a
dealer-market setting). The literature commonly uses the term marketmaker for hybrid cases like this. Henceforth, I will also use the
term marketmaker in the context of the Kyle model.
Insights
The Kyle model presented below generates many insights. Relative to the rational expectations model, three of the most important
include:
1. Marketmakers are vote counters, not analysts of fundamentals.
The votes they count are the order ow.
2. Marketmakers cannot separate informative orders from uninformative ones, and informed traders can use this to their advantage.
3. Liquidity and market efciency are deeply related: In efcient
markets, there are forces that drive liquidity (depth) toward a constant, unchanging level.
Though the model itself will enrich each of these, let me offer a few
thoughts at this stage. Insight (1) comes out of the feature common to
microstructure models that information processing has two stages.
The rst stage is the analysis or observation of fundamentals by
market participants other than marketmakers (fund managers, proprietary traders, analysts, etc.). The second stage is the marketmaker's
that is, the price setter'sinterpretation of analysis/observation
from the rst stage. That interpretation comes from reading the order
ow, as gure 4.4 illustrates.

78

Chapter 4

Figure 4.4
Two stages of information processing.

Order ow communicates information about fundamentals because it contains the trades of those who analyze/observe fundamentals. Naturally, though, these informative trades are mixed
with uninformative trades, making the task of ``vote counting'' more
complex than the term might suggest. Note too that in the Kyle
model the marketmaker can only learn about fundamentals from
order ow. This is clearly too strong. This complete dependence on
learning from order ow arises in the model because the information
being learned is not public information (by public information I
mean information that is shared by everyone, and whose implication
for the exchange rate is agreed upon by everyone).16 In the case of
public information, marketmakers obviously do not need to learn
from order ow. Though some information relevant to FX is public,
much is not, so learning from order ow is important.
Insight (2) introduces strategic behavior that is not present in the
rational expectations model. There is room for informed traders to be
strategic because marketmakers cannot separate informative and
uninformative orders. The strategic behavior of the informed traders
takes the form of camouaging their trades using the uninformed
order ow. This hides their information from the marketmaker,
thereby reducing the degree to which price moves against them (e.g.,
price rises less as a result of their trying to buy).
Insight (3)that there is a deep relation between liquidity and
market efciencyis a fascinating message. The basic idea is that in
efcient markets there are forces pushing to keep liquidity from
moving predictably over time. Although dened inconsistently in
the literature, liquidity refers here to an order's price impact
what Kyle (and practitioners) calls depth. To understand the stableliquidity insight, let us suppose that liquidity is not constant, and
that we can predict how it will change. A simple example shows that

Theoretical Frameworks

79

Figure 4.5
Timing of Kyle Model.

this predictable liquidity change can result in an inefciencya

trading opportunity. Suppose the $/yen market will be especially
illiquid this morning, but this afternoon it will be very liquid. If I
buy $1 billion this morning, this will push the yen price of dollars
upward. (If my trades can communicate information about fundamentalsas in the Kyle modelthen these price increases from my
buy orders should persist.) In the extreme, suppose liquidity is so
high in the afternoon that unwinding my position by selling $1 billion will not have any price impact. I expect to make money because
I expect to sell at a price greater than my average buying price. In an
efcient market, this opportunity should not exist. More precisely,
predictable variation in liquidity in an efcient market should not be
large enough to generate excessive risk-adjusted returns.17
The Model
The following one-period version of the Kyle (1985) auction model
is a workhorse within the microstructure literature. The model has
three types of traders: a risk-neutral informed trader, a risk-neutral
marketmaker, and many uninformed traders. The uninformed traders are nonstrategic and trade for motives other than information
(such as hedging). There is a single risky asset, and a single trading
period. Within the single trading period there are four events, detailed in gure 4.5.
The value of the risky asset's end of period payoff is V, a Normally
distributed random variable with mean zero and variance sV2 .18 The
informed trader observes the realization of this random payoff V
before trading, but the marketmaker does not.19 After the informed
trader observes V, he and others submit market orders to the marketmaker to be executed at a single market-clearing price P. These
submitted orders are of two kinds: the order from the informed
trader, D I , and orders from the uninformed traders, which sum to

80

Chapter 4

D U . (If D U is negative then the uninformed are, on balance, selling.)

D U is a Normally distributed random variable that is independent of
V, with mean zero and variance sU2 . The informed trader does not
observe D U before submitting his order D I . (Effectively, this precludes the informed trader from conditioning on the market-clearing
price, a stark contrast from rational expectations models, where all
trades are conditioned on the market-clearing price.) For setting the
price P, the marketmaker observes only the sum of the two types of
orders, D I D U .
The marketmaker's pricing rule is pinned down by the assumption
that he expects to earn a prot of zero. This assumption is consistent
with free entry of competing marketmakers, a condition under which
the single marketmaker cannot exercise monopoly power. (This zero
prot condition is important to the model and is shared by many
other models within microstructure.) Expected prot of zero implies
that the marketmaker sets price P as a function of the sum D I D U
such that
P EV j D I D U :

4:6

Price depends on the sum because the marketmaker does not observe
D I and D U individually. The D U component of that sum is exogenous in this simple version of the model, which simplies inference.
The complication comes from the D I component, which depends on
the trading strategy of the informed trader.
An important feature of the Kyle model is that the informed trader
trades strategically, meaning that he takes into account the effect of
his orders on price. This involves conditioning on the behavior of
both other player typesuninformed traders, whose trades are exogenous, and the marketmaker. Recall that in the rational expectations model the informed trader does not trade strategically. In that
model the demand of the informed trader, D I in equation (4.3), takes
the market price P as given, and thus does not consider the effect that
D I has on equilibrium price. Because the informed trader in the Kyle
model is risk neutral (recall the discussion in the introduction to this
chapter), he will choose a trading strategy that maximizes his expected prot. That is, he chooses a demand D I that maximizes the
following:
ED I V P j V ;

4:7

Theoretical Frameworks

81

Figure 4.6
Summary of Kyle Auction Model.

for each possible realization of V. The interaction between the marketmaker's problem and the informed trader's problem is clear from
these last two equations. The marketmaker's pricing rule depends on
the contribution of D I to order ow, but the informed trader's choice
of D I depends on the impact orders have on the marketmaker's price
P. In equilibrium, this circularity is resolved. (See gure 4.6 for a
summary of the model's key features.)
The pricing and trading rules that produce equilibrium convey the
model's essential lessons. The equilibrium analyzed by Kyle is the
unique linear equilibrium, with a marketmaker pricing rule
P lD I D U ;

4:8

and a trading rule for the informed trader of 20

D I bV

4:9

with strictly positive parameters l and b that take the following form
(not derived here):
l 12 sV2 =sU2 1=2
b sU2 =sV2 1=2 :

4:10

Notice that the pricing and trading rules depend on the same
two parametersthe variance of the uninformed order sU2 and the
variance of the payoff sV2 . This is a natural consequence of being determined jointly and is analogous to the consistency criterion that

82

Chapter 4

governs the pricing rule in the rational expectations model. Notice

also that the ratio of these two parameters is inverted in the two
rules. This too is quite natural: When l is high, meaning orders have
high price impact, then b is low, meaning the informed trader trades
less aggressively (to avoid the impact of his own trades). The constituent variance parameters are also easily interpreted. When sV2 is
high, other things equal, the informed trader's information is more
likely to be substantial, inducing the marketmaker to adjust price
more aggressively. When sU2 is high, the informed trader's order is a
less conspicuous (better camouaged) component of the total order
ow, inducing him to trade more aggressively.
Some Intuition for the Equilibrium
One way to provide some intuition for the equilibrium described
above is to show that a marketmaker who faces an insider with the
trading strategy in equation (4.9) would indeed set price according to
the pricing rule in equation (4.8). To show this, we need to examine
the learning problem the marketmaker faces. The marketmaker starts
with a prior belief about the true value V that is the unconditional
distribution of V, or N0; sV2 . The marketmaker then updates his belief using information in the order ow D I D U and sets price equal
to his best estimate of V. The marketmaker knows that the D U component of D I D U is exogenous and is distributed N0; sU2 . The informative component D I depends on the informed trader's strategy.
Suppose the marketmaker conjectures that the informed trader's
trading rule will be the linear rule in equation (4.9): D I bV. We
want to show that an optimizing marketmaker would indeed set
price according to the pricing rule in equation (4.8).
If the informed trader trades according to D I bV, then we can
write the total order ow as
D I D U bV D U :
We can now employ a technique introduced in the previous section:
we can transform the marketmaker's signal into a new signal, one
distributed about the variable of interestthe value V. The marketmaker divides D I D U by b to yield
D I D U =b V D U =b:

4:11

Theoretical Frameworks

83

As in the last section, we call this transformed signal Z, as in

Z V D U =b:
Note that Z is distributed Normally about V with a variance of sU2 =b 2
(from the second term). Using the updating tools in the appendix to
this chapter makes it is easy to show that after seeing Z, the marketmaker's best estimate of V is proportional to Z:
!
b 2 sU2
Z:
EV j Z
b 2 sU2 sV2
Inserting the value of b from equation (4.10) yields
EV j Z Z=2:
The risk-neutral marketmaker then sets his price equal to his best
estimate of V:
P EV j Z Z=2 V D U =b=2 1=2bD I D U ;
where the last step uses equation (4.11). Note that this exactly
matches the pricing rule in equation (4.8) because
1=2b 12 sV2 =sU2 1=2 l:
We have therefore shown what we set out to show: Faced with the
trading rule in equation (4.9), the marketmaker does indeed set price
according to the pricing rule in equation (4.8). Though this is not a
proof of equilibriumwe have not shown that the b we inserted
corresponds to optimal trading on behalf of the informed trader
this intuition for the marketmaker's problem provides valuable perspective on how he learns. Learning from the order ow is also an
essential feature of the two remaining models of this chapter.
Another interesting result is the degree to which the informed
trader's information is revealed by the equilibrium price. A common
way to measure this is from the market's expectation of V. Specically, after observing price, how much more precise is the market's
expectation? (In the context of the model, the ``market'' is the uninformed traders because they have no information other than that
conveyed by price.) Initially the marketand the marketmaker
have an expectation that is distributed about V with a variance I will
denote as s02 . After seeing D I D U , the marketmaker's expectation is

84

Chapter 4

distributed about V with a variance s12 of

s12 12 s02 ;
a fact that is easy to show using the tools in the appendix for updating conditional variances. This is a striking result: Regardless of the
realizations of V and D U , the updated variance is exactly one-half of
the prior variance. The informed trader's strategy results in exactly
half of his information being revealed by the market price. Intuition
for this result starts with the fact that one would not expect s12 to
collapse to zero: in that case, the informed trader would make no
prot, because all the trades would clear at the perfectly revealing
price V. Nor would one expect s12 to remain at s02 : In that case, the
marketmaker learns nothing, and the informed trader could make
innite prots. That s12 should settle at precisely 12 s02 is not obvious,
but it certainly is appealing on aesthetic grounds.
We have not yet addressed the link between liquidity and market efciencythe topic of this model's insight (3). Recall that we
dened liquidity as the sensitivity of price to order ow, which in the
model is measured by l (the coefcient that relates order ow to
price in the marketmaker's pricing rule). Recall too that the point
of insight (3) is the constancy of liquidity over time, as opposed to
liquidity at a point in time. Addressing liquidity over time clearly
requires something more than the one-period Kyle model. The original Kyle (1985) paper also presents a multiple-period version of the
model, which consists of a sequence of batch auctions; Kyle examines
the limit as the number of trading rounds within a xed time interval
goes to innity. In this version of the model, the informed trader
must consider the effects of current trades on future trading opportunities: if the informed trader trades too much too soon, price will
adjust rapidly, and his total prot will be smaller. Instead, the trader
chooses to trade gradually, hiding trades among the uninformed
trades as they arrive over time. Though not obvious at rst blush,
this generates the constant-liquidity property noted in insight (3). To
understand why, note that the informed trader simply wants to trade
with as little price impact as possible. If price impact is not constant,
however, the informed trader could earn higher prot by trading
more when impact is low and less when it is high. This incentive to
reallocate trading across time will persist until the marginal price
impact is equalized. The result in equilibrium is constant liquidity.

Theoretical Frameworks

85

Discussion
The Kyle (1985) model has been extended in many directions (e.g.,
allowing the informed trader to be risk averse, allowing uninformed
traders to be strategic, and allowing multiple informed traders).21 It
is designed to capture the strategic trading of traders with superior
information, and it does that very well. Naturally, though, there are
aspects of markets about which it has less to say. For analyzing the
FX market (and most other asset markets for that matter), three features of the Kyle model limit its applicability.
No Marketmaker Risk-Aversion
The Kyle model is a pure information model. It has none of the features that can produce price effects from marketmaker inventory or
imperfect substitutability (per chapter 2).22 Because the marketmaker
is risk neutral, he always sets price at his conditional expectation of
V (i.e., this story is purely about expected future payoffs). Not only
does this preclude inventory or portfolio balance effects on price, it
also precludes any interaction between information effects and these
other effects. Empirically, there is evidence that FX dealers manage
inventory intensively. One consequence is that this can alter the
composition of order ow: if a greater share of order ow is unrelated to expectations of Vdue to intensive inventory management
then the order ow's signal-to-noise ratio is lowered.
No Spread
Because all orders in the batch auction are executed at the single
auction-clearing price, the Kyle model does not generate a bid-ask
spread. The major FX markets are dealership markets, which do
generate spreads, so the Kyle model is not directly useful for analyzing those spreads. One way researchers have used the model for
spread analysis is by calculating an ``implicit'' spread (e.g., Madhavan 1996). The implicit spread is calculated from the marginal price
impact of a single-unit trade, captured in the model by the parameter
l (equation 4.8). If the one-way price impact is l, then the roundtrip
price impactthe implicit spreadcan be measured as 2l. This estimate raises its own issues, however, because the equilibrium value
of l is not derived under the possibility that additional trades can be
executed.

86

Chapter 4

No Individual Trades
Another consequence of the batch auction format of this model is
that the impact of individual orders cannot be analyzed. This is
unfortunate because the trade data available for major FX markets
include individual orders. One might apply the insights of the Kyle
model by attempting to ``batch'' the individual orders by aggregating
over time. However, the economics of aggregating orders over time
is potentially different from the batching of orders at a point in time.
4.3

Single Dealer: The Sequential-Trade Model

The Kyle model and the sequential-trade model share many features, primarily involving the specication of dealers. First, both
models include a single optimizing dealer whose prices are conditioned on information available specically to him or her.23 Second,
in both models the dealer is risk-neutral. Third, in both models the
dealer learns from order ow and has no other source of fundamental information.
The Kyle and sequential-trade models also have several important
differences that primarily involve the trading protocol. Unlike the
auction market of the Kyle model, the sequential-trade model describes a dealership market. All trades in a pure dealership market
have a dealer on one side of the transaction. In the sequential-trade
model, this takes the form of individual traders who, sequentially,
are selected from a ``pool'' and given a chance to trade at the dealer's
posted bid and offer. This protocol implies three important differences from the Kyle model. First, an explicit bid-offer spread arises in
the sequential-trade model, as opposed to the single market-clearing
price of the Kyle model. This is an attractive feature, particularly for
empirical implementation, because the spread is readily measured in
most markets. Second, the information content of individual orders
can be analyzed in a way not possible with the Kyle model because
in the Kyle model, all orders clear in a single batch. Third, and not
so attractive, the process of selecting traders from a pool in the
sequential-trade model is random. Random selection means that informed traders have no ability to adjust their trading intensity the
way the informed trader does in the Kyle model. Indeed, incorporation of strategic trading was an important advantage of the Kyle
model.

Theoretical Frameworks

87

Insights
The sequential-trade model presented below generates many insights. Relative to the single-auction version of the Kyle model
above, three of the most important include:24
1. Spreads arise even with competitive risk-neutral dealers
2. Price discovery is a gradual process by which price adjusts over
time to impound all private information (i.e., to become strong-form
efcient)
3. Dealers learn about private information from the sequential
arrival of distinct orders.
Insight (1) is truly fundamental. The equilibrium spread in this
model is such that when the dealer happens to trade with an informed trader he loses money (as in the Kyle model, here the informed trader knows the true value exactly). To prevent overall
losses, the dealer must offset these losses with gains from trading
with uninformed traders. The equilibrium spread balances these
losses and gains exactly so that expected prot is zero. Information
alone is thus sufcient to induce spreads; it is not necessary that the
dealer be risk-averse, face other dealing costs, or have monopoly
power in order to generate positive spreads.
The term ``price discovery'' used in insight (2) is not common outside microstructure; this model is a nice illustration of its meaning.
(The Kyle model in its multiple-auction version also generates this
insight.) In macro exchange rate models, it is traditional to use the
term ``price determination.'' All information is public in these models
so price is not really discoveredit is determined by a consensus
linear combination of fundamentals. In contrast, when a dealer is
trading with individuals with private information, the task is indeed
one of discoverydiscovery of the private information. These models capture the process by which this private information is embedded in price. (Recall that strong-form efciency means that all
information, including private information, is embedded in price.)
Dealers' learning about private information from individual
ordersinsight (3)is at the center of this model. Though the Kyle
auction model also focused on learning from order ow, it could
not address the link between learning and individual orders. In most
nancial markets, this individual-order process is much closer to

88

Chapter 4

reality than the batch trading described in the Kyle model. If a trader
wants to sell, the reason could be that the trader knows something
negative that the dealer does not. However, it could also be that the
trader is selling for reasons unrelated to fundamentals (such as when
the trader is a pension fund selling to meet pension obligations). The
dealer cannot identify which is the case for any individual trade. But
if a preponderance of sales occurs over time, the dealer adjusts his or
her beliefs and prices downwardit is unlikely so many sales could
occur for non-fundamental reasons. In this way the dealer's price
gradually embeds the private information in order ow.
The Model
The following simple version of the Glosten and Milgrom (1985)
model islike the Kyle modela workhorse within the microstructure literature. The model has three familiar trader types: riskneutral informed traders, a risk-neutral dealer, and uninformed
traders. There is a single risky asset, whose terminal payoff is either
high V H or low V L . All informed traders initially observe whether
the terminal payoff is high or low. Initially, the dealer knows only
the unconditional probability of V H and V L , which we denote as p
and 1 p, respectively.
The model organizes trading as a sequence of bilateral trading
opportunities. Each trading opportunity involves a single potential
trader selected at random from an unchanging pool of potential
traders. The dealer knows that q percent of the traders in the pool are
informed, and 1 q percent are uninformed. The dealer then presents the selected trader with bid and offer prices that are good for
a single transaction of one unit. The selected trader can buy at
the offer, sell at the bid, or choose not to trade (gure 4.7 presents
the model's time line). For simplicity, I will assume that when an
uninformed trader is selected, the probability of buying and selling
are equal at 0.5 (arising, for example, from idiosyncratic hedging
demands).

Figure 4.7
Timing of a single trade in sequential-trade model.

Theoretical Frameworks

89

This trading protocol is more elaborate than that for the previous
two models. The sequential-trade model's more explicit dealership setting requires more structure (in much the same way that the
Kyle model's explicit auctioneer required more structure than the
rational expectations model). This sequential protocol is what allows
us to analyze trades individually. It also prevents the informed
traders from trading aggressively when prices do not yet reect all
information.
The dealer's pricing rule is the essence of the model. Two key features pin it downrisk neutrality and zero prots. (The zero prot
condition can be supported by incipient entry of competing dealers,
as in the Kyle model.) These features mean the dealer will provide
bid and offer prices to the next trader that conform to
Bid EV j next trader sells
Offer EV j next trader buys:

4:12

These expectations internalize the effect that a sell/buy of one unit

would have on the dealer's expectation of the payoff V. They are
therefore regret free in the sense that the dealer does not regret having done (or not done) the transaction after the fact in either case.
Figure 4.8 summarizes the model's key features.
The tree diagram in gure 4.9 claries how these expectations are
determined. The protocol of the model provides just enough struc-

Figure 4.8
Summary of sequential-trade model.

90

Chapter 4

Figure 4.9
Probability of different trade types: Sequential-trade model. There are eight possible
trade types. The probability of each occurring appears in the far right column (the sum
of the eight probabilities equals 1). Each of these probabilities has three components.
First, nature produces either a high payoff value V H or a low payoff value V L , with
probabilities p and 1 p respectively. Then a trader is selected from a pool who is
either informed or uninformed, with probabilities q and 1 q respectively. Informed
traders know whether the realized value is V H or V L . Finally, the selected trader
chooses to buy or sell. If the selected trader is uninformed he buys with probability 12
and sells with probability 12. If the selected trader is informed he buys with probability
1 if payoff value is high and sells with probability 1 if payoff value is low.

ture to make this easy. There are eight possible trade types, which
are the product of two different states of the world, two different
trader types, and two different transaction directions 2  2  2. The
far right column shows the probability of each trade type occurring.
Because these eight types span all possibilities, the sum of the eight
probabilities must equal one. Each of these probabilities has three
components. First, nature produces either a high payoff value V H or
a low payoff value V L , with probabilities p and 1 p, respectively.
Then a trader is selected from the pool who is either informed or
uninformed, with probabilities q and 1 q, respectively. Informed
traders know whether the realized value is V H or V L . Finally, the

Theoretical Frameworks

91

selected trader chooses to buy or sell. If the selected trader is uninformed he buys with probability 12 and sells with probability 12 (one
can think of the uninformed in this model as trading for idiosyncratic
hedging purposes). If the selected trader is informed he buys with
probability 1 if the realized value is high and sells with probability 1
if the realized value is low.
Setting Bid and Offer Prices
Let us walk through an example of how the dealer would set the
offer price shown in equation (4.12). We can expand that equation to
Offer EV j buy V L ProbfV L j buygV H ProbfV H j buyg;

4:13

where ProbfV L j buyg denotes the probability that V V L conditional on the next trader choosing to buy. The dealer knows the values of V L and V H ; he just does not know which nature has selected.
We need expressions for ProbfV L j buyg and ProbfV H j buyg. These
two probabilities are calculated using the same method. Going
through the rst only, ProbfV L j buyg, is therefore sufcient for understanding the complete solution.
To calculate the probability ProbfV L j buyg, we need to use a
handy rule called Bayes Rule. The appendix to this chapter provides
a review of Bayes Rule and its underlying intuition. Let me simply
assert here that applying Bayes Rule provides us with the following
expression:
ProbfV L j buyg

ProbfV L gProbfbuy j V L g
ProbfV L gProbfbuy j V L g ProbfV H gProbfbuy j V H g

4:14

The dealer knows ProbfV L g and ProbfV H g are p and 1 p respectively. The only other components to the dealer's problem are
Probfbuy j V L g and Probfbuy j V H g. These are easily determined
from the probabilities in the righthand column of the tree diagram.
For example, Probfbuy j V L g is the sum of the buy probabilities in the
fth and seventh cells, divided by 1 p. The 1 p comes from the
sum of the probabilities in the fth through eighth cellsthe probability of V L .
The determination of expectations and prices for subsequent periods follows the same procedure. The dealer uses all available infor-

92

Chapter 4

mation to update these expectations over time in a Bayesian manner.

The resulting prices follow a Martingale with respect to that information. (Following a Martingale means that unless one had more
information than the dealer, the best estimate of the future price is
the current price.25) Note too that over time, trading and price adjustment occur even though there is no news beyond that in the
trading process itself. This point should speak to anyone who has
observed hours of frenzied FX trading with no apparent changes in
macro fundamentals (see also Romer 1993 on this issue).
Discussion
Like the Kyle model, the sequential-trade model has been used
extensively within microstructure and has been extended in many
directions (e.g., allowing different trade sizes, allowing the possibility of no new information, and allowing the intensity of trading
by informed traders to vary with the amount of information).26 It is
designed to capture how priceboth bid and offerevolves in a
dealership market as orders arrive in sequence. This ability to characterize the spread is a big advantage as a guide for empirical work.
Like any model, though, the sequential-trade model has its drawbacks. Three in particular deserve mention.
Order Arrival Process
The sequential-trade model's ction of random draws from a pool of
potential traders is a powerful simplifying assumption. The cost, of
course, is that it does nothing to explain the process of order arrival.
The timing and sequencing of trades are important choice variables
in actual markets.
No Strategic Behavior
This drawback is directly related to the rst. Informed traders in the
basic sequential-trade model have no ability to vary their trading
frequency. In the Kyle model, in contrast, strategic variation of trading intensity is the focus. This point applies to uninformed traders
as well: Though the uninformed have no information advantage, in
actual markets they may be able to lower their trading costs by
trading at particular times (see, e.g., the Kyle-type model analyzed
by Admati and Peiderer 1988).

Theoretical Frameworks

93

No Dealer Risk Aversion

Like the Kyle model, the sequential-trade model is a pure information model. Because the dealer is risk neutral (but has no monopoly
power), he always sets price at his conditional expectation of V. It is
therefore a story purely about expected future payoffsthere is no
possibility of price effects from inventory or imperfect substitutability. For FX markets, however, this does not square with empirical
work showing that dealers do adjust prices based on their inventory.
In addition, in multiple-dealer markets, if the intensity of inventory
control trading by dealers varies over time, this will affect the composition of the ``pool'' of potential trading partners.
4.4

Multiple Dealers: The Simultaneous-Trade Model

This section of chapter 4 is more extensive than the previous three for
several reasons. First, the simultaneous-trade model, with its multiple-dealer structure, is a better proxy for FX market institutions than
the other models, so it provides perspective on FX trading that the
others cannot provide (see also Lyons 1997a). Second, many papers
in the literature extend the simple versions of the three models presented above. Readers therefore have ample references for these
models. Because the simultaneous-trade model is more recent, references are few. Indeed, the theory of multiple-dealer trading is, on the
whole, rather underdeveloped.27 Third, I employ variations on the
simultaneous-trade model to guide much of the empirical analysis
in later chapters; a solid introduction to how it works will provide
background for understanding those empirical models.
Of the FX market features reviewed in chapter 3, there are three in
particular that the preceding models do not address, but which the
simultaneous-trade model is designed to capture. The rst of these
is interdealer trading. Most trading in the foreign exchange market
is interdealerroughly two-thirds. Single-dealer (and dealerless)
models cannot address the causes and implications of such trades.
The second feature is the way in which private information arises.
Customer order ow in FX is an important source of information
advantage that accrues to dealers. That said, it is not the case that all
information about payoffs is revealed to the dealers who enjoy this
advantage. Dealers therefore have less information than the informed
traders in the preceding single-dealer frameworks. The third feature

94

Chapter 4

is dealer risk aversion, which makes inventory imbalances less

acceptable to dealers. FX dealers manage inventory quite intensely.
The dealers in the Kyle and sequential-trade models, however, are
risk-neutral. Risk neutrality is a powerful simplifying assumption for
analyzing information revelation, but it also rules out the possibility
that information revelation and dealer risk management interact.
This interaction can be analyzed within the simultaneous-trade
model.
Modeling multiple-dealer interaction necessarily involves game
theory. The simultaneous-trade model draws from a class of games
called simultaneous-move games (versus sequential-move games).
There are two key effects of the simultaneous-trade feature. The
rst is to constrain dealers' conditioning information: Dealers cannot
condition on other dealers' trades when they occur simultaneously.
The second key effect of the simultaneous-trade feature is that it
introduces inventory shocks. Because dealers cannot condition on
incoming trades when simultaneously placing their own orders,
these incoming orders can alter the dealers' inventory in unpredictable ways. The inventories these shocks engender are integral to
the hot potato phenomenon described in chapter 1. Our three other
microstructure models cannot capture this hot potato phenomenon
because undesired inventories do not arise. (They do not arise because in the two single-dealer models the dealer is risk-neutral, and
in the rational-expectations model all trades are conditioned on the
market-clearing price.) Empirically, undesired inventories are crucial
for understanding the immense trading volume in FX, a point echoed
by Flood (1994, 147): ``The large volume of interbank trading is not
primarily speculative in nature, but rather represents the tedious
task of passing undesired positions along until they happen upon a
marketmaker whose inventory discrepancy they neutralize.''
Insights
Multiple-dealer models of the FX market present many challenges,
but also produce insights not accessible with the simpler singledealer models. Relative to the two preceding single-dealer models,
three of the most important insights from the simultaneous-trade
model are as follows:

Theoretical Frameworks

95

1. Dealer inventories and customer order ow are sources of private

information that drive dealer speculation.
2. Private information and strategic dealer behavior have interacting
effects that reduce the information revealed by price.
3. Private information and dealer risk-aversion have interacting
effects that reduce the information revealed by price.
The model provides further perspective on these, but let me offer
a few initial thoughts. To appreciate the signicance of insight (1),
recall that within exchange rate economics it is common to view
exchange rates as entirely determined by publicly available macroeconomic information. A corollary is that private information is
irrelevant. Taken literally, this view has profound implications for
the plausibility of various modeling approaches. The simultaneoustrade model provides examples of types of information that are not
publicly available and shows that this information can help individuals forecast prices (e.g., the dispersed information about current
account balances noted in chapter 2). This information is different
than in the three preceding models. In particular, there are no traders
with the type of ``inside'' information associated with the preceding
models. In the FX market, this type of concentrated information
about, say, future interest rates, is not plausible. Instead, superior
information in the simultaneous-trade model concentrates itself at
the dealer level and derives from the order ow each dealer observes.
It is therefore the dealers who have information advantages in this
model, which they use as a basis for speculation. Because the number
of dealers is not large, strategic behavior naturally arises. The strategic trading is between dealers, however, not between dealers and
an informed nondealer as in the Kyle model.
Insight (2)that private information and strategic dealer behavior
interactis a fascinating aspect of multiple-dealer markets. In effect,
dealers play two conicting roles in the market. On one hand, they
are information intermediaries, standing between the information
in their customer order ow and subsequent market prices. On the
other hand, they are also rational speculators. Their speculative behavior distorts information intermediation, reducing the informational efciency of price.
Insight (3) relates directly to hot potato trading, which arises out
of dealer risk aversion. When coupled with private information,
though, this trading is not innocuousit reduces the information

96

Chapter 4

revealed by price. This result follows from two underlying points.

First, hot potato trading reduces the information in interdealer trades
because it lowers those trades' signal-to-noise ratio. Second, it is
precisely these interdealer trades that determine price: price cannot
depend directly on customer-dealer trades because in the FX marketand as modeled in the simultaneous-trade modeldealers
cannot observe each others' customer trades (i.e., transparency of
this type of trade is zero, whereas transparency of interdealer trades
is not zero).
Consider an example that illustrates the rst of these two points
that hot potato trading lowers the signal-to-noise ratio in interdealer
trades. There are two dealers, A and B. Dealer A receives two signals
of the value of the asset (denote this value V ). One of dealer A's
signals is in the form of a customer trade, C, and the other is a signal
S that is not directly related to order ow (but may be indirectly
related; for example, it may be a signal that helps dealer A better
interpret the information in the customer trade C). For simplicity,
suppose the best estimate of V conditional on these signals is C S.
Now suppose A trades with B, and A's rule for determining demand
is D A 1 bC S, where b is known by B. Think of b as the additional weight put on the customer trade C to reect the laying off of
that trade to reduce inventory risk (the unit weights on C and S reect speculative demand based of information advantage). Because B
knows the weight b and observes A's demand D A , if b 0 then D A
communicates C S exactlythe best estimate of V. In contrast, if
b 0 0, then B does not learn the best estimate of V exactly; the demand D A is now a noisy signal of C S. In the simultaneous-trade
modelas in the preceding modelsall dealers know the weights in
the optimal trading rule. Hot potato passing of inventories causes
these weights to differ from weights based on fundamental information alone. This makes signal extraction less precise, thereby reducing the information in price.
The Model
The model is a two-period game with N dealers. Like previous
models, this model includes one risky asset. Each dealer has an
equal-sized customer base composed of a large number of competitive customers (e.g., fund managers, speculators, corporate treasurers,
liquidity traders, central banks, etc.). All dealers and customers have

Theoretical Frameworks

97

Figure 4.10
Timing in the simultaneous-trade model.

identical negative exponential utility dened over nominal wealth at

the end of period two. Figure 4.10 provides the timing of the model
and introduces some notation.
Let us turn to the information available to dealers. Before quoting
in period one, each dealer (indexed by i) receives a private signal Si
and all dealers receive a common signal S.28 Both signals are distributed about Vthe nal payoff on the risky asset. After quoting in
period one, each dealer receives orders from his own customers that
aggregate to C i . These customer orders C i are distributed Normal
(0, sC2 ). I will use the convention that C i is positive for net customer
purchases and negative for net sales. Importantly, each dealer's C i
is not observed by other dealers (and will be a source of superior
information available to dealer i). The variables Si , S, and C i are all
independently distributed.
The rules governing the quotes Pit are outlined in the model summary in gure 4.11 (under Institutions). The most important to note
is that there are no spreadsquotes are single prices.29 Note that
simultaneous moves are consistent with interdealer FX transactions
being initiated electronically rather than verbally, providing the capacity for simultaneous quotes, trades, or both. The assumption that
quotes are observable is tantamount to assuming that quote search is
costless. The last quoting rule (that dealers must take the other side
of any trade at the single quote) prevents a dealer from exiting the

98

Chapter 4

Figure 4.11
Summary of simultaneous-trade model.

game at times of informational disadvantage. Regarding this last

rule, in actual FX markets dealers who choose not to quote during
trading hours are viewed as breaching the implicit contract of quote
reciprocity and are punished by other dealers (e.g., breaches are met
with subsequent refusals to provide quotes, or by quoting large
spreads to that dealer).
The rules governing interdealer trading are also outlined in the
model summary in gure 4.11. Let Tit denote the net of outgoing
interdealer orders placed by dealer i in period t; let Tit0 denote the net
of incoming interdealer orders received by dealer i in period t, placed
by other dealers. Simultaneous and independent trading generates
an important role for Tit0 as an unavoidable disturbance to dealer i's
position in period t that must be carried into the following period.
For consistency with our previous denition of C i as positive for
net customer purchases, orders will always be signed according to
the party that initiates the trade. Thus, Tit is positive for dealer i
purchases, and Tit0 is positive for purchases by other dealers from
dealer i. Consequently, a positive C i or Tit0 corresponds to a dealer i
sale. Letting Dit denote dealer i's desired position in the risky asset
(net of customer and dealer orders received), we can write each
period's interdealer trade as follows:

Theoretical Frameworks

99

Ti1 Di1 C i ETi10 j W Ti1

4:15

Ti2 Di2 ETi20 j W Ti2 Di1 Ti10 ETi10 j W Ti1 ;

4:16

where W Ti1 and W Ti2 denote dealer i's information sets at the time of
trading in periods one and two, respectively. From equation (4.15) it
is clear that customer purchases (sales) must be repurchased (resold)
in interdealer trading to establish the desired position Di1 (we assume an initial position of zero). In addition, to establish Dit , dealers
must factor the expected value of the interdealer orders received Tit0
into each period's trade. Turning to equation (4.16), in period two,
the realized period-one position must be reversed, which has the
three components: Di1 , Ti10 , and ETi10 j W ti1 (recall that Ti10 > 0 corresponds to a dealer i sale in period one). The term Ti10 ETi10 j W ti1 is
the unexpected incoming orderthe inventory shock.
At the close of period one, dealers observe period-one interdealer
order ow:
X1

N
X

Ti1 :

4:17

i1

This summation over the signed interdealer trades Ti1 measures the
difference between buy and sell orders (i.e., net buying) because Ti1 is
negative in the case of a sale. The empirical analogue of X is the
signed order ow information communicated by interdealer brokers,
a statistic common to all dealers. Specifying this order ow statistic
X as an exact measure (i.e., without noise) maximizes the transparency difference across trade types because the model's customerdealer trades are unobservable to noncounterparties. As noted in
chapter 3, trades between dealers and their customers do indeed
have zero transparency. However, the actual transparency of interdealer trades is not complete. I address the role of noise in this
equation in Lyons 1996a, a paper that examines the effects of changing transparency.
Before proceeding, recall that one of the objectives of this model is
to capture the effect of different transparency levels across trade
types, customer-dealer versus interdealer. Because customer-dealer
trades are not generally observable, they are not aggregated in price
until they are later reected in interdealer tradeswhich are observable. The result is a ``two-stage'' process of information aggregation, with interdealer trading as the second, crucial stage (see also
Gersbach and Vogler 1998).

100

Chapter 4

Dealer Objectives and Information Sets

Each dealer determines his or her quotes and demands for the risky
asset by maximizing a negative exponential utility function dened
over nominal wealth at the close of period two. Letting Wit denote
the end-of-period t wealth of dealer i, we have:
MAX

fPi1 ;Pi2 ;Di1 ;Di2 g

EexpyWi2 j Wi

4:18

s.t.
Wi2 Wi0 C i Pi1 Pi10 Di1 ETi10 j W Ti1 Pi20 Pi10
Di2 ETi20 j W Ti2 V Pi20 Ti10 Pi20 Pi1 Ti20 V Pi2 ;
where Pi1 is dealer i's period-one quote, a prime ( 0 ) denotes a quote
or trade received by dealer i, and V is the payoff value of the risky
asset at the end of period two. The second term in nal wealth is the
dealer's roundtrip prot on his customer orderif C i is positive,
then the dealer sold to the customer at Pi1 , and bought the same
amount back from other dealers at Pi10 . The third term in nal wealth
is the capital gain on the dealer's period-one speculative and hedging demands (recall that ETi10 j W ti1 is the dealer's hedge against incoming orders). The fourth term in nal wealth is the same as the
third term, but dened for the second period. The last two terms in
nal wealth capture the position disturbances that arise due to the
simultaneous trade feature: dealer i does not know the value of
the order he receives Ti10 when he chooses his trade Ti1 (similarly for
period two). The conditioning information Wi at each decision node
(two quotes and two trades) is as follows:
Quoting Pi1 : fC i ; Si ; Sg
Trading Ti1 : fC i ; Si ; S; P11 ; . . . ; PN1 g
Quoting Pi2 : fC i ; Si ; S; P11 ; . . . ; PN1 ; Ti1 ; Ti10 ; Xg
Trading Ti2 : fC i ; Si ; S; P11 ; . . . ; PN1 ; Ti1 ; Ti10 ; X; P12 ; . . . ; PN2 g
Equilibrium Quoting Strategies
In this model, rational quotes must be the same across dealers at any
given time. If not common across dealers, arbitrage opportunities
would exist (quotes are single prices, are available to all dealers, and

Theoretical Frameworks

101

are good for any size). Further, given that quotes are common, then
they must be based on common information. Because there is only
one piece of common informationthe signal Sthis yields a linear
quoting strategy:
P1 L s S

4:19

where the constant L s is a signal-extraction coefcient that produces

an unbiased estimate of the value V conditional on S.
Given that dealers' quotes are common, I need to introduce a rule
for how interdealer trades are assigned. It is simplest to think of the
dealers as arranged in a circle and to assume that each dealer directs
his trade to the dealer on his left. This assumption is not substantive.
Relaxing it, for example, by allowing a split into m < n equal fractions is straightforward as long as m is known.30 Given this assumption, then, each dealer receives exactly one order in period one
from another dealer, from the dealer on his right. This order corresponds to the position disturbance Ti10 in the dealer's problem in
equation (4.18).
Period-two quotes are also common across dealers. In this case
they depend on two pieces of common information:
P2 L 2 S LX X:

4:20

The no-arbitrage argument for why these quotes must be common

across dealers is the same as that for period one. And like P1 , P2 is
pinned down by public information because common quotes cannot
be based on information that is not common. The public information
available in period two includes the interdealer order ow X in addition to the public signal S. Intuition for the effect of X on P2 is important. Because X is the sum of interdealer orders, it conveys some
information embedded in those orders that is not yet public (i.e., X is
correlated with the individual customer orders C i and the private
signals Si ). Order ow X does not convey all of the information in the
C i 's and the Si 's, however, so P2 does not fully impound all of that
information. Any private information not reected in P2 becomes a
basis for speculative dealer demands in period two.
Equilibrium Trading Strategies
Given the quoting strategies above, the optimal trading strategy in
periods one and two for all dealers is the following:

102

Ti1 b 11 C i b21 Si b31 S b41 P1

Ti2 b 12 C i b22 Si b32 S b42 Ti10 b52 X b62 P2 :

Chapter 4

4:21

Though busy, these two trading rules have a convenient linear

structure that is familiar from the earlier models, deriving from the
assumption of exponential utility and Normally distributed returns.
For readers interested in a detailed derivation, see Lyons (1997a).
Hot Potato Trading
Due to the simultaneous-trade feature, the model produces hot
potato trading of the kind introduced in chapter 1. Inventory imbalances are passed from dealer to dealer, independently of whether
or not they offset the imbalance of the receiving dealer. This hot
potato property arises because dealers have no knowledge, ex-ante,
of which dealers are long and which dealers are short, and they
cannot condition on other dealers' trades in any given trading round.
Consequently, perfectly efcient risk sharing between dealers is not
possible. This property is in sharp contrast to the rational expectations model (which is the only model of the preceding three that is
relevant in this respect because agents in that model are also risk
averse). In the rational expectations model, inventory imbalances
are impossible at the close of trading because agents condition on
market-clearing pricesthere are no surprises.
I view these imbalances as an important dimension of any model
that addresses trading in FX. By producing hot potato trading in
equilibrium, the simultaneous-trade model provides a rationale for
the skeptic who asks ``Why doesn't price simply adjust?'' In equilibrium, price does adjust, but only to reect the aggregate inventory
imbalance. Idiosyncratic imbalances remain because arbitrage limits
idiosyncratic dispersion in price, and dealers cannot condition on
other dealers' trades at ultra-high frequencies. Though people have
attributed the extraordinary volume in FX to the repeated passing of
inventory imbalances, this process had never before been modeled in
the context of rational agents.
The Information in Price
The hot potato passing of inventories is not innocuous; rather, it
hampers information aggregation because the passing of inventories

Theoretical Frameworks

103

dilutes the information content of order ow, which reduces the information in the period-two price (insight 3 from this model). This
is true even if dealers behave competitively. To see why, recall that
information is aggregated on the basis of signal extraction applied to
X, the interdealer order ow. The greater the noise relative to signal,
the less effective the signal extraction. Passing hot potato liquidity
trades increases the noise in the interdealer order ow.
On top of this hot potato effect, strategic dealer behavior causes a
further reduction of information in period-two price (insight 2). This
result comes from dealers recognizing that their own orders will
have a subsequent price impact. This induces each to alter speculative demand to prot from the forecastable effect on price. These
altered speculative demands exacerbate the reduced efciency of
signal extraction due to hot potato trading.
Let me return to the earlier example I used to illustrate these
insights (2) and (3), using the model's specics to make it more concrete. Suppose that dealer i receives a combination of signals and
customer order ow fS; Si ; C i g f0; 0; 1g, where S is the public signal, Si is the private signal, and C i is dealer i's customer order (the
latter being uncorrelated with the payoff V ). Because the value of the
public signal S pins down the rst period price P1 at L s S, that S 0
implies that P1 0. Further, because both signals S and Si are zero,
the expected value of the payoff EV j S; Si ; C i 0 and the incoming
interdealer trade that dealer i expects ETi10 j S; Si ; C i 0. (The latter
relation holds because the customer orders C i each dealer receives
are uncorrelated across dealers; thus, using equation (4.21), dealer i's
realization of fS; Si ; C i g gives him no information useful for forecasting the incoming interdealer trade.) Despite the realization
f0; 0; 1g, however, dealer i's speculative demand is not zero. To see
this, note that dealer i's interdealer trade Ti1 includes his customer
order C i one-for-one due to inventory control. Note alsofrom the
pricing rule in equation (4.20)that P2 will move in the same direction as the sign of interdealer ow X, one component of which is Ti1 .
Dealer i can therefore forecast the effect that the passing along of his
customer order C i will have on the price P2 . This induces him to take
a positive (long) position in order to prot from the forecastable
market move. Part of what dealer i is able to forecast is the market's
misinterpretation of dealer i's interdealer trade Ti1 : The market can
only presume that a positive Ti1 partly reects a positive Si , which, in
this example, it does not. Note that the added weight that dealer i

104

Chapter 4

puts on his customer order C i in determining his speculative demand

is not an artifact of the specic realization f0; 0; 1g: The weights in the
trading rule are not realization dependent. To summarize, the dealer
alters his speculative demand to prot from the forecastable error in
price. This behavior makes it more difcult for the market to extract
information from order ow.
Discussion
The simultaneous-trade model captures several FX market features
not captured in the three other models of this chapter.31 At the same
time, introducing multiple dealers makes the analysis much more
complicated. Tractability necessitates leaving out some important
features. Three in particular deserve mention.
No Spread
Like the Kyle and rational expectations models, this model is unable
to characterize the bid-ask spread. It is straightforward to add a
spread or commission for the initial customer trade (which provides
a means of determining the number of dealers endogenously). Adding a spread to interdealer trading, however, means that potential
arbitrage no longer rules out price dispersion. This would add considerably to the model's technical complexity.
No Signaling via Price
In the model, the only way dealers learn about others' private information is through observing interdealer order ow, so private information is reected in price only if rst reected in interdealer orders.
In theory, dealers might also communicate information through their
quoted prices. Signaling via quotes does not arise in this model because dispersion in price with no spread would produce arbitrage.
As an empirical matter, individuals in multiple-dealer markets
surely learn about private information through both order ow and
dispersion in prices (see Chakrabarti 2000).
Brokered Trading Is Unmodeled
Brokered interdealer trading is an important category of trading in
the foreign exchange market but remains unmodeled. One might be
tempted to sweep this under the models-cannot-capture-everything
rug. The omission is more egregious here, though, because the order

Theoretical Frameworks

105

ow signal that drives the model parallels most closely that provided
by brokered trading. How individuals choose a particular trading
mechanism in markets where there is choice is a burgeoning area of
research (for empirical work in FX, see Bjonnes and Rime 2000; for
theoretical work on modeling brokers, see for example Werner 1997.)
Direct versus brokered interdealer trading in FX has particularly
interesting implications given the difference in transparency across
trading mechanisms.
4.5

Appendix: A Tool Kit

There are many standard results useful for working with microstructure models. In this appendix I present three of the most useful.
Before launching into details, however, I begin at a more general
level by providing some perspective on the so-called CARA-Normal
framework, a norm within microstructure. In this framework, utility
exhibits constant absolute risk aversion (CARA) and random variables are Normally distributed. Then I turn to the three specic
results: the rst is that in the standard CARA-Normal framework,
preferences are mean-variance, which makes them particularly easy
to work with; the second is that in the CARA-Normal setting, demands for risky assets take a particularly simple form; third, I show
that in the CARA-Normal setting, conditional expectations, too, take
a particularly simple form. For each of the last three results I provide
derivations, with the intention to alert readers that the techniques
used in the derivations are themselves quite useful for solving more
complex models. Another good reference for useful tools used in
microstructure analysis is O'Hara 1995, in particular the appendix to
chapter 3, where she addresses Bayesian learning with discretely
distributed random variables.
The CARA-Normal Setting: Exponential Utility and Normal Returns
The negative exponential utility function I introduce in equation (4.1)
of this chapter is a standard feature of trading models. It produces
simple expressions for risky asset demand as long as returns are
conditionally Normally distributed. The word conditionally is important: it is not enough that all random variables in the model are
Normally distributed. There are many cases in which Normally distributed random variables do not produce conditionally Normal

106

Chapter 4

returns. Examples of non-Normal returns include products of (discrete) Normal random variables and mixtures of Normal random
variables. (An example of a mixture of Normals X1 and X2 is pX1
1 pX2 , where p is a probability between zero and one. Specications with a mixture of Normals are intuitively appealing, but technically more complex.) Trading model specications rely heavily on
the fact that the sum of two Normally distributed variables is itself
Normally distributed.
It may be helpful to review the denition of the coefcient of absolute risk aversion, y:
yW

U 00 W
;
U 0 W

where the utility function UW is dened over wealth W, and `` 0 ''

denotes derivative. With risk aversion, U 00 W < 0, which implies
that yW > 0 (because U 0 > 0). When the coefcient of absolute risk
aversion is constantthat is, independent of wealth Wthen we
have CARA.
Utility functions that satisfy CARA can be represented by the
negative exponential
UW expyW;
where here y is a positive constant, and therefore not a function of
wealth (assumed equal to 1 in equation 4.1).
For those who prefer a graphic representation for this utility specication, note that the function UW expW is simply the
graph of UW expW rotated twice: once around the x-axis to
account for the minus sign in front of the exponential, and then again
around the y-axis to account for the minus sign in front of wealth W
(see gure 4.12). The curvature of this graph is a measure of risk
aversion; it increases with the coefcient y.32
I should note before moving on that a common critique of CARANormal modeling is that the assumption of a Normally distributed
payoff on the risky asset implies prices are not bounded. This implies unlimited downside liability, which may not be appropriate for
analyzing certain issues. Also, the specication implies that prices can
be negative. However, by increasing the mean of the payoff 's distribution, the probability of a negative price becomes arbitrarily small.
Let us return to the three specic results noted above. I will state
all three at the outset, followed by derivations.

Theoretical Frameworks

107

Figure 4.12
Negative exponential utility.

Result 1:

Mean-Variance Preferences

In the CARA-Normal setting, with utility UW dened over wealth

W, and W @ Nm; s 2 , agents' preferences can be represented as a
function of the mean and variance of wealth:
gm; s 2 m 12 ys 2 ;
where y is the (constant) coefcient of absolute risk aversion.
Result 2:

Risky Asset Demand

In a one-period CARA-Normal setting, with one risky asset whose

payoff V is distributed Nm; s 2 , agents' demand for the risky asset is
given by
DP

m PR
;
ys 2

where P is the price of the risky asset, DP is the number of units

demanded, and R is the gross return on the riskless asset (i.e., 1
plus the net return; the net return is set equal to 0 in most work).
Note that this demand is not a function of the level of wealth. An
implication is that redistributions of wealth do not affect aggregate

108

Chapter 4

risky-asset demanda useful property because it means that one

does not have to keep track of the wealth of individual participants
when modeling.33
Result 3:

Conditional Expectations

With Normally distributed random variables, conditional expectations are Normally distributed and take a convenient form. Specically, dene the following:
y y e0
x i y ei

i 1; . . . ; n;

where the variable y is the variable of interest and the variables xi are
signals of y. Let each ei , i 0; . . . ; n, be distributed independently
and Normally N0; si2 . Then
Ey j x1 ; . . . ; xn

ys02 x1 s12


xn sn2

;
s02 s12


sn2

and
V y j x1 ; . . . ; xn

1
;
s02 s12


sn2

and these two moments fully characterize the conditional expectation because the conditional distribution is Normal. Note the simplicity: The mean of the posterior distribution is the sum of the prior
and signals, each weighted by its own precision (the inverse of its
own variance), divided by the sum of the precisions. The conditional
variance is just one over the sum of the precisions.
Derivation of Result 1:

Mean-Variance Preferences

To derive the result that preferences are mean-variance, start from

rst principles with
y
expyWfW dW;
EUW
y

where fW is the Normal density function:

!
1
W m 2
:
fW p exp
2s 2
2ps 2

Theoretical Frameworks

109

The trick to solving this is to factor the integral into two parts, one
part that depends on the random variable W and one part that does
not. Once factored, we will nd that maximizing EUW is as simple as maximizing the part that does not depend on W.
Now, substituting in the Normal density fW and rearranging
slightly, we get
y
 z 
1
p exp 2 dW;
EUW
2s
y 2ps 2
where
z W m 2 2yWs 2
and z can be rewritten as
z W m ys 2 2 yys 4 2ms 2 :
This is the key step in factoring the integral into two parts: the second of these two parts does not depend on random wealth W. We
can now write the problem as
 

ys 2
EUW exp y m
2
!
!
y
1
W m ys 2 2
p exp
dW ;

2s 2
y 2ps 2
where I have divided the term yys 4 2ms 2 by 2s 2 and multiplied
by 1 in order to bring this term outside the second exponential.
Notice that the integral on the right-hand side is just the integral over
a Normal density (with mean m ys 2 and variance s 2 ), and is therefore equal to one. This reduces the problem to maximizing the simple
exponential expression to the left of the integral. Given the behavior
of the negative exponential function illustrated in gure 4.12, it
should clear that this expression is maximized when m 12 ys 2 is
maximized.
Derivation of Result 2:

Risky-Asset Demand

Again, let us start from rst principles. We will nd that the solution
method is quite similar in spirit to the derivation of result 1. For
notation, recall that I dened V as the risky-asset payoff, with V @

110

Chapter 4

Nm; s 2 , P as the price of the risky asset, DP as the number of units

demanded, and R as the (gross) return on the riskless asset. Further,
let W0 and W1 denote beginning and nal wealth, respectively. With
these denitions, we can write the budget constraint as
W1 RW0 DP DV RW0 DRP DV:
Using the fact that V is Normally distributed, we can write the
problem as
y
1
expyRW0 DRP DV p
EUW1
2ps 2
y
!
V m 2
 exp
dV:
2s 2
Now, like in the derivation of result 1, we want to factor the integral
into two parts, one part that depends on the random variable (in this
case V) and one part that does not. Once factored, maximizing EUW
will be equivalent to maximizing the part that does not depend on V.
Rearranging the previous equation we have
EUW1 expyRW0 DRP


! !
1
V m 2
p exp
yDV dV :
2s 2
y 2ps 2

To factor the integral, we need to factor

V m 2
yDV
2s 2

into one part that depends on V and another that does not. Simple
algebraic manipulation yields
!  
 
Vm 2
1
V myDs 2 2
1
2myDy 2 D 2 s 2 :
yDV

2
2
2s 2
s2
The second of these two terms does not depend on the random payoff V. Accordingly, we can write the problem as
EUW1 expyRW0 DRP 1=22myD y 2 D 2 s 2
! !
y
1
V m yDs 2 2
p exp

dV :
2s 2
y 2ps 2

Theoretical Frameworks

111

As was the case for result 1, the integral on the right-hand side is just
the integral over a Normal density (with mean m-yDs 2 and variance
s 2 ), and is therefore equal to one. This reduces the problem to maximizing the following expression:
yRW0 DRP 1=22myD y 2 D 2 s 2 :
The rst order condition is
yRP my Dy 2 s 2 0;
or
DP

m PR
;
ys 2

which is what we set out to show.

Before moving on to the derivation of result 3, let us return to the
budget constraint because it claries the role played by the risk-free
asset in all the models of this chapter. Because risky-asset demand is
independent of wealth, in order for each individual's demand to be
feasible, then each individual must be able to borrow and lend at the
risk-free interest rate without constraint. If risky-asset demands were
constrained by borrowing constraints, trading limits, and so on, then
actual demand would differ from desired demand, which would in
general affect price.
Derivation of Result 3:

Conditional Expectations

Result 3 follows quite directly from Bayes Theorem for updating

beliefs with continuously distributed random variables. I will go
through it in the case of a prior and one signal; the extension to
multiple signals is straightforward. Suppose that the probability
density function describing an agent's prior belief about a payoff V is
Normally distributed, with mean m and variance sV2 . We can write
this density function as
!
1
V m 2
f V q exp
:
2sV2
2ps 2
V

Suppose one observes a random variable X (e.g., order ow) that is

jointly Normally distributed with V. What we are interested in is the
density of f V j X:

112

Chapter 4

Because X and V are jointly Normally distributed, the conditional

density of observing X given V is also Normally distributed (not
shown here). Let that conditional density function of observing X
given V be denoted gX j V:
!
1
X V 2
:
gX j V p exp
2sx2
2psx2
Now we simply need to apply Bayes Theorem. This theorem is
proved in most statistics textbooks, so I use it here without derivation:
f VgX j V
:
y gX j V f V dV

f V j X y

If we make the substitutions for f V and gX j V, we nd that

0
1
1
B
C
f V j X @qA
1
2psV2 sx2
 exp

1
2psV2 sx2 1

!

ms 2 Xsx2
V V2
sV sx2

!
:

This is simply a Normal distribution with

Mean

msV2 Xsx2
sV2 sx2

and
Variance sV2 sx2 1 :
This is what we set out to show: the mean of the posterior distribution is the sum of the prior and signal, each weighted by its own
precision (the inverse of its own variance), divided by the sum of the
precisions. The conditional variance is just the inverse of the sum of
the precisions.

Empirical Frameworks

This chapter covers key topics of empirical interest in FX microstructure research. Order ow features prominently, as it does in
later chapters that cover work that is more macro-oriented. This is
because, as we have seen, order ow plays a crucial role in microstructure models. The chapter begins with a survey of data sets; recent advances in available FX datadue largely to the advent of
electronic tradinghave opened new doors for empiricists. Then I
introduce the main empirical approaches within the eld of microstructure and how they are applied in FX. The closing sections
review important empirical results. In particular, I present results
bearing on central questions from the previous theory chapter such
as ``Does FX order ow convey private information?'' Conrming
empirical evidence supports this book's continuing use of information models. Another key question is, ``Does incomplete risk sharing
affect exchange rates?'' Conrming evidence here supports the use of
inventory models as well.
For perspective, consider how the data available for research in FX
microstructure have evolved over the last twenty years. The earliest
work used futures data because those data are available at high frequencies (Grammatikos and Saunders 1986; Jorion 1996.) In FX,
however, the futures market is much smaller than the spot market; it
is unlikely that a signicant share of price determination occurs there
(Dumas 1996). Moreover, early futures data sets did not have sufcient granularity to capture agent heterogeneity, a hallmark of the
microstructure approach. Work on the spot market itself grew in the
early 1990s with the availability of quotes on an intraday basis (specically, the indicative quotes from Reuters called FXFX). 1 These
quotes provide a quite accurate picture of price dynamics. More importantly, they also speak to heterogeneity issues because the names

114

Chapter 5

and locations of the quoting banks are available. Thus, a number of

interesting questions can be addressed that earlier data did not permit. The FXFX data did not, however, leave much room for direct
testing of theory because they provide no measures of order ow
(i.e., signed quantities traded). As we saw in previous chapters,
order ow's role in determining price is central to microstructure
theory. As order ow data became available, more direct tests became possible.2
More recently, empirical work in FX microstructure has entered an
exciting new phase. Within the last few years, the availability of
transaction data on a marketwide basis has opened important new
terrain. In equity microstructure, an analogous opening of new empirical terrain occurred in the early 1990s when then NYSE made
available its Trades, Orders, and Quotes (TORQ) and Trades and
Quotes (TAQ) databases (see Hasbrouck 1992 for more on these
equity databases). These new FX data allow us to test theory and measure price determination in ways not possible only ve years ago.
Nevertheless, it remains true that the microstructure approach to
exchange rates suffers from a lack of publicly available data. Decades
of macro data are readily accessible to researchers from sources like
Datastream or the International Monetary Fund's International Financial Statistics. Not so of microstructure data. It is my hope that this
book might stimulate further availability by promoting awareness
among ofcial and private institutions of the importance of these
data. For my part, I have collected some of the publicly available
data sets described in the next section for downloading from my
Web site: hwww.haas.berkeley.edu/@lyonsi. The Web site also provides information on how to obtain other data sets from various
sources that charge fees for public use.
5.1

FX Data Sets

Before reviewing FX data sets in detail, it may be helpful to look at

the big picture. These data sets can be grouped into three basic types,
corresponding to the three trade types reviewed in chapter 3:
1. customer-dealer trades (roughly 1/3 of total trading in $/euro
and $/yen in the late 1990s)
2. direct interdealer trades (roughly 1/3)
3. brokered interdealer trades (roughly 1/3)

Empirical Frameworks

115

Let us consider each of these in turn. Data on customer-dealer trades

have been unavailable until recently. These data are difcult to obtain because banks consider their trades with customers highly
condential. (Recall from chapter 3 that banks have no regulatory
obligation to disclose this information.) One recent data set that
makes progress on this front includes all the signed customer-dealer
trades at a single bank (Fan and Lyons 2000) over a sample period of
seven years. The bank (Citibank) is one of the top three banks in the
world in terms of trading volume, handling more than 10 percent of
worldwide customer order ow in the major FX markets. Another
available data set that relates to customer-dealer trading is FXFX. As
noted above, FXFX data are indicative quotes (no trades) that provide nondealer customers with real-time information about current prices. It is believed that dealers provide these prices to attract
customer business. Because the quotes are indicativemeaning they
are not committed pricesthey are in effect a form of advertising.
These indicative quotes lag the interdealer market slightly, and
spreads are roughly twice the size of interdealer spreads. Consequently, from a dealer's perspective, the price information the FXFX
quotes convey is dominated by the rm prices observable from interdealer brokers.
Data covering the second of the three categoriesdirect interdealer tradeswere limited until recently to the trades of individual
dealers (versus summarizing the trades of all dealers, or some large
subset). In contrast to customer-dealer trades, data on direct interdealer trades are available because dealers trade directly with one
another using a bilateral electronic trading system called Dealing
2000-1 (a Reuters product). Because quotes and trades are executed
electronically, an electronic record is produced that can be used for
empirical analysis. In this case, researchers get their data directly
from individual banks (the banks keep these electronic records on
hand, temporarily, to help resolve trading disputes, etc.). Evans
(1997) introduced a data set, obtained from Reuters, that includes all
the interdealer trades executed through D2000-1 during four months
in 1996. This data set spans a much larger slice of the total trading
pie than any preceding it.
Brokered interdealer trading has also recently become measurable.
The newfound accessibility of this data category corresponds to the
shift in brokered trading from voice-based systems to electronic systems (see chapter 3 for details). Early data from brokered trades over

116

Chapter 5

the voice-based system was rather spotty (e.g., Lyons 1995). More
recently, data sets have emerged from the records of the newer electronic brokers (e.g., Goodhart, Ito, and Payne 1996; Goodhart and
Payne 1996; Killeen, Lyons, and Moore 2000a). The main electronic
brokers in the major spot markets are EBS and Dealing 2000-2.
(Dealing 2000-2 is Reuters product for brokered interdealer trading;
Dealing 2000-1 is for direct interdealer trading.) These electronic
broker data have the advantage that they reect the activities of
multiple dealers who are trading simultaneously. At present, however, the available data sets do not span the full brokered interdealer
segment because they reect only the trading on either Dealing 20002 or EBS, but not both.
With this as background, gure 5.1 offers a model for organizing
one's thinking regarding available data.
The inner ring in gure 5.1 represents direct interdealer trading,
the most liquid part of the market. Dealing 2000-1 data is from this
inner ring. The middle ring contains brokered interdealer trading.

Figure 5.1
Three data groupings.

Empirical Frameworks

117

EBS and D2000-2 data are from this middle ring. The outer ring represents customer-dealer trading. Data from this ring come directly
from banks' own order ow records.3
The Inner Ring: Direct Interdealer Transactions and D2000-1
Data Set 1: Lyons 1995
The Lyons data set, like others from this inner ring of the market,
comes from Dealing 2000-1 trading records. For direct interdealer
trading in the major currencies, the Dealing 2000-1 system dominates
the market: it is believed to account for about 90 percent of the
world's direct interdealer volume.4 Trades on this system take the
form of electronic bilateral conversations initiated when one dealer
calls another dealer on the system, asking for a quote. Users of the
system are expected to provide a fast quote with a tight spread,
which is in turn accepted or declined quickly (i.e., within seconds).
Acceptance of a quote constitutes a trade.
The Lyons data set chronicles the activity of a dealer in the $/DM
market at a major New York investment bank. The sample spans
ve trading days over the week August 37, 1992 (from 8:30 a.m. to
on average 1:30 p.m., Eastern Standard Time). These data come from
the bank's own dealing records from the Dealing 2000-1 system. This
dealer uses Dealing 2000-1 for more than 99 percent of his nonbrokered interdealer trades.
Each record from the Dealing 2000-1 system includes the rst ve
of the following seven variables. The last two are included only if a
trade takes place:
1. a time stamp (to the minute)
2. which of the two dealers is requesting the quote
3. the quote quantity
4. the bid quote
5. the ask quote
6. the quantity traded
7. the transaction price
Note that these records provide rm bilateral quotes rather than
just transaction prices. They also identify which counterparty is the
aggressor. As such, they allow one to measure signed order ow

118

Chapter 5

precisely (rather than simply unsigned trading volume). Finally,

these records also include rm bilateral quotes that do not generate
trades. These nondealt quotes account for roughly 80 percent of all
quotes over the week.
In addition to these direct interdealer records from D2000-1, the
Lyons data set also includes data from the dealer's position sheets.
These sheets include all his trades (i.e., in addition to his direct
trades with dealers, they include his trades with customers and his
brokered trades with dealers). Because the position sheets include all
transactions, they provide an exact measure of the dealer's position
through time. For every trade these sheets provide the following:
1. the signed quantity traded
2. the transaction price
3. the trade type: direct, brokered, or customer
4. the counterparty name
Note that the bid-ask quotes at the time of the transaction are not
included on the position sheets. This part of the data set includes all
1,720 of the dealer's transactions over the week-long sample,
amounting to $7 billion, or $1.4 billion per day on average.
The following two gures help make these data more concrete.
Figure 5.2 provides an example of a Dealing 2000-1 communication,
with details provided in the notes to the gure. Figure 5.3 provides a
diagram of the data ow from the 2000-1 communications through
time. (I do not include a gure with the position sheet's structure
here because that was presented in chapter 3table 3.1.)
Data Set 2: Yao 1998a
The Yao data set is similar in its structure to that of Lyons but has
two important advantages. Like Lyons's, it includes the D2000-1
records and position sheets of a spot $/DM dealer who trades in
New York. The two advantages over the Lyons data are that the Yao
data set covers
1. 25 trading days (November 1 to December 8, 1995)
2. a dealer with substantial customer order ow
The Yao data set thus covers ve times the number of trading days.
In terms of daily average volume, the Yao and Lyons dealers are
similar: $1.5 billion versus $1.4 billion per day, respectively (note,

Empirical Frameworks

119

Figure 5.2
Example of Dealing 2000-1 Communication. The opening word, ``From,'' establishes
this as an incoming quote request (outgoing quote requests begin with ``To''); this information is crucial for signing trades. The caller's four-digit code and institution
name follow; ``GMT'' denotes Greenwich Mean Time; the date follows, with the day
listed rst; the ``1080'' at the end of line one is simply a record number. ``SP DMK 10''
identies this as a request for a spot DM/$ quote for up to $10 million; ``8891'' denotes
a bid of 88 and an offer of 91. Only the last two digits are quoted because it involves
fewer keystrokes; dealers are well aware of the rst digits of the pricesometimes
called the ``handle.'' From the conrmation that follows, one can see that the earlier bid
quote was in fact 1.5888 DM/$ and the offer quote was 1.5891 DM/$. The conrmation
also provides the transaction price and veries the transaction quantity. ``THKS N
BIFN'' is shorthand for ``thanks and bye for now.''

though, that roughly three years separate these data, and the $/DM
spot market grew in dollar terms by about 50 percent over those
three years).
The more important difference between the two data sets is the
composition of these trades; in particular, the extent to which the
two dealers trade with non-dealer customers differs dramatically.
About 14 percent of the Yao dealer's volume comes from customer
trades, whereas less than 1 percent of the Lyons dealer's volume
comes from customers. In this respect, the mix of trades executed by
the Yao dealer is more representative of market averages (in 1995,
roughly 25 percent of total trading was customer-dealer). The fact
that the Yao dealer is from a commercial bank rather than an investment bank helps explain why his customer order ow is so much

120

Chapter 5

Empirical Frameworks

121

heavier: for spot FX, commercial banks have a more natural customer base than investment banks.
Data Set 3: Evans 1997
The data set introduced by Evans (1997) covers direct interdealer
trading over a four-month period (May 1August 31, 1996). It contains time-stamped tick-by-tick data on all transactions for nine currencies against the U.S. dollar. These data were collected from the
Dealing 2000-1 system via a customized feed at the Bank of England.
Reuters keeps a temporary archive of all conversations on the system
to settle disputes; this archive is the source of these data. For every
D2000-1 transaction, the data set includes
1. a time-stamp
2. the transaction price
3. a bought or sold indicator (for signing the trade)
For condentiality reasons, Reuters was unable to provide the identity of the trading partners. The following list shows the nine currencies covered and the number of transactions for each over the
four-month period:
German Mark

257,398

Japanese Yen

152,238

Swiss Franc

67,985

British Pound

52,318

French Franc

20,553

Italian Lira

8,466

Figure 5.3
Diagram of data structure. Denitions: Q o is an outgoing interdealer quote (i.e., a
quote made) and, if the quote is hit, T i is the incoming direct dealer trade. Q i is an
incoming interdealer quote (i.e., a quote received) and, if the quote is hit, T o is the
outgoing direct trade. T b is a brokered interdealer trade. Brokered trades do not align
vertically with a quote because the data for brokered trades in the Lyons (1995) data
set come from the dealer position sheets, and the broker-advertised quotes at the time
of the transaction are not recorded. ``jj '' appears whenever a trade occurs; ``jj '' appears
whenever a nondealt quote occurs. The disjoint segment below the top timeline presents a hypothetical path of the dealer's position over the same interval; it changes with
trades only. The timeline at the bottom claries the denition of ``periods'' within the
Lyons (1995) analysis: incoming trades dene an event, not all trades (that model is
presented in chapter 5).

122

Chapter 5

Belgian Franc

5,256

Dutch Guilder

3,646

Danish Krona

1,488.

Several features of the data are noteworthy. First, they provide

transaction information for the whole interbank market over the
full twenty-four-hour trading day. This contrasts with earlier transaction data sets covering individual dealers over some fraction of the
trading day (Lyons 1995; Yao 1998a). Such a comprehensive data set
makes it possible, for the rst time, to analyze order ow's role in
price determination at the level of ``the market.'' (Recall that roughly
half of interdealer trading is direct, and, for the largest spot markets,
roughly 90 percent of the world's direct interdealer transactions take
place through the 2000-1 system.)
Second, these marketwide transactions data are not observed by
individual FX dealers as they trade. Though dealers have access to
their own transaction records, they do not have access to others'
transactions on the system. These transactions data therefore represent a history of market activity that market participants can only
infer indirectly (a kind of latent variable that is observed by the
econometrician). This provides extra power for econometric analysis
of how dealers learn about pressure for price adjustment.
Third, the data cover a relatively long time span (four months).
This is important for two reasons. First, with a longer time span and
multiple currencies, the data set can address exchange rate determination from more of an asset-pricing perspective than previously
had been possible. Second, the longer time span permits estimation
of intraday patterns in transaction activity with much more precision
than is possible from other transaction data sets. Existing studies
using the FXFX quote data have noted the importance of controlling
for these intraday patterns when analyzing other features of the data
(see, e.g., Baillie and Bollerslev 1991; Dacorogna et al. 1993, Guillaume et al. (1995), and Andersen and Bollerslev 1998).
Against all these positive features one needs to acknowledge two
drawbacks. First, unlike the Lyons and Yao data sets, the Evans data
set does not include rm bilateral quotes communicated between
dealers, but instead has only the transaction prices. This prevents
direct observation of spreads. As such, the Evans data set does not
help us understand the ``search'' process behind the trade. The fact
that bilateral quotes typically do not generate a transaction is infor-

Empirical Frameworks

123

mative from the perspective of analyzing price discovery. (As noted

above, in the Lyons data set, for example, roughly 80 percent of
all bilateral quote requests over D2000-1 did not generate a trade.)
The second drawback is that the Evans data set does not provide
the inventory positions of the trading dealers. This precludes direct
analysis of inventory control at the individual level. It also precludes
estimation of the type of ``structural'' models of dealer behavior that I
introduce in section 5.2.
The Middle Ring: Brokered Interdealer Transactions and D2000-2
The second type of trading for which data sets are available is
brokered interdealer trading. Like the Evans data set, these data sets
cover the trading activity of multiple dealers.
Data Set 4: Goodhart et al. 1996 and Payne 1999
The rst data set of this kind, introduced by Goodhart, Ito, and
Payne (1996), covers only a single day and accounts for only a small
fraction of daily trading volume. The authors' source was a screen
from Dealing 2000-2, a system that at the time was relatively new to
the market (June 1993). A later data set, used by Payne (1999), is also
from the Dealing 2000-2 system, but spans a longer period (one
trading week) and represents a larger fraction of brokered interdealer trading.
Recall from chapter 3 that interdealer broking systems provide
prices that are advertised to dealers generally. Dealers have a choice
of submitting a quotetypically on one side of the market onlyor
hitting the quote of another dealer. (Remember that brokers in the FX
market are interdealer only; customers cannot trade on these prices,
and cannot in general observe them because their access to screens is
restricted by EBS and Reuters.)5
Data Set 5: Killeen, Lyons, and Moore 2000a, b
More recently, a new data set has appeared that includes two years
of data from the larger of the two electronic interdealer brokers
EBS (at least, it is larger in the two largest FX markets, $/euro and $/
yen). This data set includes, on a daily basis, all the order ow passing through the EBS system from January 1998 to December 1999.
This is an important period because it spans both the year preceding
the launch of the euro and the year following the launch. The rst

124

Chapter 5

year of the sample includes order ow for $/DM, $/FF, and DM/FF.
The second year includes order ow for $/euro. These order ow
series are valuable for addressing hypotheses about the launch of the
new currency.
The Third Ring: Customer-Dealer Trades and Indicative Quotes
Data Set 6: Customer Orders (Fan and Lyons 2000)
This data set includes all the customer-dealer tradessigned according to the customer's directionreceived by Citibank over a period
of seven years (19931999). Citibank is among the top three banks in
the world in terms of trading volume (handling more than 10 percent
of worldwide customer order ow in the major FX markets).
Traditionally, banks have been reluctant to provide researchers
access to this kind of data, given its proprietary nature. One reason
Citibank was more accommodating is that the data are aggregated
to daily totals (i.e., individual transactions are not available). These
time-aggregated order ows are valuable for examining the link between order ows and lower frequency price dynamics.
Because these order ows come from underlying customers, they
provide a direct connection to the underlying sources of demand in
the economy. The data set includes the $/yen market and the $/euro
market (order ow for the euro before its launch is synthesized from
its constituent currencies). These data are the basis for the material I
present in chapter 9.6
Data Set 7: FXFX Quotes (Goodhart and Figliuoli 1991)
As noted above, FXFX is a second type of data that relates to customer-dealer trading. These data are indicative quotes that provide
customers (i.e., nondealers) with real-time information about current
prices. This source of data has many strengths. First, the data are
available over long periodsfrom the late 1980s to the present. Second, they are available for many currency pairs. Third, they are
available on a tick-by-tick basis (typically with thousands of ticks per
day in a major currency pair). Fourth, each quote is time-stamped to
the second. Fifth, the bank that input each quote can be identied.
These data also have several drawbacks. First, and most important, these data do not include order owthey are quotes only.
This precludes direct testing of models like those in chapter 4 because order ow is central to those models.7 (Some authors have

Empirical Frameworks

125

used the rate at which quotes arrive to proxy for trading volume
i.e., unsigned order owbut the proxy is rather loose; see Goodhart, Ito, and Payne 1996 and Evans 1997. Hartmann 1998a, on the
other hand, nds that quote arrival provides a good proxy for volume over longer horizons.) Second, the spreads in these indicative
quotes tend to be clustered at specic spread sizes, whereas rm
quotes in the market do not exhibit such clustering (see Goodhart,
Ito, and Payne 1996; Evans 1997). Third, a displayed indicative quote
cannot be replaced with a new quote until ve seconds have passed
(Evans 1997). Fourth, the raw data are rather noisy (Zhou 1996),
and even after applying the standard lter used in the literature
(Dacorogna et al. 1993), signicant outliers remain (Andersen, Bollerslev, and Das 2001).
Closing Thoughts
It is important to note that FX microstructure is evolving rapidly
in terms of available state of the art data. In the future, sample
lengths will surely be extended, and I anticipate further integration
of data from all three market segments (direct interdealer, brokered
interdealer, and customer-dealer).8 Thus, the data sets above are
perhaps best thought of as a guide to measurable variables and their
sources. At present, the customer-dealer segment is still thinly covered. This segment is important, though, because these trades represent the outside ``shocks'' to the interdealer trading at the market's
core.
I chose in this section to present FX data sets as they appear in the
literature. This provides easy reference to additional detail in the
corresponding papers. Another way to organize the data is according to information sets, that is, according to which participants observe which data in real time, and how data that are not available in
real time are disseminated. This would have been a more difcult
task. Consider, for example, the Evans 1997 data set that contains all
the direct interdealer trading from the 2000-1 system. In that case, it
is most appropriate to think of individual dealers as observing only
part of those data, the part that corresponds to their own trading
(both prices and order ow). But individual FX dealers do not observe these marketwide data in their entirety, and nondealer customers do not observe any of these data. Turning to the data sets on
brokered interdealer trading (EBS and D2000-2), to organize those

126

Chapter 5

data according to participants' information sets one would have to

distinguish between the observability of transaction prices and the
observability of order ow. Recall from section 3.1 that although
transaction prices from brokered trades are observable, the order
ow information that all dealers receive from these trades is noisy.
Nondealer customers do not in general observe these data. As for
data on dealer trading with their customers, these data are quite
bank-specic. Dealers at other banks do not observe them directly;
other bank dealers can at best make inferences based on a given
dealer's behavior and any information gleaned from interacting with
common customers.
5.2

Statistical Models and Structural Models

This section presents the main empirical approaches within microstructure and relates them to the FX market. My objective is to introduce readers who are new to microstructure to some of the empirical
tools. It is, however, exactly thatan introduction. It is not possible
to cover these approaches in depth. Ample references to key papers
will help guide those who desire more detail.
There are two broad approaches to empirical microstructure: statistical models and structural models. Not surprisingly, the split
mirrors that in economics more generally. All of these models are
designed to characterize the joint behavior of order ow and price.
The statistical models have the benet of modest data requirements.
They also are more generally applicable across different market
structures (i.e., dealer versus auction markets). Their lack of structure makes their reduced form results more difcult to interpret,
however. Structural models are grounded more explicitly in the economic decisions that dealers face. Accordingly, they are most appropriate for dealer markets, including the FX market.
There are three specic approaches I will cover here, the rst two
statistical, and the third structural. They are
1. The Vector Auto-Regression (VAR) Approach
2. The Trade-Indicator (TI) Approach
3. The Dealer-Problem (DP) Approach
Naturally, summarizing empirical microstructure with these three
approaches runs the risk of oversimplication. Although micro-

Empirical Frameworks

127

structure empiricists are sure to feel that much has been left out, I
believe they would agree that these are the three most important of
the basic approaches.9 Each has been applied in the microstructure
literature for more than a decade now, across many different market
structures and asset types.
In essence, all three approaches are a means of relating changes
in price to order ow. The status awarded to order ow in these
approaches is in keeping with the leading role played by order ow
in microstructure theory (per the previous chapter). Naturally, then,
all three approaches depend crucially on having signed trade data.
(It is not enough to know the size of a trade; one must also know its
direction.) In dealer markets, the initiator of the trade establishes the
direction: a customer selling 10 units at a dealer's bid generates an
order ow observation of 10 (exactly analogous to the Bayesian
learning problem featured in the sequential-trade model of the previous chapter). In auction markets, trades are signed according to the
direction of the incoming market order (which is executed against
the most competitive limit order).
In some markets, access to signed trade data is difcult to obtain.
Instead, one might have access only to the sequence of trade sizes
and the transaction prices at which each trade is executed. Though
certainly less desirable than having signed order ow data, there are
techniques in the literature that are commonly used to convert
unsigned trade data into signed order ow data. The conversion can
be noisy, but for many empirical purposes it is adequate (see, e.g.,
Lee and Ready 1991).
All three of these empirical approaches are estimated at the transaction frequency (i.e., observations are the realizations of individual
trades, which are matched to the prices that correspond to those
trades). For work in empirical microstructure, this is the natural data
frequency. More recently, however, estimation of microstructureinspired models has been effected at much lower frequencies (e.g.,
daily, weekly, and monthly). We review these lower frequency applications in chapters 7 and 9.
Statistical Model 1:

The Vector Auto-Regression ( VAR) Approach

The VAR approach was pioneered by Hasbrouck (1991a, b) and has

been successfully applied to FX markets by Payne (1999) and Evans
(2001).10 The approach is not predicated on any particular microstructure model, so it is particularly exible in its use. Its applica-

128

Chapter 5

tions by Hasbrouck and Payne correspond to trading in a auction

setting with a limit order book. The application by Evans corresponds to trading in a multiple-dealer setting.
The model's inferences about the information content of order
ow are drawn from two main sources. First, the model identies
informed trades from impulse responses of prices to order ow.
Specically, informed order ow is that which induces a positive
long-run response in price. Second, variance decompositions allow
one to determine what proportion of all information that enters price
is accounted for by order ow. This attribution statistic represents
the overall contribution of order ow to price determination.
There are two important economic assumptions that underlie the
typical application of this framework:
Assumption 1: Public information is immediately reected in prices.
This assumption is tantamount to assuming that the underlying
market is semi-strong-form efcient (i.e., prices embed all information that is publicly available).
Assumption 2: Trades strictly precede quote revisions.
In the VAR, order ow is allowed to affect price contemporaneously;
reverse causality is prohibited. This is consistent with the general
tack of microstructure theory, under which order ow innovations
drive price innovations, but not the reverse, with subsequent prices
impounding the information that order ow conveys. With contemporaneous order ow in the price equation, one can interpret
innovations in the price equation as reecting effects from public
information.11
Let us turn now to the specication of the VAR model. Let rt denote the percentage change in the spread midpoint and let xt denote
the incoming signed order, where t is a transaction-time observation
counter:12
rt

p
X

ai rti

i1

xt

p
X
i1

p
X
i0

gi rti

p
X

b i xti e1t

5:1

di xti e2t :

5:2

i1

Beyond the two assumptions described above, identication also requires the following restrictions on the innovations:

Empirical Frameworks

129

Ee1t Ee2t Ee1t e2t 0

Ee1t e1s Ee1t e2s Ee2t e2s 0

5:3
Et 0 s:

5:4

From this specication, the effects of order ow information on

prices are easily retrieved. Inverting the VAR representation yields
the following vector moving-average model:
  
 
aL bL
e1t
rt

5:5
xt
e2t
cL dL
The coefcients of the lag polynomials in this moving average representation are the impulse response functions implied by the VAR
model (see, e.g., Hamilton 1994).
Most important among these lag polynomials is bL, which captures the impact of order ow information on subsequent prices. The
individual coefcient bi , for example, measures the effect of a unit
order ow innovation on the price change at the i period horizon. If
we sum these price effects over all horizons, we get a measure of the
cumulative impact of order ow on the level of price. Within this
approach, these cumulative (i.e., persistent) effects are identied as
information:
y
X

bi information content of order flow:

5:6

i0

By dening information content as persistent effects, the denition

includes the two persistent categories outlined in chapter 2: (1) information about future payoffs, and (2) information about portfolio
balance effects. If, on the other hand, the price impact of order ow
reverses over the long run, then by equation (5.6), that order ow
conveys no information. This denition of information therefore
does not include the transitory category outlined in chapter 2information about transitory inventory effects.13
To assess the overall importance of order ow information, we
want to know the share of price variance that is due to order ow.
Within the VAR framework, this is a straightforward task of variance
decomposition (Hasbrouck 1991b). More formally, decompose the
(logarithm) of the spread midpoint, denoted pt , into a random-walk
component m t and a stationary component st :
pt m t s t ;

5:7

130

Chapter 5

where
m t m t1 nt

5:8

and nt @ N0; sn2 , with Ent ns 0 for t 0 s. We shall refer to the

random-walk component m t as the permanent component; we refer
to the stationary component st as the transitory component. Now,
2
dening se12 Ee1t
and se22 Ee 22t from the VAR innovations, we
can further decompose the variance of the permanent component sn2
into a part due to public information and a part due to order ow
information:
!2
!2
y
y
X
X
2
2
a i se1
bi se22
5:9
sn
i0

i0

The second term captures the information impounded in price

through order ow.
With estimates of these parameters, one can pin down the share of
variance that comes from order ow innovations. The ndings in the
literature are quite comparable across equities and FX. For NYSE
stocks, order ow accounts on average for about 33 percent of all
permanent price variation (Hasbrouck 1991b). The same statistic for
French stocks is 40 percent (De Jong, Nijman, and Roell 1996). For
the $/DM market, the statistic is 40 percent (Payne 1999).14
Statistical Model 2:

The Trade-Indicator Approach

The second of the two statistical models I present here is the TradeIndicator (TI) approach, pioneered by Glosten and Harris (1988) and
recently generalized by Huang and Stoll (1997). This model, like the
other two approaches presented here, addresses the link between
order ow and prices. The focus in this case, however, is a bit
narrower than it is under either the VAR or DP approaches. It is
narrower in two ways. First, order ow is not measured from signed
trade size, but rather as a direction indicator variable, say Dt , that
takes on a value of 1 if the previous trade is a sell, and a value of 1
if the previous trade is a buy. This makes the approach less applicable
to questions that revolve around trade size.15 Second, the TI approach
is primarily concerned with decomposing the bid-ask spread.
In chapters 1 and 2, I hinted at the idea that spreads have different
components; now it is time to be explicit. There are three basic costs

Empirical Frameworks

131

that go into determining bid-ask spreads: adverse selection costs,

inventory costs, and order processing costs. To understand the rst
adverse selection costsconsider the sequential trade model presented in chapter 4 (Glosten and Milgrom 1985). In that model, the
dealer faces a pool of potential counterparties, some of whom have
superior information. Because the dealer cannot distinguish those
who are better informed from those who are not, the dealer quotes a
spread that is wide enough so that his losses to informed traders are
balanced against the spread revenues he generates from uninformed
traders. The spread arises in that model strictly because of this information asymmetry; there are no other reasons for a spread (the
dealer is assumed risk neutral and faces no other costs, such as settlement costs, back-ofce costs, cost of this time, etc.). This information asymmetry is what leads to adverse selection costs.
The inventory cost component of the spread is best understood
using models from the inventory-control branch of microstructure
theory (see, for example, the multiple-dealer inventory model of Ho
and Stoll 1983). In these models, dealers are not risk neutral as they
are in the sequential trade model, and therefore have to be compensated for temporarily absorbing risky positions (temporary mismatches in market supply and demand). The spread that arises in
these models is entirely due to this compensation for taking risk;
there are no information asymmetries (by assumption), and no other
costs of marketmaking. (For technical convenience, some models
replace risk-averse dealers with risk-neutral dealers who face some
unspecied ``inventory holding'' cost; qualitatively, the results are
quite similar.)
The third component, order processing cost, is more of a catch-all
category than a specic type of cost. The earliest models of dealers
focus on costs in the traditional sense of production costs: labor
costs, input costs such as back-ofce services, and so on (e.g., Demsetz 1968). Spreads arise in these models in order to produce the
revenues that cover these costs. This cost category is a catch-all in
that a dealer's production function can involve many different types
of input cost (some xed, some variable). Moreover, if the dealing
``industry'' is not perfectly competitive, then spreads (revenues) will
also include some monopoly prot (``rents''). This, too, is lumped
into the order processing cost category.
Empiricists do not have the luxury of considering spread components in isolation: their methods have to accommodate all the com-

132

Chapter 5

ponents, and, to the extent possible, to distinguish among them. Enter the TI approach. Perhaps the best way to communicate how the
TI approach accomplishes this is to do so graphically. Suppose the
only cost component that enters the spread is the order processing
cost. Consider how the sequence of transaction prices would look if
this were the only component. Panel 1 of gure 5.4 provides an
illustration. At is the ask price (often called the offer price in FX), Bt is
the bid price, and Mt is the spread midpoint. Because order ow
conveys no information in this case, nor is there any inventory cost,
the only connection between transaction prices and order ow is the
bouncing of transaction prices from bid to ask (so-called bid-ask
bounce).16
Now consider the case in which the spread arises because of inventory costs only. In this case, following the customer buy order,
the dealer's bid and offer prices change (panel 2 of gure 5.4). This is
the inventory effect described in chapter 2: after the customer sell
order, the dealer is long (or at least longer than before), so to restore
the previous position the dealer lowers his or her quotes. This induces subsequent customer purchases at those lower prices (not
shown with asterisks), which brings the dealer's position gradually
back to its previous level.17 Once the original position is restored, the
original prices are restored, illustrating the transitory nature of inventory effects.
Finally, consider the case in which the spread arises because of
adverse selection costs only. Now the customer sell order pushes
prices down, but the price adjustment is persistent (panel 3 of gure
5.4). This reects the information conveyed by the order ow. This
information can be either of the two types presented in chapter 2 that
have persistent effects: payoff information or information about persistent changes in discount rates (portfolio balance effects).
Equation (5.10) presents the basic Trade-Indicator specication.
The change in the midpoint of the spread, DMt , is the change between two transactions; St1 is the quoted bid-ask spread at the time
of the previous transaction at t 1; Dt1 is the indicator variable
introduced above, which takes values of 1 or 1, depending on the
direction of the previous trade.18 The residual et represents a random
(iid) public information shock at time t:
DMt a b

St1
Dt1 et
2

5:10

Empirical Frameworks

133

Figure 5.4
An illustration of the three spread components: order processing, inventory, and adverse selection. Transactions are marked with an asterisk. At , Mt , and Bt denote ask
price, midpoint, and bid price, respectively (at time t). The indicator variable Dt is 1 if
the trade is buyer-initiated and 1 if seller-initiated.

134

Chapter 5

The coefcients a and b capture adverse selection and inventory

costs, respectively. The sum a b measures the share of the spread
stemming from these two costs. (More precisely, this is the share
of the half spread, which is what is relevant here because it is the
cost faced by a one-way transaction captured by the indicator Dt1 .
The whole spread would be relevant for a roundtrip transaction.)
The remaining share of the spread, 1 a b, stems from orderprocessing costs. Figure 5.4 is helpful for understanding why this
decomposition makes sense. In response to a trade, the mid-quote is
adjusted to reect the inventory cost of the last trade (panel 2) and
the information revealed by the last trade (panel 3). If changes in the
midpoint are unrelated to the direction of the last trade, then the
spread is arising from the order processing cost only.
To separate the adverse selection component of the spread from
the inventory component, Huang and Stoll (1997) extend the basic
model in equation (5.10) to account for the fact that inventory effects
on the mid-quote are transitory, whereas information effects on the
mid-quote are permanent (see their paper for specics). What do
Huang and Stoll nd in the end? On average for NYSE stocks, the
order processing component of the spread is about 60 percent, the
inventory component is about 30 percent, and the adverse selection
component is 10 percent. To my knowledge, the TI approach has not
yet been applied to FX markets, though there is certainly scope for
doing so in the future.19
Structural Model:

The Dealer-Problem Approach

The structural dealer-problem (DP) approach focuses explicitly on

the dealers' decision problem. The solution to that problem generates the equation to be estimated. Though the DP approach draws
heavily on existing theory, the pioneering empirical implementation is due to Madhavan and Smidt (1991), who examine NYSE
stocks.20 The variation I present here for the FX market is in Lyons
1995. The variation makes a change to account for the multipledealer setting of FX: the model introduces a role for marketwide
order ow of the kind produced by interdealer brokers.21
The DP model includes features from both the information-model
and inventory-model branches of microstructure theory. Let me address the information-model features now (the inventory-model features are introduced within the specication of the model). In the

Empirical Frameworks

135

model, dealers recognize that they will sometimes be trading with

counterparties who are better informed. The dealers are assumed to
have rational expectations, in the sense that transactions are ex post
regret free (see chapter 4 for a denition). This means that when, for
example, dealer i quotes an offer price for quantity 10, built into that
quote is an answer to the question: How would my expectation of
value change if my counterparty bought 10, given that he may have
private information? Similarly, the bid price rationally incorporates
any updating a counterparty sale of 10 would engender. To accommodate all possible order sizes, the dealer needs to quote a schedule
of prices, one for each quantity, buy or sell. This schedule therefore
internalizes the proper inference for any potential order and ensures
the quoting dealer will not regret the quote ex post. I provide an
illustration of this schedule below, once the structure of the model is
clear.
The basic idea of the model is the following. Consider the decision
faced by an individual FX dealer, dealer i, when determining what
prices (bid and offer) to quote to another dealer. This decision will
depend on many variables. The model is designed to determine
which variables those are and how each should be weighted in
determining the appropriate prices. Importantly, the model also
needs to account for how the dealer's beliefs and his position (inventory) vary across time. The model will include T trading periods,
indexed by t 1; . . . ; T, to allow for this across-time aspect of the
decision.
I present the model in three parts. The rst part is the information
environment. The second part is the formation of expectations conditional on the information environment. The third part is the determination of bid-offer quotes as a function of expectations and current
inventory.
The Information Structure
The payoff to holding FX at time T is denoted by V, which is composed of a series of incrementsfor example, interest differentials
so that:
V

T
X

Ri ;

i0

where R0 is a known constant. The increments are i.i.d. Normal with

136

Chapter 5

mean zero. Each increment Rt is realized immediately after trading in

period t. Realizations of the increments represent the ow of public
information over time. The value of FX at t is thus dened as
Pt
Ri . At the time of quoting and trading in period t, that is,
Vt i0
before Rt is realized, Vt is a random variable. Given this structure,
without transaction costs or further information, the quoted price
of FX at time t would equal Vt1 , which is the expected value of the
asset conditional on public information at t.
The following three signals dene each period's information environment prior to dealer i's quote to another dealer, dealer j. The rst
two are received simultaneously, prior to the third. The rst and
third are observed by all dealers; the second is observed only by
dealer j (this second signal is the source of adverse selection faced by
the quoting dealer, dealer i):
St Vt ht

5:11

Sjt Vt o jt

5:12

Bt Vt xt

5:13

The noise terms ht , o jt , and xt are Normally distributed about zero,

are independent of one another and across periods, and have variances sh2 , so2 , and sx2 , respectively. At the outset of each period t,
all dealers receive a public signal St of the full information value Vt .
Also at the outset of each period t, dealer j receives a private signal
Sjt of Vt . One potential source of private signals at the dealer level is
order ow from nondealer customers, as described in chapter 3.
The third signal, Bt , is an additional public signal that reects the
institutions of the FX market. It is distinct from St in that it is directly
measurable. In particular, Bt represents a signal of marketwide order
ow. As indicated in chapter 3, there is one source of information on
marketwide order ow that provides transparency in an otherwise
opaque FX trading process: interdealer brokers.
The last variable dealer i uses to determine his quote schedule is
dealer j's signed trade, denoted Xjt . That is, consistent with the
regret-free property discussed above, dealer i quotes a schedule of
prices that protects him from the information in Xjt . To determine the
information in Xjt , dealer i needs to think through dealer j's trading
incentives. Under the usual assumptions (exponential utility dened
over end-of-period wealth), the quantity dealer j chooses to trade is
linearly related to the deviation between dealer j's expectation and

Empirical Frameworks

137

Figure 5.5
Timing in each period of the DP model.

the transaction price, plus a quantity representing liquidity demand

L jt that is uncorrelated with Vt (e.g., inventory adjustment trading):
Xjt ym jt Pit L jt

5:14

where m jt is the expectation of Vt conditional on information available to dealer j at t, and the value of Xjt is known only to dealer j.
Note that Xjt can be either positive or negative.
Figure 5.5 summarizes the timing of the model in each period.
Formation of Expectations
Dealer i's quote schedule is a function of his expectation of Vt , which
is denoted m it . This expectation, in turn, is conditioned on the signals
described above: St , Bt , and Xjt (the fourth variable described above,
Sjt , is the signal embedded in Xjt ).
The rst public signal, St , summarizes dealer i's prior belief regarding Vt . After observing the second public signal Bt , dealer i's
posterior belief, denoted m t , can be expressed as a weighted average
of St and Bt :
m t rSt 1 rBt ;

5:15

where r sx2 =sx2 sh2 . These posterior beliefs m t are Normally distributed with mean Vt and variance sm2 r 2 sh2 1 r 2 sx2 .
After observing Bt , dealer i then considers what he would learn
from various possible realizations of Xjt (and the schedule he quotes
internalizes what he would learn from each possible realization). In
particular, dealer i can form the statistic Zjt :


Xjt =y Pit lm t
1
Vt ojt
L jt ;
5:16
Zjt
y1 l
1l
where l so2 =sm2 so2 . This statistic is also Normally distributed, with mean Vt and variance equal to the variance of the last
2
two terms, both of which are orthogonal to Vt . Let sZj
denote this

138

Chapter 5

variance. Note that Zjt is statistically independent of m t , because Zjt is

orthogonal to both St and Bt . Thus, dealer i's posterior m it , expressed
as a function of any Xjt , takes the form of a weighted average of m t
and Zjt :
m it km t 1 kZjt ;

5:17

2
2
=sZj
sm2 . This expectation plays a central role in
where k sZj
determining dealer i's quote.

Quote Determination
Consider a prototypical inventory control model in which the transaction price Pit is linearly related to the dealer's current inventory:
Pit m it aIit Ii gDt ;

5:18

where m it is the expectation of Vt conditional on information available to dealer i at t, Iit is dealer i's current inventory position, Ii  is i's
desired position,22 and Dt is a direction indicator that picks up bidoffer bounce in the model. The inventory control effect, governed by
a, will in general be a function of rm capital, relative interest rates,
and other inventory carrying costs. The direction indicator Dt equals
1 when the transaction price Pit is the offer price and 1 when Pit is
the bid price. For a given expectation m it , Dt picks up half the spread.
(Dt should be viewed as half the spread for quantities close to zero;
for other quantities, the quoted spread widens to protect against adverse selectionthrough the effects on m it .) This term measures the
catch-all order-processing costs described in the presentation of the
trade-indicator (TI) model.
Consistent with the regret-free property of quotes, I substitute
dealer i's expectation m it in equation (5.17) into equation (5.18),
yielding:
Pit km t 1 kZjt aIit Ii gDt
which is equivalent to


 
 
1f
a
g

Xjt
Iit Ii
Dt ;
Pit 1 rBt rSt
fy
f
f

5:19

5:20

where the parameter f k l=1 l and 0 < f < 1 because

0 < k < 1, 0 < l < 1, and k < l.

Empirical Frameworks

139

An Estimable Equation
Equation (5.20) is not directly estimable because the public signal St
is not observable to the econometrician. Fortunately, though, the
assumptions about the model's signals and the evolution of Vt allow
one to express the period t prior as equal to the period t 1 posterior
from equation (5.18) lagged one period, plus an expectational error
term e it :
St m it1 e it Pit1 aIit1 Ii gDt1 e it :

5:21

Substituting this expression for St into equation (5.20) and taking the
rst difference yields:




 
a
1f
a

a Ii
Xjt
Iit aIit1
DPit
f
fy
f
 
g

5:22
Dt gDt1 1 rBt e it
f
This corresponds to a reduced-form estimating equation of: 23
DPit b0 b1 Xjt b 2 Iit b 3 Iit1 b4 Dt
b5 Dt1 b 6 Bt b7 nit1 nit :

5:23

The structural model predicts that fb1 ; b3 ; b 4 ; b6 g > 0, fb2 ; b 5 ; b7 g < 0,

jb 2 j > b3 , and b4 > jb5 j, where the latter inequalities derive from the
fact that 0 < f < 1. The following section presents results from estimating equation (5.23) in the $/DM market.
Figure 5.6 provides an illustration of the qualitative features of the
DP model's quote schedule. Note the similarity to the net-supply
gures presented in chapter 2.
5.3

Findings: Informative Orders and Imperfect Risk Sharing

Table 5.1 presents results from estimating the DP model in equation

(5.23) over the ve-day sample in Lyons (1995). The sample size is
the number of incoming direct interdealer transactions received by
the dealer being tracked (dealer i in the model), for a total of 839
observations. (The four overnight price changes are excluded because the model is intended to explain intraday quoting dynamics,
not price changes over periods when the dealer is not active.) The

140

Chapter 5

Figure 5.6
Dealer i's quote schedule in the DP model. The slope of the quoted price schedule is
determined by b1 and reects the information conveyed by order ow Xjt (Xjt is negative if the incoming order is a sell). Inventory is a shift variable: the larger Iit is relative to the desired position Ii , the lower the price schedule throughout (to induce
inventory decumulating purchases by counterparties). The bid-offer spread at quantities
near zero is pinned down by the parameter g, which multiplies the direction-indicator
variable Dt in the pricing rule of equation (5.18).
Table 5.1
Structural Model Estimates
DPit b 0 b 1 Xjt b2 Iit b 3 Iit1 b 4 Dt b 5 Dt1 b6 Bt b7 vit1 vit
b0

b1

b2

b3

b4

b5

b6

b7

R2

1.30
(0.96)

1.44
(3.10)

0.98
(3.59)

0.79
(3.00)

10.15
(4.73)

8.93
(6.12)

0.69
(2.21)

0.09
(2.55)

0.23

1.34
(0.99)

1.40
(3.03)

0.97
(3.56)

0.78
(2.95)

10.43
(4.86)

9.16
(6.28)

0.09
(2.61)

0.22

>0

<0

>0

>0

<0

>0

<0

T-statistics in parentheses. The last row indicates the signs predicted by the structural
model. DPit is the change in the incoming transaction price (DM/$) from t 1 to t. Xit
is the incoming order transacted at dealer i's quoted prices, positive for purchases (i.e.,
effected at the offer) and negative for sales (at the bid). The units of Xit are such that
b1 1 implies an information effect on price of DM0.0001 for every $10 million. It is
dealer i's inventory at the end of period t. Dt is an indicator variable with value 1 if the
incoming order is a purchase and value 1 if a sale. Bt is the net quantity of third-party
brokered trading over the previous two minutes, positive for buyer-initiated trades
and negative for seller-initiated trades. All quantity variables are in $ millions. All
coefcients are multiplied by 10 5 . Sample: August 37, 1992, 839 observations.

Empirical Frameworks

141

rst row presents estimates of the full reduced form, that is, including the information effects of brokered order ow. The second row
excludes the brokered order ow variable to make it directly comparable to results from work on the stock market (Madhavan and
Smidt 1991). Row three indicates the signs of the coefcients predicted by the structural model under the null that both information
and inventory control effects are present.
The main results from table 5.1 are the signicant and properly
signed coefcients on the information (order ow) variables Xjt and
Bt , and the inventory variables Iit and Iit1 . The size of b1 implies that
the dealer being tracked widens his spread about 2.8 pips (0.00028
DM, or 1.4 doubledadjusted for units) per $10 million to protect
against adverse selection. The size of the inventory control coefcient
b 3 (which equals a in equation 5.18) implies that the dealer motivates
inventory decumulation by shading his DM price of dollars by 0.8
pips for every $10 million of net open position.
The coefcients on the indicator variables Dt and Dt1 , which
measure the effective spread when Xjt close to zero, are signicant,
and have the predicted relative size as well, b4 > jb5 j. (Recall that
there is a one to one correspondence between Dt and the sign of Xjt ,
so Dt controls perfectly for whether Xjt is at the bid or the offer.) The
coefcient b 4 suggests that once one controls for the information and
inventory effects, the baseline spread for this dealer is about two pips
(2b4 =10 5 ). Note too that the moving average coefcient b7 is signicant and properly signed, providing further support for the model.
Finally, the levels of the R 2 's reect that Xjt and Bt together account
for only a small fraction of the trading activity in the wider market.
Qualitatively, the main difference in the results from those for
NYSE stocks (Madhavan and Smidt 1991) is the signicance of the
inventory effect on price, captured by b2 and b 3 . Despite a large theoretical literature on price effects from inventory, work on equity
marketmakers does not detect it. In this respect, the model ts better
in FX.24 From the perspective of exchange rate economics, however,
the more striking result is the nding of a signicant b1 the private
information conveyed by order ow. It is more striking because most
exchange rate economists are quite comfortable with the idea that
there is private information in stock markets. At the same time, they
were not taught to believe that private information is relevant in the
FX market.

142

Chapter 5

The results in table 5.1 belong in the category of traditional

microstructure analysis in the sense that they arise at the transaction
frequency. Despite their high-frequency nature, we have seen that
these resultsas well as other high-frequency results noted in this
chapterpoint toward some larger lessons. I want to highlight two
in particular that will resurface in later chapters.
1. Asymmetric information is present in the FX market. There is a
long-standing convention within exchange rate economics to treat all
relevant information as common knowledge among market participants (central banks excepted). The results of empirical papers
described in this chapterand in chapter 2suggest that this common knowledge paradigm is incomplete. Moving away from this
paradigm is a realignment of how we think about this market's information environment.
2. Order ow is an important proximate determinant of exchange
rates. The biggest puzzle in exchange rate economics for nearly two
decades now has been the fact that hypothesized macroeconomic
determinants have little explanatory power. The microstructure
analysis described in this chapter suggests that order ow, in contrast, has considerable explanatory power (e.g., Payne 1999; Evans
2001). Moving toward the idea of order ow as a proximate cause of
exchange rates is another realignment of how we think about this
market; the traditional macro paradigm does not recognize the concept of order ow.
Before turning to microstructure analysis of a more macro sort (in
chapters 79), let us consider another more ``micro'' question: Where
is all that trading volume coming from?
5.4

Why Is Volume So High in FX?

Trading volume in FX marketsat $1.5 trillion per dayis extremely high. It is high relative to other asset markets, high relative
to underlying trade in goods and services, and high relative to what
standard theories would predict (be they micro theories or macro
theories).
Why, some might ask, should we care? After all, it's not clear this
volume affects prices, and if it does not affect prices, it may not have
important welfare consequences. But there are several ways in which

Empirical Frameworks

143

inexplicably high FX volume can have important consequences. For

example, misunderstanding the causes of high volume can lead to
bad policy. Consider the issue of transaction taxes. Proponents of
levying transaction taxes tend to associate high volume with excessive speculation. If instead much of this volume reects dealer
risk management, then a transaction tax wouldunintentionally
impede that risk management. Another way for high volume to have
important consequences is that it can impede order ow's information role. Per the models of chapter 4, the precision of the information that order ow conveys is a function of the underlying causes of
that order ow. It is important to understand whether those causes
contribute to, or detract from, informational precision. A third way
for volume to have important consequences is that misunderstanding its causes can lead to bad theory. The volume puzzle indicates
that the asset approach to the FX market is missing key features.
These features may not be important, but to be condent that they
are unimportant requires a tremendous leap of faith.25
Microstructure analysis suggests that resolving the volume puzzle
should rest on two features missing from the macro-asset approach:
1. Inventory control
2. Asymmetric information
Take inventory control rst. On the inventory control side, the key
mechanism is hot potato tradingpassing unwanted positions from
dealer to dealer following an initial customer trade. Burnham (1991,
135) provides a clear link between this hot potato and the volume
puzzle when he writes: ``A game of `hot potato' has begun. . . . It
is this search process for a counterparty who is willing to accept
a new currency position that accounts for a good deal of the volume
in the foreign exchange market.'' The simultaneous-trade model of
the previous chapter provides some rigor to this descriptive account
of volume amplication.
Asymmetric information contributes to resolution of the volume
puzzle by furnishing an additional and important motive for speculative trade that is not present within the macro-asset approach: Because all information is assumed to be common knowledge under
the macro-asset approach, no agent can have an information advantage on which to trade.
Do hot potato trading and asymmetric information together fully
account for the volume puzzle? They certainly both contribute, but

144

Chapter 5

we do not yet know enough to be condent that we have a full accounting. In fact, as an empirical matter, there is very little work on
hot potato trading.26 The one paper that addresses hot potato trading
head on is Lyons 1996b. The remainder of this section reviews the
methodology and results from that paper.
Volume: Sound and Fury Signifying Nothing?
Lyons 1996b addresses hot potato trading by examining whenas
opposed to whethertrades are informative. Specically, I use
transaction data to test whether trades occurring when trading intensity is high are more informativedollar for dollarthan trades
occurring when trading intensity is low. Theory admits both possibilities, depending on the posited information structure. I present
what I call a hot potato model of currency trading, which explains
why low-intensity trades might be more informative. In the model,
the wave of inventory management trading among dealers following
innovations in order ow generates an inverse relationship between
intensity and information content. Empirically, I nd that lowintensity trades are more informative, supporting the hot potato
hypothesis.
To clarify the hot potato process, consider the following crude,
but not unrealistic example. Suppose there are ten dealers, all of
whom are risk averse, and each currently with a zero net position.
One dealer accommodates a customer sale of $10 million worth of
DM. Not wanting to carry the open position, the dealer calculates
his share of this inventory imbalanceor 1/10th of $10 million
calls another dealer, and unloads $9 million worth of DM. The
dealer receiving this trade then calculates his share of this inventory imbalanceor 1/10th of $9 millioncalls another dealer, and
unloads $8.1 million worth of DM. The hot potato process continues.
In the limit, the total interdealer volume generated from the $10
million customer trade is $9m/1 0:9 $90 million. The resulting
share of wholesale trading that is interdealer is therefore $90m/
$100m, or 90 percent. This is a bit high relative to an interdealer
share these days of about two-thirds, but 10 years ago, the total share
of interdealer FX trading was close to 90 percent.
Here are two possible reactions to the example above, neither of
which vitiates its message. Reaction one: Shouldn't the multiplier be
innite because risk-averse dealers would not choose to retain any of

Empirical Frameworks

145

the imbalance? The answer is this: In equilibrium, price will adjust to

induce dealers to hold some of the perceived excess supply. (The 10
percent rule of the example is a crude approximation of a much
richer trading process, e.g., that in the simultaneous-trade model of
chapter 4.) Reaction two: Interdealer trades can reduce idiosyncratic
inventory imbalanceswhich reduces idiosyncratic risk rather than
simply bouncing itwhich will mute the multiplier. This is true,
particularly if the trades are brokered. It is therefore more reasonable
to think about the example in terms of net customer order ow (i.e.,
after offsetting buys and sells have been netted out).
As I noted above, theory admits the possibility that trades are
in fact more informative when trading intensity is highthe opposite
of the hot potato prediction. This arises in a model by Easley and
O'Hara (1992). In contrast to earlier models where private information is known to exist, in Easley and O'Hara (1992) there is simply a
probability that private information exists (which they call event
uncertainty). If private information does exist, there is some probability, say q, that an informed trader has received good news, and
probability 1 q that he has received bad news. They demonstrate
that if there is active trading at time t, then a rational dealer raises
the probability he or she attaches to the existence of private information and lowers the probability of no private information. If trading intensity is high, an incoming trade of a given size induces a
larger update in beliefs because it is more likely to be signaling news.
On the ipside, trades occurring when intensity is low should induce
a smaller update in beliefs.
Before turning to the empirical model, I should note that the role
of time in empirical microstructure has only recently emerged. Two
important contributions are papers by Hausman, Lo, and MacKinlay
(1992) and Engle and Russell (1998). Working with stocks, Hausman,
Lo, and Mackinlay test for exogeneity of the length of time between
transactions, which they reject at conventional signicance levels.
However, they also nd that their estimates do not change appreciably when they use instrumental variables to control for endogeneity. This nding motivates them to perform the rest of their
analysis under the assumption of exogenous intertransaction times.
Engle and Russell (1998) meet the role of time head on by modeling
it explicitly. Their statistical model of irregularly spaced transactions
data treats the time between transactions as a stochastic process.
(They refer to the model as an ``autoregressive conditional duration''

146

Chapter 5

model.) Applying the model to stock data, they nd evidence of

clustering in intertransaction times, which they relate to underlying
market conditions.
The Model
We introduce a role for time in the DP model of the previous section
via the liquidity demand L jt . Equation (5.14) described the role of
that liquidity demand in dealer j's trade, which I note again here for
reference:
Xjt ym jt Pit L jt

5:14

The hot potato hypothesis of order ow information associates

liquidity demand L jt with inventory adjustment trading. In FX
according to the hypothesisinnovations in nondealer order ow
spark repeated interdealer trading of idiosyncratic inventory imbalances. This rapid passing of the hot potato generates a relatively
large role for liquidity trades in periods of short intertransaction
times. The event uncertainty hypothesis, in contrast, associates short
intertransaction times with a relatively large role for informative
trading: in the presence of event uncertainty, intense trading is a
signal that an information event has occurred. To summarize, for
given precisions of the signals Sjt and St , we can characterize these
views as:
Hot potato hypothesis

High when intertransaction times are short
2
sLj
Low when intertransaction times are long
Event uncertainty hypothesis

Low when intertransaction times are short
sLj2
High when intertransaction times are long
This change in the relative intensity of liquidity trading will alter the
signal extraction problem faced by the quoting dealer. The implications within the DP model follow immediately from the analysis in
section 5.2: The hot potato hypothesis predicts that the information
effect of order ow on price will be smaller when trading is intense;
the event uncertainty hypothesis predicts the opposite, that the information effect of order ow on price will be larger when trading is
intense.

Empirical Frameworks

147

Table 5.2
Testing the Hot Potato Hypothesis: Is Order Flow Less Informative When Intertransaction Time Is Short?
DPit b0 b1 st Xjt b10 lt Xjt b 2 Iit b 3 Iit1 b 4 Dt b5 Dt1 eit
b 1 (short)

b10 (long)

Fraction short

b1 b 10
P-value

0.01
(0.01)

2.20
(3.84)

262/842

0.000

0.76
(1.63)

2.60
(3.40)

506/842

0.009

Intertransaction time short if:

Less than 1 minute
Less than 2 minutes

T-statistics in parentheses. The coefcient b1 measures the information effect of orders

for which the time from the previous transaction is short (st 1 and lt 0 in the
equation in the heading), where short is dened in the rst column. The coefcient b10
measures the information effect of those orders for which the time from the previous
transaction is long (st 0 and lt 1), where long is dened as not short. The ``Fraction
short'' column presents the fraction of observations satisfying the corresponding denition of short intertransaction times. In each case, the remaining observations fall into
the long category. The P-value column presents the signicance level at which the null
b 1 b10 can just be rejected. DPit is the change in the incoming transaction price (DM/
$) from t 1 to t. Xit is the incoming order transacted at dealer i's quoted prices, positive for purchases (i.e., effected at the offer) and negative for sales (at the bid). The
units of Xit are such that b1 1 implies an information effect on price of DM0.0001 for
every $10 million. It is dealer i's inventory at the end of period t. Dt is an indicator variable with value 1 if the incoming order is a purchase and value 1 if a sale.
Sample: August 37, 1992. Estimated using OLS with autocorrelation consistent (rst
order) standard errors.

Results
Table 5.2 presents estimates of the information content of order ow,
distinguishing between short and long intertransaction times. This is
achieved via the introduction of dummy variables st and lt (see the
equation heading the table). The dummy st equals 1 if intertransaction time is short, 0 otherwise; the dummy lt equals 0 if intertransaction time is short, 1 otherwise. Short intertransaction times are
dened two ways: less than 1 minute from the previous transaction
and less than 2 minutes. (The time stamps on the Lyons 1995 data
are very precise, because they are assigned by the computer; however, they do not provide precision beyond the minute. Hence, less
than 1 minute includes trades with the same time stamp; less than
2 minutes includes trades with time stamps differing by 1 or 0
minutes.) These categories bracket the mean intertransaction time

148

Chapter 5

of 1.8 minutes. The second category corresponds to a break at the

median intertransaction time.
The results support the hot potato hypothesis over the event uncertainty hypothesis. The coefcient b1 which measures the information effect of incoming trades with short intertransaction times
is insignicant at conventional levels. In contrast, the coefcient
b 0 1 which measures the information effect of incoming trades with
long intertransaction timesis signicant. (The model's other coefcientsnot presentedare little changed from those reported in
table 5.1.) Moreover, a test of the restriction that b 1 b 0 1 is rejected at
the 1 percent level in both cases. In summary, trades occurring when
transaction intensity is high are signicantly less informative than
trades occurring when transaction intensity is low, as the hot potato
hypothesis predicts.
Table 5.3
Testing the Hot Potato Hypothesis: Is Order Flow Less Informative When Transactions
Follow In the Same Direction?
DPit b 0 b 1 st Xjt b 10 ot Xjt b100 lt Xjt b 2 Iit b3 Iit1 b4 Dt b5 Dt1 e it
b1 (short
and same)

b 10 (short and
opposite)

b100

0.06
(0.11)

1.90
(3.01)

2.64
(3.46)

(long)

Fraction short
and same

Fraction short
and opposite

b 1 b10
P-value

276/842

230/842

0.009

T-statistics in parentheses. The coefcient b 1 measures the information effect of orders

that have (1) short intertransaction times, dened as less than the median of two
minutes, and (2) the same sign (direction) as the previous order (st 1, ot 0, and
lt 0 in the equation in the heading). The coefcient b 10 measures the information effect
of orders that have (1) short intertransaction times, dened as less than the median of
two minutes, and (2) the opposite sign (direction) of the previous order (st 0, ot 1,
and lt 0). The coefcient b100 measures the information effect of orders that have long
intertransaction times, dened as greater than or equal to the median of two minutes
(st 0, ot 0, and lt 1). The ``Fraction short and same'' column presents the fraction
of observations satisfying the corresponding denition of short and same (similarly for
the ``Fraction short and opposite'' column). The remaining observations fall into the
long category. The P-value column presents the signicance level at which the null
b1 b 10 can just be rejected. DPit is the change in the incoming transaction price (DM/
$) from t 1 to t. Xit is the incoming order transacted at dealer i's quoted prices, positive for purchases (i.e., effected at the offer) and negative for sales (at the bid). The
units of Xit are such that b 1 1 implies an information effect on price of DM0.0001 for
every $10 million. It is dealer i's inventory at the end of period t. Dt is an indicator
variable with value 1 if the incoming order is a purchase and value 1 if a sale. Sample: August 37, 1992. Estimated using OLS with autocorrelation consistent (rst
order) standard errors.

Empirical Frameworks

149

There is an additional testable implication of the hot potato hypothesis that follows directly from inventories being repeatedly
bounced from dealer to dealer. These inventory management trades
will tend to be in the same direction (i.e., have the same sign). The test
presented in table 5.3 addresses this question: In periods of intense
trading, is order ow less informative when transactions follow the
same direction? Again, we introduce dummy variables, in this case
st , ot , and lt (see the equation heading the table). The dummy st
equals 1 if (1) intertransaction time is short and (2) the previous incoming trade has the same direction, 0 otherwise; the dummy ot
equals 1 if (1) intertransaction time is short and (2) the previous
incoming trade has the opposite direction, 0 otherwise; the dummy lt
equals 0 if intertransaction time is short, 1 otherwise. A short intertransaction time is dened as less than the median of 2 minutes.
Once again, the results are consistent with the hot potato hypothesis. The coefcient b1 short intertransaction times and same
directionis insignicant. In contrast, the coefcient b 0 1 short
intertransaction times and opposite directionis signicant. A test
of the restriction that b1 b 0 1 is rejected at the 1 percent level. To
summarize, in periods of intense trading, trades occurring in the
same direction are signicantly less informative than trades occurring in the opposite direction.
It should be noted, however, that although the hot potato and
event uncertainty hypotheses make opposite predictions regarding
the relation between information and trading intensity, they are not
necessarily exclusive hypotheses. That is, both effects could be operative: hot potato trading may simply dominate when trading is most
intense in this market. We will see in chapter 8 that both effects are
indeed operative in the data.

This Page Intentionally Left Blank

Exchange Rate Models:

Macro to Microfoundations

This chapter reviews the traditional models in exchange rate economics, all of which are macroeconomic in orientation. My objective
is to provide perspective for people working in mainstream nance
who are interested in FX markets, but do not feel they know enough
exchange rate economics to work in the area. (There remains, in my
judgment, much room for intellectual arbitrage, if working on the
world's largest and arguably most important nancial market is not
motivation enough.) Accordingly, presentation of individual models
is rather compact. Fuller treatment is available in the surveys of
Frankel and Rose (1995), Isard (1995), and Taylor (1995).
Before presenting the models, though, I include a section on dening the term ``fundamentals'' as used in exchange rate economics.
Reviewing how exchange rate economists think about fundamentals
should help readers whose perspective on asset valuation comes
primarily from equity markets. At the end of the chapter, I offer a
discussion designed to help loosen the dichotomy in exchange rate
economics between ``micro'' issues and ``macro'' issues. I provide
examples of the sweep of issuesmicro to macroto which one can
apply microstructure tools.
Sections 6.3 and 6.4 address microfoundations. Because traditional
exchange rate models are macroeconomic, to many people their
lack of microeconomic foundations is problematic. Shoring up these
microfoundations is the aim of much recent work in exchange rate
economics. The microstructure approach, too, is an effort to bring
microfoundations to exchange rate economics. However, it does so in
a qualitatively different way because it emphasizes a different set of
microfoundations. Recent work within the asset market approach
emphasizes microfoundations that mirror those emphasized in macroeconomics more generally. These microfoundations are rooted in

152

Chapter 6

the ``two T's'': tastes and technology. In macroeconomics, a model

built from tastes and technology is a model with well-dened preferences dened over consumption goods (tastes) and well-dened
technology for producing those consumption goods. The microstructure approach is grounded instead in what I call the ``two I's'':
information and institutions. By information I mean that microstructure models incorporate a greater variety of information types
(most of which are ruled out in most asset-approach models, such as
information types that are not public). By institutions I mean these
models recognize the effect of market organization on how nonpublic information is learned and aggregated.
6.1

Fundamentals in FX: Goods Market and Asset Market Views

Before I survey exchange rate theory, it may be helpful to relate that

theory to the broader context of asset valuation. What, in particular,
are the fundamentals that drive valuation in FX markets? In equity
markets, the fundamental payoffs that drive valuation are future
dividends (cash ows). In bond markets, the fundamental payoffs
that drive valuation are coupons (and principal). Let me offer some
perspective on the analogue in FX markets to dividends and coupons.
I offer two core notions of FX fundamentals, one from each of the
two macro approaches introduced in chapter one. From the goods
market approach, the core notion of fundamentals comes from a
relation called Purchasing Power Parity, or PPP, which is the idea
that a given currency should have the same purchasing power over
goods everywhere in the world. The core notion of fundamentals
from the asset-market approach comes from a relation called Uncovered Interest Parity, or UIP, which is the idea that the returns on
assets that differ only in their currency denomination should be
equal (when expressed in the same currency). Let us consider each of
these in turn.
FX Fundamentals: The Goods Market View and PPP
Purchasing Power Parity is perhaps the oldest and most venerated
relation in exchange rate economics. It can be motivated intuitively
with the ``law of one price''a no-arbitrage condition, applied in this
case to international goods markets. (For a recent treatment of goods
market arbitrage, see Obstfeld and Taylor 1997.) This close link to

Exchange Rate Models

153

goods markets is what ties this PPP notion of fundamentals to the

goods market approach. In the absence of friction, the law of one
price says that the dollar price of a good in the United States should
be the same as the dollar price of the same good in another country,
say Britain:
Pi; US P$= Pi; UK

6:1

where Pi; US is the dollar price of good i in the United States, Pi; UK is
the pound sterling price of good i in Britain, and P$= is the spot exchange rate (dollars per pound).1 PPP simply generalizes this onegood relation to a price index representing multiple goods, like the
consumer price index:
P US P$= P UK ;

6:2

where P US is the dollar price of a basket of goods in the United States

and P UK is the pound price of a comparable basket in Britain. Rearranging yields an expression for exchange rate fundamentals:
P$= P US =P UK :

6:3

This goods market view of fundamentals is an important conceptual anchor within exchange rate economics. The PPP relation is
simple and appealing: the exchange rate equals the ratio of the two
national price levels. Now, empirically this relation does not hold at
all times, and departures can be substantial (departures in excess of
30 percent in major U.S. dollar markets are not uncommon). Nevertheless, there is ample evidence that departures dissipate over time,
with a half-life on the order of ve years.2 Thus, in terms of providing a ``center of gravity'' for thinking about exchange rate fundamentals, the PPP relation remains quite valuable.3
FX Fundamentals: The Asset Market View and UIP
Though fundamentals as described by PPP are an important conceptual anchor, this notion of fundamentals is a far cry from the
dividend-discount valuations common in equity markets. The assetmarket approach to exchange rates provides a more comforting
analogue. Within the asset approach, a good way to illustrate fundamentals is with the relation Uncovered Interest Parity.
UIP describes the relationship between expected returns on shortterm, interest-bearing assets denominated in different currencies.

154

Chapter 6

(For simplicity, I refer to this asset class as ``deposits.'') The relation

implies that, when measured in dollars, the expected returns on
pound sterling deposits and dollar deposits are equal. One would
expect this to hold, for example, in a risk-neutral world without
capital controls because any disparity would be immediately exploited.
In its exact version, UIP implies that


EP$=; t1 j W t
;
1 i $; t 1 i ; t
P$=; t
where i $; t and i ; t are the one-period nominal interest rates in dollars
and pounds at time t, respectively, and EP$=; t1 j W t is the expected
future spot rate. The expected spot rate is conditional on information
available at t, which I denote by ``jW t '', where W t is the relevant information set.
There is an approximate version of UIP that is a bit more intuitive,
and also easier to work with. The expected dollar return on pound
deposits has two components, corresponding to the two terms on the
righthand side of the exact version: the one-period pound interest
rate and the expected appreciation of the pound. The approximate
version of UIP expresses this in a linear form:4
i $; t i ; t Ept1 pt j W t

6:4

where pt is the natural log of P$=; t ($/ subscript suppressed) and

Ept1 pt j W t is the expected percent (log difference) appreciation
in the pound between t and t 1. The intuition of UIP is clear: the
expected dollar return on dollar deposits on the lefthand side equals
the expected dollar return on pound deposits on the righthand side.
The UIP expression in equation (6.4) generates an intuitive, assetmarket notion of fundamentals, one that will speak to those more
familiar with equity and bond valuation. Start by rewriting equation
(6.4) as
pt i ; t i $; t Ept1 j W t :
Projecting this expression one period forward, and taking expectations, we can also write:
Ept1 j W t Ei ; t1 i $; t1 j W t Ept2 j W t :
Substituting this expression for E pt1 j W t into the previous equation
yields:
pt i ; t i $; t Ei ; t1 i $; t1 j W t Ept2 j W t

Exchange Rate Models

155

Repeating this projection and substitution for dates beyond t 1

yields our asset-market version of fundamentals:
"
#
Tt1
X
i ; tt i $; tt j W t EpT j W t
6:5
pt E
t0

Thus, in a manner similar to a dividend discount model, the exchange rate fundamental is a sum of expected future (one-period)
interest differentials, plus a ``long-run'' value of the exchange rate.5
These future interest differentials are the net cash ows that accrue to
holding foreign currency deposits. The second termthe long-run
valueis typically pinned down with PPP.
6.2

Macro Models: An Overview

Most readers familiar with microstructure nance are unfamiliar

with models of exchange rate determination, so I review them here.6
There are four major models, and macro fundamentals drive all four:
1. exible-price monetary model
2. sticky-price monetary model
3. portfolio balance model
4. general equilibrium model
All four are squarely in the asset-market approach. The forwardlooking nature of the asset approach implies that current values of
macro variables are not the whole storythe market's expectation
of these variables' future values is also important. For the purposes
of this book, it is worthwhile to review the major models because
they provide valuable perspective for the more ``macro-level'' analysis of later chapters.
Flexible-Price Monetary Model
Though simple and intuitive, the PPP relation described above has
little structural content because it determines the exchange rate from
two variables that are themselves endogenous. Naturally, these features make it a ne candidate as a building block: embedding the
relation in a more fully articulated model both exploits its simplicity
and provides structure. This is exactly the strategy of the monetary

156

Chapter 6

models. Note too that PPP does not provide a direct link to policy, in
particular to monetary policy. By embedding PPP in a model that is
articulated in the monetary dimension, the monetary models provide
that direct link.
The rst of the two monetary models, the exible-price model,
starts with the assumption that PPP always holds (key references
include Mussa 1976 and Frenkel 1976). It is, in fact, a straightforward
extension of the PPP relation: It uses the PPP expression for the exchange rate and substitutes alternative expressions for the two price
levels P US and P UK . Recall that the PPP expression from equation
(6.3) is
P$= P US =P UK :
Taking logarithms of both sides yields an equivalent expression that
is a bit easier to work with:7
p $= p US p UK :
The substitute expressions for p US and p UK come from the heart of
macroeconomicsthe equation describing equilibrium in the money
market.8 The money market is in equilibrium when the supply of
money in real terms (i.e., price adjusted) equals the demand. Demand for money is modeled as a function of both real income
(transactions demand) and the nominal interest rate (an opportunity
cost of holding money). Equilibrium is typically expressed in loglinear form as
m US p US ay US bi $ ;

6:6

where m US p US is the (log) real money supply and ay US biUS is the

(log) real money demand, with m US as log nominal money supply,
p US as log price level, y US as log real output, and iUS as the short-term
nominal interest rate. Solving for the price level in separate U.S. and
U.K. equations, we have:
p US m US ay US bi $
p UK m UK ay UK bi

6:7

Plugging our price-level expressions in equation (6.7) into the PPP

relation gives us the exible-price monetary model:
p $= m US m UK ay US y UK bi $ i :

6:8

Exchange Rate Models

157

This is the simplest version of the model. The link to money supplies
and therefore monetary policyis clear. The link to the ``exibleprice'' title is less clear. Indeed, in a strict sense we have not used the
assumption of perfectly exible exchange rates and prices: It is not a
necessary condition for PPP, nor is it required for the price-level
expressions in equation (6.6). Instead, the name comes from the need
to distinguish this model from its sticky-price cousin (which we return to below).
Another link that is not yet clear is the model's connection to
expected future macro fundamentals, the hallmark of the asset approach. This link is introduced by adding the UIP relation described
in equation (6.4) above, which pins down expected returns on shortterm interest-bearing assets. Using the UIP relation that i US; t i UK; t
EDp $=; t1 j W t , and suppressing the $/ subscript, we can write
pt f t bEDpt1 j W t ;

6:9

where we have dened f t , the current fundamental, as

f t m US m UK ay US y UK :

6:10

The rational expectations solution to equation (6.9) brings the link to

future fundamentals into sharp focus:9
pt

y
X
i0

gi E f ti j W t ;

6:11

where gi 1 b=1 b i =1 b. Today's spot rate therefore depends

not only on the current state of monetary fundamentals, but also on
the stream of expected future fundamentals. This asset-approach
representation is analogous to the dividend-discount model used in
nance for equity valuation. The ``discount'' rate here, gi , is wholly a
function of b, the sensitivity of money demand to the interest rate
(equation 6.6).10
Sticky-Price Monetary Model
The sticky-price monetary model, also called the overshooting
model, has the same conceptual underpinnings as its exible-price
cousin (the key reference is Dornbusch 1976). The models share two
essential ingredients. First, at their core is equilibrium in the money
market, described above in equation (6.6). Second, the sticky-price

158

Chapter 6

model also assumes UIP, described above in equation (6.4); this is

the ingredient that places the model, like the exible-price model,
squarely in the asset approach. Two of its ingredients, however, are
substantial departures from the exible-price model. First, prices are
sticky over the short run, adjusting only gradually to the long-run
exible-price equilibrium. (The level of aggregate demand governs
this adjustment.) Second, PPP does not hold in the short run, though
it does hold in the long run, once the price level has fully adjusted to
its exible-price level. 11
The equation summarizing the sticky-price model is similar to that
for the exible-price model in equation (6.9):
pt f t bEDpt1 j W t wt

6:12

The only difference is the additional term wt , a wedge term that does
not arise when prices are exible (see, e.g., Flood and Taylor 1996).
The wedge term captures the short-run departure from PPP caused
by sticky prices. Consider for example a shock that shifts the money
supply m, starting from a steady state in which wt 0. This shift can
have an immediate and substantial impact on the exchange rate pt
because that variable is perfectly exible. Because the price level p US
is not perfectly exible, however, this results in a departure from
PPPwt is no longer zero. This effect on wt is transitory, though,
because in the long run prices adjust fully. The difference between
the exible- and sticky-price models is therefore transitory as well.
The sticky-price model is appealing not only because it relaxes the
exible-price model's uncomfortable assumptions, but also because
it can amplify the effect of a change in fundamentals, referred to as
exchange rate overshooting. The appeal of this overshooting result is
that it squares with empirical ndings that exchange rates are excessively volatile when compared with the volatility of fundamentals as
specied in, say, the exible-price model. (Note that high volatility
relative to exible-price fundamentals should not be confused with
high absolute volatility; indeed, changes in major spot rates have
a standard deviation two-thirds that of equity returns: roughly 12
percent per annum, versus 18 percent for the S&P 500.)
The overshooting result is important enough in exchange rate
economics that it deserves attention. A simple experiment provides
some clear intuition. Consider the case of an unexpected, permanent
increase in the money supply, m US . First, in the exible-price model,
the effect is straightforward: a 10 percent increase in money induces

Exchange Rate Models

159

Figure 6.1
Illustration of overshooting.

an immediate 10 percent increase in the price level p US (price exibility) and an immediate 10 percent depreciation of the dollar (to restore PPP, which holds at all times). Thus, there is no overshooting,
nor undershooting, because the exchange rate jumps to its long-run
10 percent-depreciated level.
In the sticky-price model, overshooting is best understood using
the UIP relation (equation 6.4). Suppose that before the increase in
money, the economy is in steady state, so i $; t i ; t , EDpt1 j W t 0,
and the price level is not adjusting. Figure 6.1 shows this steady
state, before the time labeled t. Now, a key difference in the stickyprice model is that by assumption the increase in money has no immediate effect on the price level. Therefore, from equation (6.6), the
increase in money supply requires an immediate fall in i $ to clear the
money market, so that now i $ < i ; this is the usual liquidity effect of
an increase in the money supply (output y is xed in simple versions
of the model). Consider the immediate effect of the fall in i $ on the
exchange rate: the dollar should depreciate (an increase in p $= ), but
by how much? Recall that the exible-price model produces an immediate 10 percent dollar depreciation to its new long-run level
neither overshooting nor undershooting. Suppose the sticky-price
model is the samean immediate 10 percent dollar depreciation to
the same long-run level (pinned down by PPP), with EDpt1 j W t 0
thereafter, labeled as point A in gure 6.1. But point A cannot be an

160

Chapter 6

equilibrium: if i $ < i , then UIP, which holds by assumption, is violated (i.e., i $ < i and EDpt1 j W t 0 are incompatible). This rules
out equilibrium at point A. Either overshooting or undershooting
must occur.12
To see which is the right answer (overshooting or undershooting),
we need to determine which is consistent with UIP. Note from
equation (6.4) that with i $ < i , it must be that EDpt1 j W t < 0, that
is, the dollar must be expected to appreciate in the future (a lower
dollar price of the pounda negative Dpt1 is dollar appreciation).
Intuitively, this future dollar appreciation is needed to compensate
dollar depositors for a lower interest rate.13 But the only way that the
dollar can appreciate toward its long-run 10-percent-depreciated
level in the future is if the dollar depreciates today by more than 10
percent, making room for the necessary trend appreciation. No other
path is compatible with UIP. The dollar must overshoot.
An important reason I have included additional detail on this
model is that it provides some intuition for the impact of a change in
interest rates on the exchange rate. Understanding this interest rate/
exchange rate link is important for the following two macro-oriented
chapters. Note from gure 6.1 that the drop in i $ induces immediate
dollar depreciation (an increase in the dollar price of a pound). This
is consistent with most people's intuition: lower dollar interest
rates make dollar deposits less attractive, other things equal, causing
portfolios to shift out of the dollar.
Portfolio Balance Model
Though still squarely anchored in the asset approach, the portfolio
balance model departs from the monetary models in two essential
ways (key references include Kouri and Porter 1974 and the survey
by Branson and Henderson 1985). First, of the four asset-approach
models, this is the only one without PPP as an ingredient. This
means the long-run exchange rate must be pinned down some other
way. Second, this model does not impose UIP, and thus, expected
dollar returns on different-currency deposits need not be equal. This
leaves room for a currency risk premium, that is, an expected return
bonus for holding particular currencies. The macro literature refers
to this as imperfect substitutability between domestic and foreign
assets (the same term introduced in the micro material in chapter 2).
The portfolio balance model is, as its name suggests, a model
which balances demand for various asset classes against supply. The

Exchange Rate Models

161

exchange rate brings them into balance. The exchange rate achieves
this by affecting demand/supply in two ways. First, an expected
change in the exchange rate affects foreign-asset demand because it
has a direct effect on the expected dollar return on foreign assets (the
right side of equation 6.4). Second, the level of the exchange rate
affects foreign-asset supply. This works over time through a traditional macro channel, the trade balance: a lower value of the dollar
pushes the trade balance toward surplus, which, through the balance
of payments, increases foreign assets in domestic portfolios.14
A simple specication of this model includes three demand functions, one for each of three available asset classes: money (M), domestic bonds (B), and foreign bonds in foreign currency (B  ). Each of
these demands depends on the same two variables: the domestic
nominal interest rate i and the expected dollar return on foreigncurrency bonds i  E%DP, where i  is the foreign nominal interest
rate and E%DP is the expected percent depreciation of the dollar. (It
is best to think of these as short-term bonds with no capital gains/
losses from interest rate changes.)
Money Demand M D i; i  E%DP
M1D

< 0;

M2D

with

<0

6:13

Bond DemandaaDomestic B D i; i  E%DP

B1D > 0;

B2D < 0

6:14


Bond DemandaaForeign B D i; i  E%DP



B1D < 0;

with

B2D > 0

with
6:15

An asterisk denotes foreign variables, and subscripts denote partial

derivatives. These equations embed the second exchange rate effect
noted abovethe effect of expected changes in the spot rate P. The
signs of the partial derivatives that appear below each equation are
sensible. Other things equal, an increase in the domestic interest
rate i reduces money demand (higher opportunity cost), increases
the demand for domestic bonds (higher return), and reduces the demand for foreign bonds (lower relative return). An increase in the
expected dollar return on foreign-currency bonds reduces money
demand (lower relative return), reduces the demand for domestic
bonds (lower relative return), and increases the demand for foreign
bonds (higher return). Note that wealth does not affect relative
demands here: By assumption, changes in wealth induce equal per-

162

Chapter 6

cent changes in each of the three demands, so their portfolio shares

are unaffected.
We need two additional equations to complete the model:
W M D B D PB D


DB S TP i  B D

6:16
with

T1 > 0:

6:17

Equation (6.16) is the wealth constraint: wealth must be allocated to

these three asset classes. Equation (6.17) imposes a constraint on the
supply of foreign-currency assets held by domestic residents: from
balance of payments identities, changes in domestic holdings of foreign assets must equal the current account (per note 14); the current
account in turn has two components, the trade balance TP and net

interest receipts on foreign bonds, i  B D . From the partial derivative
T1 > 0, we see that the trade balance depends positively on the level
of the exchange rate P. That is, a depreciation (more dollars per
pound) pushes the trade balance toward surplus, which must be
matched with an increase in net foreign assets. This is the second of
the two exchange rate effects noted abovethe effect of the exchange
rate's level on foreign asset supply.
What drives the exchange rate in this approach are changes in relative asset supplies. To see this, note that a market-clearing exchange
rate can be derived for particular supply levels of M, B, and B  .
Subsequent changes in these supplies have both short-run effects on

the exchange rate (allowing a non-zero DB S in equation 6.17) and

long-run effects (after steady state has been achieved with DB S 0).
Note, too, that there is nothing in the model that equates expected
dollar returns on different currency deposits, that is, nothing that
equates i with i  E%DP. Indeed, it is precisely these departures
from UIP that are needed to clear markets. These departures are
the currency risk premium introduced abovethe expected return
bonus for holding particular currencies.
For estimating this model from real-world data, more structure
is added to both the demand and supply sides. On the demand
side, one typically assumes Normally distributed returns and meanvariance optimization. On the supply side, determining which asset
classes to measure is the main issue. Typically, instead of using
bonds, one uses outstanding government debt. The rationale for
focusing on government debt is that changes in supply necessarily
imply changes in net private holdings. For many privately issued

Exchange Rate Models

163

asset classes (such as bank deposits or corporate bonds), in contrast,

changes in net supply are not possible because for every private issuer
there is a private holder. Another difcult issue on the supply side is
that, whatever asset classes one chooses, it is hard to measure their
levels with any precision. Because measuring supply changes is easier,
this model is typically tested for its ability to account for exchange
rate changes, rather than its ability to account for exchange rate
levels. Even as an account of changes, though, the model does not
fare well empirically (see Branson and Henderson 1985; Lewis 1988).
An Analogy: Exchange Rate Models and Microstructure Models
I close this section with an analogy that links the microstructure
models of chapter 4 with the exchange rate models of this chapter.
The analogy is based on whether a model's focus is primarily on
return, primarily on risk, or balanced between the two. As noted in
the introduction to chapter 4, there are two traditional modeling
approaches in microstructure, the information approach and the inventory approach. The information approach focuses more on return; risk plays no direct role in these models because players are
risk neutral (as in the Kyle and sequential-trade models of chapter 4),
so prices contain no risk premium. The inventory approach, in contrast, focuses more on risk. Under the inventory approach, imperfect
risk sharing takes center stage and information effects on returns are
not present. More recently, a hybrid approach has emerged within
microstructure with a more balanced treatment of both risk and return (the simultaneous-trade model of chapter 4 being one example).
There is a similar pattern in the asset-approach models of this
chapter. Like microstructure's information models, the two monetary
models focus on how information affects returns; risk, on the other
hand, has no effect on those returns. Like microstructure's inventory
models, the portfolio balance model focuses on risk and the determination of equilibrium risk premia. In terms of information and returns,
basic portfolio balance models are rather simplisticthey abstract
from key categories of fundamentals such as output, relative goods
demand, and to a great extent monetary policy. Finally, the analogue
of microstructure's hybrid model is the general equilibrium model of
exchange rates (the focus of the next section): in both cases there is
considerable richness in both the impounding of information in price
and the determination of risk premia.

164

6.3

Chapter 6

Microfoundations: Tastes and Technology

The terms ``microfoundations'' and ``tastes and technology'' are common in macroeconomics, but less familiar outside. Microfoundations refers to the grounding of analysis in well-dened decision
problems faced by individuals or groups. Grounding analysis this
way is standard within microeconomics, hence the prex micro. In
microeconomic analysis, individuals have a clear sense of both their
objectives and constraints, and they do their best to fulll their
objectives. These well-dened objectives are called ``tastes'' because
they embody the preferences of individuals for various outcomes.15
The constraints in macro models with microfoundations are called
``technology.'' Technology encompasses the production side of the
real economy (or the endowment process in a pure exchange economy) and denes the feasible choicesthe size of the pie.
Though the three macro models reviewed in the previous section
are not grounded in well-dened problems faced by individuals, the
fourth of our macro modelsthe general equilibrium (GE) model
is. The microfoundations of the GE exchange rate model are the same
as those adopted more generally in macroeconomicsthe ``two T's''
of tastes and technology. As one might imagine, though, a grounding in tastes and technology can still accommodate many different
specications. Here I review the specication of the GE model only
at a broad level. Though details are missing, a broad perspective is
sufcient for clarifying how the GE model's microfoundations differ
from the microfoundations of the microstructure approach.16
General Equilibrium Model
The GE model of exchange rate determination begins with maximization of a representative individual's utility.17 Specication of
utility as a function of various consumption opportunities has, naturally, important implications for equilibrium exchange rates. (The
microstructure approach also includes utility maximization, but as
we saw in chapter 4, utility is specied very simply, typically in
terms of terminal nominal wealth.) The rm grounding in utility is
the taste part of the model's microfoundations. The technology part
of the model's microfoundations lies in the models' tightly specied
production and transaction technologies. The production technology
summarizes all input-output relationships in the production of real

Exchange Rate Models

165

goods. The transaction technology summarizes the rules that apply

to transactions. For example, early GE exchange rate models require
individuals to hold domestic currency for purchasing domestic or
foreign goods (in the literature this is called a ``cash-in-advance''
constraint). This requirement is important for exchange rate determination because it drives the demand for currencies. (Later GE
models produce demand for money from the tastes side by putting
money directly in the utility function.)
Unlike the three macro models above, however, the focus in GE
models is on the real exchange rate, dened as
Real exchange rate P$= P UK =P US :
The real exchange rate is the nominal exchange rate P$= adjusted by
the relative price of goods. (Though I have not used it yet, the term
nominal exchange rate is customarily used for P$= when there is any
possibility of confusion.) The real exchange rate is determined in this
model as the relative price of foreign-to-domestic goods that matches
the marginal rate of substitution in consumption among those goods
(or, as in Obstfeld and Rogoff 1995, the marginal rate of substitution
among inputs in nal-good production). Given this real rate, to produce an expression for the nominal exchange rate P$= one can then
substitute for P US and P UK using the expressions for money demand
in each country from equation 6.6. (Recall that this is the same procedure we followed for the exible-price model, except that in that
case the real exchange rate was assumed xed at one, a direct implication of assuming PPP always holds.)
Early GE exchange rate models are generalizations of the exibleprice monetary model, allowing multiple goods and real shocks
(such as shocks to demand or productivity). More recent GE models
sometimes called new open-economy macro modelsare generalizations of the sticky-price monetary model (see Obstfeld and Rogoff
1995, 1996). These models allow for sticky nominal prices, which
permits departures from PPP, in much the same spirit as the departures from PPP that arise in the original sticky-price model.
One major contribution of the GE models is their value for
conducting welfare analysis. This comes from their well-specied
microfoundations on both the demand (tastes) and on the supply
(technology) sides of the economy. The recent GE models, for example, allow welfare analysis of xed versus oating exchange rate
regimes (see, e.g., Devereux and Engel 1999).

166

Chapter 6

Another feature of GE models that warrants attention is their

treatment of risk. The representative individual in this model is risk
averse, so asset classes, including currency, can earn risk premia. The
representative-individual framework, however, requires that risks be
perfectly shared, so there are no unsystematic risks borne in equilibrium, only systematic risks. This is natural for a general equilibrium
approach. But within microstructure, for example, there are risks
that affect prices that are not systematic in this sense (e.g., the inventory risk covered in chapter 2).
Empirically, GE modelslike the other three macro modelshave
not yet produced exchange rate equations that t the data. They have
been unable to overturn or explain the negative ndings of Meese
and Rogoff (1983a). Within the context of the macro-asset approach,
why oating exchange rates behave as they do over horizons up to a
couple years remains a puzzle.
6.4

Microfoundations: Information and Institutions

Unlike the micro-foundations in exchange rate economics, with its

focus on well-dened tastes and technology, the microstructure
approach is grounded in what I call the ``two I's'': information and
institutions. This distinction is one of emphasis. General equilibrium
exchange rate models certainly have an information dimension, but
it is typically simple, incorporating only public information. Microstructure models certainly include tastes and utility maximization,
but the specication of those tastes is simple, typically limited to
nominal wealth levels rather than including factors that are, by their
nature, real quantities, such as consumption and leisure.
Let me clarify the difference between technology and institutions.
Recall from the GE macro model that ``technology'' refers primarily
to the supply side of the real economy, whereas ``tastes'' refers to the
demand side. In microstructure modeling, however, the supply side
of the real economy plays no direct role.18 Institutions, on the other
hand, do play a central role in microstructure models. Institutions
embody the ``rules of the game'' that govern the interaction between
agents and, therefore, affect outcomes. (By rules of the game, I have
in mind chapter 4's summaries of each model's key features, in particular the part labeled ``institutions.'' These institutional features
also inuence the other two parts of the model summaries, i.e.,
which players participate and what information is available to each

Exchange Rate Models

167

player.) Because there is a sense in which the GE model's technology

``precedes'' interaction, it is arguably deeper, more ``primitive,'' than
institutions. That is a theoretical plus. But the proof is in the empirical pudding.19
Continuing on this line of comparing/contrasting the macro- and
microstructure approaches, gure 6.2 relates the models of this
chapter to the microstructure approach. I introduced the rst of the
models shown, purchasing power parity (PPP), in section 6.1 as a
goods market perspective on exchange rate fundamentals. It was
not, however, grounded in microeconomic analysis of well-dened
problems faced by individuals. The same is true of the rst three of
the four asset-approach models: the exible-price monetary model
(FPM), the sticky-price monetary model (SPM), and the portfolio
balance model (PBM). In contrast, the fourth asset-approach model
the GE modelis grounded in well-dened problems faced by
individuals, with an emphasis on tastes and technology. The microstructure models of chapter 4 emphasize information and institutions
instead.
Let me close this section on the two ``I's'' of the microstructure
approach by drawing attention to their interaction. Consider a simple example of how institutions can affect the processing of information, and ultimately whether that information is impounded in
price. In chapters 1 and 5, I addressed the volume amplication
effects of the FX hot potato, which results from the passing of
unwanted inventories from dealer to dealer following an initial customer trade (i.e., it results from risk management in a multipledealer setting). Now, if (1) information is impounded in price via
learning from order owas is common to all the microstructure
models of chapter 4and (2) order ow is noisier as a result of
dealer risk management, then the institutional structure can produce
less informative prices. This was one of the main insights of the
simultaneous-trade model (section 4.4). Interaction between risk
management and information processing is an exciting area of study
within microstructure research.
6.5

A False Dichotomy: Micro Issues versus Macro Issues

Though the notion of microfoundations helps to organize thinking, the orientation can be misleading because issues in exchange
rate economicsas elsewhere in nancial economicsdo not divide

168

Chapter 6

Figure 6.2
Three approaches to exchange rates and their models. PPP denotes purchasing power parity model. FPM denotes exibleprice monetary model. SPM denotes sticky-price monetary model. PBM denotes portfolio balance model. GE denotes
general equilibrium model.

Exchange Rate Models

169

Figure 6.3
Issues spectrum.

neatly into ``micro'' issues and ``macro'' issues. Instead, issues fall
along a spectrum, as gure 6.3 shows.
The more micro end of the spectrum includes issues such as optimal market design, efcient regulation, and the determination of
transaction costs. These are bread and butter issues within microstructure research. The more macro end includes issues such as
medium/long-run exchange rate determination, forward-discount
bias, and home bias (the rst two of which I address in detail in
the next chapter). Between the extremes are issues such as volatility
determination, central bank intervention, and transaction taxes
(the rst of which I address in chapter 7 and the second in chapter
8).20
It is important, in my judgment, to distinguish between tools and
the issues to which they are applied. I introduce the issues spectrum
not only to frame the later chapters, but also because it helps explain
the gap between exchange rate economics on the one hand and
microstructure nance on the other. Within microstructure nance,
the tools have been applied to bread and butter microeconomic
issues that, for the most part, do not interest macro-oriented exchange rate economists. Though one might suppose that because
these tools were developed for micro issues they are not likely to
help with macro issues, they are actually quite useful. The analysis in
chapters 7 and 8 (on longer horizon exchange rate determination,
central bank intervention, etc.) will demonstrate this.
An example may clarify why exchange rate economists view
microstructure with skepticism. The example comes from the call
for papers that initiated a market microstructure research group
in the United States (the group is part of the National Bureau of
Economic Research). The call for papers came with a list of topics
a list that was important for dening the scope and spirit of
the group's agenda. Though all eleven topics are bread and butter
microstructure, macroeconomists might quickly dismiss them. The
topics include:

170

.
.
.
.
.
.
.
.

Chapter 6

minimum tick-size and competitiveness

measurement and control of transaction costs
regulation and the evolution of market design
denition, measurement, and determinants of liquidity
costs and benets of alternate trading mechanisms
cross-border listing and trading of securities
trading technology and information systems

implications of market design for risk management and nancial

engineering
These topics do not exactly make the macroeconomist's heart utter.
The remaining three topics were

. globalization of nancial markets

. the role of information in price discovery
. domestic and international comparisons of trading costs
Of the eleven topics, only the rst two of these last three would catch
the eye of most macroeconomists, but these two are broad enough to
catch most anybody's eye. It is not clear how microstructure tools
could be relevant to these topics' macro aspects; the link between the
last topic and macroeconomics is not obvious either, unless one is
familiar with the literature linking trading costs to the home bias
puzzle. That literature shows that foreign holdings are turned over
more than domestic holdings, which is difcult reconcile with home
bias being due to higher foreign trading costs (see Tesar and Werner
1995, and note 3 of chapter 1).
The upshot is that researchers working with microstructure tools
tend not to apply them to issues on the macro end of the issues
spectrum, and those who are working on the macro end have not felt
it worthwhile to invest in apparently inappropriate microstructure
tools. For these tools to be applied to more FX issues on the macro
end, either macroeconomists need introduction to the tools or those
with the tools need to extend their study rightward on the spectrum.
Both, in fact, are occurring. As the process continues, the uneasy
dichotomy between macro and micro approaches will fade.

Macro Puzzles: The

Challenge for
Microstructure

This chapter examines traditional exchange rate puzzles and shows

how microstructure tools are used to address them. The chapter is
not intended to put these puzzles to rest: they wouldn't be traditional if they weren't stubborn. My intent is to provide a sense for
how to address ``macro'' puzzles by looking under the ``micro lamppost.'' For those unfamiliar with the lamppost metaphor, it comes
from a parable of a guy who drops his keys while getting into his car,
in the middle of a dark parking lot. Another fellow sees him searching over in a corner of the lota well-lit cornerand asks,
``What are you doing?''
``Looking for my keysI dropped them getting into my car.''
``Then why are you looking way over here?''
``Because this is where the light is.''
Is there any hope of nding the exchange rate keys under the
microstructure lamppost? Chapter 1 provides some evidence that the
microstructure lamppost is not so far removed. In this chapter, I
extend that evidence and show that the lamppost has indeed cast
light on the key puzzles. I also identify directions for future application of the microstructure approach.
The three puzzles I consider are the ``big three'' introduced in
chapter 1:

. The determination puzzle: exchange rate movements are virtually

unrelated to our best measures of fundamentals

. The excess volatility puzzle: exchange rates are excessively volatile

relative to our best measures of fundamentals

. The forward bias puzzle: excess returns in foreign exchange are

predictable and inexplicable

172

Chapter 7

The rst section addresses the determination puzzle by using order

ows. It draws on recent empirical work by Evans and Lyons (1999).
In the second section, I contrast order ow analysis with an older
approach, the so-called ow approach to exchange rates, which uses
balance of payments ows (e.g., current accounts and capital accounts), not order ows. Section three returns to the big three
puzzles, in particular the excess volatility puzzle, drawing on empirical work by Killeen, Lyons, and Moore (2000a). The last section
addresses the forward bias puzzle. The explanation I offer for this
puzzle is new to the literature.
7.1

The Determination Puzzle

Exchange rate economics has been in crisis since Meese and Rogoff
(1983a) showed that our models are empirical failures: the proportion of monthly exchange rate changes our textbook models can
explain is essentially zero. In their survey, Frankel and Rose (1995,
1704, 1708) summarize as follows:
The Meese and Rogoff analysis at short horizons has never been convincingly overturned or explained. It continues to exert a pessimistic effect on the
eld of empirical exchange rate modeling in particular and international
nance in general. . . . Such results indicate that no model based on such
standard fundamentals like money supplies, real income, interest rates, ination rates, and current account balances will ever succeed in explaining or
predicting a high percentage of the variation in the exchange rate, at least at
short- or medium-term frequencies.1

This is the determination puzzle. Immense effort has been expended

to resolve it.2
If determinants are not macro fundamentals like interest rates,
money supplies, and trade balances, then what are they? Two alternatives have attracted a lot of attention among macroeconomists.
The rst is that exchange rate determinants include extraneous variables, which are typically modeled as speculative bubbles. (A bubble
is a component of an asset's price that is nonfundamental. A bubble
can cause price to rise so fast that investors are induced to buy, even
though the bubble may burst at any time. See, e.g., Meese 1986;
Evans 1986.) On the whole, however, the empirical evidence on
bubbles is not supportive: in their survey, Flood and Hodrick (1990)
conclude that existing evidence is unconvincing. A second alternative to macro fundamentals is nonrational behavior. For example,

Macro Puzzles

173

exchange rates may be determined, in part, from avoidable expectational errors (Dominguez 1986; Frankel and Froot 1987; Hau 1998).
On a priori grounds, many nancial economists nd this second alternative unappealing. Even if one is sympathetic, however, there is a
wide gulf between the presence of nonrational behavior and accounting for exchange rates empirically.3
This section addresses the determination puzzle using the microstructure approach, drawing heavily from work presented in Evans
and Lyons (1999). One advantage of the microstructure approach is
that it directs attention to variables that have escaped the attention of
macroeconomists. Meese (1990, 130) offers a telling quote along these
lines: ``Omitted variables is another possible explanation for the lack
of explanatory power in asset market models. However, empirical
researchers have shown considerable imagination in their specication searches, so it is not easy to think of variables that have escaped
consideration in an exchange-rate equation.'' Among the variables
escaping consideration, order ow may be the most important. As
we saw in chapter 4, order ow is the proximate determinant of price
in microstructure models, regardless of institutional structure. This
ensures that the causal role played by order ow is robust to different market structures.
A Hybrid Model with Both Macro and Micro Determinants
To establish a link between the micro and macro approaches, chapter
1 introduces models with components from both. The hybrid model I
suggest in that rst chapter takes the form:
DPt f i; m; z gX; I; Z et
where the function f i; m; z is the macro component of the model
and gX; I; Z is the microstructure component. Chapters 4, 5, and 6
provide explicit specications for how interest rates i, money supplies m, order ows X, and inventories I might enter these two functions. (Recall that z and Z denote unspecied other determinants.)
An important take-away from those chapters is that f i; m; z and
gX; I; Z depend on more than just current and past values of
their determinantsthey also depend, crucially, on expectations of
determinants' future values. This stands to reason: rational markets
are forward looking, so these expectations are important for setting
prices today.

174

Chapter 7

Though I have split this stylized hybrid model into two parts, the
two parts are not necessarily independent. This will depend on the
main micro determinantorder ow Xand the type of information it conveys. Per chapters 1 and 2, order ow conveys two main
information types: payoff information and discount rate information. In macro models, information about future payoffs translates to
information about future i; m; z. One way order ow can convey
information about future i; m; z is by aggregating the information
in individuals' expectations of i; m; z. (Recall that as a measure of
expectations, order ow reects people's willingness to back their
beliefs with money; and like actual expectations, this measure
evolves rapidly, in contrast to measures derived from macro data.)
When order ow conveys payoff information, macro and micro
determinants are interdependent: order ow acts as a proximate determinant of price, but standard macro fundamentals are the underlying determinant.4
If order ow X conveys discount rate information only, then the
two sets of determinants i; m; z and X; I; Z can indeed be independent. To understand why, suppose the discount rate information
conveyed by order ow X is about portfolio balance effects (e.g.,
persistent changes in discount rates, due to changing risk preferences, changing hedging demands, or changing liquidity demands
under imperfect substitutabilitysee chapter 2).5 Now, consider the
two monetary macro models (exible and sticky-price). Portfolio
balance effects from order ow X are unrelated to these models'
specications of f i; m; z. This is because the monetary models assume that different-currency assets are perfect substitutes (i.e., they
assume that Uncovered Interest Parity holds: assets differing only in
their currency denomination have the same expected return). Thus,
effects from imperfect substitutability are necessarily independent of
the f i; m; z of these monetary models. In the case of the macro
portfolio balance model, in contrast, portfolio balance effects from
order ow X are quite likely to be related to the determining variables i; m; z. Indeed, in that model, price effects from imperfect
substitutability are the focus of f i; m; z.6
Consider the hybrid model from a different perspectivea graphical perspective. The top panel of gure 7.1 illustrates the connection
between fundamentals and price under the traditional macro view
(i.e., as reected in the models of chapter 6). Information about
fundamentals is publicly known, and so is the mapping from that

Macro Puzzles

175

Figure 7.1
Spanning macro and microstructure graphically. The top panel illustrates the connection between fundamentals and price under the traditional macro view (i.e., as
reected in the models of chapter 6): information about fundamentals is public, and so
is the mapping to price, so price adjustment is direct and immediate. The middle panel
shows the traditional microstructure view (as reected in the models of chapter 4). The
focus in that case is fundamental information that is not publicly known. This type of
information is rst transformed into order ow, which becomes a signal to the price
setter (e.g., dealer) that price needs to be adjusted. Actual markets include both, which
is illustrated in the bottom panelthe hybrid view.

176

Chapter 7

information to price. Consequently, price adjustment is direct and

immediate. The middle panel shows the traditional microstructure
view (as reected in the models of chapter 4). The focus in that case
is on fundamental information that is not publicly known. In those
models, information is rst transformed into order ow, which
becomes a signal to the price setter that price needs to be adjusted.
The bottom panel presents the hybrid view. Here, the model accommodates both possibilities: information that affects price directly
and information that affects price via order ow. Armed with
models that allow for both, we can let the data determine their relative importance.
Before describing the hybrid model estimated by Evans and Lyons
(1999), let me address some front-end considerations in modeling
strategy. First, the determination puzzle concerns exchange rate behavior over months and years, not minutes. As we saw in chapter 5,
most empirical work in microstructure is estimated at the transaction
frequency. The rst order of business is to design a trading model
that makes sense at lower frequencies. Several features of the EvansLyons model contribute to this (as I will note specically below, as
I present the features). Second, because interdealer ow is more
transparent, it is more immediately relevant to FX price determination than customer-dealer order ow. The hybrid model should reect this important institutional feature. (Of the chapter 4 models,
only the simultaneous-trade model has this feature.) Third, the
model should provide a vehicle for understanding the behavior of
order ow in gure 1.2. That gure presents cumulative interdealer
order ow in the $/DM and $/yen markets over the four-month
Evans (1997) data set, the same data set used by Evans and Lyons
(1999). A puzzling feature is the persistence: there is no obvious evidence of mean reversion in cumulative order ow. How can this be
consistent with the ndings reviewed in chapter 5, where individual
dealer inventories have a very short half-life (i.e., their positions
revert to zero rapidly)? The Evans-Lyons model accounts for this
seeming incongruity.
A Daily Frequency Model
The model of is a variant of the simultaneous-trade model. Because I
reviewed that model in chapter 4, I will focus on substantive differences. For reference, note that I also summarize the Evans-Lyons

Macro Puzzles

177

Figure 7.2
Daily timing in the Evans-Lyons model.

model in gure 7.3 (in the same manner that I use gures to summarize each of the models of chapter 4).
There are N dealers in the model, indexed by i, a continuum of
nondealer customers (the public), and an innite number of trading
days. Dealers and customers all have negative exponential utility.
Within each day, there are three rounds of trading:
Round 1: Dealers trade with the public
Round 2: Dealers trade among themselves to share risk
Round 3: Dealers trade with the public to share risk more broadly.
Figure 7.2 shows the timing within each day.
At the beginning of each day, the payoff to holding foreign exchange is Rt , which is composed of a series of increments DRt , so that
Rt

t
X

DRt :

7:1

t1

The payoff increments DRt are i.i.d. Normal0; sR2 and are observed
publicly at the beginning of each day. These realized increments
represent the ow of public macroeconomic informationthe macro
component of the model f i; m; z. For concreteness, one can think of
this abstract payoff increment DRt as changes in interest rates.
Per gure 7.2, after observing the payoff Rt , each dealer sets a
quote for his public customers. As in the simultaneous-trade model,
quotes are scalar two-way prices, set simultaneously and independently.7 Denote this dealer i quote in round 1 of day t as Pit1 . Evans
and Lyons show that in equilibrium, all dealers choose to quote the
same price, denoted Pt1 . Each dealer then receives a customer-order
realization Cit1 that is executed at his quoted price Pt1 , where Cit1 < 0 denotes a customer sale (dealer i purchase). Each of these N customer-

178

Chapter 7

Figure 7.3
Summary of the Evans-Lyons (1999) model.

order realizations is distributed:

Cit1 @ Normal0; sC2 ;
and they are uncorrelated across dealers. Importantly, the Cit1 realizations are not publicly observable. It will be useful to dene the
aggregate public demand in round 1 as
Ct1

N
X
i1

Cit1 :

One important choice in specifying the model is the correlation

between customer orders Cit1 and the stream of payoff increments
DRt . This choice determines whether the macro and micro components of the modelthe f i; m; z and gX; I; Zare interdependent.
If there is no correlation, it is not possible for order ow to convey
payoff information. Because Evans and Lyons (1999) have only four

Macro Puzzles

179

months of order ow data, they are unable to determine empirically

whether order ow conveys payoff information, discount rate information, or both. Because of this, they choose to model the customer
orders Cit1 as distributed independently of the payoff stream Rt
arguably, a less controversial choice. This means that, in their model,
the only kind of information that order ow can convey is discount
rate information. Because their model rules out inventory effects at
the daily frequency (as we shall see below), the discount rate information in their model is necessarily about portfolio balance effects.
Round 2 is the interdealer trading round. Each dealer simultaneously and independently quotes a scalar two-way price to other
dealers Pit2 . These interdealer quotes are observable and available to
all dealers in the market. Evans and Lyons show that, like in round 1,
all dealers choose to quote the same price, denoted Pt2 . Each dealer
then simultaneously and independently trades on other dealers'
quotes. (Orders at a given price are split evenly across any dealers
quoting that price.) Let Tit denote the (net) interdealer trade initiated
by dealer i in round 2 of day t.8
Importantly, at the close of round 2 all dealers observe the net
interdealer order ow on that day:
Xt

N
X

Tit :

7:2

i1

This order ow information is important to the model because it

conveys the size and sign of the public order ow in round 1. To
understand why, consider the interdealer trading rule derived by
Evans and Lyons:
Tit aCit1 ;

7:3

where a is a constant (positive) coefcient. Each dealer's trade in

round 2 is proportional to the customer order he receives in round 1.
This implies that when dealers observe the interdealer order ow
Xt Si Tit aCt1 , they can infer the aggregate public order ow Ct1 in
round 1.
In round 3, dealers share overnight risk with the non-dealer public. This feature is important in distinguishing this model from its
high-frequency cousin, the simultaneous-trade model. (In the simultaneous-trade model the public does not re-enter, and therefore
cannot share risk borne by dealers.) Unlike round 1, the public's
trading in round 3 is nonstochastic. To start the round, each dealer

180

Chapter 7

simultaneously and independently quotes a scalar two-way price

Pt3 (also common across dealers). These quotes are observable and
available to the public at large.
A crucial assumption made by Evans and Lyons is that dealers set
prices in round 3 such that the public willingly absorbs all dealer
inventory imbalances, so that each dealer ends the day with no
net position.9 As an empirical matter, it is common practice for FX
dealers to end each day with no net position, which squares with the
empirical ndings in chapter 5. Note too that this assumption rules
out inventory effects on price at the daily frequency (because dealers
do not hold overnight positions that require compensation). The
round-3 price that dealers actually quote to induce public absorption
of these imbalances depends on the round-2 interdealer order ow
Xt : this interdealer order ow informs dealers of the size of the total
position that the public needs to absorb (as noted, Xt aCt1 ).
More precisely, to determine the round-3 price, dealers need to
know two things: the total position that the public needs to absorb
(which they learn from Xt ), and the public's risk-bearing capacity.
Regarding the latter, the public's capacity for bearing foreignexchange risk is assumed less than innite (i.e., Evans and Lyons
assume that foreign- and domestic-currency assets are not perfect
substitutes.) This is the key assumption: it makes room in the model
for portfolio balance effects on price. Consistent with negative exponential utility, the public's total demand for foreign exchange in
round 3, denoted Ct3 , is a linear function of its expected return conditional on public information:
3
Rt1 j Wt3 :
Ct3 gEDPt1

7:4

The positive coefcient g captures the aggregate risk-bearing capacity

of the public: A larger g means the public is willing to absorb a larger
foreign-exchange position for a given expected return. Wt3 is the
public information available at the time of trading in round 3 (which
includes all past Rt and Xt ).
The Pricing Relation
Evans and Lyons (1999) show that the price at the end of day t is:10
Pt b 1

t
X

t1

DRt b 2

t
X
t1

Xt :

7:5

Macro Puzzles

181

The change in price from the end of day t 1 to the end of day t can
therefore be written as:
DPt b 1 DRt b2 Xt :

7:6

where b 2 is a positive constant (that depends on g and a).11 It is

not surprising that this price change includes the payoff increment
DRt : upon realization, the increment DRt becomes a known (i.e., riskfree) component of the continuing daily payoff Rt , and its discounted
value is impounded in price (b1 ).
Let me provide some intuition for the portfolio balance effectthe
b 2 Xt term. This term is the required price adjustment that induces reabsorption of the random order ow Ct1 that occurred at the beginning of the day. The value of the parameter b2 insures that at the
round-3 price:
Ct1 Ct3 0;
that is, that the dealers have no net overnight position. To understand the link to order ow, recall that the round-3 price depends on
two things: the public's risk-bearing capacity (summarized by g), and
the total position that the public needs to absorb. As noted, dealers
learn about the total position the public needs to absorb from order
ow Xt . This produces the relation between interdealer order ow
and the subsequent price adjustment.
Let's walk through an example. Consider the price at the close of
day t, as described by equation (7.5). The next day's increment to the
daily payoff R, DRt1 , is uncertain, but all previous realizations of
the payoff increment DR are known and are impounded in price.
(Expectations of future realizations do not enter equation 7.5 due to
the simple specication of DRt and Ct1 as independently distributed
across time with mean zero.) To understand the portfolio balance
term, b2 St Xt , recall that
Xt 1

N
X
i1

Tit aCt1 :

Therefore, we can write

t
X
t1

Xt z

t
X
t1

Ct1 :

The sum of the portfolio shifts Ct1 represent changes in ``effective''

182

Chapter 7

asset supply, in the sense that shifts out of FX are an increase in the
net supply that the remainder of the public must absorb. (I couch this
in terms of supply to connect with traditional portfolio balance intuition, as outlined in chapter 6.) The total increase in net supply is
the sum of past portfolio shifts out of FX:
Increase in net supply

t
X
t1

Ct1 :

As is standard in portfolio balance models, increases in supply lower

price, whereas decreases in supply raise price. This is why a negative
cumulative Xt in equation (7.5) lowers price: If cumulative Xt is negative, this implies that cumulative Ct1 is also negative, which is an
increase in net supply, requiring an decrease in price to clear the
market. Xt is the variable conveying this information about the decrease in net supply (Ct1 is unobservable). Pt depends on the sum of
the Xi because each additional decrease in supply Ct1 requires a persistent incremental increase in price.
Here is a simple diagram that illustrates the basic economics of the
model. Figure 7.4 presents a one-period example, where the uncer-

Figure 7.4
Portfolio balance effects: one-period example. The market-clearing gap EV P0 is a
function of the risky asset's net supply. In traditional portfolio balance models, variation in gross supply is the driver. In the Evans-Lyons model, gross supply is xed, but
net supply is moving over time due to shifts in demand that are unrelated to EV P0 .
These demand shifts are the exogenous realizations of Cit1 . In contrast to the dissipation
of the portfolio balance effect on price in the one-period example, the price effects do
not dissipate in the Evans-Lyons model because payoff uncertainty is resolved
smoothly over time.

Macro Puzzles

183

tain payoff V is realized at time 1. The market-clearing gap EV P0

the risk premiumwill be a function of the risky asset's net supply. In traditional portfolio balance models (see chapter 6), demand
D is a function of relative returns, and supply S is time-varying. That
is, price P0 is determined according to
~
DEV P0 S;
where the tilde denotes random variation. In these traditional models,
changes in supply S are unrelated to EV; market-clearing price is
moving in a way similar to the way the price of a bondwith a
xed, known payoffhas to adjust to alter its expected return.
The Evans-Lyons model looks different. In their model (gross)
supply is xed. But what I am calling ``net supply'' is moving over
time, due to demand shifts that are unrelated to EV P0 . These
demand shifts are the realizations of Cit1 .12 Conceptually, their model
looks more like
~ S;
DEV P0 ; C
where S denotes xed gross supply and C~ denotes shifts in net supply, that is, shifts in demand unrelated to EV P0 . In this one~ the lower the
period example, the higher the t 0 realization of C,
net supply to be absorbed by the rest of the public, and the higher
the market-clearing price P0 (to achieve stock equilibrium). In a
sense, then, the multi-period model of Evans and Lyons is akin to a
single-period model in which net supply is ``shocked'' multiple times
before trading takes place. At the time of each ``shock,'' price needs
to adjust a bit more.
Let me contrast this pricing relation in equation (7.5) with the
pricing relation in the simultaneous-trade model of chapter 4 (equation 4.20). The key difference, though not obvious from the equations, is that Xt is conveying a different type of information. In the
Evans-Lyons model, Xt conveys information about portfolio balance
effects only: Daily inventory effects are ruled out by the assumption
of no overnight dealer positions and payoff information is ruled out
by the assumption that there is no correlation between the Cit1 's and
the payoff increments DRt . In the simultaneous-trade model, on the
other hand, it is the portfolio balance effects that are ruled out: The
public does not trade in period 2 of that model, so dealers must hold
the inventory imbalance among themselves when V is realized. Because there is no risk sharing with the public, risk premia at the

184

Chapter 7

public level do not arise. This feature of the simultaneous-trade

model is more appropriate for intraday analysis than it would be for
lower frequency analysis. The simultaneous-trade model does, however, admit the two types of order ow information that the EvansLyons model rules out: payoff information and information about
inventory effects. Both operate in the simultaneous-trade model
through interdealer order ow X, which conveys payoff information
in that model because interdealer trades incorporate the private
payoff signals. X conveys information about inventory effects in the
same way that order ow conveys information about portfolio effects
in the Evans-Lyons model: A larger X implies a larger initial purchase by the public, and therefore a larger short position that dealers
as a whole have to bear. It is the size of this position that drives the
inventory effect on price.
Before moving to the Evans-Lyons results, I want to address
another of their model's important features. Recall that one of their
modeling objectives is to clarify the behavior of order ow in gure
1.2. Specically, cumulative order ow is puzzlingly persistent: there
is no obvious evidence of mean reversion in cumulative order ow,
yet, empirically, individual dealer inventories have a short half-life.
How can these two facts be consistent? The Evans-Lyons model
provides an explanation. First, note that dealer inventories in the
Evans-Lyons model are short-lived: no dealer carries an inventory
longer than one day. At the same time, cumulative interdealer order
ow in their model is persistentin fact, it follows a random walk
(i.e., there is no mean reversion whatsoever). Equations (7.2) and
(7.3) hold the key to this random-walk result. Interdealer order ow
each day is proportional to the public order ow that occurs at the
beginning of that day. Because this public order ow is i.i.d. across
dealers and time, cumulative interdealer order ow follows a random walk. In the end, these seemingly incongruous facts are consistent because, ultimately, dealers can only decumulate inventory by
trading with the public, so aggregate decumulation is not reected in
interdealer ow.13
Evans-Lyons Results
The equation Evans and Lyons actually estimate is the following:
Dpt b 1 Dit it  b2 Xt ht ;

7:7

Macro Puzzles

185

where Dpt is the change in the log spot rate (DM/$ or yen/$) from
the end of day t 1 to the end of day t, Dit it  is the change in the
overnight interest differential from day t 1 to day t (  denotes DM
or yen), and Xt is the interdealer order ow from the end of day t 1
to the end of day t (negative denotes net dollar sales).
There are two changes in this equation relative to equation (7.6).
First, the payoff increment DRt in equation (7.6) represents the macro
innovations in the model, or f i; m; z. For estimation, Evans and
Lyons have to take a stand on what to include in the regression for
DRt . They choose to include changes in the nominal interest differential; that is, they dene DRt Dit it  , where it is the nominal
dollar interest rate and it  is the nominal non-dollar interest rate (DM
or yen). As a measure of variation in macro fundamentals, the interest differential is obviously incomplete. The reason Evans and Lyons
do not specify a full-blown macro model is because other macro
variables (e.g., money supply, output, etc.) are not available at the
daily frequency. Accordingly, one should not view their model as
fully accommodating both the macro and micro approaches. At the
same time, if one were to choose a single macro determinant that
needs to be included, interest rates would be it: Innovations in interest differentials are the main engine of exchange rate variation in
macro models (e.g., the sticky-price monetary model).14 Moreover,
using the change in the interest differential rather than the level is
consistent with monetary macro models: in monetary models, shocks
to price are driven by unanticipated changes in the differential.15
The second difference in equation (7.7) relative to (7.6) is the replacement of the change in the level of price DPt with the change in
the log price Dpt . This difference makes their estimation more directly
comparable to previous macro specications, as those specications
use the log change (which is approximately equal to a percentage
change). As an empirical matter, using Dpt is inconsequential: the
two different measures for the change in price produce nearly identical results.
Table 7.1 presents estimates of the Evans-Lyons model (equation
7.7) using daily data for the DM/$ and yen/$ exchange rates. The
coefcient b2 on order ow Xt is correctly signed and signicant,
with t-statistics above ve in both equations. To see that the sign is
correct, recall from the model that net purchases of dollarsa positive Xt should lead to a higher DM price of dollars. The traditional

186

Chapter 7

Table 7.1
Estimates of the Evans-Lyons Model
Dpt b1 Dit it b 2 Xt ht
b1

b2

R2

DM

0.52
(1.5)

2.10
(10.5)

0.64

Yen

2.48
(2.7)

2.90
(6.3)

0.45

T-statistics are shown in parentheses. (In the case of the DM equation, the t-statistics
are corrected for heteroskedasticity; there is no evidence of heteroskedasticity in the
yen equation, and no evidence of serial correlation in either equation.) The dependent
variable Dpt is the change in the log spot exchange rate from 4 p.m. gmt on day t 1 to
4 p.m. gmt on day t (DM/$ or yen/$). The regressor Dit it is the change in the oneday interest differential from day t 1 to day t (  denotes DM or yen, annual basis).
The regressor Xt is interdealer order ow between 4 p.m. gmt on day t 1 and 4 p.m.
gmt on day t (negative for net dollar sales, in thousands of transactions). Estimated
using OLS. The sample spans four months (May 1 to August 31, 1996), which is 89
trading days. (Saturday and Sunday order owof which there is littleis included
in Monday.)

macrofundamentalthe interest differentialis correctly signed, but

is only signicant in the yen equation. (The sign should be positive
because, in the sticky-price monetary model for example, an increase
in the dollar interest rate it induces an immediate dollar appreciationincrease in DM/$per chapter 6.)
The overall t of the model is striking relative to traditional macro
models, with R 2 statistics of 64 percent and 45 percent for the DM
and yen equations, respectively. Moreover, the explanatory power of
these regressions is almost wholly due to order ow Xt : regressing
Dpt on Dit it  alone, plus a constant, produces an R 2 statistic less
than 1 percent in both equations and coefcients on Dit it  that
are insignicant at the 5 percent level.16 That the interest differential
regains signicance once order ow is included, at least in the yen
equation, is consistent with omitted variable bias in the interestrates-only specication.
The size of the order ow coefcient is consistent with estimates
based on single-dealer data reviewed in chapter 5. The coefcient of
2.1 in the DM equation of table 7.1 implies that a day with one
thousand more dollar purchases than sales induces an increase in the
DM price by 2.1 percent.17 Given an average trade size in the sample
of $3.9 million, this implies that

Macro Puzzles

187

$1 billion of net dollar purchases increases the DM price of a dollar

by 0.54 percent.
Equivalently, at a spot rate of 1.5 DM/$, $1 billion of net dollar
purchases increases the DM price of a dollar by 0.8 pfennig. Turning
now to the chapter 5 estimates, those results at the single-dealer level
show that information asymmetry induces the dealer to increase
price by one one-hundredth of a pfennig (0.0001 DM) for every incoming buy order of $10 million. That translates to 1 pfennig per $1
billion, versus the 0.8 pfennig per $1 billion found by Evans and
Lyons. Though linearly extrapolating the single-dealer estimate
(based on individual order sizes around $10 million) to $1 billion of
order ow is certainly not an accurate description of single-dealer
price elasticity, with multiple dealers it may be a good description of
price elasticity marketwide.
Robustness Checks
To check robustness, Evans and Lyons examine several obvious
variations on the model. For example, they include a constant in the
regression, even though the model does not call for one; the constant is insignicant for both currencies and has no substantive effect
on the other coefcients. Second, in the spirit of Uncovered Interest
Parity, they include the level of the interest differential in lieu of its
change; the level of the differential is insignicant in both cases.
Third, they test for simple forms of nonlinearity, such as adding a
squared order ow term, or testing for piece-wise linearity. Though
the squared order ow term is insignicant in both equations and
they nd no evidence of piece-wise linearity in the DM equation,
they do nd some evidence of piece-wise linearity in the yen equation (there is a greater sensitivity of the yen/$ price to order ow in
the downward direction, though estimates for both directions are
positive and signicant). Fourth, they test whether the order ow/
price relation depends on the gross level of activity. They nd that
it does: in the DM equation, the order ow coefcient is lowest on
days when the number of transactions is at a middling level (i.e., the
pattern is U-shaped); in the yen equation, they nd that the order
ow coefcient is lowest on days when the number of transactions
are at a low level (i.e., the coefcient increases with activity level).18
Their model is not rich enough to account for these coefcient varia-

188

Chapter 7

tions structurally. Fifth, Evans and Lyons decompose contemporaneous order ow into expected and unexpected components (by
projecting order ow on past ow). In their model, all order ow Xt
is unexpected, but this need not be the case in the data (in fact, daily
order ow is essentially unforecastable using past ow). They nd,
as one would expect, that order ow's explanatory power comes
from its unexpected component.
Isn't This Just Demand Increases Driving Price Increases?
At rst blush, it might appear that the Evans-Lyons results are right
out of Economics 101: of course when demand goes up, price goes
up. But this misses the most important lesson. A premise of textbook
exchange rate models (chapter 6) is that we don't need order ow to
push prices around. Rather, when public information arrives, rational markets adjust price instantaneously (i.e., excess demand from
new information causes price to adjust without trading needing to
take place). That order ow explains such a large percentage of price
moves underscores the inadequacy of this public information framework. The information the FX market is aggregating is much subtler
than textbook models assume. This we learn from our order ow
regressions.
But What Drives Order Flow?
An important challenge for the microstructure approach is determining what drives order ow, that is, the rst link in the fundamentals/
order ow/price chain (gure 7.1). Here are three promising strategies for shedding light on this question. Strategy one is to disaggregate order ow. For example, interdealer order ow can be
split into large banks versus small banks, or investment banks versus
commercial banks. Data sets on customer order ow can be split
into nonnancial corporations, leveraged nancial institutions (e.g.,
hedge funds), and unleveraged nancial institutions (e.g., mutual
and pension funds). Do all these trade types have the same price
impact? Someone believing that order ow is just undifferentiated
demand would predict that they do. In fact, they do not, as we shall
see in chapter 9. Certain types of orders (e.g., those from nancial
institutions) convey more information, and therefore have more
price impact. People who view order ow as undifferentiated de-

Macro Puzzles

189

mand overlook this level of analysisthey overlook the fact that

order ow is a vehicle for conveying information. Understanding the
information intensity of different trade types brings us closer to this
market's underlying information structure.
Strategy two for determining what drives order ow focuses on
public information intensity. Consider, for example, periods encompassing scheduled macro announcements. Does order ow account
for a smaller share of price variation within these periods? Or is
order ow an important driver of price even at these times, perhaps
helping to reconcile differences in people's mapping from public information to price? Work along these lines, too, will shed light on the
forces driving order ow.19
Strategy three for determining what drives order ow focuses on
discriminating payoff information from discount rate information.
If order ow conveys payoff information, then it should forecast
important macro variables like interest rates, money supplies, and
trade balances. New order ow data sets that cover many years of
FX tradingsuch as the data set I examine in chapter 9provide
enough statistical power to test this. At a broad level, separating
these two types of nonpublic information has implications for how
we dene the concept of ``fundamentals.'' Order ow that reects
information about payoffslike expectations of future interest rates
is in keeping with traditional denitions of exchange rate fundamentals. But order ow that reects changing discount rates may
encompass nontraditional exchange rate determinants (e.g., changing risk tolerance of nancial institutions or changing hedging demands), calling perhaps for a broader notion of fundamentals.
Comments on Causality
Under the Evans-Lyons model's null hypothesis, causality runs
strictly from order ow to price. Accordingly, under the null, their
estimation is not subject to simultaneity bias. (Unlike the classic
supply-demand identication problem, Evans and Lyons are not
simply regressing price on quantity; quantityi.e., volumeand
order ow are fundamentally different concepts.) Within microstructure theory more broadly, this direction of causality is the norm:
it holds in the models outlined in chapter 4 (i.e., the Kyle auction
model, the sequential-trade model, and the simultaneous-trade
model), despite the fact that price and order ow are determined

190

Chapter 7

simultaneously. In these models, price innovations are a function of

order ow innovations, not the other way around. Put differently,
order ow is indeed a cause of price changes, but only a proximate
cause; the underlying cause is nonpublic information (about payoffs
or about discount rates).
Although there is no simultaneity bias under the null hypothesis,
alternative hypotheses do exist under which causality is reversed.
The example that has received the most attention recently is specic
to a particular period and market: October 1998, in the dollar/yen
market (after collapse of the hedge fund Long Term Capital Management). In that case, there is evidence that feedback selling of
dollars by distressed participants fueled the dollar's fall from about
130 yen/$ to about 118 yen/$ in a single day. As the story goes,
speculators attempting to stop their already substantial losses felt
they had to sell into a falling market, thereby making it fall further.
In this case, causality appears to have been two-way. Using data on
actual order ows over that period, I examine this special episode as
a case study in chapter 9.
7.2

The Flow Approach

Chapter 6 surveys the modern approaches to exchange rates. One

``premodern'' approach not addressed in that chapter is the balance
of payments ow approach (which macroeconomists call, simply, the
ow approach; see also note 1, chapter 1). Given the microstructure
approach's emphasis on order ow, is there a relation between the
microstructure and ow approaches? Both focus squarely on the role
of transactions in determining rates. In this sense, the antecedents of
the microstructure approach were in that earlier literature (see, e.g.,
Robinson 1937 and Machlup 1939; Rosenberg 1996 provides recent
perspective). However, despite their apparent similarity, the two
approaches are distinct and differ in important ways.
First let me review the ow approach, which is a broadened version of the goods market approach (introduced in chapter 1). Under
the goods market approach, demand for currencies comes primarily
from purchases and sales of goods. For example, an increase in
exports increases foreign demand for domestic currency to pay for
those exported goods. In this simple form, the implication is rather
intuitive: countries with trade surpluses experience appreciation
(which comes from the currency demand created by the surplus).

Macro Puzzles

191

Under the ow approach, currency demand comes not only from

goods ows, but also from capital ows. (The former is captured in a
country's balance of payments by the current account; the latter is
captured by the capital account.) For example, an increase in the ow
of investment abroad increases demand for foreign currency to pay
for those investments.
This description embeds two of the three main differences between the ow and microstructure approaches. First, under the ow
approachlike its goods market cousinexchange rate determination is a kind of by-product, in the sense that the rate is not determined in its own speculative market. One consequence is that ow
approach models typically do not conform to modern standards of
nancial market efciency. (For example, publicly available information can be used in ow approach models to generate excess
speculative returns.) The microstructure approach is different in that
the exchange rate is indeed determined within its own speculative
market. All stocks of currency are willingly held, and only individuals possessing superior information can earn excess speculative
returns.20
The second main difference embedded in the description of the
ow approach is its exclusive focus on balance of payments accounts. As an empirical matter, this was part of the ow approach's
undoing because these particular ows have little explanatory power
for exchange rates (Meese and Rogoff 1983a). The microstructure
approach focuses instead on the ows the price setters (dealers) can
actually see. These need not correspond to balance of payments
transactions (see section 9.1).
The third main difference between the ow and microstructure
approaches is their treatment of beliefs. To illustrate, consider an
analogy from the material on rational expectations models in chapter
7. A key insight of the rational expectations model is that prices
do not simply clear markets, they also convey information. Early
attempts to analyze equilibrium with differentially informed traders
ignored the effect of prices on beliefs. This parallels an essential difference between the ow and microstructure approaches. Under both approaches, the exchange rate is determined from order ow. Under
the ow approach, however, the order owand subsequent price
communicates no incremental information to individuals regarding others' information. Under the microstructure approach, however, order ow does communicate information to individuals, and

192

Chapter 7

different types of order ow can communicate different amounts of

information. This is the information the market needs to aggregate, and microstructure theory describes how that aggregation is
achieved.21
Consider this point in the context of the Evans-Lyons model of the
previous section. A key feature of their model is that order ow
plays two roles. First, holding beliefs constant, order ow affects
price through the traditional process of market clearing. Second,
order ow also alters beliefs because it conveys information that is
not yet common knowledge. That is,

Price P X; BX; Z1 ; Z2 :
Price P depends both directly and indirectly on order ow X. The
indirect effect is through beliefs BX; Z1 , where Z1 and Z2 denote
other determinants. This is the information role ignored in early
attempts to analyze equilibrium with differentially informed traders.
Since the advent of rational expectations, models that ignore this information effect from order ow are viewed as less compelling.
7.3

The Excess Volatility Puzzle

This section addresses the second of the big three puzzles: the excess
volatility puzzle. By excess we mean that exchange rates are much
more volatile than our best measures of fundamentals. Though other
asset markets share this property (e.g., stock markets; see Shiller
1981), the puzzle in FX markets is in many ways distinctive.22 Consider, for example, the fact that most exchange rates are not allowed
to oat freely; many are managed through intervention by central
banks. This allows one to address the volatility puzzle in ways not
possible in other markets. To understand why, note rst that exchange rates are generally less volatile when managed. Given this,
one can compare regimes with different management intensities to
identify why volatility differs, thereby shedding light on the volatility's causes. This approach is common in the literature (e.g., Flood
and Rose 1995; Killeen, Lyons, and Moore 2000a). The analysis I
present here draws primarily on the empirical ndings of Killeen,
Lyons, and Moore (KLM).
Before reviewing the KLM ndings, let me provide more perspective on the ``cross-regime'' approach to exchange rate volatility.23
Why is it that similar macro environments produce more volatility

Macro Puzzles

193

when exchange rates oat freely? There are two main approaches
to this question, one theoretical and one empirical. The theoretical
approach was pioneered by Dornbusch (1976) in his sticky-price
monetary model (see chapter 6). Dornbusch shows that when goods
prices are sticky but the exchange rate is free to jump, then economic
shocks have a disproportionately large effect on the exchange rate
so-called overshooting. From the perspective of excess volatility, the
sticky-price monetary model generates the kind of ``amplication''
that might explain why oating rates are more volatile than fundamentals. This theoretical explanation is not borne out empirically,
however: the sticky-price model does not t the data.
The second main approach to why oating rates are more volatile
is empirical. A good example is Flood and Rose (1995, 5), who put
the cross-regime logic as follows:
Intuitively, if exchange rate stability arises across regimes without corresponding variation in macroeconomic volatility, then macroeconomic variables will be unable to explain much exchange rate volatility. Thus existing
models, such as monetary models, do not pass our test; indeed, this is also
true of any potential model that depends on standard macroeconomic variables. We are driven to the conclusion that the most critical determinants of
exchange rate volatility are not macroeconomic.

The central idea here starts with the Flood-Rose nding that managing rates does not change the volatility of fundamentals (fundamentals as described by the models of chapter 6). So if the volatility
reduction from managing rates is not coming from changed behavior
of these fundamentals, then it is unlikely these are critical fundamentals. In a sense, then, the Flood-Rose conclusion deepens the puzzle.
KLM take a different tackthey exploit a natural experiment,
using the switch from the European Monetary System (EMS) to
European Monetary Union (EMU), which in terms of regimes is a
switch from a target zone to a rigidly xed rate.24 Starting in January
1999, the euro country currencies have been rigidly xed to one
another. Before January 1999, howeverparticularly before May
1998there was still uncertainty about which countries would participate in EMU. There was also uncertainty about the timing of interest rate harmonization (which had to occur among the countries
adopting the euro).
KLM's analysis of this experiment leads them to the following
punchline: exchange rates are more volatile under exible rates because of order ow. Order ow conveys more information under

194

Chapter 7

exible rates, which increases volatility. Fixed exchange rates prevent order ow from conveying informationas a driver of returns,
it is ``turned off.'' The intuition for why this happens is tied to demand elasticity. Under oating, the elasticity of public demand is
(endogenously) low, due to higher volatility and aversion to the risk
this higher volatility entails. This makes room for the types of portfolio balance effects that arise in the Evans-Lyons model, and allows
order ow to convey information about those effects. Under (perfectly credible) xed rates, the elasticity of public demand is innite:
return volatility shrinks to zero, making the holding of foreign exchange effectively riskless. This eliminates portfolio balance effects
and precludes order ow from conveying this type of information.
Consequently, order ow as a return driver is shut down.25
Figure 7.5 provides an initial, suggestive illustration of the KLM
results. It shows the relationship between the FF/DM exchange rate
and cumulative order ow (interdealer order ow from EBSsee
below). The vertical line is May 4, 1998, the rst trading day after the
announcement of the conversion rates of the euro-participating currencies. The relationship between the two series before May 4 is
clearly positive: the correlation is 0.69. After May 4indeed, even a
bit before May 4there is a sharp unwinding of long DM positions
with no corresponding movement in the exchange rate. The effect of
order ow on the exchange rate appears to have changed from one of
clear impact to one of no impact. (Though total variation in the exchange rate is small, from a trading perspective the variation in the
January to May period is signicant.) The model KLM develop provides a more formal framework for addressing this issue (to which I
now turn).
The Killeen-Lyons-Moore Model
The KLM model is a variation on the Evans-Lyons model of section
7.1. The intraday trading structures of the two models are identical.
(Accordingly, my exposition is fullest where the two models depart.)
The key difference is the interday structure: in the KLM model,
trading days fall into one of two regimes, a exible-rate regime followed by a xed-rate regime. The shift from exible to xed rates is a
random event that arrives with constant probability p at the end of
each trading day (after all trading). Once the regime has shifted to
xed rates it remains there indenitely. This formulation has two

Figure 7.5
The level of the FF/DM exchange rate (solid) and cumulative interdealer order ow (dashed) over EBS in 1998.

196

Chapter 7

Figure 7.6
Two trading regimes. Under the exible-rate regime, payoff increments DRt are distributed Normally, with mean zero and variance SR . The shift from exible to xed
rates is a random event, the arrival of which is shown here at the end of day T. Once
the regime has shifted to xed rates it remains there indenitely. On the rst morning
of the xed-rate regime, the central bank (credibly) commits to pegging the exchange
rate at the previous day's closing price and maintains DRt 0 thereafter.

important advantages. First, the effective horizon over which foreign

exchange is priced in the exible-rate regime remains constant. Second, the parameter p provides a compact means of describing regime
shifts as far or near.
Each day t, foreign exchange earns a payoffs Rt , which is composed of a series of increments:
t
X
DRt :
7:8
Rt
t1

The DRt increments are observed publicly each day before all trading. As before, these increments represent innovations over time in
public macroeconomic information (e.g., changes in interest rates).
Under the exible-rate regime, the DRt increments are i.i.d. Normal
0; sR2 . On the rst morning of the xed-rate regime, the central bank
(credibly) commits to pegging the exchange rate at the previous
day's closing price and maintains DRt 0 thereafter. FX dealers and
customers all have identical negative exponential utility, with coefcient of absolute risk aversion y. The gross return on the risk free
asset is equal to one.
Figures 7.2 and 7.6 describe the intraday and interday timing of
the model, respectively. (Figure 7.2 appears above in my review
of the Evans-Lyons model.) Within each day, there are three rounds
of trading: rst dealers trade with the public, then dealers trade
among themselves (to share the resulting inventory risk), and nally
dealers trade again with the public (to share inventory risk more

Macro Puzzles

197

broadly). Figure 7.6 describes the two trading regimes, using T to

denote the last day of exible-rate trading.
Each day begins with public observation of the payoff increment
DRt . Each dealer quotes a price to his customers, denoted Pit1 , and
then receives a customer-order realization Cit1 that is executed at his
quoted price. (Cit1 < 0 denotes net customer sellingdealer i buying.)
The individual Cit1 's are distributed:
Cit1 @ Normal0; sC2 :
They are uncorrelated across dealers and uncorrelated with the payoff increment DRt . These orders represent exogenous portfolio shifts
of the nondealer public. Their realizations are not publicly observable, and they arrive every day, regardless of regime. It will be useful
for us to dene the aggregate public demand in round 1 as
Ct1

N
X
i1

Cit1 :

In round 2the interdealer roundeach dealer quotes a price to

other dealers. These interdealer quotes are observable and available
to all dealers. Each dealer then trades on other dealers' quotes.
(Orders at a given price are split evenly across dealers quoting that
price.) Let Tit denote the (net) interdealer trade initiated by dealer i in
round 2 (negative for dealer i selling). At the close of round 2, all
dealers observe the interdealer order ow Xt from that period:
Xt

N
X

Tit :

7:9

i1

As in the Evans-Lyons model, in round 3 of each day dealers share

overnight risk with the nondealer public. Unlike round 1, the public's motive for trading in round 3 is nonstochastic. Initially, each
dealer quotes a scalar price Pit3 at which he agrees to buy and sell any
amount. These quotes are observable and available to the public.
Public demand for the risky asset in round 3, denoted Ct3 , is less than
innitely elastic. With the earlier assumptions, this allows one to
write public demand as a linear function of the expected return:
3
Rt1 j Wt3 ;
Ct3 gEDPt1

where

7:10

198

Chapter 7

3
g z Var1 DPt1
Rt1 :

As in section 7.1, the positive coefcient g captures the elasticity of

public demandthe public's aggregate risk-bearing capacity. The
information in Wt3 is that available to the public at the time of trading
in round three of day t (which includes all past Rt and Xt ).
Equilibrium
The focus in the KLM analysis is on the exchange rate's level. The
level at the end of day t can be written as
8
t
t
X
X
>
>
>
DR

b
Xt
under exible rates t U T
b
>
t
1
2
>
< t1
t1
Pt
T
T
t
>
X
X
X
>
>
>
DR

b
X

b
Xt under xed rates t > T
b
>
t
t
1
2
3
:
t1

t1

tT1

7:11
The parameters b2 and b3 are the price impact parametersthey
govern the price impact of order ow. These b's depend inversely on
g (the return sensitivity of public demand); they also depend on the
variances sR2 and sC2 .
The message of equation (7.11) is important. It says that in the
exible regime, there should be a cointegrating relationship between
the level of the exchange rate, cumulative macro fundamentals
DRt , and cumulative interdealer order ow Xt .27 This prediction is
striking: in textbook exchange rate models there is no relationship
between order ow and exchange rates, much less a long-run relationship as implied by cointegration. An implication is that order
ow's impact on exchange rates is permanent (which is also true in
the Evans-Lyons model). This speaks directly to the theme in chapter
1 that microstructure variables can have long-lived effects. (Permanent impact does not imply that order ow cannot also have transitory effects, particularly intraday. In this model, though, inventory
effects across days are ruled out, as explained in section 7.1.)
Differences across Exchange Rate Regimes
We want to understand how the role of order ow differs under different exchange rate regimes. To do so, consider equations (7.9) and

Macro Puzzles

199

(7.10). Specically, the parameter g that represents the elasticity of

public demand is regime dependent. This comes from the regime
dependence of the return variance, VarDPt1 Rt1 j Wt3 , in the denition of g in equation (7.10). KLM show that
gxed > gexible :

7:12

Public demand is therefore more elastic in the xed-rate regime than

under the exible-rate regime. The implication for the price impact
parameter b in equation (7.11) is the following:
b2 > b3 :

7:13

This says that exchange rates react more to order ow under exible
rates than they do under xed rates. In the limit of xed rates that are
perfectly credible (i.e., for which VarDPt1 Rt1 j Wt3 0), we have
b 3 0:

7:14

The exchange rate does not respond to order ow in this case. The
intuition is simple enough: Under perfect credibility, the variance of
exchange rate returns goes to zero because public demand is perfectly elastic, and vice versa. (If the xed regime were less than 100
percent credible, then public demand would not be inniteFX
would still be a risky asset.)
Further Intuition for the Solution
Consider PT1 , the price at the close of the rst day of the xed-rate
regime. Forex is a riskless asset at this point, with return variance
equal to zero. With a return variance equal to zero, the elasticity of
public demand is innite and the price impact parameter b3 in
equation (7.11) equals zero. This yields a price at the close of trading
(round 3) on day T 1 of
PT1 b1

T
X
t1

DRt b2

T
X

Xt :

t1

The summation over the payoff increment DRt does not include an
increment for day T 1 because the central bank maintains DRt at
zero in the xed regime. Though XT1 is not equal to zero, this has
no effect on price because b 3 0, as noted. This logic holds throughout the xed-rate regime.

200

Chapter 7

To understand why order ow has price impact under oating,

consider the price at the close of the nal day of oating, day T.
According to equation (7.11),
PT b1

T
X
t1

DRt b2

T
X

Xt :

t1

At the close of day T, the payoff RT1 is uncertain because DRt1

is uncertain. (Expectations of future variables do not enter this expression due to the simple specication of DRt and Ct1 as independently distributed across time with mean zero.) The second term that
depends on Xt is the portfolio balance term. The logic of this term
under oating is exactly the same as the portfolio balance term in the
Evans-Lyons model of section 7.1. That is, the sum of the portfolio
shifts Ct1 represent changes in ``effective'' asset supply, in the sense
that they are a change in net supply that the remainder of the public
must absorb. The total decrease in net supply is the sum of past
portfolio shifts into foreign exchange:
Decrease in net supply

T
X
t1

Ct1 :

As is standard in portfolio balance models, increases in supply lower

price and decreases in supply raise price. This is why a positive
cumulative Xt in equation (7.11) raises price: if cumulative Xt is positive, this implies that cumulative Ct1 is also positive, which is a decrease in net supply, requiring an increase in price to clear the
market. Xt is the variable that conveys this information about net
supply (Ct1 is unobservable). PT depends on the sum of the Xt because each additional decrease in supply Ct1 requires an incremental
increase in price. These portfolio balance effects disappear as payoff
uncertainty shrinks to zero (as in the xed-rate regime).
The KLM Data Set
The KLM data set includes daily order ow in the German mark
French franc market for one year, 1998. The data are from EBS, the
electronic interdealer broking system described in chapter 5. (At the
time, EBS accounted for nearly half of interdealer trading in the major currencies, which translates into about a third of total trading in
these currencies; the Evans-Lyons data that are the basis of section

Macro Puzzles

201

7.1 reect the other half of interdealer tradingthe direct portion.)

By KLM's estimate, their sample accounts for about 18 percent of
trading in the DM/FF market in 1998. Daily order ow includes all
orders passing through the system over twenty-four hours, starting
at midnight gmt (weekdays only).
The data set is rich enough to allow measurement of order ow Xt
two ways: number of buys minus number of sells (as in Evans and
Lyons 1999) and amount bought minus amount sold (in DM). KLM
nd that the two measures behave quite similarly: the correlation
between the two Xt measures in the exible-rate portion of the sample (the rst four months) is 0.98. They also nd that substituting one
measure for the other in their analysis has no substantive effect on
their ndings.
Let me provide a bit more detail on EBS. As noted in chapter 3,
EBS is an electronic broking system for trading spot foreign exchange
among dealers. It is limit-order driven, screen-based, and ex ante
anonymous (ex-post, counterparties settle directly with one another).
The EBS screen displays the best bid and ask prices, together with
information on the cash amounts available for trading at these prices.
Amounts available at prices other than the best bid and offer are not
displayed. Activity elds on this screen track a dealer's own recent
trades, including price and amount, and also track the recent trades
executed on EBS systemwide.
There are two ways that dealers can trade currency on EBS. Dealers
can either post prices (submit ``limit orders''), which does not insure
execution, or dealers can ``hit'' prices (submit ``market orders''),
which does insure execution. To construct a measure of order ow,
trades are signed according to the direction of the latterthe initiator of the transaction.
When a dealer submits a limit order, he or she is displaying to
other dealers an intention to buy or sell a given cash amount at a
specied price.28 Bid prices (limit order buys) and offer prices (limit
order sells) are submitted with the hope of being executed against
the market order of another dealerthe ``initiator'' of the trade. To
be a bit more precise, not all initiating orders arrive in the form of
market orders. Sometimes, a dealer will submit a limit order buy that
is equal to or higher than the current best offer (or will submit a limit
order sell that is equal to or lower than the current best bid). When
this happens, the incoming limit order is treated as if it were a market order and is executed against the best opposing limit order im-

202

Chapter 7

mediately. In these cases, the incoming limit order is the initiating

side of the trade.
Results
Figure 7.5 illustrates the relationship between cumulative order ow
and the exchange rate. We saw that the effect of order ow on the
exchange rate appears to have changed from one of clear impact to
one of no impact. The results that follow address this more formally,
based on the KLM model's testable implications.
The analysis proceeds in two stages. First, KLM address whether
there is evidence of a cointegrating relationship between interdealer
order ow and price, as the model predicts. This rst stage also
examines the related issues of stationarity and long-run coefcient
sizes. The second stage addresses the degree to which order ow is
exogenous (as their model assumes). This stage includes a test for
reverse Granger causality, that is, statistical causality running from
the exchange rate to order ow.
Stage 1:

Cointegration and Related Issues

Let us begin by repeating equation (7.11) from the model, which

establishes the relationship between the level of the exchange rate Pt ,
a variable summarizing public information SDRt , and accumulated
order ow SXt :
8
t
t
X
X
>
>
>
b
DR

b
Xt
under exible rates t U T
>
t
2
>
< 1 t1
t1
Pt
T
T
t
>
X
X
>
> X
>
DR

b
X

b
Xt under xed rates t > T
b
>
t
t
1
2
3
:
t1

t1

tT1

Like Evans and Lyons (1999), KLM use the interest differential as the
public-information variable (the Paris interbank offer rate minus the
Frankfurt interbank offer rate).
The KLM model predicts that before May 4, 1998, all these variables are nonstationary and are cointegrated. After May 4, the model
predicts that the exchange rate converges to its conversion rate, and
should be stationary. During this latter period (May to December),
therefore, equation (7.11) only makes sense if the price-impact co-

Macro Puzzles

203

efcient, b3 , goes to zero (as the model predicts), or if accumulated

order ow becomes stationary. Otherwise, the regression is unbalanced, with some stationary variables, and some nonstationary
variables.
Stationarity
The rst step is to test whether the relevant variables are nonstationary. KLM nd that, in the rst four months of 1998, they were
(inference based on Dickey-Fuller tests). In the remaining eight months,
the exchange rate was stationary, as expected, but both cumulative
order ow and the interest differential remained nonstationary. These
results are consistent with a price impact parameter b3 in the latter
period of zero. It is important to determine, however, whether
equation (7.11) actually holds for the exible-rate period in January
April, that is, whether the variables are cointegrated as the model
predicts.
Cointegration
KLM use the Johansen procedure to test for cointegration (Johansen
1992). The unrestricted vector autoregression (VAR) is assumed to
consist of the three variablesthe exchange rate, cumulative order
ow, and the interest differentialas well as a constant and a trend.
After testing various possible lag lengths, KLM nd evidence that a
lag length of four is appropriate.
The cointegration tests show that there is indeed one cointegrating
vector. (They reject the null of no cointegrating vectors in favor of the
alternative of at least one cointegrating vector, but they cannot reject
the null of one cointegrating vector in favor of the alternative of at
least two.) This implies that a linear combination of the three variables is stationary, as the KLM model predicts.
KLM go one step further and implement the test for cointegration
without the interest differential. They nd evidence of one cointegrating vector in that case, too, now between the exchange rate and
cumulative order ow. The nding of one cointegrating vector in
both the bivariate and trivariate systems suggests that the interest
differential enters the trivariate cointegrating vector with a coefcient of zero. When KLM estimate the parameters of the cointegrat-

204

Chapter 7

ing vector directly, this is exactly what they nd: they cannot reject
that the interest differential has a coefcient of zero. By contrast, the
coefcient on cumulative order ow is highly signicant and correctly signed. (The size of the coefcient implies that a 1 percent increase in cumulative order ow moves the spot rate by about 5 basis
points.29) These ndings of cointegration and an order ow coefcient that is correctly signed are supportive of their model's emphasis on order ow in the long run. At the same time, the lack of
explanatory power in the interest differential suggests that this specialization of the payoff increment DRt is decient (in keeping with
the negative results of the macro literature more generally).
Exogeneity of Order Flow
An important question facing the microstructure approach is the degree to which causality can be viewed as running strictly from order
ow to the exchange rate, rather than running in both directions. The
KLM framework provides a convenient way to address this question. In particular, if a system of variables is cointegrated, then it
has an error-correction representation (see Engle and Granger 1987;
also Hamilton 1994, 580581). These error-correction representations
provide clues about the direction of causality. Specically, the errorcorrection representation allows one to determine whether the burden of adjustment to long-run equilibrium falls on the exchange rate,
on cumulative order ow, or both. If adjustment falls at least in part
on order ow, then order ow is responding to the rest of the system
(i.e., it is not exogenous in the way specied by the Evans-Lyons and
KLM models).
The KLM ndings suggest that causality is indeed running strictly
from order ow to price, and not the other way around. KLM test
this by estimating the error-correction term in both the exchange rate
and order ow equations (exible-rate period, JanuaryApril). They
nd that the error-correction term is highly signicant in the exchange rate equation, whereas the error-correction term in the order
ow equation is insignicant. This implies that adjustment to longrun equilibrium is occurring via the exchange rate. More intuitively,
when a gap opens in the long-run relationship between cumulative order
ow and the exchange rate, it is the exchange rate that adjusts to reduce the
gap, not cumulative order ow. In the parlance of the literature, the in-

Macro Puzzles

205

signicance of the error-correction term in the order ow equation

means that order ow is weakly exogenous. Further, KLM show that
there is no evidence of Granger causality running from the exchange
rate to order ow (i.e., there is no evidence of feedback trading). This
combination of weak exogeneity and absence of Granger causality
implies that cumulative order ow is strongly exogenous. Finally, the
KLM error-correction estimates suggest that about one-third of departures from long-run equilibrium is dissipated each day.
Concluding Thoughts
To summarize, the KLM analysis addresses the excess volatility
puzzle on two fronts, one theoretical and one empirical. On the theoretical front, they provide a new approachbased on order ow
for why volatility is high when exchange rates oat freely. The
punchline of their approach is that an important source of volatility
is order ow, or, more precisely, the information order ow conveys.
Under oating, the elasticity of public demand is (endogenously)
low, due to higher volatility and aversion to the risk that higher
volatility entails. This makes room for the portfolio balance effects
that arise in the Evans-Lyons model and allows order ow to convey information about those effects. Under (perfectly credible) xed
rates, the elasticity of public demand is innite: return volatility
shrinks to zero, making the holding of foreign exchange effectively
riskless. This eliminates portfolio balance effects and precludes order
ow from conveying this type of information. Thus, under xed
rates, order ow as a return driver is shut down. Under managed
rates (i.e., intermediate regimes), order ow as a return driver is
muted.
A nice feature of the KLM approach to excess volatility, relative to
other approaches, is that its implications can be brought to the data.
There are many theoretical papers on excess exchange rate volatility
(see, e.g., Hau 1998 and Jeanne and Rose 1999, and references to
earlier work contained therein). But in general, little of the existing
theoretical work is easily implemented empirically. The order ow
focus of the KLM approach makes it readily implementable. That
said, the specic results that KLM offer are only suggestively supportive of their particular story. Much more empirical analysis along
these lines remains to be done.

206

Chapter 7

Two of the KLM empirical ndings are especially relevant for

interpreting work on order ow more generally. First, they nd that
Granger causality runs from order ow to the exchange rate, but not
vice versa. True, Granger causality is not the same as economic causality; nevertheless, the result does help assuage concern. Second,
they nd that gaps in the relationship between cumulative order
ow and the level of the exchange rate engender an exchange rate
response but not an order ow response. This result, too, helps assuage concern about the direction of causality between these two
variables.
One might be tempted to conclude that four months of data is too
little to produce reliable analysis of cointegration; however, an important aspect of the KLM results should assuage this concern. Recall
that KLM nd rapid adjustment back to the cointegrating relationship (their error-correction estimates suggest that about one-third of
departures from long-run equilibrium is dissipated each day). The
half-life of these departures is therefore only about two days. Four
months of data is enough to cover about forty-ve of these half lives,
quite a lot in the context of estimating cointegrating relationships.
For comparison, estimates of adjustment back to the cointegrating
relationship of Purchasing Power Parity generate half-lives around
ve years. One would need over two hundred years of data to estimate PPP error correction with as many half-lives in the sample.
Finally, note that the KLM model provides a different perspective
on exchange rate credibility. In their model, a credible xed rate is
one in which the private sector, not the central bank, willingly absorbs innovations in order ow.30 The textbook treatment of xedrate regimes, in contrast, is centered on the willingness of the central
bank to buy and sell domestic currency at a predetermined price;
that is, the central bank absorbs the order ow. If the central bank
needs to intervene, the xed exchange rate regime is already in
difculty because the private sector's demand for order ow is no
longer perfectly elastic. It may be useful to revisit currency crises and
our analysis of them with this possibility in mind.
7.4

The Forward Bias Puzzle

The systematic bias in currency forward rates has attracted more

attention than any other puzzle in international nance (surveys in-

Macro Puzzles

207

clude Hodrick 1987; Froot and Thaler 1990; Lewis 1995; Engel 1996).
The bias is such that when Ft; 1 > Pt today's one-period forward
rate is ``predicting'' that the spot rate will risethe spot rate does not
tend to rise as much as predicted (i.e., Pt1 typically ends up below
Ft; 1 ).31 In fact, among the major oating currencies against the dollar,
when the forward rate predicts the spot rate will rise, it usually falls!
The forward rate systematically gets the direction wrong.
In this section, I consider the puzzle from three angles: the statistician's perspective, the economist's perspective, and the practitioner's perspective. My explanation for the bias draws heavily from
the practitioner's perspective.
The Statistician's Perspective
Viewed as a regression, the test for bias in forward rates is based on
the following equation:
pt1 pt a b ft; 1 pt ut1

7:15

where pt1 denotes the spot rate realized at time t 1 ($/other),

ft; 1 denotes the time-t forward rate for transactions settled at t 1
($/other), and ut1 is a purely random error term. (Lower case denotes natural logarithm, so the difference in natural logs pt1 pt is
roughly equal to dollar depreciation in percent. Similarly, the gap
between the forward and spot rates ft; 1 pt is roughly equal to a
percent difference, and is typically referred to as the forward discount on the dollar.)32 Under the null hypothesis of an unbiased
forward rate, estimates of the coefcient b should be one: when the
forward rate predicts the spot rate will rise by Y percent, then the
spot rate does rise, on average, by Y percent, for any given Y.
Actual estimates of the coefcient b using monthly data are signicantly different from one. In fact, they are typically negativeas
suggested aboveaveraging about 0.9 (the average from some 75
studies surveyed by Froot and Thaler 1990). Consider the graphic
depiction in gure 7.7. The two-standard-error band G2s around
the null hypothesis, denoted H0 , is not even close to including the
typical estimate of 0.9. Indeed, most studies even reject the hypothesis that b 0. (The size of the band shown should not be taken
too literally; it is drawn to be consistent with the fact that estimates
often do not include zero within the band.)

208

Chapter 7

Figure 7.7
Statistician's perspective on forward bias. H0 denotes the null hypothesis.

The Economist's Perspective

For economists, the statistical rejection of unbiasedness is all the
more uncomfortable because the forward rate systematically predicts
the wrong direction. This discomfort is partly due to the strikingly
counterintuitive nature of the result. At a more substantive level,
though, the result is economically uncomfortable because getting the
direction wrong makes explaining the bias difcult.
To understand why, consider the fact that for any ``anomaly''
documented in nancial markets, economics offers two standard explanations. One is inefciencymarkets are making systematic errors
in neglecting prot opportunities. The second explanation is that
the bias represents a risk premiumexcess return that compensates
for risk borne in taking advantage of the bias. If the bias simply
compensates for risk, it is not a violation of market efciency.33
The trickfor those partial to this risk-premium explanationis to
show empirically that the bias is consistent with models of the degree to which bearing risk should be compensated. So large a bias
makes this difcult.
Researchers' empirical attempts to link the bias to risk models
have not been successful, despite hundreds of papers that, collectively, put all the standard models to the test. In his survey, Hodrick
(1987) concludes that ``we do not yet have a model of expected
returns [i.e., a model of compensation for risk] that ts the data.'' In
their survey, Froot and Thaler (1990) conclude that among the main
approaches to evaluating the risk-premium explanation, ``none offer
the hypothesis much support.'' In a similar vein, Engel's (1996) survey concludes that ``models of the risk premium have been unsuccessful at explaining the magnitude of this failure of unbiasedness.''
From the economist's perspective, the bias remains a puzzle.

Macro Puzzles

209

It is noteworthy that standard economic models share the property

that risk premiums arise due to covariance between returns and
some other variable. What differentiates the models is what they
dene that other variable to be. There are three basic approaches:
real economy models, asset market models, and portfolio balance
models. In real economy models (e.g., Obstfeld and Rogoff 1996), the
relevant covariance is with growth in aggregate consumption. In
asset market modelslike the capital asset pricing modelthe key
covariance is with returns on a market portfolio. In portfolio balance
models (e.g., Branson and Henderson 1985), the key covariance is
typically that across exchange rates (though covariance with other
asset returns can be included). In the end, Engel's statement that
``models of the risk premium have been unsuccessful'' applies to all
three of these covariance approaches.
Despite little evidence for the risk-premium hypothesis, the alternative hypothesis of inefciency is uncomfortable. Part of this discomfort is philosophical: to most nancial economists, inefciency
isby its natureuncomfortable. Further discomfort arises in this
instance, however, because the forward bias was rst documented
some twenty years ago, at which time market participants might
have been unaware of the bias. Twenty years and hundreds of
studies later, it is no longer reasonable to consider the bias as anything but public information. Yet the bias continues. Why?
The Practitioner's Perspective: Limits to Speculation
This book addresses the bias puzzle by adopting a practitioner's
perspective, which leads naturally to an explanation of bias based on
what I call limits to speculation. This line of reasoning falls within
the broader microstructure approach in two ways. First, the explanation relies on the microstructure approach's central linkthe link
between price adjustment and order ow. Second, the explanation
relies on institutional realities assumed irrelevant within traditional
approaches.
Before presenting the explanation, let me rst introduce a question
that a statistician or economist would not naturally ask, but practitioners would: How large would the forward bias have to be before currency trading strategies yield a Sharpe ratio equal to, say, a buy-and-hold
equity strategy?34 Financial institutions commonly use Sharpe ratios

210

Chapter 7

to measure the performance of their proprietary trading. The ratio is

dened as:
Sharpe ratio

ERs Rrf
;
ss

7:16

where ERs is the expected return on the strategy, Rrf is the risk-free
interest rate, and ss is the standard deviation of the returns to the
strategy.35 Over the last fty years, the Sharpe ratio for a buy-andhold strategy in U.S. equities has been about 0.4 on an annual basis
(the excess return in the numerator is about 7 percent and annualized
return standard deviation in the denominator is about 17 percent).
Under the null hypothesis of no bias in the forward rate (b 1 in
equation 1), the Sharpe ratio of currency strategies is zerothere is
no expected return differential (i.e., ft; 1 does not tend to over-predict
subsequent spot rate changes). As b departs from 1, however, the
numerator becomes positive. The larger the bias, the larger the
numerator. The denominator, on the other hand, is not a function of
the bias. Rather, it is determined by exchange rate variances and
covariances. Another determinant of the denominator is the number
of exchange rates included in the currency trading strategya
greater number of currencies provides more diversication, other
things equal.
Strikingly, it is only when b equals about 1 or 3 that the Sharpe ratio
for currency strategies is about the same as that of equities, or 0.4. (This
assumes a currency strategy that diversies across the six largest
markets; I provide further details below.) If b falls anywhere in the
interval 1; 3, then a currency strategy designed to exploit bias has
a lower return per unit risk than a simple equity investment.
From the practitioner's perspective, the two-standard-error band
that the statistician draws around the null of b 1 misses the point.
If instead, I draw a band around the null of b 1 that corresponds to
speculative signicanceas opposed to statistical signicancethen
the reality looks more like that shown in gure 7.8. If a Sharpe ratio
of 0.4 is the practitioner's threshold for determining tradable opportunities, then the interval from b 1 to b 3 would not be attractive. These values for b dene an inaction rangea range within
which the forward rate bias does not attract speculative capital.
Viewed this way, the puzzle of forward bias is not a glaring prot
opportunity.

Macro Puzzles

211

Figure 7.8
Practitioner's perspective on forward bias. H0 denotes the null hypothesis.

An Explanation in Four Parts

My explanation for the persistence of the bias has four parts, which I
sketch at the outset so the structure is clear. Then I address each of
the parts in more detail, providing supporting evidence and theory.
Finally, I offer thoughts on why the bias is in the direction of b < 1,
an issue distinct from the question of why the bias persists.
Part 1: Anomaly Persistence: If speculative capital is not allocated
to exploiting the forward bias, then the forward bias will persist.
Part 2: Institutional Allocation of Speculative Capital: Institutions
with a comparative advantage in exploiting forward bias allocate
their speculative capital based, in large part, on Sharpe ratios.
Part 3: Sharpe Ratio of Currency Strategies: Sharpe ratios of reliable (i.e., non-data-snooped) currency strategies that exploit forward
bias are roughly 0.4 on an annual basis, similar to that of a simple
strategy of buying and holding an equity index.
Part 4: Underallocation of Speculative Capital: Because a Sharpe
ratio of 0.4 is well below most institutions' minimum threshold for
inducing capital allocation, the anomaly persists.
Part 1:

Anomaly Persistence

Part one of the explanation may seem intuitively obvious: ``If speculative capital is not allocated to exploiting the forward bias, then the
forward bias will persist'' But it deserves a closer look. A skeptic
would rightly assert that order ow is not necessary to move price.

212

Chapter 7

By this logic, if the current spot rate or the current forward rate (or
both) are at an incorrect level, and this is common knowledge within
the market, then these rates should adjust immediately to the correct
level, without any role for order ow. This logic would seem to hold
particularly well in the foreign exchange market. After all, goes the
traditional view, all fundamental information in this market is public
information. This is not a market that needs order ow to move price
(as compared to, say, the market for a particular stock, where order
ow can communicate inside information about earnings).
There is a big leap of faith, however, between the true statement
that ``order ow is not necessary to move price'' and the belief that,
in fact, ``order ow plays little role in moving price.''36 Consider two
points that clarify why this leap of faith is so bigone theoretical
and the other empirical.
Theory claries the conditions under which order ow plays no
role, conditions that do not hold in the FX market (see also chapter
2). Specically, unless both of the following are true, order ow will
play a role in moving price: (1) all information relevant for exchange
rates is publicly known, and (2) the mapping from that information
to prices is also publicly known. The second of these conditions is
patently violated. Even the rst is unlikely to hold (e.g., it does not
hold in the Evans-Lyons model, which includes a rather natural information structure).
Empirical evidence also runs counter to the view that ``order ow
plays little role in moving price.'' Section 7.1 shows that the effects of
order ow on price are quite strong. Moreover, the R 2 statistics
reported in that section indicate that order ow accounts for the
lion's share of exchange rate variation.
Part 2:

Institutional Allocation of Speculative Capital

Part 2 of the explanation is the most contentious: Institutions with a

comparative advantage in exploiting forward bias allocate their
speculative capital based, in large part, on Sharpe ratios. Let me
stress at the outset that this assertion is based on my discussions
with proprietary traders at banks and hedge funds. The message
from those discussions is clear: proprietary traders think in terms of
Sharpe ratios, and as a result, they consider the forward bias trade
relatively unattractive.38 Irrespective of the theoretical questions this
focus on Sharpe ratios raises, its empirical predominance is undeni-

Macro Puzzles

213

able (so its impact on prices and returns cannot be ruled out on simple theoretical grounds).
This part of the explanation is the most contentious because, absent other frictions, capital should not be allocated based on ``variance'' risk. Rather, capital should be allocated based on covariance
with some larger objective (e.g., covariance with the market return,
as in the Capital Asset Pricing Model). Empirically, the covariance
between returns on, say, the MSCI World Equity Index and returns
on bias-exploiting currency strategies is small. This suggests that the
risk of currency strategies is largely diversiable. Thus, though their
return-per-unit total risk may be low, their return-per-unit systematic
risk is hightoo high to be consistent with standard covariance
models (i.e., too high for people who are concerned with covariance
not to take advantage). Hence, from the covariance perspective, the
puzzle endures.
For completeness, let me address why institutions might choose
to behave this way, because it appears as though trading opportunities are not fully exploited. (I reiterate, however, that institutions
do behave this way. It is this empirical fact that is most essential to
my argument, not the theoretical rationale for why this behavior
arises.)39 The model I have in mind is in the spirit of Shleifer
and Vishny's (1997) model, in which agency frictions in professional
money management lead to less aggressive trading, or what they call
limits to arbitrage. (I intentionally use the term ``limits to speculation'' in this section to establish a conceptual link to that earlier
work.) Specically, consider an agency-friction modelwhich I
only sketch herewhere the optimal contract between a rm and a
proprietary trader is one that holds the trader responsible for ownvariance, not covariance. Traders do not want their compensation
tied to securities they do not trade; doing so with a covariance-based
contract reduces prot incentives and opens the door to disruptive
levels of idiosyncratic risk.40 Given this ``constrained-optimal'' contract, limited speculative capital is allocated to trading opportunities
with the highest Sharpe ratio.
One way around these agency frictions is to imagine a market with
a very large number of individual speculators, each taking a small
position against the forward bias. In this case, agency friction does
not arise, and speculators' collective actions eliminate the bias in
forward rates. The trouble with this approach is that, in reality, individual speculators do not have specialized knowledge, low trans-

214

Chapter 7

action costs, low monitoring costs, and diversied portfolios (see,

e.g., Goodhart and Taylor 1992, who argue that minimum positions
in currency futures are too large for most individuals to maintain a
diversied portfolio). Individuals are at a comparative disadvantage
with respect to these factors, which more than offsets the benet of
the eliminated agency costs. When considering this tiny-speculator
model and why it is not applicable, a reader might ask why it is that
he/she is not trading the forward rate bias. The answer will surely
include several factors noted here.
Part 3:

Sharpe Ratio of Currency Strategies

Part 3 of the explanation contends that Sharpe ratios of strategies for

exploiting forward bias are roughly 0.4 on an annual basis, similar to
that of a simple strategy of buying and holding an equity index.37
This part is straightforward. Table 7.2 presents Sharpe ratios (annual
basis) for various simple strategies applied to the six most liquid
currencies over the period from January 1980 to December 1998
Table 7.2
Sharpe Ratios (Annual Basis) from Pure Currency Strategy
Strategy 1:
Equal Weighted

Strategy 2:
> Median Discount

Strategy 3:
< Median Discount

Sharpe Ratio:
No Costs

0.48

0.46

0.49

Sharpe Ratio:
With Costs

0.37

0.39

0.41

Strategies for proting from forward bias entail selling foreign currency forward when
ft; 1 > pt and buying foreign currency forward when ft; 1 < pt . The three strategies shown
are implemented using the six largest currency markets: $/DM, $/yen, $/, $/Swiss,
$/FF, and $/C$. The ``Equal Weighted'' strategy has an equal position weight each
month in each of the six forward markets. The ``>Median Discount'' strategy only takes
a position in a forward market in a given month if that month's forward discount is
greater than the median forward discount for that currency over the sample (weights
are equal across forward positions taken). The ``<Median Discount'' strategy only takes
a position in a forward market in a given month if the month's forward discount is less
than the median forward discount for that currency over the sample (weights are equal
across forward positions taken). The Sharpe ratio estimate with costs assumes a cost of
ten basis points per transaction (includes price impact of trade). The sample is monthly
data, from January 1980 to December 1998 (1980 is about the time when forward bias
was rst documented in the literature).
Source: Datastream.

Macro Puzzles

215

(monthly data). The second column presents Sharpe ratios before

netting transaction costs. The third column presents Sharpe ratios
after netting transaction costs. The transaction cost adjustment is
based on the strategy's turnover and is set at ten basis points per
transaction. (This is conservative: if this were a $1 billion currency
fund, then the price impact of exchanging one-sixth of the portfolio
could double or triple this cost; see section 7.1.) But do not be distracted by transaction costs; even without transaction costs, the
Sharpe ratios do not rise above 0.5. Thus, part 3 of the explanation
does not depend on transaction costs.
Naturally, by selecting more complicated strategies than those
presented in table 7.2, we can raise in-sample Sharpe ratios. That is
why my wording of part 3 includes the word ``reliable,'' by which I
mean not generated from data snooping. The equal-weighted strategy exploits most of the gains from currency diversication, without
introducing unwarranted degrees of freedom. I consider it a sound,
conservative benchmark.
Part 4:
Range

Underallocation of Speculative CapitalThe Inaction

Part 4 of the explanation closes the loop with part 1. Because a

Sharpe ratio of 0.4 is well below most institutions' minimum threshold for inducing capital allocation, the anomaly persists. As an
empirical matter, most large nancial institutions do not devote
their proprietary capital to currency strategies that exploit forward bias (what I call ``pure currency strategies''). Those that do,
typically devote only a small share of their capital to this type of
speculation.
But how far below the minimum Sharpe ratio threshold is ``well
below''? I have interviewed several proprietary traders and proprietary desk managers regarding the threshold Sharpe ratios they use
for determining speculative capital allocation, and the most common
response is a Sharpe ratio of 1 (annual basis). Healthy skepticism
is warranted here, however: suggesting to others that one's own
threshold is high is an obvious means of enhancing industry reputation. From my interviews, though, I feel safe in asserting that there is
limited interest at these major institutions in allocating capital to
strategies with Sharpe ratios below 0.5.

216

Chapter 7

Empirical Support
Several empirical ndings are consistent with my inaction band explanation based on Sharpe ratios.41 One testable implication is the
following. If my explanation is true, then the coefcient b should be
closer to one in periods when forward discounts ft; 1 pt are further from zero, other things equal (see equation 7.15). To understand
why, note that the numerator of the Sharpe ratio is a function of two
variables: the size of the forward discount and the value of the coefcient b. Holding xed both the value of b and the denominator of
the Sharpe ratio (equation 7.16), a forward discount further from
zero implies a larger Sharpe ratio. A larger Sharpe ratio, in turn,
attracts more speculative capital, which induces adjustment in prices
toward consistency with unbiasedness (i.e., toward the level consistent with no inaction range). The data bears out this prediction.
Huisman et al. (1998), for example, nd that unbiasedness holds
much more tightly in periods when the forward discounts are further
from zero. Using time-averaged data, Flood and Taylor (1996) nd
the same result.
Consider a second implication of my explanation: currencies with
lower variances should have a coefcient b closer to one, other things
equal. This implication works from the denominator of the Sharpe
ratio: for given level of b and a given forward discount size, a lower
variance implies a larger Sharpe ratio. As above, a larger Sharpe
ratio attracts more speculative capital, which induces adjustment in
prices toward consistency with unbiasedness.42 The data bears out
this prediction, too. Flood and Rose (1996), for example, nd that
estimates of b within the lower volatility European Monetary System
are about 0.6.
Why Isn't the Coefcient b in the Middle of the Inaction Range?
The four-part explanation above claries why the bias can persist,
without violating speculative efciency. It does not explain why the
coefcient b (at 0.9) lies near one edge of the inaction range 1; 3.
For this, one needs a meta-modela model that determines the exchange rate when pure currency speculation does not occur. The
meta-model must confront the fact that inaction in terms of pure
currency speculation does not imply an absence of FX order ow.

Macro Puzzles

217

Sources of order ow that remain include customers in the FX market that do not engage in pure currency speculation, which includes
the great majority of mutual funds, pension funds, and nonnancial
corporations.
To begin our formulation of a meta-model, let us divide the customer portion of the order ow pie (see section 3.1) into three slices.
The three slices are
1. Leveraged investors
2. Unleveraged investors
3. Nonnancial corporations
The rst slice is what practitioners call leveraged investors. These are
the proprietary bank traders and hedge funds that are the focus of
the previous paragraphs. As noted, this is the slice that has a comparative advantage in implementing pure currency strategies to exploit forward bias. Despite their comparative advantage, however,
on average they devote little of their capital to this type of speculation. The second slice is what practitioners call unleveraged investors. These are institutions like mutual funds, pension funds,
and insurance companies. The last slice is nonnancial corporations.
With this breakdown, we can write aggregate customer order ow as
Ct CtL CtU CtN ;

7:17

where the superscripts L, U, and N denote leveraged investors,

unleveraged investors, and nonnancial corporations, respectively.
Turning now to exchange rate determination, previous chapters
provide a basis for writing
pt1 pt f Ct ; Zt ;

7:18

where Zt denotes other determinants. Let us specialize this by

writing:43
pt1 pt gCtL ; CtU ; CtN ; Zt :

7:19

This provides a vehicle for explaining exchange rate behavior when

the pure currency strategy no longer drives the aggregate order ow
Ct (through the CtL channel).44 This is a departure from standard
macro models because those models' assumption of Uncovered Interest Parity (UIP) is tantamount to assuming that aggregate order
ow is driven primarily by pure currency strategies.

218

Chapter 7

The meta-model relies instead on the portfolio shifts of the latter

two slicesthe unleveraged investors and nonnancial corporations.
These two institution types have a comparative disadvantage in
implementing pure currency strategies. They are better characterized
by limited participation in currency markets, as they do not monitor
and trade in currency markets continuously. Their trades are incidental to, or bundled with, activities directed primarily at other
objectives (e.g., a nonnancial corporation that times its repatriation
of overseas earnings for tax reasons, or an international mutual fund
that shifts its country allocations based on a changing view of localcurrency equity returns). Suppose that each period a subset of these
institutionsa different subset each periodconsiders trading in the
FX market. (This description of periodic adjustment draws from a
large literature on ``limited participation,'' which started with Grossman and Weiss 1983. They assume that agents adjust their cash
balances periodically, with a different subset of agents returning to
the bank to replenish each period. Periodic adjustment is typically
interpreted as shorthand for unmodeled adjustment costs, e.g., monitoring costs.45) Limited participation in FX markets by unleveraged
investors and nonnancial corporations implies that their portfolio
shifts across currencies are gradual.
Consider in this setting the effect of an increase in the foreign
interest rate.46 Given limited participation, portfolios adjust toward
the higher expected returns only gradually. This contrasts sharply
with standard asset market models, such as the sticky-price monetary model, in which a (real) foreign rate increase causes instantaneous appreciation of the foreign currency (as we saw in chapter 6).
In the limited participation model, the instantaneous effect is spread
over time, producing a period through which the foreign interest rate
is higher, and the foreign currency continues to appreciate due to the
order ow from gradual portfolio adjustment toward the foreign
currency. This yields a coefcient b in equation (7.15) at the negative
end of the inaction range (which is what we set out to explain).47
This model qualies as a meta-model in the sense that the order
ow driving the exchange rateCtU and CtN is not from pure currency strategies, as is implicit in UIP. It comes instead from other
motivations (per the examples above). If the width of the inaction
band from pure currency strategies were zero, order ow from
gradual portfolio shifts would not affect exchange rates because it
would be swamped by order ow from pure strategies in CtL . In the

Macro Puzzles

219

absence of the ow from pure currency strategies, the ow from

gradual adjustment becomes an important determinant.
This raises a fascinating question: How much order ow is required to ``enforce'' forward rate unbiasedness? Section 7.1 suggests
the amount is quite large. Specically, the estimated impact of interdealer order ow on price is 50 basis points per $1 billion; this
translates into an impact of customer order ow of 100 basis points
per $1 billion (assuming customer order ow is multiplied by two in
interdealer tradingnot unreasonable given the hot potato amplication of volume examined in chapter 5; see also chapter 9). An exchange rate that is 20 percent away from a level consistent with
unbiased forward rates would require $20 billion of order ow to
correct.48 That may not seem like a lot, but remember that to maintain the exchange rate at the new level, the $20 billion order ow
cannot be reversed. An indenite commitment of $20 billion of capital is large relative to what nancial institutionseven collectively
are willing to commit to this particular trade. In a sense, then, this
market's immense depth may actually impede correction of misalignments by making Milton Friedman's celebrated ``stabilizing
speculators'' an institutional impracticality.
New Notion of the ``Risk Premium''
I close this section with some comments that bring us back to the
beginning of the forward bias discussion and the two traditional
explanations of inefciency versus risk premium. As noted above,
the explanation provided here for the bias' persistence does not violate market efciency in the speculative sense of that term: there are
no supernormal prot opportunities going unexploited (given that
the Sharpe ratio criterion is the solution to a constrained optimization problem). Does this mean that the explanation falls in the risk
premium category? In a sense, yes; risk aversion is preventing market participants from exploiting the return differentials implied by
the bias. That said, participants are not responding to risk as they
would in a frictionless environment. That is why traditional risk
premium modelsbased on covarianceare unable to account for
return differentials empirically.
If my Sharpe ratio driven, inaction-band explanation is correct,
it suggests that there is far more ``room'' in nancial market equilibrium than frictionless models predict. Other empirical evidence

220

Chapter 7

supports this view. An oft-cited example is the pricing of shares

in ``Siamese twin'' companies, such as Royal Dutch and Shell (see,
e.g., Froot and Dabora 1999). Royal Dutch and Shell are a controlled
experiment in relative valuation. The companies are linked to one
another by corporate charter, which mandates that cash ows to the
equity holders of each company should be distributed in a strict 60/
40 ratio. One would expect that equivalent share classes should trade
at prices in xed 60/40 proportions, but this is not the case: deviations from the 60/40 ratio have ranged from 30 percent to 15
percent over the last twenty years (the 60/40 dividend split is applicable throughout). This is not a case of establishing a relative valuation between, say, IBM and Intel. The relative value of Royal Dutch
and Shell should be determined much more precisely, yet large disparities persist because, in the end, the trade is not appealing enough
to attract sufcient order ow.49 If there is this much room in pricing
in so controlled a setting, what does this say about asset pricing
more generally?

Microstructure and
Central Bank Intervention

Exchange rates directly affect international competitiveness and the

performance of national economies. Indeed, they are arguably the
world economy's most important prices. Add to this importance
the fact that exchange rate volatility appears excessive, and it is not
surprising that government intervention is more common than in
other capital markets.
This chapter applies microstructure tools to central bank intervention. Within exchange rate economics, intervention is a classic topic.
It is also a natural topic for micro analysis. I begin by describing
traditional methods of intervention analysis, which use the tools of
macroeconomics. This provides perspective on how and where a
micro approach can add value. Section 8.2 reviews theoretical work
on intervention that lies within the microstructure approach. Section
8.3 presents recent empirical work whose goal is to measure the price
impact of central bank trades. The focus in that section is recent work
by Evans and Lyons (2000), which builds on the model presented in
chapter 7 (making it especially accessible with that earlier material in
hand).1
Before I begin, I want to address an issue that runs throughout this
chapter: the issue of whether different-currency assets are imperfect
substitutes. Imperfect substitutability is important because it governs
whether portfolio balance effects are driven by order ow (irrespective of whether those orders are from the central bank or the private
sector). Iffor a given expected returninvestors are completely
indifferent to holding dollar- or yen-denominated assets (perfect
substitutability), then portfolio shifts like those in the Evans-Lyons
model (Cit1 ) will have no effect. If investors care about currency denomination, then risk premiums and portfolio balance effects should
arise (see also the portfolio balance model described in chapter 6).

222

Chapter 8

As an empirical matter, we have not yet reached consensus on the

relevance of FX portfolio balance effects.2 Views about their presence
have shifted over the last twenty years from negative to moderately
positive. The earlier negative view was based on early empirical
work that nds no evidence of portfolio balance effects (e.g., Dooley
and Isard 1982; Frankel 1982; Frankel and Engel 1984; Lewis 1988;
for an overview see Dominguez and Frankel 1993b, 105). These
studies examine FX markets at a broad level and address whether
different-currency returns are driven by changing asset supplies. In
general, these studies suffer from lack of statistical power because
changing asset supplies are notoriously difcult to measure. Studies
that focus narrowly on the effects of central bank interventiona
kind of ``event study'' on changing asset suppliesare more successful in nding effects from portfolio balance (e.g., Loopesko 1984;
Dominguez 1990; Dominguez and Frankel 1993a). But even with this
narrow focus on intervention events, results are not exclusively positive (e.g., Rogoff 1984).
The Evans and Lyons (2000) analysis of portfolio balance effects
reviewed in section threeis a new approach, one that measures
portfolio balance effects directly from the order ow of private market participants (i.e., not central banks). Though this new approach is
not without drawbacks, it does avoid several of the drawbacks of the
earlier literature. For example, the Evans-Lyons approach is arguably
more powerful (statistically) than the approach of the early broadlevel studies because it does not rely on measuring changing assets
supplies, making it less vulnerable to measurement error. It may also
be more powerful than the event-study intervention approach because it does not rely on the relatively small number of intervention
events available for that type of analysis. (Moreover, the average size
of interventions is small: For the United States, average size reported
by Dominguez and Frankel 1993b is only $200 million; U.S. intervention since then has been larger, typically in the $300 million to
$1.5 billion range, though less frequentsee Edison 1998).
Let me try to make this point about statistical power more vivid.
Envision the FX market as a ``choke point'' where portfolios are
actually being balanced (which includes the portfolios of central
banks). Measures of order ow provide precise measures of this
rebalancing and the price effects that arise as a result. As a choke
point, the FX market is the venue where market participants (effectively) say to one another, ``Here, hold this,'' where ``this'' might be

Microstructure and Central Bank Intervention

223

Figure 8.1
Fed balance sheet: Unsterilized purchase of $100 million with yen. FXR is foreign exchange reserves, DGB is domestic government bonds, and MS is money supplythat
is, money in circulation (currency plus monetary base).

10 billion euros. If we are to detect portfolio effects anywhere, this

may be the right place to look.
8.1

Macro Motivation for a Micro Approach

Exchange rate regimesfor example, oating, managed oating,

and xedare dened by the degree and type of central bank intervention used. There are two basic types:3
1. Unsterilized intervention
2. Sterilized intervention
As we shall see, unsterilized intervention affects the domestic money
supply and, as a result, the interest rate. Sterilized intervention, on
the other hand, has no effect on the money supply and therefore no
effect on the (short-term) nominal interest rate.
Consider the following example, which claries how sterilization
neutralizes the effect of an intervention on the money supply. Figure
8.1 presents a stylized version of a central bank balance sheet. It has
three basic categories, two on the asset side and one on the liability
side. Central bank assets include foreign exchange reserves (FXR)
and domestic government bonds (DGB). The main central bank liability is money in circulationthe money supply (MS).4
First, consider the effect of an unsterilized intervention. The unsterilized intervention in gure 8.1 is a purchase of $100 million by
the U.S. Federal Reserve Bank (Fed) using Japanese yen. The Fed
would do this if it wanted to support the value of the dollar: it adds
to the market demand for dollars.
To carry out intervention, the Fed could contact a dealing bank,
say Citibank, and buy the $100 million at Citibank's offer quote.5

224

Chapter 8

Settling this trade involves a transfer of $100 million worth of yen

to Citibank. This reduces Fed foreign exchange reserves, FXR, and
money in circulation, MS, by the same amount.6 The reduction in the
money supply is Citibank's payment of $100 million to the Fed. (In
reality, this payment by Citibank takes the form of reducing by $100
million Citibank's account at the Fed, rather than an actual transfer
of cash. Private banks maintain such accounts to fulll regulatory
requirements.) The contraction of the dollar money supply pushes
up short-term dollar interest rates, other things equal.
Note that this interest rate increase from the intervention reinforces the dollar-strengthening effect of the Fed's order ow (the order
to buy dollars in the FX market). To understand why, recall from
chapter 6 that asset approach models (such as the sticky-price monetary model) predict that an increase in the dollar interest rate should
cause the dollar to immediately appreciate. (The intuition for this effect
is that, other things equal, a higher interest ratenominal and real
on dollars makes dollar-denominated assets more attractive.) Interest
rates are an important part of fundamental value in these models.
Indeed, from the perspective of chapter 6's monetary models, it is the
interest rate that causes dollar appreciation when intervention is
sterilized, not the order ow.
Consider now what happens when the intervention is sterilized. In
this case, the intervention trade in the FX market is offset by another
trade designed to nullify the effect of the rst on the money supply
and interest rate. This offsetting transaction is called an open-market
operationa Fed purchase of domestic government bonds, DGB. In
this case, the offsetting transaction is a purchase of $100 million
worth of domestic government bonds, which re-injects 100 million
dollars back into the economy. This sterilized intervention is shown
on the central bank's balance sheet in gure 8.2.
To keep them straight, I have labeled the two transactions in
gure 8.2 with (1)s and (2)s. The rst transaction is the same as the
unsterilized intervention in gure 8.1: foreign exchange reserves
fall and money supply falls. The second transactionthe purchase of
$100 million worth of domestic government bondsincreases the
money supply by $100 million. The net effect on the money supply
(and therefore the interest rate) after both transactions is zero.
Let's walk through the transactions in more detail. Suppose that
this offsetting sterilization trade in domestic bonds is transacted
with Bank of America (B of A). Settling this trade involves a transfer
of $100 million worth of domestic government bonds to the Fedthe

Microstructure and Central Bank Intervention

225

Figure 8.2
Fed balance sheet: Sterilized purchase of $100 million with yen. FXR is foreign exchange reserves, DGB is domestic government bonds, and MS is money supplythat
is, money in circulation (currency plus monetary base). Transaction (1) is the unsterilized intervention; transaction (2) is the offsetting sterilization.

increase in the asset category DGBand credit of $100 million

to B of A's account with the Fed. This credit is an increase in Fed
liabilitiesthe increase in MS shown in the T-account. Taken
together, these two transactions leave the liability side of the Fed's
balance sheet unchanged: The reduction in Citibank's deposits at the
Fed are offset by the increase in B of A's deposits at the Fed, leaving
Fed deposits from the banking system as a whole unchanged. Even
though an FX transaction has taken place, the key fundamentals of
money supply and interest rate remain unaltered.
How Can Sterilized Intervention Work?
How can sterilized intervention inuence the exchange rate if it does
not alter the money supply and interest rate? Viewed from the macro
models presented in chapter 6 (e.g., the exible-price monetary
model and the sticky-price monetary model), the answer is not clear.
With no changes in money supply or interest rates, there is no
change in fundamentals (as dened by these models), so there is no
change in the exchange rate.
There are, however, two channels identied in the macro literature
through which sterilized intervention might still work:
1. The signaling channel: Sterilized intervention may signal future
money supply and interest rate changes, thereby affecting the exchange rate today.
2. The portfolio-balance channel: Sterilized intervention may affect
the exchange rate because it changes the currency denomination of
the supplies of assets held by the public.

226

Chapter 8

The signaling channel is easy to understand within the macro models

of chapter 6: if the central bank is signaling future money supply changes, then it is indeed changing fundamentals, so forwardlooking markets will move the exchange rate today. (The future
change in interest rates need not involve another FX transaction, i.e.,
it need not be the result of future unsterilized intervention.) 7 Though
easy to understand theoretically, empirically the signaling channel is
not well supported: subsequent changes in money supply tend not to
be consistent with the changes ``predicted'' by interventions (see, e.g.,
Kaminsky and Lewis 1996).
The portfolio balance channel is a bit subtler. To understand this
channel we need to move among the models of chapter 6 from the
monetary models to the portfolio balance model. Unlike the signaling channel, portfolio balance effects of sterilized intervention do
not operate through expected future payoffs (using the terminology
introduced in chapter 2); they operate through discount rates. To
understand why, note that when central bank intervention is sterilized, something has changed (though not the interest rate). From the
balance sheet example above, we see that the sterilized intervention
increases the Fed's holdings of dollar bonds and decreases the Fed's
holdings of yen bonds (foreign exchange reserves, FXR, are held in
interest-bearing bonds rather than currency). Thus, the intervention
reduces the public's total holding of dollar bonds and increases the
public's total holding of yen bonds. If dollar bonds and yen bonds
were perfect substitutes, then the public would be happy to make
this change and no adjustment in expected returns would be necessary (i.e., no adjustment in risk premia would be necessary). If these
bonds are imperfect substitutes, then changes in expected returns are
necessary to induce the public to make the change. These expected
return changes must occur from changes in the current level of the
exchange rate because payoff expectationsfor example, interest
ratesare not moving when intervention is sterilized.8
Though similar in spirit, there is a difference between this macro
portrayal of imperfect substitutability and the micro portrayal I introduced in the chapter 2 when discussing order ow information.
The difference is how the market learns about the size of the needed
exchange rate change. Under the macro story, the size of the portfolio shift forced on the public is known publicly (or, if not, it is
estimated using public information). All agents then agree on the
proper market-clearing exchange rate. Under the microstructure

Microstructure and Central Bank Intervention

227

story, in contrast, the market learns about the market-clearing exchange rate from the engendered order ow itself. Non-zero order
ows imply that more adjustment is required.
Announced, Unannounced, and Secret Intervention
The previous paragraph made a distinction between intervention
that is publicly known and intervention that needs to be estimated.
In the intervention literature, the terms used are announced intervention and unannounced intervention. Announced interventions
are released by the central bank or other ofcials over major news
wires simultaneously with the intervention trade. (These ofcial
news releases do not typically provide full information: they often
specify only that a central bank is trading without divulging the size
of the order.) Unannounced intervention is executed with a private
bank dealer or dealers without any ofcial news release. From an
information perspective, unannounced intervention comes in two
varieties. When the central bank itself places unannounced orders,
the counterparty dealer(s) know that the central bank is on the
other side. Consequently, though there is no ofcial news release, the
market does learn over the day that there has been intervention, and
subsequent reports by newswires typically convey what the market
has learned. Alternatively, central banks sometimes place orders
through agents who do not reveal that a central bank is the source of
the orderso-called secret intervention. One reason a central bank
might want to do this is to mimic the trades of private players. (This
mimicking strategy was described to me by former Fed ofcial Scott
Pardee).
Let me summarize the above discussion of intervention types.
There are nine types embedded in the discussion, illustrated by the
3  3 diagram in gure 8.3.
Intervention in Practice
In practice, intervention in major FX markets these days is typically
(1) sterilized, (2) unannounced, (3) effected through brokers, and (4)
infrequent. (That said, practices do differ across central banks, and
even across time at the same central bank.) As a consequence of features (1) and (2), most of the literature on intervention is focused on

228

Chapter 8

Figure 8.3
Types of intervention.

the lower left 2  2 block of gure 8.3.9 Let me address each of these
four features in more detail.
Intervention is typically sterilized in managed oat systems. This
is certainly the case for the managed oat systems that govern the
two largest spot markets: $/euro and $/yen. (These two systems are
in fact much closer to the pure oat model than to the pure xed
model.) The Fed, for example, sterilizes all its intervention as a matter of standard operating procedure (per Dominguez and Frankel
1993b: ``The stated U.S. policy is to sterilize its foreign exchange
intervention operations always and immediately''). The Bank of
Japan and the Bundesbank also typically sterilize, though not always
(Edison 1993). In pure xed-rate systems, on the other hand, interventions are frequently unsterilized; to be credible, central banks
need to adjust their monetary policies to defend the peg.
Most interventions are also not ofcially announced over major
news wires simultaneously with the trade (Dominguez and Frankel
1993b), although there is a trend toward more announcements. Even
without ofcial announcement, varying degrees of information gets
revealed. Consider the following description by ofcials at the Fed
(Smith and Madigan 1988):
Depending upon the degree of intervention visibility that is desired, we
will either call banks and deal directly or operate through an agent in the
broker's market. Most operations are conducted in the brokers' market,
though at the beginning of a major intervention episode we have sometimes
chosen to deal directly with several banks simultaneously to achieve maximum visibility. Within the brokers' market we can be more or less aggressive, hitting existing quotes or leaving trailing quotes.

Microstructure and Central Bank Intervention

229

When the central bank operates through an agent, the intervention

can be kept completely secret: the counterparty dealer trades and
settles with the agent, so the central bank source can remain unknown. Hung (1997) reports that about 40 percent of U.S. intervention between 1985 and 1989 was conducted secretly (secret by
Hung's denition means never reported in newspapers/newswires).
Klein's (1993) ndings are similar; moreover, he nds that many of
the interventions reported in newspapers never actually occurred,
casting some doubt on the accuracy of information in these reports.
To summarize, though we do not know the exact breakdown between secret intervention and partially revealed intervention, both
types are well represented in the major market data.
The last of the four empirical features is that central bank intervention in major currency markets is infrequent. One has to be careful on this score, however, because intervention tends to be episodic.
For example, from September 1995 to the end of 1997, the United
States never intervened, and it has intervened on only a handful of
days since then. The Bank of Japan, in contrast, has intervened more
frequently over the same period (though ofcial data are not available). The United States also completely refrained from intervention
from 19811984 (during the rst Reagan administration, which followed a laissez faire approach to currency markets). U.S. intervention was much more active from 19851992, however: over this
period the United States intervened thirty-six days per year, on
average.
The Importance of Intervention: Fixed versus Floating Rates
One can frame the importance of intervention by considering what
is perhaps the most central issue in exchange rate economics: the
choice between xed and oating exchange rates (see also section
7.3). As with most policy alternatives, policymakers are not limited
to choosing one or the other. Rather, most exchange rate systems are
hybrids, including the systems governing $/euro and $/yen. Nevertheless, understanding the polar extremes of pure xed and pure
oating helps to clarify the tensions inherent in hybrid systems. It is
precisely these tensions, and policymakers' attempts to relax them,
that generate the interesting open issues regarding intervention.
Consider rst a pure oat, which is dened by the absence of intervention: the central bank leaves the determination of rates wholly

230

Chapter 8

to the markets. The central bank therefore never trades in the FX

market and never adjusts exchange rate fundamentals (e.g., interest
rates) in order to inuence the exchange rate. Monetary policy is free
to tackle other policy objectives, such as stimulating growth or reducing ination.10
The mechanics of maintaining a xed exchange rate are more
involved. With open, unfettered capital markets, maintaining a xed
exchange rate requires that fundamentals (interest rates, money
supplies, etc.) are kept at levels consistent with the peg. When fundamentals are inconsistent, nancial markets react rapidly by, for
example, selling overvalued domestic currency to the central bank,
thereby draining the central bank's FX reserves.
This simple description of what happens under xed rates when
fundamentals are inconsistent with the peg begs two important
questions:
1. What exactly are these ``fundamentals'' that may be inconsistent
with the peg?
2. How does the central bank know when these fundamentals and
the peg are consistent?
The macro models in chapter 6 provide (theoretical) answers to
both of these questions: the fundamentals are the driving variables
in the macro models' exchange rate equations, and the equations
themselves provide the combinations consistent with a given peg.
But empirically, these equations are a poor description of actual exchange rate behavior. If not from macro models, then, how in the
real world does a central bank know when consistency between
the peg and fundamentals has been achieved? The answer is disarmingly simple: the central bank learns whether consistency has
been achieved from order ow. That is,
Fundamentals (broadly dened) are consistent with the peg when there is no
longer a signicant imbalance between buyer-initiated and seller-initiated
ordersthat is, private order ow is not signicantly different from zero.

This is not a trivial point; it is an indication that even central banks

have always done real-time monitoring of movements in fundamentals at the microstructural levelby monitoring order ow.11
Indeed, one can view changes in central bank reserves during crisis
episodes as a summary measure of private order ow (because the

Microstructure and Central Bank Intervention

231

central bank is typically the only buyer of the local currency during
crises).
Order Flow and the Macro Approach
Though the term ``order ow'' as used in microstructure nance was
not used in the macro literature on intervention before the mid1990s, the idea that trades are important for pushing prices around
was certainly present. Consider, for example, the following paragraph from Dominguez and Frankel (1993b):
The next question is how much of intervention policy's inuence on the exchange rate is the ``news'' effect and how much is the effect of actual ofcial
purchases and sales of foreign currency occurring at the same time? It is clear
that news has an effect to the extent that it causes investors to revise their
expectations of future rates of return. They buy or sell foreign currency in
response to the change in expected returns and thereby drive up or down the
current price of foreign exchange.

Perhaps most interesting about this passage is that even when the
channel is the news effect on expectations, the proximate cause of the
price change is still order owthrough the induced orders of private sector investors. For both effects thenthe news effect and the
direct effect of ofcial transactionsthe authors have order ow in
mind as the proximate driver.12 In 1993, when the DominguezFrankel book was published, we had little empirical sense of the
per-dollar price impact of orders. This is precisely what the microstructure approach is now providing.
8.2

Microstructure Models of Intervention

This section reviews recent theoretical work that applies microstructure models to central bank intervention. Two themes that distinguish this theoretical work are that it (1) recognizes asymmetric
information of different types, and (2) it addresses transparency
(secrecy) head on.
Addressing the asymmetric information theme rst, there are in
fact two basic types that are relevant for intervention (see also Evans
and Lyons 2000). The rst type is well recognized in the macro literature: asymmetric information between the central bank and the
public. The mere fact that the central bank controls several exchange

232

Chapter 8

rate fundamentals (e.g., interest rates and ination) gives it superior

information about the paths of those fundamentals. Another example of central bank information advantage is its access to privileged
information (e.g., data that are not yet public). This ``type 1'' information asymmetry between the central bank and public is at the
center of traditional intervention theory.
The second type of asymmetric information is the type that exists
between private agents. Traditional models of intervention based on
asymmetric information of the rst type do not include this second
type; all private agents in those earlier models learn about changes in
fundamentals simultaneously, and the mapping from these changed
fundamentals to price is common knowledge.
This ``type 2'' asymmetric information between private agents is
arguably more general than the rstat least as it relates to interventionbecause it plays a role in determining the effectiveness of
intervention irrespective of the channel through which the intervention effects operate. Consider, for example, the signaling channel:
irrespective of whether an intervention is sterilized, some agents observe the central bank's trade before others, and the market's understanding of this will affect the price response.13 This is also true for
portfolio balance channel: the portfolio shift that the intervention
trade is forcing on the public is not publicly observed, so price adjustment occurs via learning from the trading process.
Theoretical work within microstructure focuses more on asymmetric information of type 2. This is the case for two recent papers on
intervention: Montgomery and Popper (1998) and Bhattacharya and
Weller (1997). Montgomery and Popper use a variation of the rational expectations model to address whether, by using intervention,
the central bank can contribute to the efciency of the market's aggregation of type-2 information. In effect, intervention in their model
helps to achieve efcient sharing of information that would otherwise not occur. The focus of the Bhattacharya and Weller paper is a
bit different. They use a hybrid Kyle/rational expectations model to
show why, in the presence of type-2 asymmetric information, a central bank may prefer not to reveal its exchange rate target, nor the
size of its intervention. Their analysis provides a micro-based perspective on secrecy, with information elements that are missing from
earlier work.
Let me turn to the secrecy theme more directly by providing a brief
review of the literature on central bank secrecy that preceded Bhat-

Microstructure and Central Bank Intervention

233

tacharya and Weller's microstructure approach. Dominguez and

Frankel (1993b, 60) provide insight on that earlier literature when
they write:
[Many] reasons have been offered to explain why central banks may want
to keep their intervention operations secret. We recount them here, without
regard to how convincing they may be. These explanations of central bank
secrecy can, in turn, be grouped into three broad categories: reasons based
on a central bank's convictions regarding the efcacy of intervention at a
point in time, reasons based on the perceived depth or underlying volatility
of the foreign exchange market at a point in time, and reasons based on
[central bank] portfolio adjustment.

The rst of these broad categories is based on the fact that, at times,
central banks may act secretly because they actually prefer intervention to be ineffective. For example, the decision to intervene may be
coming from another branch of government, over the central bank's
objections. The second category of reasons for secret intervention is
based on the idea that the central bank may want, for example, to
calm so-called ``disorderly'' markets by providing a sense of a twoway market. The notion here is that if intervention were observable,
this would signal that private participants are not willing to provide
liquidity on one side of the current market (presumably leading to
greater disorder). The third category of reasons for intervention secrecywhich Dominguez and Frankel consider more persuasiveis
based on the fact that central banks frequently adjust their portfolio
holdings without any intent to move the exchange rate; they don't
want the market to confuse these transactions with purposeful intervention, so they intervene secretly.
Note that for both the rst and third of these categories, the central bank is choosing secrecy so that its price impact is minimized,
whereas ``effective'' intervention is typically associated with maximized price impact. For the second category, too, the central bank
is not choosing secrecy to maximize price impact, but to prevent
undesirable movements that might otherwise occur. In sum, none of
the three categories provides a rationale for increasing the efcacy of
intervention in the usual sense.
Like Bhattacharya and Weller, Vitale (1999) addresses why central
banks prefer secrecy using a variation of the Kyle model. To provide
a deeper view on this new line of intervention theory, let me describe
the Vitale model and his analysis in a bit more detail. (It is quite accessible on the basis of the Kyle model material presented in chapter

234

Chapter 8

4.) The added perspective on secrecy is helpful background for the

empirical analysis I present in section 8.3.
The Vitale (1999) Model
Vitale's model uses a one-period Kyle framework. At rst blush, it
may seem that a Kyle model of major FX markets is inappropriate: It
posits a trading structure that is centralized and batched, whereas
the major markets trade in a decentralized dealer structure. Vitale
offers a clever and compelling motivation for this modeling choice:
the batch market serves as a metaphor for lack of transparency.
When actual central bank orders are routed through an individual
dealer, the subsequent trades of that dealer are interpreted as noisy
signals of the intervention (noisy due to low transparency). Kyle's
batch framework captures this aspect of intervention because the Kyle
marketmaker cannot observe the identity of his or her clients and
therefore cannot distinguish between informed and uninformed
traders. Nevertheless, the orders the Kyle marketmaker receives still
convey information about fundamental value.
The Vitale model's single risky asset is foreign exchange, whose
random payoff V is Normally distributed, with mean P0 and variance
sV2 . (Vitale refers to the payoff V as the fundamental, and I adopt this
same terminology in describing his model.14) There is a single, riskneutral dealer who transacts foreign exchange with a group of
liquidity traders and a central bank. The central bank is informed,
in the sense that it knows the realization of V. All orders from the
liquidity traders and the central bank are grouped together
batchedso it is not possible for the dealer to observe the source of
orders. The dealer sees only the total order ow, which he uses to
extract information about the fundamental V and to set the marketclearing price P (according to a zero-prot condition, as in Kyle). In
the Vitale model, then, the central bank is the informed insider,
replacing the private informed trader of the Kyle model. As was the
case for the Kyle insider, the central bank can inuence the exchange
rate only by altering the dealer's expectations about V.
The model has three other key ingredients:
1. Intervention does not affect the value of V (i.e., it is sterilized).
2. Intervention targets the exchange rate P to some level P.

Microstructure and Central Bank Intervention

235

3. The central bank minimizes a loss function that depends on the

gap between P and its target P, and on the cost of intervention.
The loss L takes the form:
L XP V bP P 2 ;

8:1

where X is the central bank's (market) order. The parameter b indicates the degree of commitment to the predetermined target P (b is
nonnegative and is common knowledge). The rst term in the loss
function reects the capital commitment or cost of intervention: if the
central bank is buying foreign currency at a local-currency price
above V, this is costlya loss.
This loss function makes it clear that if the central bank were to
reveal both its intervention trade X and its target P, then the dealer
could back out the fundamental value V (because V would be the
only unknown). With the dealer knowing V, no attempt to target the
exchange rate would be effectiveonly the dealer's expectation of V
matters. Suppose instead that the central bank attempts to fool the
dealer by announcing false intervention trades. Vitale shows that
these announcements would not be credible (implying that the central bank would derive no benet).
The essence of the model is contained in the loss function, particularly in the fact that the target P and the fundamental V are not in
general equal. This creates the central tension. It is exactly when the
target is not consistent with the fundamental and announcements are
not credible that sterilized intervention is useful. Vitale shows that,
in effect, the central bank can buy credibility via costly trading,
which allows it to push the exchange rate toward its target.
Vitale's main result is that the central bank always prefers to conceal its target. When the target is secret, the realized exchange rate
is distributed more tightly around the target, which minimizes the
expected loss. (That the exchange rate has a distribution comes from
the realization of uncertain liquidity trader demands.) The basic intuition is the following: If the marketmaker knew anything about the
target, he could adjust the signals derived from order ow to make
them more informative about the fundamental V (in much the same
way that signals were adjusted in chapter 4 to form new, more informative signalstypically denoted with Z). The more informative
the order ow, the more tightly the price P is distributed about V,
which is not what the central bank wants. A secret target gives the

236

Chapter 8

central bank the greatest possible room to manipulate the beliefs of

the marketmaker.
8.3

Measuring the Price Impact of Intervention Trades

An important difference between private trades and central bank

trades is that private traders typically want to minimize price impact, whereas central banks want to maximize price impact. Central
bank trading desks do their best to ``time'' their intervention trades
to increase their price impact. In the past, this has been more a matter
of feel than anything else: data were just not available to determine
rigorously which market states are the high-impact states. 15
With the advent of electronic trading and the data that electronic
trading provides, research on this topic is at a turning point. Central
bankers enjoy detailed data on their own trades (timing, size, execution method, etc.). This allows them to generate a real-time picture of
exactly how intervention trades are absorbed in the markets, including changes in liquidity provision (e.g., limit orders at brokers), the
path of induced private order ow, and so on. From the perspective
of central bank research, this type of direct microstructure analysis is
an exciting prospect. Outside central banks, however, direct analysis
of individual intervention trades is generally not possible because
detailed intervention data are not released. (The Swiss National Bank
data analyzed by Payne and Vitale 2000 is a notable exception.)
A recent paper by Evans and Lyons (2000) circumvents these data
constraints by measuring intervention's price impact using an indirect approach. They address the following question: What does the
price impact of private trades teach us about the likely impact of
intervention trades?16 Private trade data include comprehensive
coverage of trade characteristics like time of day and state of the
market, which is not the case of intervention dataeven data available to central banks (interventions are relatively infrequent and
are clumped in terms of time of day). Another important feature
of the Evans-Lyons paper is that their data are not limited to highfrequency prices, but also include, crucially, order ow. Measuring
order ow is essential for determining trades' impact on prices.
Evans and Lyons's indirect approach is based on the fact that
central bank trades of a certain typesterilized, secret, and conveying no signaland private trades are indistinguishable.17 This allows them to analyze rather precisely the likely price impact of this
intervention-trade type.

Microstructure and Central Bank Intervention

237

Figure 8.4
Intervention transparency spectrum.

The Evans-Lyons analysis shifts attention toward a type of intervention that has received little attention of late. Since the publication
of Dominguez and Frankel (1993b), something of a consensus has
arisen that for intervention to be effective, it should be announced
and coordinated across multiple central banks. The Evans-Lyons
analysis does not dispute the effectiveness of this type of intervention; indeed, because their analysis is based on private trades only, it
has nothing to say about this type of intervention. Rather, the EvansLyons analysis highlights the effectiveness of intervention that lies on
the other end of the transparency spectrum (see gure 8.4). Private
trades areby their very nature``sterilized,'' so why not learn all
we can from the price impact of this type of trade?
Why Price Impact Might Be State Dependent
To understand why a given sized trade may have different price
impact in different market states, let us revisit the concept of event
uncertainty (introduced in chapter 5). The term refers to uncertainty
about whether or not private information is present (in contrast
to most trading models, where it is known with certainty that it
is present).18 Under event uncertainty, the information content of
trades varies depending on the market's state. In the Easley and
O'Hara (1992) model, trades are more informativethat is, have more
price impactwhen trading intensity is high. To understand why, consider an environment in which new private information may exist,
but does not necessarily exist. For example, suppose that with probability p no new information exists and with probability 1 p some
traders have observed some new information (either good or bad,
with known probabilities). Easley and O'Hara demonstrate that
rational dealers in this context will view a lack of trading as evidence
that new private information is not present. The upshot is that if
trading intensity is low, then an incoming trade induces a smaller
update in beliefs because it is more likely to be purely liquidity
motivated. On the ip side, trades occurring when trading intensity
is high are more likely to be signaling private information. Though

238

Chapter 8

Evans and Lyons (2000) do not analyze state-dependent price impact

explicitly, they do allow for it in their empirical specication.
A Model with Central Bank Trades
The intervention model that Evans and Lyons (2000) employ is
a variation on the Evans-Lyons model introduced in chapter 7.
Accordingly, I focus here on the salient differences. The model is
designed to show how the trading process reveals information conveyed by different types of intervention.19 It accommodates the main
intervention types (gure 8.3) and the main channels through which
intervention can be effective.
Though the model can accommodate all nine intervention types,
the empirical implementation in Evans and Lyons (2000) focuses on
the lower left cell of the matrix in gure 8.3: sterilized intervention
that is secret and conveys no signal about future central bank policy.
Evans and Lyons do not address the issue of coordinated intervention, though the model can be extended in this direction. (When
extending the model to address coordination, one should recognize
that coordination per se is more likely to operate through the signaling channel than the portfolio balance channel.)
Intervention Trades
Each day, one FX dealer among N is selected at random to receive an
order from the central bank. Let It denote the intervention on day t,
where It < 0 denotes a central bank sale (dealer purchase). The central bank order arrives with the public orders at the end of round 1,
and like the public orders, it is observable only to the dealer receiving the order. (Recall that in the model there are three trading rounds each day: customer-dealer trade, interdealer trade, and
another round of customer-dealer trade.) The central bank trade is
distributed Normally:20
It @ Normal0; sI2 :
In the case of sterilized intervention with no signaling, the central
bank trade It and the daily payoff increments DRt are uncorrelated:
CorrIt ; DRt 0;
for all t and t.

Microstructure and Central Bank Intervention

239

Model Solution
The model's solution is quite similar to that in the model of chapter
7. Specically, the equilibrium trading rules for round 2 interdealer
trading take the form
Tjt aCjt1

8:2

Tit aCjt1 It ;

8:3

where I use the subscript j to denote all dealers not receiving the
central bank order, and subscript i for the one dealer who does receive the central bank order. With these trading rules, the change in
price from the beginning of day t to the beginning of day t 1 takes
the familiar form:
Pt1 Pt b1 DRt b2 Xt ;

8:4

where Xt isas beforethe interdealer order ow observed at the

end of each day's round 2 interdealer trading.
As in the Evans-Lyons model of chapter 7, the price impact of
order ow Xt arises here purely for portfolio balance reasons. It
reects the price adjustment required to induce the public (in round
3) to absorb the total portfolio shift from the day's round 1 trading.
The only difference here is that the central bank's portfolio shiftin
the form of It must be absorbed by the public as well.
The punchline is as powerful as it is unsurprising: when sterilized
intervention is secret and conveys no signal of future payoffs Rt , the
market treats it like it would any other private trade. Put differently,
private trades in this model are equivalent to secret, sterilized interventions that have no impact on future monetary policy. Central
banksby choosing to intervene in this mannercan expect their
intervention to be as effective in moving price as the representative
private trade.
Summary of Key Intervention Results
Evans and Lyons (2000) present several results. First, using hourly
data from the $/DM and $/yen markets, they nd that private
orders (interdealer) have an immediate effect of about 1 percent per
$2 billion.21 This is similar in magnitude to the impact measured
from daily order ow in chapter 7. All of this immediate effect does

240

Chapter 8

not persist, however. (Persistent effects are surely more important to

ofcials who consider the current exchange rate misaligned.) Intervention strategy needs to be adjusted to maximize persistent effects:
the maximum occurs when trading intensity is about one standard
deviation above its mean. When intervention strategy is adjusted
accordingly, Evans and Lyons nd that buy orders have a persistent
effect of about 1 percent per $3 billion. These results suggest, for example, that if European ofcials want a 10 percent more valuable
euro, and they want that adjustment to persist, they would have to
sell about $30 billion of their foreign exchange reserves, and do so in
a way that mimics private transactions (subject to the caveat noted in
note 21).
The Evans-Lyons analysis also provides explicit guidance for
intervention in the form of both time-dependent rules and statedependent rules. Their time-dependent rules indicate that dollar for
dollar, order ow has the largest positive price impact during the
London afternoon (126 p.m., British Summer Timebst). During
their sample, New York begins signicant trading at about 3 p.m. bst
(8 a.m. est), so this interval includes the rst three hours of the New
York trading day (which is, incidentally, the interval within which
most Fed intervention occurs). The state-dependent rules indicate
that order ow has the largest positive price impact (1) when trading
volume is highconsistent with event uncertaintyand (2) when
markets are not especially turbulent or calm.22 These results are
consistent with those of Dominguez (1999), who also nds larger intervention effects during periods of heavy trading (and also around
macro events).
Evans and Lyons interpret their ndings as evidence of imperfect
substitutability, a necessary condition for the efcacy of the portfolio
channel. There is something of a consensus among nancial economists that portfolio balance effects from intervention are nonexistent,
or at least too small to be readily detectable. This has led some people to dismiss portfolio balance theory as irrelevant. Evans and
Lyons's approach harnesses additional statistical power that allows
them to detect the effects. The upshot: portfolio balance theory may
be more relevant than was thought.
Let me close this chapter by considering where the microstructure
approach to intervention is headed. The type of market data that are
now coming available allow very precise tracking of how the market
absorbs actual central bank trades and any information in them.

Microstructure and Central Bank Intervention

241

Central bank's with precise knowledge of their own tradesfor example, timing, announcements, stealth level, and so onwill be able
to estimate the impact of these various ``parameter'' settings. Armed
with appropriate data, a central bank could learn exactly how trading is affected in all three market segments (direct interdealer, brokered interdealer, and customer-dealer). This would include learning
about liquidity provision (on both sides of the market), transaction
activity, and the process of price adjustment. It is something like a
doctor who has a patient ingest blue dye to determine how it passes
through the digestive systemthe whole process becomes transparent. Such is the future of empirical work on this topic.

This Page Intentionally Left Blank

Customers: Underlying
Demand in the Economy

This chapter is an introduction to the trading of FX customers.

Understanding the demands of customersthe investors, the importers, the exporters, the corporate treasurers, and so onis an
important frontier for the microstructure approach. My goal is to
present some rst ndings along this frontier. I refer to these ndings as ``rst'' because the data necessary for this analysis have become available only very recently. Much work remains to be done.
In section 9.1, I provide background on the available customer
trade data. Though my empirical focus thus far has been on order
ow between dealers, section one should help to bring customerdealer trading into perspective. Section 9.2 describes the data in detail, including descriptive statistics. Because data such as these do
not appear elsewhere in the literature, description beyond what is
provided in chapter 5 is useful. Section 9.3 turns to regression analysis. It presents results from regressing price changes on customer
order ow and tests some hypotheses introduced in section one. Because the sample is sufciently long (6.5 years), these regressions
use monthly data, in contrast to the regressions of the previous
chapter that used daily data. (Recall that macroeconomic exchange
rate models are also typically estimated using monthly data.) The
nal section, section 9.4, presents a case study on the remarkable
drop in the yen/$ rate that occurred in October 1998 (around the
time of the Long Term Capital Management collapse). In a single
day, the yen/$ rate fell from about 130 to about 118, a change of
roughly 10 percent.
9.1

Background on Customers

The role of customer order ow is central to microstructure theory.

This is borne out by the models presented in earlier chapters: customer

244

Chapter 9

ow is the essence of each, in that it is the customer orders that catalyze a market response. By extension, it is not unreasonable to view
microstructure models in this respect as similar: their broad implications for the relation between exchange rates and customer ow
are the same (though the path of price adjustment may differ across
models).1
The importance of customer orders is obvious to practitioners
as well. Any FX trader or trading-desk manager would agree. One
trader I spoke with put it rather colorfully when he said that customer trades are the market's ``crack cocaine,'' by which he meant
that the customer orders are the market's catalyst, and that catalyst is
quite powerful.2 In keeping with this notion of customer ow as the
market's catalyst, proprietary information on those ows is a prime
driver of proprietary trading at the largest banks. (Smaller banks see
too little of the marketwide customer ow to make this information
useful.) Embedded in this behavior is the fact that banks nd customer ow information valuable for predicting exchange rate movements. Thus far, this book has concerned itself only with explaining
movements, that is, accounting for movements using concurrent
ow. That customer ow has predictive power as well (i.e., today's
ow predicts future movements) adds a new dimension. It is this
predictive dimension that most interests the practitioner audience of
the microstructure approach.3
So why in previous chapters have I focused so much on order ow
between dealers? There are two reasons. The rst is the simple fact
that until the dataset described below became available, researchers
had no alternative but to work with order ow between dealers.
The second reason is that despite the constraint on data availability,
there is justication for focusing on ow between dealers, due to the
differential transparency of customer-dealer and interdealer ow
(introduced in chapter 3). The reality of this market is that dealers
can observe some order ow from interdealer trades in which they
are not involved (e.g., from brokered trades). Customer-dealer trades,
on the other hand, are not observable except by the bank that receives them. Dealers therefore learn about other dealers' customer
orders as best they can by observing other dealers' interdealer trades,
and they set market prices accordingly. Although this learning from
interdealer orders is consistent with earlier chapters' empirical models, the ultimate driver of that interdealer ow is customer ow.
Let me provide a bird's eye view of how the customer ow data of
this chapter relates to the order ow data analyzed in earlier chap-

Customers

245

Figure 9.1
Trading volume pie.

ters. As noted in chapter 3, for the data sets of the 1990s volume in
the major spot FX markets splits into three roughly equal categories:
customer-dealer trades, direct interdealer trades, and brokered
interdealer trades. Figure 9.1 provides an illustration. Chapter 7's
analysis is based on data from the two interdealer categories. The
work of Evans and Lyons (2000), for example, uses data wholly
from the direct interdealer category. The work of Payne (1999) and
Killeen, Lyons, and Moore (2000a) uses data wholly from the brokered interdealer category.
As noted in chapter 5, data on customer orders in the major spot
markets are difcult to obtain. The only possible source given the
market's current structure is private banks themselves, but in general
these banks consider these data to be highly proprietary. Recently,
however, Fan and Lyons (2000) obtained customer trade data from
Citibank, a leading FX trading bank. (Citibank is among the top three
worldwide, with a market share in major-currency customer business in the 1015 percent range.) Citibank made these data available
only on a time-aggregated basisall the customer orders received
by this bank worldwide are aggregated into daily order ows (executed trades only). The data set therefore does not include individual
trades. Consequently, transaction-level analysis along the lines of
that introduced in chapter 5 is not possible.
Against this drawback, this data set has many advantages:
1. The data span more that ve years, so analysis at longer horizons
(e.g., monthly) is possible.

246

Chapter 9

2. The data cover the two largest markets: $/euro and $/yen.
(Before the launch of the euroJanuary 1999order ow data for
the ``euro'' are constructed from ows in the constituent currencies
against the dollar.)
3. The data include both spot and forward trades, but are netted of
any trades in FX swaps (because FX swaps do not have net order
ow implications; see chapter 3).
4. The data are split into three customer-type categories, corresponding to the three categories introduced in section 7.4: nonnancial
institutions (e.g., corporations), unleveraged nancial institutions
(e.g., mutual funds), and leveraged nancial institutions (e.g., hedge
funds).
Advantage (4) provides considerable statistical power for uncovering the underlying causes of order ow's impact on price. Do all
orders have the same price impact? Or might some order typessay
the orders of hedge fundsconvey more information than others?
Our ability to disaggregate order ow to answer these questions
brings us closer to a specication of the underlying information
sources.
Order Flows versus Capital Flows
When macroeconomists hear the word ``ows,'' they think of balance
of payments ows: real trade ows in the current account and capital
ows in the capital account.4 This book, in contrast, has hammered
(mercilessly?) on the concept of order ow. This is an appropriate
point to loop back on the relation between the two because we have
left the world of interdealer trading and are now considering the
currency demands of customers, which represent the underlying demands in the economy. They are the players whose demand shifts
matter for persistent price movements. (As an empirical matter,
dealers' net demands are too short-lived to matter over longer horizons, beyond the information they convey about underlying customer demands.5) Because balance of payments ows are familiar
territory for macroeconomists, this becomes a natural contact point
for readers from that background.
Now that we are examining underlying demands in the economy,
are we closer to macroeconomic notions of balance of payments
ows? At rst blush, it might appear that we are not much closer.

Customers

247

Consider, for example, the goods market approach introduced in

chapter 1 and revisited in section 7.2. True, that approach provides a
link to the balance of payments becauseunder that approach
demand for currencies depends on two balance of payments categories, imports and exports (e.g., imports increase the demand for foreign currency to pay for those imports, and that increase in demand
causes the foreign currency to appreciate). But we also learned in
chapter 1 that less than 5 percent of spot transactions comes from
imports and exports. The goods market approach is but a small piece
of the total order ow picture.
Another reason why our examination of underlying demands in
the economy may not bring us much closer to balance of payments
notions is because balance of payments transactions and FX transactions are not the same; a one-to-one relationship between the two
does not exist. To understand why, consider an import of $100 million of Japanese goods into the United States by a U.S. multinational.
(One could also use a capital account transaction for this example.)
Suppose the transaction is invoiced in yen, but the U.S. multinational
pays the invoice from yen it already holds at its Japanese subsidiary.
An import is logged, but there is no corresponding order in the FX
market; the link is not one-to-one. In time, one might expect some
rebalancing of the multinational's ``portfolio,'' but that need not occur immediately, and it need not involve order ow in the FX market
(e.g., suppose the Japanese subsidiary responds by increasing its
working capital borrowing in yen).6 The bottom line is that balance
of payments ows do not necessarily generate corresponding order
ow in the FX market. If learning from order ow is how dealers
determine price, then portfolio shifts in the form of balance of payments ows will not be counted unless and until they generate order
ow.
Though the preceding paragraphs suggest that the links between
customer ows and the balance of payments is loose, there are indeed links. At the very least, we are certainly closer conceptually to
the balance of payments in this chapter than in any of the previous
chapters. The empirical analysis of this chapter helps tighten the link
by separating order ow components according to ``current account
intensity'' and ``capital account intensity.'' For example, our customer
data allow us to separate the orders of nonnancial corporations
which includes the demands from current account transactions
from the orders of nancial institutions, who in relative terms are far

248

Chapter 9

more involved in capital account transactions. (Foreign direct

investments by nonnancial corporations, which enter the capital
account, are an important exception.) In section 9.3, we test whether
these two different order ow types have similar price impact. The
results shed light on which trader types tend to move price.7
Predictions about Customer Flow from Earlier Models
One of the conceptual contributions of the Evans-Lyons model was
its explanation for why the cumulative interdealer ow in gure 1.2
can follow a random walk, whereas at the same time the positions of
individual dealers return to zero at the close of each trading day. The
essential feature of the model that produces this return-to-zero result is that the aggregate position of dealers is fully absorbed by the
public each day. This market-clearing mechanism has strong implications for total customer ow. For example, it implies that:
hypothesis 1:
net to zero.

Marketwide, customer order ow each day should

This follows from the model's result that dealers nish each day with
no net position, which was itself a consequence of the assumption that dealers' risk-bearing capacity is small relative to the whole
market, that is, relative to all nondealers together.
The data presented below on customer order ow data represent
the orders received by one bank, not the customer ow received by
all banks. One would not therefore expect it to net exactly to zero
each day, even if the portfolio shifts model were literally true. With
these limited data, then, hypothesis 1 is untestable. Suppose, however, that the single-bank data represent a random sample, say 10
percent, of the marketwide customer order ow on any given day. In
this case, the Evans-Lyons model predicts that
hypothesis 2: For a single bank, customer order ow each day
should differ from zero due only to random sampling error.
hypothesis 3: For a single bank, customer order ow each day
should be uncorrelated with changes in the exchange rate.
Hypothesis 3 follows from the fact that the customer-order sample is
assumed here to be random. (It should therefore contain on average
as many realizations of the model's beginning of day ``shock'' orders

Customers

249

as it does end of day ``absorption'' ordersthe model's C1 's and C3 's,

respectively.)
It is also possible, however, that all customer orders are not
equally informative of subsequent market movements. Suppose, for
example, that customer order ows are not alike in terms of their
market impact. One might imagine two categories of customers:
high-impact customers and low-impact customers. If this were the
right description of the world, then a bank's customer orders might
not be representative of the customer-order population because the
bank could have a disproportionate share of high-impact customers.
9.2

A First Look at Customer Order Flow

Table 9.1 presents summary statistics for the three main customer
categories: nonnancial corporations, leveraged nancial institutions
(e.g., hedge funds), and unleveraged nancial institutions (e.g., mutual funds).8 The sample covers January 1993 to June 1999. For the
euro, the total trading volume across the three customer categories is
roughly balanced. For the yen, this is not the case: nonnancial corTable 9.1
Customer Trades: Volumes and Order Flow
Euro
Total
Trading
Volume

Yen
Cumul.
Order
Flow

Daily
Standard
Deviation

Total
Trading
Volume

Cumul.
Order
Flow

Daily
Standard
Deviation

Nonnancial
Corporations

539

25.7

0.09

259

3.3

0.07

Leveraged
Financial

667

2.5

0.16

681

16.1

0.16

Unleveraged
Financial

507

11.8

0.13

604

1.8

0.15

1,713

11.4

0.23

1,544

17.6

0.23

Total

Billions of euros for Euro and billions of dollars for Yen. Euro denotes the $/euro
market. Yen denotes the $/yen market. The sample for both currencies is January 1993
to June 1999. (Before the launch of the euro in January 1999, volume and order ow are
constructed from trading in the euro's constituent currencies.) Positive order ow in
the case of the euro denotes net demand for euros (following the convention in that
market of quoting prices in dollars per euro). Positive order ow in the case of the yen
denotes net demand for dollars (following the convention in that market of quoting
prices in yen per dollar). Daily standard deviation measures the standard deviation of
daily order ow.

250

Chapter 9

porate trading is less than half that for the other two categories (these
breakdowns may be bank specic, however). For both markets, the
daily order ow of the nonnancial corporations is the least volatile.
Cumulative order ow displays quite different characteristics across
the three customer categories. For the euro, unleveraged nancial
institutions are the largest net buyers and nonnancial corporations
are the largest net sellers. In the yen market, leveraged nancial
institutions are the largest net buyers (of dollars) and unleveraged
nancial institutions are the largest net sellers (though slight).
I turn now to plots of the customer ows, which provide a rst
glimpse of the possible link to exchange rate movements. Figure 9.2
shows cumulative customer order ows and the level of the exchange rate in both the $/euro and $/yen markets.9 Positive correlation is evident. Comparing these plots to gure 1.2which uses
the four months of daily interdealer data from Evans 1997one sees
that the correlation in gure 9.2 is not as tight at higher frequencies.
At lower frequencies, say monthly, the relation is manifested clearly.
The next section addresses this lower frequency relation more formally using regression analysis.
These plots also have implications for the three hypotheses introduced in section 9.1. Hypothesis 1 stated that marketwide customer
ow each day should net to zero. Because these plots show only one
bank's customer ow, not the whole market's, one is not able to reject this hypothesis based on these data alone. Hypotheses 2 and 3,
on the other hand, are directly testable. Hypotheses 2 and 3 stated
that this bank's daily customer ow should differ from zero due
only to random sampling error, and should therefore be uncorrelated
with exchange rate movements. These hypotheses appear to be
rejected: cumulative order ow received by this bank is correlated
with exchange rate movements.
What could explain this positive correlation? One possibility is
that it is not really therethe correlation is not statistically signicant. But I show in the next section that the relation is statistically
signicant. Another possibility is that hypothesis 1from which
hypotheses 2 and 3 are derivedmay not hold (i.e., marketwide
customer ow each day does not net to zero). For example, collectively, dealers may be maintaining nonzero positions. Though this
would not be surprising from day to day, it would indeed be surprising at weekly frequencies and lower (and these lower frequencies
are more relevant for the correlation in gure 9.2). Accordingly, I do

Customers

251

Figure 9.2
Cumulative customer ow and exchange rates. The plots show the spot exchange rate
and cumulative customer order ow received by the source bank. The sample for the
$/euro plot is January 1993 to June 1999. The sample for the yen/$ plot is January 1996
to June 1999 (the January 1993 to December 1995 period is not included due to the lack
of Tokyo ofce data). The spot exchange rate is expressed on the lefthand scale. The
cumulative customer order ow is expressed on the righthand scale (in billions of
euros for the $/euro plot and in billions of dollars for the yen/$ plot).

252

Chapter 9

not consider this a compelling reason to believe that the positive

correlation is due to a rejection of hypothesis 1. Another reason that
hypothesis 1 may not (appear to) hold is that dealers are achieving
their collective zero position, but do so by hedging with instruments
that do not enter our sample (e.g., currency futures or options). Per
chapters 3 and 5, however, the evidence suggests that FX dealers use
these methods of risk management rarely, if at all (cf. Naik and
Yadav 2000 for dealer hedging in other nancial markets).
Without any convincing reason to reject hypothesis 1, there remain
at least two other possibilities consistent with that hypothesis that
can explain the correlation in gure 9.2. (I offer these as suggestions
for future work; I cannot settle the issue based on analysis presented
here.) First, it is possible that the customers of this bank are on
average better informed. For example, Citibank is one of the very top
FX trading banks in the world, so it may attract a disproportionate
share of the most informative customer business. (More concretely,
suppose the orders of hedge funds are the most informative and this
bank receives more than its share of hedge fund orders.) A second
possibility consistent with hypothesis 1 that can explain the correlation relies instead on this bank's sheer size. Suppose this bank's customers are the same as customers marketwide, but because the bank
has such a large slice of total customer ow, its trades in the interdealer market generate disproportionate price impact. (A model
along these lines could include a cost of ``monitoring'' the trading
activity of various banks; in this setting it may be cost efcient to
place disproportionate weight on the interdealer trades of the largest
banks, despite their customers being no better informed than the
average customer; see, e.g., Calvo 1999.) Further theoretical work
will undoubtedly produce additional explanations. As additional
customer ow data become available, empiricists will be able to distinguish among them.
The Evans-Lyons Sample: MayAugust 1996
Before turning to regression analysis, let me provide an additional
plot that focuses on the $/euro over the same four-month period
analyzed by Evans and Lyons (1999). Recall that the Evans-Lyons
data reect order ow in the direct interdealer segment over a fourmonth period, May through August 1996. Those data, and their re-

Customers

253

Figure 9.3
Cumulative customer ow and exchange rates over the Evans-Lyons sample. The plot
shows the cumulative customer order ow in the $/euro market received by the
source bank from May 1 to August 31, 1996, and the spot exchange rate over the same
period. The spot rate is expressed in DM/$ on the lefthand scale. The cumulative customer order ow is expressed in millions of euros on the righthand scale (positive for
net dollar purchases).

lation to the exchange rate, appear in gure 1.2. Figure 9.3 presents
the May to August slice from the longer order ow series shown
in gure 9.2. They are plotted against the DM/$ rate for comparison purposes because the DM/$ rate is the rate plotted in gure 1.2.
Despite these data being the composite order ow for all the euro's
constituent currencies (against the dollar), the ow series tracks the
DM/$ rate quite closely. (Most of the constituent ow against the
dollar is in the DM market.) It is heartening that the customer-dealer
and interdealer ows tell a similar story. Integrated analysis of
series from these different segments is an important area for future
research.
9.3

Institution Types, Price Impact, and Information Structure

I turn now to regression analysis of the price impact of customer

order ow. These results draw from the analysis presented in Fan
and Lyons (2000). Table 9.2 presents results for the following model:
Dpt b0 b 1 Aggregate Customer Flowt et ;

9:1

where Dpt is the monthly change in the log spot exchange rate. Esti-

254

Chapter 9

Table 9.2
Price Impact of Aggregate Customer Orders
Dpt b0 b1 Aggregate Customer Flowt et
b1

R2

Euro

0.8
(3.8)

0.16

Yen

1.2
(3.0)

0.15

Monthly Data

T-statistics are shown in parentheses. The dependent variable Dpt is the monthly
change in the log spot exchange rate (the $/euro rate and the yen/$ rate, respectively).
The order ow regressors are measured over the concurrent month (in billions of euros
for the euro equation and billions of dollars in the yen equation). Estimated using OLS
(standard errors corrected for heteroskedasticity). The sample is January 1993 to June
1999. Constants (not reported) are insignicant in both equations.

mates of this model provide a more rigorous measure of the correlations displayed in gure 9.2. Fan and Lyons estimate the model at
the monthly frequency, which is the most common frequency for
estimating macro exchange rate models.10
Order ow is signicant in both regressions, with t-statistics well
above three. The R 2 statistics are respectable when compared to
those typically found for empirical macro models (typically in the 0
10 percent range), but remain far below those produced at the daily
frequency by Evans and Lyons (1999) using direct interdealer order
ow.
In the euro equation, the coefcient estimate of 0.8 implies that a
net purchase of one billion euros increases the dollar price of a euro
by about 0.8 percent. Similarly, in the yen equation, the estimate of
1.2 implies that a net purchase of $1 billion increases the yen price of
a dollar by about 1.2 percent. These price-impact coefcients are
roughly twice the size of those estimated by Evans and Lyons (1999),
who found that a net purchase of $1 billion increases the DM price of
a dollar by about 0.5 percent. (In fact Evans and Lyons nd that
some of the price impact at the daily frequency dissipates, suggesting that a comparable monthly frequency impact would be less than
0.5 percent.)
One might expect the customer ow to have a larger coefcient
for two reasons. First, the source bank's customer ow may be correlated with other banks' customer ow, and because those other

Customers

255

Table 9.3
Price Impact of Disaggregated Customer Orders
Dpt b0 b 1 Unlev: Fin: Flowt b 2 Lev: Fin: Flowt b3 Non-fin: Corp: Flowt et
b1

b2

b3

R2

Euro

1.5
(4.6)

0.6
(1.6)

0.2
(0.5)

0.27

Yen

1.1
(1.9)

1.8
(4.9)

2.3
(3.5)

0.34

Monthly Data

T-statistics are shown in parentheses. The dependent variable Dpt is the monthly
change in the log spot exchange rate. The three order ow regressors are the order
ows from unleveraged nancial institutions, leveraged nancial institutions, and
nonnancial corporations. Order ows are measured over the concurrent month (in
billions of euros for the euro equation and billions of dollars in the yen equation).
Estimated using OLS (standard errors corrected for heteroskedasticity). The sample is
January 1993 to June 1999. Constants (not reported) are insignicant in both equations.

banks' ow is not included in the regression, the source bank's ow

is getting ``credit'' for the impact. Second, if customer ow generates
rounds of knock-on interdealer ow via hot potato trading, then this
should induce an offsetting reduction in the interdealer ow's price
impact (because each dollar of interdealer ow corresponds to less
customer ow).
For the future agenda of the microstructure approach, a more important model is the following:
Dpt b0 b 1 UF Flowt b2 LF Flowt b3 NF Flowt et ;

9:2

where Dpt is the monthly change in the log spot rate, as before, and
the three regressors are the three customer ow categories introduced above: UF denotes unleveraged nancial institutions, LF denotes leveraged nancial institutions, and NF denotes nonnancial
corporations. As Fan and Lyons point out, this regression is important because it addresses whether orders of some participants are
more informative than those of others (as opposed order ow simply
being undifferentiated demand). Analyzing order ow's parts illuminates the information structure that underlies trading in this
market.
The results in table 9.3 indicate that these three different types of
order ow do indeed have different price impact. The orders of nonnancial corporations have no price impact in the $/euro market
and, strikingly, appear to be negatively correlated with price changes

256

Chapter 9

in the $/yen market. The orders of nancial institutions, on the other

hand, have uniformly positive price impact. In the $/euro market,
the key players (at least over this period) appear to be the unleveraged investors, or what practitioners call the ``real money'' accounts.
(The ``real money'' nickname for these playerse.g., mutual funds,
pension funds, life insurance companies, etc.is based on their
having real money to invest, as opposed to borrowed money.)
Though the unleveraged investors are important in the $/yen market
as well, they are not as important as the leveraged investors. The
ows of leveraged investors in the $/yen market show the largest
price impact in the table, as well as the greatest level of statistical
signicance. Note, too, that the R 2 statistics are roughly double those
produced in the univariate model presented in table 9.2.
These results suggest that order ow is not simply undifferentiated
demand. Rather, the source of the order ow matters quite a lot. Put
differently, if some banker were to subscribe to the undifferentiated
demand view of the world, then I know other banks that would be
delighted to exchange certain types of order ow information for
other types, dollar for dollar. Looking forward, an interesting avenue
for further research is whether the three parts of order ow have
different properties. For example, does cumulative order ow from
unleveraged investors follow something close to a random walk, as
their real money nickname suggests? In contrast, might the cumulative order ow of leveraged investors like hedge funds mean revert
over horizons of a couple months, as they unwind their temporary
speculative positions? Might the orders of non-nancial corporations
show the opposite tendency, that is that positive ow this period
means likely positive ow next period? These are important questions, the answers to which may help uncover not only the deeper
concurrent relationships between order ow and price changes, but
also the forecasting relationships that many banks appear to have
identied.11
9.4

Case Study: The Collapse of the Yen/$ Rate, October 1998

One of the most remarkable events in the postBretton Woods era of

oating exchange rates is undoubtedly the remarkable drop in the
yen/$ rate that occurred in October 1998. In a single day, the rate fell
from about 130 to about 118, a change of roughly 10 percent. On that
day, bid-offer spreads were said to have topped one yen, that is, one

Customers

257

percent or more in a market that usually trades with a spread of 12

basis points (and is arguably the second most liquid market in the
world, behind the $/euro market). There was no identiable macroeconomic news, at least not news that is usually associated with
exchange rate fundamentals. The nancial news at the time was
concentrated on the collapse of Long Term Capital Management
(LTCM), a hedge fund whose positions around the world had become so illiquid that unwinding them became impossible without
driving the fund's capital below zero.
Major banks attribute the yen/$ rate's drop to ``the unwinding of
positions by hedge funds that had borrowed in cheap yen to nance
purchases of higher-yielding dollar assets''the so-called yen carry
trade (The Economist, 10/10/98).12 This portfolio shiftand the selling of dollars that came with itwas forced by the scaling back of
speculative leverage following the LTCM crisis. Though received
wisdom suggests that this particular mechanism was at work, we
still have little direct evidence. One paper, Cai et al. (1999), provides
a rst cut on the issue. Its authors model volatility around the event
using an aggregate order ow measure and a comprehensive list of
macro announcements. (Their aggregate order ow measure is the
same weekly data from the U.S. Treasury used by Wei and Kim
1997.) They do nd that there is an independent role for order ow,
even after accounting for an extensive list of public news. But their
volatility model and aggregate ow measures cannot determine
which players were pushing prices in which direction.
Here I adopt a case study approach. I examine the behavior of
order ows by different players around the time of the event. Which
institution types were doing the dollar selling? Identifying the
sellers' types gives us insight into why the selling occurred (e.g.,
were they institutions that may have been ``distressed,'' in the sense
of being compelled to sell due to institutional constraints like loss
limits?).
In gure 9.4, panel a, I plot the daily yen/$ exchange rate and the
total customer order ow, which includes all three of the customer
categories. The vertical line marks the beginning of the yen/$ rate's
fall. An order ow ``trigger'' for the collapse is not present, at least
from the aggregate data. There is a small, downward blip in total
order ow at the very beginning of the rate's collapse, but it quickly
reverses itself. If anything, the aggregated ow received by this bank
suggests that the dollar selling began well before the actual collapse,

258

Chapter 9

Figure 9.4. Panel a

Cumulative total customer ow and the yen/$ rate around the October 1998 collapse.

Figure 9.4. Panel b

Cumulative ow of leveraged nancial institutions and the yen/$ rate around the
October 1998 collapse.

Customers

259

Figure 9.4. Panel c

Cumulative ow of nonnancial corporations and the yen/$ rate around the October
1998 collapse.

Figure 9.4. Panel d

Cumulative ow of unleveraged nancial institutions and the yen/$ rate around the
October 1998 collapse.

260

Chapter 9

around September 15. From September 15 to October 6, about $3

billion were sold by the source bank's customers.
The story becomes much richer as we disaggregate the customer
categories. Figure 9.4, panel b, shows that much of the advance selling of dollars in September was due to leveraged investors such as
hedge funds. These leveraged investors did not, however, account
for any abrupt selling at the time of the collapse (contrary to received
wisdom). Rather, they appeared to have provided liquidity, buying
about $1 billion at the time the dollar's price was collapsing. Panel c
shows that nonnancial corporations were also buying dollars at the
time of the price collapse; their total buying was about half the size
of the buying by leveraged nancial institutions, or $0.4 billion. The
story takes shape in panel d, which shows the powerful selling by
unleveraged nancial institutions at the time of the collapse. Not
only do these institutions appear to have been the trigger, they also
added fuel to the re, selling about $2.5 billion over the days preceding and including the collapse.
It will be interesting in the future to identify which of these unleveraged nancial institutions were the most important. Were they
U.S. institutions or Japanese? Did they become distressed in some
way precipitated by the LTCM collapse, or is the ``distressed players''
view of the event misguided? Some practitioners believe, for example, that the apparent heading among real-money instititions comes
instead from increasing emphasis on benchmarks (and the correlated
retreats toward benchmark weights that sharp market moves can
engender). Perhaps the unleveraged nancial institutions' portfolio
shift was not uncommonly large, but it happened to be coupled with
a particularly distressed set of leveraged nancial institutions who
could not provide the liquidity they would normally provide. There
is much room for future work to address these possibilities.
What are some of the larger implications of this yen/$ case study?
So abrupt a shift to a new exchange rate level (without macro news,
and persistent) leads one to consider the possibility of path dependence, which means that the actual path followed by a variable plays
a role in determining its ultimate resting point. (For example, a large
exchange rate swing can alter the structure of industries within a
countryforeign suppliers versus domesticwhich can change the
long-run equilibrium exchange rate.) The type of path dependence I
have in mind in this case comes from the sequencing of trades by
various customer types and how that sequencing may have mattered

Customers

261

for the ultimate exchange rate. For example, keeping the path of total
customer ow the same, if it had been the unleveraged nancial
institutions that had gradually ed dollars in early September, rather
than the leveraged institutions, might the new level of the yen/$ rate
in late October have been different? Though path dependence of this
kind is not a property of the earlier chapters' models, it is an interesting possibility for future work to consider.
Another larger implication of the case study is that liquidity in FX
markets varies over time, sometimes quite substantially. (In chapter
4, I provided a working denition of liquidity as an order's price
impact.) The order ow coefcients in tables 9.2 and 9.3 imply that
the relatively small portfolio shifts in gure 9.4 should have had
price impact that was much smaller, perhaps a few percent, rather
than the roughly 10 percent change that occurred. Though timevarying liquidity is also not a property of the earlier chapters' models, it is an issue that future work must consider. What triggers
liquidity changes? Might what appears on the surface to be changing
liquidity be due instead to changing order ow composition (given
the differential impact shown in the previous section)?

This Page Intentionally Left Blank

10

Looking Forward

This nal chapter is prospective. It addresses four topics: (1) what we

have learned, (2) what we still need to learn, (3) policy implications,
and (4) the FX market's future. Section 10.1 summarizes what we
have learned, focusing on the larger lessons. These larger lessons
can be divided into two broad groups, those that are more macrooriented and those that are more micro-oriented. In section 10.2, I
turn to what we don't yet know. What, for example, are the major
open questions? These questions are important directions for future
research. Beyond the major open questions, I also summarize other
topics that warrant additional work.
My review of what we've learned and what we have yet to learn
provides a nice segue to policy implications, section 10.3. Though
policy implications are implicit in the earlier chapters, they are easily
overlooked. Issues to which these implications apply include data
collection by ofcial institutions (for measuring and monitoring
market liquidity), central bank intervention, transaction taxes, capital
controls, and trading institution design. My main goal is to clarify
where additional work on the policy front is likely to be fruitful.
Section 10.4 closes the chapter with a discussion of how the institutions of the FX market are evolving. As noted in chapter 3, over the
last ten years these institutions have changed considerably (e.g., the
major role now played by electronic brokers), and the pace has not
slowed. To what extent might these near-term changes render the
microstructure approach obsolete? In fact, the approach can be successfully applied even if the structure does change. A goal of section
10.4 will be to explain why.

264

10.1

Chapter 10

What We Have Learned

Application of microstructure tools to FX markets has generated

insights that are new even to the broader microstructure literature.
For example, almost none of the work within microstructure nance
addresses macro asset pricing puzzles, in the spirit of the three big
puzzles I address in chapter 7. The microstructure approach to FX,
on the other hand, has been oriented toward macro-level puzzles
from the beginning, largely because people working in the area are
trained as macroeconomists.
Another example of cross-fertilization between FX microstructure
and the broader microstructure eld relates to information. The information structure of traditional models is relevant primarily for
equity markets. In the case of single stock, for example, earnings
announcements are a critical information event. Insiders with private
information about earnings would try to trade in advance of the
announcement (unless prohibited from doing so), hence the role for
order ow as an intermediate link between information and price.
But inside information of this kind in major FX markets is unlikely
market participants do not in general have inside information about
macroeconomic changes (e.g., interest rates, public announcements,
etc.). Despite this, order ow in FX markets does move prices, and
does so persistently. The search for precisely what this nonpublic
information might be is enlarging perspective within microstructure
regarding the nature and role of information.
The lessons learned thus far from work on order ow in FX markets can be divided into two broad groups: those that are more
macro-oriented and those that are more micro-oriented.1 The eld is
still young, so more work is required before these lessons can be
considered denitive. As data sets covering longer time periods become available, these lessons surely will be rened.
Let me begin with the more macro-oriented lessons. Though they
appear in earlier chapters, it is useful to collect them here. Six lessons
in particular seem especially important as we move forward.
Macro-Oriented Lessons
Order Flow Drives Long Horizon Prices
Order ow is the proximate cause of a large share of longer horizon
exchange rate movements. Even when based on order ow data sets

Looking Forward

265

that include only a fraction of marketwide ow, the concurrent impact of these ows accounts for 4070 percent of the persistent
movements in prices (i.e., at monthly horizons and longer, see, e.g.,
Payne 1999; Evans and Lyons 1999; Evans 2000; Rime 2000). As
richer order ow data sets become available (e.g., as they span a
larger share of the market and sign the ows more precisely), that
percentage may rise still higher.
Even Macro Announcements Affect Price via Order Flow
The ipside of the rst lesson is that concurrent macro announcements and other readily identiable macro changes do not directly
explain a large share of longer horizon price movements. Direct
mapping from concurrent macro changes to prices appears to be
limited (Evans 2000; Evans and Lyons 1999). Rather, order ow
appears to mediate most price movements (Evans and Lyons 2001).
(Recall the ``hybrid model'' in gure 7.1 that allowed information to
affect prices either directly, or indirectly via the order ow link.)
Though in some sense this result is a rediscovery of the well-known
empirical failure of macro models, work in FX microstructure is
clarifying the factors that supplement concurrent macro changes as a
driver of prices.
Price Elasticity with Respect to Order Flow is High
The elasticity of the exchange rate with respect to customer order
ow is roughly 0.8 percent per $1 billion (e.g., in the $/euro market,
per table 9.2; elasticity with respect to interdealer order ow is
roughly half that size). With world nancial wealth measured in
trillions of dollars, this is puzzlingly high. The result is consistent
with a common view that Milton Friedman's ``stabilizing speculators'' are not bold enough. Why this boldness might be lacking remains an open question. From an information-theoretic perspective,
however, high elasticity may not be so puzzling: small net ows may
be conveying signicant amounts of information.
Order Flow is a Factor in Floating Rate Volatility
We now have substantial evidence that order ow is an important
proximate factor driving volatility and may account for apparently excessive volatility under oating regimes (see, e.g., Evans and
Lyons 1999; Killeen, Lyons, and Moore 2000a). Though work on
order ow as a driver of price is focused on the sign of the relation-

266

Chapter 10

ship, there are also implications for volatility: a good model of return
rst moments is a good model of return second moments (but not
vice versa). Also relevant for this lesson is recent work by Osler
(2001). She nds that stop-loss orders on the buy (sell) side tend to
cluster at prices just above (below) round numbers, which can cause
trends to gain momentum once support and resistance levels are
crossed.
Acceleration of Information in Price
Empirical results in FX markets are not consistent with the ``accelerationist view'' that order ow simply accelerates the impounding
of information in price by, say, a few minutes (gure 3.4). As noted
above, our best measures of signed public-information ow are
virtually uncorrelated with the direction of exchange rate movements (at horizons of one year or less). This is incompatible with the
accelerated-by-a-few-minutes view.
At the same time, a fascinating possibility is that an accelerationist
story is indeed operating, but over much longer horizons. Suppose
order ow conveys individuals' expectations about macro fundamentals that are more distant (i.e., beyond the next month, quarter,
or even year). In that case, order ow serves to telescope this forward-looking information into today's spot rate. Note too that this
possibility is consistent with ndings that over longer horizons (e.g.,
three to ve years), macro variables do begin to account for a substantial share of exchange rate variation (despite concurrent macrofundamentals being virtually uncorrelated with exchange rates; see,
e.g., Mark 1995; Flood and Taylor 1996).
Order Flow Does Not Have to Sum to Zero
The Evans and Lyons (1999) model shows why order ow between
dealers does not have to sum to zero. This is important conceptually
because many people are under the mistaken impression that order
ow must sum to zero (and therefore that any ow measure that
correlates positively with price must be unrepresentative in some
way). This is not always the case.
Let me turn now to more micro-oriented lessons. By microoriented, I mean that they are based on intraday analysis of individual dealers. As such, they parallel more closely the bread and butter
work within microstructure nance.

Looking Forward

267

Micro-Oriented Lessons
Order Flow is Private Information
The behavior of individual dealers shows that they consider FX
order ow to be informative and that they set prices accordingly
(Lyons 1995, and other references in chapter 2). This empirical result
at the micro level accords well with the importance of order ow at
lower frequencies mentioned earlier. Moreover, all orders are not
alike in terms of their information content. Identifying which orders
are the most informative and who is behind them is illuminating
the market's underlying information structure. This type of analysis
goes well beyond the question of whether or not institutions affect
prices; it uses microstructure models to uncover new facets of nancial information.
Dealer Inventories Affect Price
Inventory control among spot FX dealers is strong relative to that
found for other markets. Most spot FX dealers prefer to end their day
atthat is, with no net position. Accordingly, the half-life of the
typical dealer's inventory is signicantly less than one day, and has
been estimated to be as low as ten minutes (e.g., Lyons 1998). These
half-lives are much shorter than those found in equity and futures
markets, where half-lives longer than one week are common. 2 Not
only do FX dealers control their inventories intensively, some also
adjust their prices to induce inventory-decumulating order ow.
(Lyons 1995 nds these inventory effects on price, but Yao 1998 does
not nd them for the dealer he track; see also Bjonnes and Rime 2000
and Romeu 2001.) This nding of inventory effects on price is important: they are the linchpin of the whole inventory branch of
microstructure theory, despite the fact that empiricists working on
markets other than FX have not found them.3
Hot Potato Trading Contributes to Trading Volume
Dealers describe hot potato trading as an important source of the FX
market's enormous trading volumes. The large share of trading between dealers that we nd in FX relative to other markets is consistent with a signicant role for hot potato trading. On the theoretical
front, our models show that hot potato trading is consistent with
optimizing behavior (e.g., Lyons 1997a). On the empirical front, we

268

Chapter 10

also nd direct evidence that hot potato trading is present (e.g.,

Lyons 1996b).
Central Banks in Microstructure Analysis
As a result of work on FX markets, analysis is being applied to types
of institutions not traditionally studied in microstructure, most notably central banks (see, e.g., Dominguez 1999; Vitale 1999; Kirilenko
1997). Central banks are not prot motivated, which introduces an
interesting (and policy relevant) dimension to modeling.
10.2

Directions for Research

Let me begin with what are, in my judgment, the four most important open issues in FX microstructure research.
Open Issue 1: Does the Information in FX Order Flow Reect
Payoff Information, Portfolio Balance Information, or Both?
Chapter 2 covered the economics of order ow information. In that
chapter, we saw that order ow can have persistent effects on price
through two basic channels: payoffs and discount rates. Evans and
Lyons (1999, 2000) model the persistent effects wholly as discount
rate effects, which macroeconomists typically call portfolio balance
effects.4 This feature of the Evans-Lyons model comes from their assumption that customer order ow Ct is uncorrelated with current
and future payoffs Rt . We cannot rule out, however, the possibility
that order ow conveys information about future payoffs. Consider a
simple variation on the Evans-Lyons model that claries how this
might work. Suppose customer orders are correlated with future
payoffs Rtk , where k is, say, one to ve years. This could occur if
payoff expectations held by individual customers are changing,
based on a constant ow of dispersed bits of information relevant
to forming those expectations (bits that not all customers share).
Dealers wishing to aggregate the information in those changing
expectations would respond to order ow the same way as in the
original Evans-Lyons model. If this variation on the model were
correct, then one would expect order ow to forecast future macroeconomic variables (as suggested by the long horizon accelerationist
view mentioned earlier). With the longer order ow series now
available (e.g., the seven years of $/euro data described in chapter

Looking Forward

269

9), we are just now getting enough statistical power to test this. Work
along these lines will help close the current gap between order ow
analysis and macro analysis.5
Open Issue 2: To What Extent Is Reverse Causality Presentfrom
Prices to Order Flowand Under What Circumstances (e.g.,
Distressed Institutions) Is It More Acute?
The full body of microstructure theory treats causality as running
from order ow to price. To rationalize the reverse causality, one
would have to develop an optimizing model in which investors at
the price-determining margin trade with positive feedback. This is
not an easy task, particularly since there is no compelling evidence of
ongoing positive momentum in exchange rate returns. Moreover, as
an empirical matter, work such as that of Killeen, Lyons, and Moore
(2000a) nds that in statistical terms, (Granger) causality does indeed
run from order ow to price, and not vice versa. The above theoretical and empirical arguments notwithstanding, causality almost
surely runs both directions, at least part of the time. Chapter nine's
analysis of the October 1998 plummeting of the dollar's value against
the yen provides some suggestive evidence of falling prices inducing
additional selling. If this selling was indeed ``distressed'' selling
brought about by loss limits or other institutional constraints, it will
be important to model these types of institutional constraints, and to
determine empirically to what extent they aggravate extreme market
movements.
Open Issue 3:

Why Doesn't Customer Order Flow Sum to Zero?

In the last of the ve macro-oriented lessons in section 10.1, I noted

that the Evans-Lyons model shows why total order ow between
dealers does not need to sum to zero. At the same time, that model
does predict at the daily frequency (and lower) that the sum of customer orders should be zero. Yet the private bank customer data
analyzed by Fan and Lyons (2000)and in chapter 9does not sum
to zero. Rather, it is positively correlated with exchange rate movements, in much the same way that order ow between dealers is
correlated with exchange rate movements. What accounts for this?
The answer is not clear, but the question is vitally important. Empirical work on the links between customer ow and exchange rates

270

Chapter 10

is the next frontier for the microstructure approach. Customers do,

after all, represent the underlying demand in the economy. Might it
be the case that Citibank, as a clear market leader, attracts a disproportionate share of informative customer ow, leading to the positive correlation? This is a possibility, but is as yet unsubstantiated.
Another possibility is that Citibank is just like other commercial
banks in terms of order ow composition, but because it is so large,
other banks tend to put additional weight on each dollar of its trading. As additional bank data sets with customer ows become available, researchers should be able to discriminate among the various
hypotheses.
Open Issue 4: Why Is the Price Impact of Trades From Different
Customer Types So Different?
This question, distinct from open issue 3, is at the heart of the research strategy that aspires to illuminate the underlying information
structure by disaggregating order ow. Do some types of order ow
convey more information about future macro paths? Do some types
convey more information about their own future ow or the future
ow of others? Answers to these questions will soon be in our grasp.
Other Directions for Future Work
The following list summarizes several issues that in my judgment
warrant attention in future work.
Data Integration
It will be valuable to integrate order ow data from all three trade
types (direct interdealer, brokered interdealer, and customer-dealer).
Two issues come immediately to mind that integrated data sets
would allow us to address.6 First, integrated data across the two
interdealer trade types would allow us to examine when and why
dealers prefer one trading mechanism to the other. Does the share of
trading executed in the direct interdealer market increase in times
of market turbulence because dealers are reluctant to expose limit
orders on broker screens at these times? This is an important question for institution design: if liquidity in the broker market tends to
dry up when the market is turbulent, then increasing reliance on
brokered trading may mean that the market is becoming less resilient

Looking Forward

271

(i.e., more vulnerable to liquidity breakdowns at times of extreme

price movements). Unlike work on equity markets (Goldstein and
Kavajecz 2000; Reiss and Werner 1999), work on FX microstructure
has not yet made progress on this issue.
A second issue that integrated data sets would allow us to address
is how customer trades feed into trades between dealers. To date,
much of the empirical work on trades' price impact is focused on
trades between dealers. But from a policy perspective (e.g., central
bank intervention), this is less important than measuring the price
impact of trades by customers. Perhaps more important, integrated
data sets would clarify the process by which the information in
customer ow is impounded in prices. The impounding occurs at
the interdealer stage (e.g., the Evans-Lyons model), but interdealer
orders are only noisy signals of the underlying customer demands. It
is unclear how much noise this two-stage process introduces into
prices.
Time Necessary to Achieve Strong-Form Efciency
Related to that last point is the issue of strong-form efciency. A
strong-form efcient market is one in which price impounds all
information, public and private. An important source of private information to banks is the order ow they receive from their own
customers. How long does it take the FX market to impound the information in customer ows in price? One day? A week? Banks that
rely on these customer ows to forecast exchange rates contend that
these ows have forecasting power at the one-month horizon. If true,
it will be a challenge to model where that longer horizon forecasting
power is coming from.
When Is Technical Analysis Rational?
Technical analysis is typically dened as valuation analysis that
relies wholly on past prices. Conceptually, this type of analysis
asserts that patterns embedded in past prices may help to forecast
future prices. In the FX market, it is well documented that some
participants rely on technical analysis, at least some of the time (see,
e.g., Taylor and Allen 1992). Is this a rational strategy? Perhaps. The
interesting question here is akin to that asked by Muth (1960) in his
seminal article on rational expectations, ``When are adaptive expectations rational?'' Muth solved for the class of time-series processes
for which adaptive expectations are efcient in an information sense.

272

Chapter 10

In our context, we might ask, ``Under what information structures

(and processes by which agents learn from trading) might specic
technical rules be rational?'' For work that is moving in this direction, see Osler (2000, 2001).
This issue of rational feedback tradinga type of technical analysiscertainly has implications for the direction of causality issue.
Future work might draw from existing microstructure analysis of
crashes and portfolio insurance because these are models in which
feedback trading is rational.
Why Is the Conventional Quote Size $10 Million?
In direct interdealer trading, the convention is for two-way quotes to
be good for up to $10 million (or 10 million euros in the $/euro
market). This size has important consequences for market liquidity,
but we know virtually nothing about how the convention came
about. It could simply be that FX dealers' cost functions are Ushaped as a function of trade size, with the $10 million point being
the cost minimizing size (see O'Hara 1995, 28). No work to date has
addressed this question.
Credit Risk and Market Structure
It is common for FX market participants to suggest that we have the
current, commercial bank driven dealer structure because commercial banks are ideally suited to manage credit risk. It is true that the
counterparty credit risk in FX transactions is enormous, but if this
really is what is supporting the dealer-market structure, it would be
a radical departure from the way the eld of microstructure has traditionally thought about ``equilibrium'' institutions. The traditional
focus of work on equilibrium institutions is on the management of
asymmetric information and price risk, not default risk.7
10.3

Policy Implications

There are ve broad areas where I envision microstructure analysis

of FX markets having impact on policy. The rst of these, central
bank intervention, was addressed in chapter 8, so I will not include it
here. I introduce the remaining four areas with an eye toward future
work that is likely to be useful. Though in some areas I make specic
recommendations, in other areas recommendations will have to wait
for further analysis of these policy questions.

Looking Forward

Policy Area 1:

273

The Price Impact of Order Flow

It would be useful for ofcial institutions to begin collecting data on

FX order ows. This will be particularly valuable for policymaking
in developing countries. Specically, order ow data allow one to
quantify the price impact of currency trades (both transitory and
persistent), as demonstrated repeatedly in earlier chapters. This has
a direct bearing on market liquidity. Indeed, price impact is what
liquidity is all about: the lower the price impact, the higher the
liquidity, other things equal. It would be interesting to get a sense for
how price impact in developing-country markets changes as a function of the state of the market (devaluation likelihood, etc.). Also, one
could determine whether customer forward trades have the same
price impact as customer spot trades of similar size. If not, one could
quantify the difference. (Many developing countries restrict or even
forbid forward trading on the belief that such trading is more ``speculative'' in nature than spot trading and is therefore more destabilizing.) One might also compare price impact across countries, to
determine which institutional structures are better at promoting
liquidity.8
The question of price impact is related to the issue of stability.
Policymakers in some developing countries believe that additional
liquidity is destabilizing. In theory, it is less liquidity that is destabilizing, not more liquidity: the less the liquidity, the larger the price
impact and the more prices move, other things equal. To make the
case that other things are not equal, in a way that might reverse the
relationship between liquidity and stability, one might use the discipline of microstructure trading models to identify the countervailing
forces at work.
Another issue that is particularly relevant in the developingcountry context is the stability of pegged rates (though this issue
extends to pegged rates in the industrialized world as well). Microstructure-style trading models help us to understand how and why
particular types of orders have price impact when exchange rates are
pegged (see, e.g., Calvo 1999; Corsetti, Morris, and Shin 1999). As an
empirical matter, we have a lot to learn about the types of order ow
that cause pegs to collapse (see Carrera 1999). A better understanding of these issues will aid in the design of more resilient pegged
regimes.

274

Policy Area 2:

Chapter 10

Emerging Market Design

A bread and butter policy issue in microstructure nance is how best

to design markets. This issue is relevant to FX as well, though treatment of it is limited by the fact that major currencies are traded in
a truly worldwide market, which makes it difcult for any single
regulatory authority to impose structural changes. Any country
attempting to alter or constrain the structure of trading within its
own borders would nd order ow migrating rapidly to other trading venues. As a practical matter, worldwide harmonization of this
type of policy change is infeasible at present.
The place where design of currency markets remains a hot topic
is in emerging markets. Most of these currencies are not traded on
a worldwide basis, due to lack of convertibility of one form or another. Because trading in these currencies is largely within-country,
it is feasible to legislate market design in a way that is not possible
in major markets. Microstructure analysis is well-suited to address
questions concerning whether edgling FX markets should be
organized as auction markets, as dealer markets, or both (for analysis along these lines, see Kirilenko 1997), as well as the level of
transparency that should be required. Institutions like the International Monetary Fund confront this type of policy question regularly.
The microstructure approach can provide valuable guidance.
Policy Area 3:

International Currencies

What role should specic currencies play in the international monetary and nancial system? Recent introduction of the euro has
brought this question to a level of policy relevance not seen since the
early 1970s (when the Bretton Woods xed-rate system collapsed).
Discussion of the role for an international currencythat is, a single
currency that acts as a universal means of exchangecenters on
three key aspects: (1) use as a reserve currency by central banks, (2)
use as an invoicing currency for international transactions, and (3)
use as a vehicle currency for currency transactions (vehicle currencies are used when the transaction cost of trading two currencies
directly is higher than the cost of trading them indirectly, via two
transactions through the vehicle currency). In aspects (2) and (3),
a single currency's success as an international currency is heavily
dependent on the level of transaction costs. Therefore, to predict

Looking Forward

275

whether the euro will be successful as an international currency, one

needs to model the transaction costs that will arise once it is fully
adopted. Though not an easy task, it is one that microstructure analysis is well suited to address. This type of analysis can also identify
which institutional features of the new euro will help reduce those
transaction costs. In this analysis, a central issue is the degree to
which a reduction in transaction costs stimulates trading, both directly, and indirectly through adoption as a vehicle currency. These
volume and liquidity responses to various policy alternatives are
the focus of recent work by Hartmann (1998a, b, 1999), Portes and
Rey (1998), and Hau, Killeen, and Moore (2000).
Policy Area 4:

Transaction Taxes

The issue of transaction taxes has attracted much attention among

exchange rate economists. Proponents of levying transaction taxes
tend to associate high volume with excessive speculation. As the literature has shown, however, much of this volume reects dealer risk
management (hot potato trading), rather than speculation. Imposing
a transaction tax would therefore impede risk management. Though
unintentional, this misunderstanding of the causes of high volume
could lead to bad policy. I emphasize the word ``could'' here because
microstructure analysis only adds a new dimension to this important
policy question, it does not invalidate the arguments of transaction
tax supporters. Looking forward, I expect that the microstructure
approach has a good deal more to contribute to this policy issue. (For
recent treatments using microstructure tools, see Hau and Chevallier
2000; Habermeier and Kirilenko 2000.)
10.4

Where FX Is GoingImplications for This Book

The major FX markets have undergone important changes over the

last ten years. Perhaps the most important change is the shift from
voice-based interdealer brokers to electronic interdealer brokers. This
trend away from human-intermediated transactions is evident in
many securities markets throughout the world.9 It shows no sign of
abating.
The shift from voice-based interdealer brokers to electronic brokers
is important in itself because, as described in chapter 3, electronic
brokers provide a different (mostly higher) level of order ow trans-

276

Chapter 10

parency than was provided by the voice-based brokers. This alters

dealers' information sets, which affects their trading strategies.
A larger implication of the shift to electronic brokers, however, is
that it suggests a future for spot FX trading that is more centralized,
electronic, andthis is the crucial partopen to customers. Major
markets are likely to shift toward a structure with open, electronic
limit order books that are accessible to a large number of market
participants.10 Under this scenario, institutions we have been calling
``customers'' would be able to provide liquidity to one another,
rather than having to depend on dealers. At that point, they would
cease being customersin the sense of always demanding liquiditytransformed instead into both liquidity demander and supplier.
Why might one believe that the market is going in this direction? Three pieces of evidence support the view. First, in June 2000,
three investment banks (Goldman Sachs, Merrill Lynch, and Morgan
Stanley Dean Witter) announced that they will be launching an electronic system of this kind for the U.S. bond markets (government
and corporate bonds). The U.S. bond market is currently organized
as a dealer market with characteristics similar to those in FX. Though
other security marketsfor example, equity and derivatives markets
have already shifted to a centralized electronic structure, those
other markets do not share the same FX market characteristics that
the bond markets share, and are therefore not as appropriate as
models. Second, in the last couple years, many new companies have
introduced forms of centralized trading for customers (e.g., FXall,
Atriax, FXchange, FXconnect, FXtrade, Gain.com, MatchbookFX.com,
among others). These new companies typically promote themselves
as operating at the fringes of the dealer-market structure. But there is
nothing obvious that prevents one of them from growing to a scale
that captures the network externalities inherent in concentrating
liquidity into a single pool.11
The third piece of evidence that, in my judgment, points to more
centralized FX trading in the future is that systems like EBS can be
opened up to larger customer-companies quite easily. The customer
relationships are there: the banks that own EBS are the same banks
that have customer relationships via their dealing services. The technology is not a major hurdle. The switch could be ipped in much
the same way as it promises to be ipped in the U.S. bond markets.
What might the catalyst be? A natural catalyst would be signicant
growth in market share by one of the new electronic entrants. If EBS

Looking Forward

277

decided to open its system to customers, it would be difcult for any

competitor to beat it. From the EBS perspective, it is essential to
maintain the threshold effects of network externalities in its favorif
the market is going in the direction of centralized customer trading,
EBS cannot afford to wait.
Does immanent market structure change threaten the relevance of
the microstructure approach? This is an important questioncrucial,
really, for the future relevance of this book. My answer will not surprise anyone who has read the previous nine chapters. The microstructure approach is concerned with much more than whether the
market is organized with a single dealer, multiple dealers, or a limit
order book. The role of order ow in conveying information transcends market structure, and the types of information that order ow
conveysparticularly the types with persistent price effectsare
unlikely to change radically when (if) the FX market changes in the
future. Put another way, the underlying information structure of this
market has more to do with the properties of the asset being traded
foreign exchangethan it does with the market structure per se.
Order ow will continue to tell us something about people's views
on how public information should be mapped into price. It will
continue to tell us something about current risk tolerances and
expectations about the future. In short, it will continue to convey
dispersed information that needs to be aggregated. And that is what
this book is all about.

This Page Intentionally Left Blank

Notes

Chapter 1.

Overview of the Microstructure Approach

1. Another approach in exchange rate economicsthe ``ow'' approachis a variant

of the goods market approach, so I do not include it separately in my three-approach
taxonomy. In addition to currency demand from goods ows, the ow approach
includes currency demand from the other main category of a country's balance of
paymentsthe capital account. Like its goods market cousin, the ow approach does
not typically model the exchange rate in a way consistent with nancial market efciency. I return to the parallels between the ow and microstructure approaches in
chapter 7.
2. The eld of macroeconomics is also moving in the direction of relaxing these three
assumptions. See, for example, the literatures on asymmetric information (e.g., Gordon
and Bovenberg 1996; Morris and Shin 2000), non-representative-agent macro (e.g.,
Caballero, Engel, and Haltiwanger 1997), microfoundations of monetary policy (e.g.,
Rotemberg and Woodford 1997), and new open-economy macro (Obstfeld and Rogoff
1996).
3. I continue with this three-point characterization in chapter 4 (which reviews
microstructure theory) by providing model summaries in three parts: Information,
Players, and Institutions.
4. Markets with limit order books include the Paris and Hong Kong stock exchanges,
both of which are operated electronically. Chapter 3 provides more detail along these
lines.
5. That dealers learn only from order ow is too extreme. It arises in standard microstructure models because these models assume that all information is private. Macro
models go to the other extreme, assuming all information is public. I return to this
in chapter 7, where I present an empirical model that admits both public and private
information.
6. My own view is that this central questionHow does altering the trading mechanism alter price?has unwittingly narrowed the scope of questions to which microstructure tools have been applied. I return to this issue in section 3.4, after laying more
groundwork.
7. Regarding spreads, let me communicate an experience that may speak to the reader.
I recall years ago presenting a paper on exchange rates that was rmly embedded in

280

Notes

the macro-orientation of the asset approach. Someone in the audience asked me a

question about bid-ask spreads. What went through my mindthough not my lips
was ``I couldn't care less about bid-ask spreads.'' To me, at the time, bid-ask spreads
were simply a nuisance parameter, bad manners next to an otherwise elegant approach. At the time, spreads meant microstructure to me, and microstructure meant
spreads. I no longer subscribe to this view.
8. Walrasian-style mechanisms do exist in real-world nancial markets; they are often
used to open trading (see, e.g., work on the pre-opening of the Paris Bourse by Biais,
Hillion, and Spatt 1999). Indicative orders are collected for a period before actual
trading begins, and these orders are used to determine the opening price (the orders
become executable at the opening, but typically can be retracted before the opening).
9. Actual orders come in different types, which I cover in chapter 3.
10. Andersen and Bollerslev (1997) establish another link between microstructure and
lower frequency relevance. They show that forecasts of low-frequency volatility are
more precise when based on high-frequency data.
11. Within international nance more broadly, there are four main puzzles, the three
listed above plus the ``home bias'' puzzle, which is that investors under-invest internationally. I do not include the home bias puzzle above because it is not clear it stems
from the exchange rate, whereas the other three puzzles directly involve the exchange
rate. Nevertheless, microstructure tools have already proven valuable for addressing
home bias (e.g., the trading model of Brennan and Cao 1997 and the empirical results
of Kang and Stulz 1997, among others).
These three exchange rate puzzles have analogs in other markets. For equities,
papers recognizing the three puzzles include Roll 1988, Shiller 1981, and Mehra and
Prescott 1985, respectively. (However, the equity-market version of the forward bias
puzzlethe so-called equity premium puzzleis a much looser analog than the other
two: the large risk premium on equity is rather stable over time and remains positive,
whereas the large risk premium in FX changes over time, including frequent changes
in sign.) Microstructure tools are just beginning to be applied to those major equity
puzzles (see, for example, Easley, Hvidkjaer, and O'Hara 1999).
12. Examples of information that needs to be aggregated by the FX market include
information about differential interpretation of news, about changing institutional risk
tolerance, and about changing hedging demands, among others (reviewed in chapter
2).
13. Of course, the microstructure approach also has its drawbacks (e.g., lack of publicly available data over long periods). I consider those drawbacks in later chapters.
14. The precise list of determinants depends on which model within the larger asset
approach is selected. Here our interest is simply a broad-brush contrast between the
asset and microstructure approaches. I return to precise specications of the asset
approach in chapters 6 and 7.
15. Work using structural models includes Glosten and Harris (1988), Madhavan and
Smidt (1991), and Foster and Viswanathan (1993), all of which address the NYSE;
structural models in a multiple-dealer setting include Snell and Tonks (1995) for
stocks, Lyons (1995) for currencies, and Vitale (1998) for bonds.
16. I use the notation ``DM/$'' here because I am referring specically to the value of
an exchange rate, which is quoted in the marketplace as the deutschemark price per

Notes

281

dollar. When referring in general terms to an FX market, practitioners typically list the
dollar rst: orally, they refer to this market as ``dollar-mark,'' typically written as the
``$/DM market,'' or simply as ``$/DM'' (the Bank for International Settlements also
follows this dollar-rst convention in its FX market surveys; see section 3.2). In this
book, I remain true to these differing conventions. When describing an actual rate,
precision requires me to write it as it is traded (e.g., DM/$, yen/$, and $/). When
referring in general terms to a particular market, I always list the dollar rst ($/DM,
$/yen, and $/). Finally, when I use the symbol P in equations to denote the exchange
rate as a price, P always denotes the dollar price of the other currency ($/other).
17. Because the Evans (1997) data set does not include the size of every trade, this
measure of order ow is in fact the number of buys minus sells. That is, if a dealer initiates a trade against another dealer's DM/$ quote, and that trade is a dollar purchase
(sale), then order ow is 1 (1). These are cumulated across dealers over each 24hour trading day (weekend tradingwhich is minimalis included in Monday).
18. Readers familiar with the concept of co-integration will recognize that it offers a
natural means of testing for a long-run relationship. In chapter 7, I present evidence
that cumulative order ow and the level of the exchange rate are indeed cointegrated
(i.e., the relationship between order ow and price is not limited to high frequencies). I
also present models in chapter 7 that show why a long-run relationship of this kind is
what one should expect.

Chapter 2.

The Economics of Order Flow Information

1. A related, less extreme view is that order ow may convey some non-public information but the information it conveys is likely to become public soon. In this case,
order ow advances the impounding of information in price by only, say, a few
minutes. This less extreme view is not consistent with the data, however. To a rst
approximation, our best measures of public information ow are uncorrelated with the
direction of exchange rate movements (at annual or higher frequencies), whereas order
ow is correlated with exchange rate movements. I return to this issue in chapter 3.
2. I have in mind here the major FX markets. This kind of private information is more
reasonable in emerging markets. For example, the IMF (1995) reports that Mexican
investors were the rst to ee Mexico in the period immediately prior to the December
1994 devaluation. It is not unreasonable to believe that certain people close to the devaluation decision had inside information and were able to act on it.
3. This section emphasizes empirical work that uses order ow data, as opposed to
empirical work that uses price data only.
4. A third method for testing whether order ow effects are persistentrelated to the
rst twois testing for cointegration between cumulative order ow and the level of
the exchange rate (see Killeen, Lyons, and Moore 2000a; Bjonnes and Rime 2000).
5. Hartmann (1999) nds that FX spreads widen with unexpected volume, which is
consistent with an adverse selection component.
6. In the case of FX, volatility over the lunch period can still be calculated because
trading continues in other trading locations, such as Singapore and Hong Kong.
Andersen, Bollerslev, and Das (2000) verify the signicance of this volatility increase
over lunch. They do nd, however, that some other results in the Ito et al. paper are

282

Notes

sensitive to an outlier in the data set. Covrig and Melvin's analysis (1998) of the same
Tokyo experiment excludes the outlier observation, but still nds evidence that Japanese banks are relatively well informed, corroborating the Ito et al. interpretation. The
Covrig-Melvin evidence is based on price leadership by Japanese banks. For related
evidence on volatilities during the LondonNew York trading overlap, see Hsieh and
Kleidon (1996).
7. By ``better'' I mean lower conditional variance.
8. Empirical evidence of inventory effects on FX prices is in Lyons (1995).
9. To understand why the change in the risk premiumi.e., the expected return
changes price, think of a pure discount bond: when the market interest ratethe
expected returnchanges, the price of the bond must change, even though the cash
ow at maturity does not change.
10. This logic is also behind my choice of V in this book to denote payoffs, rather than
the more customary F; the symbol F is too easily interpreted as the whole of fundamentals, an interpretation that is too narrow in my judgment.
11. This expectations story is also applicable to information about discount rates, in
which case it would fall into type (2) or type (3).
12. Other proxies for expectationssuch as time-series measures like ARIMA or VAR
modelsdo not share this backed-by-money, information-encompassing property. A
quote by Frankel and Rose (1995, 1701) provides some perspective on these time-series
measures. In their words, ``To use an ARIMA or VAR process as a measure of what
agents expect, is to ascribe to them simultaneously not enough information, and too
much. It does not ascribe to them enough information, because it leaves out all the
thousands of bits of information that market investors use. . . . It ascribes to them too
much information . . . because it assumes that they know the parameters of the statistical process from the beginning of the sample period.'' Also relevant is the discussion
in Engel (1996), where he describes ``peso problems'' in exchange rates as a case where
the market had more information than empiricists, and ``learning'' as a case where the
market had less information than empiricists (ex-post).
13. Within the literature on asset pricing more broadly (e.g., the pricing kernel
approach), work on time-varying discount rates is sometimes referred to as addressing
``stochastic discount factors.'' As is mine, the focus of this literature is on variation in
the risk-premium component of discount rates, not the risk-free-rate component.
14. Though this discount rate effect on price emerges in most models as a risk premium, for technical convenience sometimes models are specied with risk-neutral
dealers who face some generic ``inventory holding cost,'' which produces similar
results.
15. Readers familiar with microstructure theory will recognize that this gure assumes that there is no xed component to the spreadi.e., the effective spread shrinks
to zero as the size of the incoming order shrinks to zero. This is a detail that need not
concern the more macro-oriented reader.
16. For evidence of imperfect substitutability across stocks, see Scholes (1972); Shleifer
(1986); Bagwell (1992); and Kaul et al. (2000), among others. For at least two reasons,
imperfect substitutability may be more applicable to currency markets than to markets
in individual stocks. First, note that the size of the order ows that the $/euro market needs to absorb are on average more than 10,000 times those absorbed in a

Notes

283

representative U.S. stock (e.g., the average daily volume on individual NYSE stocks
in 1998 was about $9 million, whereas the average daily volume in $/DM spot
was about $150 billion). Second, there are far more individual stocks that can substitute for one another in portfolios than there are individual currencies (particularly
major currencies).

Chapter 3.

The Institutional Setting

1. This statistic is from the Bank for International Settlements, BIS (1999a). I examine
BIS survey data in more detail in section 3.2.
2. BIS (1999a), 17.
3. Briey, covered interest parity is a no-arbitrage condition that implies that
F$= =P$= 1 i$ =1 i , where F$= is today's one-period forward rate ($/), P$= is
today's spot rate, i$ is today's one-period nominal interest rate in dollars, and i is
today's one-period nominal interest rate in pound sterling. In economic (and order
ow) terms, taking offsetting positions on the lefthand side of this equation is equivalent to taking offsetting positions on the righthand side (in markets free of capital
controls).
4. One fascinating branch of related work on derivatives involves derivative and spot
market interaction (as opposed to analysis of a derivatives market in isolation). The
question is whether introducing a derivatives market can rectify specic market failures. See, for example, Brennan and Cao (1996) and Cao (1999), among many others.
5. There is a large literature on limit-order auction markets. See, for example, Glosten
(1994); Biais, Hillion, and Spat (1995); Chakravarty and Holden (1995); Harris and
Hasbrouck (1996); Handa and Schwartz (1996); Seppi (1996).
6. The terms ``dealer'' and ``marketmaker'' are typically interchangeable in the academic
literature. When there is a distinction, the term dealer is used for dealership-market
settings (like the FX market) and marketmaker is used for hybrid auction-dealership
settings (like the New York Stock Exchange). My use of these terms will be consistent
with this distinction throughout the book. I should also note that FX practitioners
typically use the term ``trader'' to describe a dealer. Because the term trader is rather
general, I opt for more specic terms when possible.
7. There is a screen called Reuters FXFX that displays dealer quotes, but these quotes
are not rm. (I provide more detail on the FXFX screen in chapter 5.) Though quotes
on the electronic brokerage screens that I describe later in this section are rm, these
brokerage quotes reect only a subset of rm quotes in the market at any given time.
For currency options, though some trading occurs on exchanges, most occurs in a
decentralized, multiple-dealer setting. For currency futures, trading around the world
tends to be centralized in various futures exchanges.
8. Though the spot FX market and the U.S. government bond market currently share
a similar structure, the way the bond market trades is evolving toward a more centralized auction structure (like the Paris Bourse and Hong Kong Stock Exchange, as
described above). See section 10.4 for more detail. Other countries' bond markets have
already moved to a centralized auction format (e.g., the Italian government bond
market). Recent work in the burgeoning literature on bond market microstructure
includes Fleming and Remolona (1999) and Vitale (1998).

284

Notes

9. See BIS (1999a). The interdealer share of volume for NASDAQ and SEAQ (London's
Stock Exchange Automated Quotations System) is less than 40 percent (see Reiss and
Werner 1995).
10. One of the big changes in the FX market over the last few years is the shift away
from voice-based brokers, where prices are advertised over intercoms at dealers'
desks, and toward electronic brokers, where prices are advertised over a screen. Indeed, there is evidence that the electronic brokers have also taken market share away
from direct interdealer trading (BIS 2001). Currently, the dominant electronic broker in
$/euro and $/yen is EBS. The other major electronic broker is Reuters 2000-2. For
more on this shift to electronic brokerage, see the Financial Times ``Survey: Foreign
Exchange,'' 5 June 1998. Interdealer brokers are also quite important in U.S. bond
markets.
11. I should mention two important variations on this story. First, banks do their best
to ``internalize'' as much customer order ow as possible. That is, they want to match
the incoming customer buy orders with incoming customer sell orders. When successful, any net ows that get passed on to the interdealer market are much reduced.
Second, the order ow lifecycle does not really ``end'' once nondealers reabsorb net
balances. Rather, that reabsorption moves through a ``chain'' of liquidity providers, the
rst nondealer link being hedge funds and banks' proprietary trading desks, and the
last link being so-called ``real money accounts.'' The real money accounts, such as
mutual funds and pension funds, are institutions that absorb positions over longer
horizons. Chapter 9 on customer trading provides more perspective.
12. One important source of noise comes from the fact that brokered trading systems
do not indicate the size of all individual transactions in real time, so it is not possible
from information about transactions' signs to construct an exact order ow measure.
For example, on the D2000-2 system, dealers only see an r on both the bid side and the
offer side if the quantity available at the best price is $10 million or more; otherwise, they see the quantity available. If, after a trade, the screen is still showing r on
both the bid and offer sides, then one cannot infer the size. Often, though, a trade
causes the size to drop below $10 million, or the quantity on one side is exhausted, in
which case dealers have received information about size. EBS, the other major electronic broker, is similar. Both systems also provide a high-frequency (but not complete) listing of deals as either ``paid'' or ``given,'' where paid indicates buyer-initiated
and given indicates seller-initiated.
Research data sets are a bit different. (In chapter 5 I review the data sets that cover
brokered interdealer trading.) These data sets, constructed from records well after
actual trading, provide exact measures of order ow, but these measures were not in
the dealers' information sets while trading.
13. For comparison, a stock selling at $50 per share with a spread of only 10 cents
would still translate into a spread of 20 basis points10 times that in $/euro.
14. There is a common misconception that dealers within a given bank pass their
positions from Tokyo to London to New York during each trading day. This is not true
in generalindividual dealers are responsible for managing their own positions. What
is true is that the bank's book of customer limit orders is passed from Tokyo to London
to New York. These unexecuted orders are not the same as dealer positions.
15. There are two nice multimedia resources that bring still more perspective on the
life of an FX dealer: (1) Goodhart and Payne 1999, and (2) Citibank's ``Bourse Course.''
The former provides a visual account of trading in an electronic interdealer broker

Notes

285

system (Reuters Dealing 2000-2). The latter is a simulated trading game designed to
replicate the FX market; it includes trading of all three types depicted in gure 3.1.
(Contact Citibank directly for more information.)
16. Chapter 5 describes the Reuters Dealing 2000-1 system in detail, including the rich
data it generates for empirical work.
17. For the equity dealer comparison, I use the numbers from Hansch et al. 1999 for
the London Stock Exchange because, unlike the NYSE, the LSE was a pure dealership
market and therefore more comparable to FX. The authors nd that dealers make a
prot of roughly 10 basis points on the average transaction. Though the authors do not
provide an average turnover by dealer, they do provide data that allow a rough estimate. The average daily turnover for FTSE-100 stocks is about $10 million (6.9 million). This total turnover is divided among dealers (active dealers make markets in
many stocks). Given the market shares the authors report for the more active dealers,
and given the number of stocks in which each makes markets, the estimated average
turnover of $10 million per dealer is about right.
18. The median quoted spread in the sample of 0.0003 DM/$ is the mode as well: that
spread size accounts for about three-quarters of all the dealer's bilateral interdealer
quotes.
19. The story is a bit subtler here, though, because this dealer does not have much
customer business. In my judgment, the right way to think about this dealer is that he
was supplying liquidity and inventory management services to other dealers that have
more customer business. So, in effect, there is a kind of ``tiering'' in the interdealer
marketFX dealers are not a homogeneous lot. My understanding is that dealers of
this type are much less protable now that electronic interdealer brokers play such an
important role (table 3.2 corresponds to a trading week in 1992).
20. An important determinant of whether $100,000 per day is ``large'' is the amount of
bank capital this dealer ties up when trading. If the capital required were $1 billion
per dayequal to his total volumethen $100,000 would be rather small. In reality,
the capital required to support a dealing operation like this one, which involves only
intraday positions, is far smaller than $1 billion. For more on whether banks' FX profits come from positioning versus intermediation, see Ammer and Brunner (1997).
21. Beyond the BIS, there are several institutions that serve as semi-ofcial coordinators in establishing FX trading practices and serving as forums for debate. In
the United States, that institution is the Foreign Exchange Committee of the Federal
Reserve Bank of New York. For more information, see their Web site at hwww.ny.
frb.org/fxci.
22. The role of derivatives in determining market resilience and efciency was a topic
of increasing public policy concern in the 1990s. The considerable emphasis afforded
derivatives in the BIS survey is in keeping with this policy concern. One should not
lose sight, however, of the fact that in FX it is the spot markets that generate most of
the order ow (per my point in the text about the largest of the FX derivative marketsthat for forex swapsgenerating lots of turnover, but no order ow).
23. The statistic of 10 percent for 19891992 is (390/356) 1, with 390 being the corrected spot turnover for 1992 of 400(800/820) and 356 being the corrected spot turnover for 1989 of 350(600/590). Corrected spot turnovers for the other years are
calculated similarly. Note that this correction should be viewed as approximate for the
spot turnover because the mix of currencies that are traded spot does not match exactly the mix of currencies in total turnover.

286

Notes

24. Part of this large amount of dollar trading (but only part) is due to use of the dollar
as a vehicle currency. When the dollar is used as a vehicle, then when trading two nondollar currencies one does so indirectly, by going through the dollar rst. With the
launch of the euro, the dollar's role as an international currency has received much
attention recently. See, for example, Hartmann 1998a and Portes, and Rey 1998. Much
of the analysis in this area turns on microstructural matters; it is a natural application
of microstructure tools to what has traditionally been a macro topic.
25. Of the major central banks, only the Bank of Japan was intervening much in the
major markets at this time. See chapter 8 for more on intervention.
26. Do not be misled by the expression ``transacted by brokers'' that appears in the
rst sentence of section 7 (BIS 1999a): brokers do not themselves transact; they merely
facilitate the transactions of dealers (for a fee).
27. The U.K. market is the biggest in terms of spot trading, accounting for about 28
percent of the worldwide spot total (table E-9).
28. There is a second major policy issue in equity-market transparency that is more
relevant to pre-trade information: Is the public entitled to see all the price quotes that
the dealers observe? As for pre-trade transparency in FX, an issue that warrants
greater policy attention is whether customer limit orders at individual banks should be
exposed to the market at large (as was recently imposed on the U.S. NASDAQ market;
see Weston 2000). Aggregation of these limit orders represents, in some sense, the
wider market's latent demand.
29. I should add, though, that transparency in centralized exchange markets (stock
and futures exchanges) was historically imposed by the members on themselves, prior
to government regulation. For work on transparency in equity markets see, for example, Naik, Neuberger, and Viswanathan (1999) and Pagano and Roell (1996).
30. In June 2000, three investment banks (Goldman Sachs, Merrill Lynch, and Morgan
Stanley Dean Witter) announced that they were launching a centralized electronic
system for the U.S. bond markets (government and corporate bonds). This system, if
successful, represents a fundamental transformation of that market, not just in terms of
increased transparency, but also in terms of access and cost of liquidity.
31. As an aside, there is an interesting argument why increasing transparency of the
trading process might reduce the information in prices: increased transparency might
reduce incentives to invest in information production. This effect does not arise in
standard trading models because the amount of private and public information is
assumed xed.
32. I consider this possibility further in chapter 10.

Chapter 4.

Theoretical Frameworks

1. For a broader treatment of theoretical microstructure, see O'Hara (1995). For a recent survey, see Madhavan (2000). For historical perspective on trading models, see
Keynes (1936); Hicks (1939); Working (1953); and Houthakker (1957). For early
reviews of microstructure theory's relevance for the FX market, see Flood (1991) and
Suvanto (1993).
2. Even in quote-driven markets, though, order ow still drives price, not the other
way around (as we shall see in the models of this chapter).

Notes

287

3. See chapter 1 for a discussion of this hypothetical Walrasian auctioneer.

4. Regarding the role of order ow in inventory models, O'Hara (1995, 16) puts it this
way: ``What is required here is that order ow be stochastic without being informative
about future market or price movements. This is the general view taken in virtually all
inventory-based microstructure models.''
5. I do not mean to suggest that inventory models are trivial or unimportant. Quite the
contraryinventory models gave microstructure nance its start (see, e.g., Garman
1976; Stoll 1978; Amihud and Mendelson 1980; Ho and Stoll 1983), and they are vital
for understanding certain issues. For understanding the issues I address in this book,
however, they need to be coupled with information models. See O'Hara (1995) for
an in-depth treatment of inventory models.
6. Price P is a sufcient statistic for the value V if the conditional density of V, given P
and a private signal Si , is independent of Si . Note that this does not imply that P
reveals V (i.e., it does not imply that the variance of V conditional on P is zero). It
means only that signal Si contributes no additional information.
7. Varian (1985) analyzes equilibria of this kind. He uses the term ``opinion'' to denote
uninformative differences in belief. One implication of being uninformative is that
when others learn opinions, via price or otherwise, this does not induce them to alter
their beliefs. See also Harris and Raviv (1993).
8. In all the models of this chapter there is also a riskless asset, but its economic role is
trivial, so I omit it from exposition. See the appendix for a discussion.
9. For expositional clarity, I use zero-mean payoff values for all the models of this
chapter. A positive mean adds nothing to the economics of the problem and can, in
fact, obscure the economics. In the appendix I include non-zero means when presenting the basic inference tools. For those concerned about the possibility that prices can
be negative, one can set the mean of fundamentals arbitrarily high, making the probability of negative prices in equilibrium arbitrarily small. For empirical work on the
distribution of (oating) exchange rate returns, see Hsieh (1988).
10. Like the process for fundamentals, it is more natural to think of the endowment
process as having a positive mean. I choose the zero-mean formulation in this case too
for expositional claritythe underlying economics is unchanged.
11. When applied in an international context, this specication implies that agents
from different countries are maximizing the utility of wealth measured in the same
numeraire, namely dollars. If instead investors in different countries care about returns
measured in their own currency, then (small) expected return differentials arise as a
result of Jensen's Inequality (sometimes referred to as Siegel's Paradox; see Siegel
1972). As an empirical matter, though, there is wide agreement within the international nance literature that these Jensen's Inequality effects play a negligible role in
exchange rate determination. Given this, I do not introduce them in this book. For
more detail, see the nice discussion in Obstfeld and Rogoff (1996), 586588.
12. Though trading is simultaneous with price determination, the uninformed trader
can calculate his demand in advance as a function of every possible price. Think of the
order he submits as being conditonal on the market-clearing price.
13. Transforming signals this wayso they are distributed about a variable of interest, in this case the signal Sis a technique I will use repeatedly in this chapter.

288

Notes

14. The Grossman and Stiglitz (1980) version of the model has a nite number of
traders i i; . . . ; n, so the assumption of perfectly competitive behavior is less of a
stretch than in the two-trader case here (but still problematic). One technique commonly used to address the issue is to assume that the informed and uninformed represent separate continuums of traders, so that no one trader has a measurable impact
on price.
15. The literature on rational expectations models of trading is vast. Papers on how
prices aggregate information include Grossman (1977); Grossman and Stiglitz (1980);
Hellwig (1980); and Diamond and Verrecchia (1981). For papers on the existence of
equilibria and their degree of revelation, see the overview in Jordan and Radner (1982).
Papers that address connections between rational expectations models and microstructure models include Hellwig (1982); Kyle (1989); and Rochet and Vila (1995). For a
multiple-period rational expectations model and its application to the home-bias puzzle in international nance, see Brennan and Cao (1997). For applications to FX markets, see Bhattacharya and Weller (1997) and Montgomery and Popper (1998).
16. Though much macro information satises the rst of these two criteria, very little
macro information satises the second. See chapter 2 for more detail on specic types
of nonpublic information.
17. In general, opportunities for protable manipulation require situations where
markets are more liquid when investors are unwinding their trades than when the
original trades are made. For models that do permit some manipulation in equilibrium, see, e.g., Allen and Gorton (1992); Lyons (1997a); and Vitale (2000).
18. For expositional clarity, I continue to use this zero-mean specicationit has no
impact on the underlying economics.
19. The original Kyle paper (1985) has this same direct observation of V, rather than a
signal of V, as in the rational expectations auction model earlier in this chapter.
Changing the Kyle specication so that the informed trader observes a signal of V is
straightforward. Note that changing the specication of the rational expectations
model so that the informed trader observes V directly would result in full revelation:
the demand of the informed trader in equation (4.3) would be either positive or negative innity at any price other than the fully revealing price.
20. Note that the demand function introduced earlier in equation (4.3) does not appear
here. That demand function is inappropriate because the informed trader does not take
price as given (i.e., he is strategic).
21. See Subrahmanyam (1991); Admati and Peiderer (1988); and Holden and Subrahmanyam (1992), respectively. For a recent application of the Kyle (1985) model in
FX, see Vitale (1999). References to other variations on the Kyle model are available in
O'Hara (1995).
22. Recall from the introduction to this chapter that the traditional focus of inventory
models is transitory variation in price around xed expected payoffs, caused by some
inventory cost (the cost typically arises from risk aversion). The focus of information
models, in contrast, is permanent price adjustment toward changing future payoffs.
23. Most of the ``single-dealer'' models in the theoretical literature actually admit
multiple dealers, but Bertrand (price) competition sufces to restrict the problem to
one dealer.

Notes

289

24. For expositional clarity I am associating these insights with the sequential-trade
model. Given the sequencing of models in this chapter this is certainly the place that
they enter. These ideas have a history, however, that predates Glosten and Milgrom
(1985), the paper that develops the sequential-trade model. See, for example, Bagehot
(1971) and Copeland and Galai (1983).
25. More formally, price follows a Martingale with respect to an information set Wt if
EPt1 j Wt Pt .
26. See Easley and O'Hara (1987, 1992); and Easley et al. (1996), respectively.
27. This is especially true when considering multiple-dealer models that capture important features of the foreign exchange market. Work on multiple-dealer theory more
generally includes Ho and Stoll (1983); Leach and Madhavan (1993); Biais (1993); Perraudin and Vitale (1996); Vogler (1997); Werner (1997); Hau (1998); and Viswanathan
and Wang (1998). Most of these papers model a centralized interdealer market rather
than a decentralized market like that in FX. Also, unlike the simultaneous-trade model,
most of these models include either fundamental private information or risk aversion,
but not both.
28. Think of the common signal S as, say, a public macro announcement. The private
signal Si allows for information advantage beyond that which will arise from observing order ow. An example often cited by dealers is that Si could capture knowledge
of customers' identity and trading motiveshedge fund trades may have more price
impact than corporate trades. I choose to model this more metaphorically with Si ,
rather than introducing identity explicitly. Evidence that customer identities do indeed
matter is provided in chapter 9.
29. It is a simple matter to add spreads to the quotes that apply to the trading between
customers and dealers (e.g., to determine the number of participating dealers endogenously). Adding spreads to the interdealer quotes is more complex.
30. An unknown m becomes intractable, however, because it generates a non-Normal
position disturbance.
31. Other features are captured in various extensions of the model, including variable order ow transparency (Lyons 1996a) and interday trading (Evans and Lyons
2000).
32. That utility is negative under the negative exponential is not problematic: utility
functions capture the ordinal ranking of outcomes, not the absolute utility attached to
any given outcome. Adding a large positive constant to all these utility values would
not change individuals' decisions.
33. This is not to suggest that wealth redistribution is irrelevant in nancial markets.
One needs to use judgment about whether the question being addressed is one for
which redistribution is likely to matter.

Chapter 5.

Empirical Frameworks

1. See Goodhart and Figliuoli (1991) and Bollerslev and Domowitz (1993), among
many others. At the daily frequency, early work includes Glassman (1987); Bossaerts
and Hillion (1991); and Wei (1994).

290

Notes

2. See Lyons (1995); Goodhart et al. (1996); Yao (1998a); and Evans (1997). For an
emerging experimental literature on markets organized like FX, see Flood et al. (1999).
For a simulation methodology, see Flood (1994). For work on information networks
embedded in FX trading technologies, see Zaheer and Zaheer (1995). For work on
emerging-market currencies, see Goldberg (1993), Goldberg and Tenorio (1997), Carrera (1998); Galati (2000); and Becker, Chadha, and Sy (2000).
3. Two Web sites are valuable for obtaining more information on the systems that
produce the data. The rst covers the Reuters dealing systems, D2000-1 and D2000-2:
hwww.reuters.com/transactions/tran00m.htmi. The second covers the EBS system
used for brokered interdealer trading: hwww.ebsp.comi.
4. Though in number, fewer than 90 percent of the world's dealers in major spot
markets use the Dealing 2000-1 system, a higher percentage of the dollar value goes
through the system because the most active dealers use the system quite intensively.
5. Speaking of ``electronic trading systems'' without rst separating direct and brokered
trading can be quite misleading in terms of the information available to dealers while
trading. The D2000-2 system, like EBS, competed with traditional voice-based brokerage, and now these systems dominate the FX brokerage business (the voice-based
brokers have been driven out). Dealing 2000-1 is the electronic means for direct trading. In terms of information dissemination, though, electronic trading in these two
different segments is very different: a communication over D2000-1 is strictly bilateral,
whereas brokered trading communicates much more information to other dealers.
6. A data set of customer orders that hit the literature just as this book was going to
print is that in Osler (2001). Her data include roughly 10,000 customer limit orders
(including stop loss orders) from a ``large'' bank from September 1999 to April 2000
(mostly $/yen and $/euro).
7. There is a large and important body of empirical work that is based on these FXFX
data. My focus in this book, however, is order ow and its effects on price. Accordingly, work based on FXFX data does not gure as prominently here as it does in the
literature. For a survey that includes much of this FXFX work, see Goodhart and
O'Hara (1997).
8. A recently introduced data setnot summarized abovethat makes progress in
integrating data from multiple market segments is that in Bjonnes and Rime (2000).
9. A more recent approach that is not on my list is the structural sequential-trade
approach pioneered by Easley et al. (1996). This approach has not yet been applied to
FX markets.
10. The cited papers provide a good deal more detail on the approach for interested
readers. For an application to the Tokyo Stock Exchange, see Hamao and Hasbrouck
(1995).
11. When this approach is applied in a limit order auction setting, order ow is measured as described in chapter 1: trades are signed according to the direction of market
orders, with limit orders being the passive side of each trade. This is tantamount to
assuming that information is conveyed by market orders, whereas liquidity-providing
limit orders convey no information. In general, in models where informed traders can
choose between market and limit orders, they choose market orders, in part because
they do not want to advertise superior information in advance of execution (see, e.g.,
Harris and Hasbrouck 1996).

Notes

291

12. The variable xt takes different denitions in different applications, and can even be
dened more generally as a vector of characteristics. One common variation on the use
of signed order ow for xt is the use of a signed trade indicator, which takes the value
of one for an incoming buy and minus one for an incoming sell. Use of this trade indicator parallels closely the second of the two empirical approaches presented here
the Trade-Indicator approach.
13. In chapter 2, I dened private information as information that (1) is not known by
all people, and (2) produces a more precise price forecast than public information
alone. By this denition, superior information about FX dealer inventories that allows
one to more accurately forecast inventory effects is private information. However, this
type of private information may not help forecast price in the long run, which precludes it from being counted as information by the denition employed in the VAR
approach (equation 5.6).
14. The estimate by Payne (1999) may in fact be an underestimate because he uses
order ow from brokered interdealer trading (versus direct interdealer ow, such as
that used by Evans and Lyons 1999). Bjonnes and Rime (2000) nd that brokered
interdealer order ow conveys less information than direct ow. Reiss and Werner
(1999) nd a similar result for trading among dealers on the London Stock Exchange.
See also Saporta (1997).
15. To be fair, Huang and Stoll (1997) do estimate their TI model for three different
categories of trade size. This amounts to modeling size via splitting one's sample,
rather than modeling size explicitly.
16. Price can of course be moving over time for other reasons. I have chosen to
hold these other reasons constant in the gure to highlight this specic source of
transaction-price variation.
17. In this case, too, prices can be moving over time for other reasons. Also, the gure
assumes that the width of the spread does not change. This is the simplest case, and it
holds in some inventory models; however, it is not a general property of inventory
models. See O'Hara (1995) for more detail along these lines.
18. In implementing the model for NYSE stocks, the indicator Qt typically takes on
values f1; 0; 1g, where 0 is assigned if the trade is executed at the posted bid-ask
midpoint. Execution at the midpoint does not occur in pure dealer markers. But the
NYSE is not a pure dealer market (see chapter 3). It is possible on the NYSE for incoming orders to be executed at the midpoint (e.g., if two limit ordersone buy and
one sellarrive simultaneously with limit prices at the posted midpoint).
19. Let me provide an example of the type of question that spread decomposition can
address (one that is relevant to FX, given increasing concentration of the market
among the largest dealing banks). The example comes from a recent paper by Weston
(2000). He asks whether competition-increasing reforms on the Nasdaq reduced the
order-processing component of the spread. (Recall that the order-processing component also includes ``rents.'') He nds that they did, suggesting that the pre-reform
spreads were less than perfectly competitive.
20. Other work using structural models includes Foster and Viswanathan (1993),
which, like Madhavan and Smidt (1991), addresses the NYSE; structural models in a
multiple-dealer setting include Snell and Tonks (1995) for stocks and Vitale (1998) for
bonds.

292

Notes

21. The model in Lyons (1995) also makes another change to accommodate FX institutions: it incorporates the fact that FX dealers use inventory control methods that are
not available to a specialist (e.g., laying off inventory at another dealer's price). To
keep things simple, I omit this second change from the version of the model presented
here.
22. Estimation of the model can also accommodate variation in desired inventory. See
Lyons (1995).
23. The moving-average error term here comes from the inference problem embedded
in the model; see Madhavan and Smidt (1991) for details. I should also add that one
might be tempted to believe that estimation of this equation using OLS would be
biased because of correlation between Xjt and eit due to correlation between St and eit
(from equation 5.21). This is not true, however: because dealer i knows both St and eit
at the time of quoting, correlation between Xjt and eit is inconsistent with quotes being
regret free (i.e., it is inconsistent with rational expectations).
24. Recent work on markets other than FX may provide an explanation why the inventory effects on the dealer's own prices are so weak: non-FX dealers appear to hedge
a lot of their inventory risk using derivatives rather than winding down the inventory
itself (Reiss and Werner 1998; Naik and Yadav 2000). FX dealers, on the other hand, do
not use derivatives much; they nd it less expensive to control inventory with actual
spot trades (see Lyons 1995).
25. The work on transaction volume in nancial markets is vast. See, for example,
Karpoff (1986, 1987) and Wang (1994).
26. We do not yet know, for example, what share of interdealer trading is due to hot
potato trading.

Chapter 6.

Exchange Rate Models

1. When I use the symbol P in equations to denote the exchange rate, it will always
represent the dollar price of other currencies (i.e., $/other).
2. See the survey by Froot and Rogoff (1995) in the Handbook for International Economics.
3. Note that, though I motivated the model with the law of one price, it is not necessary that the law of one price hold for every good. PPP might still hold if, for example,
policymakers are successfully intervening in the FX market to maintain PPP. That said,
there are many reasons why PPP does not hold in the short run (such as nontradable
goods, trade barriers, and many others).
4. The approximation is close over shorter periods of time as long as interest rates and
rates of currency appreciation are not too large. Over periods of years or over periods
of hyperination the approximation can be quite imprecise.
5. This formulation is intended to speak to readers from the eld of nance who are
more familiar with valuation based on discounted cash ows. Readers from international economics will see their more familiar formulation based on discounted monetary and real variables in the following subsection.
6. For more detail, see the surveys by Taylor (1995) and Isard (1995). In particular,
these surveys provide ample evidence of these models' lack of empirical success
(which I revisit in chapter 7).

Notes

293

7. Readers familiar with exchange rates will recognize this version of PPP as the absolute version. The relative version is expressed in rst differences in logs, or Dp$=
DpUS DpUK , where the difference operator D denotes the change from time t 1 to
time t.
8. This is the equation for the ``LM'' curve in the familiar IS-LM model. Note that here
the ``money market'' is used in the narrow, macroeconomic sense of the market for
actual money, not in the broader practitioner sense of the market for xed-income
securities with maturity of one year or less.
9. I present here only the solution that excludes rational bubbles. As is standard,
equation (6.9) has an innite number of rational expectations solutions, but the others
include a bubble component in the price path. For more on solution techniques in
rational expectations models, see Blanchard and Fischer (1989), 261266.
10. The poor empirical performance of this and other macro models of this chapter is
documented in Meese and Rogoff (1983a, b), and surveyed more recently in Frankel
and Rose (1995).
11. Strictly speaking, PPP holds in the long run only in simpler versions of this model.
It does not hold in versions with real shocks, or other shocks requiring adjustment in
the long-run real exchange rate.
12. Overshooting does not necessarily occur in versions of the model that allow output y to vary.
13. In fact, iUS < iUK throughout the adjustment period in the standard model, which
means that the dollar is expected to appreciate monotonically toward its long-run
level.
14. More precisely, as a matter of balance of payments accounting, a current account
surplus must be offset in the capital account with a net increase in domestic claims
against the foreign economy (assuming no reserve transactions on the part of central
banks).
15. Another dimension of ``tastes'' in these models is the preferences of policymakers,
e.g., over various monetary policy rules.
16. For an extensive, nontechnical overview see Stockman (1987). For more recent GE
models, which include sticky prices, see Obstfeld and Rogoff (1996).
17. The word ``general'' in the name of this model stresses that the model includes
a wide array of different markets, all of which must clear simultaneously in equilibrium (e.g., individual goods markets, labor markets, bond markets, FX markets, etc.).
18. Whether this statement is strictly true depends on the breadth of one's denition of
microstructure. One literature that links to the real economy using microstructure
models is that on insider trading. In Leland (1992), for example, insider trading can
affect a rm's real investment.
19. Recall from the previous section that I referred to the GE model's cash-in-advance
constraint as ``technology.'' In contrast to production technology, it might be more
precise to refer to the cash-in-advance constraint as an institution, albeit a rather simple one.
20. Volatility determination has traditionally attracted a lot of attention from the
macro end of the spectrum. More recently, though, the literature has begun to address

294

Notes

FX volatility from the micro end of the spectrum, which is why I list this issue in the
middle group. Recent contributions from the micro end include Wei (1994); Andersen
(1996); Andersen and Bollerslev (1998); Hau (1998); Jeanne and Rose (1999); and Cai
et al. (1999).

Chapter 7.

Macro Puzzles

1. Per earlier chapters, the relevant literature is vast. At longer horizons, e.g., longer
than two years, macro models begin to dominate the random walk (e.g., Chinn 1991;
Mark 1995). But exchange rate determination remains a puzzle at horizons less than
two years (except in cases of hyperination, in which case the ination differential
asserts itself as a driving factor, in the spirit of PPPsee chapter 6).
2. The determination puzzle exists in equity markets as wellsee Roll (1988). Roll can
account for only 20 percent of daily stock returns using traditional equity fundamentals, a result he describes as a ``signicant challenge to our science.'' The microstructure approach had not been applied directly to the determination puzzle in equity
markets when Evans and Lyons (1999) applied it in the FX market.
3. Another alternative to traditional macro modeling is the recent ``new open-economy
macro'' approach (e.g., Obstfeld and Rogoff 1995see chapter 6). I do not address this
alternative here because, as yet, the approach has not produced empirical exchange
rate equations that alter the Meese-Rogoff (1983a) conclusions.
4. If order ow is an informative measure of macro expectations, then it should forecast surprises in important variables (like interest rates). New order ow data sets
that cover up to six years of FX tradingsuch as the data set I examine in chapter
9provide enough statistical power to test this. The Evans (1997) data set used
by Evans and Lyons (1999) is only four months, so they are not able to push in this
direction.
5. Chapter 2 introduces two subcategories of discount rate information: information
about inventory effects and information about portfolio balance effects. I do not consider information about inventory effects in this chapter because inventory effects are
transitory, and are therefore unlikely to be relevant for longer horizon macro puzzles.
6. As a practical matter, however, the underlying causes of order ow X may not
captured by traditional specications of i; m; z in empirical portfolio balance models.
Note, too, that I do not consider the general equilibrium (GE) macro model in this
discussion. For the GE model, it is difcult to make meaningful statements about
whether X and i; m; z are independent when X is conveying information about portfolio balance effects: it depends on the GE model's specic features. (To clarify terms,
when X is conveying this type of information in GE models, it is conveying information about ``pricing kernels''the general-equilibrium analogue of discount rates.)
7. As in the simultaneous-trade model, introducing a bid-offer spread (or price schedule) in round one to endogenize the number of dealers is a straightforward extension.
8. Recall from chapter 4 that simultaneous and independent interdealer trades Tit
implies that dealers cannot condition on one another's trades in a given round.
9. This is tantamount to assuming thatwhen it comes to bearing overnight riskthe
dealers' capacity is small relative to the capacity of the whole public.

Notes

295

10. Note that this equation describes a cointegrating (i.e., long-run) relationship that
includes the level of price and cumulative interdealer order ow, which we revisit in
section 7.3.
11. This model can also be used to generate multiple equilibria. Introducing multiple
equilibria obscures the essential portfolio balance logic, however, so I do not pursue
this direction here.
12. In the Evans-Lyons model, these are exogenous. If one were to model them
explicitly, they could arise from any number of motives, including hedging demand,
liquidity demand, and changing risk preferences.
13. Consider an example. Starting from Xt 0, an initial customer sale to a dealer does
not move Xt from zero because Xt measures interdealer order ow only. After the
customer sale (say of one unit), then when dealer i unloads the position by selling to
another dealer, dealer j, Xt drops to 1. A subsequent sale by dealer j to another
dealer, dealer k, reduces Xt further to 2. If a customer happens to buy dealer k's
position from him, then the process comes to rest with Xt at 2. In this simple scenario,
order ow measured only from trades between customers and dealers would have
reverted to zero: the concluding customer trade offsets the initiating customer trade,
putting a stop to the hot potato. The interdealer order ow, however, does not revert
to zero.
14. Cheung and Chinn (1999b) corroborate this empirically: their surveys of foreign
exchange traders show that the importance of individual macroeconomic variables
shifts over time, but ``interest rates always appear to be important.''
15. As a diagnostic, though, Evans and Lyons also estimate the model using the level
of the differential, in the manner of Uncovered Interest Parity, and nd similar results.
16. There is a vast empirical literature that attempts to increase the explanatory power
of interest rates in exchange rate equations (by introducing individual interest rates as
separate regressors, by introducing nonlinearities, etc.). Because these efforts have not
been successful, it is very unlikely that variations on the interest rate specication
could alter the relative importance of order ow.
17. Recall from chapter 5 that one of the shortcomings of the Evans (1997) data set is
that it does not include the size of each trade, so that order ow is measured as the
number of buys minus the number of sells. (However, the data set does include the
total volume over the sample, so that an average trade size can be calculated.) This
shortcoming must be kept in perspective, however: if the Evans-Lyons results were
negative, then data concerns would be serious indeedthe negative results could
easily be due to noisy data. But their results are quite positive, which noise alone could
not produce. Indeed, that there is noise in the data only underscores the apparent
strength of the order ow/price relation.
18. These higher coefcients at higher activity levels are consistent with the ``event
uncertainty'' model described in chapter 5 (see Easley and O'Hara 1992). It should be
noted, though, that the daily frequency of the Evans-Lyons analysis is lower than the
typical transaction frequency tests of event uncertainty. Results from intraday analysis
of the Evans (1997) data paint a different picturesee chapter 8.
19. A direct role for macro announcements in determining order ow warrants exploring as well. Another possible use of macro announcements is to introduce them
directly into an Evans-Lyons-type regression. However, this tack is not likely to be

296

Notes

fruitful: there is a long literature showing that macro announcements are unable to
account for exchange rate rst moments (though they do help to account for second
momentssee Andersen and Bollerslev 1998).
20. See Osler (1998) and Carlson and Osler (2000) for models in which current account
ows are central to exchange rate determination. Unlike traditional ow approach
models, the exchange rate in these models is determined in its own speculative market.
21. I would add that this analogy also sheds light on the asset approach. In the asset
approach, too, there is no inference effect of order ow on price because all information is public, by assumption (i.e., public information is a sufcient statistic for price).
22. Contrary to popular belief, in an absolute sense, exchange rates are less volatile
than stock prices: the annual standard deviation of exchange rate returns is in the 10
12 percent range for major currencies against the dollar, whereas the annual standard
deviation of equity market returns is in the 1520 percent range (and for individual
stocks it is higher still).
23. Exchange rate regimes are not limited to oating and xed. They fall along a
spectrum. Ordered in terms of increasing commitment to the exchange rate target,
these regimes include: (1) free oat, (2) dirty oat, (3) target zone, (4) pegxed or
crawling, (5) currency board, and (6) monetary union. A dirty oat involves some
limited intervention. A currency board is an institutional commitment to dedicate
monetary policy to the exchange rate target. For more on the differences between these
regimes, see, for example, Krugman and Obstfeld (2000).
24. The transition from EMS to EMU was indisputably a transition toward exchange
rate xity. KLM assume that EMU was perfectly credible after the weekend of May 2
3, 1998the date the eleven ``in'' countries were selected and the date the internal
conversion rates for the euro-zone were determined. Extending their model to environments of imperfectly credible xed rates is a natural direction for further research.
25. The logic of this examplebased on the polar extreme of perfectly xed rates
also holds for intermediate regimes under which some volatility remains.
26. Nonstationary variables are cointegrated if there exists a linear combination of
them that is stationary. Cointegration means that although many developments can
cause permanent changes in these three series, there is an equilibrium relation that ties
them together in the long run. A good reference on cointegration is Hamilton (1994),
chapter 19.
27. EBS has a prescreened credit facility whereby dealers can only see prices for trades
that would not violate their counterparty credit limits, thereby eliminating the potential for failed deals due to these limits.
28. In their sample, the mean value of cumulative order ow is DM1.38 billion.
29. This is a theoretical point. Empirically, it appears that there was little intervention
by the national central banks or the ECB in the period from May to December 1998
(verication is difcult because these banks are not terribly forthcoming with intervention data over this period).
30. This type of bias is often referred to as ``conditional bias'' to distinguish it from
``unconditional bias.'' Conditional bias refers to the fact that to predict whether Pt1
will be lower than Ft; 1 , one needs to ``condition'' (i.e., base the prediction) on whether
Ft; 1 > Pt . Forward rates would be unconditionally biased if the average of Ft; 1 Pt1 is

Notes

297

statistically different from zero (there is no conditioning in this case on time t information). Unconditionally, major-currency forward rates tend to be unbiased (see, for
example, Perold and Schulman 1988).
31. A no-arbitrage relation called covered interest parity (introduced in an early footnote in chapter 3) requires that the forward discount equal the interest differential (as
long as there are no capital market controls that prevent arbitrage trades). This yields a
regression that is equivalent to equation (7.15) with the one-period interest differential
on the righthand side (i$ iother ) in lieu of the forward discount ft; 1 pt . Many
authors in this literature estimate equation (7.15) in this interest-differential form (in
effect, a test of uncovered interest parityintroduced in chapter 6).
32. Though I have introduced the two main economic explanations for bias
inefciency versus risk premiumthere is a third explanation that any anomaly must
confront, namely measurement error (broadly dened), i.e., that the anomaly doesn't
really exist. In the forward bias literature, the measurement error explanation is typically referred to as the ``peso problem.'' The name comes from a period through which
the Mexican peso was consistently selling at a forward discount because the market
felt devaluation was possible. When, over many years, the devaluation didn't occur,
it appeared in-sample as though the forward rate was biased, consistently overpredicting the exchange rate change. Estimates of the coefcient b are biased in this
case because the measured changes in exchange rates do not match what the market
expected (a small sample problem). Engel (1996) provides several arguments that
make the peso problem explanation difcultbut not impossibleto defend in the
context of the major exchange rates. (See Lewis 1995 for a thorough survey in which
she rightly points out that alternative hypotheses to the bias anomaly are not mutually
exclusive.)
33. The currency trading strategy here would entail selling foreign currency forward
when ft; 1 > pt and buying foreign currency forward when ft; 1 < pt . To understand
why, recall that when ft; 1 > pt , then on average pt1 ends up below ft; 1 . One can expect
to prot from locking in a sale of foreign currency at a dollar price of ft; 1 when the
expected market price (value) at t 1 is below ft; 1 . Similarly for the case when ft; 1 < pt :
one can expect to prot from having locked in a purchase of foreign currency at a
dollar price of ft; 1 when the expected market price (cost) at t 1 is above ft; 1 . (One
could trade this in an equivalent way by borrowing in the low interest rate currency
and investing in the deposit of the high interest rate currency.) Finally, by ``buy-andhold equity strategy'' I mean a simple strategy of buying and holding an equity index
fund (e.g., a fund that tracks the S&P 500 index).
34. The Sharpe ratio, named after Nobel prize winner William Sharpe, has a natural
interpretation in the standard diagram used to illustrate the Capital Asset Pricing
Model (CAPM): with expected return on the vertical axis and return standard deviation on the horizontal axis, every security (or portfolio) that lies on a ray originating
from the vertical axis at the risk-free rate has the same Sharpe ratio. For example, in a
CAPM with lending and borrowing, every point on the ray from the risk-free rate
through the tangency point on the efcient frontier has the same Sharpe ratio. In a
frictionless environment, using a Sharpe ratio criterion to select among investments is
optimal only if the investments are mutually exclusive.
35. Remember that order ow and volume are quite different: order ow is signed
volume (i.e., sell orders have a negative sign).
36. It is true that the high historical returns on U.S. equities are themselves a puzzle
(the so-called equity premium puzzle). Given those high risk-adjusted equity returns,

298

Notes

underallocation of speculative capital to forward bias is unsurprising. See also the

Sharpe ratio discussion in Backus, Gregory, and Telmer (1993), based on an earlier
sample.
37. There is scope for additional research here. One could design a survey in the spirit
of Cheung and Chinn (1999a) and Cheung and Yuk-Pang (2000) that asks decisionmakers at these institutions why they commit so little capital to trading against forward bias.
38. For more empirical evidence of the central role played by Sharpe ratios at hedge
funds and proprietary trading desks, see the leading practitioner publication from
Managed Account Reports, Inc., e.g., MAR/Hedge 2000. MAR/Hedge uses the Sharpe
ratio to rank hedge funds' performance. The ranking is provided each month.
39. Corporate nance provides many reasons why idiosyncratic risk can be value
destroying (e.g., damage to relationships with employees, customers, and suppliers,
costs of nancial distress, and so on; see the accessible survey by Bishop 1996). Formalization of my sketch of why Sharpe ratiobased contracts can be optimal need not
rely, however, on this particular agency friction. Other frictions commonly employed
in nance that produce interesting constrained-optimal decision rules include borrowing constraints, solvency constraints, and short-sale constraints. (See, for example,
He and Modest 1995.) My view from discussions with practitioners is that Sharpe
ratiodriven allocation of capital is optimal, given the institutional constraints they
face, and not an example of near-rationality, or suboptimal rules of thumb.
40. Baldwin (1990) presents what is to my knowledge the rst inaction-range model
for the forward bias. In his model, transaction costs alone generate the inaction range.
His mechanism is therefore quite different from that proposed here, in that my inaction
range is based on the evaluation of risk, whereas his is purely a function of expected
return. For a recent model of inaction ranges in international goods markets, see
Obstfeld and Taylor (1997).
41. This idea provides a direct link to the two other major puzzles addressed in this
chapter: the determination puzzle (section 7.1) and the excess volatility puzzle (section
7.3). The connecting link is the elasticity of public demand. A lower conditional return
volatility means a higher elasticity of public demand, other things equal (and vice
versa). In this section, I focus on Sharpe ratios; those previous sections focus on the
broader, more abstract notion of imperfect substitutability.
42. A big advantage of carving up public order ow this way (as opposed to, say,
chartists versus fundamentalistssee Frankel and Froot 1990) is that one can readily
identify these trade categories in empirical work. Indeed, I present estimates of models
along the lines of equation (7.19) in chapter 9.
43. As an empirical matter, just because leveraged investors do not implement pure
currency strategies does not mean they do not trade in FX markets. The FX trading of
leveraged investors is, for the most part, focused in areas where they enjoy informational advantage. Take, for example, the proprietary trading desks at major banks.
These banks have order ow information that is private, which they use to determine
speculative positions. This type of position taking need bear no particular relation to
the trades that would result from pure currency strategies.
44. See Moore and Roche (1999) for a general-equilibrium application of limited participation to the forward bias puzzle. Their model also includes habit persistence and
consumption externalities.

Notes

299

45. This particular thought experiment is essentially the same as that of Froot and
Thaler (1990, 188). Missing from their story, however, is a reason why the seemingly
large prot opportunity can persist (indeed, they describe this missing feature as an
``apparently serious aw'' in their story). My explanation does provide persistence: it
comes from institutions' Sharpe ratio based evaluation of risk. In the last sentence of
their article, Froot and Thaler point in the general direction of my approach when they
write, ``Although much of the risk in these strategies may be diversiable in principle,
more complex diversied strategies may be much more costly, unreliable, or difcult
to execute.'' See also Carlson (1998) for a gamblers' ruin model of why institutions put
tight limits on FX position taking. For a microstructure analysis of interest rate
increases at times of xed-rate crisis, see Lyons and Rose (1995).
46. As an empirical matter, this story of gradual adjustment to monetary policy
changes is borne out in the datasee Eichenbaum and Evans (1995).
47. One might wonder whether the bias from a coefcient b of 0.9 is large enough to
be consistent with an exchange rate that is misaligned by 20 percent. To understand
why it can be consistent, note that the 20 percent misalignment is in the current exchange rate level, which impounds the whole stream of future one-month bias realizations; 20 percent is not an indicator of bias over short horizons (otherwise it truly
would be an extraordinary trading opportunity).
48. Though order ow is clearly insufcient, I do not mean to suggest that the disparity is not being traded. Long Term Capital Management, for example, was one of
several institutions that selectively opened positions based on this disparity.

Chapter 8.

Microstructure and Central Bank Intervention

1. Readers interested in surveys of the macro literature on intervention should see

Dominguez and Frankel (1993b); Edison (1993); and Sarno and Taylor (2000b). Beyond
surveying the literature, Dominguez and Frankel (1993b) also contains considerable
analysis of its own, much of it based on survey expectations of exchange rate movements. As noted in chapter 1, one can think of order ow as another way to proxy for
the same expectations.
2. For evidence of imperfect substitutability across stocks, see Scholes (1972); Shleifer
(1986); Bagwell (1992); and Kaul, Mehrotra, and Morck (2000), among others. For at
least two reasons, imperfect substitutability may be more applicable to individual
currency markets than to markets in individual stocks. First, the size of the order ows
that major currency markets need to absorb are on average more than 10,000 times
those absorbed in a representative U.S. stock (the average daily volume on individual
NYSE stocks in 1998 was about $9 million, whereas the average daily volume in $/DM
spot was about $150 billion). Second, there are far more individual stocks that can
substitute for one another in portfolios than there are individual currencies (particularly major currencies).
3. I use the term ``intervention'' here to mean any trade by a central bank in the FX
market with the objective of inuencing the exchange rate.
4. Referring to the liability side of the central bank balance sheet as ``money supply'' is
a bit loose, but is valuable for expositional clarity. A more precise breakdown would
distinguish between currency in circulation and monetary base. The latter gives rise to
the money multiplier familiar from macroeconomics.

300

Notes

5. Central banks now also trade via interdealer brokers like EBS. Because brokermarket counterparties are dealing banks, the balance sheet effects are the same. One
difference that arises from trading via brokers that is not reected in balance sheets is
the degree of transparency: central bank trades done directly with a single dealer are
less transparent than central bank trades done via brokers (more on this below,
including the use of agents to protect central bank secrecy).
6. Though not essential to the story, it should be noted that foreign exchange reserves
held by central banks are not held in the form of currency, which earns no interest, but
mostly in the form of foreign government bonds. Foreign bonds have to be liquidated
to generate the yen used for payment by the Fed. Making this point explicit helps one
understand why there is not also an effect on the yen money supply from the Fed
``injecting'' $100 million worth of yen: the net effect of these transactions on the Fed's
holding of yen currency is zero.
7. Harder to understand is why the central bank would want to intervene in an
unannounced way if it intends to send a signal to the market. (I describe unannounced
intervention below.) There is no obvious rationale for this. See Dominguez and Frankel
(1993b) for a discussion.
8. Equations (6.14) and (6.15)which describe the macro portfolio balance model
provide a sense for how the two bond demands would have to adjust to accommodate
a change in supply: with xed interest rates, expected dollar depreciation, E%DP,
must change. Holding the long-run nominal value of the dollar xed, an increase in the
current value of the dollar will increase the expected rate of dollar depreciation. This
effect is akin to changing the price of a bondwhich has a xed terminal payoffin
order to change its expected return (yield).
9. I do not address the distinction between coordinated and uncoordinated central
bank intervention (i.e., coordination across central banks). For more on this topic, see
the surveys cited at the outset of this chapter.
10. Or, as is common more recently, monetary policy can be dedicated to achieving
explicit ination targets.
11. It is important to remember that order ow being near zero does not imply that
trading volume must be near zero.
12. Less clear from the context of the quoted paragraph is whether Dominguez and
Frankel consider the news effectwhich ``causes investors to revise their expectations
of future rates of return''to include only the signaling channel (i.e., payoff information), or whether it includes the portfolio balance channel (which also changes
expectations of future returns). If the latter, then their direct effect on price is probably
best thought of as representing the third information category of chapter 2: information about inventory effects.
13. This is true even in the case where intervention is announced. For announced intervention to eliminate information asymmetry, it would have to: (1) be fully credible;
(2) be fully informative regarding the size, direction, and timing of the trade; and (3)
occur in advance of, or simultaneously with, the trade. There are no ``announced''
interventions in the existing empirical literature that fulll these conditions. Moreover,
even these conditionsas strong as they areare not sufcient to eliminate an order
ow role: it would also have to be true that market participants agree on intervention's
implications.

Notes

301

14. Recall from the discussion of the Kyle model in section 4.2 that the risk neutrality
of the dealer precludes portfolio balance effects on price. Because of this narrow context, use of the term ``fundamental'' is appropriate. Regarding broader notions of fundamentals, see the discussion in section 2.3.
15. It is worth noting that, at times, the objective of central banks is to minimize the
price impact of their trades. This is the case, for example, when they are trying to
rebalance their portfolio of foreign exchange reserves, without any intention of moving
exchange rates. The intervention rules from the Evans and Lyons (2000) analysis are
useful for this objective as well.
16. For other empirical papers on intervention that use microstructure-style approaches, see Bossaerts and Hillion (1991); Goodhart and Hesse (1993); Peiers (1997);
Naranjo and Nimalendran (2000); Chang and Taylor (1998); Fischer and Zurlinden
(1999); Dominguez (1999); De Jong et al. (1999); Payne and Vitale (2000) and Chaboud
and LeBaron (2001).
17. Though indistinguishable, this is not to say that trades of these two types come
from the same distribution. Indeed, an interesting issue for future work is the extent to
which the market would adjust the mean of expected central bank orders (in the spirit
of the Lucas critique) if the central bank began to intervene secretly in a more aggressive way.
18. Recall from chapter 2 that my denition of private information is broad: information qualies as private as long as it is not known by all people and produces a better
forecast of price than public information alone. Private information does not have to be
inside information about the Fed's next interest rate move.
19. The model also claries why this learning is based on trades between dealers.
Interdealer ow is, in reality, the variable that price setters (dealers) are reacting to
when intervention is secret.
20. The model treats the central bank trade as exogenous, with expected value zero.
This might be viewed as a normalization around an ``expected'' intervention trade.
Note too that the central bank here is initiating all its trades. This is reasonable because
central banks are like other customers in this sense. In a xed rate regime, however,
where a central bank is obligated to absorb ow, the central bank will often be on the
passive side: dealers and customers initiate trades with the central bank to lay off
unwanted positions.
21. Evans and Lyons present this estimate as a ``lower bound.'' In their model, interdealer ow is equal to customer ow times a constant, a b 1. Thus, by measuring the
price impact of interdealer order ow they are underestimating the price impact of
customer tradesincluding the trades of the central bankby a factor of a. Be that as
it may, this 1-percent-per-$2-billion estimate provides some sense for why portfolio
balance effects from sterilized intervention have been so hard to detect empirically: the
average intervention trade of $200 million (Dominguez and Frankel 1993b) would
have a price impact of only about 10 basis points.
22. Because I am referring to secret intervention here, there is no way for the public to
respond to a change in central bank strategy and fully undo it. That said, one would
expect the public's trading strategiesand dealer's pricing rulesto be affected by a
changed intervention strategy. This is an interesting avenue for future theoretical work
within the microstructure approach.

302

Chapter 9.

Notes

Customers

1. By analogy, it is not unreasonable to view rms that trade on the NYSE as fundamentally the same as rms that trade on the NASDAQ, other things equal (i.e., similar
cost of capital, similar relative valuation, etc.). The distinction between them boils
down to what I called in chapter 3 the ``microstructure effects'' question. Per earlier
chapters, the microstructure effects question is not the focus of this book.
2. Recall, too, the discussion and quotations in chapter 3 regarding customer order
ows.
3. For practitioner-oriented research on order ow effects on exchange rates, see, e.g.,
Citibank's Citiows Global Flow and Volume Analysis (various issues); Deutschebank's
Flowmetrics Monthly (various issues); and Lehman Brothers' Global Economic Research
Series, particularly the issue on ``FX Impact of Cross-Border M&A.'' For evidence from
practitioner surveys on the use of ow analysis, see Gehrig and Menkhoff (2000).
4. Balance of payments accounts are based on the concept of residency. They are a
statistical record of the economic transactions taking place between a nation's residents and the rest of the world.
5. Recall that the Evans-Lyons model of chapter 8 was designed (in part) to show that
even with short-lived dealer positions, interdealer order ow still has persistent effects
on price because it conveys information about the persistent portfolio shifts of underlying customers.
6. One could argue that in frictionless general equilibrium, starting from paretooptimal allocations, it is not clear why rms' ex ante ``portfolios'' are not instantaneously restored. As an empirical matter, this objection to my example is not so
compelling. For many institution types, there are substantial (labor intensive) costs of
adjusting their net positions in the market. This may produce path dependence in
portfolio allocation, even if the relation between realized order ow and price is
unique.
7. A future direction for research in this area is to isolate categories of trades that t
neatly into a particular balance of payments category. For example, one could isolate
equity mutual funds. In this case, one could be condent that their FX trades t neatly
into the category called international portfolio investment.
8. A natural question is where the trades of central banks appear. The source of these
data is reluctant to disclose the specics. Though not reported in the table, the source
bank does maintain a small fourth category of customer, called ``miscellaneous.''
Though the trades within this category are quite small relative to the trades in the
three main categories, the category is likely to include any central bank trades for
which the source bank was the counterparty. (Recall from chapter 8 that central bank
trades tend to be small relative to private trades.)
9. The yen plot begins in January 1996 because the source bank did not include customer ow data from its Tokyo ofce in its database until late 1995. (The Tokyo ofce
is especially important for this bank's $/yen customer ow.) Note that this may account for the seemingly small share of non-nancial corporate trading in total customer trading in $/yen shown in table 9.1: If nonnancial corporations tend to trade
via their regional ofce, whereas nancial institutions tend to trade on a 24-hour basis
worldwide, then the customer trades in the database before 1996 would be tilted
toward the nancial institutions.

Notes

303

10. Integrating macro variables into this monthly regression is an important topic for
future work.
11. Because these customer ow data are not available publicly, forecasting power is
not ruled out on the basis of market efciency (e.g., on the basis of semi-strong form
efciency, under which price impounds all publicly available information). Note too
that differential price impact leaves room for the (theoretical) possibility of market
manipulation by customers who can disguise the source of trades. For a nice treatment
of traditional forecasting techniques for FX markets, see Rosenberg (1996).
12. There is a comprehensive description of events in BIS (1999b).

Chapter 10.

Looking Forward

1. Recall from chapter 1 that my focus in this book is on analysis based on order ow.
There is a lot of ne work on high-frequency exchange rates that does not integrate
order ow. For more on this branch of the literature, see the survey by Goodhart and
O'Hara (1997).
2. Resolving these differences may lie in the fact that non-FX marketmakers hedge inventory risk with instruments other than those in which they make the market (e.g.,
with related derivatives), whereas spot FX dealers nd that inventory control using
spot currencies alone is least expensive. See Reiss and Werner (1998) and Naik and
Yadav (2000) for evidence that non-FX marketmakers do indeed use derivatives for
inventory control.
3. For the NYSE, a possible resolution of these differences lies in the obligation of the
NYSE specialist to smooth prices, a task that existing inventory may facilitate.
4. If, in the end, the portfolio balance view of order ow's price effects wins out, a
subsidiary question remains: what is driving the portfolio shifts? Possibilities (mentioned in earlier chapters) include changing risk preferences, changing hedging
demands, changing liquidity demands, and changing (payoff unrelated) opinions.
5. Part of closing the gap with macro analysis will come from theoretical advances. A
natural direction along these lines is models of market incompleteness (of various
types), which t naturally with current empirical results on order ow. For a contact
point with recent work on dynamic open-economy models, see Duarte and Stockman
(2001).
6. A precondition for integrating data is availability. At present, much of the data
used in this literature is not publicly available. (As noted in chapter 5, my Web site
provides links to data sets that are publicly available.) Moving forward, one hopes that
sources of these data will recognize the value of making them available to researchers
generally.
7. Interested readers should see the material on FXNeta system for managing settlement riskon the EBS Web site (hwww.ebsp.comi).
8. One paper that addresses speculative attacks in Mexico using a microstructure
approach is Carrera (1999).
9. For an interesting article on the advent of electronic trading in FX, see Euromoney
(2000). For equity markets, Institutional Investor (2000) is a nice treatment of the electronic trading threat to more traditional trading methods.

304

Notes

10. Frankel (1996, 62) was perhaps the rst to write about such a scenario in FX,
though he was considering a Tobin tax as the possible trigger. He wrote: ``It is possible
that the imposition of a Tobin tax . . . would alter the structure of the market in a fundamental way. It might become more like other major nancial markets, in which a
sale or purchase by a customer generates only one or two transactions, rather than ve
or eight. This would be the case particularly if such a tax triggered a transition to a
new trading structure equilibrium, with the decentralized dealer network . . . replaced
by a system in which foreign currency was traded on a centralized exchange in the
manner of the NYSE.'' For more on the legal history relating to openness of broker
systems to nondealers, see Levich (2001, 100).
11. For a theoretical treatment of whether centralized limit order structures are likely
to capture liquidity and thereby dominate trading, see Glosten (1994). An issue not
addressed in that paper that is important for FX is credit risk. I noted in the previous
section that bank dealers may have a comparative advantage in managing the credit
risk arising from large transactions with customers. New entrants who want to centralize this market around an electronic trading platform need to solve this problem
because non-nancial corporations do not want to take the counterparty credit risk
that banks are comfortable taking. The standard approach is to establish a clearing
house system with margin accounts (akin to those used in futures markets).

References

Admati, A., and P. Peiderer. 1988. A theory of intraday patterns: Volume and price
variability. Review of Financial Studies 1: 340.
Allen, F., and G. Gorton. 1992. Stock price manipulation, market microstructure and
asymmetric information. European Economic Review 36: 624630.
Amihud, Y., and H. Mendelson. 1980. Dealership markets: Marketmaking with inventory. Journal of Financial Economics 8: 3153.
Ammer, J., and Brunner, A. 1997. Are banks market timers or market makers?
Explaining foreign exchange trading prots. Journal of International Financial Markets,
Institutions & Money 7: 4360.
Andersen, T. 1996. Return volatility and trading volume: An information ow interpretation of stochastic volatility. Journal of Finance 51: 169204.
Andersen, T., and T. Bollerslev. 1997. Heterogeneous information arrivals and return
volatility dynamics: Uncovering the long-run in high frequency returns. Journal of
Finance 52: 9751006.
Andersen, T., and T. Bollerslev. 1998. Deutsche mark-dollar volatility: Intraday activity patterns, macroeconomic announcements, and longer run dependencies. Journal of
Finance 53: 219266.
Andersen, T., T. Bollerslev, and A. Das. 2001. Variance-ratio statistics and highfrequency data: Testing for changes in intraday volatility patterns. Journal of Finance
56: 305328.
Backus, D., A. Gregory, and C. Telmer. 1993. Accounting for forward rates in markets
for foreign currency. Journal of Finance 48: 18871908.
Bagehot, W. 1971. The only game in town. Financial Analysts Journal 27: 1222.
Bagwell, L. 1992. Dutch auction repurchases: An analysis of shareholder heterogeneity. Journal of Finance 47: 71105.
Baillie, R., and T. Bollerslev. 1991. Intra-day and inter-market volatility in foreign exchange rates. Review of Economic Studies 58: 565585.
Baldwin, R. 1990. Reinterpreting the failure of foreign exchange market efciency tests:
Small transactions costs, large hysteresis bands. NBER Working Paper 3319.

306

References

Bank for International Settlements. 1996. Central bank survey of foreign exchange
market activity in April 1995. Publication of the Monetary and Economics Department,
BIS, March.
Bank for International Settlements. 1999a. Central bank survey of foreign exchange
market activity in April 1998. Publication of the Monetary and Economics Department,
BIS, May (available at www.bis.org).
Bank for International Settlements. 1999b. A review of nancial market events in
autumn 1998. Publication of the Committee on the Global Financial System, October.
Bank for International Settlements. 2001. BIS 71st Annual Report, June (available at
www.bis.org).
Bank for International Settlements. 2002. Central bank survey of foreign exchange
market activity in April 2001. Publication of the Monetary and Economics Department,
BIS, forthcoming (available at www.bis.org).
Bank of England. 1998. The UK foreign exchange and over-the-counter derivatives
markets in April 1998, September (available at www.bankofengland.co.uk).
Becker, T., B. Chadha, and A. Sy. 2000. Foreign exchange bid-ask spreads in the Asian
crisis. Typescript, International Monetary Fund, April.
Bessembinder, H. 1994. Bid-ask spreads in the interbank foreign exchange market.
Journal of Financial Economics 35: 317348.
Bhattacharya, U., and P. Weller. 1997. The advantage of hiding one's hand: Speculation and central bank intervention in the foreign exchange market. Journal of Monetary
Economics 39: 251277.
Biais, B. 1993. Price formation and equilibrium liquidity in fragmented and centralized
markets. Journal of Finance 48: 157184.
Biais, B., P. Hillion, and C. Spatt. 1995. An empirical analysis of the limit order book
and the order ow in the Paris bourse. Journal of Finance 50: 16551689.
Biais, B., P. Hillion, and C. Spatt. 1999. Price discovery and learning during the preopening period in the Paris Bourse. Journal of Political Economy 107: 12181248.
Bishop, M. 1996. Corporate risk management. The Economist, survey October 2, 1996.
Bjonnes, G., and D. Rime. 2000. FX trading . . . live: Dealer behavior and trading systems in foreign exchange markets. Typescript, Norwegian School of Management,
August.
Blanchard, O., and S. Fischer. 1989. Lectures on Macroeconomics. Cambridge, MA: MIT
Press.
Blume, L., D. Easley, and M. O'Hara. 1994. Market statistics and technical analysis:
The role of volume. Journal of Finance 49: 153182.
Board, J., and C. Sutcliffe. 1995. The effects of trade transparency in the London stock
exchange. Report commissioned by the London Stock Exchange, January.
Bollerslev, T., and I. Domowitz. 1993. Trading patterns and prices in the interbank
foreign exchange market. Journal of Finance 48: 14211443.

References

307

Bollerslev, T., and M. Melvin. 1994. Bid-ask spreads and volatility in the foreign
exchange market: An empirical analysis. Journal of International Economics 36: 355
372.
Bossaerts, P., and P. Hillion. 1991. Market microstructure effects of government intervention in the foreign exchange market. Review of Financial Studies 4: 513541.
Branson, W., and D. Henderson. 1985. The specication and inuence of asset markets. In Handbook of International Economics, vol. 2, edited by R. Jones and P. Kenen.
Amsterdam: North-Holland.
Brennan, M., and H. Cao. 1996. Information, trade, and derivative securities. Review of
Financial Studies 9: 163208.
Brennan, M., and H. Cao. 1997. International portfolio investment ows. Journal of
Finance 52: 18511880.
Brock, W., and A. Kleidon. 1992. Periodic market closure and trading volume: A model
of intraday bids and asks. Journal of Economic Dynamics and Control 16: 451489.
Burnham, J. 1991. Current structure and recent developments in foreign exchange
markets. In Recent Developments in International Banking and Finance, edited by
S. Khoury. Amsterdam: Elsevier North Holland.
Caballero, R., E. Engel, and J. Haltiwanger. 1997. Aggregate employment dynamics:
Building from microeconomics. American Economic Review 87: 115137.
Cai, J., Y. Cheung, R. Lee, and M. Melvin. 1999. ``Once in a generation'' yen volatility in
1998: Fundamentals, intervention, or order ow? Typescript, Arizona State University,
July.
Calvo, G. 1999. Contagion in emerging markets: When Wall Street is a carrier. Working
paper, University of Maryland.
Cao, H. 1999. The effect of derivative assets on information acquisition and price
behavior in a rational expectations equilibrium. Review of Financial Studies 12: 131
164.
Cao, H., and R. Lyons. 1998. Inventory information. Working paper, Haas School of
Business, U.C. Berkeley, August.
Carlson, J., 1998. Risk aversion, foreign exchange speculation and gambler's ruin.
Economica 65: 441453.
Carlson, J., and C. Osler. 1999. Determinants of currency risk premiums. Typescript,
New York Federal Reserve Bank, February.
Carlson, J., and C. Osler. 2000. Rational speculators and exchange rate volatility.
European Economic Review 44: 231253.
Carrera, J. 1999. Speculative attacks to currency target zones: A market microstructure
approach. Journal of Empirical Finance 6: 555582.
Chaboud, A., and B. LeBaron. 2001. Foreign exchange trading volume and Federal
Reserve intervention. Journal of Futures Markets, forthcoming.
Chakrabarti, R. 2000. Just another day in the interbank foreign exchange market. Journal of Financial Economics 56: 2964.

308

References

Chakravarty, S., and C. Holden. 1995. An integrated model of market and limit orders.
Journal of Financial Intermediation 4: 213241.
Chang, Y., and S. Taylor. 1998. Intraday effects of foreign exchange intervention by the
Bank of Japan. Journal of International Money and Finance 18: 191210.
Chen, Z. 1997. Speculative market structure and the collapse of an exchange rate
mechanism. Working paper, London School of Economics.
Cheung, Y., and M. Chinn. 1999a. Traders, market microstructure, and exchange rate
dynamics. NBER Working paper 7416, November. Journal of International Money and
Finance.
Cheung, Y., and M. Chinn. 1999b. Macroeconomic implications of the beliefs and
behavior of foreign exchange traders. NBER Working paper 7417, November.
Cheung, Y., M. Chinn, and I. Marsh. 1999. How do UK-based foreign exchange dealers
think their market operates? CEPR Discussion Paper 2230, September.
Cheung, Y., and C. Yuk-Pang. 1999. Foreign exchange traders in Hong Kong, Tokyo,
and Singapore: A survey study. U.C. Santa Cruz Department of Economics Working
Paper a425, January.
Cheung, Y., and C. Yuk-Pang. 2000. A survey of market practitioners' views on
exchange rate dynamics. Journal of International Economics 51: 401423.
Chinn, M. 1991. Some linear and non-linear thoughts on exchange rates. Journal of
International Money and Finance 10: 214230.
Chinn, M., and R. Meese. 1994. Banking on currency forecasts. Journal of International
Economics 38: 161178.
Citibank. 2001. Citiows Global Flow and Volume Analysis (various issues). London:
Citibank, NA (email: [email protected]).
Clyde, W. 1993. Intraday foreign exchange market anomalies and microstructure.
Working paper, Quinnipiac College, August.
Copeland, T., and D. Galai. 1983. Information effects and the bid-ask spread. Journal of
Finance 38: 14571469.
Corsetti, G., S. Morris, and H. Shin. 1999. Does one Soros make a difference? The role
of a large trader in currency crises. Typescript, Yale University, March.
Covrig, V., and M. Melvin. 1998. Asymmetric information and price discovery in
the FX market: Does Tokyo know more about the yen? Typescript, Arizona State
University.
Dacorogna, M., U. Muller, R. Nagler, R. Olsen, and O. Pictet. 1993. A geographical
model for the daily and weekly seasonal volatility in the FX market. Journal of International Money and Finance 12: 413438.
Danielsson, J., and R. Payne. 1999. Real trading patterns and prices in spot foreign
exchange markets. Typescript, London School of Economics, February.
Degennaro, R., and R. Shrieves. 1997. Public information releases, private information
arrival, and volatility in the foreign exchange market. Journal of Empirical Finance 4:
295315.

References

309

De Jong, F., R. Mahieu, P. Schotman, and I. van Leeuwen. 1999. Price discovery on
foreign exchange markets with differentially informed traders. CEPR Discussion Paper
No. 2296, November.
De Jong, F., T. Nijman, and A. Roell. 1996. Price effects of trading and components of
the bid-ask spread on the Paris bourse. Journal of Empirical Finance 3: 193213.
Demsetz, H. 1968. The cost of transacting. Quarterly Journal of Economics 82: 3353.
Deutschebank. 2001. Flowmetrics Monthly (various issues). London: Deutschebank AG
London (email: [email protected]).
Devereux, M., and C. Engel. 1999. The optimal choice of exchange-rate regime: Pricesetting rules and internationalized production. NBER Working Paper 6992.
Diamond, D., and R. Verrecchia. 1981. Information Aggregation in a Noisy Rational
Expectations Economy. Journal of Financial Economics 9: 221235.
Dominguez, K. 1986. Are foreign exchange forecasts rational? New evidence from
survey data. Economic Letters 21: 277281.
Dominguez, K. 1990. Market responses to coordinated central bank intervention.
Carnegie-Rochester Series on Public Policy, vol. 32.
Dominguez, K. 1999. The market microstructure of central bank intervention. NBER
Working Paper 7337, September.
Dominguez, K., and J. Frankel. 1993a. Does foreign exchange intervention matter? The
portfolio effect. American Economic Review 83: 13561369.
Dominguez, K., and J. Frankel. 1993b. Does Foreign-exchange Intervention Work? Washington, D.C.: Institute for International Economics.
Domowitz, I. 1990. The mechanics of automated trade execution. Journal of Financial
Intermediation 1: 167194.
Dooley, M., and P. Isard. 1982. A portfolio-balance rational-expectations model of the
dollar-mark rate. Journal of International Economics 12: 257276.
Dornbusch, R. 1976. Expectations and exchange rate dynamics. Journal of Political
Economy 84: 11611176.
Duarte, M., and A. Stockman. 2001. Rational speculation and exchange rates. Typescript, University of Rochester, March.
Dumas, B. 1996. Comment on Jorion. In The Microstructure of Foreign Exchange Markets,
edited by J. Frankel et al. Chicago: University of Chicago Press.
Easley, D., S. Hvidkjaer, and M. O'Hara. 1999. Is information risk a determinant of
asset returns? Typescript, Cornell University, November.
Easley, D., and M. O'Hara. 1987. Price, trade size, and information in securities markets. Journal of Financial Economics 19: 6990.
Easley, D., and M. O'Hara. 1992. Time and the process of security price adjustment.
Journal of Finance 47: 577606.
Easley, D., N. Kiefer, M. O'Hara, and J. Paperman. 1996. Liquidity, information, and
infrequently traded stocks. Journal of Finance 51: 14051436.

310

References

Edison, H. 1993. The effectiveness of central-bank intervention: A survey of the literature after 1982. Special Papers in International Economics, no. 18, Princeton University.
Eichenbaum, M., and C. Evans. 1995. Some empirical evidence on the effects of shocks
to monetary policy on exchange rates. Quarterly Journal of Economics 110: 9751009.
Engel, C. 1996. The forward discount anomaly and the risk premium: A survey of
recent evidence. Journal of Empirical Finance 3: 123192.
Engle, R., and C. Granger. 1987. Cointegration and error correction: Representation,
estimation and testing. Econometrica 55: 251276.
Engle, R., and J. Russell. 1998. Autoregressive conditional duration: A new model for
irregularly spaced transaction data. Econometrica 66: 11271162.
Euromoney. 2000. A true exchange for forex. July.
Evans, G. 1986. A test for speculative bubbles in the sterling-dollar exchange rate.
American Economic Review 76: 621636.
Evans, M. 1997. The microstructure of foreign exchange dynamics. Typescript, Georgetown University, November.
Evans, M. 2001. FX Trading and Exchange Rate Dynamics. NBER Working Paper 8116,
February.
Evans, M., and R. Lyons. 1999. Order ow and exchange rate dynamics. Typescript,
U.C. Berkeley (available at www.haas.berkeley.edu/@lyons). Forthcoming in Journal of
Political Economy.
Evans, M., and R. Lyons. 2000. The price impact of currency trades: Implications for
Secret Intervention. Typescript, U.C. Berkeley, June, presented at the NBER Summer
Institute, July.
Evans, M., and R. Lyons. 2001. Why order ow explains exchange rates. Typescript,
U.C. Berkeley, March.
Fan, M., and R. Lyons. 2000. Customer-dealer trading in the foreign exchange market.
Typescript, U.C. Berkeley, July.
Federal Reserve Bank of New York. 1998. Foreign exchange and interest rate derivatives markets survey: Turnover in the United States (available at www.ny.frb.org).
Fieleke, N. 1981. Foreign-currency positioning by U.S. rms: Some new evidence. Review of Economics and Statistics 63: 3543.
Financial Times. 1998. Survey: Foreign exchange, June 5.
Fischer, J., and M. Zurlinden. 1999. Exchange rate effects of central bank interventions:
An analysis of transaction prices. Economic Journal 109: 662676.
Fleming, M., and E. Remolona. 1999. Price formation and liquidity in the U.S. treasury
market: The response to public information. Journal of Finance 54: 19011916.
Flood, M. 1991. Microstructure theory and the foreign exchange market. Review, Federal Reserve Bank of St. Louis, 73 (November/December): 5270.
Flood, M. 1994. Market structure and inefciency in the foreign exchange market.
Journal of International Money and Finance 13: 131158.

References

311

Flood, M., R. Huisman, K. Koedijk, and R. Mahieu. 1999. Quote disclosure and price
discovery in multiple-dealer nancial markets. Review of Financial Studies 12: 3760.
Flood, R., and R. Hodrick. 1990. On testing for speculative bubbles. Journal of Economic
Perspectives 4: 85101.
Flood, R., and A. Rose. 1995. Fixing exchange rates: A virtual quest for fundamentals.
Journal of Monetary Economics 36: 337.
Flood, R., and A. Rose. 1996. Fixes: Of the forward discount puzzle. Review of Economics and Statistics 78: 748752.
Flood, R., and M. Taylor. 1996. Exchange rate economics: What's wrong with the conventional macro approach? In The Microstructure of Foreign Exchange Markets, edited by
J. Frankel, G. Galli, and A. Giovannini. Chicago: University of Chicago Press.
Foster, D., and S. Viswanathan. 1990. A theory of interday variations in volumes, variances, and trading costs in securities markets. Review of Financial Studies 3: 593624.
Foster, D., and S. Viswanathan. 1993. Variations in trading volume, return volatility,
and trading costs: Evidence on recent price formation models. Journal of Finance 48:
187211.
Frankel, J. 1982. A test of perfect substitutability in the foreign exchange market.
Southern Economic Journal 46: 406416.
Frankel, J. 1996. How well do markets work: Might a Tobin tax help? In The Tobin Tax:
Coping with Financial Volatility, edited by M. ul Haq et al. New York: Oxford University
Press.
Frankel, J., and C. Engel. 1984. Do asset demand functions optimize over the mean and
variance of real returns? A six-currency test. Journal of International Economics 17: 309
323.
Frankel, J., and K. Froot. 1987. Using survey data to test standard propositions regarding exchange rate expectations. American Economic Review 77: 133153.
Frankel, J., and K. Froot. 1990. Chartists, fundamentalists, and trading in the foreign
exchange market. American Economic Review 80: 181185.
Frankel, J., G. Galli, and A. Giovannini, eds. 1996. The Microstructure of Foreign Exchange Markets. Chicago: University of Chicago Press.
Frankel, J., G. Galli, and A. Giovannini. 1996. Introduction. In The Microstructure of
Foreign Exchange Markets, edited by J. Frankel, G. Galli, and A. Giovannini. Chicago:
University of Chicago Press.
Frankel, J., and A. Rose. 1995. Empirical research on nominal exchange rates. In
Handbook of International Economics, edited by G. Grossman and K. Rogoff. Amsterdam:
Elsevier Science.
French, K., and R. Roll. 1986. Stock return variance: The arrival of information and the
reaction of traders. Journal of Financial Economics 17: 99117.
Frenkel, J. 1976. A monetary approach to the exchange rate: Doctrinal aspects and
empirical evidence. Scandinavian Journal of Economics 78: 200224.
Froot, K., and E. Dabora. 1999. How are stock prices affected by the location of trade?
Journal of Financial Economics 53: 189216.

312

References

Froot, K., and K. Rogoff. 1995. Perspectives on PPP and long-run real exchange rates.
In Handbook of International Economics, edited by G. Grossman and K. Rogoff. Amsterdam: Elsevier Science.
Froot, K., and R. Thaler. 1990. Anomalies: Foreign exchange. Journal of Economic Perspectives 4: 179192.
Galati, G. 2000. Trading volumes, volatility, and spreads in FX markets: Evidence from
emerging market countries. Typescript, Bank for International Settlements, July.
Garman, M. 1976. Market microstructure. Journal of Financial Economics 3: 257275.
Gehrig, T., and L. Menkhoff. 2000. The use of ow analysis in foreign exchange: Exploratory evidence. Typescript, University of Freiburg, Germany, August.
Gennotte, G., and H. Leland. 1990. Market liquidity, hedges, and crashes. American
Economic Review 80: 9991021.
Gersbach, H., and K. Vogler. 1998. Learning from informed and uninformed traders
and the role of interdealer trading. Typescript, University of Heidelberg, March.
Glassman, D. 1987. Exchange rate risk and transactions costs: Evidence from bid-ask
spreads. Journal of International Money and Finance 6: 479490.
Glosten, L. 1989. Insider trading, liquidity, and the role of the monopolist specialist.
Journal of Business 62: 211236.
Glosten, L. 1994. Is the electronic open limit order book inevitable? Journal of Finance
49: 11271162.
Glosten, L., and L. Harris. 1988. Estimating the components of the bid-ask spread.
Journal of Financial Economics 21: 123142.
Glosten, L., and P. Milgrom. 1985. Bid, ask, and transaction prices in a specialist market with heterogeneously informed agents. Journal of Financial Economics 14: 71100.
Goldberg, L. 1993. Foreign exchange markets in Russia: Understanding the reforms.
International Monetary Fund Staff Papers 40: 852864.
Goldberg, L., and R. Tenorio. 1997. Strategic trading in a two-sided foreign exchange
auction. Journal of International Economics 42: 299326.
Goldstein, M., and K. Kavajecz. 2000. Liquidity provision during circuit breakers and
extreme market movements. Typescript, Babson College, November.
Goodhart, C. 1988. The foreign exchange market: A random walk with a dragging
anchor. Economica 55: 437460.
Goodhart, C., and L. Figliuoli. 1991. Every minute counts in nancial markets. Journal
of International Money and Finance 10: 2352.
Goodhart, C., and T. Hesse. 1993. Central bank forex intervention assessed in continuous time. Journal of International Money and Finance 12: 368389.
Goodhart, C., and M. O'Hara. 1997. High frequency data in nancial markets: Issues
and applications. Journal of Empirical Finance 4: 73114.
Goodhart, C., and R. Payne. 1996. Micro structural dynamics in a foreign exchange
electronic broking system. Journal of International Money and Finance 15: 829852.

References

313

Goodhart, C., and R. Payne. 1999. The foreign exchange market: A multimedia representation, compact disk, distributed by Enterprise LSE Ltd. (www.enterprise-lse.
co.uk).
Goodhart, C., and R. Payne, eds. 2000. The Foreign Exchange Market: Empirical Studies
with High-Frequency Data. New York: St. Martin's Press.
Goodhart, C., and M. Taylor. 1992. Why don't individuals speculate in the forward
foreign exchange market? Scottish Journal of Political Economy 39: 113.
Goodhart, C., T. Ito, and R. Payne. 1996. One day in June 1993: A study of the working
of the Reuters 2000-2 electronic foreign exchange trading system. In The Microstructure
of Foreign Exchange Markets, edited by J. Frankel, G. Galli, and A. Giovannini. Chicago:
University of Chicago Press.
Gordon, R., and L. Bovenberg. 1996. Why is capital so immobile internationally? Possible explanations and implications for capital income taxation. American Economic Review 86: 10571075.
Grammatikos, T., and A. Saunders. 1986. Futures price variability: A test of maturity
and volume effects. Journal of Business 59: 319330.
Grossman, S. 1977. The existence of futures markets, noisy rational expectations and
informational externalities. Review of Economic Studies 44: 431449.
Grossman, S. 1988. An analysis of the implications for stock and futures price volatility
of program trading and dynamic hedging strategies. Journal of Business 61: 275298.
Grossman, S., and M. Miller. 1988. Liquidity and market structure. Journal of Finance
43: 617633.
Grossman. S., and J. Stiglitz. 1980. On the impossibility of informationally efcient
markets. American Economic Review 70: 393408.
Grossman, S., and L. Weiss. 1983. A transactions-based model of the monetary transmission mechanism. American Economic Review 73: 871880.
Grundy, B., and M. McNichols. 1989. Trade and the revelation of information through
prices and direct disclosure. Review of Financial Studies 2: 495526.
Guillaume, D., M. Dacorogna, R. Dave, U. Muller, R. Olsen, and O. Pictet. 1995. From
the bird's eye to the microscope: A survey of new stylized facts of the intradaily foreign exchange markets. Finance and Stochastics 1: 95129.
Habermeier, K., and A. Kirilenko. 2000. Securities transaction taxes and nancial
markets. Typescript, International Monetary Fund, June.
Hamao, Y., and J. Hasbrouck. 1995. Securities trading in the absence of dealers: Trades
and quotes on the Tokyo stock exchange. Review of Financial Studies 8: 849878.
Hamilton, J. 1994. Time Series Analysis. Princeton, N.J.: Princeton University Press.
Handa, P., and R. Schwartz. 1996. Limit order trading. Journal of Finance 51: 1835
1861.
Hansch, O., N. Naik, and S. Viswanathan. 1998. Do inventories matter in dealership
markets? Evidence from the London Stock Exchange. Journal of Finance 53: 1623
1656.

314

References

Hansch, O., N. Naik, and S. Viswanathan. 1999. Preferencing, internalization, best execution, and dealer prots. Journal of Finance 54: 17991828.
Harris, L. 1986. A transaction data study of weekly and intradaily patterns in stock
returns. Journal of Financial Economics 16: 99117.
Harris, L., and J. Hasbrouck. 1996. Market versus limit orders: The SuperDOT evidence on order submission strategy. Journal of Financial & Quantitative Analysis 31:
213231.
Harris, M., and A. Raviv. 1993. Differences of opinion make a horse race. Review of
Financial Studies 6: 473506.
Hartmann, P. 1998a. Currency Competition and Foreign Exchange Markets: The dollar, the
Yen, and the Euro. Cambridge: Cambridge University Press.
Hartmann, P. 1998b. Do Reuters spreads reect currencies' differences in global trading activity? Journal of International Money and Finance 17: 757784.
Hartmann, P. 1999. Trading volumes and transaction costs in the foreign exchange
market: Evidence from daily dollar-yen spot data. Journal of Banking and Finance 23:
801824.
Hasbrouck, J. 1988. Trades, quotes, inventories, and information. Journal of Financial
Economics 22: 229252.
Hasbrouck, J. 1991a. Measuring the information content of stock trades. Journal of
Finance 46: 179207.
Hasbrouck, J. 1991b. The summary informativeness of stock trades: An econometric
analysis. Review of Financial Studies 4: 571595.
Hasbrouck, J. 1992. Using the TORQ database. Typescript, New York University
(available at jhasbrouck.stern.nyu.edu/main).
Hasbrouck, J. 1997. One security, many markets. Typescript, New York Univeristy.
Hasbrouck, J., and G. Soanos. 1993. The trades of market makers: An empirical
analysis of NYSE specialists. Journal of Finance 48: 15651593.
Hau, H. 1998. Competitive entry and endogenous risk in the foreign exchange market.
Review of Financial Studies 11: 757788.
Hau, H. 2000. Information and geography: Evidence from the German stock market.
Typescript, Insead, March.
Hau, H., and A. Chevallier. 2000. Evidence on the volatility effect of a Tobin tax.
Typescript, Insead, April.
Hau, H., W. Killeen, and M. Moore. 2000. The euro as an international currency:
Explaining the rst evidence. Typescript, Insead, April.
Hausman, J., A. Lo, and C. MacKinlay. 1992. An ordered probit analysis of transaction
stock prices. Journal of Financial Economics 31: 319379.
Hayek, F. 1945. The use of knowledge in society. American Economic Review 35: 519
530.

References

315

He, H., and D. Modest. 1995. Market frictions and consumption-based asset pricing.
Journal of Political Economy 103: 94117.
Heere, E. 1999. Microstructure theory applied to the foreign exchange market. Ph.D.
dissertation, Maastricht University, Faculty of Economics and Business Administration.
Hellwig, M. 1980. On the aggregation of information in competitive markets. Journal of
Economic Theory 22: 447498.
Hellwig, M. 1982. Rational expectations equilibrium with conditioning on past prices:
A mean variance example. Journal of Economic Theory 26: 279312.
Hicks, J. 1939. Value and Capital. Oxford: Clarendon Press.
Ho, T., and H. Stoll. 1983. The dynamics of dealer markets under competition. Journal
of Finance 38: 10531074.
Hodrick, R. 1987. The Empirical Evidence on the Efciency of Forward and Futures Foreign
Exchange Markets. New York: Harwood Academic Publishers.
Holden, C., and A. Subrahmanyam. 1992. Long-lived private information and imperfect competition. Journal of Finance 47: 247270.
Hong, H., and J. Wang. 1995. Trading and Returns Under Periodic Market Closures.
Working paper, MIT, July.
Houthakker, H. 1957. Can speculators forecast prices? Review of Economic Studies 39:
143151.
Hsieh, D. 1988. The statistical properties of daily foreign exchange rates: 19741983.
Journal of International Economics 24: 129145.
Hsieh, D., and A. Kleidon. 1996. Bid-ask spreads in foreign exchange markets: Implications for models of asymmetric information. In The Microstructure of Foreign Exchange
Markets, edited by J. Frankel et al. Chicago: University of Chicago Press.
Huang, R., and R. Masulis. 1999. FX spreads and dealer competition across the 24 hour
trading day. Review of Financial Studies 12: 6194.
Huang, R., and H. Stoll. 1997. The components of the bid-ask spread: A general
approach. Review of Financial Studies 10: 9951034.
Huisman, R., K. Koedijk, C. Kool, and F. Nissen. 1998. Extreme support for uncovered
interest parity. Journal of International Money and Finance 17: 211228.
Hung, J. 1997. Intervention strategies and exchange rate volatility: A noise trading
perspective. Journal of International Money and Finance 16: 779793.
Institutional Investor. 2000. Trading meets the millennium. January.
International Monetary Fund. 1995. International capital markets: Developments, prospects, and policy issues, edited by D. Folkerts-Landau and T. Ito. Washington, DC:
International Monetary Fund.
Isard, P. 1995. Exchange Rate Economics. Cambridge: Cambridge University Press.
Ito, T., R. Lyons, and M. Melvin. 1998. Is there private information in the FX market?
The Tokyo experiment. Journal of Finance 53: 11111130.

316

References

Ito, T., and V. Roley. 1990. Intraday yen/dollar exchange rate movements: News or
noise? Journal of International Financial Markets, Institutions and Money 1: 131.
Jacklin, C., A. Kleidon, and P. Peiderer. 1992. Underestimation of portfolio insurance
and the crash of October 1987. Review of Financial Studies 5: 3563.
Jeanne, O., and A. Rose. 1999. Noise trading and exchange rate regimes, NBER Working Paper a7104, April. Forthcoming in the Quarterly Journal of Economics.
Johansen, S. 1992. Cointegration in partial systems and the efciency of single equation
analysis. Journal of Econometrics 52: 389402.
Jones, C., G. Kaul, and M. Lipson. 1994. Transactions, volume, and volatility. Review of
Financial Studies 7: 631651.
Jordan, J., and R. Radner. 1982. Rational expectations in microeconomic models: An
overview. Journal of Economic Theory 26: 201223.
Jorion, P. 1996. Risk and turnover in the foreign exchange market. In The Microstructure of Foreign Exchange Markets, edited by J. Frankel et al. Chicago: University of
Chicago Press.
Kaminsky, G., and K. Lewis. 1996. Does foreign exchange intervention signal future
monetary policy? Journal of Monetary Economics 37: 285312.
Kandel, E., and N. Pearson. 1995. Differential interpretation of public signals and trade
in speculative markets. Journal of Political Economy 103: 831872.
Kang, J., and R. Stulz. 1997. Why is there a home bias? An analysis of foreign portfolio
equity ownership in Japan. Journal of Financial Economics 46: 328.
Karpoff, J. 1986. A theory of trading volume. Journal of Finance 41: 10691087.
Karpoff, J. 1987. The relation between price changes and volume: A survey. Journal of
Financial and Quantitative Analysis 22: 109126.
Kaul, A., V. Mehrotra, and R. Morck. 2000. Demand curves for stock do slope down:
New evidence from an index weights adjustment. Journal of Finance 55: 893912.
Keynes, J. 1936. The General Theory of Employment, Interest, and Money. New York:
Harcourt Brace (see especially chapter 12).
Killeen, W., R. Lyons, and M. Moore. 2000a. Fixed versus oating exchange rates:
Lessons from order ow. Typescript, U.C. Berkeley, July.
Killeen, W., R. Lyons, and M. Moore. 2000b. The puzzle of the euro. Typescript, U.C.
Berkeley, July.
Kim, O., and R. Verrecchia. 1991. Trading volume and price reactions to public
announcements. Journal of Accounting Research 29: 302321.
Kirilenko, A. 1997. Endogenous trading arrangements in emerging foreign exchange
markets. Working paper, University of Pennsylvania (Wharton).
Klein, M. 1993. The accuracy of reports of foreign exchange intervention. Journal of International Money and Finance 12: 644653.
Kouri, P., and M. Porter. 1974. International capital ows and portfolio equilibrium.
Journal of Political Economy 82: 443467.

References

317

Krugman, P., and M. Miller. 1993. Why have a target zone? Carnegie-Rochester Conference Series on Public Policy 38: 279314.
Krugman, P., and M. Obstfeld. 2000. International Economics, 5th ed. Reading, Mass.:
Addison-Wesley.
Kurz, M. 1997. Foreign exchange rates under alternative expectational paradigms: A
resolution of the excess volatility puzzle. Working paper, Stanford University.
Kyle, A. 1985. Continuous auctions and insider trading. Econometrica 53: 13151335.
Kyle, A. 1989. Informed speculation with imperfect competition. Review of Economic
Studies 56: 317356.
Leach, J., and A. Madhavan. 1993. Price experimentation and security market structure. Review of Financial Studies 6: 375404.
Lee, C., and M. Ready. 1991. Inferring trade direction from intradaily data. Journal of
Finance 46: 733746.
Lehman Brothers. 2000. Global Economic Research Series: FX Impact of Cross-Border MA.
London: Lehman Brothers International (20 April).
Leland, H. 1992. Insider trading: Should it be prohibited? Journal of Political Economy
100: 859887.
Levich, R. 2001. International Financial Markets: Prices and Policies, 2nd ed. New
York: McGraw-Hill Irwin.
Lewis, K. 1988. Testing the portfolio balance model: A multilateral approach. Journal of
International Economics 24: 109127.
Lewis, K. 1995. Puzzles in international nancial markets. In Handbook of International
Economics, edited by G. Grossman and K. Rogoff. Amsterdam: Elsevier Science.
Loopesko, B. 1984. Relationships among exchange rates, intervention, and interest
rates: An empirical investigation. Journal of International Money and Finance 3: 257
277.
Luca, C. 2000. Trading in the Global Currency Markets, 2nd ed. New York: Prentice
Hall.
Lui, Y., and D. Mole. 1998. The use of fundamental and technical analyses by foreign
exchange dealers: Hong Kong evidence. Journal of International Money and Finance 17:
535545.
Lyons, R. 1995. Tests of microstructural hypotheses in the foreign exchange market.
Journal of Financial Economics 39: 321351.
Lyons, R. 1996a. Optimal transparency in a dealer market with an application to foreign exchange. Journal of Financial Intermediation 5: 225254.
Lyons, R. 1996b. Foreign exchange volume: Sound and fury signifying nothing? In The
Microstructure of Foreign Exchange Markets, edited by J. Frankel et al. Chicago: University of Chicago Press.
Lyons, R. 1997a. A simultaneous trade model of the foreign exchange hot potato.
Journal of International Economics 42: 275298.

318

References

Lyons, R. 1998. Prots and position control: A week of FX dealing. Journal of International Money and Finance 17: 97115.
Lyons, R., and A. Rose. 1995. Explaining forward exchange bias . . . intraday. Journal of
Finance 50: 13211329.
Machlup, F. 1939. The theory of foreign exchanges. Economica. Reprinted in Readings
in the Theory of International Trade, edited by H. Ellis and L. Metzler. Philadelphia:
Blakiston, 1949.
Madhavan, A. 1992. Trading mechanisms in securities markets. Journal of Finance 47:
607642.
Madhavan, A. 1996. Security prices and market transparency. Journal of Financial
Intermediation 5: 255283.
Madhavan, A. 2000. Market microstructure: A survey. Journal of Financial Markets 3:
205258.
Madhavan, A., and S. Smidt. 1991. A bayesian model of intraday specialist pricing.
Journal of Financial Economics 30: 99134.
Madhavan, A., and S. Smidt. 1993. An analysis of changes in specialist inventories and
quotations. Journal of Finance 48: 15951628.
Madrigal, V. 1996. Non-fundamental speculation. Journal of Finance 51: 553578.
MAR/Hedge. 2000. Monthly issues, Managed Account Reports, Inc., New York
(available at www.marhedge.com).
Mark, N. 1995. Exchange rates and fundamentals: Evidence on long-horizon predictability. American Economic Review 85: 201218.
Meese, R. 1986. Testing for bubbles in exchange markets. Journal of Political Economy 94:
345373.
Meese, R. 1990. Currency uctuations in the postBretton Woods era. Journal of Economic Perspectives 4: 117134.
Meese, R., and K. Rogoff. 1983a. Empirical exchange rate models of the seventies.
Journal of International Economics 14: 324.
Meese, R., and K. Rogoff. 1983b. The out-of-sample failure of empirical exchange
rate models. In Exchange Rate and International Macroeconomics, edited by J. Frenkel.
Chicago: University of Chicago Press.
Mehra, R., and E. Prescott. 1985. The equity premium: A puzzle. Journal of Monetary
Economics 15: 145161.
Melvin, M., and B. Peiers. 1997. The global transmission of volatility in the foreign
exchange market. Working paper, Arizona State University.
Melvin, M., and X. Yin. 2000. Public information arrival, exchange rate volatility, and
quote frequency. Economic Journal 110: 644651.
Mendelson, H. 1987. Consolidation, fragmentation, and market performance. Journal of
Financial and Quantitative Analysis 22: 189208.

References

319

Menkhoff, L. 1998. The noise trading approachQuestionnaire evidence from foreign

exchange. Journal of International Money and Finance 17: 547564.
Montgomery, J., and H. Popper. 1998. Information sharing and central bank intervention in the foreign exchange market. Typescript, Santa Clara University, December.
Moore, M., and M. Roche. 1999. Less of a puzzle: A new look at the forward forex
market. Typescript, Queen's University of Belfast, June.
Morris, S., and H. Shin. 2000. Rethinking multiple equilibria in macroeconomic modeling. NBER Macroeconomics Annual 2000.
Mussa, M. 1976. The exchange rate, the balance of payments, and monetary and scal
policy under a regime of controlled oating. Scandinavian Journal of Economics 78: 229
248.
Muth, J. 1960. Optimal properties of exponentially weighted forecasts. Journal of the
American Statistical Association 55: 290306.
Naik, N., A. Neuberger, and S. Viswanathan. 1999. Trade disclosure regulation in
markets with negotiated trades. Review of Financial Studies 12: 873900.
Naik, N., and P. Yadav. 2000. Do market intermediaries hedge their risk exposure with
derivatives? Typescript, London Business School, February.
Naranjo, A., and M. Nimalendran. 2000. Government intervention and adverse selection costs in foreign exchange markets. Review of Financial Studies 13: 453477.
New York Federal Reserve Bank. 1995. The Foreign Exchange Committee Annual Report,
1995. New York: Federal Reserve Bank.
Obstfeld, M., and K. Rogoff. 1995. Exchange rate dynamics Redux. Journal of Political
Economy 103: 624660.
Obstfeld, M., and K. Rogoff. 1996. Foundations of International Macroeconomics. Cambridge, MA: MIT Press.
Obstfeld, M., and A. Taylor. 1997. Non-linear aspects of goods-market arbitrage and
adjustment: Heckscher's commodity points revisited. Journal of the Japanese and International Economies 11: 441479.
O'Hara, M. 1995. Market Microstructure Theory. Cambridge, MA: Blackwell Business.
O'Hara, M. 1999. Making market microstructure matter. Financial Management 28: 83
90.
O'Hara, M., and G. Oldeld. 1986. The microeconomics of market making. Journal of
Financial and Quantitative Analysis 21: 361376.
Osler, C. 1995. Exchange rate dynamics and speculator horizons. Journal of International
Money & Finance 14: 695720.
Osler, C. 1998. Short-term speculators and the puzzling behavior of exchange rates.
Journal of International Economics 45: 3757.
Osler, C. 2000. Support for resistance: Technical analysis and intraday exchange rates.
Economic Policy Review 6: 5368.

320

References

Osler, C. 2001. Currency orders and exchange-rate dynamics: Explaining the success of
technical analysis. Typescript, Federal Reserve Bank of New York, March.
Pagano, M. 1989. Trading volume and asset liquidity. Quarterly Journal of Economics
104: 255274.
Pagano, M., and A. Roell. 1996. Transparency and liquidity: A comparison of auction
and dealer markets. Journal of Finance 51: 579612.
Payne, R. 1999. Informed trade in spot foreign exchange markets: An empirical investigation. Typescript, London School of Economics, January.
Payne, R., and P. Vitale. 2000. A transaction level study of the effects of central bank
intervention on exchange rates. Typescript, London School of Economics, July.
Peiers, B. 1997. Informed traders, intervention, and price leadership: A deeper view of
the microstructure of the foreign exchange market. Journal of Finance 52: 15891614.
Perold, A., and E. Schulman. 1988. The free lunch in currency hedging: Implications
for investment policy and performance standards. Financial Analysts Journal MayJune,
4550.
Perraudin, W., and P. Vitale. 1996. Interdealer trade and information ows in a
decentralized foreign exchange market. In The Microstructure of Foreign Exchange Markets, edited by J. Frankel et al. Chicago: University of Chicago Press.
Portes, R., and H. Rey. 1998. The emergence of the euro as an international currency.
Economic Policy 26: 307343.
Reiss, P., and I. Werner. 1995. Transaction costs in dealer markets: Evidence from the
London Stock Exchange. In Industrial Organization and Regulation of the Securities Industry, edited by A. Lo. Chicago: University of Chicago Press.
Reiss, P., and I. Werner. 1997. Interdealer trading: Evidence from London. Stanford
Business School Research Paper No. 1430, February.
Reiss, P., and I. Werner. 1998. Does risk sharing motivate interdealer trading? Journal of
Finance 53: 16571704.
Reiss, P., and I. Werner. 1999. Adverse selection in dealers' choice of interdealer trading system. Dice Working Paper no. 99-7, Ohio State University, June.
Rey, H. 2001. International trade and currency exchange. Review of Economic Studies 68:
443464.
Rime, D. 2000. Private or public information in foreign exchange markets? An empirical analysis. Typescript, University of Oslo, March (available at www.uio.no/
@dagnri).
Robinson, J. 1937. The foreign exchanges. Reprinted in Readings in the Theory of International Trade, edited by H. Ellis and L. Metzler. Philadelphia: Blakiston, 1949.
Rochet, J., and J. Vila. 1995. Insider trading without normality. Review of Economic
Studies 61: 131152.
Rogoff, K. 1984. On the effects of sterilized intervention: An analysis of weekly data.
Journal of Monetary Economics 14: 133150.
Roll, R. 1988. R 2 . Journal of Finance 43: 541566.

References

321

Romer, D. 1993. Rational asset-price movements without news. American Economic

Review 83: 11121130.
Romeu, R. 2001. Parameter stability in foreign exchange market microstructure. Typescript, University of Maryland, February.
Rosenberg, M. 1996. Currency Forecasting: A Guide to Fundamental and Technical Models
of Exchange Rate Determination. Chicago: Irwin Professional Publishing.
Rotemberg, J., and M. Woodford. 1997. An optimization-based econometric framework for the evaluation of monetary policy. In NBER Macroeconomics Annual, edited by
B. Bernanke and J. Rotemberg. Cambridge, MA: MIT Press.
Rubinstein, M. 2000. Rational markets: Yes or no? The afrmative case. U.C. Berkeley
Research Program in Finance Working Paper RPF-294, June (available at www.haas.
berkeley.edu/nance/WP/rfplist.html). Forthcoming in Financial Analysts Journal.
Saar, G. 1999. Demand uncertainty and the information content of the order ow.
Typescript, Cornell University, July.
Saporta, V. 1997. Which interdealer market prevails? An analysis of interdealer trading
in opaque markets. Bank of England, working paper 59.
Sarno, L, and M. Taylor. 2000a. The microstructure of the foreign exchange market.
Special Papers on International Economics, Princeton University (available at http://
www.princeton.edu/@ies).
Sarno, L., and M. Taylor. 2000b. Ofcial intervention in foreign exchange markets.
Typescript, University of Oxford, April.
Scholes, M. 1972. The market for securities: Substitution versus price pressure and the
effect of information on sharp price. Journal of Business 45: 179211.
Seasholes, M. 2000. Smart foreign traders in emerging markets. Typescript, Harvard
University, January.
Seppi, D. 1996. Liquidity provision with limit orders and a strategic specialist. Review
of Financial Studies 10: 103150.
Shiller, R. 1981. Do stock prices move too much to be justied by subsequent changes
in dividends? American Economic Review 71: 421436.
Shleifer, A. 1986. Do demand curves for stocks slope down? Journal of Finance 41: 579
590.
Shleifer, A., and R. Vishny. 1997. The limits to arbitrage. Journal of Finance 52: 3556.
Siegel, J. 1972. Risk, interest rates, and the forward exchange. Quarterly Journal of Economics 86: 303309.
Slezak, S. 1994. A theory of the dynamics of security returns around market closures.
Journal of Finance 49: 11631211.
Smith, R., and B. Madigan. 1988. Exchange market management and monetary policy
in the United States. In Exchange Market Intervention and Monetary Policy, edited by
Bank for International Settlements. Basel: BIS.
Snell, A., and I. Tonks. 1995. Determinants of price quote revisions on the London
stock exchange. Economic Journal 105: 5573.

322

References

Spiegel, M. 1998. Stock price volatility in a multiple security overlapping generations

model. Review of Financial Studies 11: 419447.
Spiegel, M., and A. Subrahmanyam. 1995. On intraday risk premia. Journal of Finance
50: 319339.
Stockman, A. 1987. The equilibrium approach to exchange rates. Federal Reserve Bank of
Richmond Economic Review 73, 2: 1231.
Stoll, H. 1978. The supply of dealer services in securities markets. Journal of Finance 33:
11331151.
Subrahmanyam, A. 1991. Risk aversion, market liquidity, and price efciency. Review
of Financial Studies 4: 417441.
Suvanto, A. 1993. Foreign exchange dealing: Essays on the microstructure of the
foreign exchange market. ETLAThe Research Institute of the Finnish Economy.
Helsinki: Taloustieto Oy.
Taylor, M. 1995. The economics of exchange rates. Journal of Economic Literature 83: 13
47.
Taylor, M., and H. Allen. 1992. The use of technical analysis in the foreign exchange
market. Journal of International Money and Finance 11: 304314.
Tesar, L., and I. Werner. 1995. Home bias and high turnover. Journal of International
Money and Finance 14: 467493.
Varian, H. 1985. Divergence of opinion in complete markets. Journal of Finance 40: 309
317.
Viswanathan, S., and J. Wang. 1998. Why is inter-dealer trading so pervasive in
nancial markets? Working paper, Duke University, February.
Vitale, P. 1998. Two months in the life of several gilt-edged market makers on the
London Stock Exchange. Journal of International Financial Markets, Institutions & Money
8: 301326.
Vitale, P. 1999. Sterilized central bank intervention in the foreign exchange market.
Journal of International Economics 49: 245267.
Vitale, P. 2000. Speculative noise trading and manipulation in the foreign exchange
market. Journal of International Money and Finance 19: 689712.
Vogler, K. 1997. Risk allocation and inter-dealer trading. European Economic Review 41:
417441.
Wang, J. 1994. A model of competitive stock trading volume. Journal of Political Economy 102: 127168.
Wei, S. 1994. Anticipations of foreign exchange volatility and bid-ask spreads. NBER
Working Paper 4737.
Wei, S., and J. Kim. 1997. The big players in the foreign exchange market: Do they
trade on information or noise? NBER Working Paper 6256, November.
Werner, I. 1997. A double auction model of interdealer trading. Stanford Business
School Research Paper, no. 1454.

References

323

Weston, J. 2000. Competition on the Nasdaq and the impact of recent market reforms,
Journal of Finance 55: 25652598.
Wolinsky, A. 1990. Information revelation in a market with pairwise meetings. Econometrica 58: 123.
Working, H. 1953. Futures trading and hedging. American Economic Review 43: 314
343.
Yao, J. 1998a. Market making in the interbank foreign exchange market. New York
University Salomon Center Working Paper aS-98-3.
Yao, J. 1998b. Spread components and dealer prots in the interbank foreign exchange
market. New York University Salomon Center Working Paper aS-98-4.
Zaheer, A., and Zaheer, S. 1995. Catching the wave: Alertness, responsiveness, and
market inuence in global electronic networks. Typescript, University of Minnesota,
December.
Zhou, B. 1996. High frequency data and volatility in foreign exchange rates. Journal of
Business and Economic Statistics 14: 4552.

This Page Intentionally Left Blank

Index

Accelerationist view, 61, 266

Agency frictions, 212214
Announcements, macroeconomic, 189,
265, 295n19
Anomaly persistence and order ow,
211212
Asset market approach to exchange rate
determination, 3, 15, 153155
Auctioneer. See Walrasian auctioneer
Balance of payments ows, 190192,
247248, 279n1, 293n14
Bond markets, 41, 57, 282n9, 283n8
Brokers, interdealer spot, 4246, 105, 123,
275277, 284nn10, 12, 285n19
Bubbles, 172173, 293n9
Capital ows versus order ows, 246
248
CARA-Normal framework, 69, 105112
Causality, 189190, 204205, 269, 286n2
Central bank intervention
announced, unannounced, and secret,
227
and asymmetric information, 232233
and exchange rate regimes, 223225
and xed versus oating rates, 229231
in foreign exchange markets (FX), 227
229
and the Kyle model, 234236
microstructure models of, 231236
and order ow, 223225, 231
price impact of, 236241
and secrecy, 232235
sterilized, 223225, 225227
unsterilized, 223225
and the Vitale 1999 model, 234236

Citibank, 124, 245246, 252, 270

Cointegration, persistent price impact,
202206, 281nn4, 18, 295n13, 296n26
Covered interest parity, 283n3, 297n31
Crashes and collapses, market, 12, 59.
See also Fixed exchange rates
Credit risk and market structure, 272
Customer-dealer trades
and customer orders received by banks,
124
and FXFX indicative quotes, 124125
Customers
background on, 243249
and balance of payments, 247248
categories of, 249252, 270
centralized trading for, 276277
and the collapse of the Yen/$ rate, 256
261
and the Evans-Lyons model, 248249
and information structure, 253256
and institution types, 253256
leveraged investors as, 217
nonnancial corporations as, 217
and order ows, 246249, 252253, 269
270
spot market, 42, 44, 45
unleveraged investors as, 217
Data sets, foreign exchange market (FX)
customer orders received by banks, 124,
290n6
and equity data, 114
Evans, 121123
Evans-Lyons, 176190
FXFX indicative quotes, 113114, 124
125, 283n7, 290n7, 303n1
Goodhart and Payne, 123

326
Data sets, foreign exchange market (FX)
(cont.)
and information sets, 125126
integration, 125, 270271, 290n8
Killeen, Lyons, and Moore, 123124,
192198, 200202
Lyons, 117118
Yao, 118121
Dealer markets, 6365
multiple, 4041, 93105, 289n27
single, 40, 8693
Dealer-problem structural model, 134
139
formation of expectations in the, 137
138
quote determination, 138
Dealers, foreign exchange market, 68,
40, 4551, 93
and brokered interdealer transactions,
123124
and customer-dealer trades, 124125
and the daily frequency model, 176190
data integration among, 270271
and direct interdealer transactions, 93
94, 9899, 117123
inventories, 46, 141, 267
and marketmakers, 283n6
prots, 4951, 285n20, 288n23
position sheets, 4749
structural microstructure models, 134
139
and types of markets, 6366
vote counters, 7778
Demand and order ow, 7, 73, 188, 198
199
Derivatives, 3839, 52, 246, 283n4,
285n22
Determination puzzle, exchange rates,
172173
and asset payoff, 176180
and causality, 189190
daily frequency model, 176180
and dealers, 176190
and equity markets, 294n2
and the Evans-Lyons model, 176190
hybrid micro-macro models, 173176
and order ow, 187189
and the portfolio balance model, 180
184
and price, 180187
and robustness checks, 187188
Direct interdealer trading, 115123
and the Dealing 2000-1 system, 117123

Index
Evans data set on, 121123
Lyons data set on, 117118
Yao data set on, 118121
Discount-rate information, 3031
and imperfect substitutability, 3334
transitory inventory effects on, 3233
Distressed selling, 257, 269
Efciency
semi-strong form, 3, 128, 303n11
strong form, 87, 271
Electronic Broking System (EBS), 116
117, 123124, 201202, 276277, 290n3,
300n5
Electronic trading, 55, 113114, 290n5
Emerging markets, 274, 281n2, 290n2
Equilibrium
model of exchange rate determination,
164166
Killeen, Lyons, and Moore model, 198
Kyle auction model, 8081, 8284
rational expectations auction model, 66
67, 7071
simultaneous-trade model, 100102
Error correction, 204205
European Monetary Union (EMU), 53,
193
Evans data set, 121123, 176190
Evans-Lyons model
and asset payoff, 176179
and causality, 189190
and customer ow, 248249, 252253
and the ow approach to exchange
rates, 190192
and interdealer trading, 179180
and intervention trades, 236241
and order ow, 187189
and portfolio balancing, 181184, 222
and price, 180184
results, 184187
robustness checks, 187188
simultaneous-trade model links, 183
184
Event uncertainty, 145, 237238, 295n18
Excess volatility puzzle, the, 192206
Exchange rate economics
asset market approach to, 3, 15, 153155
and the collapse of the Yen/$ rate, 256
261
and the determination puzzle, 172190,
294n3
and the excess volatility puzzle, 192
206

Index
and expectations, 14, 173176
and xed versus oating rates, 192206,
229231, 296n23
exible-price monetary model, 155157
and oating rate volatility, 265266
ow approach to, 190192
and the forward bias puzzle, 206220
fundamentals, 28, 151155, 157, 301n14
general equilibrium model, 164166,
294n3, 298n44, 303n5
goods market approach to, 2, 152153
hybrid micro-macro models of, 1618,
173176
and information dispersion, 1213
and the issues spectrum, 167, 169
and macroeconomic models, 12, 294n1
and microfoundations, 151152, 164
167, 303n5
microstructure approach to, 45, 1011,
1314, 1518
and micro versus macro issues, 167
170
models, 155163
and order ow, 68, 910, 139142,
198199
overshooting model, 155157
portfolio balance model, 160163
and purchasing power parity, 152153,
155157
and spread, 89
sticky-price monetary model, 157160
and trading volume, 3, 1314, 5155
and uncovered interest parity, 153
155
and utility maximization, 164166
and the wealth constraint, 161162
Expectations
and exchange rate determination, 14,
173176, 282n12, 295n4, 300n12
and the rational expectations auction
model, 7173
tools for calculating, 72, 83, 108, 111
112, 137138
Fixed exchange rates, 12, 192206, 229
231, 273274
Flexible-price monetary model, 155157
Flow approach to exchange rates, 190
192, 279n1
Foreign exchange markets (FX), 25. See
also Spot market, the
asset market view of, 153155
and book passing, 284n14

327
brokered interdealer trades, 114117,
123124
brokers, 4245
central bank intervention, 223225, 227
229
choke point, 222
and the collapse of the Yen/$ rate, 256
261
compared to other markets, 41
credit risk and market structure, 272,
296n27
customer-dealer trades, 114117
customers, 42, 243261
daily volume, 3839
dealers, 68, 2326, 4041, 4551
derivatives, 3839, 52, 246, 283n4,
285n22
description of, 3738
direct interdealer trades, 117123
directions for research, 268272
Evans data set, 121123
and expectations, 14
exible-price monetary model, 155157
future of, 275277
Goodhart and Payne data set, 123
goods market view of, 152153
and hot potato trading, 144149, 267
268
and international currencies, 274275
and inventory control, 143144
Killeen, Lyons, and Moore data set, 123
124
large dealer positions, 4651
Lyons data set, 117118
market design, 274
microstructure effects, 5962, 302n1
modeling strategies, 4551
order ow, 3132, 139142
overshooting model, 157160
players, 4145
policy implications in, 272275, 286n28
portfolio balance model, 160163, 222
and price, 142149
protability, dealer, 4951, 285n20,
288n23
and public information, 2122, 25,
281n1
and purchasing power parity, 152153,
155157
research data sets, 113117
and the simultaneous-trade model, 93
105
and spreads, 4344, 5051

328

Foreign exchange markets (FX) (cont.)

statistical models, 126134
sticky-price monetary model, 157160
technical analysis in, 271272, 298n42
trading volume, 41, 5155, 142149
transaction taxes, 275
and transparency of order ow, 4546,
5557, 275276, 289n31
turnover, 5155
typical trading day, 4751
and uncovered interest parity, 153155
Yao data set, 118121
Forward bias puzzle, the
and anomaly persistence, 211212
description of, 206207
economist's perspective on, 208209
empirical support for, 216219
and institutional allocation of
speculative capital, 212214
and order ow, 219
practitioner's perspective on, 209210
and risk premium, 209, 219220
and Sharpe ratios, 212213
statistician's perspective on, 207
and uncovered interest parity, 217218
and underallocation of speculative
capital, 215216
Fundamentals, 28, 151155, 157, 301n14
Future of foreign exchange markets (FX),
275277
Future payoffs and order ow, 3435
Future research, 268272, 302n7, 303n10
General equilibrium model, 164166,
294n3, 298n44, 303n5
Goodhart and Payne data set, 123
Goods market approach to exchange rate
determination, 2, 152153
Granger causality, 205206, 269
Handbook of International Economics, 2
Hong Kong Stock Exchange, 40, 59
Hot potato trading, 13, 9596, 102104,
144149, 267268
Imperfect substitutability, 3334, 174,
221, 180, 240, 282n16, 292n26, 299n2
Inaction range, 215220
Information
accelerationist view, 61, 266
aggregation of dispersed, 12, 14, 3132,
36, 99, 232

Index

and central bank intervention, 231236

discount-rate, 3031
and exchange rate determination, 173
176
and focus of book, 4
and foreign exchange dealers, 2526,
5557
fundamental, 28, 151155, 157, 301n14
graphical interpretation of, 3335
and hot potato trading, 146149
intermediation, 95
investing in, 6768, 286n31
and the Kyle auction model, 7879, 82
84
mapping to price, 14, 21, 23, 5960, 174
176, 189, 232, 265
and microfoundations, 166167
and the microstructure approach to
exchange rates, 4, 1213
and opinion, 287n7
and order ow, 78, 2022, 5557, 139
142, 146149, 265
and payoffs and non-payoffs, 27, 2930,
152, 268269
and price, 1112, 2223, 253256
private, 20, 2628, 8788, 95, 104, 271,
280n12, 291n13, 301n18, 303n11
processing, two stages, 7
public, 34, 89, 1213, 128, 266, 279n5,
281n1, 289n28, 296n21
and the rational expectations auction
model, 6667, 70
and revelation, 6667, 8384, 102103
and the sequential-trade model, 8788
and the simultaneous-trade model, 95,
9699
and spot market dealers, 45
and spreads, 2324
and structural microstructure models,
139142
taxonomy, 2832
and transitory inventory effects on
discount rates, 3334
types of private, 2835
and types of market models, 6366
and volatility responses to trading halts,
2425
Institutions
allocation of speculative capital, 212
214
design, 58
and microfoundations, 166167

Index

and the microstructure approach, 4, 57

59
and the rational expectations auction
model, 70
types, price impact, and information
structure, 253256
International currencies, 274275
Intervention. See Central bank
intervention
Inventory models, 46, 66, 141, 291n17,
292nn21, 24, 294n5, 303n2
Jensen's inequality, 287n11
Killeen, Lyons, and Moore model
and asset payoff, 194198
and cointegration, 202205
data set, 123124, 200202
equilibrium, 198
and excess volatility, 205206
Kyle auction model
asset payoff, 7982
and central bank intervention, 234236
description of, 6465, 7677
details, 7982
equilibrium, 8284
and individual trades, 86
and information, 8081
insights, 7779
and institutions, 8081
and liquidity and market efciency, 77
79
marketmakers, 7779, 7982
and order ow, 78
players, 8081
and risk-aversion, 85
and spread, 85
Lamppost metaphor, 171
Limit orders, 67, 290nn6, 11, 304n11
Liquidity, 7779, 84, 260261, 273,
284n11
London Stock Exchange, 56
Lyons data set, 117118, 176190
Macroeconomics
and asset approach to exchange rates,
15, 153155
assumptions, 4, 279n2
and central bank intervention, 223227
determination puzzle, 172190
excess volatility model, 192206

329

and exchange rate models, 155163

exible-price monetary model, 155157
forward bias model, 207220
and goods market approach to exchange
rates, 2, 152153
hybrid micro-macro models, 1518,
173176
and microfoundations, 164170, 294n3,
303n5
and microstructure approach to
exchange rates, 12, 14, 1518, 264268
and micro versus macro issues, 167170
and purchasing power parity, 152153,
155157
sticky-price monetary model, 157160
and tastes and technology, 164170
trends in, 279n2
and uncovered interest parity, 153155
Markets, auction and dealer, 40, 6365,
274
Martingale, 289n25
Meta models, 217218
Microfoundations, 164170, 294n3, 303n5
Microstructure effects, 6062, 302n1
Microstructure models
and analysis of crashes and collapses,
12, 59
and CARA-normal framework, 69, 105
112
of central bank intervention, 231236
dealer-problem, 134139
denitions in, 59
and the determination puzzle, 172190
directions for research, 268272
and discount-rate information, 3031
effects versus tools, 5962, 301n1
and exchange rate determination, 25,
1314, 1516
versus exchange rate economics, 12
and the ow approach to exchange
rates, 190192
and foreign exchange markets, 2, 45,
113117, 125126, 264268, 267268
hallmarks, 59
hybrid micro-macro, 1618, 173176
and imperfect substitutability, 221
and information economics, 4, 1213
and institution design, 5759
institutions, 4
inventory models, 46, 66, 141, 291n17,
292nn21, 24, 294n5, 303n2
and the Kyle auction model, 7786

330

Microstructure models (cont.)

lessons learned, 264268
long-lived implications of, 1112
lower frequency relevance of, 1112,
176, 198, 293n18
and macroeconomics, 12, 14, 1518
and manipulation, 288n17
and order ow, 68, 910, 267
players, 4, 68
policy implications of, 272275
and the rational expectations auction
model, 6676
and risk aversion, 105112, 163, 166,
219220
and the sequential-trade model, 8693
and the simultaneous-trade model, 93
105
and spreads, 89, 2324, 130134
statistical, 126134
statistical versus structural, 126
types of, 6366
Monetary models
exible-price, 155157
and institutions and information, 166
167
overshooting, 157160
portfolio balance, 160163, 180184
sticky-price, 157160
Multiple-dealer markets, 4041, 93105
Multiple equilibria, 11, 295n10
NASDAQ, 56, 59, 284n9
NBER microstructure program, 169170
New York Stock Exchange (NYSE), 24
25, 40, 56, 291n18
Nonrational behavior, 172173
Notation, DM/$ versus $/DM, 280n16
Open issues, 268270
Order ow, 610
and anomaly persistence, 211212
and asset approach to exchange rates,
15, 294nn4, 6
and auction markets, 6, 290n11
brokered versus direct, 114117, 291n14
versus capital ows, 246248
and central bank intervention, 223225,
231
and cointegration issues, 202206,
281nn4, 18, 295n13, 296n26
and customer orders, 124, 243252, 269
270, 295n13

Index

dened, 67
and demand, 7, 73, 188, 198199
and distressed selling, 257, 269
and the Evans-Lyons model, 187189
and the excess volatility puzzle, 194
195
exogeneity of, 204205, 269
and expected future payoffs, 3435
expected versus unexpected, 188
and oating rate volatility, 265266
and foreign exchange dealers, 2526,
3132, 139142
and FX swaps, 3839, 5152, 246
and forward rate bias, 219
graphical interpretation of, 3235
and hot potato trading, 146149
hybrid micro-macro models of, 1618,
173176
and imperfect substitutability, 3334
information role for, 2022
and information structure, 253256, 265
and institution types, 253256
and the Killeen, Lyons, and Moore
model, 200202
and the Kyle auction model, 78
and limit orders, 67, 290nn6, 11,
304n11
measuring, 127, 284n12
and microstructure approach to
exchange rates, 1518, 6062, 173176,
267
and monetary models, 167, 295n4
and price, 1112, 2223, 2930, 198200,
264266, 273, 291nn13, 14
price impact measured, 141, 147, 186
187, 236, 239240, 254255, 265, 270,
273, 301n21
and random walk, 184
and the rational expectations auction
model, 76
and reverse causality, 269
and spot market dealers, 4546
and spreads, 2324
and structural models, 139142
and supply and demand, 3132
and survey data, 2526
time-aggregated, 23
and trading volume, 67
and transitory inventory effects on
discount rates, 3233
transparency of, 4546, 5557, 275276,
289n31

Index

types of information conveyed by, 28

35
and vector auto-regression models
(VAR), 22
and volatility responses to trading halts,
2425
what drives, 1718, 188189
Overshooting model, 157160
Over-the-counter markets (OTC), 56
Paris Bourse, 40, 59
Path dependence, 260261
Preferences, 69, 105112
Policy implications of microstructure
analysis, 272275, 286n28
Portfolio balance model, 160163, 180
184
and price effects, 3035, 222, 226, 240,
300n8
and sterilized intervention, 225227
Price
and asset payoffs, 2930
and brokers, 4245
and customer types, 270
and the daily frequency model, 176
190
and dealer inventories, 267
discovery, 87
elasticity, 265
and error correction, 204205
and the Evans-Lyons model, 180184
and the exible-price monetary model,
155157
and the goods market approach to
exchange rates, 152153
and institution types, 253256
and intervention trades, 236241
market-clearing, 7374
and order ow, 2223, 199200, 264
266, 273
and the overshooting model, 157160
and path dependence, 260261
and the portfolio balance model, 160
163, 180184
and private information, 2628
and public information, 128, 266
and purchasing power parity, 152153
and the rational expectations model, 67
68, 70, 7374, 76
regret free, 89, 135, 138
and the sequential-trade model, 8788,
89, 9192

331

and the simultaneous-trade model, 102

104
and the spot markets, 42
state dependency of, 237238
and the sticky-price monetary model,
157160
and trading volume, 142149
and uncovered interest parity, 153155
and vector auto-regression (VAR)
models, 22, 127130
and the wealth constraint, 161162
Private information, 20, 2628, 8788, 95,
104, 271, 280n12, 291n13, 301n18,
303n11
Protability, dealer, 4951, 285n20,
288n23
Purchasing power parity, 152153, 155
157, 293n11
Puzzles, exchange rates, 1314, 171
determination, 172190
excess volatility, 192206
forward bias, 206220
home bias, 170, 280n11
volume, 1314, 142144
Quotes
conventions, 272
FXFX indicative, 113114, 124125,
283n7, 290n7, 303n1
and the handle, 119
regret free, 89, 135, 138
Rational expectations auction model
asset payoff, 6870
description of, 6364, 6667
details, 6870
equilibrium, 6667, 7071
generality and the, 75
implicit auctioneer, 7475
insights, 6768
and knowledge of the pricing rule, 76
and market-clearing price, 7374
and nonstrategic behavior, 75
and order ow, 76
players, 70
and price, 6768, 70
and traders' demands, 73
and traders' expectations, 7173
trading period, 6870
utility function, 6970
Real money accounts, 256
Regret free quotes, 89, 135, 138

332

Revelation, full versus partial, 6667

Risk aversion and microstructure models,
105112, 163, 166, 219220
Risk premium, 2636, 160163, 181184,
209, 219220, 282n13
Riskless asset, 111, 287n11
Siegel's paradox, 287n11
Sequential-trade model
and dealer pricing, 89
dealers, 8788
description of, 65, 86
details, 8891
insights, 8788
and institutions, 89
and the order arrival process, 92
players, 89
and price discovery, 8788
and private information, 8788
and risk aversion, 93
setting bid and offer prices in the, 9192
and spreads, 8788
and strategic behavior, 92
trade types, 9091
Sharpe ratio of currency strategies, 210
216, 297n34, 298nn38, 44
Signaling channel and sterilized
intervention, 225227
Simultaneity, 89, 135, 138, 189190, 204
205, 292n23
Simultaneous-trade model
and brokered trading, 104105
and dealer inventories, 95
dealer objectives and information sets,
100
dealer quotes, 100102
description of, 6566, 9394
details, 9699
equilibrium, 100102
Evans-Lyons model links, 183184
and foreign exchange markets (FX), 93
94
and hot potato trading, 9596, 102104
and information, 95, 9699
insights, 9496
and institutions, 98
players, 98
and price, 102104
and private information, 95, 104
and risk aversion, 9596
and signaling, 104
and spread, 104

Index

Single-dealer markets, 40, 8693

Speculative capital, 213215
Speculative efciency model, 3
Spot market, the. See also Foreign
exchange markets (FX)
and auction markets, 40
and brokered interdealer trades, 4344,
4546, 55
brokers, 4245
and customer-dealer trades, 4344
customers, 42, 44, 45
dealers, 4551
and direct interdealer trades, 4344
distinguishing characteristics, 41
and electronic brokers, 55
features of, 3951
and multiple-dealer markets, 4041
players, 4145
positions, dealer, 4651
prices, 42
and single-dealer markets, 40
trade types, 43
trading volume, 3839, 41
and transparency of order ow, 4546,
5557, 275276, 289n31
turnover, 5155
types of trades, 4345
Spreads
components, 2324, 130133
and costs, 131134, 291n19
and exchange rates, 89
and foreign exchange markets, 4344,
5051
and informative order ow, 2324
and inventory component, 131134
and the Kyle auction model, 85
and microstructure research, 8
and the sequential-trade model, 8788
and the simultaneous-trade model, 104
and the trade-indicator model, 130134
Statistical microstructure models
trade-indicator, 130134
vector auto-regression (VAR), 22, 127130
Sterilized central bank intervention, 223
225, 225227
Sticky-price monetary model, 157160
Strategic trading, 75, 80, 85, 92
Structural microstructure models
dealer-problem, 134139
and hot potato trading, 144149
information structure, 135137
and order ow, 139142

Index

Supply and order ow, 3235

Systematic versus diversiable risk, 31,
213214
Tastes and technology, 164170
Taxes, transaction, 143, 275, 304n10
Technical analysis in foreign exchange
markets (FX), 271272, 290n6, 298n42
Tool kit, 105112
Trade-indicator model, 130134
Trading halts, 2425
Trading intensity, 84, 86, 145
Trading mechanisms
and the asset market approach to
exchange rates, 45
and customer-dealer trades, 124125
data integration among, 270271
direct interdealer, 9394, 9899, 115123
and the Evans data set, 121123
and FXFX indicative quotes, 124125
and the Goodhart and Payne data set,
123
and the goods market approach to
exchange rates, 2
hot potato, 9596, 102104, 144149,
267268
intervention, 236241
and the Killen, Lyons, and Moore data
set, 123124
and limit orders, 67, 290nn6, 11,
304n11
and the Lyons data set, 117118
and the microstructure approach to
exchange rates, 45
and order ow, 68, 910
and private information, 20, 2628, 87
88, 95, 104
spot market, 4345, 4551
and spreads, 89, 4344
state-dependency of, 237238
and trading volume, 3, 1314, 41, 5155,
142149
and volatility responses to trading halts,
2425
and the Yao data set, 118121
Transaction taxes, 275
Transaction volume, 3, 6, 1314, 41, 51
55, 142149, 292n25
Transitory inventory effects on discount
rates, 3233
Transparency of order ow, 4546, 55
57, 275276, 289n31

333

Uncovered interest parity, 153155, 217

218, 295n15
Unsterilized central bank intervention,
223225
Utility maximization, 105112, 164166
Vector auto-regression models (VAR), 22,
127130
Vehicle currencies, 286n24
Vitale 1999 model, the, 234236
Volatility, 11, 2425, 158, 169, 192206,
293n20, 296n22
Volume, trading, 3, 6, 1314, 41, 5155,
125, 142149, 292n25
Walrasian auctioneer, 910, 6465, 74
77, 280n1
Wealth constraint, 106108, 161162
Yao data set, 118121
Yen, Japanese, in fall 1998, 256261