Order Flow and the Monetary Model of Exchange Rates:

Evidence from a Novel Data Set

Menzie D. Chinn*
University of Wisconsin, Madison
and NBER

and

Michael J. Moore**
Queen’s University, Belfast,
Northern Ireland, United Kingdom

This Version: September 14, 2010

Abstract
We propose an exchange rate model which is a hybrid of the conventional
specification with monetary fundamentals and the Evans-Lyons microstructure
approach. We argue that the failure of the monetary model is principally due to
private preference shocks which render the demand for money unstable. These shocks
to liquidity preference are revealed through order flow. We estimate a model
augmented with order flow variables, using a unique data set: almost 100 monthly
observations on inter-dealer order flow on dollar/euro and dollar/yen. The augmented
macroeconomic, or “hybrid”, model exhibits greater in-sample stability and out of
sample forecasting improvement vis a vis the basic macroeconomic and random walk
specifications.

JEL classification: D82; E41; F31; F47

Keywords: Exchange rates; Monetary model; Order flow; Microstructure; Forecasting
performance

Acknowledgments: We are grateful to Maurice Roche, Ken West, our discussant

Martin Evans, participants at the IMF Conference on International Macro Finance
(April 2008), the NBER International Finance and Macro meetings (March 2009),
seminars at University of North Carolina and University of Pittsburgh, and two
anonymous referees, and the editor for their very helpful comments.

* Corresponding author: Robert M. LaFollette School of Public Affairs; and Department of

Economics, University of Wisconsin, 1180 Observatory Drive, Madison, WI 53706-
1393. Email: mchinn@lafollette.wisc.edu .
** Queens University School of Management, Queens University of Belfast, 25 University
Square, Belfast BT7 1NN. Email: m.moore@qub.ac.uk .
1. Introduction

One of the most enduring problems in international economics is the ‘exchange rate

disconnect’ puzzle. Numerous structural or arbitrage approaches have been tried.

Prominent among them are:

a) the sticky price monetary model

b) the Balassa-Samuelson model

c) the portfolio balance model

d) purchasing power parity

e) uncovered interest parity.

The in-sample and forecasting goodness of fit of these models were evaluated by

Cheung, Chinn and Garcia Pascual (2005 (a) and (b)). Their conclusions are not

unfamiliar:

“the results do not point to any given model/specification combination

as being very successful. On the other hand, some models seem to do
well at certain horizons, for certain criteria. And indeed, it may be that
one model will do well for one exchange rate, and not for another.”

Recently, Gourinchas and Rey (2007) have used the external budget constraint to

devise a sophisticated measure of external imbalance which has forecasting power for

exchange rate changes over some horizons.1 However, the framework seems to be

limited to some of the institutional features of the US dollar and is ex-ante silent on

the timing and the composition of external adjustment between price and quantity.

The most theoretically and empirically startling innovation in the literature has been

the introduction of a finance microstructure concept – order flow – to explain

exchange rate movements. In a series of papers Evans and Lyons2 (2002, 2005,

2008), have shown that order flow contemporaneously explains a significant

1
See an extended analysis on bilateral exchange rates using this framework in Alquist and Chinn
(2008).
2
These are just examples of their work. For a fuller account, see
http://www9.georgetown.edu/faculty/evansm1/Home%20page.htm

1
proportion of the high-frequency variation in exchange rates. Though their theoretical

framework is also very convincing, it was difficult for them to evaluate its merit at

standard macroeconomic frequencies because they were working with a daily data set

over four month period. Our data set is monthly over eight years. The variables in

the monetary model – money stocks, prices, measures of output are only available at

monthly or lower frequencies. The span and frequency of our data set enables us to

nest both the Evans Lyons model and the monetary within a hybrid general

specification. Other writers, most notably Berger et al. (2008), have also obtained

access to a long run of EBS order flow data – 6 years from 1999 to 2004 – but they

do not integrate this into the conventional monetary analysis.

In Section 2 we discuss the theoretical motivation for the hybrid monetary

fundamentals-order flow model we adopt. In Section 3 we outline the characteristics

of the data we employ in this study. Section 4 replicates the Evans and Lyons (2002)

results at the monthly frequency, confirming the fact that the order flow data we use

(and the sample period examined) are representative. Our empirical methodology and

basic in-sample results are discussed in Section 5. The next section reports some of

the robustness tests implemented. Section 7 reports the preliminary results of our in-

sample and out-of-sample validation exercises that demonstrate the predictive power

of the hybrid model. The final section makes some concluding remarks.

2. Theoretical Background

The Evans-Lyons model (Evans and Lyons, 2002) introduces the portfolio shifts

model which argues that changes in exchange rates are determined by a combination

of innovations in public and private information. The latter is revealed through order

2
flow which is measured as the net of buyer over seller initiated trades in the foreign

exchange market. In their paper, new public information is empirically implemented

as innovations in the international interest differential though Evans and Lyons are at

pains to emphasise that, in principle, they are referring to all public information

relevant to exchange rate determination. They imply that this includes international

money, output and inflation differentials and all of the variables that might be

considered in the standard monetary model. The only reason that they this is not

reflected in their empirical work is that their data is daily and the only type of public

information that is available at that frequency is interest differentials. Their model is

fully solved out in Killeen, Lyons and Moore (2006) which expresses the level of

exchange rates in that paper’s equation (5):

t t
Pt = λ1 ∑ ΔRτ + λ2 ∑ X τ (1)
τ =1 τ =1

In their notation (which we do not use in the rest of this paper), Pt is the level of the

exchange rate, ΔRτ is the public information innovation at time τ , X τ is order flow at

time τ and the λi , i = 1, 2 are parameters which are explicitly solved for in that paper.

A key feature of equation (1) to which we appeal in this paper is that order flow and

public information innovations are cumulated over time. In other words, λ1 governs

the level of public information while λ2 is the sensitivity to cumulative order flow. In

short equation (1) is a cointegrating relationship.

The discussion in the Introduction summarises the consistent lack of success of the

monetary model in explaining exchange rates. Nevertheless the economics profession

persists in believing that it must hold in some form at least in the long-run. In the

theoretical appendix, a model is sketched which suggests how this might be achieved.

3
In essence, it argues that one of the parameters of the utility function is privately

known and can only be revealed through trading. In the appendix example, the

parameter which governs the individual demand for money is stochastic and indeed

follows a unit root process. Because of its non-stationarity, this effect does not wash

out in the aggregate. Because it is idiosyncratic, its impact can only be observed

through trading i.e. through order flow. This is a simplified way of thinking about the

role in exchange rate determination of portfolio balance shocks as put forward by

Flood and Rose (1999). More specifically, the existence of shocks to liquidity

demands is one of the motivations offered for the link between order flow and

exchange rates in the seminal paper by Evans and Lyons (2002). The contention of

this paper is that cumulative shocks to liquidity demand, as specified by equation (5)

in the theoretical appendix, are captured by cumulative foreign exchange order flow.

Bjønnes and Rime (2005) and Killeen, Lyons and Moore (2006) provide evidence that

exchange rate levels and cumulative order flow are cointegrated in high frequency

data. If equation (9) in the theoretical appendix were correct, exchange rate levels

should be cointegrated with both cumulative order flow and the traditional vector of

‘fundamentals’ of the monetary model at all frequencies. It has been impossible to

test this up to this point because of lack of data.

The model is merely illustrative. Any latent variable or preference shift could

conceivably perform the same function. In addition, the model must be incomplete

because it does not specify a Bayesian updating process about private information.

What it does achieve is to show that equation (1) and the monetary model are not

incompatible.

4
3. Data

The data is monthly from January 1999 to January 2007 (see the Data Appendix for

greater detail, and summary statistics). Two currency pairs are considered:

dollar/euro and dollar/yen.

An attractive feature of the data is its long span of inter dealer order flow. It is by no

means the first paper to have a long span of order flow type data – see Bjønnes, Rime,

and Solheim (2005) and its citations – but to the best of our knowledge, it is the

longest span of inter dealer order flow to be used in an academic setting. The data

was obtained from Electronic Broking Services (EBS), one of the two major global

inter-dealer foreign exchange trading platforms. It dominates spot brokered inter

dealer trading in dollar/yen and is responsible for an estimated 90% of dollar/euro

business in the same category. The two series are:

• Order Flow: Monthly buyer initiated trades net of seller initiated trades, in

millions of base currency (OFEURUSD, OFUSDJPY)

• Order Flow Volume: Monthly sum of buyer-initiated trades and seller-initiated

trades, in millions of base currency.

For dollar/euro, the base currency is the euro while the dollar is the base currency for

dollar/yen. In the empirical exercise, we standardize the data by converting

OFEURUSD into dollar terms so that the order flow variable enters into each

equation analogously.3 In some of the robustness checks, the order flow variables are

normalized by volume (also adjusted into dollar terms). The untransformed order flow

and order flow volume data are depicted in Figures 1 and 2.

3
OFUSDJPY is multiplied by a negative sign to generate the corresponding yen variable.

5
A note of caution about the definition of order flow is worth entering at this point.

We follow the convention of signing a trade using the direction of the market order

rather than the limit order. For the current data set, this is carried out electronically by

EBS and we do not need to rely on approximate algorithms such as that proposed by

Lee and Ready (1991). The reason why the market order is privileged as the source

of information is that the trader foregoes the spread in favor of immediacy when she

hits the bid or takes the offer in a limit order book. Nevertheless, an informed trader

can optimally choose to enter a limit order rather than a market order though she is

less likely to do so. For a fuller discussion of this issue, see Hollifield, Miller and

Sandas (2004) and Parlour (1998).

The other data are standard. Monthly data were downloaded from the IMF’s

International Financial Statistics. The exchange rate data used for prediction are end-

of-month. The exchange rate data used to convert order flow, as well as the interest

rate data, are period average, which is most appropriate given the order flow data are

in flow terms. In our basic formulation, money is M2 (the ECB-defined M3 for Euro

area), and income is industrial production.4

The key variables, the exchange rates and transformed order flow series are displayed

in Figures 3 and 4 for the dollar/euro and dollar/yen, respectively. Note that in these

graphs, the exchange rates are defined (dollar/euro and dollar/yen) and order flow

transformed so that the implied coefficient is positive5.

4
As noted in Section 6, we also check to see if the results are robust to use of M1 as a money variable,
or real GDP (at the quarterly frequency) as an activity variable. M1 and real GDP are also drawn from
IFS.
5
Note that we have also run the regressions with the raw order flow and cumulative demeaned raw
order flow data. The qualitative aspects of the regression results do not change – order flow remains
important in both a statistical and economic sense.

6
4. Replicating the Evans-Lyons Results

In order to verify that the results we obtain are not driven by any particular

idiosyncratic aspects of our data set, we first replicate the results obtained by Evans

and Lyons (2002). They estimate regressions of the form

Error! Reference source not found..

∆ ∆ (2)

Where i are short term nominal interest rates and of is order flow. The estimates we

obtain are reported in Table 1. Several observations are noteworthy. First, the

proportion of variation explained goes up substantially when order flow in levels is

included. Second, the interest differential coefficient is only statistically significant

(with the anticipated sign6) when the order flow variables are omitted, and then only

in the dollar/euro case. Inclusion of the order flow variables reduces the economic and

statistical significance of the interest rate differential in this case. In short, any

suspicion that the Evans-Lyons result is an artefact of high-frequency data is firmly

dispelled. The results are, however, consistent with those of Berger et al. (2008) who

argue that the Evans Lyons result is relatively weaker at lower frequencies.

5. Empirics

We implement the rest of the portion of the paper in the following manner.

a) The Johansen Procedure is applied to test for cointegration between the

exchange rates, cumulative order flow and conventional monetary model

6
The negative slope is consistent with a sticky price monetary model story, though not, of course with
uncovered interest parity.

7
 fundamentals (here taken to be the flexible-price model determinants – money,

income and interest rates).

b) Weak exogeneity of order flow is tested for.

c) The implied single equation error correction model is estimated.

d) Out of sample forecasts for different models are compared.

5.1 Testing for Cointegration

All the monetary fundamentals – money, industrial production and interest rate

differentials – and cumulative order flow, appear to be integrated of order one (see the

Data Appendix).7 The first step in the cointegration test procedure is to determine the

optimal lag length. We evaluated the VAR specifications implied by the monetary

model and the monetary model augmented by the order flow variable (in this case

cumulated). We term this latter version the “hybrid” model.

The Akaike Information Criterion (AIC) typically selects a fairly short lag length of

one or two lags in the VAR specification. However, these specifications also typically

exhibit substantial serial correlation in the residuals, according to inspection of the

autocorrelograms up to lag 6. In contrast, the residuals appear serially uncorrelated

when three lags are included in the VARs. Hence, we fix on the three lag

specification.

Using this lag length, we applied the Johansen (1988) maximum likelihood procedure

to confirm that the presence of cointegration, and to account for the possibility of

7
We use the Elliott-Rothenberg-Stock DF-GLS test (Elliott, Rothenberg and Stock, 1996), allowing for
constant and trend. The ERS unit root test is more powerful than the standard ADF test. In no case is
the unit root null rejected for the levels data. In all cases – save the US-euro area interest differential –
the unit root null is rejected for the first differenced data. Even in the case of the interest differential,
the non-rejection is borderline.

8
multiple cointegrating vectors. Table 2 reports the results of these tests. The first

three columns of Table 2 pertain to specifications including only flexible price

monetary fundamentals. Columns 4-6 pertain to the monetary model augmented with

cumulative order flow. Columns [1] and [4] pertain to model specifications allowing a

constant in the cointegrating equation, columns [2] and [5] to ones allowing a constant

in both the cointegrating equation, and in the VAR, and columns [3] and [6] allowing

intercept and trend in the cointegrating equation, and a constant in the VAR (in all but

columns [1] and [4], deterministic time trends are allowed in the data).

The numbers pertain to the implied number of cointegrating vectors using the trace

and maximal eigenvalue statistics (e.g., “3,1” indicates the trace and maximal

eigenvalue statistics indicate 3 and 1 cointegrating vectors, respectively). Since the

number of observations is not altogether large relative to the number of coefficients

estimated in the VARs, we also report the results obtained when using the adjustment

to obtain finite sample critical values suggested by Cheung and Lai (1993). Hence,

“Asy” entries denote results pertaining to asymptotic critical values, and “fs”, to finite

sample critical values.

Inspection of Table 2 indicates that it is not easy to find evidence of cointegration,

using only monetary fundamentals (money, income and interest rate differentials).

The specification selected by the AIC for the monetary model is one that includes a

constant in the cointegrating equation and the VAR equation for the dollar/euro, and

one including a constant and trend in the cointegrating vector and a constant in the

VAR, for the dollar/yen. In both instances, there is no evidence of cointegration

detected.

9
In contrast, for the hybrid model, the AIC indicate the presence of a constant in both

the cointegrating relation for the dollar/euro, and a constant in both the cointegrating

relation and VAR for the dollar/yen. Using the finite sample critical values does not

change the conclusions. The evidence for cointegration is relatively strong for the

dollar/euro, and limited for the dollar/yen.

The resulting conclusions are suggestive that there is one cointegrating vector in

almost all cases, at least insofar as the hybrid model is concerned. Hence, we proceed

in our analysis assuming only one cointegrating vector.8 This conclusion points to an

important role for cumulative order flow in determining long term exchange rates but

only in combination with monetary fundamentals.

5.2 Estimating the Error Correction Models

We estimate the short run and long run coefficients in an error correction model

framework, focusing on the exchange rate equation.

Δst = ΔX t −1Γ + ρ1Δst −1 + ρ 2 Δst −1 + φ ( st −1 − X t −1 B ) + vt (2)

Where X is a vector of monetary fundamentals and cumulative order flow, and φ

should take on a negative value significantly different from zero, if the exchange

rate responds to disequilibria in the fundamentals. B is the vector of

cointegrating coefficients.

8
Note that while we could rely upon the Johansen procedure to obtain estimates of the long run and
short run coefficients, we decided to rely upon estimation of the single equation error correction
specification, in large part because the estimates we obtained via the Johansen procedure were so
implausibly large, and sensitive to specification.

10
Equation (2) invokes the Granger representation Theorem (Engle and Granger, 1987).

This enables us to assert that a cointegrating regression of the kind discussed in

section 5.1 has an error correction representation. Note that equation (2) does not

provide for contemporaneous order flow nor indeed any contemporaneous first

differenced variable to enter the ECM. However, this specification is implied if order

flow is weakly exogenous for the cointegrating vector (Johansen, 1992). We can test

for this condition using a likelihood ratio test on the restriction that order flow does

not respond to deviations from equilibrium (Johansen and Juselius, 1990).

The test results are reported in Table 2.2. Using asymptotic critical values, the weak

exogeneity assumption is rejected in a couple of cases, using the 5% significance

level. However, Bruggeman (2002) notes that in small samples, the likelihood ratio

test is mis-sized; using the suggested adjustment to the Chi-squared statistic9, we

obtain the test statistics reported in Table 2.3. Now, we fail to reject weak exogeneity

in all cases.

Using the theorem of Johansen (1992), we implement equation (2) by adding the

current value of order flow as a right hand side variable. Conveniently, the Johansen

result also enables us to estimate using OLS without instrumenting the current value

of a weakly exogenous variable.

We estimate (3) using nonlinear least squares, with two lags of first differenced

monetary fundamentals. When the order flow fundamentals are introduced, they are

incorporated first contemporaneously, then as a contemporaneous variable and a

9
Bruggeman (2002) suggests adjusting the likelihood ratio test statistic by (T-k)/T, where T is sample
size, and k = r+r+K(p-1)+1; r is the rank, p is the number of lags in the model, K is the number of
variables.

11
lagged cumulative variable, and then finally with both these variables, as well as two

lags of the order flow variable.

One could adopt a general-to-specific methodology with the objective of identifying a

parsimonious specification. Typically, such an approach leads to error correction

models with short lags (a lag or at most two of first differenced terms), with perhaps

income variables omitted. In order to maintain consistency of specifications across

models, we present the results of models incorporating two lags of the differenced

monetary fundamentals.

5.3 Long- and Short-Run Coefficients

The results of estimating these equations for the dollar/euro and dollar/yen are

reported in Tables 3 and 4, respectively.10,11 Note that the error correction term is in

all cases negative and statistically significant. This implies that the exchange rate

reverts to a conditional mean, confirming some form of long run linear relationship.

Since the estimation procedure does not necessarily lead to consistent estimates of the

standard errors for the long run coefficients, we report the coefficient estimates

obtained by implementing dynamic OLS, or DOLS (Stock and Watson, 1993).12

10
We rely upon a single equation estimation methodology focused on the exchange rate as the
dependent variable, which is appropriate if the “fundamentals” are weakly exogenous. We tested for
this condition, and this is typically the case.
11
In all cases, the specifications pass diagnostics for serial correlation, as indicated by the Q-statistics
and Breusch-Godfrey LM test.
12
We use two leads and two lags of the right hand side variables in the DOLS regressions. The long
run covariance estimate incorporates a Bartlett kernel, with Newey-West bandwidth set to 4). Point
estimates and standard error estimates obtained using the Phillips-Hansen FMOLS procedure are
similar to these DOLS estimates.

12
Turning first to Table 3, columns [1]-[2], one finds little evidence that the exchange

rate reacts to the long run monetary fundamentals, at least in the manner indicated by

the simple monetary model (note that while order flow is included in columns [2] , it

is not in the cointegrating relation). The money stock variable coefficient points in the

wrong direction. All the other coefficients are not statistically significant.

In column [2], order flow is included contemporaneously. It enters into the

determination of the exchange rate in an important manner; the proportion of variation

explained rises dramatically, from 0.02 to 0.34. The estimated short run effect is

1.889, indicating that a $1 billion dollar increase in order flow leads to a dollar

appreciation of approximately 0.2 percent, or 20 basis points. This is somewhat lower

than Evans and Lyons’ (2002) estimate.

The cointegration tests suggest that cumulative order flow does enter into the

cointegrating relationship. The specification in column [3] conforms to that

specification. In this case, money is now non-significant, while income is wrong-

signed.

The error correction specification, allowing the cumulative order flow to enter into the

long run relationship, explains an even larger proportion of variation in the exchange

rate change (37%). Finally, allowing the inclusion of two lags of order flow slightly

raises the proportion of variation explained (38%), although the lagged order flow and

lagged cumulative order flow coefficients are not significant.

13
Turning to the dollar/yen results in Table 4, one finds in column [1] a significant error

correction term, although the money coefficient is again wrong-signed. However, the

equation does not explain a large proportion of variation. Only when the

contemporaneous order flow variable is included (column [2]) does the fit improve

substantially, to 57%. Interestingly, in the case of dollar/yen rate, the inclusion of the

cumulative order flow in the long run relationship (columns [3]-[4]) does not have a

substantial impact on the equation’s explanatory power. While the specification in

column [4] is consistent with the cointegration test results for the hybrid model, it is

interesting that cumulative order flow fails to exhibit statistical significance.

To sum up the results from this section, there does appear to be significant evidence

of a long run relationship between exchange rates and monetary fundamentals

augmented by cumulative order flow. Even when cumulative order flow might be

argued to not enter into the long run relationship (i.e., in the case of the dollar/yen), it

is clear that order flow does enter into the short run relation.

6. Robustness Tests

We have investigated a number of variations to the basic specifications, to check

whether the empirical results are robust.

• Order flow vs. normalized order flow

• M1 vs M2

• Inclusion of inflation

• Quarterly vs. monthly data

We deal with each of these issues in turn.

14
Order flow issues. The order flow variables are included in dollar terms. It is

reasonable to scale net order flow variable by the volume of order flow. The results in

the Evans and Lyons regressions are basically unchanged. Using this normalized

order flow variable in the hybrid model specifications (conforming to columns [2]-[3]

in Tables 3 and 4) does not result in any appreciable change in the results.13

Money measures. While the substitution of narrow money for M2 results in slightly

different results, particularly with respect to the short- and long-run coefficients on

the money variable, the impact on the general pattern of estimates is not significant. In

particular, the coefficient on the cumulative order flow variable remains significant.

Inclusion of inflation. Inclusion of an inflation measure would be consistent with a

sticky-price monetary model. Over the given sample period, the inflation differential

appears to be stationary, so inclusion in the long run cointegrating relationship would

not be justified. However, as a practical matter, re-estimating the cointegrating

relationships with inflation leads to slightly greater evidence of cointegration, but no

substantive changes in the findings regarding the importance of order flow variables.14

Quarterly data. At the cost of considerable reduction in the number of observations,

one can switch to quarterly data. The benefit is that one can then use real GDP as a

measure of economic activity, rather than the more narrow industrial production

13
Another point related to order flow is that net order flow is positive in the raw data. This could be
ascribed to a data recording error. As long as the level of order flow enters in the level in the error
correction specification, then only the constant is affected. However, when the cumulated order flow
enters into the long run relationship, a deterministic trend is introduced. We can address this by
allowing a deterministic trend in the data. A direct way to address this issue is by demeaning the raw
order flow data. Using demeaned order flow has no impact on the order flow coefficient, but changes
substantially the long run coefficient on cumulative order flow.
14
A previous version of this paper incorporated sticky-price monetary fundamentals, with inflation
measured as annualized month-to-month CPI growth rates. Using 3 month growth rates yielded similar
results.

15
variable. As a check, we re-estimated the error correction models (both in a

constrained version, using nonlinear least squares, and in an unconstrained version

using OLS). What we find is that we recover the same general results as that obtained

using the monthly data. While money coefficients remain wrong-signed (as do income

variables for the yen), the order flow and cumulative order flow variables show up as

economically and statistically significant.

7. Model Validation

We approach model validation in two ways. First, is to examine the in-sample

stability of the monetary versus hybrid models. The second is a comparison of out-of-

sample forecasting performance.

One way to assess in sample stability is to conduct tests on recursive one-step-ahead

residuals. In recursive least squares the equation is estimated repeatedly, incrementing

the sample observation by observation, with the parameter estimates updated with

each additional observation. The recursive residual for period t is the actual minus

predicted based on the parameter estimates obtained on the sample up to t-1. This

process of recursive estimation is repeated until all the sample points have been used.

If the estimated model is valid, then the resulting errors should be i.i.d., and normally

distributed. The one-step ahead forecast error resulting from these sequential

predictions can then be tested, after scaling by the standard deviation, to see if they

conform to the posited distribution. Rejection of the null hypothesis of independence

and normality indicate parameter instability (Kianifard and Swallow, 1996).

16
In Figures 5-8, the recursive residuals and ±2 standard error bands are illustrated for

the error correction models.15 We compare the monetary against hybrid models

(respectively, specifications in columns [1] and [4] of Tables 3 and 4), for the

dollar/euro and dollar/yen. In Figure 5, the dollar/euro monetary model exhibits

substantial instability, with nine structural breaks indicated by the one-step ahead

recursive residuals, using the 10% msl. In contrast, the hybrid model residuals, shown

in Figure 6, indicate only five breaks (only two, using the 5% msl). For the

dollar/yen, the differences are not as striking. Nonetheless, using the 15% msl, the

hybrid model (in Figure 8) exhibits fewer breaks than the monetary model (Figure 7).

Furthermore, the n-step ahead recursive residuals test (essentially a sequence of Chow

tests) indicates instability at the beginning and end of the samples for the monetary

model. No such instability is indicated for the hybrid model.

As is well known in the exchange rate literature, findings of good in-sample fit do not

often prove durable. Hence, we adopt the convention in the empirical exchange rate

modeling literature of implementing “rolling regressions.” That is, the error correction

models are estimated over an initial data sample up to 2003(12) , out-of-sample

forecasts produced, then the sample is moved up, or “rolled” forward one observation

before the procedure is repeated. This process continues until all the out-of-sample

observations are exhausted. Note that this is sometimes referred to as a historical

simulation, as the ex post realizations – as opposed to ex ante values – of the right

hand side variables are used. In this sense, our exercise works as a model validation

exercise, rather than a true forecasting exercise.

15
These tests are applied to the unconstrained error correction models estimated by ordinary least
squares, rather than the constrained equations estimated by nonlinear least squares.

17
To standardise the results, we generate our forecasts for the monetary model from the

simple specifications of column (1) in both Tables 3 and 4. For the hybrid model, we

use column (4) from both Tables16.

Forecasts are recorded for horizons of 1, 3, and 6 months ahead. We could evaluate

forecasts of greater length, but we are mindful of the fact that the sample we have

reserved for the out of sample forecasting constitutes only three years worth of

observations. Forecasts at the 3 month horizon for the random walk, monetary and

hybrid models are presented in Figures 9 and 10.

One key difference between our implementation of the error correction specification

and that undertaken in some other studies involves the treatment of the cointegrating

vector. In some other prominent studies, the cointegrating relationship is estimated

over the entire sample, and then out of sample forecasting undertaken, where the short

run dynamics are treated as time varying but the long-run relationship is not. This

approach follows the spirit of the Cheung, Chinn and Garcia Pascual (2005b) exercise

in which the cointegrating vector is recursively updated.

The results for the dollar/euro are reported in Table 5.1. Mean error, standard errors,

Theil U statistic, and the Clark-West statistic (2007) are reported. The Theil U

statistic is the ratio of the model RMSE to the benchmark model (in this case random

walk) RMSE. Ratios greater than unity indicate the model is outpredicted by the

benchmark model. The Clark-West statistic is a test statistic that takes into account

16
All of these regressions contain the lagged interest differential as regressors. Consequently a
forecasting model based on carry trade returns is nested within the exercise.

18
estimation error, and is normally distributed at 0 under the null that the forecasts from

the model and benchmark model are of equal predictive capability.

The first two rows pertain to the no-drift random walk forecast. The next two blocks

of cells pertain to the monetary model, and the hybrid model. The final block is the

Evans-Lyons model, which we include for purposes of comparison. Note that the

Evans-Lyons model does not incorporate a long run relationship incorporating

cumulated order flow.17

Turning first to the dollar/euro exchange rate, notice that monetary model does very

badly relative to the random walk at all horizons. The ratio of the monetary model to

the random walk RMSE (the Theil U-statistic) is 1.6, 1.6 and 2.0 at the 1, 3 and 6

month horizons. In contrast, the mean error is smaller for the hybrid model at all

horizons, and Theil statistic (vis a vis the random walk) is much smaller: 1.1, 0.9, and

1.0. The relative performance of the 3-month-ahead forecasts (random walk,

monetary, hybrid) are shown in Figures 9 and 10.

Perhaps more remarkable, the RMSE for the hybrid model is smaller than the random

walk at the 3 and 6 month horizons. The upward bias in the model-based RMSE

versus the random walk RMSE (see Clark and West, 2007) suggests that the hybrid

models exhibit noticeable improvement vis à vis the random walk benchmark.

Unfortunately, inspection of the Clark-West statistic indicates that the hybrid model

never outperforms the random walk at conventional significance levels.18 The Evans-

17
The particular specification we use conforms to columns [3] and [7] in Table 1.
18
An order flow augmented sticky-price hybrid model does outperform a random walk at the 17%
significance level, at the 3 and 6 month horizons.

19
Lyons model does particularly badly at all horizons, but the performance is only

statistically worse than that of the random walk at the 6 month horizon.

The results are somewhat different in the case of the dollar/yen. There, by the RMSE

criterion, the hybrid model substantially outperforms the monetary model at all

horizons. However, the Evans-Lyons specification in this case does best, with the

lowest Theil statistic at horizons at all horizons. The specification in column [4] can

be beaten in certain cases. For the dollar/yen, a specification conforming to column

[2] outperforms a random walk, according to the Clark-West statistic, at the 5% msl.

Interestingly, all the structural models outperform the random walk benchmark – after

accounting for estimation error. That being said, only the monetary model at the one

month horizon comes close to significantly outperforming a random walk. A

noticeable feature is that the hybrid model is the only one that that returns a positive

Clark-West statistic at all horizons for both currency pairs.

8. Conclusion

We have laid out a simple and transparent framework in which non-stationary private

liquidity preference shocks give rise to instability in the demand for money and the

apparent failure of the monetary model of exchange rates. Cumulative order flow

tracks these shocks and is a candidate for the ‘missing link’ to augmenting the

explanatory power of conventional monetary models. We show that the hybrid model

beats both the monetary model and a random walk in a simple forecasting exercise.

Berger et al. (2008) concluded that while order flow plays a crucial role in high-

20
frequency exchange rate movements, its role in driving long-term fluctuations is much

more limited. We contend that this conclusion is premature.

In summary, we find substantial evidence to support our proposition that order flow is

an important variable in exchange rate determination, whose role can be rationalized

on the basis of a straightforward macroeconomic model. One of the appealing

implications of the household optimizing problem as specified in equations (4) to (7)

is that consumption in country j also depends on the unit root parameter ξ t j . This

means that the international consumption differential depends on ξ tH − ξ tF and

therefore on order flow from our interpretation. This may go some distance to explain

the international consumption correlations puzzle. However, we leave this to later

work.

21
 Data Appendix

For the conventional macroeconomic variables, monthly frequency data were

downloaded from International Financial Statistics (accessed November 4, 2007).

End of month data used for exchange rates when used as a dependent variable.

Interest rates are monthly averages of daily data, and are overnight rates (Fed Funds

for the US, interbank rates for the euro area, and call money rate for Japan). In the

basic regressions, money is M2 (the ECB-defined M3 for Euro area), although

specifications using M1 were also estimated. Income is proxied by industrial

production. Money and industrial production are seasonally adjusted.

Order flow was obtained from Electronic Broking Services (EBS). In order to make

the specifications consistent across currencies, the order flow data is converted to

dollar terms by dividing by the period-average exchange rate (for OFEURUSD) and

by putting a negative in front (for OFUSDJPY). Hence, the exchange rates are defined

(USD/EUR, USD/JPY) and order flow transformed so that the implied coefficient is

positive19. In the regression results (Tables 1, 3 and 4), the order flow variable is

divided by 1,000,000.

In some unreported regressions, the order flows are normalized by volume. Order

flow volume was also converted to dollar terms, in the same manner that order flow

was converted.

19
Any differences in results caused by the choice of numeraire would arise from Jensen’s inequality.
This is, of course, second order.

22
For the quarterly regressions (not reported), we use end-of-period exchange rates, and

the last month of each quarter for interest rates and inflation rates. The income

variable is US GDP (2000$), and for Euro area and Japan, GDP volume (1995 ref.).

Table A1: Summary Statistics for Dollar/Euro

Sample: 1999M01 2007M01 
DLXEU  DM2_EU  DY_EU  DI_EU  Z1EU 

 Mean  0.00134  ‐0.00087 ‐0.00014 2.02E‐05 10925.54 

 Median  ‐0.00131  ‐0.00089 0.000887 0.000155 10871.6 
 Maximum  0.069179  0.010421 0.018158 0.00308 33243.27 
 Minimum  ‐0.05029  ‐0.01655 ‐0.02321 ‐0.00793 ‐18058.8 
 Std. Dev.  0.027138  0.006239 0.008525 0.001907 9391.482 
 Skewness  0.413441  ‐0.3023 ‐0.31197 ‐1.55544 ‐0.46969 
 Kurtosis  2.695157  2.526386 2.634583 7.260148 3.405607 

 Observations  96  96 96 96 96 

Note: D denotes first difference; LXEU is log dollar/euro exchange rate; M2_EU is US-euro
area M2 log difference; Y_EU is US-euro area industrial production log difference; I_EU is
US-euro area overnight interest differential, in decimal form. Z1EU is order flow; Order flow
variables here expressed in trillions of USD per month.

Table A2: Summary Statistics for Dollar/Yen

Sample: 1999M01 2007M01 
DLXJP  DM2_JP  DY_JP  DI_JP  Z1JP 

 Mean  ‐0.00053  0.003197 1.66E‐05 5.71E‐05 ‐13227.2 

 Median  ‐0.00215  0.002828 0.000868 9.00E‐05 ‐13359 
 Maximum  0.051271  0.017703 0.045435 0.00266 5844 
 Minimum  ‐0.06808  ‐0.00706 ‐0.03498 ‐0.00589 ‐32780 
 Std. Dev.  0.02504  0.004129 0.012944 0.0019 8208.162 
 Skewness  ‐0.06316  0.513087 0.238788 ‐1.33065 ‐0.02836 
 Kurtosis  2.865523  4.148955 4.003145 4.863387 2.744249 

 Observations  97  97 97 97 97 

Note: D denotes first difference; LXJP is log dollar/yen exchange rate; M2_JP is US-Japan
area M2 log difference; Y_JP is US-Japan industrial production log difference; I_JP is US-
Japan overnight interest differential, in decimal form. Z1JP is order flow; Order flow variables
here expressed in trillions of USD per month.

For each variable EU, denotes Euro Area, and JP denotes Japan, relative to the United

States variable. LX## is the log exchange rate, M2_## is the relative log M2 money

23
stock, Y_## is the relative log industrial production, I_## is the relative short term

interest rate, Z1## is order flow, and CUMZ1## is cumulated order flow.

Tables A3 and A4 report unit root tests for the variables and their first differences.

Table A3: Unit Root Tests for Dollar/Euro Variables

LXEU  M2_EU  Y_EU  I_EU  CUMZ1EU
Levels 
ERS  ‐1.1623  ‐0.6522  ‐1.5631  ‐1.6081 0.2052
Lag   0  0  3  3 1
First Differences 
ERS  ‐7.6123  ‐6.9473  ‐4.4548  ‐2.5583 ‐6.9923
Lag   0  0  2  2 0
Notes: Elliot-Rothenberg-Stock DF-GLS test statistics, allowing for constant, trend. Lag
length selected using Schwartz Bayesian Criterion, allowing up to maximum 11 lags.
Bold face denotes significant at 10% msl. Critical values are *(**)[***] for -2.752, (-3.043) [-
3.595].

Table A4: Unit Root Tests for Dollar/Yen Variables

LXJP  M2_JP  Y_JP  I_JP  CUMZ1JP
Levels 
ERS  ‐1.8134  ‐1.8644  ‐1.9935  ‐0.94 ‐1.1361
Lag   0  0  1  1 0
First Differences 
‐
ERS  ‐6.0799  ‐5.2745  14.2133  ‐4.2849 ‐8.8864
Lag   0  1  0  0 0
Notes: Elliot-Rothenberg-Stock DF-GLS test statistics, allowing for constant, trend. Lag
length selected using Schwartz Bayesian Criterion, allowing up to maximum 11 lags.
Bold face denotes significant at 10% msl. Critical values are *(**)[***] for -2.752, (-3.043) [-
3.595].

24
 Theoretical Appendix

Let the utility function be the following special case of a CES function:

θ
θ −1 θ −1
⎡ θ −1 ξj ⎤
⎢( C j ) θ + e θ
t
⎛ M j
⎞ θ
⎥
⎜ j ⎟ ⎥
t
⎢ t ⎝ t ⎠ ⎥
P
∞
⎢
E0 ∑ δ t ⎣ ⎦ (3)
t =0
θ
θ −1

Where j = H , F for home and foreign respectively; Ct j is consumption at time t; M t j

is nominal money balances and Pt j is the price of Ct j . θ , δ and ξt j are parameters.

The CES parameter, θ , and the discount rate, δ , are common knowledge but the

parameter governing the demand for money is idiosyncratic and follows a unit root

process as follows:

ξt j = ξt −j 1 + ε t j (4)

Where ε t j is an i.i.d. random error with the property that Cov ( ε rH , ε sF ) = 0 ∀r , s .

The idea that preference shocks can used to explain asset pricing is not eccentric.

This is the main concept behind Campbell and Cochrane (1999) which has already

been applied to an exchange rate setting by Moore and Roche (2010) as well as

Verdelhan (2010).

Equation (3) is maximised subject to the budget constraint:

Bt j
Wt j = Pt j Ct j + M t j + (5)
1 + it j

Where it j is the nominal return on one period riskless bonds and Bt j is the number of

bonds held. Wt j is wealth, the only state variable and the control variables are Ct j , M t j

and Bt j ., The equation of motion for Wt j is:

25
 Wt +j1 = Pt +j1Yt +j1 + Bt j + M t j (6)

Where Yt j is labor income.

The solution to this is straightforward and the demand for money (using lowercase

symbols to represent the natural log of a variable) is20:

mtj − ptj = ξt j + ctj − θ rt j (7)

Denoting the home price of foreign currency as st and using PPP, st = ptH − ptF , we

have:

st = ⎡⎣( mtH − mtF ) − ( ctH − ctF ) + θ ( rt H − rt F ) ⎤⎦ − {ξtH − ξtF } (8)

The terms in the square brackets on the right hand side of equation (8) constitute a

standard way of expressing the monetary model. The novel feature is the final term in

curly brackets. Assuming the substitution semi-elasticity of the demand for money, θ

, is ‘small’, variations in velocity for each country’s will be largely driven by ξt j . The

‘exchange rate disconnect’ puzzle is here explained by instability in the demand for

money itself. Since the parameters ξt j (and their relation), are unknown in advance,

they can only be revealed though the act of trading itself i.e. through foreign exchange

order flow.

⎛ ij ⎞
20
In equations (7) and (8), rt j = Log ⎜ t j ⎟ .
⎝ 1 + it ⎠

26
References

Alquist, Ron and Menzie Chinn (2008). “Conventional and Unconventional

Approaches to Exchange Rate Modeling and Assessment,” International Journal of
Finance and Economics 13, pp. 2-13.

Berger, David, Alain Chaboud, Sergei Chernenko, Edward Howorka, and Jonathan
Wright (2008). “Order Flow and Exchange Rate Dynamics in Electronic Brokerage
System Data,” Journal of International Economics, 75, pp. 93–109.

Bjønnes, Geir H and Dagfinn Rime (2005), “Dealer Behavior and Trading Systems in
the Foreign Exchange Market", Journal of Financial Economics, vol. 75, issue 3
(March), pp. 571-605.

Bjønnes, Geir H., Dagfiin Rime and Haakon O.Aa. Solheim (2005), "Liquidity
Provision in the Overnight Foreign Exchange Market", Journal of International
Money and Finance, 24(2), pp. 177-198.

Brüggeman, Ralf (2002), “On the Small Sample Properties of Weak Exogeneity
Tests in Cointegrated VAR models,” Sonderforschungsbereich 373 2002-2, (Berlin:
Humboldt University, January).

Cheung, Yin-Wong, Menzie D. Chinn, and Antonio Garcia Pascual (2005), “What do
we know about recent exchange rate models? In-sample fit and out-of-sample
performance evaluated,” In: DeGrauwe, P. (Ed.), Exchange Rate Modelling: Where
Do We Stand? MIT Press for CESIfo, Cambridge, MA, pp. 239-276 (a).

Cheung Yin-Wong, Menzie D. Chinn and Antonio Garcia Pascual (2005), “Empirical
exchange rate models of the nineties: Are any fit to survive?” Journal of
International Money and Finance ,24, pp 1150-1175 (b).

Cheung, Yin-Wong and Kon. S. Lai (1993), “Finite-Sample Sizes of Johansen's

Likelihood Ratio Tests for Cointegration,” Oxford Bulletin of Economics and
Statistics, 55(3), pp. 313-328.

Clark, Todd E. and Kenneth D. West (2007), “Using Out-of-Sample Mean Squared
Prediction Errors to Test the Martingale Difference Hypothesis,” Journal of
Econometrics, forthcoming.

Diebold, Francis X. and Roberto Mariano (1995), “Comparing Predictive Accuracy,”

Journal of Business and Economic Statistics 13, pp. 253-265.

Elliot, Graham, Thomas J. Rothenberg, and James H. Stock (1996), “Efficient Tests
for an Autoregressive Unit Root,” Econometrica 64(4): 813-36.

Engle, Robert F., and Clive W.J. Granger (1987), “Co-Integration and Error
Correction: Representation, Estimation and Testing”, Econometrica, 55, pp 251-276.

27
Evans, Martin and Richard K. Lyons (2002) “Order Flow and Exchange Rate
Dynamics. Journal of Political Economy, 110(1), pp. 170–180.

Evans, Martin and Richard K. Lyons (2005), “Meese-Rogoff Redux: Micro-based

Exchange Rate Forecasting”, American Economic Review Papers and Proceedings,
pp. 405-414.

Evans, Martin and Richard K. Lyons (2008), “How Is Macro News Transmitted to
Exchange Rates?”, Journal of Financial Economics, 88, pp. 26-50.

Flood, Robert P. and Andrew K. Rose (1999), “Understanding Exchange Rate

Volatility without the Contrivance of Macroeconomics”, The Economic Journal
109(459), Features. (Nov.), pp. F660-F672.

Gourinchas, Pierre-Olivier and Helene Rey (2007), “International Financial

Adjustment,” Journal of Political Economy 115(4), pp. 665-773.

Hollifield, B., R. Miller, and P. Sandas, (2004), “Empirical Analysis of Limit Order
Markets”, Review of Economic Studies, 71(4), pp. 1027 - 1063.

Johansen, Søren, (1992). “Co-integration in Partial systems and the Efficiency of

Single Equation analysis”, Journal of Econometrics, 52(3), pp. 389-402.

Johansen, Søren, and Katerina Juselius (1990), “Maximum likelihood estimation and
inference on cointegration – with applications to the demand for money,” Oxford
Bulletin of Economics and Statistics 52(2): 169–210.

Kianifard, Farid, and William H. Swallow (1996), “A Review of the Development

and Application of Recursive Residuals in Linear Models,” Journal of the American
Statistical Association 91 (433): 391-400.

Killeen, William, Richard K. Lyons and Michael Moore (2006), “Fixed versus
Flexible: Lessons from EMS Order Flow”, Journal of International Money and
Finance 25, pp. 551-579.

Lee, C., and M. Ready, (1991), “Inferring Trade Direction from Intraday Data”,
Journal of Finance, 46(2), 733-46.

Meese, Richard, and Kenneth Rogoff (1983), “Empirical Exchange Rate Models of
the Seventies: Do They Fit Out of Sample?” Journal of International Economics 14,
pp. 3-24.

Moore, Michael J. and Roche, Maurice J. (2010), “Solving Exchange Rate Puzzles
with neither Sticky Prices nor Trade Costs”, Journal of International Money and
Finance 29(6), pp. 1151-1170.

Parlour, Christine A., (1998) “Price Dynamics in Limit Order Markets”, Review of
Financial Studies, Volume 11(4), pp. 789-816.

28
Stock, James, and Watson, Mark, 1993, “A Simple Estimator of Cointegrated Vectors
in Higher Order Integrated Systems,” Econometrica 61: 783-820.

Verdelhan, Adrien (2010), “A Habit-Based Explanation of the Exchange Rate Risk

Premium.” Journal of Finance, , 65(1), pp 123-145.

West, Kenneth D. (1996), “Asymptotic Inference about Predictive Ability”

Econometrica 64, pp. 1067-1084

29
Table 1: Evans-Lyons specification, 1999M02-2007M01

coefficient [1] [2] [3] [4] [5] [6] [7] [8]

USD/EUR USD/JPY
constant 0.003 -0.012 -0.009 -0.012 0.005 0.023 0.030 0.024
(0.003) (0.005) (0.005) (0.005) (0.006) (0.004) (0.007) (0.004)
Int. diff. -0.410 -0.270 -0.172 -0.186
(0.169) (0.182) (0.147) (0.145)
OF 1.179 1.080 1.182 1.799 1.807 1.857
(0.333) (0.333) (0.332) (0.301) (0.312) (0.296)
Δ(Int. diff.) 0.590 1.308
(0.988) 1.099

adj.R sq. 0.05 0.16 0.17 0.15 0.01 0.34 0.35 0.34
N 96 96 96 96 96 96 96 96
Notes: Dependent variable: First log difference of exchange rate, dollars per foreign currency
unit. OLS regression coefficients (Newey-West robust standard errors in parentheses). Bold
face denotes coefficients significant at the 10% marginal significance level. “Int. diff.” is the
money market interest differential, in decimal form, OF is net order flow measured in trillions
of USD.

30
Table 2.1: Johansen Cointegration Test Results, 1999M04-2007M01
[1] [2] [3] [4] [5] [6]
Monetary Fundamentals Hybrid
USD/EUR asy 0,0 0,0 0,0 3,1 3,0 3,0
fs 0,0 0,0 0,0 1,0 0,0 1,0

USD/JPY asy 2,2 0,1 0,0 2,2 0,1 0,0

fs 1,2 0,0 0,0 1,1 0,0 0,0

Notes: Implied number of cointegrating vectors using Trace, Maximal Eigenvalue statistics
and 1% marginal significance level. “Asy” (“fs”) denotes number of cointegrating vectors using
asymptotic (finite sample) critical values (Cheung and Lai, 1993). Columns [1] and [4] indicate
a constant is allowed in the cointegrating equation and none in the VAR; columns [2] and [5]
indicate a constant is allowed in the cointegrating equation and in the VAR; columns [3] and
[6] indicate an intercept and trend is allowed in the cointegrating equation and a constant in
the VAR. “Monetary fundamentals” include the exchange rate, money, income, and interest
differentials. “Hybrid” includes the exchange rate, money, income and interest differentials,
and cumulative order flow. Bold italics denote the specification with the lowest Akaike
Information Criterion for the single cointegrating vector case. All results pertain to
specifications allowing for 3 lags in the levels-VAR specification.

Table 2.2: Tests for weak exogeneity of order flow

[4] [5] [6]
Hybrid
USD/EUR Chi-Sq 3.465 0.481 0.041
p-val. 0.063 0.488 0.840

USD/JPY Chi-Sq 3.857 3.879 1.260

p-val. 0.050 0,049 0.262

Notes: For an explanation of the three specifications [4], [5] and [6], see the notes to Table
2.1. Likelihood ratio test statistic for restriction that order flow does not respond to
disequilibrium, distributed Chi-squared. [p-value for restriction α=0, in brackets].

Table 2.3: Tests for weak exogeneity of order flow, adjusted statistics
[4] [5] [6]
Hybrid
USD/EUR Chi-Sq 2.581 0.358 0.031
p-val. 0.108 0.5496 0.860

USD/JPY Chi-Sq 2.873 2.900 0.939

p-val. 0.090 0,089 0.333

Notes: For an explanation of the three specifications [4], [5] and [6], see the notes to Table
2.1. Likelihood ratio test statistic for restriction that order flow does not respond to
disequilibrium, adjusted for small sample (Bruggeman, 2002), distributed Chi-squared. [p-
value for restriction α=0, in brackets].

31
Table 3: USD/EUR Monetary/Order Flow Hybrid Exchange Rate Regression
Results, 1999M04-2007M01

coefficient [1] [2] [3] [4]

Error correction
term ‐0.059  ‐0.046 ‐0.085 ‐0.083
(0.038)  (0.029) (0.034) (0.036)
OF   1.889 1.873 1.802
  (0.351) (0.320) (0.324)
lag OF   0.637
      (0.429)
2nd lag OF       0.571†
  (0.297)
Long Run Coeffs.  
lag money ‐3.772  ‐3.772 ‐0.629 ‐0.629
(0.723)  (0.723) (1.413) (1.413)
lag income 5.558  5.558 5.160 5.160
(2.734)  (2.734) (2.149) (2.149)
lag int rate ‐7.835  ‐7.835 ‐3.220 ‐3.220
(3.013)  (3.013) (2.930) (2.930)
lag cumulative OF     0.306 0.306
  (0.093) (0.093)
       
adj.R sq. 0.023  0.341 0.365 0.381
SER 0.027  0.022 0.022 0.021
N 94  94 94 94
Q(6) 2.9751  3.9282 3.1184 1.5985
[0.812]  [0.686] [0.794] [0.953]
Q(12) 4.9065  6.2252 7.1515 4.5493
[0.961]  [0.904] [0.847] [0.971]
LM(6) 0.7687  1.4496 1.0080 0.5217
[0.597]  [0.208] [0.427] [0.790]

Notes: Dependent variable: First log difference of exchange rate, dollars per foreign currency
unit. Estimates from error correction model, estimated using nonlinear least squares, (Newey-
West robust standard error in parentheses), except for lagged long run coefficients, which are
estimated used DOLS(2,2), using Bartlett kernel, and Newey-West bandwidth set to 4.
Coefficients for first difference terms for monetary fundamentals not reported. OF is order flow
measured in trillions of USD. Adj-R sq., SER, and serial correlation diagnostics for error
correction regression. Q(6) and Q(12) are Box Q-statistics for test of serial correlation of order
6 and 12, respectively. LM(6) is the Breusch-Godfrey LM test statistics for serial correlation of
order 6. [p-values in brackets]. Bold face denotes significance at 10% msl.

32
Table 4: USD/JPY Monetary/Order Flow Hybrid Exchange Rate Regression
Results, 1999M04-2007M01

coefficient [1] [2] [3] [4]

Error correction
term ‐0.210  ‐0.155 ‐0.155 ‐0.154
(0.054)  (0.044) (0.046) (0.046)
OF   2.100 2.104 2.110
  (0.252) (0.252) (0.259)
lag OF       ‐0.129
      (0.311)
2nd lag OF   ‐0.140
  (0.444)
Long Run Coeffs.      
lag money ‐0.110  ‐0.110 0.077 0.077
(0.099)  (0.099) (1.747) (1.747)
lag income ‐1.864  ‐1.864 ‐1.876 ‐1.876
(0.523)  (0.523) (0.550) (0.550)
lag int rate ‐0.511  ‐0.511 ‐0.230 ‐0.230
(0.539)  (0.539) (1.241) (1.241)
lag cumulative OF   0.054 0.054
    (0.426) (0.426)
       
adj.R sq. 0.169  0.570 0.561 0.556
SER 0.023  0.017 0.017 0.017
N 94  94 94 94
Q(6) 4.2276  0.9741 1.6187 1.6120
[0.646]  [0.987] [0.951] [0.952]
Q(12) 8.6855  9.1572 9.9178 9.6516
[0.730]  [0.689] [0.623] [0.646]
LM(6) 2.0609  0.1653 0.4016 0.4828
[0.068]  [0.985] [0.876] [0.819]

Notes: Dependent variable: First log difference of exchange rate, dollars per foreign currency
unit. Estimates from error correction model, estimated using nonlinear least squares, (Newey-
West robust standard error in parentheses), except for lagged long run coefficients, which are
estimated used DOLS(2,2), using Bartlett kernel, and Newey-West bandwidth set to 4.
Coefficients for first difference terms for monetary fundamentals not reported. OF is order flow
measured in trillions of USD. Adj-R sq., SER, and serial correlation diagnostics for error
correction regression. Q(6) and Q(12) are Box Q-statistics for test of serial correlation of order
6 and 12, respectively. LM(6) is the Breusch-Godfrey LM test statistics for serial correlation of
order 6. [p-values in brackets]. Bold face denotes significance at 10% msl.

33
Table 5.1: USD/EUR Out of Sample Forecasting Performance, 2004M02-
07M01

model statistic 1 month 3 month 6 month

random
walk mean error -0.0012 -0.0053 -0.0106
std error 0.0040 0.0113 0.0198
monetary mean error -0.0166*** -0.0431*** -0.0945***
std error 0.0062 0.0142 0.0276
Theil 1.6185 1.6161 2.0532

Clark-West -0.5006 -0.0967 -0.1369

hybrid mean error 0.0010 0.0018 0.0027
std error 0.0048 0.0107 0.0185
Theil 1.0649 0.9423 0.9517
Clark-West 0.4114 0.4244 0.1151
Evans-
Lyons mean error -0.0101 -0.0243* -0.0616***
std error 0.0073 0.0145 0.0212
Theil 1.3993 1.3365 1.4834
Clark-West -1.1979 -1.2973 -1.9597**

Notes: Mean error for out-of-sample forecasting. Newey-West robust standard errors. ***(**)
denotes significance at 1%(5%) marginal significance level. Theil U-statistic is the ratio of the
model RMSE relative to random walk RMSE. A U-statistic > 1 indicates the model performs
worse than a random walk. Clark-West is the Clark-West statistic distributed Normal (0,1).
CW statistic > 0 indicates the alternative model outperforms a random walk.

Table 5.2: USD/JPY Out of Sample Forecasting Performance, 2004M02-

07M01

model statistic 1 month 3 month 6 month

random
walk mean error 0.0046 0.0107 0.0184
std error 0.0035 0.0087 0.0149
monetary mean error 0.0138*** 0.0301** 0.0474***
std error 0.0044 0.0085 0.0157
Theil 1.2714 1.2814 1.3380
Clark-West 1.1803 0.5100 0.1645
hybrid mean error -0.0012 0.0003 0.0059
std error 0.0065 0.0141 0.0276
Theil 1.1191 1.3607 1.6261
Clark-West 0.0355 0.2832 0.4513
Evans-
Lyons mean error 0.0011 0.0037 0.0054
std error 0.0041 0.0086 0.0151
Theil 0.7891 0.8059 0.8776
Clark-West 0.5583 0.7010 0.6301

Notes: Mean error for out-of-sample forecasting. Newey-West robust standard errors. ***(**)
denotes significance at 1%(5%) marginal significance level. Theil U-statistic is the ratio of the
model RMSE relative to random walk RMSE. A U-statistic > 1 indicates the model performs
worse than a random walk. Clark-West is the Clark-West statistic distributed Normal (0,1).
CW statistic > 0 indicates the alternative model outperforms a random walk.

34
 50,000 1,800,000

40,000 1,600,000

30,000 1,400,000

20,000 1,200,000

10,000 1,000,000

0 800,000

-10,000 600,000

-20,000 400,000
99 00 01 02 03 04 05 06

OFEURUSD VOLEURUSD

Figure 1: EUR/USD monthly order flow and order flow volume, in millions of euros. Order
flow, left axis; order flow volume, right axis.

60,000 1,100,000

50,000 1,000,000

40,000 900,000

30,000 800,000

20,000 700,000

10,000 600,000

0 500,000

-10,000 400,000

-20,000 300,000

-30,000 200,000
99 00 01 02 03 04 05 06

OFUSDJPY VOLUSDJPY

Figure 2: USD/JPY monthly order flow and order flow volume, in millions of dollars. Order
flow, left axis; order flow volume, right axis.

35
 .08

.06

.04

.02
DLXEU

.00

-.02

-.04

-.06
-20,000 0 20,000 40,000

Z1EU

Figure 3: First difference of log USD/EUR exchange rate and monthly net order flow in
millions of USD (purchases of euros)

.06

.04

.02

.00
DLXJP

-.02

-.04

-.06

-.08
-40,000 -30,000 -20,000 -10,000 0 10,000

Z1JP

Figure 4: First difference of log USD/JPY exchange rate and monthly net order flow in
millions of USD (purchases of yen)

36
 .08

.04

.00

-.04
.000
.025 -.08

.050
.075
.100
.125 Equation [1]
USD/EUR
.150
2000 2001 2002 2003 2004 2005 2006

One-Step Probability
Recursive Residuals

Figure 5: Recursive one step ahead recursive residuals for monetary model, USD/EUR.

.06
.04
.02
.00
-.02
-.04
.000
-.06
.025
.050
.075
.100
Equation [4]
.125
USD/EUR
.150
2000 2001 2002 2003 2004 2005 2006

One-Step Probability
Recursive Residuals

Figure 6: Recursive one step ahead recursive residuals for hybrid model, USD/EUR.

37
 .08

.04

.00

.000 -.04

.025
-.08
.050

.075

.100
Equation [1]
.125
USD/JPY
.150
2000 2001 2002 2003 2004 2005 2006

One-Step Probability
Recursive Residuals

Figure 7: Recursive one step ahead recursive residuals for monetary model, USD/JPY.

.06
.04
.02
.00
-.02

.000 -.04

.025 -.06

.050
.075
.100
.125 Equation [4]
USD/JPY
.150
2000 2001 2002 2003 2004 2005 2006

One-Step Probability
Recursive Residuals

Figure 8: Recursive one step ahead recursive residuals for hybrid model, USD/JPY.

38
.35

Actual
.30 Hybrid
Random
walk
.25

.20

.15

.10
Monetary

.05
I II III IV I II III IV I II III IV I
2004 2005 2006
Figure 9: Out-of-sample forecasts of USD/EUR, 3 month horizon

-4.55

Monetary
-4.60

Actual
-4.65

Hybrid
-4.70
Random
-4.75 walk

-4.80

-4.85
I II III IV I II III IV I II III IV I
2004 2005 2006
Figure 10: Out-of-sample forecasts of USD/JPY, 3 month horizon

39