NBER WORKING PAPER SERIES

EXCHANGE RATE FUNDAMENTALS AND ORDER FLOW

Martin D. D. Evans
Richard K. Lyons

Working Paper 13151

http://www.nber.org/papers/w13151

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
June 2007

We thank the following for valuable comments: Anna Pavlova, Andrew Rose and seminar participants
at the NBER (October 2004 meeting of IFM), the Board of Governors at the Federal Reserve, the European
Central Bank, the London Business School, the University of Warwick, the Graduate School of Business
at the University of Chicago, UC Berkeley, the Bank of Canada, the International Monetary Fund,
and the Federal Reserve Bank of New York. Both authors thank the National Science Foundation for
financial support, which includes funding for a clearinghouse for recent micro-based research on exchange
rates (at georgetown.edu/faculty/evansm1 and at faculty.haas.berkeley.edu/lyons). The views expressed
herein are those of the author(s) and do not necessarily reflect the views of the National Bureau of
Economic Research.

© 2007 by Martin D. D. Evans and Richard K. Lyons. All rights reserved. Short sections of text, not
to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including
© notice, is given to the source.
Exchange Rate Fundamentals and Order Flow
Martin D. D. Evans and Richard K. Lyons
NBER Working Paper No. 13151
June 2007
JEL No. F31,G12,G14

ABSTRACT

We address whether transaction flows in foreign exchange markets convey fundamental information.
Our GE model includes fundamental information that first manifests at the micro level and is not symmetrically
observed by all agents. This produces foreign exchange transactions that play a central role in information
aggregation, providing testable links between transaction flows, exchange rates, and future fundamentals.
We test these links using data on all end-user currency trades received at Citibank over 6.5 years, a
sample sufficiently long to analyze real-time forecasts at the quarterly horizon. The predictions are
borne out in four empirical findings that define this paper's main contribution: (1) transaction flows
forecast future macro variables such as output growth, money growth, and inflation, (2) transaction
flows forecast these macro variables significantly better than the exchange rate does, (3) transaction
flows (proprietary) forecast future exchange rates, and (4) the forecasted part of fundamentals is better
at explaining exchange rates than standard measured fundamentals.

Martin D. D. Evans
Department of Economics
Georgetown University
Washington, DC 20057
evansm1@georgetown.edu

Richard K. Lyons
Goldman Sachs
85 Broad Street, Floor 20
New York, NY 10004
and NBER
Richard.Lyons@gs.com
Introduction
Exchange rate movements at frequencies of one year or less remain unexplained by observable
macroeconomic variables (Meese and Rogo¤ 1983, Frankel and Rose 1995, Cheung et al. 2005). In
their survey, Frankel and Rose (1995) describe evidence to date as indicating that "no model based
on such standard fundamentals ... 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." Seven years later,
Cheung et al.’s (2005) comprehensive study concludes that "no model consistently outperforms a
random walk."
This paper addresses this long-standing puzzle from a new direction. Rather than attempting
to empirically link macro variables to exchange rates directly, we address instead the intermediate
market-based process that impounds macro information into exchange rates. Our approach is
based two central ideas: First, only some of the macro information relevant for the current spot
exchange rate is publicly known at any point in time. Other information is present in the economy,
but it exists in a dispersed microeconomic form in the sense of Hayek (1945). The second idea
relates to determination of the spot rate through the operation of the foreign exchange market.
Speci…cally, since the spot rate literally is the price of foreign currency quoted by foreign exchange
dealers, it can only re‡ect information that is known to dealers. Consequently, the spot rate will
only re‡ect dispersed information once it has been assimilated by dealers, (collectively called “the
market”) –a process that takes place via trading. We shall argue that this trade-based mechanism
is economically important because much information about the current state is dispersed, and
because it takes a considerable time for dispersed information to be completely assimilated by “the
market”.
To make these ideas concrete, we present a two-country general equilibrium model in which the
spot rate is determined via the optimal trading activities of dealers in the foreign exchange market.
Our model contains three essential ingredients. First, it includes information that is not publicly
observed, at least initially. Second, transaction ‡ows are correlated with this information. Third,
the equilibrium spot rate is not fully revealing. The model not only provides a theoretical rationale
for the strong empirical link between spot rate changes and transaction ‡ows (see, for example,
Evans and Lyons 2002a,b), but it also delivers two new testable implications: First, transaction
‡ows should have more power to forecast future fundamentals than current spot rates. Second,

1
insofar as the transaction ‡ows received by individual dealers predict what the rest of “the market”
will learn about fundamentals in the future, those ‡ows should have forecasting power for future
exchange rate returns.
We investigate these empirical predictions using a new data set that comprises USD/EUR spot
rates, transaction ‡ows and macro fundamentals over six and a half years. The transaction ‡ows
come from Citibank and represent propriety information of an important Bank in the USD/EUR
market. A novel and important feature of our empirical analysis is that it utilizes high-frequency
real-time estimates of macro variables. These data are estimates of the underlying macro variables
based on contemporaneously available public information. As such, they provide a more precise
measure of public expectations regarding fundamentals than realizations of the variables themselves.
This greater precision is re‡ected in the strong statistical signi…cance of our …ndings.
The implications of our model are strongly supported by our data. In particular we …nd that:

1. Transaction ‡ows in the USD/EUR market have signi…cant forecasting power for future out-
put growth, money growth, and in‡ation in both the US and Germany.

2. Transaction ‡ows have incremental forecasting power for macro variables beyond that con-
tained in the history of exchange rates and the variable itself.

3. Propriety transaction ‡ows forecast future exchange rate returns, and do so much more ef-
fectively than forward discounts.

4. The forecasting power of propriety transaction ‡ows re‡ects their ability to predict how “the
market” will react to the ‡ow of subsequent information concerning macro fundamentals.

To the best of our knowledge, these are the …rst …ndings to link macro fundamentals, transaction
‡ows and exchange rate dynamics. Taken together, they provide strong support for the idea that ex-
change rates vary as “the market”assimilates dispersed information regarding macro fundamentals
from transaction ‡ows.
Our analysis is related to several strands of the international …nance literature. From a theoreti-
cal perspective, our general equilibrium model includes two novel ingredients: dispersed information
and a micro-based rationale for trade in the foreign exchange market. Dispersed information does

2
not exist in textbook models: relevant information is either symmetric economy-wide, or, some-
times, asymmetrically assigned to a single agent – the central bank. As a result, no textbook
model predicts that market-wide transaction ‡ows should drive exchange rates. In recent research,
Bacchetta and van Wincoop (2006) examine the dynamics of the exchange rate in a rational ex-
pectations model with dispersed information. Our model shares some of the same informational
features, but derives the equilibrium dynamics from the equilibrium trading strategies of foreign
exchange dealers. Our focus on the role of transaction ‡ows as conveyors of information concerning
macro fundamentals also di¤ers from Bacchetta and van Wincoop (2006).
From a empirical perspective, our analysis is closely related to the work of Engel and West
(2005). They …nd that spot rates have forecasting power for future macro fundamentals as textbook
models predict. Indeed, our model makes the same empirical prediction. The novel aspect of our
analysis, relative to Engel and West (2005), is that we investigate whether the exchange rate
responds to transaction ‡ows because they induce a change in “the market’s” expectations about
future fundamentals. From this perspective, our …ndings should be viewed as complementing theirs.
Our analysis is also related to earlier research by Froot and Ramadorai (2005), hereafter F&R.
These authors examine VAR relationships between real exchange rates, excess currency returns,
real interest di¤erentials, and the transaction ‡ows of institutional investors. In contrast to our
results, they …nd little evidence that these ‡ow can forecast fundamentals. Our analysis di¤ers from
F&R in three respects. First, and most substantively, transaction ‡ows should be driven not by
changes in fundamentals, but by changes in fundamentals expectations. The F&R analysis focuses
on the former, whereas ours focuses on the latter. Second, we analyze transaction ‡ows which fully
span the demand for foreign currency, not just institutional investors. This facet of our ‡ow data
proves to be empirically important. Third, we require no assumption about exchange rate behavior
in the long run, whereas the variance decompositions F&R use are based on long run purchasing
power parity.
The rest of the paper is organized as follows. Section 1 provides an overview of our model and
presents the key equations determining the spot exchange rate. Section 2 derives the theoretical link
between transaction ‡ows and exchange rate fundamentals. Section 3 describes the data. Section
4 presents our empirical analysis. Section 5 concludes.

3
1 The Model

Our model is a two-country, two-good dynamic general equilibrium model that incorporates explicit
microfoundations of how trading takes place in the foreign exchange market. For this purpose, we
need to model the behavior of households, …rms, central banks and foreign exchange dealers who
act as market-makers. In this section, we …rst present the preferences and constraints facing house-
holds and …rms and describe the role of central banks. We then lay out the problem facing foreign
exchange dealers and provide intuition for their equilibrium behavior. Finally we present the equi-
librium equation for the spot exchange rate that plays a central role in our analysis. The Appendix
describes the complete structure of the model and provides detailed mathematical derivations of
our key results.

1.1 Households, Firms and Central Banks

There are two countries, each populated by a continuum of households arranged on the unit interval
[0,1]. For concreteness, we shall refer to home and foreign countries as the US and Europe and use
^ 2 [1=2; 1] to denote European households.
the index h 2 [0; 1=2) to denote US households and h
All households derive utility from consumption and real balances. The preferences of US household
h are given by:
1
X Mh;t+i 1
1
Uht = Eht i
1
1
Ch;t+i + 1 Pt+i ; (1)
i=0

where 0 < < 1 is the discount factor, > 0 and 1. Eht denotes expectations conditioned on
US household information, h;t : Mh;t is the stock of dollars held by household h; and Ch;t is a CES
consumption index de…ned over the two consumption goods:

Ch;t (Ch;t (us)( 1)=

+ Ch;t (eu)( 1)=
) =( 1)
; (2)

where Ch;t (i) is the consumption of the i-country good by household h: is the elasticity of sub-
stitution between the two goods, which we assume to be greater than one (see below). The price
us(1 ) eu(1 ) 1=(1 );
index corresponding to (2) is Pt (Pt + Pt ) where Pti are the prices of good i.
The preferences of European households are de…ned in an analogous manner with respect to the
foreign consumption index, C^h;t ; and real balances, M
^ h;t =P^t , where P^t is the European price level.

4
Hereafter, we use “hats” to indicate European variables.
In addition to domestic currency, households can hold one-period nominal dollar bonds, B;
^ and the equities issued by US and European …rms, A and A:
nominal euro bonds B; ^ Let Rt and R
^t

be the US and European one period gross nominal interest rates and let St denote the spot exchange
rate, speci…cally, the dollar price of euros ($/e). The budget constraint facing US household h is

^h;t + St Q
Bh;t + Qt Ah;t + St B ^ t A^h;t + Mh;t + Pt Ch;t =

(Qt + Dt ) Ah;t 1
^t + D
+ St (Q ^ t )A^h;t 1 + Rt 1 Bh;t 1
^t
+ St R ^
1 Bh;t 1 + Mh;t 1 (3)

^ t are the local currency prices of US and European equities with dividends per
where Qt and Q
^ t respectively. The problem facing US household h in period t is to choose
share of Dt and D
^h;t ; A^h;t ; Ah;t ; Mh;t ; and Ch;t (i) for i = fus, eug given prices fQt ; Q
Bh;t ; B ^ t ; Ptus ; Pteu g; dividends
^ t g; interest rates {Rt ; R
fDt ; D ^ t g; and the spot exchange rate St ; that maximize (1) subject to (3).

There are two representative …rms; a US …rm producing good Y; and a European …rm producing
good Y^ . Each …rm has monopoly power in the US and European market for its good and issues
equity claims to its dividend stream. To introduce consumer price-stickiness, we assume that …rms
set prices in local currencies before they have complete information about the state of demand in
each national market.
Consider the pricing problem facing the US …rm. The period t output of the us good is
Yt = t Kt with > 0; where Kt and t denote the current stock of …rm-speci…c capital and the
state of productivity. This output can be costlessly transported to meet demand in the US and
European market or used to augment the existing capital stock. Let Ptus and P^tus denote the
period t dollar and euro retail prices for the us good. Given the form of household preferences,
the US and European demands for the us good are given by (Ptus =Pt ) Ct and (P^tus =P^t ) C^t where
Ct and C^t denote aggregate US and European consumption. We assume that prices are chosen to
maximize the real value of the …rm’s dividend stream. If the total number of outstanding shares is
normalized to unity, the pricing problem facing the US …rm is

1
X
Qus us
t = max Et t+i;t (Dt+i =Pt+i ) (4)
Ptus ;P^tus i=0

subject to

5
 Dt =Pt = (Ptus =Pt )1 Ct + (St P^t =Pt )(P^tus =P^t )1 C^t ; and (5)

Kt+1 = (1 %)Kt + t Kt (Ptus =Pt ) Ct (P^tus =P^t ) C^t : (6)

where Eus
t denotes the …rm’s expectations conditioned period-t information. t+i;t is the stochastic
discount factor between t and t + i that the …rm uses to value the stream of real dividends. Firms
cannot hold …nancial assets or claims, so real dividends, Dt =Pt ; must equal the the sum of US
and European sales measured in terms of US aggregate consumption as shown in (5). Equation
(6) describes capital accumulation with depreciation rate % > 0:2 Notice that the …rm faces three
(potential) sources of uncertainty when choosing period t prices: uncertainty about aggregate
consumption, Ct and C^t ; the aggregate price levels, Pt and P^t ; and the spot exchange rate, St : The
European …rm producing the eu good faces an analogous problem in choosing prices, Pteu and P^teu :
The Federal Reserve (FED) and European Central Bank (ECB) play a simple role in our model.
Both central banks set one period nominal interest rates so as to achieve a target level for their
^ t are set at the beginning of period
national money supplies. Speci…cally, we assume that Rt and R
t such that
mt = Efed
t mt ; and ^ t = Eecb
m t m ^t
R 1=2 R1
where mt ln Mh;t dh and m
^t ^
0 1=2 ln Mh;t dh are the aggregate log demands for dollars and
euros and mt and m
^ t denote the targets for the US and European log money supplies. (Hereafter
we denote aggregates by dropping the h subscript and use lowercase variables to denote natural
logs, e.g. st = ln St , ch;t ln Ch;t ; etc.). Notice that interest rates are set on the basis of the FED’s
and ECB’s expectations concerning the demand for currency, Efed ecb ^ ; rather than
t mt and Et m t

the actual demand. Insofar as central banks are unable to exactly predict the aggregate demand
for currency, because individual household demands are a function of private information, excess
demand is accommodated at the chosen interest rates.

1.2 Foreign Exchange Dealers

A key distinction between our model and traditional international …nance models is that the spot
exchange rate is determined as the foreign currency price quoted by dealers in the foreign exchange

2
The …rm’s problem is not well-posed if the elasticity parameter is less than one because real dividends and
future capital would be increasing functions of current relative prices.

6
market. We assume that there are d dealers (indexed by d) who act as market-makers in the spot
market for foreign currency. As such, each dealer quotes prices at which they stand ready to buy
or sell foreign currency to households and other dealers.3 Each dealer also has the opportunity to
initiate transactions with other dealers at the prices they quote. We now described the decision
problem facing a typical dealer in detail.
For simplicity, we assume that all dealers are located in the US. The preferences of dealer d are
given by:
1
X 1
Udt Edt i 1
1 Cd;t+i ; (7)
i=0

where Edt denotes expectations conditioned on the dealer’s period t information, d;t , and Cd;t
represents the dealers consumption of the 2 goods aggregated via the CES function shown in (2).
Dealers have the same preferences as US households except that real balances have no utility value.
As a consequence, they will not hold currency in equilibrium –a feature that proves useful in the
deriving equations for the equilibrium exchange rate below. We assume that dealers are prohibited
from holding equities for the same reason.
Trading in period t is split into two rounds. In round i, dealers quote prices at which they are
willing to trade with households. In round ii, dealers quote prices at which they will trade with
other dealers and they initiate trades against other dealer’s quotes. More speci…cally, at the start
i ; at which he is willing to buy or sell
of round i, each dealer d quotes a dollar price for euros, Sd;t
euros. These price quotes are publicly observed and good for any quantity of euro (i.e. there is
no bid-ask spread). Each dealer then receives orders for euros from a subset of households. We
i : Household orders
denote the net household order to purchase euros received by dealer d as Td;t
are only observed by the recipient dealer and so represent a source of private information. At the
ii : These prices, too, are good for any
start of round ii, each dealer quotes a price for euros of Sd;t
quantity and publicly observed, so that trading with multiple partners (e.g., arbitrage trades) is
feasible. Each dealer d then chooses the quantity of euros he wishes to purchase, Td;t ; (negative
values for sales) by initiating a trade with other dealers. Interdealer trading is simultaneous and,
to the extent trades are desired at a quote that is posted by multiple dealers, those trades are
divided equally among dealers posting that quote. We denote the net quantity of euros purchased

3
More precisely, the price dealers quote is for the euro bond, which can be thought of as an interest-baring euro
deposit account.

7
 ii : After round ii trading is
from dealer d as a result of the trades initiated by other dealers by Td;t
complete, dealers make their period-t consumption decisions.
^ i denote dealer d0 s holdings of dollar and euro bonds at the start of round i
i and B
Let Bd;t d;t

trading in period t: At the end of round i trading, the dealer’s bond holdings are

ii
Bd;t i
= Bd;t i
+ Sd;t i
Td;t ; and ^ ii = B i
B i
Td;t ; (8)
d;t d;t

i is the price quoted by dealer d; and T i are the incoming household orders to purchase
where Sd;t d;t
ii ; receives incoming order for euros of T ii and initiates euro
euros. In round ii, dealer d quotes Sd;t d;t

purchases of Td;t at the price of Stii ; the price quoted by other dealers. (In equilibrium all dealers
quote the same price so we need not worry about the identity of the other dealers.) To …nance his
desired basket of consumption goods, dealer d then exchanges US bonds worth Pt Cd;t for dollars at
the US central bank, and makes his consumption purchases in the US markets for the two goods.
The dealer’s bond holdings at the start of period t + 1 are therefore given by

^i
B ^ ^ ii ii
d;t+1 = Rt (Bd;t + Td;t Td;t ); and
i ii ii ii
Bd;t+1 = Rt (Bd;t + Sd;t Td;t Stii Td;t Pt Cd;t ): (9)

i ; that
The problem facing dealer d at the start of round i is to choose the price quote, Sd;t
maximizes Udt based on current information, i ;
d;t subject to (8) and (9). By assumption, all
i cannot be conditioned on the quotes of
dealers choose quotes simultaneously, so the choice of Sd;t
i for n 6= d: At the start of round ii, dealer d faces the analogous problem of
other dealers, i.e., Sn;t
ii that maximizes Ud based on
choosing Sd;t ii ; subject to (9). After all the dealers have quoted their
t d;t

round ii prices, dealer d must determine his interdealer euro order, Td;t , to maximize Udt based on
ii ii gD
and fSd;t
d;t d=1 subject to (9). Once again, the choice of Td;t cannot be conditioned on incoming
ii ; because interdealer trading takes place simultaneously. After
euro orders from other dealers, Td;t
round ii trading is complete, dealer d then chooses his consumption of the US and EU goods,
Cd;t (us) and Cd;t (eu), to maximize Udt based on current information and the sequence of future
constraints in (8) and (9).

8
1.3 The Equilibrium Exchange Rate

An equilibrium in this model is described by a set of: (i) market-clearing equity prices, (ii) consump-
tion and portfolio rules that maximize the expected utility of households, (iii) local currency pricing
rules for …rms that maximize the value of their dividend streams, (iv) optimizing quote, trade and
consumption rules for dealers, and (v) interest rates consistent with both central banks monetary
targets. To characterize this equilibrium, we need to specify how market clearing is achieved in
the equity markets and how the information used in decision-making di¤ers across agents. For this
purpose, we make the following assumptions:

A1 Households within each county have the same information.

A2 Households cannot hold the equity issued by foreign …rms.

Assumption A1 rules out intranational di¤erences in the information available to individual

households. It does not rule out di¤erences between the information available to dealers, and
b to identify
households, or between households in di¤erent countries. We use the index h and h
a representative US and European household and denote their common information sets at the
h; b
h
start of period t by t and t respectively. With this simpli…cation, we can use the currency
orders of representative US and European households to describe how information concerning the
macroeconomy is transmitted to the exchange rate. Trade in the equity markets is ruled out by A1
and A2. Taken together, these assumptions imply that all the equities issued by US and European
…rms are held the domestic representative household.4 As a result, the market clearing real price
of US equity, Qt =Pt ; must equal the value of Qus
t Dt =Pt under an optimal period t pricing policy
where t+i;t is the discount factor of US households.5 The market clearing price of European equity,
^ t =P^t ; is analogously identi…ed from the solution to the European …rm’s pricing problem. Notice
Q
that all other goods and asset prices are set by either …rms, central banks or dealers.
Let us now focus on the determination of the equilibrium exchange rate. For this purpose we
must consider the optimal choice of dealers’ quotes in the two rounds of trading. As in Lyons

4
Obviously, this implication of A1 and A2 is at odds with the degree of international …nancial integration we
observe in world equity markets. We use it here to avoid having to model market-making activity in both currency
and equity markets – an extension we leave for future research.
5
Note that Qt =Pt is the ex-dividend real price of us equity in period t; while Qus
t is the period t present value of
current and future real dividends valued using the us household’s stochastic discount factor. Hence Qus t = Qt =Pt +
Dt =Pt :

9
(1997), our trading environment constitutes a game played over two trading rounds each period
by the d dealers. As such, we identify optimal dealer quotes and trades by the Perfect Bayesian
Equilibrium (PBE) strategies. The resulting quotes for dealer d are given by

i ii d
Sd;t = Sd;t = St = F( t ); (10)

where d = \d i is the information set common to all dealers at the beginning of round i in
t d;t

period t:
Equation (10) shows that optimal quotes have three features: First, each dealer quotes the
same prices in rounds i and ii. Second, quotes are common across all dealers. Third, all quotes
are a function, F(:); of common information at the start of period t; d: The intuition behind
t

these features is straightforward: Recall that round ii quotes are available to all dealers, are good
for any amounts, and that each dealer can initiate trades with multiple counterparties. Under
these conditions, any dealer quoting a di¤erent price from Stii would expose himself to arbitrage.
A similar argument applies to the round i quotes. Again, these quotes are publicly observed and
households are free to place orders with several dealers. Consequently, all dealers must quote the
same prices to avoid arbitrage trading losses. Dealers must also have an incentive to …ll their share
of incoming orders at the quoted common price (i.e., they must be willing to participate in round
i). This rules out di¤erences between the round i and round ii common quote. Finally, recall that
quotes must be chosen simultaneously at the beginning of each trading round. As such, round i
quotes will only be common across all dealers if they depend on common dealer information, d:
t

Dealers may posses private information at the start of period t, but they cannot use it in their
choice of quote without exposing themselves to arbitrage losses.
The relationship between the common period t quote, St , and dealers’common information,
d; implied by the PBE of our model is identi…ed in the following proposition:
t

Proposition 1 The log spot rate implied by the PBE quote strategies of dealers in period t is

1
X i
st = 1
1+ Edt 1+ ft+i ; (11)
i=0

10
where is a positive constant and Edt denotes expectations conditioned on dealers’common period-t
information, d: ft denotes exchange rate fundamentals, which are de…ned as
t

ft c^t ct + mt m
^ t + "t (12)

where "t ln(St P^t =Pt ) is the log real exchange rate and is a risk premium.

The Appendix provides a detailed derivation of these equations from the log linearized equilib-
rium conditions as well as the results reported in the propositions that follow. Here, we provide
some intuition. In the equilibrium of our model, dealers must be willing to …ll incoming orders
for euros at the price they quote. This means that the period-t quote must be set such that the
expected excess return on euros between t and t+1 compensates the dealers for the risk of …lling in-
coming currency orders during period t: In other words, all dealers must quote a price, St exp(st );
such that
Edt st+1 + r^t rt = ; (13)

where st+1 st+1 st and is the risk premium that depends on the conditional second moments
of dealers’ marginal utility of wealth and the future spot rate:6 Notice that Edt st+1 + r^t rt
will di¤er from the expectations of (log) excess returns held by an individual dealer d when he
has private information about the future spot rate (i.e., Edt st+1 6= Edt st+1 ): Individual dealers
use this private information when making the round ii trading decisions, not when choosing St .
Proposition 1 follows easily from (13) and the implications of money market clearing. In particular,
our speci…cation for household preferences implies that the expected demand for dollars conditioned
on d is approximately Edt mt = $ + pt + Edt ct rt : The expected demand for euros is similarly
t

approximated by Edt m
^ t = $ + p^t + Edt c^t r^t : Under the reasonable assumption that central banks
expectations concerning aggregate money demand are at least as precise as expectations based
on d; Edt mt = Edt mt and Edt m
^ t = Edt m
^ t by the law of iterated expectations. Combining these
t

expressions with (13) gives us the equations in Proposition 1.

Equation (11) plays a central role in our analysis. It shows that the log price of euros quoted
by all dealers is equal to the present value of fundamentals, ft : There are two noteworthy di¤er-

6
For the sake of clarity, we shall take this risk premium to be constant in the analysis that follows. Allowing for
time-variation does not a¤ect the focus of our analysis.

11
ences between this speci…cation and the exchange rate equations found in traditional monetary
models. First, the de…nition of fundamentals in (12) includes the di¤erence between foreign and
home consumption rather than income. This arises because household preferences imply that the
demand for national currencies depends on consumption rather than income. Second, equation (11)
shows that fundamentals a¤ect the spot rate only via dealers’expectations. This is a particularly
important feature of the model: Since the current spot rate is simply the common price of euros
quoted by dealers before trading starts, it must only be a function of information that is common
to all dealers at the time, d. This means that exchange rate dynamics in our model are driven by
t

the evolution of dealers’common information.

To further emphasize the importance of dealers’ information, it is useful to consider the im-
plications of (11) for the rate of depreciation, st+1 . Speci…cally, if we iterate (11) forward to
get st = Edt ft + Edt st+1 ; and rearrange, we can write the depreciation rate implied by the PBE
quotes as
st+1 = 1 (st Edt ft ) + et+1 ; (14)

X1 i
where et+1 1
(Edt+1 Edt )ft+i+1 : (15)
1+ i=0 1+

Equation (14) shows that the evolution of dealers’ information can a¤ect the depreciation rate
through two channels: First, it can a¤ect the di¤erence between the current spot rate and dealers’
estimate of current fundamentals, st Edt ft . Second, it can lead to revisions in dealers’common
knowledge forecasts of future fundamentals, (Edt+1 Edt )ft+i+1 for i 0; which as (15) shows,
contribute to dealer errors in forecasting next period’s spot rate, et+1 st+1 Edt st+1 : Since the
…rst term in (14) is multiplied by the reciprocal of the semi-interest elasticity of money demand,
1= ; a small number, the second channel is more likely to be empirically relevant. Indeed, because
depreciation rates are very hard to forecast over short time periods, any attempt to make progress on
understanding the origins of high-frequency spot rate dynamics must focus on the second channel.7
This is exactly the strategy of this paper. Speci…cally, our aim is to investigate whether transaction
‡ows in the foreign exchange market convey information about fundamentals to dealers that they

7
This point holds outside the context of our speci…c model. Engel and West (2005) note that forecasting the
depreciation rate implied by several standard models will be hard because the value of the coe¢ cient in the present
value representation of the equilibrium exchange rate is very large. Thus, the lack of forecastability does not, in itself,
imply that spot exchange rates are disconnected from fundamentals (see, also, Evans and Lyons 2005).

12
then incorporate into their price quotes. In other words, we ask: Do transaction ‡ows act as a
proximate driver of spot exchange rates because they convey information that leads to revisions in
dealers’forecasts of fundamentals, (Edt+1 Edt )ft+i+1 ?
Before we address this question in detail, it proves useful to have an overview of how information
contained in customer orders becomes incorporated into the equilibrium spot rate. Recall that the
i ; represent private information to the dealer. In our
customer orders received by each dealer d; Td;t
model, the PBE strategy for each dealer is to use this information when initiating trades with other
dealers (i.e., when choosing Td;t ). As a result, interdealer trading in round ii e¤ectively aggregates
the information contained in customer orders received by dealers across the market. Indeed, it is
the information conveyed by interdealer trading that augments dealer’s common information by
the start of period t + 1; and hence a¤ects dealers’ PBE choice for st+1 . This does not mean
that dealers necessarily have complete information about the current fundamentals by the end of
interdealer trading. As the model of Evans and Lyons (2004) shows, they will under some special
circumstances, but in general the inference problem facing dealers is too complex for them to make
precise inferences about current fundamentals from their observations of interdealer trading. We
will have more to say about dealers’assimilation of information below.
Finally, a few comments about the structure of the model are in order. Our speci…cation
for the household and production sectors deliberately does not include many of the features to
be found in recent two-country general equilibrium models. Our aim, instead, is to present a
minimal speci…cation that provides microfoundations for the key macroeconomic factors that a¤ect
the behavior of the spot exchange rate. These are: (i) household demands for foreign currency
motivated by optimal portfolio choice, and (ii) pricing decisions by …rms that imply variations
in the real exchange rate. While richer speci…cations for preferences and the production sector
would clearly improve the empirical relevance of the model along many dimensions, they would not
qualitatively a¤ect the links between exchange rates, fundamentals and transaction ‡ows which are
the focus of this paper.

13
2 Fundamentals and Order Flow

We now examine the link between transaction ‡ows, fundamentals and the spot exchange rate. More
speci…cally, our aim is to identify the conditions under which the customer order ‡ows reaching
i ; convey new information about fundamentals that dealers incorporate into their price
dealers, Td;t
quotes for euros. We proceed in two steps. First we identify the factors driving customer order
‡ows. Second, we show why order ‡ows may convey information about fundamentals.

2.1 Customer Order Flow

Let xt denote aggregate customer order ‡ow de…ned as the dollar value of aggregate household
purchases of euros from dealers during period t trading. The contribution of US households to this
^h;t
order ‡ow is St (B ^h;t
B 1) = ^
t Wh;t Rt
^h;t
St B 1 where t denotes the desired share of euro
^hb;t
bonds in the US households’wealth. Similarly, European households contribute St (B ^hb;t
B 1) =
^ hb;t R
^ t St W ^t ^hb;t
St B 1 where ^ t is the desired share of euro bonds in European wealth. Market
clearing requires that aggregate holdings of euro bonds by households and non-households (i.e.,
^t
central banks and dealers) sum to zero, so that B 1
^hb;t
+B 1
^hb;t
+B 1
^ denotes the
= 0 where B
aggregate holdings of non-households. Hence, aggregate order ‡ow can be written as

xt = [ t `t + ^ t (1 ^ t + St B
`t )] Wt R ^t 1; (16)

where Wt ^ hb;t is world household wealth in dollars, and `t

Wh,t + St W Wh,t =Wt : Thus, order ‡ow
depends upon the portfolio allocation decisions of US and European households (via t; and ^ t ),
the level and international distribution of household wealth (via Wt and `t ) and the outstanding
^t
stock of foreign bonds held by non-households from last period’s trading, B 1: These elements
imply that order ‡ow contains both pre-determined (backward-looking) and non-predetermined
(forward-looking) components. The former include the level and distribution of wealth, the latter
are given by the portfolio shares because they depend on households’ forecasts of future returns.
We formalize these observations in the following proposition.

14
Proposition 2 The utility-maximizing choice of portfolios by US and European households im-
plies that aggregate order ‡ow may be approximated by

xt = rEht st+1 + ^ rEhtb st+1 + ot ; (17)

with ; ^ > 0; where rEt! st+1 Et! st+1 Etd st+1 for ! = fh,h
bg and ot denotes terms involving the
distribution of wealth, non-household bond holdings, and the consumption of European households.

Equation (17) describes the second important implication of our model. It relates order ‡ow to
the di¤erence between households’ forecasts for the future spot rate, Et! st+1 for ! = fh,h
bg; and
dealers’ forecasts, Etd st+1 : In particular, there will be positive order ‡ow for euros if households
are more optimistic about the future value of the euro than dealers, so that rEt! st+1 > 0 for
bg:
! = fh,h
To understand why di¤erences in expectations play this role, we need to focus on how households
choose their portfolios. In the appendix we show that the optimal share of US household wealth
held in the form of euro bonds is increasing in the expected log excess return, Eht st+1 + r^t rt :
Now, when dealers’ foreign currency quotes satisfy (11) and (12), the log spot rate also satis…es
Edt st+1 + r^t rt = : We can therefore write the excess return on European bonds expected by
US households as

Eht st+1 + r^t rt = Edt st+1 + r^t rt + rEht st+1 = rEht st+1 + :

Thus, when US households are more optimistic about the future value of the euro than dealers,
they expect a higher excess return on euro bonds. These expectations, in turn, increase the desired
fraction of US household wealth in euro bonds, so US households place more orders for euros with
dealers in round i of period t trading. Optimism concerning the value of the euro on the part of
European households (i.e. rEhtb st+1 > 0) contributes positively to order ‡ow in a similar manner.
Of course household portfolio choices are also a¤ected by risk. The ot variable in (17) summarizes
the e¤ects of risk, the distribution of wealth and non-household bond holdings. These terms will
not vary signi…cantly from month to month or quarter to quarter under most circumstances, and so
will not be the prime focus of the analysis below. We shall concentrate instead on how the existence

15
of dispersed information, manifest through the existence of the forecast di¤erentials, rEht st+1 and
rEhtb st+1 ; a¤ects the joint behavior of order ‡ow, spot rates and fundamentals.

2.2 How is Order Flow Related to Fundamentals?

To address this question, we …rst characterize the equilibrium dynamics of fundamentals. Let yt
denote the vector that describes the state of the economy at the start of period t: This vector
includes the variables that comprise fundamentals (i.e. consumption, money targets and the real
exchange rate) as well as those variables needed to describe …rms’behavior, and the distribution
of wealth across households and dealers. In Evans and Lyons (2004), we describe in detail the
equilibrium dynamics of a model with a similar structure. Here our focus is on the empirical
implications of the model, so we present the equilibrium dynamics in reduced form:

yt+1 = A yt + ut+1 ; (18)

where yt yt yt 1 with ut+1 a vector of mean zero shocks. This speci…cation for the equilibrium
dynamics of the state variables is completely general, yet it allows us to examine the link between
order ‡ow and fundamentals in a straightforward way.
We start with the behavior of the spot exchange rate. Let fundamentals be a linear combination
of the elements in the state vector: ft = Cyt : When dealers quote spot rates according to (11) in
Proposition 1, and (18) describes the dynamics of the state vector yt ; the spot exchange rate can
be written as
st = Edt yt ; (19)

where yt0 [yt0 ; yt0 ] and C{1 + 1+ C(I 1+ A) 1 A{

2, with yt = {1 yt and yt = {2 yt : is a
vector that relates the log spot rate to dealers’current estimate of the state vector yt : We can now
write the US forecast di¤erential as:

rEht st+1 = Eht Edt+1 yt+1 Edt Edt+1 yt+1

= Eht Edt+1 yt+1 Edt yt+1 : (20)

Suppose that US households collectively know as much about the state of the economy as dealers

16
do. Under these circumstances, the right hand side of (20) is equal to Eht Edt+1 Edt yt+1 : In
other words, the forecast di¤erential for the future spot rate depends on households’expectations
regarding how dealers revise their estimates of the future state, yt+1 : As one might expect, this
di¤erence depends on the information sets, h and d: Clearly, if h = d; then Eht (Edt+1 Edt )yt+1
t t t t

must equal a vector of zeros because (Edt+1 Edt )yt+1 must be a function of information that is
not in d: Alternatively, suppose that households collectively have superior information so that
t
h = f d;
t t tg for some vector of variables t: If dealers update their estimates of yt+1 using
elements of t, then some elements of (Edt+1 Edt )yt+1 will be forecastable based on h:
t

We formalize these ideas in the following proposition.

Proposition 3 If US and European households are as well-informed about the state of the economy
d h d b
h
as dealers, so that t t and t t; then US and European forecast di¤ erentials for spot
rates are

rEht st+1 = (Eht yt+1 Edt yt+1 ); (21a)

rEhtb st+1 = ^ (Ehtb yt+1 Edt yt+1 ); (21b)

and order ‡ow follows

xt = rEht yt+1 + ^ ^ rEhtb yt+1 + ot : (22)

for some matrices, and ^ :

The intuition behind Proposition 3 is straightforward. If US households are collectively as

well-informed about the future state of the economy as dealers, then rEht st+1 = Eht (Edt+1
Edt )yt+1 ; so the forecast di¤erential depends on the speed at which US household expect dealers
to assimilate new information concerning the future state of the economy. We term this the pace
of information aggregation. If dealers learn nothing new about yt+1 during period t trading,
Edt+1 yt+1 = Edt yt+1 : Hence, if US households expect that period t trading will reveal nothing
new to dealers, Eht Edt+1 Edt yt+1 = 0 and there is no di¤erence between dealer and household
forecasts of future spot rates. Under these circumstances, there is no information aggregation
so and ^ are equal to null matrices. Alternatively, if households expect dealers to assimilate
information from period t trading, the forecast di¤erentials for spot rates will be non-zero. In

17
the extreme case where period-t trading is su¢ ciently informative to reveal to dealers all that
households know about the future state of the economy, (Edt+1 Edt )yt+1 will equal E!t yt+1 Edt yt+1
bg: In this case, information aggregates quickly, so
for ! = fh,h and ^ equal the identity matrices.
Under other circumstances where the pace of information aggregation is slower, the and ^ matrices
will have many non-zero elements. (Exact expressions for and ^ are provided in the Appendix.)
Equation (22) combines (17) from Proposition 2 with (21). This equation expresses order ‡ow
in terms of forecast di¤erentials for the future state of the economy and the speed of information
aggregation. Since fundamentals represent a combination of the elements in yt ; (22) also serves to
link dispersed information regarding future fundamentals to order ‡ow. In particular, if households
have more information about the future course of fundamentals than dealers, and dealers are
expected to assimilate at least some of this information from transaction ‡ows each period, order
‡ow will be correlated with variations in the forecast di¤erentials for fundamentals.
We should emphasize that the household currency orders driving order ‡ow in this model
are driven solely by the desire to optimally adjust portfolios. Households have no desire to inform
dealers about the future state of the economy, so the information conveyed to dealers via transaction
‡ows occur as a by-product of their dynamic portfolio allocation decisions. The transaction ‡ows
associated with these decisions establish the link between order ‡ow, dispersed information, and
the speed of information shown in equation (22).
One aspect of our model deserves further clari…cation. Our model abstracts from informational
h; b
h
heterogeneity at the household level, so t and t represent the information sets of the repre-
sentative US and European households. This means that the results in Proposition 3 are derived
under the assumption that representative households have strictly more information than dealers
d h d b
h
( t t and t t ): Clearly this is a strong assumption. Taken literally, it implies that every
household knows more about the current and future state of the economy than any given dealer.
Fortunately, our central results do not rely on this literal interpretation. To see why, suppose, for
example, that each household receives its own money demand shock and is thereby privately mo-
tived to trade foreign exchange. In this setting, no household would consider itself to have superior
information. But the aggregate of those realized household trades would in fact convey information
about the average household shock, i.e., the state of the macroeconomy. For the sake of parsimony,
we have not modelled heterogeneity at the US and European household levels. Instead, we assume

18
that households in any given country share the same information about the macroeconomy. Ex-
tending the model to capture heterogeneity is a natural extension, but not one that would alter the
main implications of our model that are the focus of the empirical analysis below.8

3 Data

Our empirical analysis utilizes a new data set that comprises end-user transaction ‡ows, spot
rates and macro fundamentals over six and a half years. The transaction ‡ow data di¤ers in two
important respects from the data used in earlier work (e.g., Evans and Lyons 2002a,b). First, they
cover a much longer time period; January 1993 to June 1999. Second, they come from transactions
between end-users and a large bank, rather than from inter-bank transactions. Our data covers
transactions with three end-user segments: non-…nancial corporations, investors (such as mutual
funds and pension funds), and leveraged traders (such as hedge funds and proprietary traders).
The data set also contains information on trading location. From this we construct order ‡ows
for six segments: trades executed in the US and non-US for non-…nancial …rms, investors, and
leveraged traders. Though inter-bank transactions account for about two-thirds of total volume in
major currency markets at the time, they are largely derivative of the underlying shifts in end-user
currency demands. Our data include all the end-user trades with Citibank in the largest spot
market, the USD/EUR market, and the USD/EUR forward market.9 Citibank had the largest
share of the end-user market in these currencies at the time, ranging between 10 and 15 percent.
The ‡ow data are aggregated at the daily frequency and measure in $m the imbalance between
end-user orders to purchase and sell euros.
There are many advantages of our transaction ‡ow data. First, the data are simply more
powerful, covering a much longer time span. Second, because the underlying trades re‡ect the
world economy’s primitive currency demands, the data provide a bridge to modern macro analysis.
Third, the three segments span the full set of underlying demand types. We shall see that those not

8
As is standard in literature, we use “households” as a metaphor for a wide class of agents that constitute the
private sector. In particular, households represent the class of non-dealer agents that observe some component of
macro fundamentals. One way to introduce heterogeneity would be to di¤erential between the information available
to di¤erent members of this class, e.g., …nancial institutions and individuals.
9
Before January 1999, data for the Euro are synthesized from data in the underlying markets against the Dollar,
using weights of the underlying currencies in the Euro.

19
covered by extant end-user data sets are empirically important for exchange rate determination.10
Fourth, because the data are disaggregated into segments, we can address whether the behavior
of the individual segments is similar, and whether they convey the same information concerning
exchange rates and macro fundamentals.
Our empirical analysis also utilizes new high-frequency real-time estimates of macro variables
for the US and Germany: speci…cally GDP, consumer prices, and M1 money. As the name implies,
a real-time estimate of a variable is the estimated value based on public information available on a
particular date. These estimates are conceptually distinct from the values that make up standard
macro time-series. Importantly, because they are computed from information available to market
participants contemporaneously, real-time estimates are relevant for understanding the link between
the foreign exchange market (or any other …nancial market) and the macroeconomy.
A simple example clari…es the di¤erence between a real-time estimate of a macro variable
and the data series usually employed in empirical studies. Let { denote a variable representing
macroeconomic activity during month ; that ends on day m( ), with value {m( ) . Data on the
value of { is released on day r( ) after the end of month with a reporting lag of r( ) m( )
days. Reporting lags vary from month to month because data is collected on a calendar basis, but
releases issued by statistical agencies are not made on holidays and weekends. (For quarterly series,
such as GDP, reporting lags can be as long as several months.) The real-time estimate of { on day
t in month is the expected value of {m( ) based on day t information: Formally, the real-time
estimate of a monthly series { is

{m( )jt E[{m( ) j t] for m( 1) < t m( ); (23)

where t denotes an information set that only contains data known at the start of day t: In the
case of a quarterly series like GDP, the real-time estimate on day t is

{q(i)jt E[{q(i) j t] for q(i 1) < t q(i); (24)

where q(i) denotes the last day of quarter i:

10
Froot and Ramadorai (2002), consider the transactions ‡ows associated with portfolio changes undertaken by
institutional investors. Osler (2003) examines end-user stop-loss orders.

20
 Real-time estimates are conceptually distinct from the values for {m( ) or {q(i) found in standard
macro time series. To see why, let v( ) denote the last day on which data on { for month was
revised. A standard monthly time series for variable { spanning months = 1; ::T comprises the
sequence {{m( T 11
)jv( ) g =1 . This latest vintage of the data series incorporates information about
the value of { that was not known during month . We can see this more clearly by writing the
di¤erence between {m( )jv( ) and real-time estimate as

{m( )jv( ) {m( )jt = {m( )jv( ) {m( )jr( ) + {m( )jr( ) {m( )jm( ) + {m( )jm( ) {m( )jt :
(25)
The …rst term on the right hand side represents the e¤ects of data revisions following the initial data
release. We denote the value for {m( ) released on day r( ) by {m( )jr( ) so {m( )jv( ) {m( )jr( )

identi…es the e¤ects of all the data revisions between r( ) and v( ): Croushore and Stark (2001),
Faust, Rogers, and Wright (2003) and others have emphasized that these revisions are signi…cant
for many series. The second term in (25) is the di¤erence between the value for {m( ) released on
day r( ) and the real-time estimate of {m( ) at the end of the month. This term identi…es the
impact of information concerning {m( ) collected by the statistical agency before the release date
that was not part of the m( ) information set. This term is particularly important in the case of
quarterly data where the reporting lag can be several months. The third term on the right of (25)
is the di¤erence between the real time estimate of {m( ) at the end of month and the estimate on
a day earlier in the month.
In this paper we construct real time estimates of GDP, consumer prices, and M1 for the US
and Germany using an information set based on 35 macro data series. For the US estimates our
speci…cation for t includes the 3 quarterly releases on US GDP and the monthly releases on 18
other US macro variables. The German real-time estimates are computed using a speci…cation for

t that includes the 3 quarterly release on German GDP and the monthly releases on 8 German
macro variables. All series come from a database maintained by Money Market News Services that
contains details of each data release. We use the method developed in Evans (2005) to compute
the real-time estimates. Specially, for each variable { we use the Kalman Filter to calculate the

11
For the sake of notational clarity, we have implicitly assumed that the statistical agency uses the t information
set when computing data revisions. Relaxing this assumption to give the agency superior information does not a¤ect
the substance of our argument. For a further discussion, see Evans (2005).

21
conditional expectations in (23) and (24) from estimates of a state space model that speci…es a
daily time series process for {t and its relation to the sequence of data releases (i.e. the elements
of t ): The Appendix provides an overview of the state space model and the estimation method.
Our real time estimates have several important attributes. First our speci…cation insures that
the information set used to compute each real-time estimate, t; is subset of the information
available to participants in the foreign exchange market on day t: This means that the real-time
estimate of monthly variable {; {m( )jt ; can be legitimately used as a variable a¤ecting market
actively on day t: By contrast, the values for {m( ) found in either the …rst or …nal vintage of a
time series (i.e., {m( )jr( ) or {m( )jv( ) ) contain information that was not known to participants on
day t:
The second attribute of the real-time estimates concerns the frequency with which macro data
is collected and released. Even though the macro variables are computed on a quarterly (GDP) or
monthly (prices and money) basis, real-time estimates vary day-by-day as the ‡ow of macro data
releases augments the information set t: This attribute is illustrated in Figure 1, where we plot
the real-time estimates of log GDP for the US and Germany. The real-time estimates (shown by
the solid plots) clearly display a much greater degree of volatility than the cumulant of the data
releases (shown by the dashed plots). This volatility re‡ects how inferences about current GDP
change as information arrives in the form of monthly data releases during the current quarter and
GDP releases referring to the previous quarter. A further noteworthy feature of Figure 1 concerns
the di¤erence between the real-time estimates and the ex post value of log GDP represented by the
vertical gap be the solid and dashed plots. This gap should be small if the current level of GDP
could be precisely inferred from contemporaneously available information. However, as the …gure
clearly shows, there are many occasions where the real-time estimates are substantially di¤erent
from the ex post values.
A third attribute of the real-time estimates concerns their variation over our sample period.
Although our data covers only six and a half years, Figure 1 shows that there is considerable
variation in our GDP measures within this relatively short time span. The vertical axis shows that
real-time estimates of US GDP have a range of approximately 2.4 percent around trend, while the
range for German GDP is more than 4.5 percent.
Figure 2 displays the variation in the other real-time estimates. The left hand panel shows that

22
Figure 1: Real-time estimates of log GDP (solid line) and cumulant of GDP releases (dashed
line). The right hand panel shows plots for US GDP, the left panel plots for German GDP. All
series are detrended and multiplied by 100.

while the real-time estimates of US prices varied very little from their trend, German prices varied
by almost 3 percent. In the right hand panel the real-time estimates for M1 have a range of almost
16 percent in the US and 7 percent in Germany. Because the reporting lag for both prices and
money are much shorter than that for GDP, the di¤erences between these real-time estimates and
the ex-post values are much smaller than those shown in Figure 1. (We omit ex-post values from
Figure 2 for clarity.) Real-time uncertainty about current consumer prices and M1 is far less than
the degree of uncertainty surrounding GDP.
In sum, all but one of the real-time estimates varies signi…cantly over our sample period. This is
important if we want to study how macroeconomic conditions a¤ect the foreign exchange market. If
all of our real-time estimates were essentially constant over our sample, there would be no room for
detecting how perceived developments in the macroeconomy are re‡ected in the foreign exchange
market.
In the analysis that follows we consider the joint behavior of exchange rates, order ‡ows and
the real-time estimates of macro variables at the weekly frequency. This approach provides more
precision in our statistical inferences concerning the high frequency link between ‡ows, exchange
rates and macro variables than would be otherwise possible. The weekly timing of the variables
is as follows: We take the log spot rate at the start of week t; st ; to be the log of the o¤er rate
(USD/EUR) quoted by Citibank at the end of trading on Friday of week t 1 (approximately

23
Figure 2: Left hand panel: Real-time estimates of log US consumer prices (solid line) and
Germany consumer prices (dashed line). Right hand panel: Real-time estimates of US M1 (solid
line) and German M1 (dashed line). All series are detrended.

17:00 GMT). This is also the point at which we sample the week t interest rates from Datastream.
The week-t ‡ow from segment j; xj;t ; is computed as the total value in $m of dollar purchases
initiated by the segment against Citibank’s quotes between the 17:00 GMT on Friday of week
t 1 and Friday of week t: Positive values for these order ‡ows therefore denote net demand for
euros by the end-user segment: The week t change in the real-time estimates are computed as
the di¤erence between the Friday estimates on weeks t 1 and t 2: This timing insures that the
week-t change in the real-time estimates are derived using a subset of the information available to
foreign exchange dealers when quoting spot rates at the start of week-t trading. In other words,
our timing assumptions insure that the information used to compute {m( )jt or {q(i)jt is a subset of
the information available to all dealers when quoting the spot rate st :12
Summary statistics for the weekly data are reported in Table 1. The statistics in panel A show
that weekly changes in the log spot rate, st st st 1; have a mean very close to zero and
display no signi…cant serial correlation. These statistics are typical for spot exchange rates and
suggest that the univariate process for st is well-characterized by a random walk. Two features
stand out from the statistics on the six ‡ow segments shown in Panel B. First, the order ‡ows are
large and volatile. Second, they display no signi…cant serial correlation. At the weekly frequency,

12
More precisely, our timing assumptions imply that the real-time estimates of {m( )jt or {q(i)jt incorporate macro
data releases that are only a few hours old by the time dealers quote st :

24
the end-user ‡ows appear to represent shocks to the foreign exchange market arriving at Citibank.
This is not to say that ‡ows are unrelated across segments. The (unreported) cross-correlations
between the six ‡ows range from approximately -0.16 to 0.16, but cross-autocorrelations are all
close to zero.
Summary statistics for the weekly changes in the real-time estimates are reported in Panel C
of Table 1. The most notable feature of these statistics concerns the estimated autocorrelations.
These are generally small and insigni…cant at the 5% level except in the case of the M1 real-time
estimates. For perspective on these …ndings, consider the weekly change in the monthly series {:
If the weekly change falls within a single month, the change in real-time estimate is

{m( )jw(j) {m( )jw(j-1) E[{m( ) j w(j) ] E[{m( ) j w(j 1) ];

where w(j) denotes the last day of week j: In this case the weekly change simply captures the ‡ow
of new information concerning the value of { in the current month, {m( ) ; and so should not be
correlated with any elements of w(j 1) ; including past changes in the real-time estimates. If the
weekly change occurs at the end of the month, the change in the real-time estimate can be written
as

{m( +1)jw(j) {m( )jw(j-1) = E[{m( +1) j w(j) ] E[{m( +1) j w(j 1) ] + E[{m( +1) {m( ) j w(j 1) ] :

Here the …rst term on the right hand side represents the the ‡ow of new information concerning
{m( +1) : Once again this should not be correlated with any elements in w(j 1) : The second term
identi…es initial expectations about the growth in { from month to + 1: This term is a function
of elements in w(j 1) and so may be correlated with past changes in the real-time estimates.
The autocorrelations in Table 1 are computed from all weekly changes in our sample, and
so capture the characteristics of both the within and cross-month changes. The small amounts of
positive serial correlation we see re‡ect the fact that forecasts for monthly M1 growth are positively
correlated with past growth, a feature that is evident from the plots in Figure 2. That said, the
over-arching implication of the estimated autocorrelations is that the weekly changes in each real-
time estimates primarily re‡ects the arrival of new information concerning the current state of the
corresponding macro variable. Our real-time estimates will therefore enable us to capture changing

25
 Table 1: Summary Statistics

mean max skewness Autocorrelations

Std. min kurtosis 1 2 4 8
A: Exchange Rate
(i) st (x100) -0.043 3.722 0.105 -0.061 0.027 0.025 -0.015
1.234 -3.715 3.204 (0.287) (0.603) (0.643) (0.789)
B: Order Flows
(ii) Corporate US -16.774 549.302 -0.696 -0.037 -0.040 0.028 -0.028
108.685 -529.055 9.246 (0.434) (0.608) (0.569) (0.562)
(iii) Corporate Non-US -59.784 634.918 -0.005 0.072 0.089 -0.038 0.103
196.089 -692.419 3.908 (0.223) (0.124) (0.513) (0.091)
(iv) Traders US -4.119 1710.163 0.026 -0.021 0.024 0.126 -0.009
346.296 -2024.275 8.337 (0.735) (0.602) (0.101) (0.897)
(v) Traders Non-US 11.187 972.106 0.392 -0.098 0.024 0.015 0.083
183.36 -629.139 5.86 (0.072) (0.660) (0.747) (0.140)
(vi) Investors US 19.442 535.32 -1.079 0.096 -0.024 -0.03 -0.016
146.627 -874.15 11.226 (0.085) (0.568) (0.536) (0.690)
(vii) Investors Non-US 15.85 1881.284 0.931 0.061 0.107 -0.030 -0.014
273.406 -718.895 9.253 (0.182) (0.041) (0.550) (0.825)
C: Real-Time Data
(viii) US Output -0.001 0.711 0.060 0.072 0.107 -0.015 0.058
0.201 -0.610 0.134 (0.084) (0.056) (0.788) (0.329)
(ix) US Prices 0.000 0.250 1.527 0.006 -0.034 0.091 0.004
0.030 -0.104 18.673 (0.695) (0.135) (0.142) (0.963)
(x) US Money -0.007 5.679 -0.230 0.076 0.065 0.132 0.032
1.368 -6.981 9.160 (0.003) (0.012) (0.131) (0.595)
(xi) German Output 0.002 2.840 -0.298 0.072 -0.039 -0.009 0.019
0.514 -4.087 20.437 (0.138) (0.193) (0.873) (0.671)
(xii) German Prices 0.002 4.090 0.105 0.069 0.005 0.009 -0.044
0.817 -3.988 8.632 (0.111) (0.918) (0.864) (0.444)
(xiii) German Money 0.022 7.447 1.073 0.116 0.083 0.100 0.042
1.421 -6.263 13.120 (0.000) (0.000) (0.339) (0.473)
Notes: The table reports summary statistics for the following variables sampled at the
weekly frequency between January 1993 and June 1999: (i) the weekly change in the log
spot rate x100, (ii)-(vii) order ‡ows from end-user segments cumulated over a week, and
(viii)- (xiii) weekly changes in real-time estimates measured in annual percent. The last
four columns on the right report autocorrelations i at lag i and p-values for the null that
i = 0 in parentheses.

perceptions concerning the current state of the macroeconomy rather than its actual evolution. It
is the link between the changing perceptions of market participants and the behavior of exchange

26
rate that is the focus of our empirical analysis.

4 Empirical Analysis

In this section we examine the empirical implications of Propositions 1 - 3. First, we consider the
implications of our model for the correlation between order ‡ows and changes in spot exchange
rates. Next, we examine the links between spot rates and fundamentals. Our model identi…es
conditions under which order ‡ow should have incremental forecasting power beyond spot rates.
We …nd strong empirical support for this prediction, implying that order ‡ows convey information
about macro fundamentals to the market. Finally, we investigate whether this informational role
can account for the forecasting power of order ‡ows for future changes in exchange rates.

4.1 The Order Flow/Spot Rate Correlation

Evans and Lyons (2002a,b) show that order ‡ows account for between 40 and 80 percent of the
daily variation in the spot exchange rates of major currency pairs. Propositions 1 - 3 provide a
structural interpretation of this …nding. Recall that when dealers’foreign currency quotes satisfy
(11) and (12) in Proposition 1, the log spot rate satis…es Edt st+1 + r^t rt = : Combining this
restriction with the identity st+1 Edt st+1 + st+1 Edt st+1 gives

st+1 = rt r^t + + st+1 Edt st+1 ;

= rt r^t + + Edt+1 yt+1 Edt yt+1 ; (26)

where the second line follows from the relation between the spot rate and state vector described
by equation (19). Thus, Proposition 1 implies that the rate of depreciation is equal to the interest
di¤erential, a risk premium, and the revision in dealer forecasts concerning the future state of
the economy between periods t and t + 1: This forecast revision is attributable to two possible
information sources. The …rst is public information that arrives right at the start of period t + 1,
before dealers quote st+1 . The second is information conveyed by the transaction ‡ows during
period t: It is this second information source that accounts for the correlation between order ‡ow
and spot rate changes in the data.

27
Proposition 4 When dealer quotes for the price of foreign currency satisfy (11), and order ‡ow
follows (22), the rate of depreciation can be written as

st+1 = rt r^t + + b (xt Edt xt ) + t+1 : (27)

t+1 represents the portion of Edt+1 yt+1 Edt yt+1 that is uncorrelated with order ‡ow, and b is
a projection coe¢ cient equal to

^ V rEhb yt+1 ^ 0 0
CV (yt+1 ; ot ) V (rEht yt+1 ) 0 0 t
+ + ; (28)
V(xt ) V(xt ) V(xt )

where V (:) and CV(:; :) denote the population variance and covariance:

Inspection of expression (28) reveals that the observed correlation between order ‡ow and the
rate of depreciation can arise through two channels. First, if the distribution of wealth and dealer
bond holdings a¤ect order ‡ow (via ot in equation 17) and has forecasting power for fundamentals,
order ‡ow will be correlated with the depreciation rate through the …rst term in (28). Since there
is little variation in ot from month to month or even quarter to quarter, it is unlikely that this
channel accounts for much of the order ‡ow/spot rate correlation we observe at a daily or weekly
frequency. The second channel operates through the transmission of dispersed information. If
household expectations for the future state vector di¤er from dealers’expectations, and information
aggregation accompanies trading in period t; both the second and third terms in (28) will be
positive. Notice that the depreciation rate is correlated with order ‡ow in this case not just
because households and dealers hold di¤erent expectations, but also because households expect
some of their information to be assimilated by dealers from the transaction ‡ows they observe in
period t: In this sense, the correlation between order ‡ow and the depreciation rate informs us
about both the existence of dispersed information and the pace at with information aggregation
takes place.
Now we turn to the empirical evidence. Table 2 presents the results of regressing currency
returns between the start of weeks t and t + for = f1; 4g on a constant, the interest di¤erential
at the start of week t; rt r^t ; and the order ‡ows from the six segments between the start of weeks
t and t + . These regressions are the empirical counterparts to (27) with the six ‡ows proxying for

28
 Table 2: Contemporaneous Return Regressions

Horizon Interest Corporate Traders Investors R2 2

Di¤erential US Non-US US Non-US US Non-US (p-value)

1 week
-0.2 -0.326 -1.096 0.03 12.627
(0.391) (0.584) (0.309) (0.002)
-0.193 1.018 0.63 0.094 38.139
(0.364) (0.170) (0.350) (0.000)
-0.134 1.194 1.441 0.131 30.818
(0.341) (0.576) (0.327) (0.000)
-0.297 -0.321 -0.817 0.791 0.632 1.108 1.254 0.213 88.758
(0.325) (0.535) (0.291) (0.170) (0.337) (0.572) (0.312) (0.000)
4 weeks
-0.182 -0.006 -0.340 0.058 16.101
(0.252) (0.165) (0.085) (0.000)
-0.168 0.279 0.11 0.113 23.354
(0.247) (0.061) (0.118) (0.000)
0.001 0.144 0.49 0.251 68.471
(0.204) (0.121) (0.063) (0.000)
-0.19 0.027 -0.202 0.177 0.046 0.218 0.41 0.323 109.571
(0.209) (0.138) (0.071) (0.060) (0.101) (0.119) (0.066) (0.000)
Notes: The table reports coe¢ cients and standard errors from regressions of returns measured over
one week and one month, on a constant (estimates not reported), the lagged interest di¤erential
and order ‡ows cumulated over the same horizon. The interest di¤erential is computed from the
one month rates on Euro Dollar and DM deposits. Estimated coe¢ cients on the order ‡ows are
multiplied by 1000. The right hand column reports 2 statistics for the null that all the coe¢ cients
on order ‡ows are zero. Estimates are calculated at the weekly frequency. The standard errors
correct for heteroskedasticity and the moving average error process induced by overlapping forecasts
(4 week results).

xt Edt xt : Several points emerge from the table. First, the coe¢ cients on the order ‡ow segments are
quite di¤erent from each other. Some are positive, some are negative, some are highly statistically
signi…cant, others are not. Second, while the coe¢ cients on order ‡ow are jointly signi…cant in
every regression we consider, the proportion of the variation in returns that they account for rises
with the horizon: the R2 statistic in regressions with all six ‡ows rises from 21 to 32 percent as
we move from the 1 to 4 week horizon.13 Third, the explanatory power of the order ‡ows shown

13
Froot and Ramadorai (2002) also …nd stronger links between end-user ‡ows and returns as the horizon is extended
to 1 month; their ‡ow measure is institutional investors, however, not economy-wide.

29
here is much less than that reported for interdealer order ‡ows. Evans and Lyons (2002a), for
example, report that interdealer order ‡ow accounts for approximately 60 percent of the variations
in the $/DM at the daily frequency. Finally, we note that none of the coe¢ cients on the interest
di¤erential are statistically signi…cant, and many have an incorrect (i.e. negative) sign.14 This is
not surprising in view of the empirical literature examining uncovered interest parity. However, the
estimated coe¢ cients on the order ‡ows are essentially unchanged if we re-estimate the regressions
with a unity restriction on the interest di¤erential, as implied by equation (27).
The key to understanding these results lies in the distinction between unexpected order ‡ow in
the model, xt Edt xt ; and our six end-user ‡ows: According to the model, realized foreign exchange
returns re‡ect the revision in dealer’s quotes driven by new information concerning fundamentals.
This information arrives in the form of public news, macro announcements and inter-dealer order
‡ow, but not the end-user order ‡ows of individual dealers such as Citibank: Any information
concerning fundamentals contained in the end-user ‡ows received by individual banks a¤ects the
FX price quoted by dealers only once it is inferred from the inter-dealer order ‡ows observed by all
dealers. In Evans and Lyons (2006) we study the relationship between end-user ‡ows and market-
wide inter-dealer order ‡ow (i.e., the counterpart to xt Edt xt ). This analysis shows that individual
coe¢ cients have no structural interpretation in terms of measuring the price-impact of di¤erent
end-user orders, they simply map variations in end-user ‡ows into an estimate of the information
‡ow being used by dealers across the market. This interpretation also accounts for the pattern of
explanatory power: As the horizon lengthens, the idiosyncratic elements in Citibank’s’ end-user
‡ows become relatively less important, with the result that the ‡ows are more precise proxies for
the market-wide ‡ow of information driving quote revisions.
To summarize, the results in Table 2 show that end-user ‡ows are contemporaneously linked
with changes in spot rates, but the strength of the link is less than that reported elsewhere for
inter-dealer order ‡ows. Once one recognizes that Citibank’s end-user ‡ows are an imperfect proxy
for inter-dealer order ‡ows, our …ndings are consistent with the theoretical link between exchange
rates and order ‡ow implied by the model.

14
We report results using 4 week rates on Euro-dollar and Euro-mark deposits in both panels of the table because
1 week euro-current rates were unavailable over the entire sample period. Re-estimating the regressions in the upper
panel with 1 week rates when they are available over the second half of the sample gives very similar results.

30
4.2 Forecasting Fundamentals

According to Proposition 3, changes in the exchange rate are correlated with order ‡ow because the
latter contains information concerning fundamentals. If this is the mechanism responsible for the
results reported in Table 2, order ‡ows ought to have forecasting power for future fundamentals.
We now examine whether this implication of our model applies to the end-user ‡ows. First we
derive the model’s implications for forecasting fundamentals with spot rates and order ‡ows. We
then examine the forecasting power of spot rates and the end-user ‡ows for future changes in our
real-time estimates.
The model’s implications for forecasting fundamentals with spot rates follow straightforwardly
from Proposition 1. In particular equation (11) can be rewritten as

1
X i
st = Edt ft + Edt 1+ ft+i : (29)
i=1

Thus, the log spot rate quoted by dealers di¤ers from dealers’ current estimate of fundamentals
by the present value of future changes in fundamentals. One implication of (29) is that the gap
between the current spot rate and estimated fundamentals, st Edt ft ; should have forecasting power
for future changes in fundamentals. This can be formally shown by considering the projection:

ft+ = s (st Edt ft ) + "t+ ; (30)

1
X i
where s = 1+ CV(Edt ft+i ; Edt ft+ )=V (st Edt ft ) ;
i=1

and "t+ is the projection error that is uncorrelated with st Edt ft . The projection coe¢ cient s

provides a measure of the forecasting power of st Edt ft for the change in fundamentals periods
ahead.
Now we turn to the forecasting power of order ‡ow. According to Proposition 3, order ‡ow is
driven in part by di¤erences between dealers’forecasts and household forecasts concerning future
fundamentals. Consequently, if households have more precise information concerning future funda-
mentals than dealers, order ‡ows should have incremental forecasting power beyond that contained
st Edt ft : We formalize this idea in the following proposition.

31
Proposition 5 When dealer quotes for the price of foreign currency satisfy (11), and order ‡ow
follows (22), changes in future fundamentals are related to spot rates and order ‡ows by

ft+ = s (st Edt ft ) + x (xt Edt xt ) + t+ ; (31)

where t+ is the projection error. s is the projection coe¢ cient identi…ed in (30) and x is equal
to
1 0 C 0 {0 ^ ^ V rEhb yt+1 A 1 0 C 0 {0
CV (ot ; ft+ ) V (rEht yt+1 ) A 2 t 2
+ + :
V (xt Edt xt ) V (xt Edt xt ) V (xt Edt xt )

The intuition behind Proposition 5 is straightforward. Recall from (29) that st Edt ft is equal to
the present value of future changes in fundamentals. The …rst term in (31) is therefore a function of
dealers’information at the start of period t; d: Period-t order ‡ow will have incremental forecasting
t

power of future changes in fundamentals, beyond st Edt ft ; when it conveys information about
ft+ that is not already known to dealers (i.e. in d ): The expression for shows that this will
t x

happen when: (i) the distribution of wealth and dealer bond holdings a¤ect order ‡ow and have
forecasting power for fundamentals, and (ii) when there is dispersed information concerning future
fundamentals and information aggregation occurs via period-t trading. Proposition 4 showed that
order ‡ow would be correlated with the depreciation rate under these same conditions. Thus, if
our theoretical rationale for the results in Table 2 holds true, we should also …nd that order ‡ow
has incremental forecasting power for future changes in fundamentals.
To assess the empirical evidence on this prediction, we consider forecasting regressions of the
form:
k k P6 k
{t+ = a1 {t + a2 st + n=1 j xj;t + t+ ; (32)

where {t+ denotes the week change in the real-time estimate of variable { ending at week
t+ ; ks is the rate of depreciation between weeks t k and t; and xkj;t is the order ‡ow from
t

segment j in weeks t k to t: The …rst two terms on the right hand side are known to dealers at the
start of week t and are used to proxy for st Edt ft in equation (31). Estimates of the j coe¢ cients
will reveal whether our end-user ‡ows have incremental forecasting power for future fundamentals.
Table 3 presents the results from estimating (32) in weekly data with horizons ranging from
one month to two quarters. We report results where k is set equal to ; but our …ndings are not

32
sensitive to the number of cumulation weeks k: There are a total of 284 weekly observations in
our sample period, so there are 11 non-overlapping observations on the dependent variable at our
longest forecasting horizon (e.g. 2 quarters). In each cell of the table we report the R2 statistic as a
measure of forecasting power and the signi…cance level of a Wald test for the joint signi…cance of the
forecasting variables. These test statistics are corrected for conditional heteroskedasticy and the
moving-average error structure induced by the forecast overlap using the Newey-West estimator.
The results in Table 3 clearly show that order ‡ow has considerable forecasting power for all
of the six macro variables, and this forecasting power is typically a signi…cant increment over the
forecasting power of the other variables considered. Consider, for example, the case of US GDP.
At the two-quarter forecasting horizon, order ‡ow produces an R2 statistic of 24.6 percent, which
is signi…cant at the one-percent level. In contrast, forecasting US GDP two months out using both
past real-time estimates of GDP and the spot rate produces an R2 statistic of only 9.6 percent,
a level of forecasting power that is insigni…cant at conventional levels. In general, the forecasting
power of order ‡ow is greater as the forecasting horizon is lengthened.15
Our …ndings in Table 3 are robust to the inclusion of other variables as proxies for st Edt ft : In
particular, we have estimated versions of (32) that include multiple lags of k{ and ks as well
t t

as the term spread, default spread and the commercial paper spread.16 We found that the term
spread predicts US GDP and M1, and German prices and M1, while the default and commercial
paper spreads predict US GDP. However, the marginal forecasting contribution of these variables
is small. Moreover, in all cases, the forecasting contribution of the six ‡ow segments remains highly
signi…cant at one and two-quarter horizons. These …ndings indicate that the results in Table 3 are
indeed robust to the inclusion of di¤erent variables proxying for st Edt ft :
Although the longest horizon we consider in Table 3 is short compared to the span of our data,
our asymptotic inferences concerning forecasting power over 1 and 2 quarters may not be entirely

15
Our theoretical model indicates that US and German consumption are components of fundamentals. Unfortu-
nately, we were not able to compute real-time estimates for both consumption series because the sequence of data
releases for German consumption are unavailable. We did compute real-time estimates of US consumption and found
that the forecasting power of order ‡ows is similar to that we report for US GDP.
16
The term spread is the di¤erence between the 3-month and 5-year yields on US bonds. We compute the default
spread as the di¤erence between Moody’s AAA corporate bond yield and Moody’s BAA corporate bond yield. The
commercial paper spread is the di¤erence between the 3-month commercial paper rate and the 3-month T-Bill rate.
Before September 1997 we use the 3-month commercial paper rate, thereafter the 3-month rate for non-…nancial
corporations. We obtained the term structure data from CRSP, and the other interest rates from the FRED database
at the St Louis Fed.

33
 Table 3: Forecasting Fundamentals

Forecasting US GDP German GDP

Variables 1M 2M 1Q 2Q 1M 2M 1Q 2Q
GDP 0.002 0.003 0.022 0.092 0.004 0.063 0.089 0.006
(0.607) (0.555) (0.130) (0.087) (0.295) (0.006) (0.009) (0.614)
Spot Rate 0.001 0.005 0.005 0.007 0.058 0.029 0.003 0.024
(0.730) (0.508) (0.644) (0.650) (0.002) (0.081) (0.625) (0.536)
GDP and Spot 0.003 0.007 0.031 0.096 0.059 0.083 0.099 0.033
(0.802) (0.710) (0.287) (0.224) (0.007) (0.021) (0.024) (0.709)
Order Flows 0.032 0.080 0.189 0.246 0.012 0.085 0.075 0.306
(0.357) (0.145) (0.002) (0.000) (0.806) (0.227) (0.299) (0.000)
All 0.052 0.086 0.199 0.420 0.087 0.165 0.156 0.324
(0.383) (0.195) (0.011) (0.000) (0.021) (0.037) (0.130) (0.000)
US Prices German Prices
Prices 0.003 0.024 0.005 0.053 0.007 0.037 0.053 0.024
(0.461) (0.146) (0.487) (0.213) (0.402) (0.067) (0.040) (0.232)
Spot Rate 0.005 0.007 0.013 0.016 0.081 0.000 0.000 0.033
(0.351) (0.419) (0.391) (0.457) (0.000) (0.962) (0.858) (0.305)
Prices and Spot 0.007 0.028 0.015 0.06 0.088 0.038 0.053 0.051
(0.505) (0.352) (0.636) (0.441) (0.002) (0.214) (0.112) (0.364)
Order Flows 0.025 0.050 0.116 0.212 0.050 0.116 0.178 0.271
(0.773) (0.629) (0.052) (0.000) (0.429) (0.010) (0.025) (0.000)
All 0.031 0.082 0.124 0.240 0.127 0.158 0.258 0.511
(0.788) (0.151) (0.010) (0.000) (0.005) (0.021) (0.005) (0.000)
US Money German Money
Money 0.071 0.219 0.253 0.329 0.05 0.111 0.122 0.041
(0.009) (0.000) (0.000) (0.000) (0.023) (0.005) (0.017) (0.252)
Spot Rate 0.021 0.001 0.003 0.005 0.002 0.044 0.036 0.065
(0.054) (0.778) (0.732) (0.619) (0.558) (0.031) (0.123) (0.343)
Money and Spot 0.086 0.22 0.267 0.333 0.05 0.13 0.129 0.08
(0.002) (0.000) (0.000) (0.000) (0.075) (0.004) (0.040) (0.403)
Order Flows 0.034 0.119 0.280 0.424 0.026 0.082 0.152 0.578
(0.466) (0.239) (0.026) (0.000) (0.491) (0.147) (0.037) (0.000)
All 0.096 0.282 0.417 0.54 0.074 0.175 0.284 0.624
(0.056) (0.000) (0.000) (0.000) (0.244) (0.020) (0.001) (0.000)
The table reports the R2 statistic from the forecasting regression (32) for the real-time estimate
of the fundamental listed in the header of each panel, using the forecasting variables reported on
the left. The regressions are estimated in weekly data (284 observations). Signi…cance levels for
statistics testing the null hypothesis of no predictability (corrected for heteroskedasticity and the
forecast horizon overlap) are reported in parentheses. The forecasting horizons listed at the head of
each column are 1 month ( = 4); 2 months ( = 8); 1 quarter ( = 13) and 2 quarters ( = 26):

34
reliable.17 To insure that our forecasting …ndings are robust, we supplemented our analysis at these
two horizons with the following Monte Carlo experiment: First we estimated an AR(4) process for
weekly change in the real-time estimate, {t and a fourth-order VAR for the weekly change in log
spot rate, st ; and the six ‡ow segments, xj;t : Next, we generated a pseudo data series spanning 284
weeks for {t by combining a bootstrap sample from the {t residuals with estimates of AR(4)
process. Pseudo data series for st and xj;t are similarly generated by bootstrap sampling from the
VAR residuals and estimates. Notice that under this data generation process, realizations of {t
are independent from the other variables. We then used the pseudo data to estimate equation (32)
at the 1 quarter ( = 13) and 2 quarter ( = 26) horizons. This process was repeated 5000 times to
construct a bootstrap distribution for the regression estimates under the null hypothesis that both
spot rates and order ‡ows have no forecasting power for the real-time estimates of fundamentals.

The results of our Monte Carlo experiment are shown in Table 4. To conserve space we only
report the results from regressions that include the real-time estimates, spot rates and the end-
user ‡ows. The upper entry in each cell is the OLS coe¢ cient on the variable listed at the head
of each column estimated from our data. The lower entry is the percentage contribution of the
variable to the variance of the future change in the fundamental, again estimated from our data.18
The variance contribution of all six end-user ‡ows is shown in the right hand column. We compare
these estimated variance contributions to the Monte Carlo distribution of the contributions generate
under the null of no forecastability, and denote by “*”, “**” and “***” cases where the variance

17
Estimates of long-horizon forecasting regressions like (32) are susceptible to two well-known econometric problems.
First, the coe¢ cient estimates may su¤er from …nite sample bias when the independent variables are predetermined
but not exogenous. Second, the asymptotic distribution of the estimates provides a poor approximation to the
true distribution when the forecasting horizon is long relative to the span of the sample. Finite-sample bias in the
estimates of j is not a prime concern here because our six ‡ow segments display little or no autocorrelation and
are uncorrelated with past changes in the real-time estimates. There should also be less of a size distortion in the
asymptotic distribution than found elsewhere. For example, Mark (1997) considers a case where the data span is less
than …ve times the length of his longest forecasting horizon. Here, we have 11 non-overlapping observations at the
2-quarter horizon.
18
To compute the contribution, we take the …tted values of (32),
P
{t+ = a ^2 k st + 6n=1 ^ j xkj;t + ^t+
^ 1 k {t + a

multiply both sides by {t+ ; and take second moments:

k k P6 ^ CV(xkj;t ;
V( {t+ ) = a
^1 CV( {t ; {t+ ) + a
^2 CV( st ; {t+ ) + n=1 j {t+ )xkj;t
+CV( {t+ ; ^t+ ):

Notice that by least squares, CV( {t+ ; ^t+ ) = V(^t+ ); so we end up with a decomposition for V( {t+ ). The
^2 CV( k st ;
variance contribution of the spot rate is therefore a {t+ )=V( {t+ ) and so on.

35
 Table 4: Contributions to Fundamentals’Forecasts

Forecast Spot Fund Corporate Traders Investors All Flows

US non-US US non-US US non-US
US
GDP
1Q -0.013 -0.111 1.476 0.329 -0.235 0.283 -0.099 -0.116
0.230 1.689 8.797 3.579 2.946 0.483 0.684 1.575 18.064
2Q -0.109 -0.449 0.709 0.283 0.607 -0.218 -0.662 0.281
-2.332 13.544 6.300 2.763 15.921 0.458 3.600 1.839 30.882
Prices
1Q 0.005 0.048 -0.144 0.040 0.037 -0.068 -0.072 0.020
0.987 0.330 1.874 1.829 2.406 1.955 1.301 1.695 11.060
2Q -0.006 0.184 0.141 -0.010 -0.045 0.083 -0.020 0.100
-1.247 4.150 0.664 -0.602 1.027 1.525 0.993 17.236 20.842
M1
1Q 0.360 0.425 0.919 -0.957 2.129 -5.184 -9.573 -2.477
0.577 21.427 0.633 -0.826 -0.045 3.020 14.798 2.160 19.740
2Q -1.039 0.439 1.391 -3.944 5.292 -1.806 -10.537 -0.056
1.962 30.842 1.259 -2.929 1.657 0.855 19.764 0.009 20.615
German
GDP
1Q 0.095 -0.281 -0.983 -0.712 0.302 -0.997 -0.726 -0.426
0.617 8.814 0.257 2.108 0.409 2.646 0.368 0.394 6.181
2Q -0.042 -0.106 -1.677 0.260 0.024 -0.845 1.402 1.170
-0.859 0.837 2.630 -0.730 0.095 3.026 6.511 19.995 31.527
Prices
1Q -0.192 -0.286 2.315 0.167 -0.068 -3.479 -2.701 1.027
-0.307 6.483 2.485 0.065 0.052 9.421 5.384 2.171 19.578
2Q -0.531 -0.491 1.764 0.714 0.104 -3.242 -4.703 1.394
7.156 7.04 2.263 3.142 -0.327 3.287 25.076 3.355 36.797
Money
1Q 0.724 0.396 -3.224 2.408 -0.233 3.210 5.180 -5.215
-3.982 13.88 0.443 1.928 0.154 1.046 1.749 13.15 18.469
2Q 1.143 0.243 5.670 -3.091 0.203 8.813 -4.129 -11.436
-6.683 4.935 0.938 0.349 -0.067 10.483 1.530 50.605 63.838
Notes: The upper entry in each cell is the OLS coe¢ cient in the forecasting equation
computed at the 1 quarter ( =13) and 2 quarters ( = 26) horizon. The lower entry is the
percentage contribution to the variance of the forecast variable. Estimated contributions
falling in the 10, 5, and 1 percent tails of the bootstrap distribution are denoted by *, **,
and ***.

36
contribution lies in 10, 5 and 1 percent tails of the bootstrap distribution.
The results in Table 4 complement our earlier …ndings in two important ways. First, the Monte
Carlo results con…rm that our end-user ‡ows have signi…cant incremental forecasting power for
fundamentals. Although many of the variance contributions from the individual ‡ow segments
do not appear statistically signi…cant when compared against the Monte Carlo distribution, the
joint contribution of all six ‡ows are signi…cant at the 5 percent level at either the one or two
quarter horizon. Moreover, judged by the estimated size of the variance contributions, the order
‡ows contain information that accounts for an economically meaningful fraction of variance in the
variable being forecast. In fact, with the exception of US M1, the order ‡ows account for more
of the variance than do spot rates or fundamentals. The second noteworthy feature of Table 4
concerns the forecasting contribution of spot rates. Although the estimates are small, they are
highly statistically signi…cant at the two quarter horizon for all six variables. Consistent with the
present value equation in (11), changes in the spot rates do appear to contain information about
the future course of macro fundamentals. We also note from Table 4 that there is a good deal
of heterogeneity in the estimated coe¢ cients and contributions of the individual order ‡ows.19
Imperfect classi…cation of end-user orders into our six segments probably accounts for some of this
heterogeneity. Recall that our ‡ow segments are classi…ed according to trade location rather than
the nationality of the end-user. Nevertheless, we do note that the largest and most statistically
signi…cant contributions come from US-located trades for US variables, and non-US-located trades
for German variables.
The results in Tables 3 and 4 contrast quite sharply from the …ndings of Froot and Ramadorai
(2005). They found no evidence of a long run correlation between real interest rate di¤erentials
(their measure of fundamentals) and the transaction ‡ows of institutional investors. One likely
reason for this di¤erence is the wider span of end-users generating the order ‡ows in our data. The
estimates in Table 4 suggest that transactions from di¤erent end-users convey di¤erent information.
Our use of the real-time estimates is also important. Recall that the change in real-time estimate
comprises an ex ante forecast and an information ‡ow. For example, the change in the real-time

19
It is tempting to interpret the coe¢ cients on the individual ‡ows as measuring the information content of an
unexpected order from a particular segment. However, as we note earlier, our six ‡ows are correlated with one-another,
so the information content of an unexpected order cannot be measured by a single coe¢ cient.

37
estimate of GDP over the …rst quarter of the year can be written as

q
ln GDPq(2)jw(13) ln GDPq(1)jw(1) = E ln GDPq(2) w(1) + ln GDPq(2)jw(13) ln GDPq(2)jw(1) ;

where q ln GDP ln GDPq(2) ln GDPq(1) and w(j) denotes the …rst day on week j: Thus the
q(2)

change in the real-time estimate comprises the ex ante forecast of GDP growth in the …rst quarter,
and the ‡ow of information concerning second-quarter GDP over the …rst quarter. Now according
P
to (29), variations in st Edt ft re‡ect changes in Edt 1
i=1 ( 1+ )
i f
t+i : So if is large, as Engel and
West (2005) argue, and log GDP is correlated with fundamentals ft ; then variations in the ex ante
q ln GDP d d
P1 i f
forecasts, E q(3) w(1) ; should track changes in st Et ft = Et i=1 ( 1+ ) t+i : This

is the element in the real-time forecasts picked up by the spot rate and lagged fundamentals. Table
4 showed that the estimated variance contributions from these variables are small yet statistically
signi…cant –exactly what we should expect to …nd if there is little variation in the ex-ante forecasts.
The forecasting power of the order ‡ows for the real-time estimates works through a di¤erent
mechanism. Recall that our order ‡ows are not public information, so their forecasting power for
the change in the real-time estimates cannot come via changes in the ex-ante forecasts. Instead,
the order ‡ows must be correlated with the ‡ow of public information concerning the fundamental
over the forecast horizon. For the case of GDP, this is the second term in the decomposition above.
The only di¤erence between ln GDPq(2)jw(13) and ln GDPq(2)jw(1) is that the former estimate in-
corporates the information in public data releases between week 1 and 13. With this perspective,
our results in Tables 3 and 4 imply that the end-user ‡ows convey information about future funda-
mentals that is subsequently revealed by the arrival of public data releases. Clearly, these releases
represent information that is incremental to the information embedded in spot rates at the be-
ginning of the forecast period. Our empirical …ndings therefore provide strong corroboration for
Proposition 5.

4.3 Exchange Rate Dynamics and Information Flow

One notable feature of the results in Tables 3 and 4 is that the forecasting power of our end-user
‡ows for fundamentals appears stronger at longer forecasting horizons. We interpret this …nding
as evidence that some of the information conveyed by the order ‡ows only shows up in public news

38
releases many months later. In this section we investigate two implications of this interpretation.
First, we examine whether end-user ‡ows have forecasting power for changes in the exchange rate.
Second, we consider whether the forecasting power of ‡ows is consistent with their ability to forecast
the future ‡ow of information concerning exchange rate fundamentals.
To understand how our forecasting results for fundamentals relate to the forecastability of the
exchange rate, we return to the model. In particular, we consider the implications of Proposition
1 for the change in the log spot rate.

Proposition 6 When dealer quotes for the price of foreign currency satisfy (11), the change in
the log spot rate between the start of period t and t + is

1
X
st+ st+ st = ' (st Edt ft; ) + 1
1+ ( 1+ )i Edt+ Edt ft+ +i (33)
i=0

P 1
where ' ( 1+ ) 1 > 0 and ft; ' +1
' (1+ ) i=0 ( 1+ )i ft+i :

Equation (33) shows us that the change in spot rate comprises two components. The …rst
term on the right identi…es the expected depreciation rate Edt [st+ st ]; which is proportional to
the gap between the current spot rate and expected “near-term” fundamentals, ft; : The second
term identi…es the impact of new information regarding future fundamentals received by dealers
between the start of periods t and t + ; (Edt+ Edt )ft+ +i : This will be the only term making a
signi…cant contribution to the change in sport rates over short and medium horizons. The reason
is that reasonable estimates of the semi-interest elasticity, ; fall between 20 to 60 (Engel and West
2005), so ' will be close to zero until becomes very large. Any variation in st Edt ft; will
therefore have little impact on the realized change in spot rates. Consequently, we should expect
short- and medium-term changes in spot rates to be mainly driven by the arrival of new information
concerning the future course of fundamentals.
The implications of our …ndings in Tables 3 and 4 for forecasting returns should now be clear.
If our end-user ‡ows forecast changes in the real-time estimates of variable { because they contain
information about the future ‡ow of public information concerning {; the ‡ows should also have
forecasting power for future changes in spot rates if { is correlated with exchange rate fundamentals.
In other words, our results in Tables 3 and 4 suggest that end-user ‡ows ought to predict (Edt+

39
Edt )ft+ +i if the macro variables we examined are correlated with fundamentals.
To examine this hypothesis, we estimate the following forecasting regression:

6
X
st+ = a0 + a1 (rt r^t ) + j xj;t + ! t+ ; (34)
j=1

where rt r^t is the interest di¤erential between one month Eurodollar and Euromark deposits and
xj;t is the order ‡ow from segment j in weeks t to t: We include the interest di¤erential to control
for any variations in expected depreciation (i.e., ' (st Edt ft; ) in equation 33). The regression
errors ! t+ pick up news concerning future fundamentals that is not correlated with the end-user
‡ows.

Table 5: Forecasting Exchange Rate Changes

Weeks r^ r Corporate Traders Investors R2 2

US NUS US NUS US NUS (p-value)

1 0.102 0.482 -0.033 0.089 -0.153 -0.346 0.142 0.027 8.056

(0.409) (0.317) (0.136) (0.102) (0.198) (0.238) (0.141) (0.234)

2 0.147 0.509*** -0.037 0.088 -0.09 -0.449*** 0.163** 0.074 17.239

(0.324) (0.263) (0.104) (0.082) (0.148) (0.188) (0.096) (0.008)

3 0.176 0.615*** -0.034 0.095* -0.084 -0.432*** 0.145** 0.121 24.500

(0.305) (0.215) (0.090) (0.073) (0.137) (0.177) (0.082) (0.001)

4 0.202 0.544*** -0.042 0.094* -0.097 -0.517*** 0.137** 0.163 30.738

(0.302) (0.177) (0.084) (0.068) (0.125) (0.158) (0.072) (<0.001)
Notes: The table reports coe¢ cient and standard errors from regressions of future returns
measured over horizons of one to four weeks, on an (unreported constant), the current
interest di¤erential and order ‡ows cumulated over the last 4 weeks. The left hand column
reports Wald statistics for the null that all the coe¢ cients on order ‡ow are zero. Estimates
are calculated at the weekly frequency. The standard errors correct for heteroskedasticity
and the moving average error process induced by overlapping forecasts (2 - 4 week results).
*, **, and *** denote signi…cance at the 10%, 5% and 1% levels.

Table 5 reports the results of estimating (34) for horizons of one to four weeks. Two features of
the table are striking. First, many of the i coe¢ cients on the end-user ‡ows are highly statistically
signi…cant, particularly the US corporate and long-term investor ‡ows. The right hand column

40
shows Wald statistics for the joint signi…cance of all six ‡ow segments that are highly signi…cant
beyond the one week horizon. By contrast, none of the coe¢ cients on the interest di¤erential
are statistically signi…cant (although they do have the correct positive sign). The second striking
feature concerns the degree of forecastability as measured by the R2 statistics. The forecasting
power rises with the horizon, reaching 16 percent at four weeks. By comparison, the R2 statistics
from Fama-type regressions (where the rate of depreciation is regressed on the interest di¤erential)
are generally in the 2-4 percent range. Here all the forecasting power comes from the order ‡ows.
If we omit the interest di¤erentials and re-estimate the regressions, the estimated coe¢ cients on
the ‡ows and the R2 statistics are essentially unchanged.
The results in Table 5 point to a remarkably strong within-sample relation between order ‡ows
and future exchange rate changes. However, there is a long tradition in the exchange rate literature
of considering out-of sample forecasting performance. In Evans and Lyons (2005) we examined the
out-of-sample forecasting performance of the six order ‡ows for st+ with the restrictions a0 = 0
and a1 = 1: At the four week horizon the out-of-sample forecasts accounted for a highly signi…cant
15.7 percent of the variation in excess returns. This degree of forecastability closely matches the
in-sample R2 statistic in Table 5. So even though the statistics in the table relate to the within-
sample relation between order ‡ows and changes in the exchange rate, they are representative of
the true out-of-sample forecasting power of order ‡ows.
In view of these results, it is now natural to ask whether the predictive power of order ‡ows
for exchange rate changes is consistent with their ability to forecast the future ‡ow of information
concerning fundamentals. To address this question, we need to take a stand on the relation between
true fundamentals, ft ; and our real-time estimates. We consider 6 di¤erent measures based on the
variables for which we have real-time estimates, ftm : (i) the di¤erence between log GDP in the US
and German, y y^; (ii) the US-German log price ratio, p p^; (iii) the US-German log M1 ratio,
m m;
^ (iv) the US log M1 to GDP ratio, m y; (v) the German log M1 to GDP ratio, m
^ y^;
and (vi) the log M1-GDP di¤erential between the US and Germany, (m y) (m
^ y^): For each
measure of fundamentals, ftm , we …rst calculate the projection of the quarterly change in ftm on the
six order ‡ows, Proj ( q f m jfx g); as the …tted value from the regression
t+q j;t

6
X
q m m
ft+q ft+13 ftm = 0+
4
j xj;t + t+13 :
j=1

41
We then estimate

q m
st+ = b0 + b1 (rt r^t ) + b2 Proj( ft+q jfxj;t g) + t+ : (35)

If the predictive power of order ‡ows for future changes in the exchange rate are due to their ability
to forecast the future ‡ow of information concerning measured fundamentals, i.e. Edt+ ft+13
m

Edt ft+13
m ; then the estimates of b in (35) should be positive and signi…cant. Moreover, if our measure
2

of fundamentals, ftm ; is closely correlated with actual fundamentals, ft ; none of the individual order
‡ows xj;t ; should have incremental predictive power for st+ beyond their role in the projection
Proj( q f m jfx g):
t+q j;t

Table 6 reports the estimates of the forecasting regression (35) for horizons of one to four weeks,
using the projections of the six di¤erent fundamentals measures. Once again, we …nd the results
rather striking. First, the coe¢ cient estimates display a similar pattern across all four forecast
horizons. The coe¢ cients on the projections involving the log GDP and price ratios are small and
statistically insigni…cant. By contrast, the coe¢ cients on projections for the log M1 ratios, M1
to income ratios, and the M1-GDP di¤erentials are all highly signi…cant. This constitutes direct
empirical evidence that the end-user ‡ows are conveying information about the future course of
fundamentals, and it is this information that gives ‡ows their forecasting power for future changes
in spot rates. The second noteworthy feature concerns the R2 statistics. A comparison of the
R2 statistics in Table 5 with the statistics in the lower three rows of each panel in Table 6 shows
that the forecasting power of the projections is almost as high as that of the underlying order
‡ows. For example, at the four week horizon the R2 statistic from (35) using the projection of
the quarterly change in (m y) (m
^ y^) is 13.3 percent, while the R2 from estimating (34) is
16.3 percent. The use of the projection places restrictions on the way that the six ‡ows enter
(35), but these restrictions do little to impair the forecasting ability of ‡ows for future exchange
rates. The right hand column of Table 6 provides more formal evidence on this idea. Here we
report LM statistics for the null hypothesis that the residuals from (35) are unrelated to the six
‡ows. If order ‡ows have forecasting power for exchange rates for reasons that are unrelated to
the role in conveying information about fundamentals, or the fundamentals measures used in the
projections are only weakly correlated with true fundamentals, we should …nd that order ‡ows have

42
some residual forecast power, so the null ought to be rejected. However, as the table shows, we fail
to reject the null in all the cases where the projection coe¢ cients appear signi…cant.

Table 6: Information Flow Regressions

Horizon r^ r y y^ p p^ m m
^ m y m
^ y^ (m y) R2 LM
(m
^ y^) (p-value)
1 week -0.229 -0.229 <0.001 7.859
(0.369) (0.949) (0.249)
-0.194 -0.290 0.001 8.525
(0.367) (0.507) (0.202)
0.161 0.589** 0.023 1.115
(0.387) (0.218) (0.981)
0.04 0.436** -0.700** 0.025 0.783
(0.398) (0.280) (0.281) (0.993)
0.110 0.585** 0.023 1.158
(0.381) (0.219) (0.979)
2 weeks -0.215 0.013 <0.001 21.882
(0.307) (0.704) (0.001)
-0.208 -0.129 0.001 23.977
(0.312) (0.420) (0.001)
0.199 0.656** 0.060 na
(0.317) (0.185)
0.055 0.467* -0.771** 0.063 3.741
(0.317) (0.227) (0.231) (0.712)
0.136 0.639** 0.059 4.787
(0.310) (0.184) (0.571)
3 weeks -0.208 -0.061 <0.001 32.498
(0.308) (0.659) (<0.001)
-0.198 -0.093 <0.001 39.596
(0.317) (0.399) (<0.001)
0.245 0.708** 0.104 4.475
(0.310) (0.169) (0.613)
0.101 0.524** -0.825** 0.109 4.364
(0.301) (0.212) (0.199) (0.628)
0.182 0.694** 0.102 5.926
(0.302) (0.172) (0.432)

43
 Table 6: Information Flow Regressions (cont.)

Horizon r^ r y y^ p p^ m m
^ m y m
^ y^ (m y) R2 LM
(m
^ y^) (p-value)
4 weeks -0.214 -0.094 <0.001 52.375
(0.315) (0.651) (<0.001)
-0.200 -0.106 <0.001 53.402
(0.327) (0.383) (<0.001)
0.248 0.709** 0.135 8.033
(0.316) (0.156) (0.236)
0.122 0.564** -0.799** 0.138 9.129
(0.302) (0.193) (0.186) (0.166)
0.186 0.697** 0.133 10.109
(0.307) (0.162) (0.120)
Notes: The table reports coe¢ cients and asymptotic standard errors from regressions of future
returns measured over horizons of one to four weeks on the current interest di¤erential, and the
projection of the future quarterly change macro fundamentals on current order ‡ows from the six
user-user segments. Fundamentals are listed at the head of each column. The left hand column
report LM statistics for the null that the regression residuals are unrelated to order ‡ows. (The
LM statistic could not be computed for the case labelled “na” because the projection was perfectly
correlated with one or more of the order ‡ows.) Standard errors correct for heteroskedasticity and
the moving average error process induced by overlapping forecasts (2 - 4 week results). *, **, and
*** denote signi…cance at the 10%, 5% and 1% levels.

5 Conclusion

The aim of this paper has been to analyze the links between macro fundamentals, transaction
‡ows and exchange rate dynamics. First, we presented a micro-founded general equilibrium model
that provides the theoretical foundation for understanding how dispersed information concerning
macro fundamentals is conveyed to spot rates via transaction ‡ows. We then examine the empirical
implications of the model. We found that transaction ‡ows have signi…cant forecasting power for
macro fundamentals –incremental forecasting power beyond that contained in exchange rates and
other variables. We also showed that proprietary transaction ‡ows have signi…cant forecasting
power for future exchange returns and that this forecasting power re‡ects their ability to predict
how “the market” will react to the ‡ow of information concerning macro fundamentals. In sum,
we …nd strong support for the idea that exchange rates vary as “the market”assimilates dispersed
information regarding macro fundamentals from transaction ‡ows.
Let us conclude with some perspective. Our results provide a qualitatively di¤erent view of

44
why macroeconomic variables perform so poorly in accounting for exchange rates at horizons of
one year or less. This view is di¤erent from both the traditional macro and the emerging “mi-
cro” perspectives. Unlike the macro perspective, we do not view all new information concerning
macro fundamentals as being immediately embedded into the exchange rate. Much information
about macro fundamentals is dispersed and it takes time for “the market” to fully assimilate its
implications for the spot exchange rate. It is this assimilation process that accounts for (much
of) the disturbances in exchange rate equations. Our approach also di¤ers from the extant micro
perspective because models o¤ered thus far (e.g., Evans and Lyons 2002a,b) have interpreted the
information conveyed by transaction ‡ows as orthogonal to macro fundamentals. This information
is viewed, instead, as relating to the other driver within the broader asset pricing literature, termed
stochastic discount factors, expected returns or portfolio balance e¤ects. Most readers of this micro
literature have adopted the same view: transaction ‡ow e¤ects on exchange rates are about pricing
errors, not about fundamentals. Our …ndings, by contrast, suggest that transaction ‡ows are cen-
tral to the process by which expectations of future macro variables are impounded into exchange
rates.

45
 References

Bacchetta, P., and E. van Wincoop (2006), Can information dispersion explain the exchange rate
disconnect puzzle? American Economic Review, 93, pp. 552-576.

Campbell, J. Y., and L. Viceira (2002), Strategic Asset Allocation: Portfolio Choice for Long Term
Investors, Clarendon Lectures in Economics, Oxford University Press.

Cheung, Yin-Wong, M. Chinn, and A. Pascual (2005), Empirical exchange rate models of the
nineties: Are any …t to survive? Journal of International Money and Finance, 24, pp. 1150-
1175.

Croushore, D. and T. Stark (2001), A real-time data set for macroeconomists, Journal of Econo-
metrics 105, pp. 111-130.

Engel, C., and K. West (2005), Exchange rates and fundamentals, Journal of Political Economy,
113: 485–517.

Evans, M. (2005), Where are we now? Real-time estimates of the macro economy, The Interna-
tional Journal of Central Banking, 1, pp. 127-175.

Evans, M., and R. Lyons (2002a), Order ‡ow and exchange rate dynamics, Journal of Political
Economy, 110, pp. 170-180.

Evans, M., and R. Lyons (2002b), Informational integration and FX trading, Journal of Interna-
tional Money and Finance, 21, pp. 807-831.

Evans, M., and R. Lyons (2004), A New Micro Model of Exchange Rate Dynamics, NBER Working
Paper 10379, March.

Evans, M., and R. Lyons (2005), Meese-Rogo¤ Redux: Micro-based Exchange Rate Forecasting,
American Economic Review Papers and Proceedings, pp. 405-414.

Evans, M., and R. Lyons (2006), Understanding Order Flow, International Journal of Finance
and Economics, 11, pp. 1-24.

Faust, J., J. Rogers, and J. Wright (2003), Exchange rate forecasting: The errors we’ve really
made, Journal of International Economics, 60, pp. 35-59.

46
Frankel, J., and A. Rose (1995), Empirical research on nominal exchange rates, in G. Grossman
and K. Rogo¤ (eds.), Handbook of International Economics, Elsevier Science, Volume III:
1689-1729.

Froot, K., and T. Ramadorai (2005), Currency returns, institutional investor ‡ows, and exchange
rate fundamentals, Journal of Finance 60, pp. 1535-1566.

Hayek, F. (1945), The use of knowledge in society, American Economic Review, 35: 519-530.

Lyons, R. (1997), A Simultaneous Trade Model of the Foreign Exchange Hot Potato, Journal of
International Economics, 42, pp. 275-298.

Mark, N. (1997), Exchange Rates and Fundamentals: Evidence on Long-horizon Predictability,

American Economic Review, 85 pp. 201-218.

Meese, R., and K. Rogo¤ (1983), Empirical exchange rate models of the seventies, Journal of
International Economics, 14, pp. 3-24.

Osler, C. (2003), Currency Orders and Exchange-Rate Dynamics: Explaining the Success of Tech-
nical Analysis, Journal of Finance 58, pp. 1791-1819

47
A Appendix

This Appendix has three sections. Appendix A.1 provides a detailed description of the model
outlined in Section 1 of the paper. Appendix A.2 derives the results presented in Propositions 1 -
6. Finally, Appendix A.3 contains a brief description of how the real-time estimates are computed.

A.1 Model Details

A.1.1 Households

Under assumptions A1 and A2 in the text, the consumption and portfolio decisions facing the
representative US household can be written as the following dynamic programming problem:

n o
1 m 1
J(Wh,t ) = max 1
Ch,t + ( t Wh,t =Pt ) + Eht J(Wh,t+1 ) ;
;Ch,t 1
a m
t; t ; t
1
m
s.t. Wh,t+1 = Rt (ERh,t+1 Wh,t Pt Ch,t );
q
m ^t
St+1 R Rt+1 a Rt 1 m
ERh,t+1 = 1+ St Rt 1 t + Rt 1 t Rt t

q
where Rt+1 (Qt+1 + Dt+1 )=Qt is the return on us equity and

Wh,t = Rt 1 Bh,t 1
^t
+ St R ^
1 Bh,t 1 + Rt Ah,t 1 + Mh,t 1

m
is the dollar value of wealth at the beginning of period t: ERh,t+1 is the excess return on wealth
between periods t and t + 1 that depends on the share of wealth held in euro bonds t
^h,t =(R
St B ^ t Wh,t ); us equity a Qt Ah,t =Wh,t ; and real balances m Mh,t =Wh,t : Solving this
t t

problem gives the following …rst-order conditions:

Ch,t+1
Ch,t : Eht Ch,t
Pt
Pt+1 Rt = 1; (A1a)

m Mh,t 1 Rt 1
t : Pt Ch,t = Rt ; (A1b)
q
a Ch,t+1 Rt+1
t : Eht Ch,t Rt = 1; (A1c)

Ch,t+1 St+1 Pt ^
t : Eht Ch,t St Pt+1 Rt = 1: (A1d)

A1
 To characterize optimal household decisions, we work with log normal approximations to the
…rst-order conditions and a log linearization of the budget constraint. We …rst combine the identity
m m
Mh,t =Wh,t with the …rst-order condition for real balances and the de…nition of ERh,t+1 : The
t

budget constraint can then be rewritten as:

Wh,t+1 Pt Ch,t
Wh,t = Rt ERh,t+1 (1 + (Rt )) Wh,t ;

1
1= R 1 1
where (R) R and

q
^t
St+1 R Rt+1 a
ERh,t+1 1+ St Rt 1 t + Rt 1 t:

Notice that the coe¢ cient on the consumption-wealth ratio includes the (Rt ) function because
increased consumption raises holdings of real balances. This, in turn, reduces the growth in wealth
because the return on nominal balances is zero. Taking logs on both sides of the budget constraint,
and linearizing the right hand side around the point where the consumption-wealth ratio and home
nominal interest rate are constant, gives:

1
wh,t+1 = rt + k + 1 (erh,t+1 &rt ) (pt + ct wh,t ); (A2)

( 1) 1
where 1 (1 + (R)) ; & (R 1)R (R); k ln + (1 ) ln + &= ; and is the steady
state value of the consumption-wealth ratio, Pt Ch,t =Wh,t : Using the de…nition of ERh,t+1 above,
we follow Campbell and Viceira (2002) in approximating the log excess return on wealth by:

a q 1 a a h q
erh,t+1 = t rt+1 rt + t( st+1 + r^t rt ) + 2 t (1 t )Vt rt+1
h a h q
+ 21 t (1 t )Vt ( st+1 ) t t CVt rt+1 ; st+1 ; (A3)

where Vht (:) and CVht (:; :) denote the variance and covariance conditioned on information h:
t This
second–order approximation holds exactly in the continuous–time limit when the spot exchange
rate and the price of other assets follow di¤usion processes.
We can now use (A2), (A3) and the log linearized …rst-order conditions to characterize the
optimal choice of consumption, real balances and the portfolio shares: Combining the log linearized

A2
versions of (A1c) and (A1d) with (A2) and (A3) we obtain:
2 3 2 3
s
t h 14
Eht st+1 + r^t rt + 12 Vht (st+1 ) h;t
4 5= ( t)
5; (A4)
q q 1 h q q
t Eht rt+1 rt + 2 Vt (rt+1 ) h;t

where v
h;t = CVht (pt+1 + ch;t+1 wh;t+1 ; vt+1 ) + (1 ) CVht ( pt+1 ; vt+1 ) ;

q
for v = fs; rq g: The matrix h
t is the conditional covariance of the vector ( st+1 ; rt+1 )0 : Eht st+1 +
s q q
r^t rt h;t and Eht rt+1 rt h;t are the risk–adjusted expected excess returns on euro bonds
v
and us equities. The variance terms arise because we are working with log excess returns. h;t

identi…es the hedging factor associated with euro bonds (v = s) and us equities (v = rq ): All that
now remains is to characterize the demand for real balances and the consumption-wealth ratio.
The former is found by log linearizing (A1b):

mh;t pt = $ + ch;t rt ; (A5)

1
where $ ln + rR and 1= (R 1) > 0: An approximation to the log consumption
wealth-ratio is found by combining (A2) with the linearized version of (A1a):

1
X 1
X
k
ch;t + pt wh;t = 1 + 1 1
Eht i+1
(rt+i pt+1+i ) + Eht i 1
(erh,t+i &rt+i 1 ):
i=0 i=1

We can characterize the behavior of the representative European household in a similar way.
Speci…cally, the linearized budget constraint is:

1 1
w
^hb;t+1 = r^t p^t+1 + k + erhb;t+1 & r^t (^
pt + c^hb;t w
^hb;t ); (A6)

where the log excess return is approximated by:

b b
q
erhb;t+1 = ^ at (rt+1 r^t ) + ~ t (it st+1 it ) + 12 ^ at (1 ^ at )Vhtb (rt+1
q
)
b
+ 21 ~ t (1 ~ t )Vhtb ( st+1 ) + a b q
h
t ~ t CVt (rt+1 ; st+1 ): (A7)

q b
with ~ t 1 ^ at ^t ^m
t as the share in dollar bonds in household wealth, and rt+1 denoting

A3
the log return on European equity. The optimal portfolio shares are:
2 3 2 3
~t 1 rt Ehtb st+1 r^t + 12 Vhtb ( st+1 ) s
b;t
4 5= ^ hb 4 h 5 (A8)
t b b b
^ qt Ehtb rt+1
q
r^t + 12 Vhtb rt+1
q q
b;t
h

where !
b;t
h = CVhtb c^hb;t+1 + p^t w
^hb;t+1 ; vt+1 + (1 ) CVhtb ( p^t+1 ; vt+1 ) ;

b
for v = f s; rqb g and ^ htb is the conditional covariance matrix for the vector ( q
st+1 ; rt+1 )0 . The
demand for log real balances is given by:

m
^ hb;t p^t = $ + c^hb;t r^t ; (A9)

and the log consumption wealth ratio by:

1
X 1
X
c^hb;t + p^t w
^hb;tt = 1
k
+ 1 1
Ehtb i+1
(^
rt+i p^t+1+i ) + Ehtb i 1
(erhb;t+i & r^t+i 1 ): (A10)
i=0 i=1

A.1.2 Firms

The pricing problem facing the us …rm can be written as the following dynamic programming
problem:

Q(Kt ) = max Eus

t fDt =Pt + t+1 Q(Kt+1 )g ;
Ptus ;P^tus
1 1
Ptus St P^t P^tus
s.t. Dt
= Ct + C^t ;
Pt Pt Pt P^t
Ptus P^tus
Kt+1 = (1 %)Kt + t Kt Ct C^t ;
Pt P^t

where t+1 t+1;t is the stochastic discount factor between periods t and t + 1: Recall that Eus
t

denotes expectations conditioned on the information available to us …rms at the start of period t;
us : The …rst-order conditions for the us …rm’s problem are
t

Ptus 0 Ptus 1
(1 ) t+1 Q (Kt+1 )
Ptus : 0 = Eus
t Pt Pt Ct + Pt Pt Ct ; and
1
(1 ) St P^t P^tus 0
t+1 Q (Kt+1 ) P^tus
P^tus : 0 = Eus
t C^t + C^t :
P^t Pt P^t P^t P^t

A4
Q0 (Kt ) is the marginal value of capital that satis…es the envelope equation Q0 (Kt ) =
1
Eus
t [
0 k
t+1 Q (Kt+1 )Rt ], where Rtk (1 %) + t Kt . Simplifying these equations and log-
linearizing gives the following expressions for the log prices set by us …rms

pus us us
t = Et pt + mt ; and p^us us
^ us
t = Et pt + m t Eus
t st ;

where mus ^ us
t and m t are the percentage markups in the price of us goods over the expected log us

price-level, Eus
t pt . Let nt t + ln Q0 (Kt ) where t = ln t We can now identify the markups by

mus
t = Eus
t nt+1 ln 1
+ 21 Vus
t (ct + pt + nt+1 )
1 us
2 Vt (( 1) pt + ct ); and

^ us
m t = Eus
t nt+1 ln 1
+ 21 Vus ct + p^t + nt+1 )
t (^
1 us
2 Vt (( 1) pt + c^t + "t ) ;

where nt = ln + Eus k
t nt+1 + rt + t + 12 Vus
t (nt+1 ) :

The pricing problem facing the eu …rm is analogous and produces the following approximations
for the log prices of eu goods:

p^eu eu
t = Et p ^ eu
^t + m t ; and peu eu
t = Et p^t + Eeu eu
t st + mt ;

with markups

^ eu
m t = ln 1
+ Eeu
t n^t+1 + 21 Veu ^t+1 )
ct + p^t + n
t (^
1 eu
2 Vt (( 1) pt + c^t ) and

meu
t = ln 1
+ Eeu
t n^t+1 + 21 Veu ^t+1 )
t (ct + pt + n
1 eu
2 Vt (( 1) pt + ct "t ) ;

^t
where n ^ t + ln Q0 (K
^ t ) and n
^t = ln + Eeu ^t+1 + r^tk + ^ t + 21 Veu nt+1 ) with r^tk
t n t (^ ln[(1 %) +
^K^ 1
]:
t t

We can now relate the real exchange rate to the pricing decisions of …rms. In particular, if we
…rst write the real exchange rate as

8 1 1
91=(1 )
>
< Ptus1 St P^tus =Ptus + Pteu1 St P^teu =Pteu >
=
St P^t
Et =
Pt >
: Ptus1 + Pteu1 >
;

A5
and then take a log-linear approximation around the symmetric steady state of E = 1; we obtain

"t = 1
2 (st + p^us
t pus 1
^eu
t ) + 2 (st + p t peu
t );

= 1
2 (st Eus 1
t st ) + 2 (st Eeu 1 ^ us
t st ) + 2 (mt mus 1 ^ eu
t ) + 2 (mt meu
t ):

Hence, the log real exchange rate varies with the expectational errors of …rms concerning the current
spot rate, and the di¤erential in the price markups between the European and US markets.

A.1.3 Dealers

i and S ii at the start of

Dealers make four decisions each period. They choose euro price quotes Sd;t d;t

trading rounds i and ii. The initiate trades, Td;t ; against other dealers’quotes in round ii trading,
and they choose consumption, Cd;t ; after trading is complete. We refer below to this consumption
decision as the round iii decision. It proves useful to …rst consider the optimal choices of Td;t and
i and S ii are determined.
Cd;t before examining how Sd;t d;t
^i
i = Bi + SiB
De…ne Wd;t d;t t d;t as the dollar wealth of dealer d at the start of round-i in period

t; :where Sti is the price quoted by other dealers. The trading problem facing dealer d in round ii
can be written as

ii ^ ii iii ^ iii ii
V(Wd;t ; Bd;t ) = max E[V(Wd;t ; Bd;t )j d;t ] (A11)
Td;t

s.t. iii
Wd;t ii
= Wd;t ii
+ (Sd;t Stii )Td;t
ii
; and ^ iii = B
B ^ ii + Td;t ii
Td;t :
d;t d;t

^ i ) denotes the value function for the dealer de…ned over wealth, W i ; and euro bond
where V(Wdi ; Bd d
^ i : Notice that T ii 62
holdings, B ii so the dealer’s choice of trade, Td;t ; cannot be conditioned
d d;t d;t
ii ; (i.e., interdealer trade takes place simultaneously). In
on incoming orders from other dealers, Td;t
round iii dealers choose consumption so solve

n o
iii ^ iii 1 1 i ^i ii ^ iii
V(Wd;t ; Bd;t ) = max 1 Cd;t + E[V(Wd;t+1 ;Bd;t+1 )j d;t ; Bd;t ] (A12)
Cd;t
i
s.t. Wd;t+1 iii
= Rt (Wd;t i
Pt Cdt ) + St+1 ^t
R ^ iii ;
Rt Stii B and ^i
B ^ ^ iii
d;t d;t+1 = Rt Bd;t :

A6
The …rst-order conditions associated with the problems in (A11) and (A12) are

iii ^ iii ii
0 = E[V2 (Wd;t ; Bd;t )j d;t ]; and (A13)

Cd;t = i
Rt Pt E[V1 (Wd;t+1 ^i
;B ii ^ iii
d;t+1 )j d;t ; Bd;t ]; (A14)

where Vi (:; :) denotes the i0 th. partial derivative of the dealer’s value function.
Next, we consider the quote problems facing the dealer at the start of round i and ii. The round
i problem can be written as

i ^ i ) = max E[V(W ii ; B
;B ^ ii )j i
V(Wd;t d;t i d;t d;t d;t ] (A15)
Sd;t
ii
s.t. Wd;t i
= Wd;t + (Stii ^i
Sti )(B i
Td;t i
) + (Sd;t Sti )Td;t
i
; and B ^i
^ ii = B i
Td;t ;
d;t d;t d;t

and the round ii problem as

ii ^ ii iii ^ iii ii
V(Wd;t ; Bd;t ) = max
ii
E[V(Wd;t ; Bd;t )j d;t ] (A16)
Sd;t

s.t. iii
Wd;t ii
= Wd;t ii
+ (Sd;t Stii )Td;t
ii
; and ^ iii = B
B ^ ii + Td;t ii
Td;t :
d;t d;t

i cannot be conditioned on
Recall that all dealers choose quotes simultaneously, so the choice of Sd;t
i for n 6= d and i = fi,ii}. Furthermore, quotes are good for any
the quotes of other dealers, i.e., Sn;t
amount and are available to all households in round i, and all dealers in round ii. Consequently,
i
(Sd;t Sti )Td;t
i will have a limiting value of i di¤ers from S i in trading round i = fi,ii}: We
1 if Sd;t t

establish below that E[V1 (Wd;t ^ ii )j

ii ; B i ] ^ iii )j
iii ; B
and E[V1 (Wd;t ii ] are positive so dealers must
d;t d;t d;t d;t
i = S i and S ii = S ii :
quote a common price in each trading round, i.e., Sd;t t d;t t

Now we turn to the determination of Sti and Stii : A dealer will only be willing to quote at the
beginning of each trading round if doing so does not reduce his expected utility. In round ii trading,
ii ; is
the marginal utility associated with incoming orders, Td;t

iii ^ iii ii iii ^ iii

E[V1 (Wd;t ; Bd;t )(Sd;t Stii ) V2 (Wd;t ; Bd;t )j ii
d;t ]:

ii = S ii and round ii trades are chosen optimally satisfying (A13).

This term equals zero when Sd;t t

Thus, incoming orders from other dealers have no e¤ect on the dealers’ expected utility at the

A7
margin provided he has the opportunity to initiate trades and quotes a common price to avoid
arbitrage.
i ; is zero when
In round i, the marginal utility associated with incoming orders, Td;t

ii ^ ii i ii ^ ii
0 = E[V1 (Wd;t ; Bd;t )[(Sd;t Sti ) (Stii Sti )] V2 (Wd;t ; Bd;t )j ii
d;t ]:

Applying the no-arbitrage restriction of common round i quotes, and substituting the envelope
condition V2 (Wd;t ^ ii ) = E[V2 (W iii ; B
ii ; B ^ iii )j ii ] from problem (A11) gives
d;t d;t d;t d;t

ii ^ ii iii ^ iii
0 = E[V1 (Wd;t ; Bd;t )(Stii Sti ) + V2 (Wd;t ; Bd;t )j ii
d;t ];

ii ^ ii
= (Stii Sti )E[V1 (Wd;t ; Bd;t )j ii
d;t ];

where the second line follows from (A13) and the fact that both Stii and Sti are a function of common
information d ii : Thus dealers will not be made worse o¤ at the margin by incoming orders
t d;t

during round i trading if the common quote is the same in each round: Sti = Stii = St :
i = S i = S for i = fi,ii}, it is straightforward
Finally, we determine the value of St : With Sd;t t t

to establish that

i
V1 (Wd;t ^ i ) = E[V1 (W ii ; B
;B ^ ii )j i iii ^ iii i
d;t d;t d;t d;t ] = E[V1 (Wd;t ; Bd;t )j d;t ];

ii ^ ii iii ^ iii ii
V1 (Wd;t ; Bd;t ) = E[V1 (Wd;t ; Bd;t )j d;t ]; and
iii ^ iii i ^i ii ^
V1 (Wd;t ; Bd;t ) = Rt E[V1 (Wd;t+1 ;Bd;t+1 )j d;t ; Bd;t ]:

Hence, the …rst order condition in (A12) implies that

iii ^ iii
Cd;t = V1 (Wd;t ; Bd;t )Pt : (A17)

Consequently,

iii ^ iii i i iii ^ iii ii ii

E[V1 (Wd;t ; Bd;t )j d;t ] = E[Cd;t =Pt j d;t ] > 0; and E[V1 (Wd;t ; Bd;t )j d;t ] = E[Cd;t =Pt j d;t ] > 0;

A8
as noted above. We also have

h i
i
V2 (Wd;t ^ i ) = E E[V2 (W iii ; B
;B ^ iii )j ii i
+ (Stii ii ^ ii
Sti )E[V1 (Wd;t i
d;t d;t d;t d;t ] d;t ; Bd;t )j d;t ]; (A18)

and
h i
iii ^ iii i ^i i ^ ^
V2 (Wd;t ; Bd;t ) = E V1 (Wd;t+1 ;Bd;t+1 )(St+1 Rt Rt Stii ) j ii
d;t ; Bd;t

^ t E[V2 (W i
+ R ^i ii ^
d;t+1 ; Bd;t+1 )j d;t ; Bd;t ]: (A19)

When Stii = Sti ; (A18) and (A13) imply that V2 (Wd;t

i ;B^ i ) = 0: Using this result and the fact that
d;t
i
V1 (Wd;t+1 ^i
;B iii ^ iii i
d;t+1 ) = E[V1 (Wd;t+1 ; Bd;t+1 )j d;t+1 ]; (A19) becomes

h i
iii ^ iii iii ^t
^ iii )(St+1 R ii ^
V2 (Wd;t ; Bd;t ) = E V1 (Wd;t+1 ;Bd;t+1 Rt St )j d;t ; Bd;t : (A20)

Taking expectations conditioned on d; using (A13) and the law of iterated expectations, we get
t

iii ; B
^ t E[V1 (Wd;t+1
R ^ iii )St+1 j dt ]
d;t+1
St = : (A21)
iii ; B
Rt E[V1 (Wd;t+1 ^ iii )j d ]
d;t+1 t

Equation (A21) identi…es the price at which dealers are willing to …ll incoming orders for euros
in rounds i and ii based on common information, d, and the trading environment of our model.
t

To gain more perspective on its implications, we take a log-normal approximation to (A21):

d
E[st+1 st j t] + r^t rt = ; (A22)

where Vdt (st+1 ) CVdt ( ln V1 (Wd;t+1 ^ iii ); st+1 ): This is the form of equation (13) in the
iii ; B
d;t+1

text. It says that log spot rate, st ; implied by the common dealer quotes must be such that the
expected log excess return based on d compensates the dealers for …lling incoming euro orders
t

from households and other dealers.

Finally, we consider the consumption and trading decisions of each dealer. Combining the
…rst-order condition in (A14) with our results on the marginal utility of wealth gives

Cd;t = Rt E[Cd;t+1 (Pt =Pt+1 )j ii ^ iii

d;t ; Bd;t ]: (A23)

A9
This is the standard consumption-Euler equation. Notice, however, that dealers can condition their
^ iii : The optimal
period-t choices on their holdings of euro bonds after round ii is complete, i.e., Bd;t

choice of round ii trade, Td;t ; is governed by the …rst-order condition in (A13). Combining this
expression with (A20) and (A17) gives

Cd;t+1 Pt ^t ii
0= E Cd;t Pt+1 St+1 R R t St d;t : (A24)

This equation takes the form of a standard …rst-order condition governing the portfolio choice
between dollar and euro bonds.

A.1.4 Equilibrium

An equilibrium in this model is described by: (i) the consumption and portfolio decisions of house-
holds, (ii) the price-setting decisions of …rms, (iii) the interest rate decisions of central banks, (iv)
the quote, trade and consumption decisions of dealers, consistent with market clearing in the equity,
bonds, money and goods markets. Assumptions A1 and A2 imply that all the equities issued by
US and European …rms are held by the domestic representative household. Thus market clearing
^ t satisfy
implies that the ex-dividend prices of us and eu equity, Qt and Q

us equity : Qt + Dt = Pt Qus
t ; and
^t + D
eu equity : Q ^ t = P^t Qeu
t ;

where Qit is the value of the real dividend stream of …rm i = fus,eu} to domestic households under
an optimal price-setting policy. Market clearing also implies that the optimal share of equities in
households’portfolios satisfy

q
t Wh;t = Qt ; and ^ qt Wh;t = Q
^ t;

because the number of outstanding shares issued by each …rm is normalized to one.
Market clearing in the euro bond market requires that the dollar value of aggregate euro orders

A10
received by dealers in round i trading equal aggregate household order ‡ow:

d
X
i
St Td;t = xt :
d=1

In round ii all trading takes place between dealers. Hence, the dollar value of incoming orders for
euros received by all dealers must equal the aggregate dealer order ‡ow:

d
X d
X
ii
St Td;t = St Td;t :
d=1 d=1

^h,t ; eu
At the end of each period, the aggregate holdings of euro bonds by us households, B
^hb,t ; and non-households (i.e. dealers and central banks), B
households, B ^t ; must sum to zero:

^h,t + B
B ^hb,t + B
^t = 0:

In the money markets, central banks accommodate households’demand for currency at a chosen
nominal interest rate. These interest rates are set as

fed : rt = 1
f$ + Efed
t [pt + ct mt ]g = 1
f$ + pt + Efed
t ct mt g ; and

ecb : r^t = 1
f$ + Eecb
t [^pt + c^t m
^ t ]g = 1
f$ + p^t + Eecb
t c ^t m
^tg:

Market clearing in goods markets is demand-determined in each national market given the prices
chosen by …rms. Aggregate us consumption comprises the consumption of us households and
P R 1=2
dealers: Ct = Ch;t + dd=1 Cd;t where Ch;t 0 Ch;t dh: Aggregate eu consumption comprises
R1
the consumption of eu households C^t = C^hb;t ^
1=2 Ch;t dh: The implications of price-setting for

dividends via their e¤ect on consumer demand are incorporated into the …rms’ decision-making
problems.

A11
A.2 Proofs of Propositions

Proposition 1 Consider the expected demand for money by us and eu households implied by
(A5) and (A9) given prices and interest rates, conditioned on dealers’common information, d :
t

Edt mh;t pt = $ + Edt ch;t rt ; (A25a)

Edt m
^ hb;t p^t = $ + Edt c^hb;t r^t ; (A25b)

If d is a subset of the period t information available to the fed and the ecb, Edt mh;t = Edt Efed
t t mh;t

and Edt m
^ hb;t = Edt Eecb
t m ^ hb;t by iterated expectations. Combining (A25) with these restrictions and
the central banks’policy rules gives

Edt mt pt = $ + Edt ch;t rt ; (A26a)

Edt m
^t p^t = $ + Edt c^hb;t r^t : (A26b)

To derive equation (11), we …rst use (A26) to substitute for rt and r^t in (A22). This gives

st = 1+ Edt st+1 + 1
1+ Edt ft ; (A27)

where fundamentals, ft ; are de…ned in (12). Solving this equation forward and applying the law of
iterated expectations gives (11). Notice that if Edt mh;t 6= Edt Efed d^
t mh;t and Et mh
d ecb ^
b;t 6= Et Et m b;t ;
h

because central banks have less information than dealers, we still get (A27) from (A25) and (A22),
but fundamentals depend on mh;t m
^ hb;t rather than mt m
^ t : The present value expression for the
log spot rate is therefore robust to di¤erent information assumptions regarding dealers and central
banks provided we adjust the de…nition of fundamentals accordingly.

q q q q
Proposition 2 Let ert+1 rt+1 rt + 21 Vht (rt+1 ) h;t be the risk adjusted log excess return on
us equities. We may now rewrite the portfolio allocation equation in (A4) as

h h h q
t = t Eht st+1 + r^t rt + 21 Vht (st+1 ) s
h;t t Et ert+1 ; (A28)

A12
 q q
where h
t Vht (rt+1 )=j hj
t and h
t CVht (rt+1 ; st+1 )=j h j:
t Households know that dealers quote
spot rates in accordance with (11). So the expected excess return on euro bonds can be written as

Eht st+1 + r^t rt = Edt st+1 + r^t rt + rEht st+1 = rEht st+1 + :

Combining this expression with (A28) gives us

h h h h q h 1 h s
t = t rEt st+1 t Et ert+1 + t 2 Vt (st+1 ) + h;t : (A29)

In the case of European households, their desired share of wealth held in euro bonds, ^ t ; is by
de…nition equal to 1 ~t ^ qt (P^t C^hb,t + M
^ hb,t )=W
^ hb,t : Substituting for ~ t and ^ a from (A8),
t

P^t C^hb,t =W
^ hb,t from (A10) and M
^ hb,t =W
^ hb,t from (A9) in this de…nition gives

b
h b
h 1 hb s b h
h b q b
^ t = 1+ t (Et st+1 +^
rt r t 2 Vt ( st+1 )+ b;t )
h t Et ert+1 exp(^
chb;t +^
pt w
^hb;t )(1+exp ($ r^t ))
(A30)
b b b h b b h b
where h
t (Vhtb (rt+1
q
)+ h b q ^ hb
t Vt (rt+1 ))=j t j and b
h
t (Vhtb (st+1 ) + h b q ^ hb
t Vt (rt+1 ))=j t j. Proceeding
as above, we obtain

b
h b
h b h
h b q b b
h s 1 hb
^ t = 1+ t rEt st+1 t Et ert+1 + t b;t
h
+ 2 Vt (st+1 ) exp(^
chb;t +^
pt w
^hb;t )(1+exp ($ r^t )):
(A31)
Equations (A29) and (A31) show that the desired portfolio shares for euro bonds depend on: (i)
the di¤erence in expectations regarding future spot rates between the households and dealers, (ii)
the risk adjusted expected excess return on equities, (iii) risk premia; and in the case of European
households; the consumption wealth and money wealth ratios. Substituting the expressions for

t and ^ t in the order ‡ow equation (16), and linearizing around a symmetric steady state where
expectations of dealers and households are the same gives (17).

Proposition 3 Let h =f d; d h:
t t tg for some vector of variables t so that t t From Bayesian
updating we known that

! ! !
E [{t+1 j t; t] = E [{t+1 j t] + B{;v ( t E [ tj t ]) ; (A32)
1
B{;v = V!t ( t ) CV!t ({t+1 ; t ):

A13
for some random variable {t+1 and information set !: Applying this equation in the case where
t

{t+1 = E[yt+1 j d ]; ! = d; and h =f d;

t+1 t t t t t g; gives

Eht Edt+1 yt+1 Edt yt+1 = BEdt+1 yt+1 ; t ( t E[ t j d

t ]):

In the case where {t+1 = yt+1 ; ! = d; and h =f d;

t t t t tg we get:

Eht yt+1 Edt yt+1 = Byt+1 ; t ( t E[ t j d

t ]):

Combining these equations we obtain:

Eht Edt+1 yt+1 Edt yt+1 = (Eht yt+1 Edt yt+1 ); (A33)

where
BEdt+1 yt+1 ; t (B0yt+1 ; t Byt+1 ; t ) 1 0
Byt+1 ; t
:

Now we combine (20) and (A33) to give rEht st+1 = rEht yt+1 which is (21a). Applying the same
technique to the foreign forecast di¤erential gives rEhtb st+1 = ^ rEhtb yt+1 where ^ is the foreign
counterpart of : This is equation (21b). Substitution for rEht st+1 and rEhtb st+1 in (17) with these
expressions gives (22).

Proposition 4 First we use (A32) with yt+1 = {t+1 ; ! = d; and d =f d;

t t t+1 t tg to give

Edt+1 yt+1 Edt yt+1 = Byt+1 ; t ( t E[ t j d

t ]):

Next we combine this expression with (26):

d
st+1 = rt r^t + + Byt+1 ; t ( t E[ t j t ]):

Now note that the vector t denotes the new information available to dealers between the start of
periods t and t + 1. Thus, period t order ‡ow xt is an element of t: We can therefore write:

st+1 = rt r^t + + b(xt Edt xt ) + t+1

A14
where b = Byt+1 ;xt and t+1 denotes the e¤ect of other elements in t that are uncorrelated with
order ‡ow. To see how the correlation between order ‡ow and spot rates depends on the degree
of information aggregation, we simply use (22) to substitute for xt in the de…nition of Byt+1 ;xt : In
particular, we …rst write

Byt+1 ;xt Vdt: (xt ) = CVdt: yt+1 ; rEht yt+1

0 0 0
+ ^ CVdt: yt+1 ; rEhtb yt+1
0
^0 0
+ CVdt: (yt+1 ; ot ) ;

and use the identity yt+1 Edt yt+1 + E!t yt+1 Edt yt+1 + (yt+1 E!t yt+1 ) for ! = fh,h
b} to give

b = Vdt: (xt ) 1
Vdt: (rEht yt+1 ) 0 0
+ ^ Vdt: (rEhtb yt+1 )^ 0 0
+ Vdt: (xt ) 1
CVdt: (yt+1 ; ot ) :

Proposition 5 Consider the projection of ft+ on st Edt ft and the unexpected component of
order ‡ow xt Edt xt :

ft+ = s (st Edt ft ) + x (xt Edt xt ) + t+ :

Order ‡ow has incremental forecasting power when x di¤ers from zero. To show that this is
indeed the case, we …rst note that x (xt Edt xt ) + t+ must equal the projection error in (30),
"t+ ; because xt Edt xt is uncorrelated with st Edt ft : Consequently, s takes the same value as it
did in (30) and:
CV ( ft+ ; xt Edt xt )
x = :
V (xt Edt xt )

Using the identity ft+ rE!t ft+ +Edt ft+ + ( ft+ E!t ft+ ) for ! = fh,h
bg to substitute
for ft+ ; and (22) to substitute for order ‡ow, we …nd that

CV (rEht yt+1 ; rEht ft+ ) + ^ ^ CV rEhtb yt+1 ; rEhtb ft+ + CV (ot ; ft+ )
x = :
V (xt Edt xt )

The …nal step is to substitute for ft+ using the fact that ft = Cyt :

Proposition 6 First we iterate (A27) forward periods to get

X1
st = 1
1+ Edt ( 1+ )i ft+i + ( 1+ ) Edt st+ :
i=0

A15
Subtracting ( 1+ ) st from both sides and re-arranging gives

X1
Edt st+ = ( 1+ ) 1 st ( 1+ ) 1
1+ Edt ( 1+ )i ft+i :
i=0

Combining this equation with the identity st+ = Edt st+ + st+ Edt st+ ; we …nd that

X1
st+ = ( 1+ ) 1 st ( 1+ ) 1
1+ Edt ( 1+ )i ft+i + st+ Edt st+ ;
i=0
= ' (st Edt ft; ) + st+ Edt st+ ;

P 1
where ' ( 1+ ) 1 > 0 and ft; ' +1
' (1+ ) i=0 ( 1+ )i ft+i : Using (11) to substitute for
st+ Edt st+ gives equation (33).

A.3 Real-Time Estimates

We provide a brief description of how we computed the real-time estimates of a monthly log series
{. Computing the real-time estimates for a quarterly series like GDP follows analogously and is
described in detail by Evans (2005). Let {t denote the increment to the monthly value for {m( ) ;
where m( ) is the last date of month : Next, de…ne the partial sum

minfm( );tg
X
g{t {t
i=m( 1)+1

as the cumulative daily contribution to {m( ) in month : Notice that when t = m( ); the monthly
change in {m( ) ; m{
q( ) =g{ m( ) : The daily dynamics of g{ t are described by

g{ t = (1 dumt ) g{t 1 + {t ; (A34)

where dumt is a dummy variable equal to one on the …rst day of each month, and zero otherwise.
To accommodates the presence of variable reporting lags, let m(j) { denote the monthly growth
t

in { ending on day m( j) where m( ) denotes the last day of the most recently completed month

A16
and t m( ): Monthly growth in the last (completed) month is given by

m(1)
{t = (1 dumt ) m(1)
{t 1 + dumt g{t 1: (A35)

When t is the …rst day of a new month, dumt = 1; so m(1) {

m( )+1 =g{ m( ) = m{
m( ) : On all
other days, m(1) { = m(1) {
t t 1: To accommodate occasions where the reporting lag is more than
a month, we track monthly growth two months back via the recursion:

m(2) m(2) m(1)

{t = (1 dumt ) {t 1 + dumt xt 1: (A36)

Equations (A34), (A35) and (A36) enable us to de…ne the link between the daily contributions,
{t ; and data releases. Suppose the reporting lag for the release on day t is less than one month.
Then if d{ t is the released value for the growth in { during the last month on day t;

d{t = m(1)
{t : (A37)

If the reporting lag is longer than a month (but less than two),

d{t = m(2)
{t : (A38)

We incorporate the information contained in the monthly data releases on other variables is a similar
manner. (Incorporating information from quarterly data releases is more complex, see Evans 2005
for details.) Speci…cally, let zti denote the value of another series, released on day t, that relates to
activity in the last completed month. We assume that

zti = i
m(1)
{t + uit : (A39)

where uit is an i.i.d.N (0; 2)

i shock. In cases where the reporting lag is two months,

zti = i
m(2)
{t + uit : (A40)

It is important to recognize that (A37) - (A40) allows for variations in the reporting lag from data

A17
release to data release.
To complete the model, we specify the dynamics for the daily increments, {t : We assume that

k
X
m(i)
{t = i {t + t; (A41)
i=1

where is an i.i.d.N (0; 2) shock.

t

Finding the real time estimates of { requires a solution to two related problems. First, there
is a pure inference problem of how to compute E[{m( ) j t] using the signalling equations (A37)
- (A40), and the {t process in (A41), given values for all the parameters in these equations.
Second, we need to estimate these parameters. The Kalman Filtering algorithm provides a solution
to both problems. In particular, given a set of parameter values, the algorithm provides the means
to compute the real-time estimates E[{m( ) j t ]: The algorithm also allows us to construct a sample
likelihood function from the data series, so that the model’s parameters can be computed by
maximum likelihood.
To use the algorithm, we write the model in state space form. For the case where k = 1;
the dynamics described by equations (A34) - (A36) and (A41) can be represented by the matrix
equation:
2 3 2 32 3 2 3
g{t 1 dumt 0 0 1 g{t 1 0
6 7 6 76 7 6 7
6 m(1) { 7 6 76 m(1) { 7 6 7
6 t 7 6 dumt 1 dumt 0 0 76 t 1 7 6 0 7
6 7=6 76 7+6 7;
6 m(2) { 7 6 76 m(2) { 7 6 7
6 t 7 6 0 dumt 1 dumt 0 76 t 1 7 6 0 7
4 5 4 54 5 4 5
{t 0 1 0 0 {t 1 t

or, more compactly

Zt = At Zt 1 + Vt : (A42)

This is the state equation of the state space form.

The link between the data releases on { and the elements of Zt are described by (A37) and
(A38):
h i
d{t = 0 ml1t ({) ml2t ({) 0 Zt ; (A43)

where mlit ({) denotes a dummy variable that takes the value of one when the reporting lag for

A18
series { lies between i 1 and i months, and zero otherwise. The link between the releases for the
other series and elements of Zt are described by (A39) and (A40):

h i
zti = 0 1 i
i mlt (z )
2 i
i mlt (z ) 0 Zt + uit : (A44)

Stacking (A43) and (A44) for series i = 1; 2:::; g gives

2 3 2 3 2 3
d{t 0 ml1t ({)
^ ml2t ({)
^ 0 0
6 7 6 7 6 7
6 1 7 6 1 1 2 (z 1 ) 0 7 6 7
6 zt 7 6 0 i mlt (z ) 1 ml t 7 6 u1t 7
6 7=6 7 6 7;
6 .. 7 6 .. .. .. .. 7 Zt + 6 .. 7
6 . 7 6 . . . . 7 6 . 7
4 5 4 5 4 5
ztg 0 1 g
g mlt (z ) g ml 2 (z g ) 0
t ugt

or
Xt = Ct Zt + Ut : (A45)

This equation links the vector of potential data releases for day t; Xt ; to elements of Zt : The vector
of actual data releases for day t; Yt ; is related to the vector of potential releases by

Yt = Bt Xt ;

where Bt is a n (g + 1) selection matrix that “picks out” the n 1 data releases for day t:
Combining this expression with (A45) gives us the observation equation:

Yt = Bt Ct Zt + Bt Ut : (A46)

Equations (A42) and (A46) describe a state space form which can be used to …nd real-time
estimates of variable { in two steps. In the …rst, we obtain the maximum likelihood estimates of
the model’s parameters. For this purpose the sample likelihood function is built up recursively by
applying the Kalman Filter to (A42) and (A46). The second step applies the Kalman Filter to
(A42) and (A46) to calculate the real-time estimates of { using the maximum likelihood parameter
estimates:
The real-time estimates for US variables use data releases on quarterly GDP and 18 monthly
releases: Nonfarm Payroll, Employment, Retail Sales, Industrial Production, Capacity Utiliza-

A19
tion, Personal Income, Consumer Credit, Personal Consumption Expenditures, New Home Sales,
Durable Goods Orders, Construction Spending, Factory Orders, Business Inventories, the Gov-
ernment Budget De…cit, the Trade Balance, NAPM index, Housing Starts, the Index of Leading
Indicators, Consumer Prices and M1. The real-time estimates for German variables use data re-
leases on quarterly GDP and 8 monthly releases: Employment, Retail Sales, Industrial Production,
Manufacturing Output, Manufacturing Orders, the Trade Balance, Consumer Prices and M1. We
allow for 10 lags in the daily increment process when estimating real-time GDP, and 7 lags for the
other variables. These speci…cations appear to capture all the time-series variation in the data. In
particular, we are unable to reject the null hypothesis of no serial correlation in the Kalman Filter
innovations evaluated at the maximum likelihood estimates for any of our models.

A20